Tesis Alexandra

Download as pdf or txt
Download as pdf or txt
You are on page 1of 298

!

Methods and Resources for


Sentiment Analysis in Multilingual
Documents of Different Text Types

Dissertation
Presented to the Department of Software and Computing
Systems, University of Alicante, in partial fulfillment of
the requirements for the title of

Doctor of Philosophy

Author: Alexandra Balahur Dobrescu


Supervisors: Prof. Dr. Andrs Montoyo
Dr. Ralf Steinberger

,PDJHQ3RUWDGD(/*5,72 IUDJPHQWR (GXDUG0XQFK 

Whereas our argument shows that the power and capacity of learning
exists in the soul already, and that just as the eye was unable to turn from
darkness to light without the whole body, so too the instrument of knowledge
can only by the movement of the whole soul be turned from the world of
becoming into that of being, and learn by degrees to endure the sight of
being and of the brightest and best of being, or in other words, of the good.
Plato, The Republic, Book 7, section 7 (The Myth of the Cave)

ACKNOWLEDGEMENTS
First of all, I would like to thank my main advisor Prof. Dr. Andrs Montoyo, for all
the knowledge and advice he has shared with me along these almost 4 years. I
would like to thank him, Prof. Dr. Manuel Palomar and Prof. Dr. Patricio MartnezBarco for all the support they have given me in the past years. I would like to thank
the University of Alicante for the PhD scholarship and all the people at the
Department of Software and Computing Systems, in the Natural Language
Processing Group, for their support and friendship. I would especially like to thank
Ester Boldrini and Elena Lloret, with whom I had a fruitful collaboration on
numerous occasions.
I would like to thank Dr. Ralf Steinberger, my co-advisor, for all the knowledge
and experience he has shared with me, as well as all the advice he has given me
during the past 2 years. Together with him, I would like to thank the entire
OPTIMA team at the European Commissions Joint Research Centre, led by Erik
van der Goot, for everything they taught me during my traineeship in Ispra (Italy),
for their valuable advice, their friendship and the collaboration in the development
of various resources and methods for sentiment analysis. I would especially like to
thank Dr. Mijail Kabadjov and Dr. Josef Steinberger, with whom I have
collaborated on various research papers in opinion summarization, during my
traineeship and after.
I would like to thank Prof. Dr. Dietrich Klakow for receiving me as an intern at
the Spoken Language Systems Department at the University of Saarbrcken
(Germany) and for all the knowledge he and his group shared with me during my
traineeship there. I would like to thank Dr. Michael Wiegand for his collaboration
on various sentiment analysis projects.
I would like to thank my wonderful family my mother Doina, my father Paul
and my sister Elvira for raising me to be curious, ambitious and perseverant, to
enjoy science and challenges, to never be afraid to ask, never be scared to look for
answers and always aim for excelsior. I would like to thank all my family and
friends, here, at home and in the world for their support and their advice.
I would like to thank Jess Hermida, a very special person in my life, for his
support, advice and the knowledge he has shared with me.
Finally, I would like to thank all the people that I have come across at different
events in the past years, who with their suggestions and critiques helped to improve
my research and/or life approach, as well as those who believed in me or doubted
me, making me wiser and stronger.
I

AGRADECIMIENTOS
En primer lugar, quisiera agradecer a mi tutor principal el Prof. Dr. Andrs
Montoyo, por todos los conocimientos y consejos que ha compartido conmigo a lo
largo de estos casi cuatro aos. Me gustara darle las gracias a l y a Prof. Dr.
Manuel Palomar y Prof. Dr. Patricio Martnez, por todo el apoyo que me han dado
en los ltimos aos. Gracias a la Universidad de Alicante por la beca de doctorado y
a toda la gente en el Departamento de Lenguajes y Sistemas Informticos, y en
especial al Grupo de Procesamiento del Lenguaje Natural, por su apoyo y amistad.
Me gustara dar las gracias a Ester Boldrini y Elena Lloret, con quien tuve una
fructfera colaboracin en numerosas ocasiones.
Me gustara agradecer al Dr. Ralf Steinberger, mi co-tutor, por todo el
conocimiento y la experiencia que ha compartido conmigo, as como todos los
consejos que me ha dado durante los ltimos 2 aos. Junto con l, me gustara dar
las gracias a todo el equipo de OPTIMA del Joint Research Centre de la Comisin
Europea, dirigido por el Erik van der Goot, por todo lo que me ense durante mi
periodo de prcticas en Ispra (Italia), por sus valiosos consejos, su amistad y
la colaboracin en el desarrollo de los diversos recursos y mtodos para el anlisis
de sentimientos. En especial me gustara dar las gracias a Dr. Mijail Kabadjov y Dr.
Josef Steinberger, con los que he colaborado en diversas investigaciones sobre las
tcnicas de resmenes de opinin, durante mi periodo de prcticas y despus.
Me gustara agradecer al Prof. Dr. Dietrich Klakow por darme la oportunidad de
hacer una estancia de investigacin en el Departamento de Sistemas de Habla (en
ingls Spoken Language Systems) de la Universidad de Saarbrcken (Alemania),
y por todo el conocimiento que su grupo y l mismo compartieron conmigo durante
mi periodo de prcticas. Me gustara dar las gracias tambin a Dr. Michael Wiegand
por su colaboracin en varios proyectos de anlisis de emociones, durante mi
estancia en Saarbrcken y despus.
Me gustara dar gracias a mi maravillosa familia mi madre Doina, mi padre
Paul y mi hermana Elvira por apoyar mi curiosidad, ambicin y perseverancia, por
ensearme a disfrutar de la ciencia y los desafos, a atreverme a preguntar, no tener
miedo a buscar respuestas y siempre tratar de buscar la excelencia en lo que hago,
llegar a un "excelsior". Me gustara dar las gracias a toda mi familia y amigos, aqu,
en casa y en el mundo por su apoyo y sus consejos.
Me gustara dar las gracias a Jess Hermida, una persona muy especial en mi
vida, por su apoyo, el asesoramiento y el conocimiento que ha compartido conmigo.
Por ltimo, me gustara dar las gracias a toda la gente que me he encontrado en
diferentes eventos en los ltimos aos, que con sus sugerencias y crticas ayudaron
III

a mejorar mi investigacin y/o enfoque de la vida. Asimismo, gracias a todos


aquellos que creyeron en m o dudaron de m, ayudndome as a ser ms sabia y
ms fuerte.

IV

ABSTRACT

The present doctoral thesis deals with the issues and challenges involved in the
development of methods and resources for the Natural Language Processing (NLP)
task of sentiment analysis.
Specifically, the first aim is to develop adequate techniques for the automatic
detection and classification of directly, indirectly or implicitly-expressed sentiment
in texts of different types (reviews, newspaper articles, dialogues/debates and
blogs), in different languages. The second aim is to apply the sentiment analysis
methods proposed in the context or jointly with other NLP tasks and propose
adequate techniques to tackle the issues raised by the peculiarities of affect
expression in these tasks.
In order to achieve the proposed objectives, the work presented has been
structured around answering five research questions. Following is a description of
the questions and a summary of the answers we have given in the present thesis.
1. How can sentiment analysis and, in a broader perspective, opinion mining
be defined in a correct way? What are the main concepts to be treated in
order to create a good definition that can be used to appropriately define
the task and subsequently propose correct methods to tackle it?
In Chapter 2, we define the main concepts we will frequently employ throughout
this thesis. We first present an overview of the definitions given in the NLP
literature to the related tasks of subjectivity analysis, sentiment analysis, opinion
mining, appraisal/attitude analysis, and emotion detection. We subsequently present
the definitions of the terms that are related to these tasks, both in well-established
dictionaries, as well as the research literature in the field. Finally, we propose an
operational definition that is consistent with the manner in which the different terms
related to sentiment analysis are defined. In Chapter 3, we present the state of the
art in the field and show that depending on the final aim of the application, the tasks
involving sentiment analysis are defined and tackled in a different manner.
The subsequent research questions we address in this thesis are:
2. Can sentiment analysis be performed using the same methods, for all text
types? What are the peculiarities of the different text types and how do they
influence the methods to be used to tackle it? Do we need special resources
for different text types?
V

3. Can the same language resources be used in other languages (through


translation)? How can resources be extended to other languages?
In Chapter 4, we present the peculiarities of different text types (reviews,
newspaper articles, blogs, political debates), analyze them and propose adequate
techniques to address them at the time of performing sentiment analysis. In the
cases where no generally-accepted definition of the sentiment analysis task exists
for a specific textual genre, we propose new definitions and annotate new resources
accordingly. We present different methods and resources we built for the task of
sentiment analysis in different text types, in different languages (English, Spanish,
German). In each of the genres studied, we evaluate our approaches
correspondingly, both in-house, as well as in international competitions. We show
that the techniques employed are robust enough to obtain good results, even in the
case where the original texts are in a language for which we do not have any
resources available and for the treatment of which we employ translation engines.
Finally, given the results obtained, we show that our approaches perform at the
level of state-of-the-art systems and in many cases outperform them.
4. How can we deal with opinion in the context of traditional tasks? How can
we adapt traditional tasks (Information Retrieval, Question Answering, Text
Summarization) in the context of opinionated content? What are the new
challenges in this context?
In Chapter 4, we only concentrate on the task of sentiment analysis as a
standalone challenge, omitting the steps required in order to obtain the texts on
which the sentiment analysis methods were applied or eliminating redundancy in
the information obtained. However, in a real-world application scenario,
automatically detecting the opinion expressed in a text is often not the first, neither
the last task to be performed. In order to analyze the sentiment found in different
texts, the documents must firstly be retrieved. Additionally, the results of the
automatic sentiment analysis may still contain a high volume of information, with
much redundancy. Bearing in mind these necessities, in Chapter 5 we study
methods to combine opinion mining with question answering and summarization.
We show that performing traditional tasks in the context of opinionated text has
many challenges and that systems that were designed to work exclusively with
factual data are not able to cope with opinion questions. Thus, we propose new
methods and techniques to adapt question answering and summarization systems to
deal with opinionated content. Additionally, we create and annotate appropriate
resources for the evaluation of the proposed methods. Finally, we evaluate our
approaches, as well as the impact of using different tools and resources in these
tasks. Our evaluations, both in in-house experiments, as well as through the
VI

participation in international competitions, show that the proposed methodologies


are appropriate for tackling question answering and summarization in the context of
opinionated texts.
The last research question we address in this thesis is:
5. Can we propose a model to detect emotion (as a component of sentiment)
from text, in the cases where it is expressed implicitly, requiring world
knowledge for its detection?
As we will see throughout this thesis, sentiments can be explicitly or implicitly
present in texts. While in the first case, lexical clues may be found in the text
indicating the presence of sentiment, through sentiment-bearing words, in the
second case, the emotion underlying the sentiment is not explicitly stated through
the use of affective words. In these situations, the emotion is only inferable based
on commonsense knowledge (i.e. emotion is not explicitly, but implicitly expressed
by the author, by presenting situations which most people, based on commonsense
knowledge, associate with an emotion, like going to a party, seeing your child
taking his/her first step etc.). Motivated by the fact that most work in sentiment
analysis has been done only in view of the existence of lexical clues for sentiment
detection and classification, and having seen the limitations of such models, in
Chapter 6 of the thesis, we present our contribution to the issue of automatically
detecting emotion expressed in text in an implicit manner. The initial approach is
based on the idea that emotion is triggered by specific concepts, according to their
relevance, seen in relation to the basic needs and motivations, underpinning our
idea on the Relevance Theory. The second approach we propose is based on the
Appraisal Theory models. The general idea behind it is that emotions are most of
the times not explicitly stated in texts, but results from the interpretation (appraisal)
of the actions contained in the situation described, as well as the properties of their
actors and objects. Thus, we set up a framework for representing situations
described in text as chains of actions (with their corresponding actors and objects),
and their corresponding properties (including the affective ones), according to
commonsense knowledge. We show the manner in which the so-called appraisal
criteria can be automatically detected from text and how additional knowledge on
the properties of the concepts involved in such situations can be imported from
commonsense knowledge bases. Finally, we demonstrate through an extensive
evaluation that such a representation is useful to obtain an accurate label of the
emotion expressed in text, without any linguistic clue being present therein,
increasing the recall of systems performing sentiment analysis from texts.

VII

TABLE OF CONTENTS
Chapter 1. Introduction............................................................................................. 1
1.1.

Background ............................................................................................... 1

1.2.

Motivation ................................................................................................. 5

1.3.

Applications............................................................................................... 7

Chapter 2. Tasks and concepts ................................................................................ 11


2.1. Subjectivity................................................................................................... 12
2.2. Opinion, sentiment, emotion. Computationally-related tasks. ..................... 13
2.3. Attitude and appraisal. Appraisal theories.................................................... 19
Chapter 3. State of the art ........................................................................................ 23
3.1. Subjectivity analysis ..................................................................................... 23
3.1.1. Resource creation for subjectivity analysis ........................................... 24
3.1.2. Applications of subjectivity analysis ..................................................... 27
3.2. Sentiment analysis (opinion mining) ............................................................ 28
3.2.1. Creation of resources for sentiment analysis ......................................... 29
3.2.2. Research in text sentiment polarity classification ................................. 31
3.2.3. Extraction of opinion from text ............................................................. 36
3.3. Applications of opinion mining to traditional tasks Opinion Question
Answering, Opinion Summarization ................................................................... 37
3.3.1. Opinion Question Answering ................................................................ 37
3.3.2. Opinion Summarization......................................................................... 39
3.4. Related tasks: emotion detection as a general AI and NLP challenge ......... 40
Chapter 4. Methods and resources for multilingual sentiment analysis in different
text types ................................................................................................................. 43
4.1. Opinion mining from product reviews feature-based opinion mining and
summarization ..................................................................................................... 45
4.1.1. Introduction ........................................................................................... 45
4.1.2. Methods for mining sentiment from product reviews ........................... 47
4.1.3. Classifying opinion on products using relatedness scores on specific
corpora and emotion expression patterns ........................................................ 58
4.1.4. Further improvements to sentiment analysis in reviews the issue of
Disambiguating Adjective Polarity using the local context ......................... 63
IX

4.1.5. Creating an annotation scheme for feature-based opinion mining ........ 66


4.1.6. Building a recommender system using feature-based opinion mining.. 76
4.2. Opinion mining from newspaper articles ..................................................... 77
4.2.1. Introduction ........................................................................................... 77
4.2.2. Initial experiments ................................................................................. 79
4.2.3. Preliminary conclusions from initial experiments ................................. 82
4.2.4. Redefining sentiment analysis from newspaper articles........................ 82
4.2.5. Experiments applying the new task definition of sentiment analysis from
newspaper articles ........................................................................................... 90
4.2. . Conclusions on sentiment analysis from newspaper articles................. 93
4.3. Opinion mining from political debates ......................................................... 94
4.3.1. Introduction ........................................................................................... 94
4.3.2. Background ........................................................................................... 95
4.3.3. Experiments ........................................................................................... 96
4.3.4. Evaluation............................................................................................ 101
4.3.5. Discussion and conclusions on sentiment analysis from political debates
....................................................................................................................... 105
4.4. Opinion mining from blogs ........................................................................ 106
4.4.1. Introduction ......................................................................................... 106
4.4.2. The EmotiBlog annotation model ....................................................... 107
4.4.3. Experiments and evaluations of EmotiBlog on English corpora ......... 109
4.4.4. Using EmotiBlog to detect and classify sentiment in Spanish texts .... 113
4.5. Conclusions on the proposed methods for sentiment analysis ................... 116
Chapter 5. Applications of sentiment analysis ...................................................... 119
5.1. Introduction ................................................................................................ 119
5.2. Opinion Question Answering and Summarization ..................................... 120
5.2.1. Participation to the TAC 2008 Opinion Pilot task ............................... 120
5.2.2. Post TAC 2008 competition experiments and results ......................... 127
5.2.3. Proposal of a Framework for opinion question answering .................. 132
The question analysis phase .......................................................................... 149
5.2.4. Evaluation of the proposed Opinion Question Answering framework 151
5.2.5. Opinion Question Answering with temporal restrictions participation
in the NTCIR 8 MOAT ................................................................................. 154
X

5.2.6. Conclusions ......................................................................................... 160


5.3. Opinion-oriented summarization for mass opinion estimation .................. 161
5.3.1. Introduction and background ............................................................... 161
5.3.2. Initial experiments ............................................................................... 162
5.3.3. Summarization based on opinion strength........................................... 168
5.3.4. Opinion summarization using a topic-sentiment analysis approach.... 172
5.3.5. Topic-sentiment analysis versus summarization as first step in opinion
summarization ............................................................................................... 176
5.3.6. Conclusions on the proposed approaches for Opinion Summarization 178
Chapter 6. Discovering implicit expressions of sentiment from text ................... 179
6.1. Introduction ................................................................................................ 179
6.1.1. Motivation ........................................................................................... 180
6.1.2. The concept of emotion ....................................................................... 180
6.1.3. Background ......................................................................................... 181
6.2. Emotion detection based on emotion triggers ......................................... 182
6.2.1. Definition of emotion triggers.......................................................... 182
6.2.2. Theories underlying the emotion triggers approach......................... 183
6.2.3. Building the emotion triggers lexicon .............................................. 185
6.2.4. Assignment of polarity and emotion to the emotion triggers .............. 186
6.2.5. A method for emotion detection in text based on emotion triggers . 187
6.2.6. Implementation of the emotion detection method based on emotion
triggers ......................................................................................................... 189
6.2.7. Evaluation of the emotion detection method based on emotion triggers
....................................................................................................................... 190
6.3. Emotion detection from text using appraisal criteria based on commonsense
knowledge ......................................................................................................... 191
6.3.1. Introduction ......................................................................................... 191
6.3.2. Context of research .............................................................................. 192
6.3.3. Issues in present approaches ................................................................ 195
6.3.4 A Method for Modeling Affective Reaction Using Appraisal Models 197
6.3.5 Building the EmotiNet Knowledge Base.............................................. 199
6.3.6 Experiments and Evaluation ................................................................. 207
6.4. Conclusions on methods for implicit sentiment detection.......................... 214
XI

Chapter 7. Contributions ....................................................................................... 215


Chapter 8. Future Work ......................................................................................... 223
References ............................................................................................................. 227
ANNEX A. Scientific contributions and merits during the PhD Studies ............. 249
1.

Scientific publications: .............................................................................. 249


Book chapters: ............................................................................................... 249
Journal articles:.............................................................................................. 249
Articles in conference proceedings: .............................................................. 250

2.

Participation in scientific competitions: .................................................... 253

3.

Participation in projects: ............................................................................ 253

4.

Participation in Program Committees: ...................................................... 254

5.

Organization of scientific events: .............................................................. 255

6.

Awards and scholarships obtained: ........................................................... 255

ANNEX B. Resumen en Castellano ...................................................................... 257


Introduccin y motivacin................................................................................. 257
Problemas del dominio y objetivos de la investigacin .................................... 259
Contribucin ...................................................................................................... 262
Conclusiones y trabajos futuros ........................................................................ 269

XII

CHAPTER 1. INTRODUCTION
Motto: Human behavior flows from three main sources: desire, emotion and
knowledge. (Plato)

1.1.

BACKGROUND

The era in which we live has been given many names. Global village,
technotronic era, post-industrial society, information society, information
age, and knowledge society are just a few of the terms that have been used in an
attempt to describe the deep changes that have occurred in the lives of societies and
people worldwide as a result of the fast development of ICT technologies, the
access to Internet and its transformation into a Social Web. In this new context,
having access to large quantities of information is no longer an issue, as there are
terabytes of new information produced on the Web every day that are available to
any individual with an Internet connection. In contrast to older times, when finding
sources of information was the key problem to companies and individuals, todays
information society challenges companies and individuals to create and employ
mechanisms to search and retrieve relevant data from the huge quantity of
available information and mine it to transform it into knowledge, which they can
use to their advantage. As opposed to the past, when this advantage was a question
of finding sources of information, in todays society, which is flooded by data that
is changing at a rapid pace, the advantage is given by the quality (accuracy,
reliability) of the extracted knowledge and its timeliness. For the era in which we
live, information has become the main trading object. In this context, having at
hand high quality and timely information is crucial to all the spheres of human
activity: social, political, and economic, to name just a few.
However, in many cases, the relevant information is not found in structured
sources (i.e. tables or databases), but in unstructured documents, written in human
language. The high quantity of such data requires the use of automatic processing
techniques. The discipline that deals with the automatic treatment of natural
language in text or speech is called Natural Language Processing (NLP). NLP is
part of the research area of Artificial Intelligence (AI), which is defined as the
science and engineering of making intelligent machines (McCarthy, 1959), by
simulating the mechanisms of human intelligence. The goal of Artificial
Intelligence, as it was stated in the 1950s, is to create machines that are capable of
passing the Turing Test. Roughly speaking, a computer will have passed the
Turing Test if it can engage in conversations indistinguishable from that of a
1

human's (Lee, 2004). In order to achieve this goal, NLP deals with the text
analysis at different levels: phonologic (sounds), lexical (words), morphologic
(parts of speech), syntactic (representation of the structure of the sequence of
lexical units based on their dependency), semantic (logical structure representing
the meaning expressed) and pragmatic (studying the influence of the context and
the world knowledge on the general meaning of the text). NLP contains many
research areas. Each of these constitute either general NLP problems, which need to
be solved in any application areas (Word Sense Disambiguation, Co-reference
resolution ), or that have been set up in the view of a specific end application
(Information Retrieval, Information Extraction, Question Answering, Text
Summarization, Machine Translation).
Traditionally, the application areas of NLP were designed for the treatment of
factual (exact) data. Nowadays, however, factual information is no longer the main
source from which crucial knowledge is extracted.
The present is marked by the growing influence of the Social Web (the web of
interaction and communication) on the lives of people worldwide. More than ever
before, people are more than willing and happy to share their lives, knowledge,
experience and thoughts with the entire world, through blogs, forums, wikis, review
sites or microblogs. They are actively participating to events, by expressing their
opinions on them, by commenting on the news appearing and the events that take
place in all spheres of the society. The large volume of subjective information
present on the Internet, in reviews, forums, blogs, microblogs and social network
communications has produced an important shift in the manner in which people
communicate, share knowledge and emotions and influence the social, political and
economic behavior worldwide. In consequence, this new reality has led to
important transformations in the manner, extent and rapidness in which news and
their associated opinions circulate, leading to new and challenging social,
economical and psychological phenomena.
In order to study these phenomena and address the issue of extracting the crucial
knowledge that nowadays is contained in opinionated data, new fields of research
were born in Natural Language Processing (NLP), aiming at detecting subjectivity
in text and/or extracting and classifying opinions into different sets (usually
positive, negative and neutral). The main tasks that were tackled in NLP are
subjectivity analysis (dealing with private states (Banfield, 1982), a term that
encloses sentiment, opinions, emotions, evaluations, beliefs and speculations)
sentiment analysis and opinion mining, although different terminologies have
been used to denote the approaches taken (e.g. review mining, appraisal extraction)
and sentiment analysis and opinion mining have been used interchangeably, as they
are considered by some authors to point to the same task (Pang and Lee, 2008). A
closely related task is also emotion detection, dealing with the classification of
2

texts according to the emotion expressed. All these research areas are part of the
wider field in Artificial Intelligence denominated affective computing (Picard,
1995).
This thesis deals with the task of sentiment analysis, in the context of
multilingual documents of different text types. Specifically, the work we will
present throughout the following chapters concentrates on answering the following
research questions:
1. How can sentiment analysis and, in a broader perspective, opinion mining
be defined in a correct way? What are the main concepts to be treated in
order to create a good definition that can be used to appropriately
delimitate the task and subsequently propose correct methods to tackle it?
In Chapter 2, we define the main concepts we will frequently employ throughout
this thesis. We first present an overview of the definitions given in the NLP
literature to the related tasks of subjectivity analysis, sentiment analysis, opinion
mining, appraisal/attitude analysis, and emotion detection. We subsequently present
the definitions of the terms that are related to these tasks, both in well-established
dictionaries, as well as the research literature in the field. Finally, we propose an
operational definition that is consistent with the manner in which the different terms
related to sentiment analysis are defined. In Chapter 3, we present the state of the
art in the field and show that depending on the final aim of the application, the tasks
involving sentiment analysis are defined and tackled in a different manner.
The subsequent research questions we address in this thesis are:
2. Can sentiment analysis be performed using the same methods, for all text
types? What are the peculiarities of the different text types and how do they
influence the methods to be used to tackle it? Do we need special resources
for different text types?
3. Can the same language resources be used in other languages (through
translation)? How can resources be extended to other languages?
In Chapter 4, we present the peculiarities of different text types (reviews,
newspaper articles, blogs, political debates), analyze them and propose adequate
techniques to address them at the time of performing sentiment analysis. In the
cases where no generally-accepted definition of the sentiment analysis task exists
for a specific textual genre, we propose new definitions and annotate new resources
accordingly. We present different methods and resources we built for the task of
sentiment analysis in different text types, in different languages (English, Spanish,
German). In each of the genres studied, we evaluate our approaches
correspondingly, both in-house, as well as in international competitions. We show
3

that the techniques employed are robust enough to obtain good results, even in the
case where the original texts are in a language for which we do not have any
resources available and for the treatment of which we employ translation engines.
Finally, given the results obtained, we show that our approaches perform at the
level of state-of-the-art systems and in many cases outperform them.
4. How can we deal with opinion in the context of traditional tasks? How can
we adapt traditional tasks (Information Retrieval, Question Answering, Text
Summarization) in the context of opinionated content? What are the new
challenges in this context?
In Chapter 4, we only concentrate on the task of sentiment analysis as a
standalone challenge, omitting the steps required in order to obtain the texts on
which the sentiment analysis methods were applied or eliminating redundancy in
the information obtained. However, in a real-world application scenario,
automatically detecting the opinion expressed in a text is often not the first, neither
the last task to be performed. In order to analyze the sentiment found in different
texts, the documents must firstly be retrieved. Additionally, the results of the
automatic sentiment analysis may still contain a high volume of information, with
much redundancy. Bearing in mind these necessities, in Chapter 5 we study
methods to combine opinion mining with question answering and summarization.
We show that performing traditional tasks in the context of opinionated text has
many challenges and that systems that were designed to work exclusively with
factual data are not able to cope with opinion questions. Thus, we propose new
methods and techniques to adapt question answering and summarization systems to
deal with opinionated content. Additionally, we create and annotate appropriate
resources for the evaluation of the proposed methods. Finally, we evaluate our
approaches, as well as the impact of using different tools and resources in these
tasks. Our evaluations, both in in-house experiments, as well as through the
participation in international competitions, show that the proposed methodologies
are appropriate for tackling question answering and summarization in the context of
opinionated texts.
The last research question we address in this thesis is:
5. Can we propose a model to detect emotion (as a component of sentiment)
from text, in the cases where it is expressed implicitly, requiring world
knowledge for its detection?
As we will see throughout this thesis, sentiments can be explicitly or implicitly
present in texts. While in the first case, lexical clues may be found in the text
indicating the presence of sentiment, through sentiment-bearing words, in the
second case, the emotion underlying the sentiment is not explicitly stated through
4

the use of affective words. In these situations, the emotion is only inferable based
on commonsense knowledge (i.e. emotion is not explicitly, but implicitly expressed
by the author, by presenting situations which most people, based on commonsense
knowledge, associate with an emotion, like going to a party, seeing your child
taking his/her first step etc.). Motivated by the fact that most work in sentiment
analysis has been done only in view of the existence of lexical clues for sentiment
detection and classification, and having seen the limitations of such models, in
Chapter 6 of the thesis, we present our contribution to the issue of automatically
detecting emotion expressed in text in an implicit manner. The initial approach is
based on the idea that emotion is triggered by specific concepts, according to their
relevance, seen in relation to the basic needs and motivations, underpinning our
idea on the Relevance Theory. The second approach we propose is based on the
Appraisal Theory models. The general idea behind it is that emotions are most of
the times not explicitly stated in texts, but results from the interpretation (appraisal)
of the actions contained in the situation described, as well as the properties of their
actors and objects. Thus, we set up a framework for representing situations
described in text as chains of actions (with their corresponding actors and objects),
and their corresponding properties (including the affective ones), according to
commonsense knowledge. We show the manner in which the so-called appraisal
criteria can be automatically detected from text and how additional knowledge on
the properties of the concepts involved in such situations can be imported from
commonsense knowledge bases. Finally, we demonstrate through an extensive
evaluation that such a representation is useful to obtain an accurate label of the
emotion expressed in text, without any linguistic clue being present therein,
increasing the recall of systems performing sentiment analysis from texts.

1.2.

MOTIVATION

The radical shift in the method employed for communication and the content of this
communication has brought with itself new challenges, but also many opportunities.
At the economic level, the globalization of markets combined with the fact that
people can freely express their opinion on any product or company on forums,
blogs or e-commerce sites led to a change in the companies marketing strategies, in
the rise of awareness for client needs and complaints, and a special attention for
brand trust and reputation. Specialists in market analysis, but also IT fields such as
Natural Language Processing, demonstrated that in the context of the newly created
opinion phenomena, decisions for economic action are not only given by factual
information, but are highly affected by rumors and negative opinions. Wright
5

(2009) claims that for many businesses, online opinion has turned into a kind of
virtual currency that can make or break a product in the marketplace 1. Studies
showed that financial information presented in news articles have a high correlation
to social phenomena, on which opinions are expressed in blogs, forums or reviews.
On the other hand, many tasks that involved extensive efforts from the companies
marketing departments are easier to perform. An example is related to market
research for advertising, business intelligence and competitive vigilance. New
forms of expression on the web made it easier to collect information of interest,
which can help to detect changes in the market attitude, discover new technologies,
machines, markets where products are needed and detect threats. On the other hand,
using the opinion information, companies can spot the market segments their
products are best associated with and can enhance their knowledge on the clients
they are addressing and on competitors. The analysis of the data flow on the web
can lead to the spotting of differences between the companies products and the
necessities expressed by clients and between the companies capacities and those of
the competitors. Last, but not least, the interpretation of the large amounts of data
and their associated opinions can give companies the capacity to support decisionmaking through the detection of new ideas and new solutions to their technological
or economic problems.
The opinionated data on the web has also produced important changes in the
manner in which communities are able to participate in the elaboration of laws and
policies. Consultations with communities that in the past were made through the use
of questionnaires are now easily made through forums of opinion. Additionally,
such data on the opinions that people have about laws, policies, and administrators
can be extracted from a variety of sources (e.g. microblogs, blogs, social networks).
The advantage and, at the same time, issue related to these new capabilities is
the large amount of information available and its fast growing rate. Lack of
information on markets and their corresponding social and economical data, leads
to wrong or late decisions and finally to important financial losses. Lack of
information of the policy makers leads to wrong decisions, affecting large
communities.
Although mostly positive, there are also downsides to the increasing
communication through the use of Web 2.0 technologies. The development of
social networks and communication between their members led to the development
of interesting phenomena, whose effects are both positive and negative and which
are difficult to assess. Within social networks gathered around the most peculiar
topics, people talk about subjects that they would not address in their everyday life
and with their friends or family. Under the hidden identity on the web, however,
they are free to express their innermost fears and desires. That is why, allowing and
1

www.nytimes.com/2009/08/24/technology/internet/24emotion.html?_r=1&ref=start-ups

supporting free communication led to the birth of sites where violence is predicated
and encouraged, where people with psychological problems or tendencies towards
suicide, addictions etc. talk to one another and encourage their negative behaviors.
Such sites must be discovered and controlled, in order to keep under control
different social issues that may arise from the described potentially conflictive
situations.
As we can see, the present reality is profoundly marked by the opinionated
information present both in traditional, as well as new textual genres. Given the
proven importance of such data, but also the challenges it raises, in terms of
volume, automatic systems, sentiment analysis has become a highly active research
field in NLP in the past years.
In the next section, we will show that this task is not only useful in order to
obtain important knowledge from non-factual data, but that it also contributes to the
improvement of systems dealing with other NLP challenges.

1.3.

APPLICATIONS

In the Motivation Section, we explained the main reasons for doing research in the
field of sentiment analysis and the applications it has in real-world scenarios.
Further on, we will present some of the fields and domains in which this task is
useful.
Research has been conducted in the field of opinion mining, aimed at improving
different social, economical, political and psychological aspects of every-day
human life. There are many applications of opinion mining systems to real-world
scenarios. Some of these applications are already available online, others are still
under research and other directions and developments in the field are merely
appearing. There are sites like swotty.com, which mine, classify and summarize
opinions from reviews on products on the e-commerce sites that people can use for
comparison, advice or recommendation. Other applications of the task, directly
related to the commerce and competition markets of companies, use opinion mining
from the web to obtain direct, sincere and unbiased market feedback, about their
own products (business intelligence), and of the products of their market
competition (competitive vigilance). Companies, as well as public figures, use
opinion mining to monitor their public image and reputation (trust). Authors can
benefit from opinion mining to track their literary reputation.
It was demonstrated that fluctuation in public opinion correlates to fluctuations
of stock prices for the targeted companies (Devitt and Ahmad, 2007). Thus, opinion
mining can be used to track opinion across time for market and financial studies, for
early action in predicted crisis situations or for the issuing of alerts.
7

Recent developments of the Social Web technologies and the growing number
of people writing and reading social media (blogs, forums etc.) also allows for the
monitoring and analysis of social phenomena, for the spotting of potentially
dangerous situations and determining the general mood of the blogosphere.
Examples of sites implementing these concepts are wefeelfine.org or
twends.com.
Yet another application of sentiment analysis is the tracking of political view, to
detect consistency and inconsistency between statements and action at the
government level. It was recently stated that election results could be better
predicted by following the discussion threads in blogs.
eRulemaking, as a democratic way of consulting the whole targeted population
when a law or policy is to be implemented, can also highly benefit from sentiment
analysis, as method to spot and classify a large quantity of subjective data. This task
is also performed when tracking views on laws from legal blogs (blawgs).
Last, but not least, studying affect related phenomena is basic for HumanComputer Interaction (Picard, 1995), as most reactions and interactions are not only
rationality-based, but heavily rely on emotion.
It was also demonstrated that opinion mining improves other Natural Language
Processing tasks, such as:
! Information Extraction, by separating facts from opinions (Riloff et al.,
2005);
! Question Answering (Somasundaran et al., 2007), where the application of
opinion mining can improve the answering of definition questions (Lita et
al., 2005)
! and Multi-Perspective Question Answering (Stoyanov et al., 2005; Yu and
Hatzivassiloglou, 2003) where there is not a single, true and correct answer,
but a set of answers describing the attitude of different persons on a given
fact;
! Summarization of multi-perspective texts (Ku et al., 2005; Ku et al., 2006),
where redundant opinion (opinion of the same polarity, given the same
arguments) must be removed;
! Authorship (source) determination (Teufel and Moens, 2000; Piao et al.,
2007);
! Word Sense Disambiguation (Wiebe and Mihalcea, 2006).
The next chapters of this thesis present different methods and approaches for
tackling the task of sentiment analysis in different text types, languages and in the
context of a variety of final applications, addressing the five research questions we
described. First of all, however, in order to ensure an understanding of the
terminology we will employ, in Chapter 2 we present an overview of the tasks and
related concepts definition. The main motivations for defining the concepts
8

involved in this task is that the issues related to the study of affective phenomena
have been studied for a long time in disciplines such as Psychology or Philosophy
and that sentiment analysis in NLP is a recent field, in which the terminology is not
yet fully established.

CHAPTER 2. TASKS AND CONCEPTS


Motto: How much has to be explored and discarded before reaching the naked
flesh of feeling?(Claude Debussy)
Having seen the high number of practical applications to the automatic processing
of subjective and opinionated language, it is of no wonder that the tasks of
subjectivity and sentiment analysis have registered a growing interest from the NLP
research community in the past few years. Vast amounts of research has been
performed, all falling within the scope of developing computational methods for
text analysis, in order to discover whether it is subjective or objective, whether it
contains opinions, sentiments, attitudes (and if so, what polarity/tonality/orientation
these have) or emotions.
Due to the large number of applications in the social, political and economic
spheres, most work has concentrated on creating and evaluating methods, tools and
resources to discover whether a specific object (person, product, organization,
event, etc.) is regarded2 in a positive or negative manner by a specific source
(i.e. a person, an organization, a community, people in general, etc.). This task has
been given many names, from opinion mining, to sentiment analysis, review
mining, attitude analysis, appraisal extraction and many others. At the same time,
the term opinion mining, which authors such as Pang and Lee (2008) consider
being equivalent to sentiment analysis, has been employed to denote work that
aims at classifying text according to different criteria:
a) the polarity of the sentiment expressed (into positive and negative;
sometimes the neutral category is also employed);
b) whether the text includes good or bad news (Ku et al., 2005);
c) whether the candidate that the text is talking about is likely or unlikely
to win (Kim and Hovy, 2005);
d) whether the text is expression support or opposition (Bansal et al.,
2008; Terveen et al., 1997) ;
e) pros and cons (Kim and Hovy, 2006);
f) determining the polarity of the outcome (e.g. improvement versus death
in medical texts) (Niu et al., 2005)
g) Whether a person agrees or disagrees with a topic (political debates)
(Balahur et al., 2009e).

2
Regard is a term we deliberately use here, in order to avoid employing any of the terminology used
so far, which is defined and detailed in this chapter. In this context, it refers to: 1) an assessment based
on a set of criteria (i.e. personal taste, convenience, social approval, moral standards etc.) or 2) on the
emotional effect it has on the person as a cause of this assessment.

11

Given the high dynamics of the field and the vast amount of research done
within its framework in the past few years, the terminology employed in defining
the different tasks dealing with subjectivity, as well as the concepts involved in
them is not yet uniform across the research community.
In order to establish the scope of our research and ensure the understanding of
the approaches presented, we first give an overview of the definitions given to the
concepts involved outside the area of Natural Language Processing. Subsequently,
we present some of the tasks proposed in the research, the concepts involved in
them and the definitions they were given. Finally, we propose a set of definitions to
the concepts we will use across this thesis and of the different tasks aiming at the
automatic processing of subjective texts.

2.1. SUBJECTIVITY
In Philosophy, subjectivity refers to the subject and his or her perspective, feelings,
beliefs, and desires. (Solomon, 2005).
In NLP, the most widely used definition is the one proposed by Wiebe (1994).
The author defines subjectivity as the linguistic expression of somebodys
opinions, sentiments, emotions, evaluations, beliefs and speculations. In her
definition, the author was inspired by the work of the linguist Ann Banfield
(Banfield, 1982), who defines as subjective the sentences that take a characters
point of view (Uspensky, 1973) and that present private states (Quirk, 1985) (that
are not open to objective observation or verification) of an experiencer, holding an
attitude, optionally towards an object. Subjectivity is opposed to objectivity, which
is the expression of facts. Wiebe et al. (2005) considers the term private state,
which is in a pragmatic sense equivalent to subjectivity and that is defined as a
general term that covers opinions, beliefs, thoughts, feelings, emotions, goals,
evaluations, and judgments. According to the definition proposed by Wiebe (1994),
an example of subjective sentence is This book is amazing!, whereas an example
of objective sentence is This book costs 10 on Amazon.com.
In view of the given definition, the Multi-Perspective Question Answering
corpus (Wiebe et al., 2005) takes into account three different types of elements for
the annotation of subjectivity: explicit mentions of private states (e.g. The U.S.
fears a spill-over, said Xirao-Nima), speech events expressing private states (e.g.
The U.S. fears a spill-over,said Xirao-Nima), expressive subjective element (e.g.
The report is full of absurdities,).
In the Handbook of Natural Language Processing (2010), Bin Liu defines
subjective versus objective sentences as follows: An objective sentence expresses

12

some factual information about the world, while a subjective sentence expresses
some personal feelings or beliefs.

2.2. OPINION, SENTIMENT, EMOTION.


COMPUTATIONALLY-RELATED TASKS.
In the case of opinion, if one were to look at the term definition given in the
Webster dictionary3 , they would find the following set of synonyms: opinion,
view, belief, conviction, persuasion, sentiment, meaning a judgment
one holds as true. Out of this definition, it is important to stress upon the fact that
these closely related, synonym terms, have slightly different meanings.
Opinion implies a conclusion thought out yet open to dispute; it is:
1. A): a view, judgment, or appraisal formed in the mind about a
particular matter; B): approval, esteem;
2. A): a belief stronger than impression and less strong than positive
knowledge; B): a generally held view;
3. A): a formal expression of judgment or advice by an expert; B): the
formal expression (as by a judge, court, or referee) of the legal reasons
and principles upon which a legal decision is based.
View suggests a subjective opinion.
Belief implies often deliberate acceptance and intellectual assent.
Conviction applies to a firmly and seriously held belief.
Persuasion suggests a belief grounded on assurance (as by evidence) of its
truth.
Sentiment suggests a settled opinion reflective of ones feelings.
The term feeling is defined as the conscious subjective experience of emotion.
(Van den Bos, 2006). This is approximately the same definition as the one given by
Scherer (2005), which states that the term feeling points to a single component of
emotion, denoting the subjective experience process, and is therefore only a small
part of an emotion.
This definition suggests that there are different types of opinions and that not all
opinions are subjective (see the definition of view), as well as not all opinions
have a sentiment associated to them. An objective opinion could be considered to
be the one of an expert (e.g. a doctor giving a diagnosis on the basis of observed
symptoms). A subjective opinion is one that is based on personal criteria
(depends on the individual taste, ideas, standards etc.). This same definition also
pinpoints to the fact that sentiments are types of opinions, namely the ones that are
reflective of ones feelings, where feeling is the conscious subjective
3

http://www.merriam-webster.com/

13

experience of emotion. Thus, sentiment relates to emotion, in the sense that it is


the expression of an evaluation based on the emotion the writer feels.
Opinion mining, as a computational task, appeared for the first time in a
paper by Dave et al. (2003), and it was defined as follows: Given a set of
evaluative text documents D that contain opinions (or sentiments) about an
object (person, organization, product etc.), opinion mining aims to extract
attributes and components of the object that have been commented on in each
document d in the set D and to determine whether the comments are positive,
negative or neutral. According to Pang and Lee (2008), the fact that this work
appeared in the proceedings of the World Wide Web (WWW) 2003 conference
explains the popularity of this terminology within the web search and retrieval
research community. This also explains the fact that Esuli and Sebastiani (2006)
define opinion mining as a recent discipline at the crossroads of information
retrieval and computational linguistics which is concerned not with the topic a
document is about, but with the opinion it expresses.
From the computational point of view, Kim and Hovy (2005) define opinion
as a quadruple [Topic, Holder, Claim, Sentiment] in which the Holder believes
a Claim about the Topic, and in many cases associates a Sentiment, such as
good or bad, with the belief. As far as sentiments are concerned, the authors define
them as: Sentiments, which in this work we define as an explicit or implicit
expression in text of the Holders positive, negative, or neutral regard toward
the Claim about the Topic. Sentiments always involve the Holders emotions or
desires, and may be present explicitly or only implicitly.
This definition relates opinion with sentiment, in the sense that it states that
some opinions carry a sentiment, while others do not. In order to illustrate the
difference between opinions with sentiment and opinions without sentiment, Kim
and Hovy (2005) provide the following examples:
(1) I believe the world is flat.
(2) The Gap is likely to go bankrupt.
These are sentences that express opinions, but they do not contain any sentiment.
The following examples, taken from the same paper, explain the difference between
explicitly versus implicitly expressed sentiment of opinions:
(3) I think that attacking Iraq would put the US in a difficult position.
(implicit)
(4) The US attack on Iraq is wrong. (explicit)
(5) I like Ike. (explicit)
(6) We should decrease our dependence on oil. (implicit)
Another definition of the term opinion was given by Bing Liu (2010). The
author is the one who defined the task of feature-based opinion mining and

14

summarization, which deals with the classification of opinions expressed on


different features of products and their summarization (Hu and Liu, 2004).
According to Liu (2010):
! An opinion on a feature f is a positive or negative view, attitude, emotion
or appraisal on f from an opinion holder.
! The holder of an opinion is the person or organization that expresses the
opinion.
! An explicit opinion on feature f is an opinion explicitly expressed on f in a
subjective sentence.
! An implicit opinion on feature f is an opinion on f implied in an objective
sentence.
! An opinionated sentence is a sentence that expresses explicit or implicit
positive or negative opinions. It can be a subjective or objective sentence.
! Emotions are our subjective feelings and thoughts.
All tasks defined within opinion mining aim at classifying the texts according to
the orientation of the opinion (usually into three classes of positive, negative
and neutral). The classes of opinion considered have been denoted using different
terms: opinion orientation, sentiment polarity, polarity, sentiment orientation,
polarity of opinion, semantic orientation.
As far as sentiment analysis as NLP task is concerned, most of the research in
the field coincides with the following definition: The binary classification task of
labeling an opinionated document as expressing either an overall positive or an
overall negative opinion is called sentiment polarity classification or polarity
classification. (Pang and Lee, 2008)
The orientation of an opinion on a feature f indicates whether the opinion is
positive, negative or neutral. Opinion orientation is also known as sentiment
orientation, polarity of opinion, or semantic orientation.(Liu, 2010)
A related concept is valence, defined as a negative or positively attitude
(Polanyi and Zaenen, 2004). In relation to this concept, Polanyi and Zaenen (2004)
define the so-called contextual valence shifters (e.g. negatives and intensifiers,
modals, presuppositional items, ironical formulations, connectors), which are
lexical items or formulations that change the orientation of the attitude.
The term sentiment in the context of a computational text analysis task is
mentioned for the first time in the paper by Das and Chen (2001). According to the
authors in this paper, sentiment takes on a specific meaning, that is, the net of
positive and negative opinion expressed about a stock on its message board. . At
the same time, Tong (2001) proposed a new task at the Workshop on Operational
Text Classification (OTC2001), which concerned the detection and tracking of

15

opinions in on-line discussions and the subsequent classification of the sentiment of


opinion.
The aim of the paper by Turney (2002) is to classify reviews as recommended
(thumbs up) or not recommended (thumbs down). The classification of a review is
predicted by the average semantic orientation of the phrases in the review that
contain adjectives or adverbs. A phrase has a positive semantic orientation when it
has good associations (e.g., subtle nuances) and a negative semantic orientation
when it has bad associations (e.g., very cavalier).
Pang et al. (2002) propose different methods to determine the sentiment, or
overall opinion towards the subject matter for example, whether a product review is
positive or negative.
Nasukawa and Yi (2003) entitled their paper, Sentiment analysis: Capturing
favorability using natural language processing. In this paper, they state that the
essential issues in sentiment analysis are to identify how sentiments are expressed
in texts and whether the expressions indicate positive (favorable) or negative
(unfavorable) opinions toward the subject.
Yi et al. (2003), in their paper Sentiment Analyzer: Extracting sentiments about
a given topic using natural language processing techniques, consider opinion an
equivalent term to sentiment. Their approach approximates the task later known as
feature-based opinion mining and summarization (Hu and Liu, 2004), as they
extract sentiment in correlation to a specific topic.
Subjectivity analysis and sentiment analysis/opinion mining have been
considered to be highly-related tasks. Pang and Lee (2003) state that subjectivity
analysis performed prior to sentiment analysis leads to better results in the latter.
Banea et al. (2010) states in this sense that while subjectivity classification labels
text as either subjective or objective, sentiment or polarity classification adds an
additional level of granularity, by further classifying subjective text as either
positive, negative or neutral.
However, according to Pang and Lee (2008): () nowadays many construe the
term (sentiment analysis) more broadly to mean the computational treatment of
opinion, sentiment, and subjectivity in text.
As we can observe, terminology employed in this field is highly variable. At
times, the definitions used to denote one task or another and their related concepts
are vague, inexact, overlap with definitions given for other terms, different terms
are used to denote the same task or concept and the definitions are not consistent
with the formal ones (that we can find, for example, in a dictionary). On top of
their inconsistencies, there is also a large body of research performing emotion
detection to improve sentiment analysis (Cambria et al., 2009), although no explicit
relation between emotion, sentiment and opinion is presented.

16

We will further on define emotion and relate it to the wider context of affect and
the term affective computing, used in Artificial Intelligence. Subsequently, we
clarify the connection between all these concepts, within the frame of the Appraisal
Theory4.
Affect is a superordinate concept that subsums particular valenced conditions
such as emotions, moods, feelings and preferences (Ortony et al., 2005), being one
of the four components whose interaction make the human organism function
effectively in the world (Ortony et al., 2005), along with motivation, cognition and
behaviour.
Affective computing is a branch of Artificial Intelligence (AI) dealing with the
design of systems and devices that can recognize, interpret, and process human
affect. The concept includes interdisciplinary work from computer science, but also
psychology and cognitive science. However, the term was introduced in the study
within AI by (Picard, 1995), who envisages both the capabilities of computers to
interpret affect from digital content, as well as imitate affect in humans.
Emotion is a complex phenomenon, on which no definition that is generally
accepted has been given. However, a commonly used definition considers emotion
as an episode of interrelated, synchronized changes in the states of all or most of
the five organismic subsystems (Information processing, Support, Executive,
Action, Monitor) in response to the evaluation of an external or internal stimulus
event as relevant to major concerns of the organism. (Scherer, 1987; Scherer,
2001).
Emotion detection and classification is the task of spotting linguistic
expressions of emotion from text and classifying them in predefined
categories/labels (e.g. anger, fear, sadness, happiness, surprise, disgust etc.).
Emotion is a much more complex phenomenon, whose expression in language may
not always have a subjective form (Ortony, 1997).
Let us consider a few examples, to note the difference between subjectivity,
opinion, emotion and sentiment, in a sense that is consistent to the definitions these
concepts are given outside the NLP world. In the following table, Y corresponds
to yes, N corresponds to no, C corresponds to context (dependence on
the context), POS to positive, NEG stands for negative and NEU for
neutral. The symbol --- stands for the lack of the corresponding element.

This set of theories have been proposed in Psychology, by De Rivera (1977), Frijda (1986), Ortony,
Clore and Collins (1988), Johnson-Laird and Oatley (1989). It has also been used in to define the
Appraisal Framework in Linguistics, by Martin and White( 2001).

17

11.
12.
13.
14.
15.
16.
17.
18.

Polarity?
POS/NEG/NE
U
(if there is a
sentiment)

10.

Sentiment?

9.

Emotion?

8.

Opinionated?

7.

Subjective?

5.
6.

It broke in two days.


Its great!
The bank is likely to go bankrupt.
The president denied the existence of
any financial crisis during the
election campaign.
The screen is really huge!
It took them three years to fix the
puthole in front of our building.
They always advertise this product
as cheap and chic. Can you actually
say that?
These headphones are perfect. If you
are deaf, that is.
I am firmly convinced they came
yesterday.
This car costs 14.000 Euros.

N
Y
Y
N

N
Y
Y
C

Y
Y
C
C

Y
Y
C
C

NEG
POS
--NEG/--- (C)

Y
N

Y
Y

Y
Y

Y
Y

POS/NEG/(C)
NEG

NEG

NEG

This car only costs 14.000 Euros.


The president should prepare his
arguments for the next negotiations.
It is rumored he might sell his
business.
This book costs 10 on Amazon.com
On Amazon, this book costs 10!
He killed all the flies.
It killed all the birds.
In the end, he killed all the bad guys
and walked into the sunset with the
girl.

Y
Y

Y
Y

C
C

Y
C

POS/NEG/NEU
(C)
POS/NEG/NEU
(C)
POS
NEG/NEU (C)

---

N
N
N
N
N

N
Y
N
Y
C

N
Y
N
Y
C

N
Y
N
Y
C

--POS/NEG (C)
--NEG
POS/NEG (C)

Example

Nr.
1.
2.
3.
4.

Table 2.1: Examples illustrating the concepts of sentiment, emotion, opinion,


subjectivity and objectivity
As we can see, these classes that are very frequently used in opinion
mining/sentiment analysis and subjectivity analysis are not related in a
straightforward manner. As stated by some of the definitions we presented, not all
opinions are neither necessarily subjective in nature, nor do they have to contain
18

sentiment in all the cases. Sentiments on different targets can also be conveyed by
presenting arguments that are factual in nature (e.g. The new phone broke in two
days), as the underlying emotion that is expressed is not always stated directly.
Moreover, subjective statements that contain an opinion must not necessarily
contain an emotion, thus they have no sentiment correlated to them (e.g. I believe
in God). Finally, a text may express and emotion without expressing any opinion
and sentiment (e.g. Im finally relaxed!). This idea can be summarized in the
following schema:

Subjective

Objective

Opinion

Sentiment
Emotion

Figure 2.1: The relation between subjectivity, objectivity, opinion, sentiment and
emotion

2.3. ATTITUDE AND APPRAISAL. APPRAISAL THEORIES.


Although the definitions of the tasks, as well as of the concepts they involve are
highly discrepant, all the tasks which are concerned with opinion, sentiment,
attitude, appraisal or emotion and which were denominated sentiment analysis or
opinion mining (and the aforementioned related terms) actually have the same
aim. This can be summarized, in a very nave manner, by the statement Does the
source of the text like/appreciate/think positively/feel good/is happy/ about the
target of what is stated or not?; What exactly does the source
like/appreciate/think positively/feel good/is happy or does not like/appreciate/think
positively/feel good/is happy about?; How does he/she express that?.
It is very interesting to note that most of the literature does not consider any
further elements of the act of communication the source of what is said and who it
is intended for. The Speech-Act theory (Austin, 1976) states that When we speak,
our words do not have meaning in and of themselves. They are very much affected
by the situation, the speaker and the listener. Thus words alone do not have a simple
fixed meaning. Therefore, apart from what is explicitly stated in a text, it is very
19

important to take into consideration the context of what is said (who is saying it,
why, what is the intention behind what he/she is saying; is the act of writing purely
to inform the reader on a specific object, is it to produce a specific emotion to the
reader, is it to convince him/her about something, is it to purely express his/her own
regard on this object he/she is writing about?) and whom it is addressed to
(would a potential reader like what he/she is reading, would he be comfortable with
it, would he resent it, would he think it is good, positive, would he feel happy about
it).
Following on this idea and based on the definitions we have seen so far, we can
claim that much of the work that has been done under the umbrella of sentiment
analysis or opinion mining is actually concerned with detecting attitudes and
classifying them, since, in fact, the act of communication (in this case, through text)
is intended for a reader. Both the writer (with his own views on the world), as well
as the reader, who is not a mere passive entity, but who actively uses his/her own
knowledge of the world, the context and his/her affect to interpret what is written
should be taken into consideration at the time of deciphering the sentiments
expressed in text. Additionally, especially in the traditional textual genres such as
newspaper articles, where writers are thought to be objective when rendering a
piece of news, expressions of sentiment cannot be direct. Opinions are expressed in
a non-subjective manner, by omitting certain facts and overly repeating others
(Balahur and Steinberger, 2009).
The work in the field of NLP that deals with the computational treatment of
attitude is called attitude analysis or appraisal analysis relating it to the
Appraisal Theory.
An attitude (Breckler and Wiggins, 1992) is a hypothetical construct that
represents an individual's degree of like or dislike for something. Attitudes are
generally positive or negative views of a person, place, thing, or event this is
often referred to as the attitude object. People can also be conflicted or ambivalent
toward an object, meaning that they simultaneously possess both positive and
negative attitudes toward the item in question. Attitudes are judgments. They
develop on the ABC model (affect, behavior, and cognition). The affective response
is an emotional response that expresses an individual's degree of preference for an
entity. The behavioral intention is a verbal indication or typical behavioral tendency
of an individual. The cognitive response is a cognitive evaluation of the entity that
constitutes an individual's beliefs about the object. Most attitudes are the result of
either direct experience or observational learning from the environment.
Work that has concentrated on attitude analysis was done by Taboada and
Grieve (2004), Edmonds and Hirst (2002), Hatzivassiloglou and McKeown (1997)
and Wiebe et al. (2001). However, only the first work considered attitude as

20

different from subjectivity, although subjectivity, as we have seen, does not


always imply an evaluation as in the case of attitude.
The work by Taboada and Grieve (2004) and the recent work by Neviarouskaya
et al. (2010) also try to link the concept of sentiment (judgment/appreciation) with
attitude, based on the Appraisal Theory (Martin and White, 2005). This is a
framework of linguistic resources inscribed in discourse semantics, which describes
how writers and speakers express inter-subjective and ideological positions (the
language of emotion, ethics and aesthetics), elaborating on the notion of
interpersonal meaning, i.e. social relationships are negotiated through evaluations of
the self, the others and artifacts. According to this theory, emotion is achieved by
appraisal, which is composed of attitude (affect, appreciation, and judgment),
graduation (force and focus), orientation (positive versus negative) and polarity
(which can be marked or unmarked).
Only the work by Neviarouskaya et al. (2010) distinguishes among the different
components of attitudes and employs different methods to tackle each of the issues
in this context (i.e. direct and indirect expressions of sentiments, through judgments
and appreciations or emotions, using emotion detection and sentiment analysis
techniques). Nonetheless, this work, too, remains at a lexical level of analysis.
Although the terminology employed in defining the tasks related to the
computational treatment of subjectivity and sentiment is still not well-established,
the body of research performed in the field makes it impossible to draw a line
between what could be wrongfully defined and what is correctly defined.
The tasks of subjectivity analysis, opinion mining or sentiment analysis, as we
will see along the following chapters, are defined and tackled very differently
depending on the final aim of the application, the type of text on which it is applied
and the context in which it is performed. As a consequence, in the next chapters,
besides the description of the tasks we aim at resolving using sentiment analysis, we
will also describe each of the definitions given in the context of the tasks sentiment
analysis was applied to.
The operational definition we will use for the concept of sentiment along this
thesis will combine the understanding that is given by the literature to sentiment
and attitude together. Thus, we consider sentiment as a settled opinion reflective of
ones feelings a single component of emotion, denoting the subjective experience
process (Scherer, 2005), implicitly or explicitly present in text through
expressions of affect, appreciation, judgment, but also an expression of behavior
and cognition. In this context, sentiments are not only present in subjective
sentences, but can also be expressed in objective sentences (e.g. It broke in two
days, implicitly describing a negative appreciation of the quality of the product
described). In the context where the term opinion is employed, we refer to the
specific type of opinions represented by sentiments (defined as above).
21

CHAPTER 3. STATE OF THE ART


Motto: Compelling reason will never convince blinding emotion.(Richard Bach)
Research in affect has a long established tradition in many sciences - linguistics,
psychology, socio-psychology, cognitive science, pragmatics, marketing or
communication science. Recently, as we have seen in the previous chapter, many
closely related subtasks were developed also in the field of affect computing, from
emotion detection, subjectivity analysis, opinion mining to sentiment analysis. As
we will show in the next sections and chapters, this research area covers many
aspects in Artificial Intelligence (AI) and Natural Language Processing (NLP), each
with its challenges and proposed solutions.
Not less important is the fact that subjectivity analysis, for example, was
demonstrated to bring substantial improvement over other NLP tasks, such as
information extraction (IE), question answering (QA), authorship determination,
text categorization, multi-document summarization and was employed as a filtering
step for sentiment analysis. Opinion mining (sentiment analysis), on the other hand,
besides being employed per se in order to seek attitude towards different entities,
has become a crucial component in Natural Language Processing systems that
answer mixed types of questions or summarize texts pertaining to new, emerging
text types (blogs, e-commerce sites reviews, forums etc.).
In this chapter, we present a general overview of the state of the art methods and
resources that were employed and created for the tasks of subjectivity and sentiment
analysis. Additionally, we present the state of the art methods employed in general
affect detection challenges and their application in the context or jointly with other
NLP tasks. We start by presenting the resources created for the task of subjectivity
analysis.

3.1. SUBJECTIVITY ANALYSIS


The study of subjectivity in text from a computational linguistics point of view
began together with the work of Janyce Wiebe (Wiebe, 1994), who, based on Ann
Banfields (Banfield, 1982) linguistic theories on narrational aspects and Quirks
definition, centered the idea of subjectivity around that of private states (Quirk,
1985), events which are not open to objective observation and verification, the
linguistic expression of somebodys opinions, sentiments, emotions, evaluations,
beliefs, speculations (Wiebe et al., 2004). The aim in her research was to recognize
opinion-oriented language, which she denominated as subjectivity indicators, and
to distinguish it from objective descriptions.
23

The aim of subjectivity analysis, as described in the task definition, is to classify


content into objective or subjective. This task is not trivial, as many expressions
carry in themselves a certain subjectivity (which we will later show is related to the
content and text intentionality and many expressions are used both in a subjective,
as well as objective manner. An example of factual text would be 100 people died
after a bomb exploded in the centre of the town and an example of subjective text
is Hundreds of prisoners were mercilessly tortured in the prison.
Research in the field of subjectivity analysis concentrated on the creation of
resources and the definition of methods to detect subjectivity indicators.

3.1.1. RESOURCE CREATION FOR SUBJECTIVITY ANALYSIS


A) MANUALLY CREATED RESOURCES
Resources containing subjective words are The General Inquirer5 (Stone et al.,
1966), which was not specifically designed for subjectivity analysis, but contains
11788 sense disambiguated words, out of which the subjective ones are annotated
correspondingly with polarity, strength and according to axes of emotion. The
resource was created manually.
Comlex (Macleod et al., 1994) is a dictionary containing 38000 words for the
English language that was specifically built for NLP systems. This dictionary also
contains a large number of attitude adverbs.
Another resource for subjectivity analysis, which is also annotated as far as
polarity is concerned, is the Multi-Perspective Question Answering (MPQA)
corpus6 (Wiebe et al., 2005; Wiebe and Wilson, 2005; Wilson and Wiebe, 2003),
containing annotations of 10000 sentences from the world press. The annotation
scheme is complex, with the intention of spotting as many of the subjectivity and
emotion-related aspects as possible. Subjectivity is defined around private states
(Wiebe, 1994). For the MPQA annotation process, each of the subjectivity
expressions is defined within a private state frame (Wiebe et al., 2005), which
includes the source of the private state (which can be nested, i.e., it is not only the
author that can be the source of the private state described, but the author may
introduce the private states of another person), the target and properties involving
the intensity of the subjectivity expressed, the significance and the type of attitude.
The second annotated type of element is objective speech event, which, in

5
6

http://www.wjh.harvard.edu/~inquirer/
http://www.cs.pitt.edu/mpqa/

24

contrast to the private frames, contain annotations of events that are objective in
nature.
The Opinion Finder lexicon (subjectivity clues) (Wilson et al., 2005) contains
over 8000 words, annotated also at the level of polarity value, and was built starting
with the grouping of the subjectivity clues in (Riloff and Wiebe, 2003) and enriched
with polarity annotated subjective words taken from the General Inquirer and the
lexicon proposed by Hatzivassiloglou and McKeown (1997). It is interesting to
notice that authors found most of the subjective words to have either a positive or a
negative polarity and only few were both positive and negative or neutral.
B) SEMI-AUTOMATICALLY AND AUTOMATICALLY CREATED
RESOURCES
Another annotation scheme and corpus for subjectivity versus objectivity
classification, as well as polarity determination at sentence level was developed by
Yu and Hatzivassiloglou (2003), in a semi-automatic manner. The authors start
from a set of 1336 seed words, manually annotated by Hatzivassiloglou and
McKeown (1997), extended by measuring co-ocurrence between the known seed
words and new words. The hypothesis on which the authors based their approach is
that positive and, respectively, negative words, tend to co-occur more than it is
expected by chance. As measure for association, the authors employ log-likelihood
on a corpus that is tagged at the part-of-speech level.
A resource for subjectivity was built semi-automatically on the basis of the
Appraisal Theory (Martin and White, 2005). The Appraisal Theory is a framework
of linguistic resources inscribed in discourse semantics, which describes how
writers and speakers express inter-subjective and ideological positions (the
language of emotion, ethics and aesthetics), elaborating on the notion of
interpersonal meaning, i.e. social relationships are negotiated through evaluations of
the self, the others and artifacts. According to this theory, emotion has achieved by
appraisal, which is composed of attitude (affect, appreciation, judgment),
graduation (force and focus), orientation (positive versus negative) and polarity
(which can be marked or unmarked). A lexicon of appraisal terms is built by
Whitelaw et al. (2005), based on the examples provided by Martin and White
(2005) and Matthiassen (1995) (400 seed terms) and patterns in which filler
candidates were extracted from WordNet (Fellbaum ed., 1999). Term filtering was
done by ranking obtained expressions and manually inspecting terms that were
ranked with high confidence. The resulting lexicon contains 1329 terms.

25

C) APPROACHES TO MAPPING SUBJECTIVITY RESOURCES TO


OTHER LANGUAGES
Most of the work in obtaining subjectivity lexicons was done for English. However,
there were some authors who developed methods for the mapping of subjectivity
lexicons to other languages.
To this aim, Kim and Hovy (2006) use a machine translation system and
subsequently use a subjectivity analysis system that was developed for English to
create subjectivity analysis resources in other languages.
Mihalcea et al. (2007) propose a method to learn multilingual subjective
language via cross-language projections. They use the Opinion Finder lexicon
(Wilson et al., 2005) and use two bilingual English-Romanian dictionaries to
translate the words in the lexicon. Since word ambiguity can appear (Opinion
Finder does not mark word senses), they filter as correct translations only the most
frequent words. The problem of translating multi-word expressions is solved by
translating word-by-word and filtering those translations that occur at least three
times on the Web.
Another approach in obtaining subjectivity lexicons for other languages than
English was explored by Banea et al. (Banea et al., 2008b). To this aim, the authors
perform three different experiments, obtaining promising results. In the first one,
they automatically translate the annotations of the MPQA corpus and thus obtain
subjectivity annotated sentences in Romanian. In the second approach, they use the
automatically translated entries in the Opinion Finder lexicon to annotate a set of
sentences in Romanian. In the last experiment, they reverse the direction of
translation and verify the assumption that subjective language can be translated and
thus new subjectivity lexicons can be obtained for languages with no such
resources.
Further on, another approach to building lexicons for languages with scarce
resources is presented by Banea et al. (Banea et al., 2008a). In this research, the
authors apply bootstrapping to build a subjectivity lexicon for Romanian, starting
with a set of seed subjective entries, using electronic bilingual dictionaries and a
training set of words. They start with a set of 60 words pertaining to the categories
of noun, verb, adjective and adverb from the translations of words in the Opinion
Finder lexicon. Translations are filtered using a measure of similarity to the original
words, based on Latent Semantic Analysis (LSA) (Deerwester et al., 1990) scores.
Yet another approach to mapping subjectivity lexica to other languages is
proposed by Wan (2009), who uses co-training to classify un-annotated Chinese
reviews using a corpus of annotated English reviews. He first translates the English
reviews into Chinese and subsequently back to English. He then performs cotraining using all generated corpora.
26

Kim et al. (2010) create a number of systems consisting of different subsystems,


each classifying the subjectivity of texts in a different language. They translate a
corpus annotated for subjectivity analysis (MPQA), the subjectivity clues (Opinion
finder) lexicon and re-train a Nave Bayes classifier that is implemented in the
Opinion Finder system using the newly generated resources for all the languages
considered.
Finally, Banea et al. (2010) translate the MPQA corpus into five other languages
(some with a similar ethimology, others with a very different structure).
Subsequently, they expand the feature space used in a Nave Bayes classifier using
the same data translated to 2 or 3 other languages. Their conclusion is that by
expanding the feature space with data from other languages performs almost as well
as training a classifier for just one language on a large set of training data.

3.1.2. APPLICATIONS OF SUBJECTIVITY ANALYSIS


Automatic subjectivity analysis was proven to be helpful as filtering step in
Information Extraction (IE) (Somasundaran et al., 2007), as subjectivity often
causes false hits for IE.
Riloff et al. (2005) present a strategy to improve IE through the use of
subjectivity. In their approach, the authors use the Opinion Finder subjectivity
lexicon in order to create subjectivity patterns, which help them build a classifier
that is able to capture feature above the word level and takes into consideration the
surrounding context of the subjective expressions. At the time of performing IE,
they discard the subjective sentences within the results and obtain a large
improvement over previous approaches that do not distinguish sentences from the
subjectivity/objectivity point of view.
Pang and Lee (2004) demonstrate that using subjectivity analysis as the first step
towards opinion mining yeilds better results.
Stoyanov et al. (2004) show that subjectivity analysis used in the context of
separating opinionated answers to multi-perspective questions improves the results
in this task.
Wiebe and Mihalcea (2006) research on the effect of subjectivity detection for
Word Sense Disambiguation (WSD). They show that distinguishing subjective
contexts can help to assign the right sense to words that can be used both as
objective as well as subjective. This distinction is also used by Rumbell et al.
(2008), where the authors identify the figurative use of language (metaphorical
senses of words) to classify sentiment in conversations.

27

3.2. SENTIMENT ANALYSIS (OPINION MINING)


Sentiment analysis, also known as opinion mining (Pang and Lee, 2008) attempts to
identify the opinion/sentiment that an entity may hold towards an object and it
involves a more profound, fine-grained analysis of the text compared to subjectivity
analysis.
Some authors consider that sentiment analysis is subsequent to the task of
subjectivity detection (Pang and Lee, 2002). In this sense, sentiment analysis
continues with the classification of identified sentiment-containing text into two or
three categories (positive, negative or positive, negative and neutral). The neutral
category corresponds, in this case, to the objective category in subjectivity analysis,
although one may imagine that neutrality can also be expressed as opinions that do
not have a clear tendency towards positive or negative.
Although intuitively subjective texts express polar sentiments, Wilson et al.
(2005) show that factual pieces of text contain indirectly expressed opinions and
opinionated texts also contain factual expressions, as a means of argumentation for
an idea. Therefore, sentiment analysis has been approached by the majority of the
research independently from subjectivity analysis.
Opinions have three basic components. The first one is the opinion holder,
which is the source of the opinion; the second element is the object, which is
the target of the opinion; finally, the third element is the opinion (a view,
attitude, or appraisal on an object from an opinion holder), which can be seen as a
private state concerning the target. There are many subtasks that were defined
within the main sentiment analysis task. For each of these, resources were created
and different methods were developed. There are three main research areas around
which these tools evolved:
! Creation of resources for sentiment analysis/ opinion mining;
! Classification of text (whose main aim is finding expressions of emotion
and classifying the texts into positive and negative), a task which has been
performed at a document, sentence, phrase and word level;
! Opinion extraction (which is concerned with finding parts of text with
opinion, identifying the polarity of the sentiment expressed and determining
the source and target of the sentiment expressed).
Recent work has also considered opinion mining as a two-stage problem
(Jijkoun et al., 2010), in an attempt to join the two communities that have been
working in this field (Information Retrieval and Information Extraction) and offer
an end-to-end solution to the opinion analysis problem from the retrieval to the
classification stages. Thus the authors believe that two different, but complementary
problems can be identified (Jijkoun et al., 2010):
28

Sentiment extraction: given a set of textual documents, identify phrases,


clauses, sentences or entire documents that express attitudes, and
determine the polarity of these attitudes (Kim and Hovy, 2004); and
Sentiment retrieval: given a topic (and possibly, a list of documents
relevant to the topic), identify documents that express attitudes toward this
topic (Ounis et al., 2007).

3.2.1. CREATION OF RESOURCES FOR SENTIMENT ANALYSIS


As we mentioned in Section 3.1, some authors consider that the task of sentiment
analysis is subsequent to that of subjectivity analysis, involving an extra step: the
classification of the retrieved opinion words according to their polarity. Thus, the
existing lexical resources for the opinion task contain words and expressions that
are subjective and that have a value for polarity assigned.
At the same time, distinguishing a text that presents facts from one that depicts
subjective statements is more straightforward, whereas classifying what is said into
valence categories involves a more profound analysis of the context, the structure of
the text, the presence of modifier expressions etc.
Thus, apart from creating lexical resources that contain words and expressions
with their corresponding a priori assigned polarity, research in sentiment analysis
also concentrated on the development of annotation schemes that can be used to
label corpora in order to capture the specificities of the different expressions of
opinion (be it direct, indirect or implicit), in the diverse types of text from which
opinion is mined (news/blogs/product reviews).
It was demonstrated that corpora labeling is a necessary step for the training and
evaluation of systems implementing sentiment analysis and that fine-grained
opinion mining requires the use of such resources. However, some approaches to
document-level sentiment analysis (such as opinion mining of movie reviews) use
as gold standard texts that are already classified, since the e-commerce sites
where they are taken from allow for product reviewers to assign stars to the
different categories describing a product (from 1 to 5 stars, 1 star being bad and 5
stars meaning very well).
There are a series of techniques that were used to obtain lexicons of subjective
words with associated polarity.
Hu and Liu (2004) start with a set of seed adjectives (good and bad) and
apply synonymy and antonymy relations in WordNet.
A similar approach was used in building WordNet Affect (Strapparava and
Valitutti, 2004), starting from a larger set of seed affective words, classified
according to the six basic categories of emotion (joy, sadness, fear, surprise, anger
and disgust) and expanding the lexicon using paths in WordNet.
29

Another related method was used in the creation of SentiWordNet (Esuli and
Sebastiani, 2005). The idea behind this resource was that terms with similar
glosses in WordNet tend to have similar polarity. Thus, SentiWordNet was build
using a set of seed words whose polarity was known and expanded using gloss
similarity.
As mentioned in the subjectivity classification section, in the collection of
appraisal terms in Whitelaw et al. (2005), the terms also have polarity assigned.
MicroWNOp (Cerini et al., 2007), another lexicon containing opinion words
with their associated polarity, was built on the basis of a set of terms (100 terms for
each of the positive, negative and objective categories) extracted from the General
Inquirer lexicon and subsequently adding all the synsets in WordNet where these
words appear. The criticism brought to such resources is that they do not take into
consideration the context in which the words or expressions appear. Other methods
tried to overcome this critique and built sentiment lexicons using the local context
of words.
Pang et al. (2002) built a lexicon of sentiment words with associated polarity
value, starting with a set of classified seed adjectives and using conjunctions
(and) disjunctions (or, but) to deduce orientation of new words in a corpus.
Turney (2002) classifies words according to their polarity on the basis of the
idea that terms with similar orientation tend to co-occur in documents. Thus, the
author computes the Pointwise Mutual Information score between seed words and
new words on the basis of the number of AltaVista hits returned when querying the
seed word and the word to be classified with the NEAR operator.
In our work in, we compute the polarity of new words using polarity anchors
(words whose polarity is known beforehand) and Normalized Google Distance
(Cilibrasi and Vitanyi, 2006) scores using as training examples opinion words
extracted from pros and cons reviews from the same domain, using the clue that
opinion words appearing in the pros section are positive and those appearing in
the cons section are negative (Balahur and Montoyo, 2008b; Balahur and
Montoyo, 2008d; Balahur and Montoyo, 2008f). Another approach that uses the
polarity of the local context for computing word polarity is the one presented by
Popescu and Etzioni (2005), who use a weighting function of the words around the
context to be classified.
The lexical resources that were created and are freely available for use in the
task of opinion mining are:
WordNet Affect (Strapparava and Valitutti, 2004);
SentiWordNet (Esuli and Sebastiani, 2006);
Emotion triggers (Balahur and Montoyo, 2008a);
MicroWNOp (Cerini et al., 2007);
General Inquirer (Stone et al., 1966);
30

Appraisal terms (annotated according to the principles of the appraisal


theory framework in linguistics) (Whitelaw et al., 2005).
On the other hand, there are several specific annotation schemes and
corresponding corpora that were created for the affect-related applications in NLP:
The ISEAR (International Survey on Emotion Antecedents and Reactions)
corpus (Scherer and Wallbott, 1997);
The MPQA (Multi-perspective Question Answering System) corpus;
The EmotiBlog corpus (Boldrini et al., 2009);
The TAC (Text Analysis Conference) Opinion Pilot data (TAC 2008) and
TREC (Text Retrieval Conference) data (from 2006 to 2010), which
consists of annotated texts for the different opinion retrieval specific tasks,
on the Blog06 collection;
The NTCIR MOAT (Multilingual Opinion Analysis Track) data (20072010), which contains both monolingual annotated data for opinion mining
in English, Chinese and Japanese, as well as cross-lingual analysis data (in
MOAT 2010).
The recent tasks proposed at NTCIR MOAT concerning the detection of the
opinion holder and opinion target, in addition to the opinion classification, has
encouraged the development of new corpora that annotates the opinion source and
target, at different granularities. In this sense, apart from the work by Boldrini et al.
(2009), the work by Toprak et al. (2010) proposes a model for sentence and
expression level annotation of opinions in user-generated discourse.

3.2.2. RESEARCH IN TEXT SENTIMENT POLARITY


CLASSIFICATION
Research in sentiment analysis, as we defined it, aims at classifying opinionated
texts according to the polarity of the sentiment expressed. However, depending on
the final use of the sentiment analysis task, classification of opinion is done at
different levels.
For example, when a user is interested in finding out opinions people gave about
a movie, the overall sentiment is enough to be able to decide whether to see it or
not. On the other hand, when a user is interested in booking a hotel or buying an
electronic product, the general opinion given in reviews might not be sufficient,
because the user might be more interested in some of the features of the hotel or the
electronic product and be indifferent with regard to others (e.g. location versus
service in a hotel or size versus GPS capabilities in a mobile phone).
Opinion mining requires different techniques and approaches, depending both on
the level of analysis that is required and interesting to the user, as well as on the
type of text analyzed.
31

Researchers have proposed different scenarios for mining opinion and have
proposed different methods for performing this task. It is important to mention that
although the general idea of classifying sentiment in text is understood as one of
assigning a piece of text (document, sentence, review) a value of positive or
negative (or neutral), other scenarios were defined in which the positive
category refers to liking, arguments brought in favor of an idea (pros) or support
of a party or political view and the negative class includes expressions of
disliking something, arguments brought against an idea expressed (cons) or
opposition to an ideology.
Sentiment analysis (opinion mining) is a difficult task due to the high semantic
variability of natural language, which we have defined according to the
understanding given to sentiments and attitudes, supposes not only the discovery of
directly expressed opinions also the extraction of phrases that indirectly or
implicitly value objects, by means of emotions or attitudes.
It is also important to note the fact that sentiment analysis does not necessarily
require as input an opinionated piece of text (Pang and Lee, 2008). Good versus bad
news classification has also been considered as a sentiment classification task,
which was approached in research such as the one proposed by Koppel and
Shtrimberg (2004).
However, it is also very important to note that a clear distinction must be made
at the time of performing sentiment analysis at the document level, namely that, for
example, the content of good versus bad news (which is factual information) should
not influence the judgment of sentiment as far as the facts are concerned or the
people involved. To exemplify, a sentence such as Great struggles have been made
by the government to tackle the financial crisis, which led many companies to
bankruptcy must not be seen as negative because it discusses the consequences and
gravity of the financial crisis, but must be seen as positive, when sentiment on the
government is analyzed. Certainly, we can see that in this case the sentiment and its
polarity arise from the manner in which the reporting is done.
We thus distinguish between document-level, sentence-level and feature-level
sentiment analysis. The tasks are defined differently at each level and involve the
performing of extra, more specialized steps. We will further show which those steps
are. According to the survey by Pang and Lee (2008), general strategies that have
been used in sentiment polarity classification were:
Classification using the representation of text as feature vectors where
entries correspond to terms, either as count of frequencies (using tf-idf), or
counting the presence or absence of a certain opinion words. In this context,
Wilson et al. (2005) have shown that rare words (that appear very
infrequently in the opinion corpus), also called hapax legomena have a very
good precision in subjectivity classification.
32

Using information related to the part of speech of the sentiment words and
applying specialized machine learning algorithms for the acquiring of such
words (adjectives, verbs, nouns, adverbs). The work in acquiring nouns
with sentiment has been proposed by Riloff et al. (2005). Here, the authors
use dependency parsing and consider as features of machine learning
algorithms the dependency relations. In this setting, information about
modifiers or valence shifters can be introduced, as dependency analysis
allows for the identification of the constituents that are modified.
For the tasks in which sentiment on a certain topic must be extracted, the
features used in machine learning for sentiment classifications were
modified to include information on the mentions of the topic or the Named
Entities mentioned in relation to it.
In the subsequent sections, we present the methods that were employed for
sentiment analysis at the different levels considered.

SENTIMENT ANALYSIS AT A DOCUMENT LEVEL


Sentiment analysis has been done at a document level for movies, book reviews
etc., starting from the assumption that each document (or review) focuses on a
single object (product, topics) and contains opinion from a single opinion holder.
This setting is true for reviews, but does not hold for newspaper articles or blog
posts.
Work at this level has been done by Turney (2002), on movie reviews. The
sentiment polarity of the individual opinion words is computed using a set of seed
adjectives whose polarity is previously known and computing the Pointwise Mutual
Information score that is obtained between the word to classify and the known word
using the number of hits obtained by querying the two words together with the
NEAR operator on the AltaVista search engine.
The final score obtained for the review is computed as sum of the polarities of
the individual opinionated words in the review, from a set of sentences that is
filtered according to patterns bases on the presence of adjectives and adverbs.
Another approach at the classifying polarity of sentiment at a document level is
presented in Pang et al. (2002), where the authors use Nave Bayes machine
learning using unigram features and show that the use of unigrams outperforms the
use of bigrams and of sentiment-bearing adjectives.
Another work in classifying documents according to their polarity is presented
by Dave et al. (2003). In this work, the authors extract patterns of opinion from a
corpus of reviews which are already graded.

33

Mullen and Collier (2004) show that classifying sentiment using Support Vector
Machines with features computed on the basis of word polarity, semantic
differentiation computed using synonymy patterns in WordNet, proximity to topic
features and syntactic relations outperforms n-gram classifications.
Another similar approach was taken by Pang and Lee (2003). In this approach,
the authors classify reviews into a larger scale of values (not only positive and
negative), seen as a regression problem, and employ SVM machine learning with
similarity features. They compare the outcome against the number of stars given to
the review.
Chaovalit and Zhou (2005) perform a comparison between different methods of
supervised and unsupervised learning based on n-gram features and semantic
orientation computed by using patterns and dependency parsing.
Goldberg and Zhu (2006) present a graph-based approach to sentiment
classification at a document level. They represent documents as vectors, computed
on the basis of presence of opinion words and then link each document to the k
most similar ones. Finally, they classify documents on the basis of the graph
information using SVM machine learning.
A similar effort is made by Ng et al. (2006), where the goal is also to classify
documents according to their polarity. The authors present an interesting
comparison between dependency-based classification and the use of dependency
relations as features for machine learning, which concludes that dependency parsing
is not truly effective at the time of performing document level sentiment analysis, as
it was previously shown in other research (Kudo and Matsumoto, 2004).
SENTENCE-LEVEL SENTIMENT ANALYSIS
At the sentence level, or part of document level, sentiment analysis is done in most
cases in two steps: the first one views the selection of subjective sentences and the
second one aims at classifying the sentiment expressed according to its polarity.
The assumption that is made in this case is that each sentence expresses one single
opinion.
Sentiment analysis at the sentence level includes work by Pang and Lee (2004),
where an algorithm based on computing the minimum cut in a graph containing
subjective sentences and their similarity scores is employed.
Yu and Hatzivassiloglou (2003) use sentence level sentiment analysis with the
aim of separating fact from opinions in a question answering scenario.
Other authors use subjectivity analysis to detect sentences from which patterns
can be deduced for sentiment analysis, based on a subjectivity lexicon
(Hatzivassiloglou and Wiebe, 2000; Wiebe and Riloff, 2006; Wilson et al., 2004).

34

Kim and Hovy (2004) try to find, given a certain topic, the positive, negative
and neutral sentiments expressed on it and the source of the opinions (the opinion
holder). After creating sentiment lists using WordNet, the authors select sentences
which contain both the opinion holder as well as carry opinion statements and
compute the sentiment of the sentence in a window of different sizes around the
target, as harmonic and, respectively, geometrical mean of the sentiment scores
assigned to the opinion words.
Kudo and Matsumoto (2004) use a subtree-based boosting algorithm using
dependency-tree-based features and show that this approach outperforms the bagof-words baseline, although it does not bring significant improvement over the use
of n-gram features.
FEATURE-LEVEL SENTIMENT ANALYSIS
Sentiment analysis at the feature level, also known as feature-based opinion
mining (Hu and Liu, 2004; Liu, 2007), is defined as the task of extracting, given
an object (product, event, person etc.), the features of the object and the opinion
words used in texts in relation to the features, classify the opinion words and
produce a final summary containing the percentages of positive versus negative
opinions expressed on each of the features. This task has been previously defined
by Dave et al. (2003).
Feature-based opinion mining involves a series of tasks:
Task 1: Identify and extract object features that have been commented on
by an opinion holder (e.g., a reviewer).
Task 2: Determine whether the opinions on the features are positive,
negative or neutral.
Task 3: Group feature synonyms.
Subsequently, once all the groups of words referring the same feature is gathered
and the polarity of the opinion is computed, the result is presented as a percentage
of positive versus negative opinion on each feature (feature-based opinion summary
of multiple reviews).
There are a series of techniques and approaches that were used in each of these
three subtasks.
For the identification of features, pros and cons reviews were used, label
sequential rules based on training sequences were employed to define extraction
rules (Popescu and Etzioni, 2005), frequent features were mined using sequential
pattern mining (frequent phrases) and patterns for part of relations were defined
(Ding et al., 2008). Infrequent features were discovered with similarity in WordNet.
Polarity classification was done using as start point the good and bad adjectives
and exploring the synonyms and antonyms of these words in WordNet (Hu and Liu,
2004), using weighting functions depending on surrounding words (Popescu and
35

Etzioni, 2005) or using local conjunction or disjunction relations with words with
priory known polarity (Ding et al., 2008). Grouping of feature synonyms was done
using relations in WordNet.
An important related research area was explored in Task 18 at SemEval 2010
(Wu and Jin, 2010). In this task, the participants were given a set of contexts in
Chinese, in which 14 dynamic sentiment ambiguous adjectives are selected. They
were: |big, |small, |many, |few, |high, |low, |thick, |thin, |deep,
|shallow, |heavy, |light, |huge, |grave. The task was to
automatically classify the polarity of these adjectives, i.e. to detect whether their
sense in the context is positive or negative. The majority of participants employed
opinion mining systems to classify the overall contexts, after which local rules were
applied, depending on which the polarity surrounding the adjective to be classified
remained the same as the overall polarity of the text, or it changed.
Recently, authors have shown that performing very fine or very coarse-grained
sentiment analysis has drawbacks for the final application, as many times the
sentiment is expressed within a context, by comparing or contrasting with it. This is
what motivated McDonald et al. (2007) to propose an incremental model for
sentiment analysis, starting with the analysis of text at a very fine-grained level and
adding up granularity to the analysis (the inclusion of more context) up to the level
of different consecutive sentences. The authors showed that this approached highly
improved the sentiment analysis performance. The same observation was done by
Balahur and Montoyo (2009) for the task of feature-based opinion mining and
subsequently confirmed by experiments in opinion question answering (Balahur et
al., 2009a; Balahur et al., 2009d; Balahur et al, 2009h; Balahur et al., 2010a;
Balahur et al., 2010c).

3.2.3. EXTRACTION OF OPINION FROM TEXT


The task of extracting opinion in text refers to the problem of spotting exact parts of
texts where a specifically sought opinion (a positive or negative sentiment) is
presented on a specific, given target. The experiments conducted were used in the
context of multi-perspective question answering (Why do people like George
Clooney?), political debates (discovering what arguments are brought in favor or a
against a certain law or policy), dialogues (determining who is pro or against a topic
and finding the reasons) and blogs (finding users who agree or disagree with
different posts and extracting their view).
The sentiment analysis tasks defined in the extraction setting generally evolve
around Named Entities (the targets) and the source of the opinion is given
beforehand. Interesting, state-of-the-art approaches to extracting opinion from text
36

were implemented by systems participating in the Text Retrieval Conference


(TREC) in 2006-2008, Text Analysis Conference (2008) and NTCIR MOAT
competitions in 2008-2010.
State-of-the-art approaches demonstrated that the tasks related to subjectivity
and sentiment analysis require the use of specialized tools, lexical resources and
methods, as well as annotated corpora and gold standards against which systems
can compare in order to evaluate their performance.
Different machine learning algorithms were employed in diverse settings, using
several types of weighting schemes and similarity measures. Other approaches used
simpler, feature vector representations of word frequencies and presence, n-gram
classifications or simple bag-of-word counting of sentiment words. The use of
dependency parsing was disputed as to whether is brings or not improvement over
the simpler methods.
Although some approaches perform better than the others, the assumptions made
when performing the different tasks make most results impossible to implement
with the same success in real applications. For the latter, extensive knowledge has
to be added on topic, target, source, context and even world knowledge. Last, but
not least, it was proved that different textual genres and formulations of the problem
of sentiment analysis require specialized approaches, either from the resource-use
point of view or the need to enclose techniques from other tasks in NLP (for topic
detection, IE, QA, IR).
From the overview on the state-of-the-art in the field until this point, we can
conclude that we are standing in front of a challenging tasks, needing creative
solutions, involving many areas of NLP and knowledge from fields outside AI or
computing, such as linguistics, psychology, pragmatics, cognitive science and
philosophy. Additionally, in real-case scenarios, sentiment analysis must be
combined with other NLP tasks. We discuss the issues involved in this context in
the next sections.

3.3. APPLICATIONS OF OPINION MINING TO TRADITIONAL


TASKS OPINION QUESTION ANSWERING, OPINION
SUMMARIZATION
3.3.1. OPINION QUESTION ANSWERING
Question Answering (QA) can be defined as the Natural Language Processing
(NLP) task in which given a set of questions and a collection of documents, an
automatic NLP system is employed to retrieve the answer to the queries in Natural
Language (NL). Research focused on building factoid QA systems has a long
37

tradition; however, it is only recently that researchers have started to focus on the
development of Opininion Question Answering (OQA) systems.
Stoyanov et al. (2005) and Pustejovsky and Wiebe (2005) studied the
peculiarities of opinion questions and have found that they require the development
of specific techniques to be tackled, as their answers are longer and the analysis of
the question is not as strightforward as in the case of factoid questions. Cardie et al.
(2004) employed opinion summarization to support a Multi-Perspective QA system,
aiming to identify the opinion-oriented answers for a given set of questions.
Yu and Hatzivassiloglou (2003) separated opinions from facts and summarized
them as answer to opinion questions.
Kim and Hovy (2005) identified opinion holders, which are a key component in
retrieving the correct answers to opinion questions.
Due to the realized importance of blog data, recent years have also marked the
beginning of NLP research focused on the development of OQA systems and the
organization of international conferences encouraging the creation of effective QA
systems both for fact and subjective texts. The TAC 20087 QA track proposed a
collection of factoid and opinion queries called rigid list (factoid) and squishy
list (opinion) respectively, to which the traditional QA systems had to be adapted.
In this competition, some participating systems treated opinionated questions as
other and thus they did not employ opinion specific methods. However, systems
that performed better in the squishy list questions than in the rigid list
implemented additional components to classify the polarity of the question and of
the extracted answer snippet. The Alyssa system (Shen et al., 2007) uses a Support
Vector Machines (SVM) classifier trained on the MPQA corpus (Wiebe et al.,
2005), English NTCIR8 data and rules based on the subjectivity lexicon (Wilson et
al., 2005). Varma et al. (2008) performed query analysis to detect the polarity of the
question using defined rules. Furthermore, they filter opinion from fact retrieved
snippets using a classifier based on Nave Bayes with unigram features, assigning
for each sentence a score that is a linear combination between the opinion and the
polarity scores. The PolyU (Li et al., 2008b) system determines the sentiment
orientation of the sentence using the Kullback-Leibler divergence measure with the
two estimated language models for the positive versus negative categories. The
QUANTA system (Li et al., 2008a) performs opinion question sentiment analysis
by detecting the opinion holder, the object and the polarity of the opinion. It uses a
semantic labeler based on PropBank9 and manually defined patterns. Regarding the
sentiment classification, they extract and classify the opinion words. Finally, for the

http://www.nist.gov/tac/
http://research.nii.ac.jp/ntcir/
9
http://verbs.colorado.edu/~mpalmer/projects/ace.html
8

38

answer retrieval, they score the retrieved snippets depending on the presence of
topic and opinion words and only choose as answer the top ranking results.
Other related work concerns opinion holder and target detection. NTCIR 7
MOAT organized such a task, in which most participants employed machine
learning approaches using syntactic patterns learned on the MPQA corpus (Wiebe
et al., 2005). Starting from the abovementioned research, the work we proposed
(Balahur et al., 2009a; Balahur et al., 2009d; Balahur et al., 2009h; Balahur et al.,
2010a; Balahur et al., 2010e) employed opinion specific methods focused on
improving the performance of our OQA. We perform the retrieval at 1 sentence and
3 sentence-level and also determine new elements that we define as crucial for the
opinion question answering scenario: the Expected Source (ES) and the Expected
Target (ET), expected answer type (EAT) and Expected Polarity Type (EPT).

3.3.2. OPINION SUMMARIZATION


Whilst there is abundant literature on text summarization (Kabadjov et al., 2009;
Steinberger et al., 2007; Hovy, 2005; Erkan and Radev, 2004; Gong and Liu, 2002)
and sentiment analysis (Pang and Lee, 2008; Hovy et al., 2005; Kim and Hovy,
2004; Turney and Littman, 2003), there is still limited work at the intersection of
these two areas (Stoyanov and Cardie, 2006; Saggion and Funk, 2010; Saggion et
al., 2010).
For the first time in 2008 there was a Summarization Opinion Pilot track at the
Text Analysis Conference organized by the US National Institute of Standards and
Technology (NIST). The techniques employed by the participants were mainly
based on the already existing summarization systems. Whilst most participants
added new features (sentiment, positive/negative sentiment, positive/negative
opinion) to account for the presence of positive opinions or negative ones
CLASSY (Conroy and Schlesinger, 2008); CCNU (He et al., 2008); LIPN (Bossard
et al., 2008); IIITSum08 (Varma et al., 2008) -, efficient methods were proposed
focusing on the retrieval and filtering stage, based on polarity - DLSIUAES
(Balahur et al., 2008) - or on separating information rich clauses italic (Cruz et al.,
2008). Finally, fine-grained, feature-based opinion summarization is defined in (Hu
and Liu, 2004).
The fact that opinion summarization was proposed in such a complex setting
(additionally requiring the determination of answers to opinion questions), lead to
very low results from the participating systems. Having realized the great challenge
such an end-to-end system proposes, research in this field has been split into two
directions Opinion Question Answering and Opinion Summarization.

39

3.4. RELATED TASKS: EMOTION DETECTION AS A


GENERAL AI AND NLP CHALLENGE
Emotion is a complex phenomenon, on which no definition that is generally
accepted has been given. However, a commonly used definition considers emotion
as an episode of interrelated, synchronized changes in the states of all or most of
the five organismic subsystems (Information processing, Support, Executive,
Action, Monitor) in response to the evaluation of an external or internal stimulus
event as relevant to major concerns of the organism. (Scherer, 1987; Scherer,
2001).
The term feeling points to a single component denoting the subjective
experience process (Scherer, 2005) and is therefore only a small part of an emotion.
Moods are less specific, less intense affective phenomena, product of two
dimensions - energy and tension (Thayer, 2001).
As defined by the Webster dictionary, sentiment is a personal belief or
judgment that is not founded on proof or certainty10.
In Artificial Intelligence (AI), the term affective computing was first introduced
by Picard (1995). Although there were previous approaches in the 80s and 90s, in
the field of NLP, the task of emotion detection has grown in importance together
with the exponential increase in the volume of subjective data on the Web in blogs,
forums, reviews, etc.
Previous approaches to spot affect in text include the use of models simulating
human reactions according to their needs and desires (Dyer, 1987), fuzzy logic
(Subasic and Huettner, 2000), lexical affinity based on similarity of contexts the
basis for the construction of WordNet Affect (Strapparava and Valitutti, 2004) or
SentiWord-Net (Esuli and Sebastiani, 2005), detection of affective keywords (Riloff
et al., 2003) and machine learning using term frequency (Pang et al., 2002; Wiebe
and Riloff, 2006). The two latter approaches are the most widely used in emotion
detection systems implemented for NLP, because they are easily adaptable across
domains and languages.
Other proposed methods include the creation of syntactic patterns and rules for
cause-effect modeling (Mei Lee et al., 2009).
Significantly different proposals for emotion detection in text are given in the
work by (Liu et al, 2003) and the recently proposed framework of sentic computing
(Cambria et al., 2009), whose scope is to model affective reaction based on
commonsense knowledge.
Danisman and Alpkocak (2008) proposed an approach based on vectorial
representations. The authors compute the set of words that is discriminatory for 5 of

10

http://www.merriam-webster.com/thesaurus/

40

the 7 emotions in the ISEAR corpus and represent the examples using measures
computed on the basis of these terms.
In the SemEval 2007 Task 18 Affective Text (Strapparava and Mihalcea,
2007), the task was to classify 1000 news headlines depending on their valence and
emotion, using Ekmans 6 basic emotions model (Ekman, 1999). The participating
systems used rule-based or machine learning approaches, employing the polarity
and emotion lexicons existent at that time (SentiWordNet, General Inquirer and
WordNet Affect), the only training set available for emotion detection at that time
(the training data containing 1000 news headlines provided by the task organizers)
or calculating Pointwise Mutual Information scores using search engines.
In this chapter, we presented the main issues that research in sentiment analysis
aims to tackle and the state-of-the-art approaches that have been proposed for them.
The methods used so far, as we have seen, do not make any distinction
according to the type of text that is analyzed. In the next chapter, we present the
methods and the resources that we created in ordet to tackle the task of sentiment
analysis. In contrast to existing approaches, we will first study the requirements of
the sentiment analysis task in the different textual genres considered, and
subsequently propose adequate techniques and resources.

41

CHAPTER 4. METHODS AND RESOURCES FOR


MULTILINGUAL
SENTIMENT
ANALYSIS
IN
DIFFERENT TEXT TYPES
Motto: But are not this struggle and even the mistakes one may make better and
do not develop us more than if we keep systematically away from
emotions?(Vincent van Gogh)
As we have seen in the state-of-the-art chapter, sentiment analysis can be applied to
different textual genres, at a coarser or finer-grained level and for different
applications. The choice in the level of analysis normally depends on the on the
type of text that one is processing and the final application i.e. on the degree of
detail that one wishes or requires in order to benefit from the process of automatic
sentiment detection.
While detecting the general attitude expressed in a review on a movie suffices to
take the decision to see it or not, when buying an electronics product, booking a
room in a hotel or travelling to a certain destination, users weigh different
arguments in favor or against, depending on the features they are most interested
in (e.g. weight versus screen size, good location versus price).
Reviews are usually structured around comments on the product characteristics
and therefore, the most straightforward task that can be defined in this context is the
feature-level analysis of sentiment. The feature-level analysis is also motivated by
the fact that on specific e-commerce sites, reviews contain special sections where
the so-called pros and cons of the products are summarized, and where stars
can be given to value the quality of a characteristic of a product (e.g. on a scale
from 1 to 5 stars).
As far as the source of opinion is concerned, in this type of text, reviews are
written on the same topic and by the same author. At the time of processing, thus,
one is not usually interested in the author of the review, but rather on being able to
extract as many opinions as possible from the reviews available.
In contrast to that, in newspaper articles, for example, sentiment can be
expressed on many topics within the same piece of news, by different sources.
Thus, in this kind of text, the source and the target of opinions are very important at
the time of analyzing opinion. Moreover, in newspaper articles, the author might
convey certain opinions, by omitting or stressing upon some aspect of the text and
by thus inserting their own opinion towards the facts. Such phenomena, analyzed as
part of work on perspective determination or news bias research, should also be
taken into consideration at the time of performing opinion mining from this textual
source. Moreover, in these texts, the news in itself is highly correlated with the
43

opinion expressed; however, the positivity or negativity of the news content should
not be mistaken for the polarity of the opinion expressed therein.
In blogs, we are facing the same difficulties i.e. of having to determine the
characteristics of the source, as well as ensure that the target of the opinions
expressed is the required one. Moreover, blogs have a dialogue-like structure, and
most of the times, the topic discussed is related to a news item that is taken from a
newspaper article. The same phenomena are also present in forums, microblogs,
social network comments and reviews, but the characteristics of these texts are
different (e.g. shorter documents, different language used, single versus multiple
targets of opinions, different means of referencing targets).
In this chapter, we present the tasks and the methods we have proposed, in a
suitable manner, to tackle sentiment analysis in different text types. For each of the
textual genres considered, we have appropriately defined the task of sentiment
analysis, identified the genre peculiarities and proposed adequate methods to tackle
the issues found.
Where previous approaches fell short of correctly identifying the needs of the
specific textual genre, we proposed adequate formulations of the problem and
proposed specific methods to tackle them. Additionally, where insufficient
resources were available, we have developed new annotation schemes and new
corpora, for English and other languages (Spanish, German, Chinese).
This chapter is structured as follows: in Section 4.1 we present the methods and
resources we proposed and evaluated in the context of sentiment analysis from
product reviews. Subsequently, in Section 4.2., we discuss the issues involved in
sentiment analysis from newspaper articles, specifically in reported speech
extracted from news (i.e. quotations). In Section 4.3., we present a method to detect
sentiments expressed in political debates, studying the needs of a generic sentiment
analysis system that is able to deal with different topics, in a dialogue framework, in
which different sentiment sources and targets are present. Finally, in Section 4.4.,
we present the methods and resources we have developed for sentiment analysis
from blogs. Summing up all the experience gathered from the analysis of the
previous text types, we design a general method to tackle sentiment analysis, in the
context of this new and complex text type with a dialogue-like structure, in which
formal and informal language styles are mixed and sentiment expressions are highly
diverse.

44

4.1. OPINION MINING FROM PRODUCT REVIEWS


FEATURE-BASED OPINION MINING AND SUMMARIZATION
4.1.1. INTRODUCTION
Presently, the consumer market is flooded with products of the most varied sorts,
each being advertised as better, cheaper, more resistant, easy to use and fault free.
But is all what is advertised actually true? Certainly, all companies will claim that
the products they make are the best. In practice, however, each product will behave
better or worse depending on the users necessities and level of expertise, will have
certain capabilities depending on the price and product class. Moreover, for most
users, the choice will depend on reasons such as lower price, brand name or ratio
between price and performance.
Therefore, how can a person that is faced with the harsh decision of having to
choose among tens of products of the same type, with the same features, finally take
a rational decision? Since the world wide web contains a large quantity of product
reviews, in specialized review sites and user blogs, a good solution seems to be that
of lending an ear to this word-of-mouth on the Web (Liu, 2007), weighing the
pros and cons of the different products and finally buying the one satisfying the
personal needs and expectations.
The clear advantage in having at hand the volume of data present on the Internet
nowadays is that one is apt to obtain almost objective information on the products
that he/she is planning to buy. This process can be accomplished by viewing
different opinions that people who have previously bought the product in question
have on it, based on the experience in using it. Such people can comment on the
reasons motivating their choice of purchase and their present attitude towards that
decision. Thus, besides the objective information concerning price and product
capabilities, a prospect buyer can also have access to the subjective information
about a product. However, a high volume of information is also bound to bring
difficulty in sifting through it.
The ideal situation is that in which one is able to read all available user reviews
and create his/her opinion, depending on the feature(s) of interest and the value or
ratio of the feature attributes. The main problem then becomes the time spent in
reviewing all available data and the language barrier the fact that product reviews
are written in different languages.
The solution is a system that automatically analyzes and extracts the values of
the features for a given product, independent of the language the customer review is
written in. Such a system can then present the potential buyer with percentages of
positive and negative opinions expressed about each of the product features and
possibly make suggestions based on buyer preferences.
45

This can be achieved by performing sentiment analysis at the feature level, an


approach that is also known as feature-based opinion mining (Hu and Liu, 2004;
Liu, 2007). Previously mentioned by Dave et al. (2003), this task aims at extracting,
given an object (product, event, person etc.), the features of the object and the
opinions expressed in texts in relation to the features, classify the opinion words
and produce a final summary containing the percentages of positive versus negative
opinions expressed on each of the features. Feature-based opinion mining involves
a series of tasks:
Task 1: Identify and extract object features that have been commented on
by an opinion holder (e.g., a reviewer).
Task 2: Determine whether the opinions on the features are positive,
negative or neutral.
Task 3: Group feature synonyms.
Subsequently, once all the groups of words referring the same feature are
gathered and the polarity of the opinion is computed, the result is presented as a
percentage of positive versus negative opinion on each feature (feature-based
opinion summary of multiple reviews).
The approach we use is grounded on the feature-based opinion mining and
summarization paradigm, whose theoretical background has been described by Liu
(2007). Relevant research in feature-driven opinion summarization has been done
by Ding et al. (2008) and Dave et al. (2003).
The issues we have identified in this context is that present research has not
included the discovery of implicit features and furthermore, it has left the problem
of explicit features dependent on the mentioning of these features in the individual
user reviews or not. The authors describe approaches that are lexicon-based and
consist in discovering frequent features using association mining and determining
the semantic orientation of opinions as polarity of adjectives (as opinion holders)
that features are described by. The classification of adjectives is done starting with a
list of seeds and completing it using the WordNet synonymy and antonymy
relations. Infrequent features are deduced using the opinion holders. However, the
fact that there is no well-organized structure of features and sub-features of
products leads to the fact that, for example, the summarization of opinions is done
for 720 features for an mp3 player (Ding et al., 2008). The question that arises is:
would a user in a real-life situation be interested on whether the edges of a camera
are round or flat and what the previous buyers think about that, or would a potential
buyer like to see if the design of the product is fine or not, according to the many
criteria developed by buyers to assess this feature? The work does not approach
implicit features and does not classify the orientation of adjectives depending on the
context. A solution to the latter problem is presented by Ding et al. (2008) where
the authors take a holistic approach to classifying adjectives, that is, consider not
46

only the local context in which they appear next to the feature they determine, but
also other adjectives appearing with the feature and their polarity in different
contexts. Popescu and Etzioni (2005) employ a more complex approach for
feature-based summarization of opinions, by computing the web PMI (Pointwise
Mutual Information) statistics for the explicit feature extraction and a technique
called relaxation labeling for the assignation of polarity to the opinions. In this
approach, dependency parsing is used together with ten extraction rules that were
developed intuitively.
We propose an initial approach to the issue of feature-based opinion mining
(Balahur and Montoyo, 2008d; Balahur and Montoyo, 2008f), which we
subsequently extended in Balahur and Montoyo (2008c), by introducing product
technical details and comparing two different measures of term relatedness
(Normalized Google Distance (Cilibrasi and Vitanyi, 2006) and Latent Semantic
Analysis (Deerwester et al., 1990). On top of this system, we propose a method to
recommend products based on the scores obtained for the different quantified
features in the opinion mining step. The method was presented in (Balahur and
Montoyo, 2008b).
In the light of the fact that no standard annotation was available for featurebased opinion mining, in Balahur and Montoyo (2009) we proposed an annotation
scheme that aimed at standardizing the labeling of reviews, so that different types of
mentions of features (direct, indirect, implicit) and the different manner of
expressing opinions (through subjective or objective statements) can be correctly
labeled. Subsequently, we studied methods to infer sentiment expressed on different
features using examples of annotations from a review corpus and Textual
Entailment (Balahur and Montoyo, 2009).
In the following sections we describe our approaches and results. The method
we propose is language and customer-review independent. It extracts a set of
general product features, finds product specific features and feature attributes and is
thus applicable to all possible reviews in a product class. The approaches we
present in this section were summarized by Balahur et al. (2010).

4.1.2. METHODS FOR MINING SENTIMENT FROM PRODUCT


REVIEWS
Our method consists of two distinct steps: preprocessing and main processing, each
containing a series of sub-modules and using different language tools and resources.

47

Retrieve
reviews

User query

Reviews
R
Extract

Identify
language

Product
Class
Product class
Obtain
product
features

English productt
features
Translate
features

ConceptNet

English
Reviews

Spanish
Reviews

WordNet

EuroWordNet

Spanish productt
features

Figure 4.1: Preprocessing Stage

A PREPROCESSING
In our approach, we start from the following scenario: a user enters a query about a
product that he/she is interested to buy.
The search engine retrieves a series of documents containing the product name,
in different languages. Further on, two parallel operations are performed: the first
one uses the Lextek 11 language identifier software to filter and obtain two
categories one containing the reviews in English and the other the reviews in
Spanish.
The second operation implies a modified version of the system proposed by
Kozareva et al. (2007) for the classification of person names. We use this system in
order to determine the category that the product queried belongs to (e.g. digital
camera, laptop, printer, book). Once the product category is determined, we proceed
to extracting the product specific features and feature attributes. This is
accomplished using WordNet and ConceptNet and the corresponding mapping to
Spanish using EuroWordNet. Apart from the product specific class of features and
feature attributes, we consider a core of features and feature attributes that are
product-independent and whose importance determines their frequent occurrence in
customer reviews. Figure 4.1 describes the components used in the preprocessing
stage.
11

http://www.lextek.com/langid/

48

Product-independent features and feature attributes


There are a series of features that are product independent and that are important to
any prospective buyer. We consider these as forming a core of product features. For
each of these concepts, we retrieve from WordNet the synonyms which have the
same Relevant Domain (Vzquez et al., 2004), the hyponyms of the concepts and
their synonyms and attributes, respectively.
Using WordNet to extract product specific features and feature attributes
Once the product category has been identified, we use WordNet to extract the
product specific features and feature attributes. We accomplish this in the following
steps:
1. For the term defining the product category, we search its synonyms in
WordNet (Fellbaum , 1999).
2. We eliminate the synonyms that do not have the same top relevant domain
(Vzquez et al., 2004) as the term defining the product category.
3. For the term defining the product, as well as each for each of the remaining
synonyms, we obtain their meronyms from in WordNet, which constitute
the parts forming the product.
4. Since WordNet does not contain much detail on the components of most of
new technological products, we use ConceptNet to complete the process of
determining the specific product features. We explain the manner in which
we use ConceptNet in the following section. After performing the steps
described above, we conclude the process of obtaining the possible terms
that a customer buying a product will comment on. The final step consists
in finding the attributes of the features discovered by applying the has
attributes relation in WordNet to each of the nouns representing product
features. In the case of nouns which have no term associated by the has
attribute relation, we add as attribute features the concepts found in
ConceptNet under the OUT relations PropertyOf and CapableOf. In case
the concepts added are adjectives, we further add their synonyms and
antonyms from WordNet.
As result we have for example, in the case of photo, the parts raster and
pixel with the attributes blurry, clarity, sharp.
Using ConceptNet to extract product-specific features and feature attributes
ConceptNet (Liu and Singh, 2004) is a freely available commonsense
knowledgebase and NLP toolkit which supports many practical textual reasoning
tasks over real-world documents. Commonsense knowledge in ConceptNet
encompasses the spatial, physical, social, temporal, and psychological aspects of
49

everyday life. It contains relations such as CapableOf, ConceptuallyRelatedTo, IsA,


LocationOf etc. In order to obtain additional features for the product in question, we
add the concepts that are related to the term representing the concept with terms
related in ConceptNet by the OUT relations UsedFor and CapableOf and the IN
relations PartOf and UsedFor. For example, for the product camera, the OUT
UsedFor and CapableOf relations that will added are take picture, take
photograph, photography, create image, record image and for the IN PartOf
and UsedFor relations shutter, viewfinder, flash, tripod.
Mapping concepts using EuroWordNet
EuroWordNet 12 (EWN) is a multilingual database with WordNets for different
European languages (Dutch, Italian, Spanish, German, French, Czeck and
Estonian). Each language has its own designed WordNet, structured as the
Princeton WordNet. Having these connections, it is possible that parting from one
word, one can consult similar words in any other language of the EWN. The main
advantage in using this lexical resource is that all the terms discovered in one
language can be easily mapped to another language. We employ EuroWordNet and
map the features and feature attributes, both from the main core of words, as well as
the product specific ones that were previously discovered for English, independent
of the sense number, taking into account only the preservation of the relevant
domain. Certainly, we are aware of the noise introduced by this mapping, however
in the preliminary research we found that the concepts introduced that had no
relation to the product queried did not appear in the user product reviews.
Discovering overlooked product features
The majority of product features we have identified so far are parts constituting
products. However, there remains a class of undiscovered features that are
indirectly related to the product. These are the features of the product constituting
parts, such as battery life, picture resolution, auto mode. Further, we
propose to extract these overlooked product features by determining bigrams made
up of target words constituting features and other words in a corpus of customer
reviews. In the case of digital cameras, for example, we considered a corpus of 200
customer reviews on which we ran Pedersens Ngram Statistics Package to
determine target co-occurrences of the features identified so far. As measure for
term association, we use the Pointwise Mutual Information (PMI) score, which is
calculated according to the following formula:

12

http://www.illc.uva.nl/EuroWordNet/

50

where x and y are two words and P(x) stands for the probability of the word x
occurring in the corpus considered. In this manner, we discover bigram features
such as battery life, mode settings and screen resolution.
B

MAIN PROCESSING

The main processing in our system is done in parallel for English and Spanish. In
the next section, we will briefly describe the steps followed in processing the initial
input containing the customer reviews in the two considered language and offer as
output the summarized opinions on the features considered. Figure 4.2 presents the
steps included in the processing.
We start from the reviews filtered according to language. For each of the two
language considered, we used a specialized tool for anaphora resolution- JavaRAP
for English and SUPAR (Ferrndez et al., 1999) for Spanish. Further on, we
separate the text into sentences and use a Named Entity Recognizer to spot names
of products, brands or shops.
Using the lists of general features and feature attributes, product-specific
features and feature attributes, we extract from the set of sentences contained in the
text only those containing at least one of the terms found in the lists.
Reviews

EN

ES

Anaphora Resolution
JavaRAP

SUPAR

Chunking & NER


LingPipe

Freeling

Sentence Parsing & Extraction


Minipar

Freeling

EN
features

ES
features

Feature Value Extraction

EN

Results

ES

Figure 4.2: System Architecture


51

Anaphora resolution
In order to solve the anaphoric references on the product features and feature
attributes, we employ two anaphora resolution tools - JavaRAP for English and
SUPAR for Spanish. Using these tools, we replace the anaphoric references with
their corresponding referents and obtain a text in which the terms constituting
product features could be found.
JavaRAP is an implementation of the classic Resolution of Anaphora Procedure
(RAP) given by Lappin and Leass (1994). It resolves third person pronouns, lexical
anaphors, and identifies pleonastic pronouns.
Using JavaRAP, we obtain a version of the text in which pronouns and lexical
references are resolved. For example, the text: I bought this camera about a week
ago, and so far have found it very very simple to use, takes good quality pics for
what I use it for (outings with friends/family, special events). It is great that it
already comes w/ a rechargeable battery that seems to last quite a while..., by
resolving the anaphoric pronominal reference, becomes I bought this camera
about a week ago, and so far have found <this camera> very very simple to use,
takes good quality pics for what I use <this camera> for (outings with friends/family,
special events). It is great that <this camera> already comes w/a rechargeable
battery that seems to last quite a while....
For the anaphora resolution in Spanish, we employ SUPAR (Slot Unification
Parser for Anaphora Resolution). The architecture of SUPAR contains, among
others, a module solving the linguistic problems (pronoun anaphora, element
extraposition, ellipsis, etc.). We use SUPAR in the same manner as JavaRAP, to
solve the anaphora for Spanish. Sentence chunking and NER Further on, we split
the text of the customer review into sentences and identify the named entities in the
text. Splitting the text into sentences prevents us from processing sentences that
have no importance as far as product features that a possible customer could be
interested in are concerned.
Chunking and Named Entity Resolution
LingPipe13 is a suite of Java libraries for the linguistic analysis of human language.
It includes features such as tracking mentions of entities (e.g. people or proteins),
part-of-speech tagging and phrase chunking.
We use LingPipe to split the customer reviews in English into sentences and
identify the named entities referring to products of the same category as the product
queried. In this manner, we can be sure that we identify sentences referring to the
13

http://alias-i.com/lingpipe/

52

product queried, even the reference is done by making use of the name of another
product. For example, in the text For a little less, I could have bought the Nikon
Coolpix, but it is worth the extra money., anaphora resolution replaces <it> with
<Nikon Coolpix> and this step will replace it with <camera>.
The FreeLing 14 package consists of a library providing language analysis
services. The package offers many services, among which text tokenization,
sentence splitting, POS-tagging, WordNet based sense annotation and rule-based
dependency parsing. We employ FreeLing in order to split the customer reviews in
Spanish into sentences and identify the named entities referring to products of the
same category as the product queried.
Sentence extraction
Having completed the feature and feature attributes identification phase, we
proceed to extracting for further processing only the sentences that contain the
terms referring to the product, product features or feature attributes. In this manner,
we avoid further processing of text that is of no importance to the task we wish to
accomplish. For example, sentences of the type I work in the home appliances
sector will not be taken into account in further processing. Certainly, at the overall
level of review impact, such a sentence might be of great importance to a reader,
since it proves the expertise of the opinion given in the review. However, for the
problems we wish to solve by using this method, such a sentence is of no
importance.
Sentence parsing
Each of the sentences that are filtered by the previous step are parsed in order to
obtain the sentence structure and component dependencies. In order to accomplish
this, we use Minipar (Lin, 1998) for English and FreeLing for Spanish. This step is
necessary in order to be able to extract the values of the features mentioned based
on the dependency between the attributes identified and the feature they determine.
Feature value extraction
Further on, we extract features and feature attributes from each of the identified
sentences, using the following rules:
1. We introduce the following categories of context polarity shifters (Polanyi
and Zaenen, 2004), in which we split the modifiers and modal operators in
two categories i.e. positive and negative:
! negation: no, not, never, etc.
14

http://nlp.lsi.upc.edu/freeling/

53

modifiers: positive (extremely, very, totally, etc.) and negative


(hardly, less, possibly, etc.)
! modal operators: positive (must, has) and negative (if, would,
could, etc.)
2. For each identified feature that is found in a sentence, we search for a
corresponding feature attribute that determines it. Further on, we search to
see if the feature attribute is determined by any of the defined modifiers.
We consider a variable we name valueOfModifier, with a default value of 1, that will account for the existence of a positive or negative modifier of
the feature attribute. In the affirmative case, we assign a value of 1 if the
modifier is positive and a value of 0 if the modifier is negative. If no
modifier exists, we consider the default value of the variable. We extract
triplets of the form (feature, attributeFeature, valueOfModifier). In order to
accomplish this, we use the syntactic dependency structure of the phrase,
we determine all attribute features that determine the given feature (in the
case of Minipar, they are the ones connected by the det or mod
relation).
3. If a feature attribute is found without determining a feature, we consider it
to implicitly evoke the feature that it is associated with in the feature
collection previously built for the product. The camera is small and
sleek. becomes (camera, small, -1) and (camera, sleek, -1), which is then
transformed by assigning the value small to the size feature and the
value sleek to the design feature.
!

Assignment of Polarity to Feature Attributes


In order to assign polarity to each of the identified feature attributes of a product,
we employ Support Vector Machines Sequential Minimal Optimization (SVM
SMO) machine learning (Platt, 1998) and the Normalized Google Distance (NGD).
The main advantage in using this type of polarity assignment is that NGD is
language independent and offers a measure of semantic similarity taking into
account the meaning given to words in all texts indexed by Google from the world
wide web. The set of anchors contains the terms {featureName, happy, unsatisfied,
nice, small, buy}, that have possible connection to all possible classes of products.
Further on, we build the classes of positive and negative examples for each of
the feature attributes considered. From the corpus of annotated customer reviews,
we consider all positive and negative terms associated to the considered attribute
features. We then complete the lists of positive and negative terms with their
WordNet synonyms. Since the number of positive and negative examples must be
equal, we will consider from each of the categories a number of elements equal to
the size of the smallest set among the two, with a size of at least 10 and less or
54

equal with 20. We give as example the classification of the feature attribute tiny,
for the size feature. The set of positive feature attributes considered contains 15
terms (e.g. big, broad, bulky, massive, voluminous, large-scale, etc.) and the set of
negative feature attributes considered is composed as opposed examples, such as
(small, petite, pocket-sized, little, etc).
We use the anchor words to convert each of the 30 training words to 6dimensional training vectors defined as v(j,i) = NGD(wi,aj), where aj with j ranging
from 1 to 6 are the anchors and wi, with i from 1 to 30 are the words from the
positive and negative categories.
After obtaining the total 180 values for the vectors, we use SVM SMO to learn
to distinguish the product specific nuances. For each of the new feature attributes
we wish to classify, we calculate a new value of the vector
vNew(j,word)=NGD(word, aj), with j ranging from 1 to 6 and classify it using the
same anchors and trained SVM model.
In the example considered, we had the following results (we specify between
brackets the word to which the scores refer to:
(small)1.52,1.87,0.82,1.75,1.92,1.93,positive
(little)1.44,1.84,0.80,1.64,2.11,1.85,positive
(big)2.27,1.19,0.86,1.55,1.16,1.77,negative
(bulky)1.33,1.17,0.92,1.13,1.12,1.16,negative

The vector corresponding to the tiny attribute feature is:


(tiny)1.51,1.41,0.82,1.32,1.60,1.36.

This vector was classified by SVM as positive, using the training set specified
above. The precision value in the classifications we made was between 0.72 and
0.80, with a kappa value above 0.45.
For each of the features identified, we compute its polarity depending on the
polarity of the feature attribute that it is determined by and the polarity of the
context modifier the feature attribute is determined by, in case such a modifier
exists. Finally, we statistically summarize the polarity of the feature attributes, as
ratio between the number of positive quantifications and the total number of
quantifications made in the considered reviews to that specific feature and as ratio
between the number of negative quantifications and the total number of
quantifications made in all processed reviews. The formulas can be summarized in:
1. pos 

2. neg 

pos feature attributes


feature atributes

neg feature attributes


feature atributes

55

The results shown are triplets of the form (feature, percentagePositiveOpinions,


percentageNegativeOpinions).

C. DISCUSSION AND EVALUATION


For the evaluation of the system, we annotated a corpus of 50 customer reviews for
each language, collected from sites as amazon.com, newegg.com, dealsdirect.com,
ciao.es, shopmania.es, testfreaks.es and quesabesde.com. The corpus was annotated
at the level of feature attributes, by the following scheme:
<attribute>(name of attribute)
<feature>(feature it determines)</feature>
<value>(positive/ negative)</value>
</attribute>

It is difficult to evaluate the performance of such a system, since we must take into
consideration both the accuracy in extracting the features that reviews comment on,
as well as the correct assignation of identified feature attributes to the positive or
negative category.
The formula used in measuring the accuracy of the system represented the
normalized sum of the ratios between the number of identified positive feature
attributes and the number of existing positive attributes and the ratio of identified
negative feature and the total number of negative feature attributes for each of the
considered features existing in the text.
Secondly, we compute the Feature Identification Precision (P) as ratio between
the number of features correctly identified from the features identified and the
number of identified features.
Thirdly, we compute the Feature Identification Recall (R) as the number of
correctly identified features from the features identified and the number of correctly
identified features. The results obtained are summarized in Table 4.1.
We show the scores for each of the two languages considered separately and the
combined score when using both systems for assigning polarity to feature attributes
of a product. In the last column, we present a baseline, calculated as average of
using the same formulas, but taking into consideration, for each feature, only the
feature attributes we considered as training examples for our method.

56

Feature
Baseline
extraction
English
Spanish
Combined
English
performance
Accuracy
0.82
0.80
0.81
0.21
Precision
0.80
0.78
0.79
0.20
Recall
0.79
0.79
0.79
0.40
Table 4.1: System results on the annotated review corpus

Baseline
Spanish
0.19
0.20
0.40

We can notice how the use of NGD helped the system acquire significant new
knowledge about the polarity of feature attributes, in the context.
There are many aspects to be taken into consideration when evaluating a system
identifying features, opinion on features and summarizing the polarity of features.
First of all, customers reviewing products on the web frequently use informal
language, disregard spelling rules and punctuation marks.
At times, phrases are pure enumerations of terms, containing no subject or
predicate. In this case, when there is no detectable dependency structure between
components, an alternative method should be employed, such as verifying if the
terms appearing near the feature within a window of specified size are frequently
used in other contexts with relation to the feature. Secondly, there are many issues
regarding the accuracy of each of the tools and language resources employed and a
certain probability of error in each of the methods used. In this initial research, we
presented a method to extract, for a given product, the features that could be
commented upon in a customer review.
Further, we have shown a method to acquire the feature attributes on which a
customer can comment in a review. Moreover, we presented a method to extract
and assign polarity to these product features and statistically summarize the polarity
they are given in the review texts in English and Spanish. The method for polarity
assignment is largely language independent (it only requires the use of a small
number of training examples) and the entire system can be implemented in any
language for which similar resources and tools as the ones used for the presented
system exist.
The main advantage obtained by using this method is that one is able to extract
and correctly classify the polarity of feature attributes, in a product dependent
manner. Furthermore, the features in texts are that are identified are correct and the
percentage of identification is high. Not lastly, we employ a measure of word
similarity that is in itself based on the word-of-mouth on the web. The main
disadvantage consists in the fact that SVM learning and classification is dependent
on the NGD scores obtained with a set of anchors that must previously be
established. This remains a rather subjective matter. Also, the polarity given in the
57

training set determines the polarity given to new terms, such that large in the
context of display will be trained as positive and in the case of size as negative.
However, there are many issues that must be addressed in systems identifying
customer opinions on different products on the web. The most important one is that
concerning the informal language style, which makes the identification of words
and dependencies in phrases sometimes impossible.

4.1.3. CLASSIFYING OPINION ON PRODUCTS USING


RELATEDNESS SCORES ON SPECIFIC CORPORA AND EMOTION
EXPRESSION PATTERNS
Subsequently to this first approach, we improved the system by adding extra
features, taking into consideration the product technical specifications and defining
patterns for indirectly expressed opinions using WordNet Affect categories
(Balahur and Montoyo, 2008c), as well as enriching our feature-dependent method
of opinion classification using Latent Semantic Analysis relatedness scores; in the
case of LSA scores, the context is given by the corpus from which the model is
learnt, as opposed to the NGD, which is computed at the web level. We show the
manner in which all these factors influence the system performance and at what
cost. Last, but not least, many of the opinions on products are expressed in an
indirect manner, that is, not relating the product or its features with polarity words,
but expressing an emotion about them. We propose a set of patterns to extract such
indirectly expressed opinions using the emotion lists from WordNet Affect.
Our solution to the problem of feature attributes classification is using machine
learning with two measures of similarity. On the one hand, we employ the
Normalized Google Distance, which gives a measure of the strength of relationship
between two considered words at the level of the entire Web and on the other hand,
we use the LSA, which gives the same measure of strength, but at a local corpus
level. Classifying the feature attributes according to these scores and taking into
consideration 6 anchor words that relate each word with the feature and known
polarities, we show how the classification of feature attributes can be done in the
feature context.
In the reviews to be mined and summarized, however, other opinion words can
be found and other manners of expressing opinion can be encountered, such as
those describing emotional states related to the product (e.g. I love this camera)
or to using it. The solution we give to this problem is to propose a list of patterns to
extract from the reviews such phrases containing emotion words, which are used to
express opinions of the different product features using descriptions of emotions
felt. The words related to emotion are taken from WordNet Affect.
In the evaluation section, we show how the use of such patterns raised with 12%
the recall of the system, while the precision of classification rose to the same
58

degree. In our previous approach, in order to assign polarity to each of the identified
feature attributes of a product, we employed SVM SMO machine learning and the
NGD. In this approach, we complete the solution with a classification employing
Latent Semantic Analysis with Support Vector Machines classification. For this, we
build the classes of positive and negative examples for each of the feature attributes
considered. From the list of classified feature attributes in the pros and cons
reviews, we consider all positive and negative terms associated to the considered
attribute features. We then complete the lists of positive and negative terms with
their WordNet synonyms. Since the number of positive and negative examples must
be equal, we will consider from each of the categories a number of elements equal
to the size of the smallest set among the two, with a size of at least 10 and less or
equal with 20.
We give as example the classification of the feature attribute tiny, for the
size feature. The set of positive feature attributes considered contains 15 terms
such as big, broad, bulky, massive, voluminous, large-scale etc. and
the set of negative feature attributes considered is composed as opposed examples,
such as small, petite, pocket-sized, little etc. We use the anchor words to
convert each of the 30 training words to 6-dimensional training vectors defined as
v(j,i) = LSA(wi,aj), where aj with j ranging from 1 to 6 are the anchors and wi, with i
from 1 to 30 are the words from the positive and negative categories. After
obtaining the total 180 values for the vectors, we use SVM SMO to learn to
distinguish the product specific nuances. For each of the new feature attributes we
wish to classify, we calculate a new value of the vector vNew(j,word) = LSA(word,
aj), with j ranging from 1 to 6 and classify it using the same anchors and trained
SVM model. We employed the classification on the corpus present for training in
the Infomap software pack. The blank lines represent the words which were not
found in the corpus; therefore a LSA score could not be computed. The results are
presented in Table 4.2. On the other hand, we employed the classification on a
corpus made up of reviews on different electronic products, gathered using the
Google API and a site restriction on amazon.com. In the table below, we show an
example of the scores obtained with LSA on the features attributes classified for the
feature size. The vector for the feature attribute tiny was classified by SVM as
positive, using the training set specified above. The results are presented in Table
4.3.
Feature
V1
V2 V3 V4
V5 V6 Polarity
attribute
small
0.76 0.74 --- 0.71
1
0.71
pos
big
0.80 0.75 --- 0.74 0.73 0.68
neg
bulky
----- --- ------pos
little
----- --- ------neg
59

Feature
V1
V2 V3 V4
V5 V6 Polarity
attribute
tiny
0.81 0.71 --- 0.80 0.73 0.72
--Table 4.2: LSA scores on non-specialized corpus (not only with product reviews)
In Table 4.3, we show an example of the scores obtained with the similarity
given by the LSA scores on a specialized corpus of reviews on products. The vector
for the feature attribute tiny was classified by SVM as positive, using the training
set specified above.
Feature
V1
V2
V3
V4
V5
V6
Polarity
attribute
small
0.83
0.77
0.48
0.72
1
0.64
pos
big
0.79
0.68
0.74
0.73
0.77
0.71
neg
bulky
0.76
0.67
0.71
0.75
0.63
0.78
pos
little
0.82
0.76
0.52
0.71
0.83
0.63
neg
tiny
0.70
0.70
0.65
0.67
0.71
0.71
pos
Table 4.3: LSA scores on a specialized corpus of product reviews
Precision values in classifications we made with NGD and LSA for different
product features for the examples of digital camera reviews and the mobile phones
reviews vary from 0.75 to 0.8 and kappa statistics shows high confidence of
classification (Balahur and Montoyo, 2008c).
The conclusion that can be drawn from the results presented is that the main
advantage in using the first method of polarity assignment is that NGD is language
independent and offers a measure of semantic similarity taking into account the
meaning given to words in all texts indexed by Google from the World Wide Web.
On the other hand, using the whole Web corpus can also add significant noise.
Therefore, we employ Latent Semantic Analysis at a local level, both on a nonspecialized corpus, as well as on a corpus containing customer reviews. As we will
show, the classification using LSA on a specialized corpus brings an average of 8%
of improvement in the classification of polarity and a rise of 0.20 in the kappa
measure, leading to an 8% overall improvement in the precision of the
summarization system. However, these results were obtained using a specialized
corpus of opinions, which was previously gathered from the Web. To this respect, it
is important to determine sources (web sites, blogs or forums) specific to each of
the working languages, from which to gather the corpus on which the LSA model
can be built.
Using LSA on a non-specialized corpus improved the classification to the same
degree as the classification on a specialized corpus in the cases where the specific
pairs of words to be classified were found in the corpus. However, in 41% of the
60

cases, the classification failed due to the fact that the words we tried to classify
were not found in the corpus. Further on, we developed a method for feature
polarity extraction using subjective phrases.
As observed before, some opinions on the product or its features are expressed
indirectly, with subjective phrases containing positive or negative emotions which
are related to the product name, product brand or its features. In order to identify
those phrases, we have constructed a set of rules for extraction, using the emotion
lists from WordNet Affect. For the words present in the joy emotion list, we
consider the phrases extracted as having a positive opinion on the product or the
feature contained. For the words in the anger, sadness and disgust emotion
lists, we consider the phrases extracted as having a negative opinion on the product
or the feature contained. Apart from the emotion words, we have considered a list
of positive words (pos list), containing adverbs such as definitely, totally,
very, absolutely and so on - as words positively stressing upon an idea - (Iftene
and Balahur-Dobrescu, 2007), that influence on the polarity of the emotion
expressed and that are often found in user reviews.
We present the extraction rules in Table 4.4 (verb emotion, noun emotion and
adj emotion correspond to the verbs, nouns and adjectives, respectively, found in
the emotion lists from WordNet Affect under the emotions joy, sadness,
anger and disgust). In case of surprise, as emotion expressed about a product
and its features, it can have both a positive, as well as negative connotation.
Therefore, we have chosen not to include the terms expressing this emotion in the
extraction patterns.
1. I [pos list*][verb emotion][this||the||my] [product name||product feature]
2. I ([am||m||was||feel||felt])([pos list**])[adj emotion][with||about||by]
[product name||product feature]
3. I [feel||felt][noun emotion][about||with][product name ||product brand]
4. I [pos list*][recommend][this||the][product name||product brand]
5. I ([dont])[think ||believe][sentence***]
6. It [s||is] [adj
feature][product action]

emotion]

[how||what][product

name||product

7.
You
||Everybody||Everyone||All||He||She||They][will||would][verb
emotion][this||the][product name brand||feature]
Table 4.4: List of patterns for opinion extraction based on emotion clues
61

We have performed a comparative analysis of the system employing the SVM


SMO polarity classification using NGD and LSA on a specialized corpus, the
subjective phrases and combined, with the corpus used by Balahur and Montoyo
(Balahur and Montoyo, 2008c) and also the corpus of 5 reviews from (Hu and Liu,
2004). Results obtained in Table 4.5 are obtained when evaluating on our own
annotated corpus, in terms of Precision (P) and Recall (R).

NGD

LSA

Rules

NGD+
Rules

LSA+
Rules

P
R
P
R
P
R
P
R
P
R
0.80 0.79 0.88 0.87 0.32 0.6
0.89
0.85
0.93
0.93
Table 4.5: System results on the review test set in Balahur and Montoyo( 2008c)
In the case of the 5-review corpus proposed by Hu and Liu (2004), the
observation that is important to make is that, as opposed to the annotation made in
the corpus, we have first mapped the features identified to the general feature of the
product (for example fit refers to size and edges refers to design), as we
relieve that in real life situations, a user benefits more from a summary on coarser
classes of product features.

NGD

LSA

Rules

NGD+
Rules

LSA+
Rules

P
R
P
R
P
P
R
P
R
0.81 0.80 0.85 0.88 0.28 0.5
0.89
0.85
0.93
Table 4.6: System results on the corpus employed by Hu and Liu (2004)

P
0.93

Also, a set of sentences that were not annotated in the corpus, such as Youll
love this camera, which expresses a positive opinion on the product. The results
shown in Table 4.6 are compared against the baseline of 0.20 precision and 0.41
recall, which was obtained using only the features determined by Balahur and
Montoyo (Balahur and Montoyo, 2008f) and the feature attributes whose polarity
was computed from the pros and cons-style reviews. As it can be seen, the best
results are obtained when using the combination of LSA with the rules for
subjective phrases extraction. However, gathering the corpus for the LSA model
can be a costly process, whereas NGD scores are straightforward to be obtained and
classifying is less costly as time and resources used.
What is interesting to study is the impact of employing LSA for gradual learning
62

and correction of a system that uses NGD for classifying the polarity of feature
attributes. In such a self-learning scheme, the online classification would be that
of NGD. However, the classification of the new feature attributes can be later
improved offline using the classification given by LSA, which can then be used
as better training for learning the polarity of new feature attributes by the online
NGD classification.
From this subsequent research, we could draw some conclusions on the
advantages and disadvantages of using different scenarios for computing opinion
polarity. The main advantage in using polarity assignment depending on NGD
scores is that this is language independent and offers a measure of semantic
similarity taking into account the meaning given to words in all texts indexed by
Google from the World Wide Web. The main advantage in using LSA on a
specialized corpus, on the other hand, is that it eliminates the noise given by the
multiple senses of words. We completed the opinion extraction on different product
features with rules using the words present in WordNet Affect, as indicative of
indirectly expressed opinions on products. We showed how all the employed
methods led to significant growth in the precision and recall of our opinion mining
and summarization system.

4.1.4. FURTHER IMPROVEMENTS TO SENTIMENT ANALYSIS IN


REVIEWS THE ISSUE OF DISAMBIGUATING ADJECTIVE
POLARITY USING THE LOCAL CONTEXT
As we have seen in the previous section, one of the challenges faced in opinion
mining is the fact that some adjectives have a different polarity depending on the
context in which they appear.
Our initial approach consisted in classifying adjective polarity using a set of
anchors and the NGD and LSA scores, respectively. However, this component was
not evaluated separately. With the participation in the SemEval-2 Task Number 18
Disambiguation of Sentiment Ambiguous Adjectives (Wu and Jin, 2010), we
aimed at extending the feature-based opinion mining system to resolve this task in a
more general opinion mining scenario. Thus, we aimed at proposing a suitable
method to tackle this issue, in a manner that is independent from the feature-based
opinion mining framework. We named the system participating in this competition
OpAL (Balahur and Montoyo, 2010).
In this task, the participants were given a set of contexts in Chinese, in which 14
dynamic sentiment ambiguous adjectives are selected. They were: |big, |small,
|many, |few, |high, |low, |thick, |thin, |deep, |shallow, |heavy,
|light, |huge, |grave. The task was to automatically classify the polarity

63

of these adjectives, i.e. to detect whether their sense in the context is positive or
negative.
Our approach is based on three different strategies: a) the evaluation of the
polarity of the whole context using an opinion mining system; b) the assessment of
the polarity of the local context, given by the combinations between the closest
nouns and the adjective to be classified; c) rules aiming at refining the local
semantics through the spotting of modifiers. The final decision for classification is
taken according to the output of the majority of these three approaches. The method
used yielded good results, the OpAL system run achieving approximately 76%
micro accuracy on a Chinese corpus. In the following subsections, we explain more
in detail the individual components employed.
A. THE OPAL OPINION MINING COMPONENT
First, we process each context using Minipar15. We compute, for each word in a
sentence, a series of features, computed from the NTCIR 7 data and the EmotiBlog
annotations. These words are used to compute vectors of features for each of the
individual contexts:
! the part of speech (POS)
! opinionatedness/intensity - if the word is annotated as opinion word, its
polarity, i.e. 1 and -1 if the word is positive or negative, respectively and
0 if it is not an opinion word, its intensity (1, 2 or 3) and 0 if it is not a
subjective word
! syntactic relatedness with other opinion word if it is directly dependent
of an opinion word or modifier (0 or 1), plus the polarity/intensity and
emotion of this word (0 for all the components otherwise)
! role in 2-word, 3-word, 4-word and sentence annotations:
opinionatedness, intensity and emotion of the other words contained in
the annotation, direct dependency relations with them if they exist and 0
otherwise.
We add to the opinion words annotated in EmotiBlog the list of opinion words
found in the Opinion Finder, Opinion Finder, MicroWordNet Opinion, General
Inquirer, WordNet Affect, emotion triggers lexical resources. We train the model
using the SVM SMO implementation in Weka16.

15
16

http://webdocs.cs.ualberta.ca/~lindek/minipar.htm
http://www.cs.waikato.ac.nz/ml/weka/

64

B. ASSESSING LOCAL POLARITY USING GOOGLE QUERIES


This approach aimed at determining the polarity of the context immediately
surrounding the adjective to be classified. To that aim, we constructed queries using
the noun found before the adjective in the context given, and issued six different
queries on Google, together with six pre-defined adjectives whose polarity is known
(3 positive - positive, beautiful, good and 3 negative negative, ugly,
bad). The form of the queries was noun+adjective+AND+pre-defined
adjective. The local polarity was considered as the one for which the query issued
the highest number of total results (total number of results for the 3 queries
corresponding to the positive adjectives or to the negative adjectives, respectively).
C. MODIFIER RULES FOR CONTEXTUAL POLARITY
This rule accounts for the original, most frequently used polarity of the given
adjectives (e.g. high is positive, low is negative). For each of them, we define its
default polarity. Subsequently, we determine whether in the window of 4 words
around the adjective there are any modifiers (valence shifters). If this is the case,
and they have an opposite value of polarity, the adjective is assigned a polarity
value opposite from its default one (e.g. too high is negative). We employ a list of
82 positive and 87 negative valence shifters.
D. EVALUATION RESULTS FOR OPAL IN THE SEMEVAL-2010
TASK 18
Table 4.7 presents the results obtained by the 16 participating systems including
OpAL- in the SemEval 2010 Task 18 competition. As it can be seen in this table,
the system ranked fifth, with a Micro accuracy of 0.76037 and sixth, with a Macro
accuracy of 0.7037. The data is reproduced from Wu and Jin (2010).
System
YSC-DSAA
HITSZ_CITYU_1
HITSZ_CITYU_2
Dsaa
OpAL
CityUHK4
CityUHK3
HITSZ_CITYU_3
QLK_DSAA_R

Micro Acc.(%) Macro Acc.(%)


94.20
92.93
93.62
95.32
93.32
95.79
88.07
86.20
76.04
70.38
72.47
69.80
71.55
75.54
66.58
62.94
64.18
69.54
65

System
Micro Acc.(%) Macro Acc.(%)
CityUHK2
62.63
60.85
CityUHK1
61.98
67.89
QLK_DSAA_NR
59.72
65.68
Twitter Sentiment
59.00
62.27
Twitter Sentiment_ext
56.77
61.09
Twitter Sentiment_zh
56.46
59.63
Biparty
51.08
51.26
Table 4.7: Results for the 16 system runs submitted (micro and macro accuracy)
Since the gold standard was not provided, we were not able to perform an
exhaustive analysis of the errors. However, from a random inspection of the system
results, we could see that a large number of errors was due to the translation
through which modifiers are placed far from the word they determine or the words
are not translated with their best equivalent.

4.1.5. CREATING AN ANNOTATION SCHEME FOR FEATUREBASED OPINION MINING


Subsequently to the initial research we have done in feature-based opinion mining,
we proposed a new annotation scheme that is appropriated for the task and the
elements that should be contemplated when tackling it (Balahur and Montoyo,
2009). Our motivations for designing and implementing this scheme, as well as
creating new corpora that is labeled using this scheme are multiple.
The first one is the lack of a fine-grained opinion annotation scheme for reviews
that takes into consideration the characteristics of this type of writing, both at the
opinion versus fact level, as well as, within the opinion category, among the
different methods to express it. Our first contribution is thus the definition of such
an annotation scheme that encloses all the elements that should be considered at the
time of opinion mining.
Further on, in order to evaluate systems implementing the paradigm of featurebased opinion mining, there are two available corpora, which are small, only
concentrate on electronic products and are annotated at a sentence level, using a
simple scheme under the form of [feature name, polarity, value]. Our second
contribution is annotating a corpus of 100 reviews in English of different product
categories (ranging from electro domestics to restaurants and books), using the
created annotation scheme. We make this corpus available to further research.
Thirdly, we evaluate and validate our corpus annotations for fact versus opinion,
using n-gram similarity to the annotated phrases and further classify opinionated
66

sentences according to the polarity expressed, using n-gram similarity to the


annotated sentences.
Finally, our contribution resides in proposing a method to use textual entailment
in the opinion mining task. As far as we are aware of, such research has not been
done so far. From the categories on which stars are given, we generate short
positive and negative phrases (e.g. Ease of Use This is easy to use versus This
is not easy to use.) To these short hypotheses that we want to verify, we add
examples of positive and negative phrases from the annotated corpus, which are
related to the category. We test the entailment relation in a window of three
consecutive phrases. The results obtained are encouraging.
OPINION ANNOTATION SCHEME
In order to train and test a system performing feature-based opinion mining, an
annotated corpus is needed. However, in this field, of feature-based opinion mining,
there are just two corpora available, annotated in a simple manner (for each
sentence, the feature mentioned and the associated positive/negative score), both
containing a small number of products and all pertaining to the same category (Hu
and Liu, 2004; Ding et al., 2008). Another corpus developed for the more general
opinion field is the Multi-Perspective Questioning Answering (MPQA) one (Wiebe
et al., 2005), separating among the subjective and objective aspects of the annotated
texts; however, this corpus only contains newspaper articles and does not take into
consideration the aspects of product characteristic annotation that we are mostly
interested in within the feature-based opinion mining task.
Starting with this observation, we decided to develop and apply an annotation
scheme that would be able to capture all the aspects involved in the opinion mining
process. Further on, we could use this annotated information for three experiments
fact versus opinion sentence classification, polarity classification for opinionated
sentences and automatic feature detection.
As in the case of MPQA, our annotation scheme is designed for the integration
within the GATE (General Architecture for Text Engineering)17 framework. The
annotation elements and their attributes are described using XML Schema files.
We describe four types of annotations: fact, opinion, feature expression and
modifier.
The first element fact (Figure 4.4)- was created for the labeling of tokens and
phrases containing factual information. The attributes we defined for this element
are source (writer, a quote etc.), target (name of product capability it refers to), id
(an identifier given by the annotator for future reference), feature (name of the
feature that the fact describes), type (direct, indirect or implicit), POS (part of
17

http://gate.ac.uk

67

speech, for factual information expressed in individual tokens, which can be


adjective, noun, verb, adverb or preposition), phrase (if the factual information is
expressed in more than one token, a sentence or a group of sentences).
The second element opinion is presented in Figure 4.3. Apart from the
attributes that are common to the fact element, this label contains the attributes
polarity (positive or negative), intensity (degree of the opinion expressed, which
can be low, medium or high) and affect (as results from the formulation of the
opinion statement).
Another defined element feature expression, with the same elements as the
opinion tag, aims at identifying phrases that indirectly express a feature of the
product under review (Figure 4.5).
Finally, the modifier element, whose structure is the same as for the opinion
element, is included to spot the words whose use lead to a change in the polarity of
the expressed opinion, but cannot be used alone to express an opinion (e.g. Its
absurd). An example of annotation is given in Figure 4.6.

Figure 4.3: Annotation scheme for the opinion element

68

Figure 4.4: Annotation scheme for the fact element

Figure 4.5: Annotation scheme for the feature expression element

69

Figure 4.6: Annotation scheme for the modifier element


REVIEW CORPUS ANNOTATION
With the created scheme, we annotated a corpus containing 100 reviews, each
containing approximately 1000 words (they range from 700 to 2000 words per
review), on a high variety of objects, pertaining to the categories of travel
(resorts, hotels, restaurants and touristic attractions), home & garden (furniture,
washing machines), electronics (laptops, PDAs, mobile phones, digital cameras),
computers and software, music, movies, books and cars. All these reviews were
taken from the epinions.com site. The advantage in using this source is that
reviewers tend to profoundly analyze each of the aspects of the products and the
reviews are long and complicated.
The idea behind the high variety of topics was, on the one hand, to detect similar
manners to express opinions and, on the other hand, to test our opinion mining
approaches at the coarse grained level, against the categories on which the users can
punctuate the object using stars and at the fine-grained level, using our
annotations. Moreover, choosing to label reviews on such a diversity of objects
allowed us to test the applicability of the schema at the level of the feature-based
opinion mining paradigm (i.e. identify, for any type of object, be it location, event,
product etc., its features and the opinion words that are used to describe it). The
70

annotation was done by two non-native speakers of English, one with a degree in
Computer Science and the second a Linguistics student.
An example of annotation is presented in Figure 4.7.
<modifier gate:gateId="46" source ="w" target="restaurant"
id="83" feature="" type="direct" polarity="positive"
intensity="medium" POS="" phrase="multiword"
affect="admiration">I was pleasantly surprised</modifier>
to learn that the food is <modifier gate:gateId="47"
source ="w" target="excellent" id="84" feature=""
type="direct" polarity="positive" intensity="medium"
POS="adverb" phrase="word"
affect="admiration">still</modifier> <opinion
gate:gateId="48" source ="w" target="food" id="85"
feature="length" type="direct" polarity="positive"
intensity="high" POS="" phrase="word"
affect="admiration">excellent</opinion>, and the staff
very <opinion gate:gateId="49" source ="w" target="staff"
id="86" feature="length" type="direct" polarity="positive"
intensity="high" POS="adjective" phrase="word"
affect="admiration">professional</opinion> and <opinion
gate:gateId="50" source ="w" target="staff" id="87"
feature="length" type="direct" polarity="positive"
intensity="high" POS="adjective" phrase="word"
affect="admiration"> gracious </opinion>.

Figure 4.7. Sample from the annotated corpus


As a result of the annotation process, we found some interesting phenomena that are
worth mentioning and that justify our annotation schema:
1. References along the argumentation line must be resolved (e.g. I went to
another restaurant from this chain and they treated us horribly. These
people were very nice.). This is the reason for which our scheme contains
the elements target and id. In this way, all factual or opinion
expressions referring to an entity can be traced along the text.
2. An idea is many times expressed along a line of sentences, combining
facts and opinions to make a point (e.g. This camera belongs to the new
generation. Quality has highly improved over the last generations.). We
did not have a problem with this phenomenon, since our annotation
scheme was designed to allow the labeling of token or expression-level
elements, as well as sentence and multi-sentence level.
3. Within a phrase, a reviewer presents facts in such a manner that the
reader is able to extract the corresponding opinion (e.g. They let us wait
71

for half an hour in the lobby and finally, when they gave us the key,
realized the room had not been cleaned yet.). While the sentence is
purely factual in nature, it contains the phrases let us wait for half an
hour in the lobby , the room had not been cleaned yet, which we
annotated as feature expressions of implicit type.
4. There is an extensive use of conditionals within reviews. (e.g. If you
dare, buy it! It was great for two weeks until it broke!; If youve got
shaky hands, this is the camera for you and if you dont, theres no need
to pay the extra $50 ) . We consider the sentence containing the
conditional expression a modifier.
5. There are many rhetoric-related means of expressing opinion (e.g. Are
you for real? This is called a movie?). We annotate these elements as
implicit feature expressions.
At the time of performing our experiments, these phenomena are important and
must be taken into consideration, since 27% of the annotated opinionated phrases in
our corpus are composed of more than one sentence. More generally, these findings
draw our attention upon the context in which opinion mining is done. Most of the
work so far concentrated on sentence or text level, so our findings draw the
attention upon the fact that more intermediate levels should also be considered.
EXPERIMENTS AND EVALUATION
The first experiment we performed aimed at verifying the quality and constancy of
the annotation as far as fact versus opinion phrases are concerned and, within the
opinionated sentences, the performance obtained when classifying among positive
and negative sentences. In the first phase, we lemmatize the annotated sentences
using TreeTagger18 and we represented each fact and opinion phrase as a vector of
characteristics, measuring the n-gram similarity (with n ranging from 1 to 4) and
overall similarity with each of the individual corpus annotated sentences, tokens
and phrases. We perform a ten-fold cross validation using the SVM SMO. The
results for fact versus opinion and positive versus negative classifications are
presented in Table 4.8.
Precision
Recall
Kappa
Fact
0.72
0.6
0.53
Opinion
0.68
0.79
0.53
Positive
0.799
0.53
0.65
Negative
0.72
0.769
0.65
Table 4.8. Evaluation of fact vs. opinion and positive vs. negative classification
18

http://www.ims.uni-stuttgart.de/projekte/corplex/TreeTagger/

72

In the second phase, we consider for classification only the phrases containing
two sentences. As in the first phase, we represent each fact and opinion phrase as a
vector of characteristics, measuring the n-gram similarity (with n ranging from 1 to
4) and overall similarity with each of the individual corpus annotated sentences and
with all other phrases containing two sentences, then perform a ten-fold cross
validation using SVM SMO. We summarize the results obtained in Table 4.9.
Precision
Recall
Kappa
Fact
0.88
0.76
0.43
Opinion
0.74
0.89
0.43
Positive
0.84
0.71
0.68
Negative
0.89
0.92
0.68
Table 4.9. Evaluation of fact versus opinion and positive versus negative 2sentences phrase classification
In the third phase, we consider for classification only the phrases containing
three sentences. The fact and opinion phrases are represented as in the first
experiments and a ten-fold cross validation using SVM SMO is done. The results
are shown in Table 4.10.
Precision
Recall
Kappa
Fact
0.76
0.6
0.80
Opinion
0.78
0.94
0.80
Positive
0.85
0.76
0.68
Negative
0.92
0.96
0.68
Table 4.10. Evaluation of fact versus opinion and positive versus negative 3sentences phrase classification
From the results obtained, we can notice that using longer phrases, we obtain an
improved classification performance in both fact versus opinion classification, as
well as positive versus negative classification. The only drop in classification
performance is in the case of longer factual phrases. We explain this by the fact that
in many of the cases, these types of phrases contain descriptions of opinionated
sentences or represent combinations of factual and opinion sentences (e.g. They
said it would be great. They gave their word that it would be the best investment
ever made. It seems they were wrong). The results show that our annotation is
constant and that labeled elements present similarity among them; this fact can be
used to automate the annotation, as well as use the labeled corpus for the training of
an opinion mining system or for its evaluation. Evaluation proved that the
annotation schema and approach are general enough to be employed for labeling of
reviews on any product.
73

APPLYING TEXTUAL ENTAILMENT FOR OPINION


CLASSIFICATION
Subsequently to annotating the review corpus, we study the manner in which
feature-based opinion mining can be done at a coarse level (Balahur and Montoyo,
2009), i.e., not discovering all the features of a product and the corresponding
opinion words that are used to describe them, but the opinion on the main aspects of
the object on which opinion is mined, which are also evaluated by the users with
stars (from 1 to 5 stars, 1 being the lowest and 5 being the highest). Comparing
against the number of stars given to these main product functionalities can also be a
useful method to evaluate opinion mining systems.
The approach taken in this section is motivated by the fact that many products
on review sites have a star assigning system associated. In this manner, people
are given the opportunity to value the product in question, besides using
opinionated phrases, by employing, for a number of default defined features, from
one to five stars. We consider this to be useful information at the time of review
mining, both for the fact that we can thus overcome the problem of automatically
discovering the distinct product features, as well as the problem of evaluating our
approach without having a previously annotated corpus on the specific product.
Textual entailment recognition (Dagan et al., 2006) is the task of deciding, given
two text snippets, one entitled Text (T) and the other called Hypothesis (H), if H
can be inferred from T.
For this task, we will use the information given by the stars category. We
consider the given criteria and generate simple sentences (hypotheses) for each of
the positive and negative categories. The idea behind this proposed method is to test
whether a textual entailment system would be able to capture and better resolve the
semantic variability given by opinionated text.
Our textual entailment system (Iftene and Balahur-Dobrescu, 2007) is based on
the tree edit distance algorithm. Each of the sentences is dependency parsed using
Minipar19 and passed through Lingpipe20 in order to detect and classify the Named
Entities it contains. Subsequently, syntactic, lexical and semantic similarities are
computed, starting from the root of the dependency tree, between all the tokens and
their corresponding dependency links in the hypothesis and the tokens in the text
using the DIRT 21 collection, eXtended WordNet 22 and a set of rules for context
modifiers.

19

http://www.cs.ualberta.ca/~lindek/minipar.htm
http://alias-i.com/lingpipe/
21
http://demo.patrickpantel.com/Content/LexSem/paraphrase.htm
22
http://xwn.hlt.utdallas.edu/
20

74

From the annotated corpus, for each of the considered products, we selected the
reviews containing sentences describing opinions on the criteria which users are
also allowed to assess using the stars system. The categories which are punctuated
with two or less stars are considered as negative and those punctuated with four or
five stars are considered as having been viewed positively.
We generated hypotheses under the form Category is good and Category is
nice., Category is not good and Category is not nice.,e.g. The price was
good.,The price was not good. , The food was good., The food was not
good. The view was nice., The view was not nice.. In case no entailment was
found with such built sentences, we computed entailment with annotated sentences
in the review corpus. The results obtained are shown in Table 4.11.
Name of product

Stars Category
Accuracy
Price
62%
Service
58%
Restaurant
Food
63%
View
53%
Atmosphere
58%
Ease of use
55%
Durability
60%
Digital camera
Battery life
80%
Photo quality
65%
Shutter lag
53%
Ease of Use
60%
Durability
72%
Washing machine
Ease of Cleaning
67%
Style
65%
Table 4.11. Polarity classification accuracy against the number of stars per
category
As we can see from the obtained results, textual entailment can be useful at the
time of performing category based opinion mining. However, much remains to be
done at the level of computing semantic similarity between opinionated texts. Such
work may include the discovery of opinion paraphrases or opinion equivalence
classes.

75

4.1.6. BUILDING A RECOMMENDER SYSTEM USING FEATUREBASED OPINION MINING


Having seen the manner in which opinion mining can be done at the level of
features, we envisage a straightforward method to recommend products based on
their fine-grained characteristics, the assessments made on these features and the
preferences a user can express. Thus, in order to recommend a product for purchase,
we present a method to compute the similarity between a product (whose features
are summarized using the feature-based opinion mining and summarization system)
and what is seen as the perfect product in the category.
For each product category, we consider a list containing the general and productspecific features. The perfect product within that category can thus be represented
as a vector whose indices correspond to the list of features and whose values are all
1, signifying that all features are 100% positive. At this point, it is interesting to
note that the semantics of positive and negative for the product category are
given by the feature attributes we extracted for each of the categories (positive for
size thus includes small, tiny, pocket-fit etc.).
In order to find recommendable products, we use the customer review
summarization system presented for each product model and its corresponding
collection of reviews. In this manner, we build vectors corresponding to the product
models, whose indices will be the same as the ones of the perfect product, and
whose corresponding values will be the percentage in which the feature is classified
as positive by the summarization system. Finally, we compute the similarity
between the each of the obtained vectors and the vector corresponding to the
perfect product using the cosine similarity measure. We recommend the top 5
matching products. In order to better understand the process of recommendation,
we will consider an example and suppose the user would like to buy a 4-Megapixel
camera. There are around 250 available models on the market and for each model
one can read an average of 10 customer reviews. Instead of having to read 2500
reviews, employing the presented system, being given the 5 best products, the user
will only have to browse through 50 reviews, in case (s)he is not confident in the
system classification; when the user is confident, (s)he has to read none.
The list of features for a 4 Megapixel camera is: (price, warranty, size, design,
appearance, weight, quality, lens, viewfinder, optical zoom, digital zoom, focus,
image resolution, video resolution, memory, flash, battery, battery life, LCD size,
LCD resolution, accessories).
The vector associated to the perfect 4-Megapixel camera will have as indices
the features in the above list and all corresponding values 1: vperf(price)=1; v
perf(warranty) =1 and so on, in the order given by the list of features. After applying
the customer review summarization system on other 4-Megapixel cameras, we
76

obtain among others the vectors v1 and v2, corresponding to Camera1 4MP and
Camera2 4MP. In this case:
v1=(0.7,0.5,0.6,0.2,0.3,0.6,0.5, 0.5,0.7,0.8,0.7,0.8,0.4,0.3,0.3,0.7,0.6,0.3,0.8,0.4,0.4)
v2 = (0.8,1, 0.7,0.2,0.2,0.5, 0.4,0.4,0.8,0.8,0.8,0.8,0.7,0.7,0.3,0.8,0.6,0.7,0.5,0.3,0.6)
Calculating the cosine similarity between v1 and vperf and v2 and vperf,
respectively, we obtain 0.945 and 0.937. Therefore, we conclude that Camera1 4MP
is better than Camera2 4MP, because it is more similar to the perfect 4-Megapixel
camera model.

4.2. OPINION MINING FROM NEWSPAPER ARTICLES


4.2.1. INTRODUCTION
Subsequently to the experiments in feature-based opinion mining, our aim was to
apply sentiment analysis to newspaper articles. As mentioned before, the task can
be formulated in the context of any textual type. Nevertheless, given the
peculiarities of the genre and the final aim of the sentiment analysis task, the
requirements of a system that automatically processes the intended kind of text to
extract opinions are different. Anticipating the following sections, this observation
was confirmed in the experiments we performed with news data, blogs, political
debates and microtext.
In a first approach, we started researching on appropriate methods to classify
sentiment expressed in news, with the aim of including an opinion mining
component to the Europe Media Monitor23 family of applications (Balahur et al.,
2009f). Such data is very different from product reviews in that sentiment is usually
expressed much less explicitly. Bias or sentiment can be expressed by mentioning
some facts while omitting others, or it can be presented through subtle methods like
sarcasm (e.g.Google is good for Google, but terrible for content providers).
Another major difference between the news and product reviews is that the target of
the sentiment is much less concrete. A camera has very well-defined and easily
identifiable features like weight, flash light, battery life, etc., but what are the
features of a named entity such as a specific person or of an organization like the
European Commission (EC)? As opposed to the features of products, which are
parts or characteristics that are easily linkable through technical details or frequent
mentioning, the features of more general topics, such as persons, events or
organizations can be considered as sub-topics (e.g. administration, policy areas,
23

http://emm.newsbrief.eu/overview.html

77

development of poorer regions, consumer protection, environment issues).


Newspaper articles are much more complex than reviews on products, as they
contain in their majority part factual descriptions of events and their participants. It
is thus tricky to detect whether any negative sentiment detected refers to a specific
organization we are interested in, to the main news, or to any other entity or topic
mentioned in the context. Furthermore, the context in itself may have a specific
sentiment associated to it (i.e. the news can be good or bad, independent on the
sentiment that is expressed on the entities or the topic it is related to). Therefore, a
clear distinction must be done in order to separate the content in which the entity
we are interested in was mentioned (e.g. a natural disaster), and the sentiment
expressed towards this entity in the given context (e.g. if the entity was helpful and
offered support, or, on the contrary, if it ignored the gravity of the situation and did
nothing). Observing these characteristics of newspaper articles, we can state that
different tasks can be performed in their context and each must be treated in a
different manner, using specific methods and resources. For example, classifying
the context as good or bad news is different from classifying the opinion expressed
on a specific entity participating in the news or classifying the sentiment expressed
on the general context. Additionally, all these tasks are distinct from the one
involving the detection of author or source bias.
Given the complexity of the phenomena identified in the preliminary analysis,
our first aim when trying to perform opinion mining from newspaper articles was to
delimit the scope of the analysis and define the task we are going to tackle.
Our first aim was to detect sentiment in quotations (direct reported speech). The
reason for this is that the text in quotes is usually more subjective than the other parts
of news articles, where sentiment is either expressed less, or it is expressed less
explicitly. We also know for quotes who the person is that made the statement
(referred to as the source of the opinion statement) and if the speaker makes
reference to another entity within the quotation we have a clue about the possible
target (or object) of the sentiment statement. (e.g. Steinmeier said: I think we can
conclude that there is a fresh wind in NATO, and also, hopefully, a new atmosphere
of cooperation,). In the first experiments, we will thus propose different methods to
classify quotations depending on whether or not they are subjective and to determine
the polarity of the sentiment expressed in the identified subjective quotations. Unlike
full articles, quotations are relatively short and often are about one subject. However,
they contain a variety of interesting phenomena, such as the combination of a short,
factual summary of the event or what the target did or a general view on the
problem, as well as the opinion or position of the source on this fact description
(e.g. It is a tough battle and those who perceive us as competitors are not going to
roll over and play dead. But again, both Branson and Fernandes are battle scarred
and, with a song and a prayer, and lots of hard work, I believe we shall prevail).
78

Another interesting aspect concerns the presence of various possible targets in the
quote, on which antonymic opinions are expressed (e.g. How can they have a case
against Fred, when he didnt sign anything?). Moreover, they contain a larger scale
of affective phenomena, which are not easily classifiable when using only the
categories of positive and negative: warning (e.g. Delivering high quality education
cannot be left to chance!), doubt (e.g. We dont know what we should do at this
point), concern, confidence, justice etc. (where doubt is generally perceived as a
negative sentiment and confidence as a positive one).
The aim we have is to determine the attitude polarity (tonality of the speech),
independent of the type of news, interpreting only the content of the text and not the
effect it has on the reader.

4.2.2. INITIAL EXPERIMENTS


For our first experiments, we chose a set of 99 quotes, on which agreement between
a minimum of two annotators could be reached regarding their classification in the
positive and negative categories, as well as their being neutral/controversial or
improperly extracted. The result of the grouping was a total of 35 positive, 33
negative, 27 neutral/controversial and 4 improperly extracted quotes. We used this
dataset to comparatively analyze the different possible methods and resources for
opinion mining and we explored the possibility to combine them in order to
increase the accuracy of the classification.
The first approach is based on a bag of words the use of different lexicons
containing positive and negative words. The second approach contemplates
measuring similarity to annotations extracted from existing corpora and machine
learning.
A. BAG-OF-WORDS APPROACH
At the present moment, there are different lexicons for affect detection and opinion
mining. The aim in the following evaluation is to test the different resources in the
quote classification scenario and assess the quality and consistency of these
lexicons.
Each of the employed resources were mapped to four categories, which were
given different scores positive (1), negative (-1), high positive (4) and high
negative (-4). The assignment of these values was based on the intuition that certain
words carried a higher affective charge and their presence should be scored
accordingly. Our intuition was supported by experiments in which we used just the
positive and negative categories and that scored lower.

79

The polarity value of each of the quotes was computed as sum of the values of
the words identified; a positive score leads to the classification of the quote as
positive, whereas a final negative score leads to the system classifying the quote as
negative. The resources used were: the JRC lists of opinion words, WordNet Affect
(Strapparava and Valitutti, 2004), SentiWordNet (Esuli and Sebastiani, 2005),
MicroWNOp (Cerini et al., 2007). WordNet Affect categories of anger and disgust
were grouped under high negative, fear and sadness were considered negative, joy
was taken as containing positive words and surprise as highly positive;
SentiWordNet and MicroWNOp contained positive and negative scores between 0
and 1 and in their case, we mapped the positive scores lower than 0.5 to the positive
category, the scores higher than 0.5 to the high positive set, the negative scores
lower than 0.5 to the negative category and the ones higher than 0.5 to the high
negative set.
As a filtering step, we first classified the quotes based on the presence of
subjectivity indicators, using the Opinion Finder lexicon (Wilson et al., 2005).
The subjective versus objective filtering had an accuracy of 0.89, as 2 of the
positive and 5 of the negative quotes were classified as neutral.
We evaluated the approaches both on the whole set of positive and negative
quotes, as well as only the quotes that were classified as subjective by the
subjectivity indicators. Subsequently, we grouped together resources that tended to
over-classify quotes as positive or negative, in an attempt to balance among their
classification. Finally, we grouped together all the words pertaining to the different
classes of positive, negative, high positive and high negative words belonging to all
the evaluated resources. The results are shown in Table 4.12 (-S/O and +S/O
indicate absence and presence, respectively, of the subjectivity filtering):
Resource
JRCLists
SentiWN
WNAffect
MicroWN
SentiWN + WNAffect
All

-S/O
X

+S/O
X

X
X
X
X
X
X
X
X
X
X

Ppos
0.77
0.81
1
1
0
0
0.62
0.73
0.22
0.24
0.68
0.73

Pneg
0.3
0.35
0
0
1
1
0.36
0.35
0.66
0.67
0.64
0.71

Rpos
0.54
0.6
0.51
0.54
0
0
0.52
0.57
0.42
0.47
0.7
0.75

Rneg
0.55
0.625
0
0
0.51
0.54
0.48
0.53
0.45
0.41
0.62
0.69

Table 4.12: Results of the classification using the different opinion and affect
lexicons
80

B. SIMILARITY APPROACH
In this approach we used two existing resources the ISEAR corpus (Scherer and
Walbott, 1997) - consisting of phrases where people describe a situation when they
felt a certain emotion and EmotiBlog (Boldrini et al., 2009), a corpus of blog posts
annotated at different levels of granularity (words, phrases, sentences etc.)
according to the polarity of the sentiments and the emotion expressed.
In the first approach, we computed the individual quotes similarity with the
sentences belonging to each of the emotions in the ISEAR corpus, using Pedersens
Similarity Package24, based on the Lesk similarity25. Subsequently, we classified
each of the quotes based on the highest-scoring category of emotion. Table 4.13
presents the results:
Class
Joy Fear Anger Shame Disgust Guilt Sadness
Positive
8
7
1
3
3
5
8
Negative 6
7
1
5
8
2
4
Table 4.13: Results of the classification using the similarity scores with the ISEAR
corpus
We consider as positive the examples which fell into the joy category and
classify as negative the quotes which were labeled otherwise. The results are
presented in Table 4.14:
Ppos Pneg Rpos Rneg Accuracy
0.22 0.82 0.58 0.5
0.514
Table 4.14: Results of the positive versus negative classification using the similarity
score with the ISEAR corpus
EmotiBlog represents an annotation schema for opinion in blogs and the
annotated corpus of blog posts that resulted when applying the schema. The results
of the labeling were used to create a training model for an SVM classifier that will
subsequently be used for the classification of opinion sentences. The features
considered are the number of n-grams (n ranging from 1 to 4) and similarity scores
with positive and negative annotated phrases, computed with Pedersens Similarity
Package. The approach was previously described by Balahur et al. (Balahur et al.,
2009b). The evaluation results are presented in Table 4.15:
24

http://www.d.umn.edu/~tpederse/text-similarity.html
http://kobesearch.cpan.org/htdocs/WordNet-similarity/WordNet/ Similarity/lesk.htm

23

81

Class
Precision Recall F-measure
Positive
0.667
0.219
0.33
Negative
0.533
0.89
0.667
Table 4.15: Results of the classification using SVM on the EmotiBlog corpus model
From the results obtained, we can infer that the use of some of the resources
leads to better performance when classifying positive or negative quotes
(SentiWordNet versus WordNet Affect), and that the combined resources produce
the best results when a vocabulary-based approach is used. Another conclusion is
that previous subjectivity filtering indeed improves the results.

4.2.3. PRELIMINARY CONCLUSIONS FROM INITIAL EXPERIMENTS


From the results in Table 4.15, we can conclude that annotations about a specific
topic cannot be applied to generic opinion mining on news. This confirms that
open-domain opinion analysis is a more difficult problem than topic-specific
sentiment classification and other sub-tasks defined in opinion mining, such as
feature-based opinion mining. Experiments showed that simple bag-of-words
approaches cannot reach a satisfactory level, even when large sets of words are
employed. Most importantly, these preliminary experiments, even if they were
performed on a very small dataset, have succeeded in shedding some light on the
sentiment analysis related phenomena in newspaper articles and the challenges
that are associated to them.
Following these findings, we also realized that the task of sentiment analysis
needs to be redefined in the context of news. This was done in a further effort,
described by Balahur and Steinberger (2009). Experiments using the new definition
provided have shown that indeed, when the task is clarified, both the annotation of
newspaper article texts according to their sentiment has a better agreement, as well
as the performance of the automatic processing increases.

4.2.4. REDEFINING SENTIMENT ANALYSIS FROM NEWSPAPER


ARTICLES
Following this first set of experiments by Balahur and Steinberger (2009), we set
the objective of annotating a larger corpus of quotations extracted from newspaper
articles. We extracted a set of 1592 quotations in English which we set out to
annotate with sentiment. The task was to decide whether a quotation was positive,
negative or neutral.

82

ISSUES REGARDING SENTIMENT ANNOTATION IN NEWSPAPER


ARTICLES
After a first effort to annotate 400 quotations by 2 different persons, we realized
that the inter-annotator agreement even for the short pieces of text was relatively
low (below 50%). The following is an example of a quotation where annotators had
difficulty to agree:
(1) Politician A said: We have declared a war on drugs.
While one annotator may feel that this is a positive statement as it shows action
and efficiency to overcome an existing drug-related problem, another annotator
may interpret this statement as being negative because (a) war is negative and
(b) the situation must be rather bad if an extreme reaction such as declaring war is
necessary. Depending on the background of the annotator, s/he may even want to
argue in a certain context (c) that drugs refers to soft drugs and that these are not a
problem at all for society, in which case politician As attitude would probably be
considered as being misled and erroneous. While the source of the quotation
(politician A) is clear, there is thus some confusion regarding what the target is. Is it
the energetic and problem-solving (or erroneous) attitude of the politician (positive
attitude), or is it the alleged fact that there is a drug problem (negative news). A
further issue is the question whether the confidence expressed in the politicians
statement should be considered as being positive, i.e. whether a statement such as
we will do something about this problem should be considered a sentiment
statement at all. It is clear that such a statement is intended to cause positive
feelings towards the politician. Some existing sentiment or affect vocabulary lists
do also include words like war and mother with the respective values negative or
positive. By adding one level to this example:
(2) Person B mocked politician As statement that We have declared a war on
drugs.
Yet another interpretation is possible, namely that Person B is the source and
politician A is the target, questioning also somehow the positive sentiment
politician A wanted to create with his/her statement. The journalist writing the
article may express his or her own opinion on this, as in:
(3) Person B unreasonably mocked politician As statement that We have
declared a war on drugs.
In this case, the journalist expresses negative sentiment towards Person B for
criticizing politician A.

83

The chain could theoretically be continued: For instance, if the newspaper were
a known defender of person A (and the corresponding) political party and attitudes,
the whole statement (3) could be interpreted as sarcasm, inverting the negative
sentiment of Person B towards Politician A, and so on. While this is clearly a
constructed example, our low inter-annotator agreement and the clarifying
discussions showed that our initial sentiment annotation instructions were underspecified and left too much leeway for interpretation.
(4) Time called on the War on Drugs?
These are real examples of texts, where we can further notice difficulties. In this
case, the journalist mocks the idea of delaying taking an action against drugs. Or the
following example:
(5) Argentina and Mexico have taken significant steps towards decriminalising
drugs amid a growing Latin American backlash against the US-sponsored war on
drugs.
In this context, US-sponsored is the key expression towards understanding the
negative opinion on the war on drugs.
For these reasons, we re-defined our task and subsequently annotated the whole
set of 1592 quotations, after which the inter-annotator agreement was 0.81%.

PREVIOUS DEFINITIONS
In order to redefine the task, we first start by looking into the definitions that were
given until this point.
Subjectivity analysis is defined by (Wiebe, 1994) as the linguistic expression of
somebodys opinions, sentiments, emotions, evaluations, beliefs and speculations.
In her definition, the author was inspired by the work of the linguist Ann Banfield
(Banfield, 1982), who defines as subjective the sentences that take a characters
point of view (Uspensky, 1973) and that present private states (Quirk, 1985) (that
are not open to objective observation or verification) of an experiencer, holding an
attitude, optionally towards an object. Subjectivity is opposed to objectivity, which
is the expression of facts. As Kim and Hovy (2004) notice, opinion is subjective,
but may not imply a sentiment. But what about our example of war on drugs?
Can facts express opinions? Is there a difference between interpretations of facts at
sentiment level and the direct expression of sentiments? Should we take them into
consideration? Therefore, in our context, this definition did not help.
Esuli and Sebastiani (2006) define opinion mining as a recent discipline at the
crossroads of information retrieval and computational linguistics which is
84

concerned not with the topic a document is about, but with the opinion it
expresses. This is a very broad definition, which targets opinions expressed at a
document level. As we have shown before, news articles contain mentions of
different persons and events, the topic in itself might involve a negative tonality and
both the author of the text, as well as the facts presented or the interpretation they
are given by the reader may lead to a different categorization of the document. So,
this definition is not specific enough for us to understand what we should be
looking for when annotating pieces of newspaper articles.
Dave et al. (2003) define an opinion mining system as one that is able to
process a set of search results for a given item, generating a list of product
attributes (quality, features, etc.) and aggregating opinions about each of them
(poor, mixed, good). Opinion mining, in this context, aims therefore at extracting
and analyzing judgments on various aspects of given products.
A similar paradigm is given by Hu and Liu (2004), which the authors entitle
feature-based opinion mining. It is, however, not clear how statements such as It
broke in two days, The night photos are blurry, that are actual fact information
(according to the definition of subjectivity, they are verifiable) could be and should
be annotated. Do they fall outside the goal of opinion mining? Since in our context,
persons, organizations or events have no definable or inferable lists of features, this
definition of the task does not work for us either.
Kim and Hovy (2005) define opinion as a quadruple (Topic, Holder, Claim,
Sentiment), in which the Holder believes a Claim about the Topic, and in many
cases associates a Sentiment, such as good or bad, with the belief. The authors
distinguish among opinions with sentiment and opinions without sentiment and
between directly and indirectly expressed opinions with sentiment. In this context,
it does not remain clear how an example such as the Local authorities have
provided no help for the victims of the accident. should be interpreted and why.
Some might even argue that a statement they claim to be opinionated but with no
sentiment Gap is likely to go bankrupt (which would probably be interesting
when assessing favorability in markets), has a sentiment and that sentiment is
negative.
In the SemEval 2007 No. 14 Affective Text Task (Mihalcea and Strapparava,
2007), the systems were supposed to classify 1000 newpaper titles according to
their valence and emotion contained. A title such as Scientists proved that mens
perspiration raises womens hormone levels or 100 killed in bomb attack were
classified as negative. However, this is factual, verifiable information. Does this
mean that when capturing the media sentiment, we should consider these results as
being negative? Do these statements refer to a fact and are we interested in the
information conveyed or in the sought effect? If so, which of these aspects would
we include in a system doing sentiment analysis from newspaper articles?
85

In other approaches, capturing favorability versus unfavorability, support versus


opposition, criticism versus appreciation, liking versus disliking, even bad versus
good news classification were considered sentiment analysis.
However, at the moment of annotating sentiment in newspaper articles, we have
seen that combining all these aspects together did not help to clear what the task
was and how annotation should be done. Even in the case of quotes, which are short
pieces of text where the source was known and the possible targets were identified,
expressions of opinion that needed some kind of interpretation or knowledge of the
situation fell short of agreement, due to personal convictions, background and so
on.
REDEFINITION OF GUIDELINES FOR SENTIMENT ANNOTATION IN
NEWSPAPER ARTICLES
We further on then present an annotation effort for newspaper quotes that shed light
on the issue and helped define guidelines for labelling that led, from level of
agreement of under 50%, to a level of agreement of 81%. We give some details on
the gold-standard quotation collection we created according to these guidelines.
Finally, we redefine the task of sentiment analysis in the news, capturing the
different aspects of sentiment in text that we identified and pinpointing what exactly
we expect a sentiment analysis system to discover from news under the different
aspects.
REDEFINING THE TASK OF SENTIMENT ANNOTATION
Although some definitions of the task were proposed, none of them, as we have
seen, could give an indication of the specific aspects that a system implementing
opinion mining in a news context should contemplate. To clarify the task, we
selected a collection of 1592 quotes (reported speech) from newspaper articles in
English, whose source and target were known (their extraction patterns are designed
in that scope) which we set out to annotate. Details on the length of the quotes are
given in Figure 4.8.

86

Figure 4.8: Histogram of the quotes length


The first experiments had an agreement lower than 50%. Specifying that just the
sentiment on the target should be annotated and separated from the good and bad
news that was described led to an increase in the agreement up to 60%. We realized
that by delimiting a few aspects, the task became much clearer. Following are the
annotation guidelines we used:
We are trying to decide whether the entity (person or organization) in the text
snippet is being talked about in a positive (POS) or in a negative light (NEG), or if
the statement is rather objective/neutral (OBJ). We thus distinguish three cases of
sentiment: two cases of subjectivity, in which case we can directly indicate the
polarity (POS, NEG), and the case of non-subjectivity, objectivity or neutrality
(OBJ). OBJ is the default, so no need to label neutral/objective examples.
Here are some more clarifications that may help:
1. If you can, try not to make use of your world knowledge, such as the
political views of the entity. If you cannot decide without knowing the
political views of the entity, just leave it neutral/objective (OBJ).
2. It may help to imagine that you are the one being talked about: would
you like or dislike the statement without using your world knowledge?
3. We are not interested in knowing whether the whole piece of text is
positive or negative, but exclusively the sentiment expressed towards the
entity.
4. Another hint to identify subjectivity those news items whose content
cannot be verified and whose content is expressly changed to induce a
negative/positive opinion should be annotated as positive or negative.
E.g. Negotiations with Turkey have been delayed factual (objective
from a sentiment point of view) vs. EU stalls negotiations with Turkey
(subjective, negative sentiment).
5. Note that, in the text snippet X supported Y for criticizing Z, there is
87

negative sentiment towards Z from X and Y, but positive sentiment


between X and Y.
6. Please try to separate good news vs. bad news from the sentiment
expressed. We should NOT annotate the good versus bad content of the
news. E.g. if the news talks about 7000 people being killed by a bomb,
the news is factual/objective (OBJ), even if there is negative news
content.
7. Expressions of attitude: EU is willing to make efforts to prevent this
from becoming a crisis (This shows positive attitude, i.e. positive
sentiment, POS); On the other hand, the sentence EU sent help to the
earthquake-affected Aquila citizens is objective from a sentiment point
of view (OBJ).
8. Should there be both positive and negative statements in the snippet,
please consider the statement to be objective (OBJ). (strictly speaking, it
would be subjective, but balanced; but we are not trying to distinguish
this case).
9. It is certain that there will be many cases of doubt. In case of doubt, just
leave the example un-annotated (neutral/objective, OBJ).
The original data set we decided to annotate contained 1592 quotes extracted
from news in April 2008. The average final agreement was 81%, between 3 pairs
of two annotators each.

Agreement

Number
of
quotes

Number
of agreed
quotes

Number
of agreed
negative
quotes

Number
of agreed
positive
quotes

Number
of agreed
objective
quotes

1592

1292

234

193

865

81%

78%

78%

83%

Table 4.16: Results of the annotations in terms of agreement per class of sentiment
The result of the annotation guidelines and labeling process what a corpus in
which we agreed what sentiment was and was not in our case. The number of
agreed sentiment-containing quotes was one third of the total number of agreed
quotes, showing that only clear, expressly stated opinion, which required no
subjective interpretation from the annotators part was done.
The result of our labeling showed that in the case of newspapers, it is mandatory
to distinguish between three different components: the author, the reader and the
text itself (Figure 4.9).

88

AUTHOR

READER1

INTERPRETATION 1

READER 2

FACTS
ATTITUDE
SENTIMENT

READER N

TEXT

INTERPRETATION 2

INTERPRETATION N

Figure 4.9: A 3-component view on sentiment expression author, text, reader


While the author might convey certain opinions, by omitting or stressing upon
some aspect of the text and by inserting their own opinion towards the facts, the
spotting of such phenomena is outside the aim of sentiment analysis as we have
defined it (and done by perspective determination, or news bias research). From the
readers point of view, the interpretations of the text can be multiple and they
depend on the personal background knowledge, culture, social class, religion etc. as
far as what is normal (expected) and what is not are concerned. Lastly, the opinion
stated strictly in the text is the one that one should concentrate on at this level, being
expressed directly or indirectly, by the target, towards the source, with all the
information needed to draw this conclusion on polarity present in the text.
From the author and the readers perspective and not from the texts pure
informational point of view, opinion is conveyed through facts that are interpretable
by the emotion they convey. However, emotions are not universal in their
signification. They are determined socially, culturally and historically. There are
general emotions, but most of the times they relate to the norms, their significance
and the cultural environment. Emotions imply an evaluation, which is both
cognitive and affective, of a behavior, with respect to a norm and the mutual
expectation it raises. Some norms are common sensical and overall accepted and
understood. Normative expectations link the behavior (reaction) to a meaning and
on this ground, by the significance it is given. From the readers point of view,
sentiment analysis would be defined as the assessment of a target, based on its
characteristics and factual information related to it, according to whether or not
the results of the assessments are according to or against the norm.
From the authors point of view, news bias or perspective determination should
be concerned with discovering the ways in which expression of facts, word choice,
89

omissions, debate limitations, story framing, selection and use of sources of quotes
and the quote boundaries, for example, conveys a certain sentiment or not. The
sentiment content of the text, finally, is what is expressly stated, and not what is left
to be understood between the lines. Although pragmatics, through the speech- act or
other theories would argue there is no text that has no intended meaning, the
sentiment or factual information conveyed is different from reader to reader and can
thus not be done at a general level, as sentiment analysis intends to. For example,
the text The results of the match between Juventus Torino and Real Madrid last
night are 3-0. would maybe be interpreted as something positive, a motive for
pride in an Italian newspaper, it would be a negative, sad thing if reported by a
Spanish source, it would be bad or good depending on whether or not an interested
reader were pro or against the two teams and it would constitute just factual news
from the strict point of view of the text. Given these three views one must be aware
of at the time of constructing a sentiment analysis system for news, we can see that
the task becomes much clearer and the agreement at the time of annotating texts,
implementing and evaluating systems is higher.
Should one want to discover the possible interpretations of texts, sources and
readers profiles must be defined and taken into consideration, for a whole
understanding of the possible sentiment effects text has or is intended to have,
and not just a general, often misunderstood one.
At this moment, having the tasks clearly defined, we have started experimenting
with adequate methods to perform sentiment analysis considering these insights.

4.2.5. EXPERIMENTS APPLYING THE NEW TASK DEFINITION OF


SENTIMENT ANALYSIS FROM NEWSPAPER ARTICLES
EXPERIMENTAL SETUP
In order to measure the impact of our defined task, we performed different
experiments on the set of 1292 quotes on which agreement has been reached
(Balahur et al, 2010d). Out of these 1292 quotations, the target was successfully
identified by the sentiment analysis system in 1114 quotes (direct mentions of the
target through the name or its title). The baseline we compare against is the
percentage of quotes pertaining to the largest class of quotes objective, which
represents 61% of our corpus.
According to the approach we settled on, we wanted to make sure that: a) we
estimate the opinion on the target of the quote (by computing the opinion in
windows of words between the mentions of the entity), b) we eliminate the bad
versus good news content (by eliminating those words which are both sentimentbearing words and words that are part of EMM category definitions, from now on
90

called category words). Given that we are faced with the task of classifying opinion
in a general context, we employed a simple, yet efficient approach, presented in
(Balahur et al., 2009f).
At the present moment, there are different lexicons for affect detection and
opinion mining. In order to have a more extensive database of affect-related terms,
in the following experiments we used WordNet Affect (Strapparava and Valitutti,
2004), SentiWordNet (Esuli and Sebastiani, 2006), MicroWNOp (Cerini et al,
2007). Additionally, we used an in-house built resource of opinion words with
associated polarity, which we denote by JRC Tonality. Each of the employed
resources was mapped to four categories, which were given different scores:
positive (1), negative (-1), high positive (4) and high negative (-4). The score of
each of the quotes was computed as sum of the values of the words identified
around the mentions of the entity that was the target of the quote, either directly
(using the name), or by its title (e.g. Gordon Brown can be referred to as Gordon,
as Brown or as the British prime-minister)26. The experiments conducted used
different windows around the mentions of the target, by computing a score of the
opinion words identified and eliminating the words that were at the same time
opinion words and category words (e.g. crisis, disaster).
EVALUATION RESULTS
Table 4.17 presents an overview of the results obtained using different window
sizes and eliminating or not the category words in terms of accuracy (number of
quotes that the system correctly classified as positive, negative or neutral, divided
by the total number of quotes).
As it can be seen, the different lexicons available performed dramatically
different and the impact of eliminating the alert words was significant for some
resources or none for others, i.e. in those cases where there were no category words
that coincided with words in the respective lexicon.

26

For the full details on how the names and corresponding titles are obtained, please see (Pouliquen
and Steinberger, 2009).

91

Word
window
Whole text
3
6
6
10

W or W/O
Alerts
W Alerts
W/O Alerts
W Alerts
W/O Alerts
W Alerts
W/O Alerts
W Alerts
W/O Alerts
W Alerts
W/O Alerts

JRC
MicroWN WNAffect SentiWN
Tonality
0.47
0.54
0.21
0.25
0.44
0.53
0.2
0.2
0.51
0.53
0.24
0.25
0.5
0.5
0.23
0.23
0.63
0.65
0.2
0.23
0.58
0.6
0.18
0.15
0.82
0.2
0.23
0.79
0.18
0.15
0.61
0.64
0.22
0.2
0.56
0.64
0.15
0.11

Table 4.17: Accuracy obtained using different lexicons, window sizes and alerts
As we can see, computing sentiment around the mentions of the entity in smaller
window sizes performs better than computing the overall sentiment of texts where
the entities are mentioned. From our experiments, we could notice that some
resources have a tendency to over-classify quotes as negative (WordNet Affect) and
some have the tendency to over-classify quotes as positive (SentiWordNet). We
have performed evaluations using combinations of these four lexicons. The best
result we obtained were using the combination of JRC Tonality and MicroWN, on a
window of 6 words; in this case, the accuracy we obtained was 82%. As we can see,
the majority of the resources used did not pass the baseline (61%), which shows
that large lexicons do not necessarily mean an increase in the performance of
systems using them.
ERROR ANALYSIS
Subsequently to the evaluation, we have performed an analysis of the cases where
the system fails to correctly classify the sentiment of the phrase or incorrectly
classifies it as neutral.
The largest percentage of failures is represented by quotes which are erroneously
classified as neutral, because no sentiment words are present to account for the
opinion in an explicit manner (e.g. We have given X enough time, He was the
one behind all these atomic policies, These revelations provide, at the very least,
evidence that X has been doing favours for friends, We have video evidence that
activists of the X are giving out food products to voters) or the use of idiomatic
expressions to express sentiment (e.g. They have stirred the hornets nest).
Errors in misclassifying sentences as positive instead of negative or vice-versa
were given by the use of irony (e.g. X seemed to offer a lot of warm words, but
92

very few plans to fight the recession).


Finally, quotes were misclassified as positive or negative (when they should in
fact be neutral) because of the presence of a different opinion target in the context
(e.g. Ive had two excellent meetings with X, At the moment, Americans seem
willing to support Y in his effort to win the war, everyone who wants Y to fail is
an idiot, because it means were all in trouble, The chances of this strategy
announced by X are far better than the purely military strategy of the past...) or the
use of anaphoric references to the real target.
All these problems require the implementation of specific methods to tackle
them. Thus, firstly, the opinion lexicons should be extended to contain concepts
which implicitly imply an assessment of the target because they are concepts we
employ in our everyday lives (e.g. hunger, food, approval). Secondly, expressions
that are frequently used in a language to describe good and bad situations have
to be added to the opinion lexical (e.g. stir the hornets nest, take the bull by the
horns). Irony is difficult to detect in text; however, when dealing with a larger
context, the polarity of such pieces of text could be determined in relation to that of
the surrounding sentences. Further on, we are researching on methods to determine
the target of the opinion using Semantic Roles; thus, the judgement on the opinion
expressed can be improved. Finally, resolving co-reference using a standard tool
should in theory lead to a higher performance of the opinion mining system.
However, in practice, from our preliminary experiments, the performance of the
opinion mining system decreases when employing anaphora resolution tool.
These improvements are studied along the next chapters.

4.2. . CONCLUSIONS ON SENTIMENT ANALYSIS FROM


NEWSPAPER ARTICLES
As we have seen, mining sentiment from newspaper articles is a different task from
feature-based opinion mining and summarization, due to the complexities of the
textual genre. We have seen that there is a need to clearly define, before the
annotation is done, what the source and the target of the sentiment are, subsequently
separate the good and bad news content from the good and bad sentiment expressed
on the target and, finally, annotate only clearly marked opinion that is expressed
explicitly, not needing interpretation or the use of world knowledge. We have
furthermore seen that there are three different possible views on newspaper articles
author, reader and text and they have to be addressed differently at the time of
analysing sentiment. We have performed experiments in this direction, by using
categories to separate good and bad news content from the opinionated parts of the
text. We also evaluated our approach using different lexicons in diverse
combinations, and word windows.
93

We have shown that this simple approach produces good results when the task is
clearly defined. The data is available for public use at:
http://langtech.jrc.ec.europa.eu/Resources/2010_JRC_1590-Quotes-annotatedfor-sentiment.zip
Subsequently to these annotation efforts, we have continued to work on the
creation of corpora for sentiment analysis for other languages. We have created a
corpus of quotations extracted from newspaper articles in German containing 2387
quotes, based on the same annotation criteria. This resource is also publicly
available upon request to the authors.

4.3. OPINION MINING FROM POLITICAL DEBATES


4.3.1. INTRODUCTION
Until now, we have seen different methods to tackle opinion mining from reviews,
on the one hand, and quotations extracted from newspaper articles, on the other
hand.
The tasks involved, as well as the methodology used to detect and classify
sentiment within these two textual genres are very different. Reviews are written by
one author and usually have only one target the product in question and its
features. In newspaper articles, we have defined the task of sentiment analysis by
separating three different components the author, the text and the reader, and the
text into the context of the news and the sentiment expressed. The fact that our
experiments were done on already-extracted quotations, whose source and target
were priori given made the task more straight-forward, in the sense that the
mentions of the target entity could be traced within the text.
However, if we are to extend the frame of texts considered and try to detect the
sentiment on a general topic or event, from a text where multiple targets and
sources of sentiments are present, the methods we presented until now would not be
directly applicable (e.g. the sentiments expressed on a law that is submitted for
approval by the parliament, its benefits and drawbacks for the community it
targets).
As a result of the participation in the Opinion Pilot track at the Text Analysis
Conference (TAC) 2008, with the system presented by Balahur et al. (2008), which
will be detailed in Chapter 5 of this thesis, we have decided to investigate on the
resources and methods that are appropriate to treat opinion in any textual genre and
independently of the nature of the source or target of the opinions (i.e. whether they
are persons, events, products etc.). In the Opinion Pilot track, opinion questions
were asked on 25 different targets (persons, events, products, topics related to
94

persons etc.), whose answers were to be found in a set of blogs. As in this scenario,
a general opinion mining system must deal with many topics, ranging from
products to brands, companies, public figures, news topics, etc., which may not be
directly stated in the text as such. Therefore, when pursuing the goal of classifying
opinions, one must first of all have a base system that is able to detect negative and
positive opinion. To this aim, we propose a general opinion mining system (Balahur
et al., 2009e), which is able to track the sentiment expressed on different topics
mentioned in political debates. In order to further generalize the nature of the text
considered, we have chosen to test our initial methods for general opinion mining
on a corpus of political debates. Taking into consideration the corpus we have at
hand, we study the manner in which opinion can be classified along dialogues,
depending on the intervening speakers. We evaluate two methods to aggregate
opinion scores in order to make a unique classification of opinions provided by a
given speaker. While this type of classification was previously done by Thomas et
al. (2006), their approach was dependent on the previous training on the same kind
of data; our approach is data-independent.
Last, but not least, we study the possibility to determine the source of the
opinion expressed taking into consideration its polarity and the affect words used in
expressing the arguments. Since the corpus is already annotated with the party the
speaker belongs to, we perform this classification among the two parties
represented democrat and republican. While this type of classification was
previously approached by Mullen and Malouf (2006), the authors took into
consideration the general vocabulary used, and not the attitude towards the topic per
se and vocabulary related to it.

4.3.2. BACKGROUND
Although in the State-of-the-art chapter we have specified the directions of research
in sentiment analysis in general, we will briefly comment on work that is related to
this particular effort. Related work includes document-level sentiment analysis and
opinion classification in political texts.
Research in sentiment analysis at a document level, relevant research was done
by Turney et al. (2002), who first select important sentences based on pre-specified
part-of-speech patterns, then compute the semantic orientation of adjectives and
subsequently sum up this orientation to determine the document polarity.
Other related research was done by Pang et al. (2002), who employ machine
learning techniques to determine the overall sentiment in user reviews. Additionally,
Dave et al. (2003) propose classifying opinion on products based on individual
opinions on product parts. Gamon (2004) that studies the problems involved in
machine learning approaches and the role of linguistic analysis for sentiment
classification in customer feedback data. Matsumoto et al. (2005) research on the
95

impact of dependency analysis in sentiment analysis. Finally, Ng et al. (2006)


analyze the role of linguistic knowledge sources in opinion mining.
On the other hand, research in sentiment analysis from political texts included
classifying texts as conservative, liberal or libertarian (Mullen and Malaouf, 2006),
placing texts on an ideological scale (Laver et al., 2003; Martin and Vanberg,
2007). Other authors proposed methods to represent opposing viewpoints of two
parties in conflict (Lin et al., 2006). The corpus used in our research was first put
forward and employed by Thomas et al. (2006). In the research presented, the
authors investigate the possibility to determine support and opposition to the
proposed legislation from the floor debates. They use subjectivity indicators to train
a SVM classifier on part of the data and then employ a minimum-cut graph
algorithm to determine the orientation of the opinions. They perform individual
evaluations, first classifying individual speech segments and secondly classifying
opinion depending on the speakers (assuming that a speaker will maintain the same
opinion throughout the debate on a topic).
Our approach differs in many aspects to the ones taken in previous work. First,
we employ a general algorithm to classify the individual speech segments of the
different persons participating in the debates on each of the topics into positive and
negative. We base our classification on similarity measures between the speech
segments and the words pertaining to the categories of affect, opinion and attitude.
In the first phase, we perform the classification without taking into consideration
the target of the opinion expressed in the speech segments and without assuming
any opinion consistency with respect to the speakers. The second classification,
performed at speaker level, is done independently of the data. While we show the
manner in which machine learning can be employed on our method to improve the
results of the classifications, we discuss the implications this brings to the
generality of the system.

4.3.3. EXPERIMENTS
The corpus we use in our experiments is made up of congressional floor debates
and was compiled by Thomas et al. (2006). The corpus is available for download
and research27. It is split into three sets: the development set, the training set and
the test set. The first one contains 702 documents (one document corresponds to
one speech segment) pertaining to the discussion of 5 distinct debate topics, the
training set contains 5660 documents organized on 38 discussion topics and the test
set contains 1759 documents belonging to 10 debates. The corpus contains three
versions of these three sets, with the difference consisting in the removal of certain
clues relating to the topic and the speaker referred to in the speech. The speech27

http://www.cs.cornell.edu/home/llee/data/convote.html

96

segment file-naming convention, ###_@@@@@@_%%%%$$$_PMV is decoded


as follows28:
1. ### is an index identifying the bill under discussion in the speech segment
(hence, this number also identifies the 'debate' to which the speech segment
belongs)
2. @@@@@@ is an index identifying the speaker
3. %%%% is the index for the page of the Congressional record on which
the speech segment appears, i.e., a number from 0001 to 3268
corresponding to one of the original HTML pages that we downloaded
from govtrack.us .
4. $$$ is an index indicating the position of the speech segment within its
page of the Congressional record. Hence, for example, a file named
055_400144_1031004_DON.txt would be the 4th speech on the 1031st
HTML page of the record.
5. 'P' is replaced by a party indicator, D or R (or X if no corresponding party
could be found). As mentioned in the paper, we purposely *did not* use
this information in our experiments.
6. 'M' is replaced by an indicator of whether the bill under discussion is
mentioned directly in the speech segment, or whether it is only referenced
by another speech segment on the same page. If the bill is directly
mentioned in the current speech, the letter M appears in the file name;
otherwise, the letter O appears.
7. 'V' is replaced by a vote indicator, Y or N, which serves as the ground-truth
label for the speech.
In the following experiments, we only make use of the decoding at positions 1,
2, 5 and 7. We also need to mention, that out of the three variants in which the data
is available, we chose to use the first stage of the data. Therefore, we can use the
references that are annotated within the individual speech segments, which is useful
for the third approach of the first task.
POLARITY CLASSIFICATION
The first experiment we performed was classifying the data on the basis of polarity
of opinion. The first approach to this task was determining the polarity of the debate
segments, taken individually. At this stage, we did not consider the information
regarding the speaker. In order to perform this, we used the similarity measure

28

Taken from the README file of the corpus.

97

given by Ted Pedersens Statistics Package 29 with affect, opinion and attitude
lexicon.
The affect lexicon consisted of three different sources: WordNet Affect - (with 6
categories of emotion joy, surprise, anger, fear, sadness, disgust), the ISEAR
corpus (Scherer and Walbott, 1997) that contains the 7 categories of emotion
anger, disgust, fear, guilt, joy, sadness and shame, from which stopwords are
eliminated) and the emotion triggers database (Balahur and Montoyo, 2008;
Balahur and Montoyo, 2008c; Balahur and Montoyo, 2008e)- which contains terms
related to human needs and motivations annotated with the 6 emotion categories of
WordNet Affect.
The opinion lexicon contained words expressing positive and negative values
(such as good, bad, great, impressive etc.) obtained from the opinion
mining corpus in (Balahur and Montoyo, and to which their corresponding nouns,
verbs and adverbs were added using Rogets Thesaurus.
Finally, the attitude corpus contains the categories of accept, approval,
confidence, importance, competence, correctness, justice, power,
support, truth and trust, with their corresponding antonymic categories
criticism, opposition, uncertainty, doubt, unimportance, incompetence,
injustice, objection, refusal , incorrectness.
After obtaining the similarity scores, we summed up the scores pertaining to
positive categories of emotion, opinion and attitude and the negative categories,
respectively. Therefore, the general positive score was computed as sum of the
individual similarity scores for the categories of joy and surprise from the
affect category, the positive values of the opinion lexicon and the accept,
approval, competence, confidence, correctness, justice, power,
support, trust and truth. On the other hand, the general negative score was
computed as sum of the anger, fear, sadness, shame from the affect
categories, the negative values of the opinion lexicon and the criticism,
opposition, uncertainty, doubt, unimportance, incompetence, injustice,
objection, refusal and incorrectness categories of the attitude lexicon. The
first classification between negative and positive speaker segments was done
comparing these two resulting scores and selecting the higher of the two as final
value for polarity. We evaluated the approach on the development, training and test
sets (Classification 1).
On the other hand, we employed the scores obtained for each of the emotion,
opinion and attitude categories, as well as the combined scores used for classifying
in the first step for the training of an SVM classifier, using the development and
training sets. We then tested the approach on the test set solely. Due to the fact that
in the affect category there are more negative emotions (4) than positive ones (only
29

http://www.d.umn.edu/~tpederse/text-similarity.html

98

2), we chose for classification only the two strongest emotions according to the
similarity scores found in the first approach. Those two categories were fear and
anger. The results are presented under Classification 2.
Further, we parsed the speaker segments using Minipar30, in order to determine
possible dependency paths between words pertaining to our affect, opinion or
attitude lexicon and the topic under discussion or mentioning of another speaker.
Our guess was that many of the speech segments that had been classified as
negative, although the ground-truth annotation had them assigned a positive value,
contained a negative opinion, but not on the topic under discussion, but on the
opinion that was expressed by one of the anterior speakers. Therefore, the goal of
our approach was to see whether the false negatives were due to the classification
method or due to the fact that the object on which the opinion was given was not
the one we had in mind when classifying. In order to verify our hypothesis, we
extracted from the files in which the opinion words from the files with similarity
higher than 0 appeared and sought dependency relations between those words and
the mention of a speaker (based on the number assigned) or the words describing
the topic discussed marked in files in which this names appear, the words bill,
legislation, amendment and measure. Affect, opinion or attitude words to
which no relation was found to the mentioned topic or a speaker were discarded. In
this approach, we did not use anaphora resolution, although, theoretically, it could
help improve the results obtained. It would be interesting to study the effect of
applying anaphora resolution on this task.
The results of the classification are summed up under Classification 3. Figure
4.10 presents an example of the dependency analysis for one of the sentences in
which an attitude word was identified. It can be see that the word support
pertaining to the attitude category, has a dependency path towards the name of the
bill under discussion h.r. 3283. Figure 4.11 shows a schematic overview of the
first approach, with the resources, tools and methods employed therein.
> (
E0
1
2
E2
3
4
5
6
7
8
30

(()
fin C *
(i
~ N
2
(rise ~ V
E0
(()
I N
2
(in
~ Prep
(strong
~ A
(support
~ N
(of
~ Prep
(h.r ~ N
6
(3283 ~ N
7

)
s
(gov rise))
i
(gov fin))
subj (gov rise) (antecedent 1))
2
mod
(gov rise))
5
mod
(gov support))
3
pcomp-n
(gov in))
5
mod
(gov support))
pcomp-n
(gov of))
num
(gov h.r))

http://www.cs.ualberta.ca/~lindek/minipar.htm

99

9
10
11
12
13
14
15
16
17
)

(,
~ U
7
punc (gov h.r))
(the ~ Det 16
det
(gov act))
(united
~ U
12
lex-mod
(gov United States))
(states
United States N
16
nn
(gov act))
(trade
~ A
16
mod
(gov act))
(rights
right N
16
nn
(gov act))
(enforcement
~ N
16
nn
(gov act))
(act ~ N
7
appo (gov h.r))
(.
~ U
*
punc)

Figure 4.10: Minipar output for a sentence in topic 421 on act h.r. 3283
WordNet
Affect

Emotion
triggers

ISEAR
corpus

Opinion
lexicon

S
I
M
I
L
A
R
I
T
Y

S
C
O
R
E

Individual
Speech
Segments

MINIPAR
Attitude
lexicon

Classification 1

SVM
CLASSIFIER

Classification 3

Classification 2

Figure 4.11: Resources and tools scheme for the first approach
The second approach on the data was aggregating the individual speaker
segments on the same debate topic into single documents we denote as speaker
interventions. We then performed, on the one hand, a classification of these
interventions using the sum-up of the scores obtained in the individual speech
segments and, on the other hand, based on the highest score in each of the
categories. Thirdly, we employed SVM to classify the speaker interventions using
the aggregated scores from the individual text segments and the highest scores of
the individual speaker segments, respectively. The training was performed on the
development and training sets and the classifications (Classification 4 and
Classification 5, respectively) were evaluated on the test set.

100

SOURCE CLASSIFICATION
The second experiment we performed was classifying the source of opinions
expressed. In the following experiments, we used the fact that the corpus contained
the name of the party the speaker belonged to coded in the filenames. The goal was
to see whether or not we are able to determine the party a speaker belongs to, by
taking into consideration the words used to express opinion on a given subject, the
arguments (the words used within the argumentation) and the attitude on the subject
in question.
Our hypothesis was that, for example, parties in favor of a certain piece of
legislation will use both a set of words that are positively speaking on the matter, as
well as a set of arguments related to the topic that are highlighting the positive side.
In order to perform this task, we used a clustering on the words pertaining to the
affect lexicon, opinion lexicon and attitude lexicon, as well as the most frequent
words appearing in the individual speaker segments of persons belonging to each of
the two parties Democrat and Republican.
As mentioned by Mullen and Malouf (2006), there are two problems that arise
when intending to classify pertainance to a political party in a topic debate. The first
one is the fact that when talking on a certain topic, all or most persons participating
in the debate will use the same vocabulary. The second issue is that a certain
attitude on a topic cannot reliably predict the attitude on another topic. Related to
the first problem, we verify whether or not attitude towards a topic can be
discriminated on the basis of the arguments given in support or against that topic,
together with the affect, opinion and attitude lexicon used in connection to the
arguments. As far as the second issue is concerned, we do not aim to classify
depending on topic, but rather predict, on the basis of the arguments and affective
words used, the party the speaker belongs to.

4.3.4. EVALUATION
We evaluated the approaches described in terms of precision and recall.
In order to exemplify the manner in which we calculated these scores, we
present confusion matrices for all individual speech segments pertaining to the 5
topics in the development set.
The yes category includes the individual speech segments whose ground truth
was yes. The no category includes the individual speech segments that whose
ground truth was no.

101

The positive category includes the individual speech segments that the system
classified as positive. The negative category includes the individual speech
segments the system classified as negative.
yes
no
positive
30
7
negative
22
16
Table 4.18: Confusion matrix for topic 199 from the development set
yes
no
positive
45
2
negative
14
26
Table 4.19: Confusion matrix for topic 553 from the development set
yes
no
positive
28
3
negative
15
29
Table 4.20: Confusion matrix for topic 421 from the development set
yes
no
positive
44
3
negative
26
58
Table 4.21: Confusion matrix for topic 493 from the development set
yes
no
positive
35
2
negative
24
26
Table 4.22: Confusion matrix for topic 052 from the development set
The following table shows the confusion matrix for the source classification, trained
on the development set and tested using a sample of 100 documents from the test
set, equally distributed among the Democrat and Republican Party.
D
R
Classified D
29
21
Classified R
10
40
Table 4.23: Results for source classification 100 documents

102

We computed precision over positive classification as the number of individual


text segments that our system classified as positive and that had the ground truth
yes divided by the number of individual text segments that our system classified
as positive and that had the ground truth yes summed with the number of
individual text segments that our system classified as positive and that had the
ground truth no.
We computed precision over negative classification as the number of individual
text segments that our system classified as negative and that had the ground truth
no divided by the number of individual text segments that our system classified
as negative and that had the ground truth no summed with the number of
individual text segments that our system classified as negative and that had the
ground truth yes.
Ppos(199) = 30/37 = 0.81; Ppos(553) = 0.95; Ppos(421) = 0.90;
Ppos(493)=0.93; Ppos(052)=0.94
Pneg(199) = 16/38 = 0.43; Pneg(553) = 0.65; Pneg(421) = 0.66;
Pneg(493) = 0.78; Pneg(052)=0.52

We computed recall over positive classification as the number of individual text


segments that our system classified as positive and that had the ground truth yes
divided by the number of individual text segments that our system classified as
positive and that had the ground truth yes summed with the number of individual
text segments that our system classified as negative and that had the ground truth
yes.
We computed recall over negative classification as the number of individual text
segments that our system classified as negative and that had the ground truth no
divided by the number of individual text segments that our system classified as
negative and that had the ground truth no summed with the number of individual
text segments that our system classified as positive and that had the ground truth
no.
Rpos(199) = 30/52 = 0.57; Rpos(553) = 0.76; Rpos(421) = 0.65;
Rpos(493) = 0.62; Rpos(052) = 0.59;
Rneg(199) = 16/23 = 0.70; Rneg(553) = 0.92; Rneg(421) = 0.90;
Rneg(493) = 0.95; Rneg(052)= 0.92

We compute the accuracy score as the sum of the number of correct positive
classifications and the number of correct negative classifications, divided by the
total number of documents on a topic.
A(199) = 46/75 = 0.62; A(553) = 0.81; A(421) = 0.76;
A(493) = 0.77; A (052) = 0.70

103

The overall precision over positive classification is computed as average of all


the precision scores of all the positive classifications. The precision over negative
classifications is computed in the same manner.
The overall recall over positive classification is computed as average of all the
recall scores of all the positive classifications. The recall scores over negative
classifications is computed in the same manner.
The overall accuracy is computed as average of all the accuracy scores of all the
topics.
P pos
P neg R pos R neg
Accuracy
Development set 0.71
0.6
0.63 0.87
0.73
Training set
0.69
0.61
0.63 0.86
0.72
Test set
0.70
0.6
0.62 0.87
0.73
Table 4.24: Classification 1 (individual speaker segments) based on sums of
similarity scores to affect, opinion and attitude lexicon categories
P pos
P neg R pos R neg
Accuracy
Development set 0.72
0.69
0.69 0.85
0.77
Training set
0.71
0.68
0.68 0.85
0.76
Test set
0.70
0.67
0.68 0.84
0.76
Table 4.25: Classification 2 (individual speaker segments) using dependency
parsing
P pos
P neg
R pos R neg
Accuracy
Test set
0. 75 0.63
0.66 0.88
0.78
Table 4.26: Classification 3 (individual speaker segments) based on SVM
Accuracy
Development set
0.68
Training set
0.66
Test set
0.67
Table 4.27: Classification 4 (speaker interventions) based on sum-up of scores
Accuracy
Development set
0.56
Training set
0.55
Test set
0.58
Table 4.28: Classification 5 (individual speaker segments) based on highest score

104

4.3.5. DISCUSSION AND CONCLUSIONS ON SENTIMENT ANALYSIS


FROM POLITICAL DEBATES
As can be noticed from the scores obtained, the system has a rather balanced
behavior on all topics in question. This is a positive fact, because the composition
of the documents in the corpus was different, with as much as double the number of
positive speaker interventions than negative ones.
We notice the tendency of the system to overly classify interventions as
negative, a fact which is rectified by SVM learning, where better results are
obtained. Dependency parsing was also found to help on the classification,
improving the systems performance noticeably. SVM performed better than the
initial method used for classification, but overall, the performance was lower than
that obtained when performing dependency analysis.
We can also notice that when uniting speaker segments, the classification is
done with better results. We believe this to be due to the fact that in the context of
the larger text segments, more of the emotion categories have assigned a value
above 0, and therefore more scores are important to the final result. Another factor
to take into consideration when we analyze the results is the fact that speaker
interventions were not equal as far as text length. From the post evaluation analysis,
we found that the system performs better in classifying longer texts, to which more
categories of emotion have similarity scores above 0.
It is interesting to note the fact that the categories that had the most importance
at the time of classifying were those of fear and joy from the affect list but
given not by the lexicon in WordNet Affect, but from the categories found in the
emotion triggers. As far as opinion source is concerned, the results shown in Table
4.23 demonstrate that a classification with 0.69% accuracy can be easily achieved
using the affect and argument specific lexicon.
In order to tackle the issue of opinion mining from a more general perspective,
we have proposed different methods to classify opinion from texts using affect,
opinion and attitude lexica. We applied the proposed approaches to Congressional
debates. We presented three methods to classify individual speech segments, the
first based on a simple sum of similarity scores to the three lexicons used. We
showed that applying dependency parsing and discovering the target of the opinion
improved the initial classification, not only in the sense of the scores obtained, but
also helped to balance between the results obtained for the positive and negative
categories, respectively. Further, we showed that using SVM machine learning, we
can improve the classification of the opinions, but that SVM performs worse than
dependency analysis in the sense that it does not help improve the misclassification
of positive opinions as negative ones. We showed that classification is dependent
on the text length and that speaker interventions are better classified than individual
speech segments alone, based only on the aggregated score obtained for each of the
105

individual segment files. As far as opinion source is concerned, we showed that


using affect, opinion and attitude lexicon in relation to the arguments given within
the speech segments can help classify the source of opinion with 69% accuracy.

4.4. OPINION MINING FROM BLOGS


4.4.1. INTRODUCTION
Following the initial efforts to propose a general method for opinion mining and the
results obtained, we realized that such an approach required not only appropriate
methods, but new resources that are appropriately annotated, so that NLP systems
could be trained on them.
As we have seen until now, opinions can be present at the level of individual
words, in sentences, phrases and even in consecutive sentences. Moreover, while in
texts such as reviews, the opinion is expressed most of the time by a single source
and about one target, in other types of texts, such as newspaper articles, there are
different components of the text that have to be analyzed separately. From the
experiments we have done until this point and the analysis performed on the results,
we can deduce that the requirements for a resource (annotated corpus) that is able to
capture the phenomena involved in opinion expression so that an automatic system
is able to detect it are the following ones. The resource should:
! Capture the difference between author intention, user interpretations and
directly expressed opinion;
! Distinguish between opinion with sentiment and opinion without
sentiment;
! Contain annotations of emotion, as a basic component of sentiment (so that
experiments can be done to determine sentiment, even in the cases where it
is present in objective sentences);
! Distinguish between opinion expressed in subjective sentences and opinion
expressed through the mention of factual data, directly, indirectly or
implicitly;
! Contain the annotation for opinion source and target (to avoid mixing the
polarity of the context with the sentiment expressed on a specific entity);
! Track mentions of the source and target, whether they are present in text
by anaphoric references, or topic-related mentions;
! Annotate opinion at the appropriate level whether it is expressed through
a word, a phrase, a sentence or various sentences;
! Capture the modalities of opinion expression both in traditional genres of
texts, as well as in genres that are specific to the Social Web e.g. blogs,
forums, reviews, microblogs, social network comments; this includes the
106

possibility to annotate the style of writing, the language employed and the
structure of writing and linking (e.g. misspellings, ungrammaticality,
shortening of words and/or repetition of letters and punctuation signs, the
use of colloquial expressions, urban acronyms, smileys).
The annotation scheme that allows for all these elements to be labeled is called
EmotiBlog and it was proposed by Boldrini et al. (2009). The corpus that was
annotated with this scheme contains blog posts in three languages: Spanish, Italian,
and English about three subjects of interest, which in part overlap with the topics
annotated in the MPQA corpus. The aim in choosing similar topics is that we can
subsequently compare the results of the systems depending on the type of texts
considered.
The first one contains blog posts commenting upon the signing of the Kyoto
Protocol against global warming, the second collection consists of blog entries
about the Mugabe government in Zimbabwe, and finally we selected a series of
blog posts discussing the issues related to the 2008 USA presidential elections. For
each of the abovementioned topics, we have manually selected 100 blog posts,
summing up a total of 30.000 words approximately for each language.

4.4.2. THE EMOTIBLOG ANNOTATION MODEL


The EmotiBlog (Boldrini et al., 2009) annotation is divided into different levels,
detailed in Table 4.29.
Element
Objective speech
Subjective speech
Adjectives
Adverbs
Verbs
Anaphora
Capital letter
Punctuation
Names

Description
Confidence, comment, source, target.
Confidence, comment, level, emotion, phenomenon, polarity,
source and target.
Confidence, comment, level, emotion, phenomenon,
modifier/not, polarity, source and target.
Confidence, comment, level, emotion, phenomenon,
modifier/not, polarity, source and target.
Confidence, comment, level, emotion, phenomenon, polarity,
mode, source and target.
Confidence, comment, type, source and target.
Confidence, comment, level, emotion, phenomenon,
modifier/not, polarity, source and target.
Confidence, comment, level, emotion, phenomenon,
modifier/not, polarity, source and target.
Confidence, comment, level, emotion, phenomenon,
107

Element
Reader
Interpretation
Author
Interpretation
Emotions

Description
modifier/not, polarity, and source.
Confidence, comment, level, emotion, phenomenon, polarity,
source and target.
Confidence, comment, level, emotion, phenomenon, polarity,
source and target.
Confidence, comment; accept, anger, anticipation, anxiety,
appreciation, bad, bewilderment, comfort, compassion,
confidence, consternation, correct, criticism, disappointment
discomfort, disgust, despondency, depression, envy, enmity,
excuse, force, fear, grief, guilt, greed, hatred, hope, irony,
interesting, important.
Table 4.29: EmotiBlog structure

For each element we are labelling the annotator has to insert his level of
confidence. In this way, each label is assigned a weight that will be computed for
future evaluations. Moreover, the annotator has to insert the polarity, which can be
positive or negative, the level (high, medium, and low) and also the emotion this
element is expressing. The phenomenon level describes whether the element is a
saying or a colloquialism or a multi-word phrase.
As suggested by Balahur and Steinberger (2009), even if the writer uses an
apparently objective formulation, he/she intends to transmit an emotion and a
sentiment. For this reason we added two elements: reader and author interpretation.
The first one is the impression/feeling/reaction the reader has reading the
intervention and what s/he can deduce from the piece of text and the author
interpretation is what we can understand from the author (politic orientation,
preferences). Another innovative element we inserted in the model is the coreference but just at a cross-post level. It is necessary because blogs are composed
by posts linked between them and thus cross-document co-reference can help the
reader to follow the conversations. We also label the unusual usage of capital letters
and repeated punctuation. In fact, it is very common in blogs to find words written
in capital letter or with no conventional usage of punctuation; these features usually
mean shouts or a particular mood of the writer. Using EmotiBlog, we annotate the
single elements, but we also mark sayings or collocations, representative of each
language. Finally we insert for each element the source and topic.

108

4.4.3. EXPERIMENTS AND EVALUATIONS OF EMOTIBLOG ON


ENGLISH CORPORA
In order to evaluate the appropriateness of the EmotiBlog annotation scheme and
to prove that the fine-grained level it aims at has a positive impact on the
performance of the systems employing it as training, we performed several
experiments.
Given that a) EmotiBlog contains annotations for individual words, as well as for
multi-word expressions and at a sentence level, and b) they are labeled with
polarity, but also emotion, our experiments show how the annotated elements can
be used as training for the opinion mining and polarity classification task, as well as
for emotion detection. Moreover, taking into consideration the fact that EmotiBlog
labels the intensity level of the annotated elements; we performed a brief
experiment on determining the sentiment intensity, measured on a three-level scale:
low, medium and high.
In order to perform these three different evaluations, we chose three different
corpora. The first one is a collection of quotes (reported speech) from newspaper
articles presented by Balahur et al. (Balahur et al., 2010d), enriched with the
manual fine-grained annotation of EmotiBlog; the second one is the collection of
newspaper titles in the test set of the SemEval 2007 task number 14 Affective
Text. Finally, the third one is a corpus of self-reported emotional response ISEAR
(Scherer and Walbott, 1999). The intensity classification task is evaluated only on
the second corpus, given that it is the only one in which scores between -100 and 0
and 0 and 100, respectively, are given for the polarity of the titles.
CREATION OF TRAINING MODELS
For the OM and polarity classification task, we first extracted the Named Entities
contained in the annotations using Lingpipe and united through a _ all the tokens
pertaining to the NE. All the annotations of punctuation signs that had a specific
meaning together were also united under a single punctuation sign. Subsequently,
we processed the annotated data using Minipar. We compute, for each word in a
sentence, a series of features (some of these features are used by Choi et al. (2005)):
! the part of speech (POS);
! capitalization (if all letters are in capitals, if only the first letter is in capitals,
and if it is a NE or not);
! opinionatedness/intensity/emotion - if the word is annotated as opinion word,
its polarity, i.e. 1 and -1 if the word is positive or negative, respectively and 0
if it is not an opinion word, its intensity (1.2 or 3) and 0 if it is not a subjective
word, its emotion (if it has, none otherwise);
109

syntactic relatedness with other opinion word if it is directly dependent of an


opinion word or modifier (0 or 1), plus the polarity/intensity and emotion of
this word (0 for all the components otherwise);
! role in 2-word, 3-word and 4-word annotations: opinionatedness, intensity and
emotion of the other words contained in the annotation, direct dependency
relations with them if they exist and 0 otherwise.
Finally, we add for each sentence as feature binary features for subjectivity and
polarity, the value corresponding to the intensity of opinion and the general
emotion. These feature vectors are fed into the Weka 31 SVM SMO machine
learning algorithm and a model is created (EmotiBlog I). A second model
(EmotiBlog II) is created by adding to the collection of single opinion and emotion
words annotated in EmotiBlog, the Opinion Finder lexicon and the opinion words
found in MicroWordNet, the General Inquirer resource and WordNet Affect.
!

EVALUATION RESULTS OF EMOTIBLOG-BASED MODELS ON TEST


SETS
In order to evaluate the performance of the models extracted from the features of
the annotations in EmotiBlog, we performed different tests. The first one regarded
the evaluation of the polarity and intensity classification task using the Emoitblog I
and II constructed models on two test sets the JRC quotes collection and the
SemEval 2007 Task Number 14 test set. Since the quotes often contain more than a
sentence, we consider the polarity and intensity of the entire quote as the most
frequent result in each class, corresponding to its constituent sentences. Also, given
the fact that the SemEval Affective Text headlines were given intensity values
between -100 and 100, we mapped the values contained in the Gold Standard of the
task into three categories: [-100, -67] is high (value 3 in intensity) and negative
(value -1 in polarity), [-66, 34] medium negative and [33, 1] is low negative. The
values between [1 and 100] are mapped in the same manner to the positive
category. 0 was considered objective, so containing the value 0 for intensity. The
results are presented in Table 4.30 (the values I and II correspond to the models
EmotiBlog I and EmotiBlog II):
Test
Corpus
JRC
quotes I
JRC
quotes II
31

Evaluation
type
Polarity
Intensity
Polarity
Intensity

http://www.cs.waikato.ac.nz/ml/weka/

110

Precision

Recall

32.13
36.00
36.4
38.7

54.09
53.2
51.00
57.81

Test
Evaluation
Precision
Recall
Corpus
type
Polarity
38.57
51.3
SemEval
I
Intensity
37.39
50.9
Polarity
35.8
58.68
SemEval
II
Intensity
32.3
50.4
Table 4.30: Results for polarity and intensity classification using models built on
EmotiBlog
The results presented in Table 4.30 show a significantly high improvement over
the results obtained in the SemEval task in 2007. This is explainable, on the one
hand, by the fact that systems performing the opinion task did not have at their
disposal the lexical resources for opinion employed in the EmotiBlog II model, but
also because of the fact that they did not use machine learning on a corpus
comparable to EmotiBlog (as seen from the results obtained when using solely the
EmotiBlog I corpus). Compared to the NTCIR 8 Multilingual Analysis Task this
year, we obtained significant improvements in precision, with a recall that is
comparable to most of the participating systems.
In the second experiment, we tested the performance of emotion classification
using the two models built using EmotiBlog on the three corpora JRC quotes,
SemEval 2007 Task No.14 test set and the ISEAR corpus. The JRC quotes are
labeled using EmotiBlog; however, the other two are labeled with a small set of
emotions 6 in the case of the SemEval data (joy, surprise, anger, fear, sadness,
disgust) and 7 in ISEAR (joy, sadness, anger, fear, guilt, shame, disgust). Moreover,
the SemEval data contains more than one emotion per title in the Gold Standard,
therefore we consider as correct any of the classifications containing one of them.
In order to unify the results and obtain comparable evaluations, we assessed the
performance of the system using the alternative dimensional structures defined by
Boldrini et al. (2009). The ones not overlapping with the category of any of the 8
different emotions in SemEval and ISEAR are considered as Other and are not
included either in the training, nor test set. The results of the evaluation are
presented in Table 4.31. Again, the values I and II correspond to the models
EmotiBlog I and II. The Emotions category contains the following emotions: joy,
sadness, anger, fear, guilt, shame, disgust, surprise.
Test
corpus
JRC
quotes I

Evaluation
type
Emotions

Precision

Recall

24.7

15.08

111

Test
Evaluation
Precision
Recall
corpus
type
JRC
Emotions
33.65
18.98
quotes II
SemEval
Emotions
29.03
18.89
I
SemEval
Emotions
32.98
18.45
II
ISEAR I Emotions
22.31
15.01
ISEAR
Emotions
25.62
17.83
II
Table 4.31: Results for emotion classification using the models built on
EmotiBlog
The best results for emotion detection were obtained for the anger category,
where the precision was around 35 percent, for a recall of 19 percent. The worst
results obtained were for the ISEAR category of shame, where precision was
around 12 percent, with a recall of 15 percent. We believe this is due to the fact that
the latter emotion is a combination of more complex affective states and it can be
easily misclassified to other categories of emotion. Moreover, from the analysis
performed on the errors, we realized that many of the affective phenomena
presented were more explicit in the case of texts expressing strong emotions such as
joy and anger, and were mostly related to common-sense interpretations of the
facts presented in the weaker ones.
As it can be seen in Table 4.30, results for the texts pertaining to the news
category obtain better results, most of all news titles. This is due to the fact that
such texts, although they contain a few words, have a more direct and stronger
emotional charge than direct speech (which may be biased by the need to be
diplomatic, find the best suited words etc.). Finally, the error analysis showed that
emotion that is directly reported by the persons experiencing is more hidden, in
the use of words carrying special signification or related to general human
experience. This fact makes emotion detection in such texts a harder task.
Nevertheless, the results in all corpora are comparable, showing that the approach is
robust enough to handle different text types. All in all, the results obtained using the
fine and coarse-grained annotations in EmotiBlog increased the performance of
emotion detection as compared to the systems in the SemEval competition.
DISCUSSION ON THE OVERALL RESULTS
From the results obtained, we can see that this approach combining the features
extracted from the EmotiBlog fine and coarse-grained annotations helps to balance
112

between the results obtained for precision and recall. The impact of using additional
resources that contain opinion words is that of increasing the recall of the system, at
the cost of a slight drop in precision, which shows that the approach is robust
enough so that additional knowledge sources can be added. Although the corpus is
small, the results obtained show that the phenomena captured by the approach are
relevant to the opinion mining task, not only for the blog sphere, but also for other
types of text (newspaper articles, self-reported affect).
Another advantage of EmotiBlog is the fact that it contains texts in three languages:
English, Spanish and Italian. That is why, in the following experiments, we will test
the usability of the resource in a second language, namely, Spanish.

4.4.4. USING EMOTIBLOG TO DETECT AND CLASSIFY SENTIMENT


IN SPANISH TEXTS
Our annotation model includes word/ phrase/ text levels of annotation. It is thus
useful for constructing similarity models for the training of machine learning
algorithms working with different values of n-grams, as well as sentences as a
whole.
EmotiBlog can be used to extrinsically evaluate systems mining opinions.
Moreover, our annotation scheme can be used either for basic tasks of sentiment
polarity classification, as well as emotion detection, either on very fine-grained
categories, as well as psychology-based emotion classes. Furthermore, most work
done in opinion mining only concentrated on classifying polarity of sentiments into
positive or negative. Thus, another contribution our work brings is the classification
according to three categories: positive, negative and neutral. The last category is
made up of both opinionated sentences in which there is no clear indication of
approval or disapproval of an idea, as well as objective sentences. Thirdly, our
research focused both on the sentiment polarity classification, as well as the ratio
between computational costs versus performance. Since our final goal is to employ
the system in a real life scenario where it would adequately respond to user
opinionated input with multimedia feedback, we must be aware of the advantages
and disadvantages the use of each resource has on the system. Therefore, another
contribution we bring is the assessment of performance versus time ratios. Last, but
not least, we contribute to the research in the field by proposing and evaluating a
method for sentiment polarity classification, based on n-gram and phrase similarity
features used with machine learning. We evaluate the method both intrinsically
by cross-fold validation of the subjective and phrases in the corpus, as well as
extrinsically, using a corpora of negative, positive and neutral opinions on
recycling. The system was trained and tested using the annotation in the created
corpus. However, given a different domain, it can be equally used, given that
training examples are available.
113

The main aim of this experiment, described by Balahur et al. (Balahur et al.,
2009b) is to obtain a system able to mine opinion from user input in real time and,
according to the inferred information on the polarity of the sentiment, offer them
corresponding feedback. The topic of our opinion mining experiment is
recycling: the computer asks about a persons opinion on recycling and then the
user answers this question generating a sentence with emotion that can be of
different intensity levels. The system reacts to the user input with some messages or
faces that correspond to the reactions for the users feedback.
For the task at hand we employ annotations from texts on the Kyoto protocol
pertaining to the EmotiBlog corpus. We use the annotated elements to train our
opinion mining system and then classify new sentences that are on the topic
recycling (to which some vocabulary similarity can be found, since they both
topics refer to environmental issues). For the task at hand, we manually created a
set of 150 sentences on recycling, 50 for each of the positive, negative and neutral
categories. The first experiment carried out aimed at proving that the corpus is a
valid resource and we can use the annotations for the training of our opinion mining
system. For this assessment, we use the same methodology we will further employ
to mine opinions from user input.
CROSS-FOLD EVALUATION OF THE ANNOTATION
As a result of the annotation, we obtained 1647 subjective phrases and 1336
objective ones. Our agreement was 0.59 for subjective phrases and 0.745 for the
objective one.
Further on, we will consider for our tests only the sentences upon which we agreed
and the phrases whose annotation length was above four tokens of the type noun,
verb, adverb or adjective. For the cross-validation of the corpus, each of the
sentences is POS-Tagged and lemmatized using FreeLing 32 . Further on, we
represent each sentence as a feature vector, whose components are unigram features
containing the positive and respectively negative categories of nouns, verbs,
adverbs, adjectives, prepositions and punctuation signs (having 1 in the
corresponding position of the feature vector for the words contained and 0
otherwise), the number of bigrams/ trigrams and 4-grams overlapping with each of
the phrases we have annotated as positive and negative or objective, respectively
and finally the overall similarity given by the number of overlapping words with
each of the positive and negative or objective phrases from the corpus, normalized
by the length of the given phrase. We test out method in two steps: first of all the
classification of sentences among subjective and objective, for which the vectors
contain as final values subjective or objective and second of all the
classification of subjective sentences into positive and negative, for which case the
32

http://www.lsi.upc.edu/~nlp/freeling/

114

classification vectors contain the values positive and negative. We perform a


ten-fold cross validation of the corpus for each of the two steps. The results are
presented in Table 4.32, in terms of precision, recall and kappa.
Precision
Recall
Kappa
Subjective
0.988
0.6
0.43
Objective
0.68
0.89
0.43
Positive
0.799
0.511
0.68
Negative
0.892
0.969
0.68
Table 4.32: Results of ten-fold corpus cross-validation

CLASSIFICATION OF NEW EXAMPLES


The second experiment we performed concerned the assessment of the systems
performance as far as the sentiment polarity classification of the sentences on
recycling is concerned. Here, the challenge we are facing is the possible difference
in the lexicon used. However, we assume the affective vocabulary to be
approximately the same. We will consider as subjective neutral the sentences that
are more similar to the objective statements in the training corpus. We will also
test the importance of the fine-grained annotations on the classification
performance.
Therefore, we will consider two scenarios: one in which we use the unigrams
given by the annotated nouns, adjectives, verbs and adverbs from the corpus that we
find in the phrases to be annotated; in the second scenario, we will only use the ngram overlaps with n greater than 2 and the overall sentence similarity. Thus, we
can obtain an overall evaluation of the importance of detecting also single words
that have an affective charge.
CLASSIFICATION USING ALL N-GRAM FEATURES
For this first classification of our test data, we first run FreeLing on the set of
positive, negative and neutral sentences on recycling in order to lemmatize and tag
each word on part of speech. We then represent each sentence as a feature vector, in
the same manner as in the first conducted experiment. Further on, we conduct two
experiments on this data. The first one aims at training an SVM classifier on the
corpus phrases pertaining to the subjective versus objective categories and test
it on the statements on the recycling topic pertaining to the positive or negative
versus neutral categories. The second experiment consists in classifying the
instances according to three polarity classes: positive, negative and neutral. The
results of the two experiments are summarized in Table 4.33.

115

Precision
Recall
Kappa
Subjective
0.977
0.619
0.409
Objective
0.44
0.95
0.409
Positive
0.881
0.769
0.88
Negative
0.92
0.96
0.88
Table 4.33: Classification results using n-grams
As we can notice from the results, using the annotated elements, it is easier to
distinguish the subjective sentences, due to the fact that we train on subjective ngrams. As far as the positive, negative and neutral classification is concerned, the
results are both high, as well as balanced, proving the correctness of our approach.
CLASSIFICATION USING N-GRAMS, N>2.
In this experiment, we test the importance of annotating affect in texts at the token
level. From our blog corpus, we have a large number of nouns, verbs, adverbs and
adjective, annotated as positive or negative and at the emotion level. We used these
words at the time of classifying examples using n-grams, with n ranging from 1 to 4
(in 5.2.1). To test their importance, we removed the vector components accounting
for their presence in the feature vectors and re-classified, both at the level of
objective versus subjective, as well as at the positive, negative, neutral level. In the
table below, we can see the results obtained.
Precision
Recall
Kappa
Subjective
0.93
0.60
0.43
Objective
0.43
0.7
0.43
Positive
0.83
0.64
0.85
Negative
0.90
0.91
0.85
Neutral
0.90
0.96
0.85
Table 4.34: Classification results using n-grams, n>2
As we can see, removing single words with their associated polarities from the
training data resulted in lower scores. Therefore, fine-grained annotation helps at
the time of training the opinion mining system and is well-worth the effort.

4.5. CONCLUSIONS ON THE PROPOSED METHODS FOR


SENTIMENT ANALYSIS
In this chapter, we presented different methods and resources we built for the task
of sentiment analysis in different text types.
We started by presenting methods to tackle the task of feature-based opinion
mining and summarization, applied to product reviews. We have analyzed the
116

peculiarities of this task and identified the weak points of existing research. We
proposed and evaluated different methods to overcome the identified challenges,
among which the most important were the discovery of indirectly mentioned
features and the computation of the polarity of opinions in a manner that is featuredependent. Subsequently, we proposed a unified model for sentiment annotation for
this type of text, able to capture the important phenomena that we had identified
different types of sentiment expressions, feature mentioning and span of text
expressing a specific opinion.
Further on, we explored different methods to tackle sentiment analysis from
newspaper articles. After the initial experiments, we analyzed the reasons for the
low performance obtained and redefined the task, taking into account the
peculiarities of this textual genre. We created an annotation model and labeled two
different corpora of newspaper article quotations, in English and German. After
redefining the task and delimiting the scope of the sentiment analysis process to
quotations small text snippets containing direct speech, whose source and target
are previously known-, the annotation agreement rose significantly. Additionally,
improving the definition of the task made it possible to implement automatic
processing methods that are appropriate for the task and significantly improve the
performance of the sentiment analysis system we had designed. In the view of
applying sentiment analysis to different types of texts, in which objective content is
highly mixed with subjective one and where the sources and targets of opinions are
multiple, we have proposed different general methods for sentiment analysis, which
we applied to political debates.
The results of this latter experiment motivated us to analyze the requirements of
a general labeling scheme for the task of sentiment analysis, which can be used to
capture all relevant phenomena in sentiment expression.
To this aim, Boldrini et al. (2009) defined EmotiBlog, an annotation scheme that
is able to capture, at a fine-grained level, all linguistic phenomena related to
sentiment expression in text. The subsequent experiments have shown that this
model is appropriate for the training of machine learning models for the task of
sentiment analysis in different textual genres, in both languages in which
experiments have been carried out using it English and Spanish.
In this chapter, we have only concentrated on the task of sentiment analysis as a
standalone challenge, omitting the steps required in order to obtain the texts on
which the sentiment analysis methods were applied. In a real scenario, however,
automatically detecting the opinion expressed in a text is not the first task to be
performed. Additionally, in many of the cases, the results obtained after
automatically processing texts to determine the sentiment they contain still pose
many problems in terms of volume. Thus, even if the sentiment is determined

117

automatically, one may still require a summarization component, in order to further


reduce the quantity of information so that it is usable by a person.
Therefore, real-world applications that contain an opinion mining component
must also contemplate the integration with other NLP systems. In the next chapter,
we describe the challenges faced when integrating opinion mining systems with
other NLP technologies, such as information retrieval (IR), question answering
(QA) and text summarization (SUM). We first propose different methods to employ
sentiment analysis in the context of NLP systems that are traditionally used for the
analysis of factual data. Starting from the low results obtained in these preliminary
efforts, we describe and evaluate new methods to tackle traditional NLP tasks (such
as IR, QA and SUM) in the context of opinionated content. Our objective is to
redefine these tasks, so that the phenomena encountered in opinionated texts are
taken into account at every step of the processing, thus improving the overall
performance of final systems.

118

CHAPTER 5. APPLICATIONS OF SENTIMENT


ANALYSIS
Motto: He who molds public opinion makes statues and decisions possible or
impossible to make. (Abraham Lincoln)

5.1. INTRODUCTION
In the previous chapter, we presented different methods to perform sentiment
analysis from a variety of text types. We have shown what the challenges related to
each of these genres are and how the task of opinion mining can be tackled in each
of them.
Nevertheless, real-world applications of sentiment analysis often require more
than an opinion mining component. On the one hand, an application should allow a
user to query about opinions, in which case the documents in which these opinions
appear have to be retrieved. In more user-friendly applications, the users can be
given the option to formulate the query into a question in natural language.
Therefore, question answering techniques must be applied in order to determine
the information required by the user and subsequently retrieve and analyze it. On
the other hand, opinion mining offers mechanisms to automatically detect and
classify sentiments in texts, overcoming the issue given by the high volume of such
information present on the Internet. However, in many cases, even the result of the
opinion processing by an automatic system still contains large quantities of
information, which are still difficult to deal with manually. For example, for
questions such as Why do people like George Clooney? we can find thousands of
answers on the Web. Therefore, finding the relevant opinions expressed on George
Clooney, classifying them and filtering only the positive opinions is not helpful
enough for the user. He/she will still have to sift through thousands of texts
snippets, containing relevant, but also much redundant information. Moreover,
when following the comments on a topic posted on a blog, for example, finding the
arguments given in favor and against the given topic might not be sufficient to a
real user. He/she might find the information truly useful only if it is structured and
has no redundant pieces of information. Therefore, apart from analyzing the opinion
in text, a real-world application for sentiment analysis could also contain a
summarization component.
The aim of the work presented in this chapter is to apply the different opinion
mining resources, tools and approaches to other tasks within NLP. The objective
was, on the one hand, to evaluate the performance of our approaches, and, on the
other, to test the requirements and extra needs of an opinion mining system in the
context of larger applications. In this chapter, we present the research we carried
119

out in order to test the manner in which opinion mining can be best combined with
information retrieval (IR), question answering (QA) and summarization (SUM), in
order to create a useful, real-life, end-to-end system for opinion analysis in text. Our
initial efforts concentrated on applying already-existing IR, QA and SUM systems
in tandem with our sentiment analysis systems. Subsequently, having realized that
directly applying systems that were designed to deal with factual data in the context
of opinionated text led to low results, we proposed new methods to tackle IR, QA
and SUM in a manner that is appropriate in the context of subjective texts.
In this chapter, we present the methods and improvements achieved in Opinion
Question Answering and Opinion Summarization.

5.2. OPINION QUESTION ANSWERING AND


SUMMARIZATION
While techniques to retrieve objective information have been widely studied,
implemented and evaluated, opinion-related tasks still represent an important
challenge. As a consequence, the aim of this section of research is to study,
implement and evaluate appropriate methods for the task of Question Answering
(QA) in the context of opinion treatment.
The experience in the TAC 2008 Opinion Pilot competition (Balahur et al., 2008),
as well as the post-competition experiments (Lloret et al., 2009), motivated us to
continue the research we started with the participation in the TAC competition with
the study of different aspects of opinion Question Answering and opinion
summarization, such as classification of fact versus opinion questions (Balahur et
al., 2009c), defining opinion answer types and methods to retrieve answers to
opinion questions (Balahur et al., 2009a; Balahur et al., 2009d; Balahur et al.,
2009h; Balahur et al., 2010a; Balahur et al., 2010b; Balahur et al., 2009e) or
opinion-driven summarization in the context of threads in blogs (Balahur et al.,
2010e; Balahur et al., 2009g; Balahur et al., 2009i; Kabadjov et al., 2009).

5.2.1. PARTICIPATION TO THE TAC 2008 OPINION PILOT TASK


The Text Analysis Conference 2008 edition (TAC 2008) contained three tasks:
Question Answering, Recognizing Textual Entailment and Summarization. The
Summarization track included two tasks: the Update Summarization and the
Opinion Summarization Pilot task. We participated in the Opinion Summarization
Pilot task within the Summarization track, with the aim of measuring the
performance of our opinion mining system (Balahur and Montoyo, 2008c; Balahur
and Montoyo, 2008d). However, given the fact that the referenced system was built
to deal only with opinions on products and their features, in the end we changed our
120

approach to a more general one, which is similar to the one used to determine
opinions from political debates (Balahur et al., 2009e).
The Opinion Summarization Pilot task consisted in generating summaries from
blogs snippets that were the answers to a set of opinion questions. The participants
were given a set of blogs from the Blog06 collection and a set of squishy list
(opinion) questions from the Question Answering track, and had as task to produce
a summary of the blog snippets that answered these questions. There were 25
targets, and on each of them one or two questions were formulated. All these
questions concerned the attitude held by specified sources on the targets given. For
example, for the target George Clooney, the two questions asked were Why do
people like George Clooney? and Why do people dislike George Clooney?).
Additionally, a set of text snippets were also provided, which contained the answers
to the questions. These snippets were selected from the answers given by systems
participating in the Question Answering track, and opinion summarization systems
could either use them or choose to perform themselves the retrieval of the answers
to the questions in the corresponding blogs.
Within our participation in the Opinion Summarization Pilot task, we used two
different methods for opinion mining and summarization. The two approaches
suggested were different only as far as the use of the optional text snippets provided
by the TAC organization was concerned. Our first approach (the Snippet-driven
Approach) used these snippets, whereas the second one (Blog-driven Approach)
found the answers directly in the corresponding blogs.
In the first phase, we processed the questions, in order to determine a set of
attributes that will further help us find and filter the answers. The process is
described in Figure 5.1. In order to extract the topic and determine the question
polarity, we define question patterns. These patterns take into consideration the
interrogation formula and extract the opinion words (nouns, verbs, adverbs,
adjectives and their determiners). The opinion words are then classified in order to
determine the polarity of the question, using the WordNet Affect emotion lists, the
emotion triggers resource (Balahur and Montoyo, 2008), a list of four attitudes that
we built, containing the verbs, nouns, adjectives and adverbs for the categories of
criticism, support, admiration and rejection and two categories of value words
(good and bad) taken from the opinion mining system proposed by Balahur and
Montoyo (Balahur and Montoyo, 2008c).
Examples of rules for the interrogation formula What reasons are:
1. What reason(s) (.*?) for (not) (affect verb + ing) (.*?)?
2. What reason(s) (.*?) for (lack of) (affect noun) (.*?)?
3. What reason(s) (.*?) for (affect adjective ||positive ||negative) opinions
(.*?)?

121

Using these extraction patterns, we identified the nouns, verbs, adjectives etc.
that gave an indication of the question polarity 33 . Further on, these indicators
were classified according to the affect lists mentioned above. The keywords of the
question are determined by eliminating the stop words. At the end of the question
processing stage, we obtain, on the one hand, the reformulation patterns (that are
eventually used to link and give coherence to the final summaries) and, on the other
hand, the question focus, keywords and the question polarity. Depending on the
focus/topic and polarity identified for each question, a decision on the further
processing of the snippet was made, using the following rules:
1. If there is only one question made on the topic, determining its polarity is
sufficient for making the correspondence between the question and the
snippets retrieved; the retrieved snippet must simply obey the criteria that it
has the same polarity as the question.
2. If there are two questions made on the topic and each of the questions has a
different polarity, the correspondence between the question and the answer
snippets can simply be done by classifying the snippets retrieved according
to their polarity.
3. If there are two questions that have different focus but different polarities,
the correspondence between the questions and the answer snippets is done
using the classification of the answer snippets according to focus and
polarity.
4. If there are two questions that have the same focus and the same polarity,
the correspondence between the questions and the answer snippets is done
using the order of appearance of the entities in focus, both in the question
and in the possible answer snippet retrieved, simultaneously with the
verification that the intended polarity of the answer snippet is the same as
that of the question.
The categorization of questions into these four classes is decisive at the time of
making the question - answer snippet correspondence, in the snippet/blog phrase
processing stage. Details on these issues are given in what follows and in Figure
5.1.

33

Later in this chapter, as we will redefine the task of question answering in the context of opinions,
we will refer to this concept as Expected Polarity Type (EPT). Although in this first experiment, we
called it question polarity, the new EPT term was necessary, as question polarity entails that the
question in itself has some sort of orientation, when in fact it is the polarity of the expected answer that
is actually computed at this stage.

122

WN
AFFECT

QUESTION
PATTERNS

REFORMULATION
PATTERNS

EMOTION
TRIGGERS

ATTITUDE
WORDS

VALUE
WORDS

QUESTIONS

QUESTION
KEYWORDS

QUESTION
FOCUS

QUESTION
POLARITY

DECISION
ON
SNIPPET

Figure 5.1: Question Processing Stage


In the first approximation, the given answer-snippets constitute the basis for looking
up the phrases these snippets were extracted from in the blogs collection. In the
second approximation, we use the question keywords to determine the phrases from
the blogs that could constitute answers to the questions. Further on, the blog
original phrases or the blog retrieved phrases, respectively, are classified according
to polarity, using the vector similarity with the set of vectors consisting of three
distinct subsets.
The first subset of vectors is built from the phrases in the ISEAR corpus
(without stop words), one vector per statement and having the phrases classified
according to the emotion it described. The second subset of vectors is built
according to the WN Affect list of words in the joy, anger, sadness and fear. The
third subset consists of the vectors of emotions from the emotion triggers resource,
on each of the 4 categories that were considered also for building the vectors from
WordNet Affect.
For each of the blog phrase, we compute the similarity with all the vectors.
Further on, each of the emotions is assigned a polarity - emotions from the fear,
anger, sadness, disgust, shame, guilt categories are assigned a negative polarity and emotions from the joy and surprise categories we considered as positive. The
final polarity score was computed as sum of the scores obtained in each of the
vector similarity computations. The higher of the two scores - positive or negative is considered as being the snippet polarity. In the second approximation, we also
perform a sorting of the phrases retrieved, in descending order, according to their
polarity scores. This is helpful at the time of building the final summary, whose
length must not surpass a given limit (in this case, the organizers set the limit to
7000 characters).

123

In the final phrases used in creating the summary we added, for coherence
reasons, the reformulation patterns deduced using the question structure. Taking
into consideration the number of characters limitation, we only included in the
summary the phrases with high positive scores and those with high negative scores,
completed with the reformulation patterns, until reaching the imposed character
limit. Thus, the score given by the sentiment analysis system constituted the main
criteria for selecting the relevant information for the summaries and the F-measure
score reflects the quality of the opinion mining process.
SNIPPETS/KEYWORDS
WN
AFFECT

ORIGINAL BLOG
PHRASES

EMOTION
TRIGGERS

ISEAR
SNIPPET/P.
/
POLARITY

SORT BY
POOLARITY
STRENGTH

DECISION ON
SNIPPET

QUESTIONQ
SNIPPET
CORRESPON
ANCE

SNIPPET/P.
TOPIC

REFORMULATION
PATTERNS

FINAL
SUMMARIES

Figure 5.2: The snippet/ blog phrase processing stage


In the first approach, we used the snippets provided by the organizers. We
automatically analyzed them in order to determine the focus and polarity in each of
the given questions for a topic, determine the associated given answer snippet (by
computing the snippets polarity and focus), locate the whole sentences snippet
within the corresponding blog, and finally use patterns of reformulation from the
questions structure to bind together the snippets for the same polarity and focus to
produce the final summary. Having computed the polarity for the questions and text
snippets, and having set out the final set of sentences to produce the summary with
their focus, we bound each sentence to its corresponding question, and we grouped
all sentences which were related to the same question together, so that we could
generate the language for this group, according to the patterns of reformulation that
were created for each question. Finally, the speech style was changed to an
impersonal one, in order to avoid the presence of directly expressed opinion
sentences. A POS-tagger tool (TreeTagger) was used to identify third person verbs
124

and change them to a neutral style. A set of rules to identify pronouns was created,
and they were also changed to the more general pronoun they and its
corresponding forms, to avoid personal opinions.
The second approach had as starting point determining the focus, keywords,
topic and polarity in each of the given questions. The processing of the question is
similar to the one performed for the first approximation. Starting from the focus,
keywords and topic of the question, we sought sentences in the blog collection
(previously processed as described in the first approximation) that could constitute
possible answers to the questions, according to their similarity to the latter. The
similarity score was computed with Pedersens Text Similarity Package 34 . The
snippets thus determined underwent dependency parsing with Minipar and only the
sentences which contained subject and predicate were kept, thus ensuring the
elimination of some of the present noise (such as section titles, dates, times etc.).
The remaining snippets were classified according to their polarity, using the
similarity score with respect to the described emotion vectors. The direct language
style was changed to indirect speech style. The reformulation patterns that were
deduced using the questions structure were added to bind together the snippets and
produce the final summary, concatenating the snippets with the added
reformulations. Since the final length of the summary could easily overpass the
imposed limit, we sorted the snippets using their polarity strength (the higher the
polarity score - be it positive or negative the higher the rank of the snippet), and
included the reformulated snippets in descending order until the final limit was
reached.
EVALUATION OF THE APPROACH IN THE TAC2008 COMPETITION
45 runs were submitted by 19 teams for evaluation in the TAC 2008 Opinion Pilot
task. Each team was allowed to submit up to three runs, but finally, due to the
difficulty involved in the evaluation of such a task, only the first two runs of each
team was evaluated, leading to 36 runs being evaluated. Table 5.1 shows the final
results obtained by the first two runs we submitted for evaluation in the TAC 2008
Opinion Pilot and the rank they had compared to the other participating systems.
The column numbers stand for the following information:
1. summarizerID ( our Run 1 had summarizerID 8 and Run 2 had
summarizerID 34)
2. Run type: "manual" or "automatic"
3. Did the run use the answer snippets provided by NIST: "Yes" or "No"
4. Average pyramid F-score (Beta=1), averaged over 22 summaries
5. Average score for Grammaticality
34

http://wn-similarity.sourceforge.net/

125

6. Average score for Non-redundancy


7. Average score for Structure/Coherence (including focus and referential
clarity)
8. Average score for Overall fluency/readability
9. Average score for Overall responsiveness
1
8
34

2
3
4
5
6
7
8
automatic Yes
0.357
4.727
5.364
3.409
3.636
automatic
No
0.155
3.545
4.364
3.091
2.636
Table 5.1: Evaluation results in the TAC 2008 competition

9
5.045
2.227

Further on, we will present the system performances with respect to all other
teams, first as an overall classification (Table 5.2) and secondly, taking into
consideration whether or not the run used the optional answer snippets provided by
NIST (Table 4).
In Table 5.2, the numbers in columns 4, 5, 6, 7, 8 and 9 correspond to the
position within the 36 evaluated submissions. In Table 5.3, the numbers in columns
4, 5, 6, 7, 8 and 9 correspond to the position within the 17 submissions that used the
given optional answer snippets (in case of Run 1) and the position within the 19
submissions evaluated that did not use the provided answer snippets.
1
8
34

2
3
4
5
6
7
8
automatic
Yes
7
8
28
4
16
automatic
No
23
36
36
13
36
Table 5.2: Classification results (overall comparison)

9
5
28

1
2
3
4
5
6
7
8
9
8
automatic
Yes
7
15
14
2
11
5
34
automatic
No
9
19
19
6
19
14
Table 5.3: Classification results (comparison with systems using/not using answer
snippets)
As it can be noticed from the results table, our system performed well regarding
Precision and Recall, the first run being classified 7th among the 36 evaluated runs
as far as F-measure. As far as the structure and coherence are concerned, the results
were also good, placing Run 1 in 4th position among the 36 evaluated runs. Also
worth mentioning is the good performance obtained as far as the overall
responsiveness is concerned, where Run 1 ranked 5th among the 36. When
comparing our approaches separately, in both cases, they did not perform very well
with respect of the non-redundancy criterion, nor the grammaticality one. An
126

interesting thing that is worth mentioning as far as the results obtained are
concerned, is that the use of reformulation patterns, in order to generated sentences
for completing the summaries, has been appropriate, leading to very good rankings
according to the structure/coherence criterion. However, due to the low results
obtained as far as the redundancy and grammaticality criteria are concerned, we
decided to test different methods to overcome these issues.

5.2.2. POST TAC 2008 COMPETITION EXPERIMENTS AND RESULTS


As we have seen from the results obtained in the TAC 2008 Opinion Pilot
competition results, when using the optional snippets, the main problem to address
is to remove redundancy. Many of the text snippets provided by the organizers
repeat the same arguments for the same target, although with slightly different
words. In order to create non-redundant summaries (and in this manner, also be able
to improve on the F-score, as eliminating useless information allows to include
more informative text snippets in the final summary, without surpassing the
imposed 7000 characters limit), we must determine which of the snippets represent
better an idea for the final summary and remove all the other snippets that express
the same thought. To this aim, in Lloret et al. (2009), we performed several
experiments, involving the use of a summarization system (Lloret et al., 2008) and a
textual entailment system (Iftene and Balahur-Dobrescu, 2007).
Several participants in the TAC 2008 edition performed the opinion
summarization task by using generic summarization systems. Their results were
significantly lower as far as F-measure or overall responsiveness was concerned, as
their systems were specifically designed to perform summarization on factual data.
Therefore, we use the summarization system and the textual entailment system prior
to classifying opinion and selecting the highest scored snippets for generating the
final summaries. The generic summarization system employed was the one
described by Lloret et al. (2008). The main idea of this proposal is to score
sentences of a document with regard to the word frequency count (WF)35, which
can be combined with a Textual Entailment (TE) module. The TE system created
by Iftene and Balahur-Dobrescu (2007) contemplates the entailment at different
levels lexical, syntactic and semantic and is based on the tree-edit distance
algorithm (Kouleykov and Magnini, 2006).
Textual entailment (Dagan et al., 2006) is the task of determining whether the
meaning of one text snippet (the hypothesis, H) can be inferred by another one (the
text, T). Systems implementing this task have been competing since 2005 in the

35

This principle states that the more times a word appears in a document, the more relevant the
sentences that contain this word are.

127

Recognising Textual Entailment (RTE) Challenges (RTE) 36 . The following


examples extracted from the development corpus provided by the 3rd RTE
Challenge show a true and false entailment relation between two text snippets:
Pair id=50 (entailment = true)
T: Edison decided to call his invention the Kinetoscope, combining the Greek
root words kineto(movement), and scopos (to view).
H: Edison invented the Kinetoscope.
Pair id=18 (entailment = false)
T: Gastrointestinal bleeding can happen as an adverse effect of non-steroidal
anti-inflammatory drugs such as aspirin or ibuprofen.
H: Aspirin prevents gastrointestinal bleeding.
Although the first approach suggested for opinion summarization obtained much
better results in the evaluation than the second one, we decided to run the generic
summarization system by Lloret et al. (2008) over both approaches, with and
without applying TE, to provide a more extent analysis and conclusions. After
preprocessing the blogs and having all the possible candidate sentences grouped
together, we considered these as the input for the generic summarizer.
The goal of these experiments was to determine whether the techniques used for
a generic summarizer would have a positive influence in selecting the main relevant
information to become part of the final summary. We re-evaluated the summaries
generated by the generic system following the nuggets list provided by the TAC
2008 organization, and counting manually the number of nuggets that were covered
in the summaries. This was a tedious task, but it could not be automatically
performed because of the fact that many of the provided nuggets were not found in
the original blog collection. After the manual matching of nuggets and sentences,
we computed the average Recall, Precision and F-measure (Beta =1) in the same
way as in the TAC 2008 was done, according to the number and weight of the
nuggets that were also covered in the summary. Each nugget had a weight ranging
from 0 to 1 reflecting its importance, and it was counted only once, even though the
information was repeated within the summary. The average for each value was
calculated taking into account the results for all the summaries in each approach.
Table 4 points out the results for all the approaches reported. We have also
considered the results derived from our participation in the TAC 2008 conference

36

The RTE Challenges have been organized between 2005 and 2007 by the Pascal Network of
Excellence and since 2008, by the National Institute for Standards and Technology, within the Text
Analysis Conference (TAC).

128

(OpSum-1 and OpSum-2), in order to analyze whether they have been improved or
not.
System

Recall Precision F-measure

OpSum-1

0.592

0.272

0.357

OpSum-2

0.251

0.141

0.155

WF-1

0.705

0.392

0.486

TE+WF -1

0.684

0.630

0.639

WF -2

0.322

0.234

0.241

TE+WF-2 0.292
0.282
0.262
Table 5.4: Comparison of the results obtained in the competition versus the results
obtained applying the summarization system proposed by Lloret et al. (2008)
From these results it can be stated that the TE module in conjunction with the
WF counts, have been very appropriate in selecting the most important information
of a document. Although it can be thought that applying TE can remove some
meaningful sentences which contained important information, results show the
opposite. It benefits the Precision value, because a shorter summary contains
greater ratio of relevant information. On the other hand, taking into consideration
the F-measure value only, it can be seen that the approach combining TE and WF,
for the sentences in the first approach, significantly improved the best F-measure
result among the participants of TAC 2008, increasing its performance by 20%
(with respect to WF only), and improving by approximately 80% with respect to
our first approach submitted to TAC 2008.
However, a simple generic summarization system like the one we have used
here is not enough to produce opinion oriented summaries, since semantic
coherence given by the grouping of positive and negative opinions is not taken into
account. Therefore, simply applying the summarization system does not yield the
desired results in terms of opinionated content quality. Hence the opinion
classification stage must be added in the same manner as used in the competition
and combined appropriately with a redundancy-removing method.
Motivated by this first set of experiments, in a second approach, we wanted to
test how much of the redundant information would be possible to remove by using a
Textual Entailment system similar to the one proposed by Iftene and BalahurDobrescu (2007), without it affecting the quality of the remaining data. As input for
the TE system, we considered the snippets retrieved from the original blog posts.
We applied the entailment verification on each of the possible pairs, taking in turn
all snippets as Text and Hypothesis with all other snippets as Hypothesis and Text,
129

respectively. Thus, as output, we obtained the list of snippets from which we


eliminated those that are entailed by any of the other snippets.
System

F-Measure

Best system

0.534

Second best system

0.490

OpSum-1 + TE

0.530

OpSum-1
0.357
Table 5.5. Comparison of the best scoring systems in TAC 2008 and the DLSIUAES
teams improved system
Table 5.5 shows that applying TE before generating the final summary leads to
very good results increasing the F-measure by 48.50% with respect to the original
first approach. Moreover, it can be seen form Table 5.5 that our improved approach
would have ranked in the second place among all the participants, regarding Fmeasure, maintaining the linguistic quality level, with which our approach ranked
high in the TAC 2008 Opinion Pilot competition. The main problem with this
approach is the long processing time. We can apply Textual Entailment in the
manner described within the generic summarization system presented, successively
testing the relation as Snippet1 entails Snippet2?, Snippet1+Snippet2 entails
Snippet3? and so on. The problem then becomes the fact that this approach is
random, since different snippets come from different sources, so there is no order
among them. Further on, we have seen that many problems arise from the fact that
extracting information from blogs introduces a lot of noise. In many cases, we had
examples such as:
At 4:00 PM John said Starbucks coffee tastes great
John said Starbucks coffee tastes great, always get one when reading New York
Times.
To the final summary, the important information that should be added is
Starbucks coffee tastes great. Our TE system contains a rule specifying that the
existence or not of a Named Entity in the hypothesis and it not being mentioned in
the text leads to the decision of NO entailment. For the example given, both
snippets are maintained, although they contain the same data. Another issue to be
addressed is the extra information contained in final summaries that is not scored as
nugget. As we have seen from our data, much of this information is also valid and
correctly answers the questions. Therefore, what methods can be employed to give
more weight to some and penalize others automatically?
130

Regarding the grammaticality criteria, once we had a summary generated we


used the module Language Tool 37 as a post-processing step. The errors that we
needed correcting included the number matching between nouns and determiners as
well as among subject and predicate, upper case for sentence start, repeated words
or punctuation marks and lack of punctuation marks. The rules present in the
module and that we switched off, due to the fact that they produced more errors,
were those concerning the limit in the number of consecutive nouns and the need
for an article before a noun (since it always seemed to want to correct Vista for
the Vista a.o.). We evaluated by observing the mistakes that the texts contained,
and counting the number of remaining or introduced errors in the output. The
results obtained can be seen in Table 5.6.
Problem

Rightly corrected Wrongly corrected

Match S-P
90%
10%
Noun-det
75%
25%
Upper case
80%
20%
Repeated words
100%
0%
Repeated .
80%
20%
Spelling mistakes
60%
40%
Unpaired /()
100%
0%
Table 5.6. Grammaticality analysis
The greatest problem encountered was the fact that bigrams are not detected and
agreement is not made in cases in which the noun does not appear exactly after the
determiner. All in all, using this module, the grammaticality of our texts was greatly
improved. In our post-competition experiments, we showed that using a generic
summarization system, we obtain 80% improvement over the results obtained in the
competition, with coherence being maintained by using the same polarity
classification mechanisms. Using redundancy removal with TE, as opposed to our
initial polarity strength based sentence filtering improved the system performance
by almost 50%. Finally, we showed that grammaticality can be checked and
improved using an independent solution given by Language Tool.
Nevertheless, many of the components of this approach still require
improvement. First of all, as we have seen, special methods must be applied in
order to treat the questions, in order to determine additional elements (expected
polarity of answer, source of opinion expected, target of opinion etc.). Additionally,
as in the case of the initial QA task at TAC 2008, real-life systems must be able to
distinguish between factual questions and questions that require an opinionated
37

http://community.languagetool.org/

131

answer, in order to apply specific, appropriate methods to retrieve the correct


answers. This motivated our subsequent research, which first concentrated on
defining a method through which opinion questions could be differentiated from
factual ones and subsequently on proposing and evaluating specific methods to
tackle opinion question answering.

5.2.3. PROPOSAL OF A FRAMEWORK FOR OPINION QUESTION


ANSWERING
Motivated by the experiments in the TAC 2008 competition and the subsequent
efforts to improve the approaches proposed, we set out to determine what the
particularities of opinion questions are, in order to be able to propose adequate
methods to tackle them.
FACT VERSUS OPINION QUESTION CLASSIFICATION
Firstly, we propose an approach towards solving the problem of question
classification in a mixed setting of opinion (multi-perspective) and fact question
answering (Balahur et al., 2009c).
Moreover, we present a method for multi-perspective question answering
based on the use of polarity classification and Textual Entailment. We show why
the classification of questions is important at the time of answer retrieval and
validation and what are the challenges we are faced with in answering opinion
questions. The difficulties within this task are explained and solutions are
proposed for each of the issues encountered. Finally, we evaluate the accuracy in
question classification using the mixed test sets of the Answer Validation
Exercise 2008, the TAC 2008 Opinion Pilot, the question set made on the OpQA
corpus (Stoyanov et al., 2005), and the set of questions we created using our own
annotated corpus EmotiBlog (Boldrini et al., 2009). The multi-perspective
question answering method is evaluated using the OpQA corpus and EmotiBlog.
We show the improvements brought by the approach in retrieving the correct
answers. We discuss the implications of the method proposed in a Question
Answering (QA) System that is able to retrieve answers for both fact and opinion
questions.
Another relevant challenge is to determine the best way to look for the
answers. In order to achieve this, we must first analyze the correct answers that
are annotated in a corpus and check if it is difficult to retrieve them (or which of
the methods and tools should be employed in order to take into account their
peculiarities).
132

Using EmotiBlog, we label each answer adding three elements. The first one
is the source, the second one is target and finally, we annotate the required
polarity. Using these annotations, we will be able to detect the author of the
sentence, the target of the sentiment and also the polarity of the expressed
opinion.
Furthermore, we also have to find an effective method to check if an answer
is correct. This is difficult for general opinion answers, because in most of the
cases answers for the same question can be different but not for this reason
incorrect. In fact, they could express different points of view about a subject. We
could say that each of them is partially correct and none of them is totally
correct.
After having retrieved the answers, answer validation is needed. We
could perform the Answer Validation Exercise (AVE) for objective answers, but
for opinion sentences we do not have at our disposal any similar tool. We would
need for example a consistent collection of paraphrases for opinion but, to the
best of our knowledge, such a collection does not exist.
In order to carry out our experiments, we created a small collection of questions
in addition to the ones made by Stoyanov et al. (2005) for the OpQA corpus. We
decided to use part of this collection because our corpus and the MPQA share the
same topic, the Kyoto Protocol; as a consequence our corpus will probably contain
the answer to the queries and our questions will also have answers in their corpus.
We used 8 questions in English, divided into two groups; the first one is
composed by factoid ones and we consider them as objective queries. They usually
require a single fact as answer and include questions such as: What is the Kyoto
Protocol? or When was the Kyoto Protocol ratified? As we can deduce, they are
asking for a definition and for a date.
In general, factoid queries ask for a name, date, location, time etc. and as a
consequence, the most relevant aspect is that they need a univocal answer.
Moreover, answers to factoid questions can be validated, using different techniques,
such as Textual Entailment, as proposed in the Answer Validation Exercise
(AVE)38.
The second group is composed of opinion questions. They are different from the
ones previously described. They are more complex in nature, as they contain
additional elements, such as required polarity, and the answer is not univocal.
Through their nature, opinion questions require more answers to be retrieved that
are equally correct and valid (i.e. What reasons do people have for liking
MacDonalds?). We will obtain a wide list of answers each of them will be correct.
Furthermore, there are opinion questions that could be interpreted as objective,
but which are actually requiring the analysis of opinionated content in order to be
38

http://nlp.uned.es/clef-qa/ave/

133

answered (e.g. Are the Japanese unanimous in their opinion of Bushs position on
the Kyoto Protocol?). Usually, opinion queries can have more than one answer.
One of the possible answers found in the corpus under the form of Japanese
were unanimous in their opinion about Bushs position on the Kyoto protocol, and
it is equally possible that answering this question requires the analysis of many
opinions on the same subject. They are evaluative or subjective because they
express sentiments, feelings or opinions.
Table 5.7 presents the collection of questions we used in order to carry out our
experiments.
Type
Questions
factoid
What is the Kyoto protocol about?
factoid
When was the Kyoto Protocol adopted?
factoid
Who is the president of Kiko Network?
factoid
What is the Kiko Network?
opinion
What is Bush's opinion about the Kyoto protocol?
opinion
What are people's feelings about Bush's decision?
opinion What is the Japanese reaction to the Kyoto protocol?
opinion
What are peoples opinions about Bush
Table 5.7: The set of questions proposed on the EmotiBlog corpus
As we can see in Table 5.7, we have a total of 8 questions, divided between
factoid and opinion. After having created the collection of questions, we labeled the
answers in our blog corpus using EmotiBlog, the annotation scheme we built for
emotion detection in non-traditional textual genres, as for example blogs or forums.
The granularity of the annotations done using this scheme could be an advantage if
the model is designed in a simple, but effective way. Using EmotiBlog, we aim at
capturing all relevant phenomena for the sentiment analysis task, in order to provide
the opinion mining system the maximum amount of information.
In order to propose appropriate methods to tackle opinion questions, we must
first determine what are the challenges associated, in comparison to factual
questions. Therefore, we first studied the answers to opinion and factual questions
in the OpQA corpus. In the case of factoid questions, we noticed that, in most of the
cases, they are answerable in a straightforward manner, i.e. the information required
is found in the corpus expressed in a direct manner. One problem we found was
related to temporal expressions. For example, for the question When was the
Kyoto Protocol ratified?, there are some answers that contain 1997, but other
answers are last November, or in the course of the year. As a consequence, we
deduce that, in order to correctly interpret the answer, additional context must be

134

used. Such extra details could be for example the date of the document. For the rest
of factoid question we did not detect other relevant obstacles.
Regarding opinion queries, the analysis revealed a series of interesting issues.
The first one is that some opinion questions could be interpreted as factoid. For
example, if we have: Are the Japanese unanimous in their opinion of Bushs
position on the Kyoto Protocol? This query could be answered with yes/no, but the
information we really need is an opinion. Thus, the first thing to do should be to
build up some patterns in order to detect if a question is about an opinion.
Another problem we detected analyzing OpQA answers is the lack of source of
the opinion. If we ask: How is Bushs decision not to ratify the Kyoto Protocol
looked upon by Japan and other US allies? This question is asking for two different
points of view, but its corresponding answers are not clear. We detected for
example: Wed like to see further efforts on the part of the U.S. or dump. As we can
notice, we do not know who is expressing each of these opinions. This is another
issue that should be solved. Therefore, when annotating the answers with
EmotiBlog, we specify the target (the entity on which the opinion is expressed) and
the source of the opinion.
In order to improve this task we could label our collection of questions using the
Expected Answer Type (EAT) and our corpus with Named Entities (NE). On the
one hand, the EAT would solve the problem of understanding each query type, and
on the other hand NE labeling could improve the entity detection.
It is worth mentioning that opinion QA is a challenging and complex task. The
first step we should be able to perform is to discriminate factoid versus opinion
questions and, after having solved this first problem, we should try to find a way to
expand each of our queries.
Finally, we should have a method to automatically verify the correctness of
partial answers.
In order to study the task of fact versus opinion classification, we use on the one
hand the sets provided within the AVE (Answer Validation Exercise) 2007 and
2008 development and test sets (fact questions) and the development and test sets in
the TAC (Text Analysis Conference) 2008 Opinion Pilot, the questions made on the
OpQA corpus (Stoyanov et al., 2005) and the questions we formulated on the
EmotiBlog corpus. Finally, we gathered a total of 334 fact and 79 opinion questions.
The first classification we performed was using the corpus annotations on the Kyoto
protocol blog posts. For each of the questions, we computed the similarity it has
with each of the objective versus subjective phrases annotated in the corpus and
built vectors of the obtained scores. We then classified these vectors using ten-fold
cross validation with SVM SMO.
The results obtained in this first phase are summarized in Table 5.8 (P denotes
the Precision score, R the Recall, Acc. is the accuracy over the classified examples
135

and Kappa measures the confidence of the classification).


Question type

Fact

0.891 0.955

Opinion

0.727 0.506

Acc. Kappa
0.86

0.52

Table 5.8: Fact versus opinion question classification results


As we can notice, opinion questions are more difficult to discriminate based
only on the similarity to the objective/subjective vocabulary.
A second experiment we performed concerned taking into consideration the
interrogation formula as clue for classifying the fact and opinion questions. We
considered, aside from the similarity features employed in the first experiment, the
type of interrogation formula employed. We had a list of interrogation formulas that
were composed of one or two words at the beginning of the questions. They were:
who, what, why, when, how long, how much, where, which, what reason(s), what
motive, are, is does, did, do, were, was and other. The reason for considering also
cases of interrogation formulas composed of two words was that many opinion and
fact questions starting with what or how can only be discriminated on the basis
of the word following these terms (e.g. how long, how much, what reason(s)
etc.). All other formulations were considered as type other. The results obtained
are summarized in Table 5.9.
Question
type

Fact

0.923

0.987

Opinion

0.93

0.671

Acc.

Kappa

0.92

0.73

Table 5.9: Fact versus opinion question classification using the interrogation
formula
As we can notice from Table 5.9, when using the clue of the interrogation
formula, the results improved substantially for the opinion question category
classification. Further on, we performed a third experiment that had as aim to test
the influence of the number of learning examples for each category on the
classification results. As category of opinion questions category was much smaller
than the number of examples for fact questions (79 as opposed to 334), we wanted
to measure the performance in classification when the training sets were balanced
for the two categories. We randomly selected 79 fact questions from the 334 initial
136

ones and repeated the second experiment.


The results are summarized in Table 5.10.
Question
type

Fact

0.908

0.747

Opinion

0.785

0.924

Acc.

Kappa

0.83

0.67

Table 5.10 Fact versus opinion question classification using the interrogation
formula and an equal number of training examples
In this case, we can see that recall significantly drops for the fact category and
increases for the opinion examples. Precision for the fact category remains around
the same values in all classification settings and drops for the opinion category
when using fewer examples of fact questions. At an overall level, however, the
results become more balanced. Therefore, a first conclusion that we can draw is that
we need larger collections of opinion questions in order to better classify opinion
and fact questions in a mixed setting.
Subsequently to this analysis of methods for fact versus opinion question
classification, we annotated the answers corresponding to the set of questions we
proposed, in the EmotiBlog corpus. Table 5.11 presents the questions and the
number of answers annotated using the EmotiBlog annotation scheme.
Number
of answers
What is Bush's opinion about the Kyoto protocol?
4
What are people's feelings about Bush's decision?
1
What is the Japanese reaction to the Kyoto Protocol
3
What are peoples opinions about Bush?
4
Table 5.11 List of opinion questions and the number of corresponding answers
annotated in the corpus
Questions

As we can see in table 5.11, we have labeled different answers for each opinion
questions. Each of these questions is different in nature from the other ones and
requires a distinct approach in order to be answered. For example, if we consider
the last question: What are people's feelings about Bush's decision? The annotated
answers are: I am just disappointed that the US never supported it/ The whole
worlds perturbed with the greenhouse effect; emission gases destroying the earth
and global warming causing terrible climate changes, except, of course President
Bush./ After years of essentially rolling over against the Bush administration on
137

Climate Change, the EU showed surprising spine. / A collection of reasons why we


hate Bush.
Analyzing the answers we labeled in our corpus we can notice different
problems. The first one is that we do not have a general opinion and, as a
consequence, there is no correct or wrong answer. Each of them represents a
different point of view about Bush.
The second one is that questions are written with words that are different from
the ones of the answer, as for example synonyms or paraphrases, causing problems
at the time of answer retrieval. Finally, we have to take into account that blog posts
are written in a informal style and thus, we could be able to contemplate linguistic
phenomena such as sayings or collocations that vary from one language to another.
As we can deduce, opinion QA is a challenging task due to the fact that we add to
the problems of general QA, the difficulties of the opinion mining research area, a
complex field where linguistic phenomena, language features and subjectivity are
involved.
DEFINING THE NECESSITIES OF OPINION QUESTION ANSWERING
SYSTEMS QUESTION ANALYSIS AND RETRIEVAL
Subsequently to defining a method to discriminate among factual and opinionated
questions, we proceeded to studying the necessities of automatic systems that are
able to correctly retrieve answers to opinionated questions (Balahur et al., 2009d).
In order to carry out our evaluation, we employed the EmotiBlog corpus (Boldrini
et al., 2009). It is a collection of blog entries in English, Spanish and Italian.
However, for this research we used the first two languages. We annotated it using
EmotiBlog and we also created a list of 20 questions for each language. Finally, we
produced the Gold Standard, by labeling the corpus with the correct answers
corresponding to the questions.
No

Type

138

Question
What international organization do people criticize for its
policy on carbon emissions?
Cul fue uno de los primeros pases que se preocup por el
problema medioambiental?
What motivates peoples negative opinions on the Kyoto
Protocol?
Cul es el pas con mayor responsabilidad de la
contaminacin mundial segn la opinin pblica?
What country do people praise for not signing the Kyoto

No

Type

10

11

12

13

14

Question
Protocol?
Quin piensa que la reduccin de la contaminacin se
debera apoyar en los consejos de los cientficos?
What is the nation that brings most criticism to the Kyoto
Protocol?
Qu administracin acta totalmente en contra de la lucha
contra el cambio climtico?
What are the reasons for the success of the Kyoto Protocol?
Qu personaje importante est a favor de la colaboracin
del estado en la lucha contra el calentamiento global?
What arguments do people bring for their criticism of media
as far as the Kyoto Protocol is concerned?
A qu polticos americanos culpa la gente por la grave
situacin en la que se encuentra el planeta?
Why do people criticize Richard Branson?
A quin reprocha la gente el fracaso del Protocolo de
Kyoto?
What president is criticized worldwide for his reaction to the
Kyoto Protocol?
Quin acusa a China por provocar el mayor dao al medio
ambiente?
What American politician is thought to have developed bad
environmental policies?
Cmo ven los expertos el futuro?
What American politician has a positive opinion on the Kyoto
protocol?
Cmo se considera el atentado del 11 de septiembre?
What negative opinions do people have on Hilary Benn?
Cul es la opinin sobre EEUU?
Why do Americans praise Al Gores attitude towards the
Kyoto protocol and other environmental issues?
De dnde viene la riqueza de EEUU?
What country disregards the importance of the Kyoto
Protocol?
Por qu la guerra es negativa?
What country is thought to have rejected the Kyoto Protocol
due to corruption?
Por qu Bush se retir del Protocolo de Kyoto?
What alternative environmental friendly resources do people
139

No
15

16

17

Type
F/O

F/O

18

F/O

19

F/O

20

F/O

Question
O
suggest to use instead of gas en the future?
Cul fue la posicin de EEUU sobre el Protocolo de Kyoto?
Is Arnold Schwarzenegger pro or against the reduction of
O
CO2 emissions?
Qu piensa Bush sobre el cambio climtico?
What American politician supports the reduction of CO2
O
emissions?
Qu impresin da Bush?
What improvements are proposed to the Kyoto Protocol?
O
Qu piensa China del calentamiento global?
What is Bush accused of as far as political measures are
O
concerned?
Cul es la opinin de Rusia sobre el Protocolo de Kyoto?
What initiative of an international body is thought to be a
O
good continuation for the Kyoto Protocol?
Qu cree que es necesario hacer Yvo Boer?
Table 5.12: List of question in English and Spanish

We created a set of factoid (F) and opinion (O) queries for English and for
Spanish, presented in Table 5.12. Some of the questions could be defined between
factoid and opinion (F/O) and the system can retrieve multiple answers after having
selected, for example, the polarity of the sentences in the corpus.
The first step in our analysis was to evaluate and compare the generic QA
system of the University of Alicante (Moreda et al., 2008) and the opinion QA
system presented by Balahur et al. (2008), in which Named Entity Recognition with
LingPipe 39 and FreeLing 40 was added, in order to boost the scores of answers
containing NEs of the question Expected Answer Type (EAT).
The open domain QA system of the University of Alicante (Moreda et al., 2008)
deals with factual questions as location, person, organization, date-time and
number in English and also in Spanish. Its architecture comprises three modules:
!

39

Question Analysis: the language object of the study is determined selecting


the language for which more words are found. Further on, the question type
is selected using a collection of regular expressions and the keywords of each
question are obtained with morphological and dependencies analysis
MINIPAR41 for Spanish and Freeling42 for English.

http://alias-i.com/lingpipe/
http://garraf.epsevg.upc.es/freeling/
41
http://www.cs.ualberta.ca/~lindek/minipar.htm
42
http://garraf.epsevg.upc.es/freeling/

40

140

Information Retrieval: the system, originally, relied on the Internet search


engines. However, in order to look for information among the Web Log
collection, a keyword-based document retrieval method has been
implemented to get relevant documents given the question keywords.
Answer Extraction: the potential answers are selected using a NE recognizer
for each retrieved document. LingPipe 43 and Freeling have been used for
English and Spanish respectively. Furthermore, NE of the obtained question
type and question keywords are marked up in the text. Once selected they are
scored and ranked using answer-keywords distances approach. Finally, when
all relevant documents have been explored, the system carries out an answer
clustering process to group all answers that are equal or contained by others
to the most scored one. Figure 5.3 presents the modules of QA system of the
University of Alicante.

Question Analysis
Question

Language

Type

Keywords

Information
ti R
Retrieval
Documentss

Document Retrieval

Answer Extraction
Answer

Language

NE

Scoring

Dist.

Clustering

--

Figure 5.3: The architecture of the QA system of the University of Alicante


INITIAL METHOD FOR OPINION QUESTION ANSWERING
In order to test the performance of an opinion QA system, we used an approach
similar to the one presented by Balahur et al. (2008).
Given an opinion question, we try to determine its polarity, the focus, its
keywords (by eliminating stopwords) and the expected answer type (EAT) (while
also marking the NE appearing in it). After having extracted this information from
the question, we split blog texts into sentences and we also mark NEs. Finally,
sentences in the blogs are sought which have the highest similarity score with the
43

http://alias-i.com/lingpipe/

141

question keywords, whose polarity is the same as the determined question polarity
and which contains a NE of the EAT. As the traditional QA system outputs 50
answers, we also take the 50 most similar sentences and extract the NEs they
contain. In the future, when training examples will be available, we plan to set a
threshold for similarity, thus not limiting the number of output answers, but setting
a border to the similarity score (this is related to the observation made by Stoyanov
et al. (2005) that opinion questions have a highly variable number of answers). The
approach is depicted in Figure 5.4.
Further on, we present the details of our method. In order to extract the topic and
determine the question polarity, we define question patterns. These patterns take
into consideration the interrogation formula and extract the opinion words (nouns,
verbs, adverbs, adjectives and their determiners). The opinion words are then
classified in order to determine the polarity of the question, using the WordNet
Affect emotion lists, the emotion triggers resource (Balahur and Montoyo, 2008), a
list of four attitudes that we built, containing the verbs, nouns, adjectives and
adverbs for the categories of criticism, support, admiration and rejection and a list
of positive and negative opinion words taken from the system by Balahur and
Montoyo (2008).
Questions

Analyze questions
Blog
sentences

NER (Freeling)
EAT patterns

Question
polarity

EAT

Calculate
similarity

Question
focus

Annotated
sentences

Question
keywords

Annotate
sentences
EmotiBlog

Retrieve
sentences

Candidate
sentences

50 top
similar
sentences

Figure 5.4: The opinion QA system


On the other hand, we pre-processed the blog texts in order to prepare the
answer retrieval. Starting from the focus, keywords and topic of the question, we
sought sentences in the blog collection (which was split into sentences and where
Named Entity Recognition was performed using FreeLing) that could constitute
possible answers to the questions, according to their similarity to the latter.
142

The similarity score was computed with Pedersens Text Similarity Package44
(using this software, both the words in the question, as well as the words in the blog
sentences are also stemmed). The condition we subsequently set was that the
polarity of the retrieved snipped be the same as the one of the question.
The polarity was computed using SVM on the trained model for the annotations
in the EmotiBlog corpus, using as features the n-gram similarity (with n ranging
from 1 to 4), as well as overall similarity to the annotated phrases, an approach
similar to Balahur et al. (2009). Moreover, in the case of questions with EAT
PERSON, ORGANIZATION or LOCATION, we required that a Named Entity of
the appropriate type was present in the retrieved snippets and we boosted the score
of the snippets fulfilling these conditions to the score of the highest ranking one. In
case more than 50 snippets were retrieved, we only considered for evaluation the
first 50 in the order of their similarity score, filtered by the polarity and NE
presence criteria (which proved to be a good indicator of the snippets importance
(Balahur et al., 2008).
EVALUATION OF THE INITIAL APPROACH FOR OPINION
QUESTION ANSWERING
Table 5.12 presents the results obtained for English and Table 5.13 for Spanish. We
indicate the id of the question (Q), the question type (T) and the number of answer
of the Gold Standard (A). We present the number of the retrieved questions by the
traditional system (TQA) and by the opinion one (OQA). We take into account the
first 1, 5, 10 and 50 answers.

1
2
3
4
5
6
7
8
9
10

F
O
F
F
O
O
O
F
F
F

Number of answers found


@1
@5
@10
@ 50
TQA OQA TQA OQA TQA OQA TQA OQA
5
0
0
0
2
0
3
4
4
5
0
0
0
1
0
1
0
3
2
1
1
2
1
2
1
2
1
10
1
1
2
1
6
2
10
4
11
0
0
0
0
0
0
0
0
2
0
0
0
0
0
1
0
2
5
0
0
0
0
0
1
0
3
5
1
0
3
1
3
1
5
1
5
0
1
0
2
0
2
1
3
2
1
0
1
0
1
1
2
1
A

44

http://www.d.umn.edu/~tpederse/text-similarity.html

143

Number of answers found


@1
@5
@10
@ 50
TQA OQA TQA OQA TQA OQA TQA OQA
11 O
2
0
1
0
1
0
1
0
1
12 O
3
0
0
0
1
0
1
0
1
13 F
1
0
0
0
0
0
0
0
1
14 F
7
1
0
1
1
1
2
1
2
15 F/O 1
0
0
0
0
0
1
0
1
16 F/O 6
0
1
0
4
0
4
0
4
17 F 10
0
1
0
1
4
1
0
2
18 F/O 1
0
0
0
0
0
0
0
0
19 F/O 27
0
1
0
5
0
6
0
18
20 F/O 4
0
0
0
0
0
0
0
0
Table 5.12: Results of the opinion question answering method for English
Q

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

F
F
F
F
F
F
F
F
O
O
O
O
O
O
O
O
O
O
O

9
13
2
1
3
2
4
1
5
2
5
2
8
25
36
23
50
10
4

144

@1
TQA
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

OQA
0
1
1
0
0
0
0
0
1
0
0
0
1
1
1
0
1
1
1

Number of answers found


@5
@10
TQA
0
2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

OQA
1
3
2
0
0
1
0
0
2
0
1
1
2
2
2
0
5
1
1

TQA
1
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0

OQA
1
6
2
0
0
1
0
0
2
0
2
1
2
4
6
0
6
2
1

@ 50
TQA
1
11
2
1
1
2
4
1
0
0
0
0
0
0
0
0
0
0
0

OQA
3
7
2
0
0
1
0
0
4
0
3
1
4
8
15
0
10
2
1

@1

Number of answers found


@5
@10

@ 50

TQA OQA TQA OQA TQA OQA TQA OQA


20
O
4
0
1
0
1
0
1
0
1
Table 5.13: Results of the opinion question answering method for Spanish

DISCUSSION ON THE PRELIMINARY RESULTS


There are many problems involved when trying to perform mixed fact and
opinion QA. The first can be the ambiguity of the questions e.g. De dnde viene la
riqueza de EEUU?. The answer can be explicitly stated in one of the blog
sentences, or a system might have to infer them from assumptions made by the
bloggers and their comments. Moreover, most of the opinion questions have longer
answers, not just a phrase snippet, but up to 2 or 3 sentences.
As we can observe in Table 5.12, the questions for which the TQA system
performed better were the pure factual ones (1, 3, 4, 8, 10 and 14), although in some
cases (question number 14) the OQA system retrieved more correct answers. At the
same time, opinion queries, although revolving around NEs, were not answered by
the traditional QA system, but were satisfactorily answered by the opinion QA
system (2, 5, 6, 7, 11, 12). Questions 18 and 20 were not correctly answered by any
of the two systems. We believe the reason is that question 18 was ambiguous as far
as polarity of the opinions expressed in the answer snippets (improvement does
not translate to either positive or negative) and question 20 referred to the title
of a project proposal that was not annotated by any of the tools used. Thus, the
OQA system must be added a component for the identification of quotes and titles,
as well as explore a wider range of polarity/opinion scales.
Furthermore, questions 15, 16, 18, 19 and 20 contain both factual as well as
opinion aspects and the OQA system performed better than the TQA, although in
some cases, answers were lost due to the artificial boosting of the queries
containing NEs of the EAT (Expected Answer Type). Therefore, it is obvious that
an extra method for answer ranking should be used, as Answer Validation
techniques using Textual Entailment.
In Table 5.13, the OQA missed some of the answers due to erroneous sentence
splitting, either separating text into two sentences where it was not the case or
concatenating two consecutive sentences; thus missing out on one of two
consecutively annotated answers. Examples are questions number 16 and 17, where
many blog entries enumerated the different arguments in consecutive sentences.
Another source of problems was the fact that we gave a high weight to the presence
145

of the NE of the sought type within the retrieved snippet and in some cases the
name was misspelled in the blog entries, whereas in other NER performed by
FreeLing either attributed the wrong category to an entity, failed to annotate it or
wrongfully annotated words as being NEs. Not of less importance is the question
duality aspect in question 17. Bush is commented in more than 600 sentences;
therefore, when polarity is not specified, it is difficult to correctly rank the answers.
Finally, also the problems of temporal expressions and the co-reference need to be
taken into account.
FINAL PROPOSAL FOR AN OPINION QUESTION ANSWERING
SYSTEM
Summarizing our efforts until this point, subsequent to the research in classifying
questions between factual and opinionated, we created a collection of both factual
and opinion queries in Spanish and English. We labeled the Gold Standard of the
answers in the corpora and subsequently we employed two QA systems, one open
domain, one for opinion questions. Our main objective was to compare the
performances of these two systems and analyze their errors, proposing solutions to
creating an effective QA system for both factoid an opinionated queries. We saw
that, even using specialized resources, the task of QA is still challenging. From our
preliminary analysis, we could see that Opinion QA can benefit from snippet
retrieval at a paragraph level, since in many cases the answers were not simple parts
of sentences, but consisted in two or more consecutive sentences. On the other
hand, we have seen cases in which each of three different consecutive sentences
was a separate answer to a question.
Therefore, our subsequent efforts (Balahur et al., 2010a; Balahur et al., 2010c)
concentrated on research to analyze the improvements that can be brought at the
different stages of the OQA process: question treatment (identification of expected
polarity EPT, expected source ES and expected target ET-), opinion retrieval
(at the level of one and three-sentences long snippets, using topic-related words or
using paraphrases), opinion analysis (using topic detection and anaphora
resolution). This research is motivated by the conclusions drawn by previous
studies (Balahur et al., 2009d). Our purpose is to verify if the inclusion of new
elements and methods - source and target detection (using semantic role labeling
(SRL)), topic detection (using Latent Semantic Analysis) and joint topic-sentiment
analysis (classification of the opinion expressed only in sentences related to the
topic), followed by anaphora resolution (using a system whose performance is not
optimal), affects the results of the system and how. Our contribution to this respect
is the identification of the challenges related to OQA compared to traditional QA.
We propose adding the appropriate methods, tools and resources to resolve the
146

identified challenges. With the purpose of testing the effect of each tool, resource
and technique, we will carry out a separate and a global evaluation. Until this point,
although previous approaches opinion questions have longer answers than factual
ones, the research done in OQA so far has only considered a sentence-level
approach. This also includes our work (Balahur et al., 2009a; Balahur et al, 2009d).
In the following experiments we will thus evaluate the impact of the retrieval at 1
and 3-sentence level and the retrieval based on similarity to query paraphrases
enriched with topic-related words. We believe retrieving longer text could cause
additional problems such as redundancy, co-reference and temporal expressions or
the need to apply contextual information.
Starting from the research in previous works (Balahur et al., 2009a; Balahur et
al., 2009d), our aim is to give a step forward and employ opinion specific methods
focused on improving the performance of our OQA. We perform the retrieval at 1
sentence and 3 sentence-level and also determine the ES and the ET of the
questions, which are fundamental to properly retrieve the correct answer. These two
elements are selected employing SR. The expected answer type (EAT) is
determined using Machine Learning (ML) using Support Vector Machines (SVM),
by taking into account the interrogation formula, the subjectivity of the verb and the
presence of polarity words in the target SR. In the case of expected opinionated
answers, we also compute the EPT by applying OM on the affirmative version of
the question. These experiments are presented in more detail in the experiment
section.
In order to carry out the present research for detecting and solving the
complexities of opinion QA systems, we employed two blog posts corpora:
EmotiBlog (Boldrini et al., 2009a) and the TAC 2008 Opinion Pilot test collection
(part of the Blog06 corpus).
The TAC 2008 Opinion Pilot test collection is composed by documents with the
answers to the opinion questions given on 25 targets. EmotiBlog is a collection of
blog posts in English extracted from the Web. As a consequence, it represents a
genuine example of this textual genre. It consists in a monothematic corpus about
the Kyoto Protocol, annotated with the improved version of EmotiBlog (Boldrini et
al., 2009b). It is well know that Opinion Mining (OM) is a very complex task due to
the high variability of the language we study, thus our objective is to build an
annotation model for an exhaustive detection of subjective speech, which can
capture the most important linguistic phenomena, which give subjectivity to the
text. Additional criteria employed when choosing the elements to be annotated were
effectiveness and noise minimization. Thus, from the first version of the model, the
elements not statistically relevant have been eliminated. The elements that compose
the improved version of the annotation model are presented in Table 5.14.

147

Elements
Obj. speech
Subj. speech

Description
Confidence, comment, source, target.
Confidence, comment, level, emotion, phenomenon,
polarity, source and target.
Adjectives/Adverbs
Confidence, comment, level, emotion, phenomenon,
modifier/not, polarity, source and target.
Verbs/ Names
Confidence, comment, level, emotion, phenomenon,
polarity, mode, source and target.
Anaphora
Confidence, comment, type, source and target.
Capital
letter/
Confidence, comment, level, emotion, phenomenon,
Punctuation
polarity, source and target.
Phenomenon
Confidence, comment, type, collocation, saying,
slang, title, and rhetoric.
Reader/Author
Confidence, comment, level, emotion, phenomenon,
Interpretatipm (obj.)
polarity, source and target.
Emotions
Confidence, comment, accept, anger, anticipation,
anxiety, appreciation, bad, bewilderment, comfort,
compassion
Table 5.14: EmotiBlog structure
The first distinction consists in objective and subjective speech. Subsequently, a
finer-grained annotation is employed for each of the two types of data.
Objective sentences are annotated with source and target (when necessary also
the level of confidence of the annotator and a comment).
The subjective elements can be annotated at a sentence level, but they also have
to be labeled at a word level. EmotiBlog also contains annotations of anaphora at a
cross-document level (to interpret the storyline of the posts) and the sentence type
(simple sentence or title, but also saying or collocation).
Finally, the Reader and the Writer interpretation have to be marked in objective
sentences. This element is employed to mark and interpret correctly an apparent
objective discourse, whose aim is to implicitly express an opinion (e.g. The
camera broke in two days). The first is useful to extract what is the interpretation
of the reader (for example if the writer says The result of their governing was an
increase of 3.4% in the unemployment rate instead of The result of their governing
was a disaster for the unemployment rate) and the second to understand the
background of the reader (i.e.. These criminals are not able to govern instead of
saying the x party is not able to govern) from this sentence the reader can deduce
the political ideas of the writer. The questions whose answers are annotated with
EmotiBlog are the subset of opinion questions in English presented in (Balahur et
al., 2009). The complete list of questions is shown in Table 5.15.
148

Question
Number
2
5
6
7
11
12
15
16
18
19
20

Question
What motivates peoples negative opinions on the Kyoto
Protocol?
What are the reasons for the success of the Kyoto Protocol?
What arguments do people bring for their criticism of media as
far as the Kyoto Protocol is concerned?
Why do people criticize Richard Branson?
What negative opinions do people have on Hilary Benn?
Why do Americans praise Al Gores attitude towards the Kyoto
protocol?
What alternative environmental friendly resources do people
suggest to use instead of gas en the future?
Is Arnold Schwarzenegger pro or against the reduction of CO2
emissions?
What improvements are proposed to the Kyoto Protocol?
What is Bush accused of as far as political measures are
concerned?
What initiative of an international body is thought to be a good
continuation for the Kyoto Protocol?
Table 5.15: Questions over the EmotiBlog corpus

The main difference between the two corpora employed is that Emotiblog is
monothematic, in fact only posts about the Kyoto Protocol compose it, while the
TAC 2008 corpus contains documents on a multitude of subjects. Therefore,
different techniques must be adjusted in order to treat each of them.

THE QUESTION ANALYSIS PHASE


In order to be able to extract the correct answer to opinion questions, different
elements must be considered. As stated by Balahur et al. (2009), we need to
determine both the expected answer type (EAT) of the question as in the case of
factoid ones - as well as new elements such as expected polarity type (EPT).
However, opinions are directional i.e., they suppose the existence of a source and
a target to which they are addressed.
Thus, we introduce two new elements in the question analysis expected source
(ES) and expected target (ET). These two elements are selected by applying SR and
choosing the source as the agent in the sentence and the direct object (patient) as the
target of the opinion. The expected answer type (EAT) (e.g. opinion or other) is
149

determined using Machine Learning (ML) using Support Vector Machine (SVM),
by taking into account the interrogation formula, the subjectivity of the verb and the
presence of polarity words in the target SR. In the case of expected opinionated
answers, we also compute the expected polarity type (EPT) by applying OM on
the affirmative version of the question. An example of such a transformation is:
given the question What are the reasons for the success of the Kyoto Protocol?,
the affirmative version of the question is The reasons for the success of the Kyoto
Protocol are X.
In the answer retrieval stage, we employ four strategies:
1. Using the JIRS (JAVA Information Retrieval System) IR engine (Gmez et
al., 2007) to find relevant snippets. JIRS retrieves passages (of the desired
length), based on searching the question structures (n-grams) instead of the
keywords, and comparing them.
2. Using the Yahoo search engine to retrieve the first 20 documents that are
most related to the query. Subsequently, we apply LSA on the retrieved
documents and extract the words that are most related to the topic. Finally,
we expand the query using words that are very similar to the topic and
retrieve snippets that contain at least one of them and the ET.
3. Generating equivalent expressions for the query, using the DIRT
paraphrase collection (Lin and Pantel, 2001) and retrieving candidate
snippets of length 1 and 3 (length refers to the number of sentences
retrieved) that are similar to each of the new generated queries and contain
the ET. Similarity is computed using the cosine measure. Examples of
alternative queries for People like George Clooney are People adore
George Clooney, People enjoy George Clooney, People prefer
George Clooney.
4. Enriching the equivalent expressions for the query in 3. with the topicrelated words discovered in 2. using LSA.
In order to determine the correct answers from the collection of retrieved
snippets, we must filter only the candidates that have the same polarity as the
question EPT. For polarity detection, we use a combined system employing SVM
ML on unigram and bigram features trained on the NTCIR MOAT 7 data and an
unsupervised lexicon-based system. In order to compute the features for each of the
unigrams and bigrams, we compute the tf-idf scores.
The unsupervised system uses the Opinion Finder lexicon to filter out subjective
sentences that contain more than two subjective words or a subjective word and a
valence shifter (obtained from the General Inquirer resource). Subsequently, it
accounts for the presence of opinionated words from four different lexicons Micro
WordNet (Cerini et al., 2007), WNAffect (Strapparava and Valitutti, 2004),
Emotion Triggers (Balahur and Montoyo, 2008) and General Inquirer (Stone et al.,
150

1966). For the joint topic-polarity analysis, we first employ LSA to determine the
words that are strongly associated to the topic. Consequently, we compute the
polarity of the sentences that contain at least one topic word and the question target.
Finally, answers are filtered using the Semrol system for SR labeling proposed
by Moreda (2008). Subsequently, we filter all snippets that have the required target
and source as agent or patient. Semrol receives as input plain text with information
about grammar, syntax, word senses, Named Entities and constituents of each verb.
The system output is the given text, in which the semantic roles information of each
constituent is marked. Ambiguity is resolved depending on the machine algorithm
employed, which in this case is TIMBL45.

5.2.4. EVALUATION OF THE PROPOSED OPINION QUESTION


ANSWERING FRAMEWORK
We evaluate our approaches on both the EmotiBlog question collection, as well as
the TAC 2008 Opinion Pilot test set. We compare them against the performance of
the system proposed by Balahur et al. (Balahur et al., 2009d) and the best (Copek et
al., 2008) and lowest-scoring (Varma et al., 2008) systems as far as F-measure is
concerned in the TAC 2008 task. For both the TAC 2008 and EmotiBlog sets of
questions, we employ the SR system in SA and determine the ES, ET and EPT.
Subsequently, for each of the two corpora, we retrieve 1-phrase and 3-phrase
snippets. The retrieval of the of the EmotiBlog candidate snippets is done using
query expansion with LSA and filtering according to the ET. Further on, we apply
sentiment analysis (SA) approach and select only the snippets whose polarity is the
same as the determined question EPT. The results are presented in Table 5.16.

Q
No.

No.
A

2
5
6
7
11
12
15

5
11
2
5
2
3
1

Baseline
(Balahur et al.,
2009d)
@ @ @
@
1
5
10 50
0
2
3
4
0
0
0
0
0
0
1
2
0
0
1
3
1
1
1
1
0
1
1
1
0
0
1
1

1 phrase + ET+SA

3 phrases +ET+SA

@
1
1
0
1
1
0
0
0

@
1
1
1
0
0
0
0
1

@
5
2
2
1
1
0
1
0

@
10
3
2
2
1
0
2
1

@
50
4
2
2
3
0
3
1

@
5
2
2
1
2
0
0
1

@
10
3
3
2
2
0
1
1

@
20
4
4
2
4
1
2
1

45

http://ilk.uvt.nl/downloads/pub/papers/Timbl_6.2_Manual.pdf and http://ilk.uvt.nl/timbl/

151

Q
No.

No.
A

16
18
19
20

6
1
27
4

Baseline
(Balahur et al.,
1 phrase + ET+SA
3 phrases +ET+SA
2009d)
1
4
4
4
0
1
1
2
1
2
2
6
0
0
0
0
0
0
0
0
0
0
0
0
1
5
6
18
0
1
1
2
0
1
1
1
0
0
0
0
0
0
1
1
0
0
1
2
Table 5.16: Results for questions over EmotiBlog

The retrieval of the TAC 2008 1-phrase and 3-phrase candidate snippets was
done using JIRS. Subsequently, we performed different evaluations, in order to
assess the impact of using different resources and tools. Since the TAC 2008 had a
limit of the output of 7000 characters, in order to compute a comparable F-measure,
at the end of each processing chain, we only considered the snippets for the 1phrase retrieval and for the 3-phases one until this limit was reached.
1. In the first evaluation, we only apply the sentiment analysis tool and select the
snippets that have the same polarity as the question EPT and the ET is found in
the snippet. (i.e. What motivates peoples negative opinions on the Kyoto
Protocol? The Kyoto Protocol becomes deterrence to economic development
and international cooperation/ Secondly, in terms of administrative aspect, the
Kyoto Protocol is difficult to implement. - same EPT and ET)
We also detected cases of same polarity but no ET, e.g. These attempts mean
annual expenditures of $700 million in tax credits in order to endorse
technologies, $3 billion in developing research and $200 million in settling
technology into developing countries EPT negative but not same ET.
2. In the second evaluation, we add the result of the LSA process to filter out the
snippets from 1., containing the words related to the topic starting from the
retrieval performed by Yahoo, which extracts the first 20 documents about the
topic.
3. In the third evaluation, we filter the results in 2 by applying the Semrol system
and setting the condition that the ET and ES are the agent or the patient of the
snippet.
4. In the fourth evaluation setting, we replaced the set of snippets retrieved using
JIRS with the ones obtained by generating alternative queries using
paraphrases. We subsequently filtered these results based on their polarity (so
that it corresponds to the EPT) and on the condition that the source and target of
the opinion (identified through SRL using Semrol) correspond to the ES and
ET.
5. In the fourth evaluation setting, we replaced the set of snippets retrieved using
JIRS with the ones obtained by generating alternative queries using
152

paraphrases, enriched with the topic words determined using LSA. We


subsequently filtered these results based on their polarity (so that it corresponds
to the EPT) and on the condition that the source and target of the opinion
(identified through SRL using Semrol) correspond to the ES and ET.
System
F-measure
Best TAC
0.534
Worst TAC
0.101
JIRS + SA+ET (1 phrase)
0.377
JIRS + SA+ET (3 phrases)
0.431
JIRS + SA+ET+LSA (1 phrase)
0.489
JIRS + SA+ET+LSA (3 phrases)
0.505
JIRS + SA+ET+LSA+SR (1 phrase)
0. 533
JIRS + SA+ET+LSA+SR (3 phrases)
0.571
PAR+SA+ET+SR(1 phrase)
0.345
PAR+SA+ET+SR(2 phrase)
0.386
PAR_LSA+SA+ET+SR (1 phrase)
0.453
PAR_LSA+SA+ET+SR (3 phrases)
0.434
Table 5.17: Results for the TAC 2008 question set
From the results obtained, we can draw the following conclusions. Firstly, the
hypothesis that OQA requires the retrieval of longer snippets was confirmed by the
improved results, both in the case of EmotiBlog, as well as the TAC 2008 corpus.
Secondly, opinion questions require the joint topic-sentiment analysis; as we can
see from the results, the use of topic-related words in the computing of the affect
influences the results in a positive manner and joint topic-sentiment analysis is
especially useful for the cases of questions asked on a monothematic corpus.
Thirdly, another conclusion that we can draw is that target and source detection is a
relevant step at the time of answer filtering, not only helping in the more accurate
retrieval of answers, but also at placing at the top of the retrieval the relevant
results. Nonetheless, as we can see from the relatively low improvement in the
results, much remains to be done in order to appropriately tackle OQA. As seen in
the results, there are still questions for which no answer is found (e.g. 18). This is
due to the fact that its treatment requires the use of inference techniques that are
presently unavailable (i.e. define terms such as improvement).
The results obtained when using all the components, for the 3-sentence long
snippets significantly improve the results obtained by the best system participating
in the TAC 2008 Opinion Pilot competition (determined using a paired t-test for
statistical significance, with confidence level 5%). Finally, from the analysis of the
errors, we could see that even though some tools are in theory useful and should
153

produce higher improvements such as SR their performance in reality does not


produce drastically higher results. The idea to use paraphrases for query expansion
also proved to decrease the system performance. From preliminary results obtained
using JavaRap46 for co- reference resolution, we also noticed that the performance
of the OQA lowered, although theoretically it should have improved.
With the objective of improving the task of QA in the context of opinion data,
we presented and evaluated different methods and techniques. From the evaluations
performed using different NLP resources and tools, we concluded that joint topicsentiment analysis, as well as the target and source identification, are crucial for the
correct performance of this task. We have also demonstrated that by retrieving
longer answers, the results have improved. We thus showed that opinion QA
requires the development of appropriate strategies at the different stages of the task
(recognition of subjective question, detection of subjective content of the question,
source, and target and retrieving of the required data). Due to the high level of
complexity of the subjective language, further improvements could be obtained by
testing higher-performing tools for co-reference resolution, other (opinion)
paraphrases collections and paraphrasing methods and the employment of external
knowledge sources that refine the semantics of queries. Another important issue to
be solved is the resolution of temporal expression. This issue is tackled in the next
section.

5.2.5. OPINION QUESTION ANSWERING WITH TEMPORAL


RESTRICTIONS PARTICIPATION IN THE NTCIR 8 MOAT
Having analyzed the needs of an opinion question answering system and proposed
adequate solutions for tackling this task, we subsequently aimed at evaluating our
approach in an open competition. This external evaluation of our approaches was
done in the NTCIR 8 MOAT (Multilingual Opinion Analysis Task).
In this competition, the participants were provided with twenty topics. For each
of the topics, a question was given, together with a short and concise query
corresponding to the question, the expected polarity of the answer and the period of
time required. For each of the topics, the participants were given a set of
documents, that were split into sentences (for the opinionated and relevance
judgments) and into opinion units (for the polarity, opinion target and source tasks).
5 different subtasks were defined both in a monolingual, as well as cross-language
setting: judging sentence opinionatedness, relevance, determining the polarity of
opinionated sentence, as well as the source and target of the opinions identified.
The monolingual subtasks were defined for 4 languages: English, Traditional
Chinese, Simplified Chinese and Japanese. In the cross-lingual task, participants
46

http://wing.comp.nus.edu.sg/~qiu/NLPTools/JavaRAP.htm

154

could retrieve answers to the set of questions in English in any of the other 3
languages.
In order to evaluate our approaches to opinion question answering, as well as
evaluate the inclusion of filtering techniques based on topic relevance and temporal
restrictions, we participated in the first 3 subtasks in an English, monolingual
setting, as well as in the cross-lingual challenge, retrieving the answers to the
question set in English in the document set in Traditional Chinese. For the
participation in this competition, we named our opinion question answering system
OpAL.
For the English monolingual subtasks, we submitted three runs of the OpAL
system, for the opinionated, relevance and polarity judgment tasks.
$ TACKLING THE ENGLISH MONOLINGUAL SUBTASKS AT
17&,5MOAT
Judging sentence opinionatedness
The opinionated subtask required systems to assign the values YES or NO (Y/N)
to each of the sentences in the document collection provided. This value is given
depending on whether the sentence contains an opinion (Y) or it does not (N).
In order to judge the opinionatedness of the sentence, we employed two different
approaches (the first one corresponding to system run number 1 and the second to
system runs 2 and 3).
Both approaches are rule-based, but they differ in the resources employed. We
considered as opinionated sentences the ones that contain at least two opinion
words or one opinion word preceded by a modifier. For the first approach, the
opinion words were taken from the General Inquirer, Micro WordNet Opinion and
Opinion Finder lexicon and in the second approach we only used the first two
resources.
Determining sentence relevance
In the sentence relevance judgment task, the systems had to output, for each
sentence in the given collection documents per topic, an assessment on whether or
not the sentence is relevant for the given question. For the sentence relevance
judgement task stage, we employ three strategies (corresponding to the system runs
1, 2 and 3, respectively):
1. Using the JIRS (JAVA Information Retrieval System) IR engine (Gmez et
al., 2007) to find relevant snippets. JIRS retrieves passages (of the desired
length), based on searching the question structures (n-grams) instead of the
keywords, and comparing them.

155

2. Using faceted search in Wikipedia and performing Latent Semantic


Analysis (LSA) to find the words that are most related to the topic. The
idea behind this approach is to find the concepts that are contained in the
query descriptions of the topics. In order to perform this task, we match the
query words, starting from the first, to a category in Wikipedia.
Subsequently we match each group of two consecutive words to the same
categories, then groups of 3, 4, etc. until the highest match is found. The
concepts determined through this process are considered as the topic
components. For each of these topic components, we determine the most
related words, applying LSA is to the first 20 documents that are retrieved
using the Yahoo search engine, given the query. For LSA, we employ the
Infomap NLP47 software.
Finally, we expand query using words that are very similar to the topic
(retrieved through the LSA process) and retrieve snippets that contain at
least two such words.
3. The third approach consists in judging, apart from the topic relevance
characteristic, the temporal appropriateness of the given sentences. In order
to perform this check, we employ TERSEO (Saquete et al., 2006). We then
filter the sentences obtained in the second approach depending on whether
or not the document in which they appear have a date matching the required
time interval or the sentence with the resolved temporal expressions
contains a reference to the required time interval.
Polarity and topic-polarity classification for judging sentence answerness
The polarity judgment task required the system to assign a value of POS, NEG or
NEU (positive, negative or neutral) to each of the sentences in the documents
provided. In order to determine the polarity of the sentences, we passed each
sentence through an opinion mining system employing SVM machine learning over
the NTCIR 7 MOAT corpus, the MPQA corpus and EmotiBlog. Each sentence is
preprocessed using Minipar48. For the system training, the following features were
considered, for each sentence word:
! Part of speech (POS);
! Opinionatedness/intensity - if the word is annotated as opinion word, its
polarity, i.e. 1 and -1 if the word is positive or negative, respectively and 0 if it
is not an opinion word, its intensity (1.2 or 3) and 0 if it is not a subjective
word, its emotion (if it has, none otherwise);

47
48

http://infomap-nlp.sourceforge.net/
http://webdocs.cs.ualberta.ca/~lindek/minipar.htm

156

Syntactic relatedness with other opinion word if it is directly dependent of an


opinion word or modifier (0 or 1), plus the polarity/intensity and emotion of
this word (0 for all the components otherwise).
The difference between the submitted runs consisted in the lexicons used to
determine whether a word was opinionated or not. For the first run, we employed
the General Inquirer, MicroWordNet and the Opinion Finder opinion resources.
For the second one, we employed, aside from these three sources, the emotion
trigger resource (Balahur and Montoyo, 2008).
!

% TACKLING THE ENGLISH-CHINESE CROSS-LINGUAL SUBTASK


$7NTCIR 8 MOAT
In the Cross-lingual setting, the task of the participating systems was to output, for
each of the twenty topics and their corresponding questions (in a language), the list
of sentences containing answers (in another language). For this task, we submitted
three runs of the OpAL system, all of them for the English- Traditional Chinese
cross-lingual setting (i.e. the topics and questions are given in English; the output of
the system contains the sentences in set of documents in Traditional Chinese which
contain an answer to the given topics).
In the following part, we explain the approaches we followed for each of the
system runs. Given that we had no previous experience with processing Chinese
text, the approaches taken were quite simple.
The first step we performed was to tokenize the Chinese texts using LingPipe49.
Further on, we applied a technique known as triangulation to obtain opinion and
subjectivity resources for Chinese. The idea behind this approach is to obtain
resources for different languages, starting from correct parallel resources in 2 initial
languages. The process is exemplified in Figure 5.5 for obtaining resources in
Chinese, starting with resources in English and Spanish.
As mentioned before, this technique requires the existence of two correct
parallel resources in two different languages to obtain correct resources for a third
language. We have previously translated and cleaned the General Inquirer 50 ,
MicroWordNet and Opinion Finder lexicons for Spanish. The emotion triggers
resource is available both for English, as well as for Spanish. In order to obtain
these resources for Traditional Chinese, we use the Google translator.

49
50

http://alias-i.com/lingpipe/
http://www.wjh.harvard.edu/~inquirer/

157

Figure 5.5: Obtaining new resources in Chinese through triangulation


We translate both the English, as well as the Spanish resources, into Traditional
Chinese. Subsequently, we performed the intersection of the obtained translations
that is, the corresponding words that have been translated in the same manner
both from English as well as from Spanish. We removed words that we translated
differently from English and Spanish. The intersection words were considered as
clean (correct) translations. We mapped each of these resources to four classes,
depending on the score they are assigned in the original resource of high
positive, positive, high negative and negative and we give each word a
corresponding value (4, 1, -4 and -1), respectively.
On the other hand, we translated the topic words determined in English using
LSA. For each of the sentence, we compute a score, given by the sum of the values
of the opinion words that are matched in it.
In order for a sentence to be considered as answer to the given question, we set
the additional conditions that it contains at least one topic word and that the polarity
determined corresponds to the required polarity, as given in the topic description.
The three runs differ in the resources that were employed to calculate the
sentiment score: in the first run, we employed the General Inquirer and
MicroWordNet resources; in the second run we added the emotion trigger
resource and the third run used only the Opinion Finder lexicon.
The following tables present the results of the system runs for the three subtasks
in English in which we took part and the cross-lingual English - Traditional Chinese
task.

158

System
RunID
OpAL 1
OpAL 2

17.99
19.44

45.16
44

25.73
26.97

OpAL 3
19.44
44
26.97
Table 5.18: Results of system runs for opinionatedness
System
RunID
OpAL 1
OpAL 2

82.05
82.61

47.83
5.16

60.43
9.71

OpAL 3
76.32
3.94
7.49
Table 5.19: Results of system runs for relevance
System
P
R
F
RunID
OpAL 1
38.13
12.82
19.19
OpAL 2
50.93
12.26
19.76
Table 5.20: Results of system runs for polarity
System
RunID
OpAL 1
OpAL 2

3.54
3.35

56.23
42.75

6.34
5.78

OpAL 3
3.42
72.13
6.32
Table 5.21: Results of system runs for the cross-lingual task agreed measures,
Traditional Chinese
System RunID
OpAL 1
OpAL 2

P
14.62
14.64

R
60.47
49.73

F
21.36
19.57

OpAL 3
15.02
77.68
23.55
Table 5.22: Results of system runs for the cross-lingual task non-agreed
measures, Traditional Chinese
DISCUSSION AND CONCLUSIONS
From the results obtained, on the one hand, we can see that although the extensive
filtering according to the topic and the temporal restrictions increases the system
159

precision, we obtain a dramatic drop in the recall. On the other hand, the use of
simpler methods in the cross-lingual task yielded better results, the OpAL crosslingual run 3 obtaining the highest F score for the non-agreed measures and ranking
second according to the agreed measures.
From the error analysis performed, we realized that, on the one hand, the LSAbased method to determine topic-related words is not enough to perform this task.
The terms obtained by employing this method are correct and useful, but they
should be expanded using language models, to better account for the language
variability.
Finally, we have seen that systems performing finer tasks, such as temporal
expression resolution, are not mature enough to be employed in such tasks. This
was confirmed by in-house experiments using anaphora resolution tools such as
JavaRAP51, whose use also led to lower performances of the system and dramatic
loss in recall.

5.2.6. CONCLUSIONS
In this section, our research was focused on solving a recent problem born with the
massive usage of the Web 2.0: the exponential growth of the opinionated data that
need to be efficiently managed for a wide range of practical applications.
We identified and explored the challenges raised by OQA, as opposed to the
traditional QA. Moreover, we studied the performance of new sentiment-topic
detection methods and analyzed the improvements that can be brought at the
different stages of the OQA process and analyzed the contribution of discourse
analysis, employing techniques such as co-reference resolution and temporality
detection. We also experimented new retrieval techniques such as faceted search
using Wikipedia with LSA, which demonstrate to improve the performance of the
task.
From the results obtained, we can draw the following conclusions. The first one
is that on the one hand, the extensive filtering according to the topic and the
temporal restrictions increases the system precision but it produces a dramatic drop
in the recall. As a consequence, the use of simpler methods in the cross-lingual task
would be more appropriate in this context. The OpAL cross-lingual run 3 obtaining
the highest F score for the non-agreed measures and ranking second according to
the agreed measures. On the other hand, we can deduce that LSA-based method to
determine topic-related words is not enough to perform this task. The terms
obtained by employing this method are correct and useful; however, as future work
our purpose is to use language models, to better account for the language
variability. Finally, we understand that co-reference or temporal resolution systems
51

http://aye.comp.nus.edu.sg/~qiu/NLPTools/JavaRAP.html

160

do not improve the performance of OQA, and as a consequence there is a need to


study the performance of other co-reference and temporal resolution systems in
order to check if the technique is not enough mature or if other systems can bring
added value to this task.

5.3. OPINION-ORIENTED SUMMARIZATION FOR MASS


OPINION ESTIMATION
5.3.1. INTRODUCTION AND BACKGROUND
The research described until this point has been motivated by the participation in
the TAC 2008 Opinion Pilot competition, where systems were supposed to create
summaries from the answers to opinion questions.
Nevertheless, as we outlined in the previous chapter, the need to summarize
opinions is not specific only to answers to opinion questions, but a requirement to
any system that aims at processing opinionated content, in order to support a users
need for information extracted from such type of data. In order to be of real help to
users, the automatic systems that are able to detect opinions must be enhanced with
a summarization component, in order to deal with the remaining issue after opinion
data has been mined and classified: large volume and high redundancy. In blogs, for
example, an opinion mining system aiming at analyzing threads (the sequence of
texts containing the post on a subject and the subsequent comments on it made by
different bloggers"), should analyze its content as far as opinion is concerned.
However, the result obtained after this processing would still contain a large
quantity of redundant information. Therefore, in order for the application to be truly
useful to user, another component is needed that subsequently summarizes the
classes of opinions expressed e.g. arguments pro and against the topic.
In the first part of this chapter, we defined, employed and evaluated different
methods for the summarization of answers to opinion questions that were retrieved
from blogs. However, there are several issues related to evaluating the methods we
proposed in this scenario:
! The evaluation of the summarization process is dependent on the retrieval
and opinion mining stages, so the opinion summarization process cannot be
directly evaluated;
! The snippets that should be contained in the final summary do not have a
polarity associated to them, so it is impossible to evaluate the correctness of
the opinion mining component;
Therefore, the aim of this subsequent research is to study the manner in which
opinion can be summarized, so that the obtained summary can be used in real-life
applications e.g. marketing, decision-making. We discuss the aspects involved in
161

this task and the challenges it implies, in comparison to traditional text


summarization, demonstrating how and why it is different from content-based
summarization. We propose the labeling of a corpus for this task, as well as three
different approaches to perform opinion summarization and test our hypotheses.
Additionally, we compare and evaluate the results of employing opinion mining
versus summarization as a first step in opinion summarization. Subsequently, we
propose an adequate method to tackle the summarization of opinions, in order to
create high-performance systems for real-world applications.

5.3.2. INITIAL EXPERIMENTS


CORPORA
The first corpus we employed in our experiments is a collection of 51 blog threads
extracted from the Web (Balahur et al., 2009a), which was labeled with EmotiBlog.
The structure of the data is presented in Table 5.23. The elements from EmotiBlog
used for the labeling of the data are presented in Table 5.24.
Number of
Number of
Total number of
words
words
words
per news item
per post
Total
1829
72.995
226.573
299.568
Average
33.87
1351.75
4195.79
5547.55
Table 5.23: Structure of the corpus annotated for blog thread summarization
Number of
Posts

Element
Attribute
Polarity
Positive, negative
Level
Low, medium, high
Source
name
Target
name
Table 5.24: Elements from the EmotiBlog scheme used in for the annotation of blog
threads
The blog threads are written in English and have the same structure: the authors
create an initial post containing a piece of news and possibly their opinion on it and
subsequently, bloggers reply, expressing their opinions about the topic (thus
forming a discussion thread). The blog corpus annotation contains the URL from
which the thread was extracted, the initial annotated piece of news and the labeled
user comments. The topics contained in the corpus are very diverse: economy,
science and technology, cooking, society and sports. The data is annotated at a
document level, with the overall polarity and topic, and at sentence level,
162

discriminating between objective and subjective sentences. Subsequently,


subjective sentences are annotated with the polarity of the sentiment expressed and
the intensity of the opinion expressed (low, medium or high). Finally, the source of
the discourse and the target of the sentence are specified. Figure 5.6 contains an
example of annotation. We would like to stress upon the fact that we indicate more
than one topic. We decided to contemplate cases of multiple topics only if they are
relevant in the blog. In this case, the main topic is the economic situation, while the
secondary ones are the government and banks.
<topic>economic situation</topic>
<topic2>government</topic2>
<topic3>banks</topic3>
<news> Saturday, May 9, 2009 My aim in this blog has largely been to
give my best and most rational perspective on the reality of the
economic situation. I have tried (and I hope) mostly succeeded in
avoiding emotive and partisan viewpoints, and have tried as far as
possible to see the actions of politicians as misguided. Of late,
that perspective has been slipping, for the UK, the US and also for
Europe. </news>
<phenomenon gate:gateId="1" target="economic crisis" degree1="medium"
category="phrase" source="Cynicus Economicus" polarity1="negative" >I
think that the key turning point was the Darling budget, in which the
forecasts were so optimistic as to be beyond any rational
belief</phenomenon>
<topic>economic situation</topic>
<topic2>government</topic2>
<topic3>banks</topic3>
<news> Saturday, May 9, 2009 My aim in this blog has largely been to
give my best and most rational perspective on the reality of the
economic situation. I have tried (and I hope) mostly succeeded in
avoiding emotive and partisan viewpoints, and have tried as far as
possible to see the actions of politicians as misguided. Of late,
that perspective has been slipping, for the UK, the US and also for
Europe. </news>
<phenomenon gate:gateId="1" target="economic crisis" degree1="medium"
category="phrase" source="Cynicus Economicus" polarity1="negative" >I
think that the key turning point was the Darling budget, in which the
forecasts were so optimistic as to be beyond any rational
belief</phenomenon>

Figure 5.6:Example of annotation of blog thread using EmotiBlog


The second dataset used in our experiments is the collection of 85 bank reviews
with fine-grained classification used by Saggion and Funk (2010). The aim in
evaluating our approach on this dataset is to verify whether, in the context of
opinion summarization, the first step should be opinion mining or summarization
and how this choice in influences the final output of the system.

163

EXPERIMENTAL SETTING
The main objective of our experiments is to design a system that is able to produce
opinion summaries, in two different types of texts: a) blog threads, in which case
we aim at producing summaries of the positive and negative arguments given on the
thread topic; and b) reviews, in the context of which we assess the best manner to
use opinion summarization in order to determine the overall polarity of the
sentiment expressed. In our first opinion summarization experiments, we adopt a
standard approach by employing in tandem a sentiment classification system and a
text summarizer. The output of the former is used to divide the sentences in the blog
threads into three groups: sentences containing positive sentiment, sentences
containing negative sentiment and neutral or objective sentences. Subsequently, the
positive and the negative sentences are passed on to the summarizer separately to
produce one summary for the positive posts and another one for the negative ones.
Next, we present the sentiment analysis system followed by a description of the
summarization system, both of which serve as a foundation for subsequent sections.
The ideas and results presented in this section were initially put forward in Balahur
et al. (Balahur et al., 2009g).
The Sentiment Analysis System
The first step we took in our approach was to determine the opinionated sentences,
assign each of them a polarity (positive or negative) and a numerical value
corresponding to the polarity strength (the higher the negative score, the more
negative the sentence and vice versa). Given that we are faced with the task of
classifying opinion in a general context, we employed the simple, yet efficient
approach, presented in Balahur et al. (2009).
In the following experiments, we used WordNet Affect (Strapparava and
Valitutti, 2004), SentiWordNet (Esuli and Sebastiani, 2005), MicroWNOp (Cerini
et al., 2007). Each of the resources we employed were mapped to four categories,
which were given different scores: positive (1), negative (-1), high positive (4) and
high negative (4).
As we have shown (Balahur et al., 2009f), these values per formed better than
the usual assignment of only positive (1) and negative (-1) values. First, the score of
each of the blog posts was computed as the sum of the values of the words that
were identified; a positive score leads to the classification of the post as positive,
whereas a negative score leads to the system classifying the post as negative.
Subsequently, we performed sentence splitting using Lingpipe and classified the
sentences we thus obtained according to their polarity, by adding the individual
scores of the affective words identified.

164

The Summarization System


The summarization system we employed in this preliminary analysis, presented by
Balahur et al. (Balahur et al., 2009g) is the one described by Steinberger and Jeek
(2008). It was originally proposed by Gong and Liu (2002) and later improved by
Steinberger and Jeek (2004). The main idea of the approach relies on the use of
Latent Semantic Analysis (LSA) to detect the topic-relevant sentences. This
approach first builds a term-by-sentence matrix from the source, then applies
Singular Value Decomposition (SVD) and finally uses the resulting matrices to
identify and extract the most salient sentences. SVD extends the latent (orthogonal)
dimensions, which in simple terms correspond to the different topics discussed in
the source.
EVALUATION OF THE INITIAL OPINION SUMMARIZATION
PROPOSAL
In order to test the performance of this first approach, we first apply the sentiment
analysis system proposed by Balahur et al. (Balahur et al., 2009f) and subsequently
the text summarizer described by Steinberger and Jeek (2008). We analyze and
discuss the performance of the sentiment recognition, followed by the overall
summarization performance. Performance results of the sentiment analysis are
shown in Table 5.25. We have also analyzed in depth the results of 14 blog threads
in order to assess the quality of the output produced by the opinion mining system,
independently of the topic relevance of the sentences classified. As only sentences
that were relevant to the topic in question were labeled in the Gold Standard, we
also assessed the sentiment of the sentences that were not annotated (see Table
5.26).
System
Sentpos
Sentneg

Precision
0.53
0.67

Recall
0.89
0.22

F1
0.67
0.33

Table 5.25: Results of the evaluation of the sentiment analysis system


As we can observe from the results presented in Table 5.25, the system we
employed had an overall relatively low precision and a high recall (measured as
average of the results obtained for the positive and negative sets of summaries),
meaning that the sentences that were classified as positive or negative by the system
were either erroneously classified (as positive, when they were in fact negative, or
vice-versa) or they were annotated in the Gold Standard as being objective.
However, the high recall suggests that the system, although simple, is capable of
distinguishing subjective sentences from objective ones. We further on analyzed the
165

performance of the sentiment analysis system independently of the topic relevance.


The results are summarized in Table 5.26.

Thread
No.
1
2
3
4
6
8
11
12
14
15
16
18
30
31

Number of sentences
with positive sentiment
OOTotal O-P
N
O
5
0
1
2
5
0
0
1
4
1
2
1
4
2
1
1
5
1
4
0
5
4
0
1
4
3
0
1
4
0
3
1
5
1
3
1
5
2
2
1
4
2
2
0
5
0
3
2
7
3
3
1
5
2
3
0

Number of sentences
with negative sentiment
OOOTotal
P
N
O
4
0
2
0
5
1
2
1
4
0
2
0
4
0
2
1
5
0
4
1
6
0
3
2
6
3
3
0
6
0
3
1
5
0
4
0
5
0
3
2
5
1
2
0
4
1
0
3
4
0
2
0
5
1
3
0

Asserted
Pos

Asserted
Neg

2
1
0
0
0
0
0
0
0
0
0
0
0
0

2
1
2
1
1
1
0
2
1
0
2
0
2
1

Table 5.26: Results of the evaluation of the sentiment analysis classification


performance on 14 topics, independently of the topic relevance
As we can observe from the results presented in Table 5.26, the opinion mining
system performed well as far as polarity classification was concerned. Thus,
although relatively few of the sentences in the summary are also present in the Gold
Standard as far as polarity and sentence importance are concerned, the sentences
were classified well, especially for the negative class (see correlation between
Negative and O-N and Positive and O-P in Table 5.26). The degree of importance
of each latent topic is given by the singular values and the optimal number of
latent topics (i.e., dimensions) r can be re-tuned on training data. Thus, some
improvement can be achieved over these results by adding a topic detection
component.
It can also be noticed that the system has a tendency to over classify sentences as
being negative, a fault which we attribute to the fact that in our approach we do not
contemplate negations and the fact that the resources used contain, in their original
form, word senses, and we do not perform any word sense disambiguation.
Additionally, most of the resources used for sentiment detection have a large
number of negative terms and a significantly lower number of positive ones.
Performance results of the summarizer are shown in Table 5.27. We used the
standard ROUGE evaluation (20) which has also been used for the Text Analysis
166

Conference (TAC) series of competitions. We include the usual ROUGE metrics:


R1 is the maximum number of co-occurring unigrams, R2 is the maximum number
of co-occurring bigrams, RSU4, is the skip bigram measure with the addition of
unigrams as counting unit, and finally, RL is the longest common subsequence
measure. In all cases, we present the average F1 score for the given metric and the
95% confidence intervals within parenthesis. There are three rows in Table 5.27:
the first one (Sent+Summneg) corresponds to the performance of the LSA
summarizer on the negative posts and the second one (Sent+Summpos) presents the
performance of the LSA summarizer on the positive posts. The last line contains the
results obtained by the best-scoring system in the TAC 2008 summarization track.
System

R1

R2

RSU4

RL

Sent + Summneg

0.22 (0.18-0.26)

0.09 (0.06-0.11)

0.09 (0.06-0.11)

0.21 (0.17-0.24)

Sent + Summpos

0.21 (0.17-0.26)

0.05 (0.02-0.09)

0.05 (0.02-0.09)

0.19 (0.16-0.23)

SummTAC08

0.348

0.081

0.12

Table 5.27: Overall sentiment summarization performance


DISCUSSION OF THE INITIAL RESULTS
The first thing to note from Table 5.27 is that the performance on negative posts is
better, though, being within the 95% confidence intervals, the difference cannot be
considered statistically significant. One possible reason for the slightly better
performance on the negative posts is that the sentiment recognition system is more
accurate with negative sentiment than with positive.
The other observation we make is that the results on our corpus are not directly
comparable with those of TAC 2008 for two reasons: Firstly, the data sets are
different and secondly, we used the F1 score to account better for the variation in
size of our model summaries. However, it is worth noting that the LSA summarizer
employing the same method as our LSA summarizer ranked in the top 20%
summarization systems at the TAC 2008 competition. Additionally, the same LSA
method has already been improved upon by incorporating higher level semantic
information such as co-reference (Steinberger et al., 2007), and hence, applying the
same method in our context would also potentially translate in performance
improvement.
In the light of this, we believe the performance results we obtained are
promising. The main problem we encountered was that the LSA-based
summarization method we adopted was originally designed to work with
grammatical sentences from news articles. In our case, however, blog posts are
often composed of ungrammatical sentences and, additionally, a high number of
unusual combinations of characters such as :-), ;), :-( etc. (corresponding to the socalled \emoticons"), which make the blog data much nosier and harder to process
167

than the standard data sets traditionally used for summarization evaluation.
Nevertheless, in our case, the LSA method, being a statistical method, proved to be
quite robust to variations in the input data and, most importantly, to the change of
domain. We used F1 score instead of recall used at TAC, because the lengths of our
model summaries vary from one thread to another.

5.3.3. SUMMARIZATION BASED ON OPINION STRENGTH


INTRODUCTION
Further to our initial experiments (Balahur et al., 2009g), we continued our efforts
to develop adequate techniques for opinion summarization (Kabadjov et al., 2009).
In the latter analysis by Kabadjov et al. (2009), we explored the impact of sentiment
intensity on the summarization performance and, in general, the relationship
between these two concepts. In other words, are comments expressing very
negative or very positive opinions also salient from the point of view of
summarization? Intuitively, sentiment summarization can be different from the
summarization of factual data, as sentences regarded as informative from the factual
point of view may contain little or no sentiment, so, eventually, they are useless
from the sentiment point of view.
The main question we address at this point is: how can one determine, at the
same time, both sentiment, as well as information-relevant sentences? In the light of
these questions, we proposed adequate methodologies and performed experiments
to study the relationship between sentiment intensity and summarization (Kabadjov
et al., 2009). This idea was originally discussed by Balahur et al. (2008).
EXPERIMENTAL SETTING
Our approach follows a simple intuition: when people express very negative or very
positive sentiment, for example, in blogs, they might be also conveying important
and valuable information that is somewhat more salient than other comments. The
sub-area of Natural Language Processing concerned with identifying salient
information in text documents is Text Summarization, hence, we decided to
formalize our intuition in the context of text summarization and make use of
standard methodology from that area. In addition, we cast the above intuition as a
statistical hypothesis test where the null hypothesis we seek to reject is the opposite
of our intuition, that is, the sentiment intensity of salient blog comments is no
different from the sentiment intensity of non-salient comments. In order to carry out
experiments to study in a quantitative manner whether sentiment intensity is a
useful summary indicator, three things are needed: a sentiment analysis system
capable of producing a sentiment intensity score for a given blog comment, a
168

summarization algorithm exploiting this sentiment intensity score and a reference


corpus annotated for both sentiment and salience (i.e., gold standard data). Next, we
describe each of those components and the design of the hypothesis test that we
used in our research (Kabadjov et al., 2009).
For the experiments we used the sentiment analysis system described by Balahur
et al. (Balahur et al., 2009f). Subsequently, we defined a straightforward
summarization algorithm that exploits sentiment intensity in the following manner.
The system should:
1. Rank all comments according to their intensity for a given polarity.
2. Select highest-scoring n comments (until the limit in the number of
sentences given by the compression rate.
At this stage, it is important to point out that positive and negative polarity
comments are treated separately, that is, we produce one summary for all positive
comments and one for all negative comments for a given blog thread.
We ran this algorithm at two commonly used compression rates: 15% and 30%.
That is, we produce two summaries for each polarity for each thread, one by
choosing the top 15% and the other by selecting the top 30% of all comments.
In addition to a standard summarization evaluation, we evaluate the hypothesis
that very positive or very negative comments are good choices to be included in a
summary, by casting the problem as a statistical hypothesis test. Student's t-test. We
define the following setting in order to execute an independent two-sample onetailed t-test of unequal sample sizes and equal variance:
1. Null hypothesis,  
Alternative hypothesis,  
2. Level of significance:
3. t statistic:



   

where

Criterion: Reject the null hypothesis in favor of the alternative hypothesis if


 where   (degrees of freedom) and

169

In the setting considered, n is the number of sample points, 1 is group one and 2
is group two. More specifically, in our case group one is composed of all the
comments annotated as salient in our corpus (i.e., gold summary comments) and
group two is composed of all the comments that were not annotated (i.e., gold nonsummary comments). Furthermore, we further slice the data upon polarity (as
produced by the sentiment analysis tool), so we have two samples (i.e., group one
and group two) for the case of positive comments and two samples for the case of
negative comments. For example, out of all the comments that were assigned a
positive score by the sentiment analysis tool, there are those that were also
annotated as positive by the annotators these constitute group one for the positive
polarity case and those that were not annotated at all these constitute group two for
the positive polarity case. The same thinking applies for the negative polarity case.
EVALUATION OF THE OPINION SUMMARIZATION PROPOSAL
BASED ON SENTIMENT INTENSITY
The performance results of the sentiment analysis are shown in Table 5.28.
System

Precision

Recall

F1

Sentneg

0.98

0.54

0.69

Sentpos

0.07

0.69

0.12

Table 5.28: Performance of the sentiment analysis system


The first thing to note in Table 5.28 is that the sentiment analysis tool is doing a
much better job at identifying negative comments (F1 = 0.69) than positive ones
(F1 = 0.12), the main problem with the latter being a very low precision (P = 0.07).
One possible reason for this is an insufficient number of annotated positive
examples (there were much more negative examples than positive ones in the
corpus). In the next section, we discuss whether this substantial difference in
performance between the negative and positive cases has an impact on the
subsequent analysis. Performance results of the summarizer are shown in Table
5.29.

170

System

R1

R2

RSU4

RL

SISummneg at 15%

0.07

0.03

0.03

0.07

SISummpos at 15%

0.22

0.03

0.03

0.19

SISummneg at 30%

0.17

0.06

0.06

0.16

SISummpos at 30%

0.19

0.03

0.03

0.17

TopSummTAC08

0.111

0.142

System

R1

R2

RSU4

RL

BottomSummTAC08

0.069

0.081

Table 5.29: Evaluation of the sentiment summarization system with ROUGE scores
We used the same evaluation metrics as the ones employed in our previous
efforts (Balahur et al., 2009g).
There are five rows in Table 5.29: the first (SISummneg at 15%) is the
performance of the sentiment-intensity-based summarizer (SISumm) on the
negative posts at 15% compression rate; the second (SISummpos at 15%) presents
the performance of SISumm on the positive posts at 15% compression rate; the
third (SISummneg at 30%) is the performance of the SISumm on the negative posts
at 30% compression rate; the fourth (SISummpos at 30%) presents the performance
of SISumm on the positive posts at 30% compression rate; and finally, the fifth and
the sixth rows correspond to the official scores of the top and bottom performing
summarizers at the 2008 Text Analysis Conference Summarization track (TAC08),
respectively. The last scores are included to provide some context for the other
results. Certainly, in order to use gold polarity alongside the score produced by the
sentiment analysis tool as we do, we had to firstly automatically align all the
automatically identified sentences with the annotated comments. The criterion for
alignment we used was that at least 70% of the words in an automatically identified
sentence are contained in an annotated comment for it to inherit the gold polarity of
that comment (and by virtue of that to be considered a gold summary sentence).

DISCUSSION
From Table 5.29 it is evident that the ROUGE scores obtained are low (at least in
the context of TAC 2008). This suggests that sentiment intensity alone is not a
sufficiently representative feature of the importance of comments for
summarization purposes. Thus, using it in combination with other features that have
proven useful for summarization, such as entities mentioned in a given comment
(Balahur et al., 2010d), certain cue phrases and surface features, or features
capturing the relevance of blog posts to the main topic, is likely to yield better
results. In particular, incorporating topic detection features would be crucial, since
at the moment off-topic, but very negative or very positive, comments are clearly
bad choices for a summary, and currently we employ no means for filtering these
out.
There is also an alternative interpretation of the attained results. These results
were obtained by using a methodology used in text summarization research, so it is
possible that the method is not particularly well-suited for the task at hand, that of
171

producing sentiment-rich summaries. Hence, the reason for the low results may be
that we addressed the problem in the context of a slightly different task, suggesting
that the task of producing content-based summaries and that of producing
sentiment-based summaries are two distinct tasks which require a different
treatment. In addition to the above results, we perform the statistical hypothesis test.
The values of the variables and the resulting t-statistic values are shown in Table
5.30.

-3.95

-4.04

1092

1381

10.13

t statistic

Negative

10.5

0.021

Positive

4.37

4.26

48

1268

9.3

28.03

0.036

Polarity

Table 5.30: Values for the variables and resulting t-statistic for the 2-sample t-test,
unequal sample sizes, equal variances
In both cases, negative and positive polarity, the t values obtained are not large
enough for us to reject the null hypothesis in favor of the alternative hypothesis.
That is, we do not have any empirical evidence to reject the null hypothesis that the
sentiment intensity of salient blog comments is any different from the sentiment
intensity of non-salient comments in favor of our alternative hypothesis that,
indeed, sentiment intensity in summary blog comments is different from that of
non-summary blog comments.
We conclude that, based on our annotated corpus, the hypothesis that very
positive or very negative sentences are also good summary sentences does not hold.
But, once again, we point out that these results are meaningful in the context of text
summarization, that is, the task of producing content-based summaries. Hence, the
observation we made above that producing content based summaries is different
from producing sentiment-based summaries and as such these tasks should be
treated differently also applies in this case. We note, however, that the results on
our corpus are not directly comparable with those of TAC08, since the data sets are
different and the tasks involved are significantly distinct. Blog posts in our corpus
were annotated as important with respect to the main topic of the respective blog
threads.

5.3.4. OPINION SUMMARIZATION USING A TOPIC-SENTIMENT


ANALYSIS APPROACH
INTRODUCTION
Subsequently to the initial efforts (Balahur et al., 2009g; Kabadjov et al., 2009), we
realized that the sentiment analysis component needs to be enhanced with a topic172

detection module, so that the performance of the opinion summarization could


increase. These subsequent efforts to include topic detection and determine the
sentiment of opinions in a topic-dependent manner were described by Balahur et al.
(Balahur et al., 2010e).
As discussed in the preceding sections, the performance of the opinion
summarization, as it was tackled so far (without taking into consideration the topic)
was rather low.
From the human evaluation of the obtained summaries, we could see that the
sentiment analysis system classified the sentences correctly as far as opinion,
polarity and intensity are concerned. However, many topic irrelevant sentences
were introduced in the summaries, leaving aside the relevant ones. On the other
hand, we could notice that in the experiments, taking into consideration the
presence of the opinion target and its co-references and computing the opinion
polarity around the mentions of the target reaches a higher level of performance.
Therefore, it became clear that first of all, a system performing opinion
summarization in blogs must first of all include a topic component. Secondly, the
research done so far in this area has not taken into consideration the use of methods
to detect sentiment that is directly related to the topic. In the experiments we have
performed, we detect sentences where the topic is mentioned, by using Latent
Semantic Analysis (LSA). Thirdly, most summarization systems do not take into
consideration semantic information or include Named Entity variants and coreferences. In our approach, also employed in the TAC 2009 summarization, we
employ these methods and show how we can obtain better results through their use.
EXPERIMENTAL SETTING
In the first stage, we employ the same technique as in the preliminary approach, but
using only the resources that best scored together (MicroWordNet Opinion, JRC
Lists and General Inquirer). We map each of these resources into four classes
(positive, negative, high positive and high negative), and assign each of the words
in the classes a value, i.e., 1, -1, 4 and -4, respectively. We score each of the blog
sentences as sum of the values of the opinion words identified in it. In the second
stage, we first filter out the sentences that are associated to the topic discussed,
using LSA. Further on, we score the sentences identified as relating to the topic of
the blog post, in the same manner as in the previous approach. Subsequently to this,
we propose a method to select the topic-related sentences based on LSA.
The aim of this approach is to select for further processing only the sentences
which contain opinions on the post topic. In order to filter these sentences in, we
first create a small corpus of blog posts on each of the topics included in our
collection. These small corpora (30 posts for each of the five topics) are gathered
using the search on topic words on http://www.blogniscient.com/ and crawling the
173

resulting pages. For each of these 5 corpora, we apply LSA, using the Infomap NLP
Software52. Subsequently, we compute the 100 most associated words with two of
the terms that are most associated with each of the 5 topics and the 100 most
associated words with the topic word. For example, for the term bank, which is
associated to economy, we obtain (the first 20 terms):
bank:1.000000;money:0.799950;pump:0.683452;switched:0.682389; interest:0.674177;
easing:0.661366; authorised:0.660222; coaster:0.656544; roller:0.656544;
maintained:0.656216; projected:0.656026; apf:0.655364; requirements:0.650757;
tbills:0.650515; ordering:0.648081; eligible:0.645723;
ferguson's:0.644950;proportionally:0.63358; integrate:0.625096; rates:0.624235

The summarization system we employed in these experiments is based on the one


described by Steinberger et al. (2009). In this approach, the source representation
(that is, the input matrix A to the LSA system) is enriched with semantic
information combining several source of knowledge, such as lexical information,
information about entities and about hypernymy relationships such as those found
in IS-A taxonomies, such as MeSH. For our experiments, we used the
NewsExplorer multilingual tools for geo-tagging and entity disambiguation
described by Pouliquen et al. (2007) and used them to augment the source entityby-sentence matrix A used in the LSA-based summarizer proposed by J.
Steinberger et al. (2009). In addition, we augmented the matrix with terms
grounded to the Medical Subject Headings (MeSH) taxonomy, created by the
Health-On-the-Net organization. The main idea behind this is to capture more
complex semantic relationships such as hypernymy and synonymy. The Medical
Subject Headings (MeSH) thesaurus is prepared by the US National Library of
Medicine for indexing, cataloguing, and searching for biomedical and health-related
information and documents. Although it was initially meant for biomedical and
health-related documents, since it represent a large IsA taxonomy, it can be used in
more general tasks (http://www.nlm.nih.gov/ mesh/meshhome.html).
EVALUATION RESULTS AND DISCUSSION
For the experimental analysis, we include the usual ROUGE metrics: R1 is the
maximum number of co-occurring unigrams, R2 is the maximum number of cooccurring bigrams, RSU4 is the skip bigram measure with the addition of unigrams
as counting unit, and finally, RL is the longest common subsequence measure. In
the cases of the baseline systems we present the average F1 score for the given
metric and within parenthesis the 95% confidence intervals. There are four rows in
Table 5.31: the first one, Sent + BLSummneg, is the performance of the baseline
52

http://infomap-nlp.sourceforge.net/

174

LSA summarizer on the negative posts (i.e., using only words), the second one,
Sent + Summneg, is the enhanced LSA summarizer exploiting entities and IS-A
relationships as given by the MeSH taxonomy, the third one, Sent + BLSummpos,
presents the performance of the baseline LSA summarizer on the positive posts and
the fourth one, Sent+Summpos, is the enhanced LSA summarizer for the positive
posts.
System

R1
R2
RSU4
RL
0.22
0.09
0.09
0.21
Sent+BLSummneg
(0.18-0.26) (0.06-0.11)
(0.06-0.11)
(0.17-0.24)
Sent+Summneg
0.087
0.087
0.268
0.253
0.21
0.05
0.19
Sent+BLSummneg
0.05 (0.02-0.09)
(0.17-0.26) (0.02-0.09)
(0.16-0.23)
Sent+Summneg
0.076
0.076
0.275
0.249
Table 5.31: Results of the opinion summarization process
Based on Table 5.31 we can say that the results obtained with the enhanced LSA
summarizer are overall better than the baseline summarizer. The numbers in bold
show statistically significant improvement over the baseline system (note they are
outside of the confidence intervals of the baseline system). The one exception
where there is a slight drop in performance of the enhanced summarizer with
respect to the baseline system is in the case of the negative posts for the metrics R2
and RSU4, however, the F1 is still within the confidence intervals of the baseline
system, meaning the difference is not statistically significant.
We note that the main improvement in the performance of the enhanced
summarizer comes from better precision and either no loss or minimal loss in recall
with respect to the baseline system. The improved precision can be attributed, on
one hand, to the incorporation of entities and IS-A relationships, but also, on the
other hand, to the use of a better sentiment analyzer than the one used to produce
the results of the baseline system.
We conclude that by using a combined topic-sentiment approach in opinion
mining and exploiting higher-level semantic information, such as entities and IS-A
relationships, in the summarization process, we obtain a tangible improvement for
the opinion-oriented summarization of blogs.

175

5.3.5. TOPIC-SENTIMENT ANALYSIS VERSUS SUMMARIZATION


AS FIRST STEP IN OPINION SUMMARIZATION
EXPERIMENTAL SETTING
In all our previous approaches, we considered by default that opinion
summarization should be done by first employing an opinion mining system and
subsequently, a summarizer. In this last experiment, we set out to demonstrate that
this order is motivated by the improved quality of the results obtained when
performing opinion mining and subsequently summarization, as opposed to firstly
applying summarization and secondly opinion mining.
In order to compare the results of these two approaches, we employ the opinion
mining system proposed by Balahur et al. (Balahur et al., 2009f) and the
summarization system described by Steinberger et al. (2009), on the corpus of bank
reviews presented by Saggion and Funk (2010) and also employed by Saggion et al.
(2010). Each of the reviews in this set has a number of stars assigned (from 1 to 5, 1
for a very negative opinion and 5 for a very positive one), corresponding to the
positive/negative assessment of the bank by the reviewer. There is a total of 89
reviews (17 one-star, 11 two-star, 9 three-star, 28 four-star and 24 five-star). Since
our opinion summarization system only uses two classes of sentiment (positive and
negative), we exclude from the review set those which are assigned three stars and
perform our experiments on the set of 80 reviews where the topic is assessed
negatively (1 and 2-star reviews) or positively (4 and 5-star reviews).
Further on, we perform four sets of experiments, designed to evaluate which
technique is appropriate for the summarization of opinions, so that in the end, the
overall polarity is present in the result.
In the first three experiments, we employ the opinion mining system as the
initial step in the process; in the last experiment, we first employ the summarization
system and subsequently process the results obtained with the opinion mining
system. In these first three experiments, we initially process the reviews with the
topic-sentiment opinion mining system described above and compute a sentiment
score for each of the sentences in the reviews. Subsequently, in our first experiment,
we computed the score of the individual reviews as sum of the scores assigned to
the sentences it contains (this is referred to as Document level in Table 5.32). In our
second approach (referred to as OM +Top-scoring sentences in 5.32), we selected
the top-scoring 15% of positive and negative sentences in each review and
computed the overall score of the review as sum of the normalized positive score
(total score of the selected positive sentences divided by the number of selected
positive sentences) and the normalized negative score (total score of the selected
negative sentences divided by the number of selected sentences). In the third
approach (OM+Summarizer in Table 5.32), we processed the positive and negative
176

sentences in each review using the summarization system described above and
obtained the 15% most salient sentences in each of these two sets. We then
computed the overall score of the reviews as sum of the normalized positive score
(total score of the selected positive sentences divided by the number of selected
positive sentences) and the normalized negative score (total score of the selected
negative sentences divided by the number of selected sentences).
In the fourth experiment (Summarizer+OM in Table 5.32), we first process each
review with the summarization system and obtain the top 15% most important
sentences. Subsequently, we compute the sentiment score in each of these
sentences, using the opinion mining system.
EVALUATION AND DISCUSSION
The results obtained in these four approaches are presented in Table 5.32. They are
also compared against a random baseline, obtained by an average of 10 random
baselines computed over a set of 80 examples to be classified into positive or
negative (we have 28 negative and 52 positive reviews).
Approach
Document level
OM+Top-scoring sentences
OM+Summarizer
Summarizer+OM
Random baseline

Accuracy
0.62
0.76
0.8
0.66
0.47

Table 5.32: Results of opinion summarization for different approaches


As it can be noticed from the results in Table 5.32, performing opinion mining
as a prior step to summarization results in a better approximation of the sentiment
expressed in the initial text.
The results obtained in the case when summarization is performed prior to
sentiment analysis are comparable to the ones obtained when computing the overall
sentiment at a document level, showing that the result of summarization does offer
a genuine image of what is expressed in the original, but fails to filter in opinionrelated information with a higher priority.
Finally, the results show that by employing only the opinion mining system and
selecting the top-scoring positive and negative sentences, we outperform the
document-level sentiment analysis and summarization as first step approaches, but
obtain lower results than in the case of using opinion mining as a first processing
step, followed by summarization. We believe that the difference is due to the fact
that when scoring the sentiment present in a sentence, its relevance as far as topic is
concerned is only taken into account as filtering factor, and has no influence on the
177

sentence score. This shortcoming can be overcome, as seen in the results, by using
the summarization system, which adds information on the importance of the
sentences as far as the information content is concerned.

5.3.6. CONCLUSIONS ON THE PROPOSED APPROACHES FOR


OPINION SUMMARIZATION
In this second part of this chapter, we presented and evaluated different methods for
opinion summarization in the context of blogs and blog threads. In these
experiments, the aim was to create a robust system that is able to assess mass
opinion on different topics and present the main arguments in favor and against
them.
Within the given setting, we showed that a mere combination of an opinion
mining system with a summarization system is not sufficient to tackle the task.
Further on, we showed that opinion summarization is different from content-based
summarization. Subsequently, we proposed a method to extend the original
approach by integrating topic-opinion analysis and semantic information, achieving
significantly better performance. We used an annotated corpus and the standard
ROUGE scorer to automatically evaluate the performance of our system. Finally,
we assessed the importance of the order in which opinion mining and
summarization are applied to texts, so that the final result of the opinion
summarization process offers an accurate image of the opinions expressed in the
initial document. The different approaches showed that in the case of opinion
summarization, performing the summarization step first can lead to the loss of
information that is vital from the opinion point of view.
All in all, we have shown that performing traditional tasks in the context of
opinionated text has many challenges. In the case of opinion questions, new
elements have to be defined (such as Expected Polarity Type, Expected Source,
Expected Target), in order for the task to be correctly tackled. In the case of opinion
summarization we have shown that the sentiment analysis system must be
employed prior to the summarization system and that the sentiment analysis
component must be enhanced with topic-detection mechanisms.
Finally, we have shown that in the case of opinionated text, relevance is given
not only by the information contained, but also by the polarity of the opinion and its
intensity. Although initial results have shown that there is no correlation between
the Gold Standard annotations and the intensity level of sentences, as output by the
sentiment analysis system, given the fact that using this method, we obtained high
results as far as F-measure is concerned in TAC 2008, we believe that more
mechanisms for opinion intensity should be studied, so that the clear connection
between sentence relevance and the opinion it contains, as well as the intensity it
has, can be established.
178

CHAPTER 6. DISCOVERING IMPLICIT


EXPRESSIONS OF SENTIMENT FROM TEXT
Motto: There are moments in life, when the heart is so full of emotion/ That if by
chance it be shaken, or into its depths like a pebble/Drops some careless word, it
overflows, and its secret,/ Spilt on the ground like water, can never be gathered
together (Henry Longfellow)

6.1. INTRODUCTION
In the previous chapters, we explored the task of sentiment analysis in different text
types and languages, proposing a variety of methods that were appropriate for
tackling the issues in each particular text type. Most of the times, however, the
approaches we took were limited to discovering only the situations where sentiment
was expressed explicitly (i.e. where linguistic cues could be found in the text to
indicate it contained subjective elements or sentiment).
Nevertheless, in many cases, the emotion underlying the sentiment is not
explicitly present in text, but is inferable based on commonsense knowledge (i.e.
emotion is not explicitly, but implicitly expressed by the author, by presenting
situations which most people, based on commonsense knowledge, associate with an
emotion). In this final chapter, we will present our contribution to the issue of
automatically detecting emotion expressed in text in an implicit manner.
Firstly, we present our initial approach, which is based on the idea that emotion
is triggered by specific concepts, according to their relevance, seen in relation to the
basic needs and motivations (Maslow, 1943; Max-Neef 1990). This idea is based on
the Relevance Theory (Sperber and Wilson, 2000). Subsequently, based on the
Appraisal Theory models (De Rivera, 1977; Frijda, 1986; Ortony, Clore and
Collins, 1988; Johnson-Laird and Oatley, 1989), we abstract on our initial idea and
set up a framework for representing situations described in text as chains of actions
and their appraisal values, in the form of a knowledge base. We show the manner in
which additional knowledge on the properties of the concepts involved in such
situations can be imported from external sources and how such a representation is
useful to obtain an accurate label of the emotion expressed in text, without any
linguistic clue being present therein.

179

6.1.1. MOTIVATION
Remembering the definition we provided in Chapter 2, sentiment 53 suggests a
settled opinion reflective of ones feelings -the conscious subjective
experience of emotion (Van den Bos, 2006). Thus, sentiments cannot be present
without an emotion being expressed in text, either implicitly, or explicitly. Due to
this reason, detecting implicit expressions of emotion can increase the performance
of sentiment analysis systems, making them able to spot sentiment even in the cases
where it is not directly stated in text, but results as a consequence of the readers
emotion, as a consequence of interpreting what is said.
Detecting emotion is a more difficult task than the mere sentiment analysis from
text, as the task includes classification between a larger number of categories (i.e
emotion labels), which are not as easily separable or distinguishable as the
positive and negative classes, because of their number (at least 6 basic
emotions54) and characteristics. Although emotion detection is a related problem to
sentiment analysis, we chose to present the work done in this area separately, as we
consider that the approaches in the first problem are more difficult and require
specific methods and tools to be tackled.
This chapter is structured as follows: we first give a brief introduction on the
concepts of emotion and the background of the work presented. Subsequently, we
describe and evaluate the emotion trigger method, put forward by Balahur and
Montoyo (2008), in which the main idea is to create a collection of terms that
invoke an emotion based on their relevance to human needs and motivations.
Finally, we present EmotiNet the framework we built for the detection of emotion
implicitly expressed in text (Balahur et al, 2011a; Balahur et al., 2011b). The
underlying mechanism of this framework is the EmotiNet knowledge base, which
was built on the idea of situation appraisal using commonsense knowledge.

6.1.2. THE CONCEPT OF EMOTION


For the readers convenience, we will first repeat some of the definitions given in
Chapter 2 to the term emotion and the most related concepts.
Emotion is commonly defined as an episode of interrelated, synchronized
changes in the states of all or most of the five organismic subsystems (Information
processing, Support, Executive, Action, Monitor) in response to the evaluation of an
53

http://www.merriam-webster.com/
One of the most widely used classifications of emotions is that of Paul Ekman (1972), which
includes 6 basic emotions. Other models, such as the ones proposed by Parrot (2001) or Plutchik
(2001) include a higher number of basic emotions, as well as secondary and tertiary emotions.
54

180

external or internal stimulus event as relevant to major concerns of the organism


(Scherer, 1987; Scherer, 2001).
The term feeling points to a single component denoting the subjective
experience process (Scherer, 2005) and is therefore only a small part of an emotion.
Moods are less specific, less intense affective phenomena, product of two
dimensions - energy and tension (Thayer, 2001).
Sentiment is the conscious subjective experience of emotion (Van den Bos,
2006).

6.1.3. BACKGROUND
Understanding the manner in which humans express, sense and react to emotion has
always been a challenge, each period and civilization giving a distinct explanation
and interpretation to the diversity of sentiments (Oatley, 2004). Societies used
emotions for the definition of social norms, for the detection of anomalies, and even
for the explanation of mythical or historical facts (e.g. the anger and wrath of the
Greek Gods, the fear of the unknown and the Inquisition in Middle Age, the
romantic love in Modern Times) (Ratner, 2000). The cultural representations
concerning emotions are generally ambivalent (Goldie, 2000, Evans, 2001, Oatley
et al., 2006). Emotions were praised for their persuasive power (Aristotle defines
pathos the ability to appeal to the audiences emotions as the second component
of the art of rhetorics), but also criticized as a weakness of the human being,
which should ideally be rational.
Different scientific theories of emotion have been developed along the last
century of research in philosophy, psychology, cognitive sciences or neuroscience,
each trying to offer an explanation to the diversity of affect phenomena. There were
different attempts to build systems that automatically detect emotion from text in
the 70s and 80s. However, it was not until 1995, when Rosalind Picard consecrated
the term affect computing in Artificial Intelligence (Picard, 1995) that the interest
computer engineers expressed towards the research in emotion increased
significantly. The need to develop systems that are able to detect and respond to
affect in an automatic manner has become even more obvious in the past two
decades, when a multitude of environments of interaction between humans and
computers has been built e.g. e-learning sites, social media applications and
intelligent robots. On the one hand, if such environments are able to detect emotion,
they can better adapt to the user needs. On the other hand, if they are able to express
emotion, they create a more natural type of interaction. Despite the fact that Picard
(1995) identified three different types of systems dealing with automatic affect
processing (systems detecting emotions, systems expressing what a human would
perceive as emotion and systems feeling an emotion), most of the research in affect
181

computing has so far concentrated solely on the first type of systems (Calvo and
DMello, 2010).
In Natural Language Processing (NLP), the task of detecting emotion expressed
in text has grown in importance in the last decade, together with the development of
the Web technologies supporting social interaction. Although different approaches
to tackle the issue of emotion detection in text have been proposed by NLP
researchers, the complexity of the emotional phenomena and the fact that
approaches most of the times contemplate only the word level have led to a low
performance of the systems implementing this task - e.g. the ones participating in
the SemEval 2007 Task No. 14 (Strapparava and Mihalcea, 2007). The first
explanation for these results, supported by linguistic studies and psychological
models of emotion, is that expressions of emotion are most of the times not direct,
through the use of specific words (e.g. I am angry.). In fact, according to a
linguistic study by Pennebaker et al (2003), only 4% of words carry an affective
content. Most of the times, the affect expressed in text results from the
interpretation of the situation presented therein (Balahur and Montoyo, 2008;
Balahur and Steinberger, 2009), from the properties of the concepts involved and
how they are related within the text. In this sense, the first experiments we
performed aimed at building a lexicon of terms whose presence in text trigger
emotion. Subsequently, we described a framework for detecting and linking
concepts (and not just words) that are used to implicitly express emotion in a text.

6.2. EMOTION DETECTION BASED ON EMOTION


TRIGGERS
6.2.1. DEFINITION OF EMOTION TRIGGERS
Most of the systems that we have presented so far base their analysis on the spotting
of linguistic cues of sentiment (e.g. finding words such as good, bad, happy,
etc.). Nevertheless, there are many examples in which emotions (and hence
sentiments) are expressed and elicited in an indirect manner.
In the example Government approves new taxes for car imports, one reader
will remain neutral, another will be infuriated, a car seller might be glad. Or in the
case of a news title such as The Spanish Civil War now a computer game, a
Spanish person could feel offended, another outraged, another amused, a person of
another nationality might feel neutral and a computer games addict very happy. On
the other hand, a title such as Children killed in bomb attack will most certainly
produce a general feeling of sadness and/or fear to the readers. While in this latter
example we can find different words that we can link to sentiment (e.g. killed,
bomb, attack), the first two examples are objective in nature.
182

In order to determine the emotion elicited by a text that is apparently objective in


its statements, in (Balahur and Montoyo, AISB 2008) we introduced the concept of
emotion triggers. They are defined as a words or concepts expressing an idea,
that depending on the readers world of interest, cultural, educational and social
factors, leads to an emotional interpretation of the text content or not.
Examples of emotion triggers are freedom, salary, employment, sale,
pride, esteem, family and so on.
In our initial efforts (Balahur and Montoyo, 2008), we described the process of
building a lexicon of such emotion triggers, classifying them according to their
polarity and integrating them in a system which spots and classifies the polarity of
the sentiment and the emotion expressed therein.

6.2.2. THEORIES UNDERLYING THE EMOTION TRIGGERS


APPROACH
In order to build the lexicon of emotion triggers, we were inspired by three different
theories:
1. The Relevance Theory (Sperber and Wilson, 2000) from Pragmatics.
2. Abraham Maslows theory of human motivation and its corresponding
pyramid of human needs.
3. Max-Neefs matrix of fundamental human needs.
THE THEORY OF RELEVANCE
The Theory of Relevance, arises from pragmatics, and states in the cognitive
principle that human cognition tends to be geared toward the maximization of
relevance, that is, from the multiple stimuli present in a communication, be it
written or spoken, a reader will choose the one with highest significance to their
world of interest.
In the case of emotions, motivated by the work done in automatic argumentation
analysis (Mazzotta et al., 2008), we considered that the basic stimuli for emotions
are related to the basic needs and motivations. Therefore, the core of our lexicon is
represented by the terms included in Maslows Pyramid of Human Needs and
Motivations and Max-Neefs matrix of fundamental human needs. We explain in
detail the components of these two representations.
MASLOWS PYRAMID OF HUMAN NEEDS AND MOTIVATIONS
Abraham Maslow (1943), classified the human needs and motivational factors into
a 5-level pyramid, from the basic, physiological ones, to the more education and
personal level of development dependent ones. Needs such as food, shelter, peace
are at the bottom of the pyramid, whereas needs for self achievement, fame, glory
183

are at the top. The basic needs are the general human ones; as we move towards the
top, we find the more individual dependent ones.
5. Self
actualization

4. Self-esteem,
confidence

3. Friendship, family,
intimacy

2. Security of body, employment,


resources

1. Survival, food, water, sleep, sex,


homeostasis

Figure 6.1. Maslows pyramid of human needs and motivations


We consider the terms in Maslows pyramid levels as primary emotion triggers,
very general notions that express ideas that are fundamental to all human beings. In
order to exploit this classification, we build a lexical database of emotion triggers at
the 5 levels. The words found in the five levels are nouns, verbs, adjectives and
adverbs.
MAX-NEEFS MATRIX OF FUNDAMENTAL HUMAN NEEDS
Among the critics of the Maslow theory of human needs is Manfred Max Neef, in
(Max-Neef, 1991). Human needs, according to Neef, are understood as a system i.e. they are interrelated and interactive. Max-Neef classifies the fundamental
human needs as: subsistence, protection, affection, understanding, participation,
recreation (in the sense of leisure, time to reflect, or idleness), creation, identity and
freedom. Needs are also defined according to the existential categories of being,
having, doing and interacting, and from these dimensions, a 36 cell matrix is
developed which can be filled with examples of satisfiers for those needs55.
Therefore, starting in parallel from the matrix of fundamental human needs as
primary emotion triggers, we see, on one hand, if a classification of emotion
triggers is better than a flat model with rules of inference, and on the other hand,
can build a fine-grained taxonomy of terms indicating the precise category in which
each type of emotion trigger influences the human affect.

55

http://www.rainforestinfo.org.au/background/maxneef.htm

184

6.2.3. BUILDING THE EMOTION TRIGGERS LEXICON


BUILDING THE LEXICON OF EMOTION TRIGGERS
The core of English emotion triggers is built, at the first stage, of the approximately
37 terms found in Maslows pyramid of human needs, structured on 5 levels
starting from the terms corresponding to the deficiency needs, found on the four
bottom levels and having on top the growth needs terms, of achieving the personal
potential, on level 5.
Since most of the words are general notions and their number is relatively small
(37), we disambiguate them with the sense numbers they have in WordNet 2.1, in
order to ensure that further on, the added words will remain with the intended
meaning. For each term, we add all the senses and all grammatical categories that
are valid in the context of Maslows pyramid levels. We then add to these words the
corresponding synonyms and hyponyms from WordNet. For the verbs considered,
we also add the entailed actions. We consider as having a negative value the
emotion triggers that are antonyms of the nouns found. For each of the nouns and
verbs, we further add the corresponding nouns and verbs, respectively, using
NomLex (Macleoud, 1998). Since NomLex does not assign sense numbers to
distinguish between the possible semantics of the nouns and verbs in the collection,
we use the Relevant Domains concept and corresponding repository (Vzquez et
al., 2007) to preserve the intended meaning, by taking the top relevant domain of
each word sense and assigning the corresponding verb or noun in NomLex the
sense number that has the same top relevant domain. If more such senses exist, they
are all added.
On the other hand, another core of English words is completed with the terms
found in Max Neefs matrix of fundamental human needs. This matrix is built
according to the four main characteristics of the individual: being, having, doing
and interacting, for which terms are assigned in order to nine categories of needs:
identity, subsistence, affection, creation, protection, freedom, participation, leisure
and understanding.
When building the core of words corresponding to the taxonomy proposed by
Neef, we start with the terms semantically disambiguated according to the WordNet
2.1 sense numbers. As in the case of the concepts extracted from Maslows 5-level
pyramid, we then add to these words the corresponding synonyms and hyponyms
from WordNet. For the verbs considered, we also add the entailed actions. We
consider as having a negative value the emotion triggers that are antonyms of the
nouns found. For each of the nouns and verbs, we further add the corresponding
nouns and verbs, respectively, using NomLex. Since NomLex does not assign sense
numbers to distinguish between the possible semantics of the nouns and verbs in the
collection, we use the Relevant domain concept and corresponding repository to
185

preserve the intended meaning, by taking the top relevant domain of each word
sense and assigning the corresponding verb or noun in NomLex the sense number
that has the same top relevant domain. If more such senses exist, they are all added.
Using EuroWordNet 56 , we map the words in the English lexical database of
emotion triggers to their Spanish correspondents, preserving the meaning through
the WordNet sense numbers.
The final step in building the lexical databases consists of adding real-world
situations, cultural-dependent contexts terms to the two lexical databases. For
English, we add the concepts in ConceptNet57 that are linked to the emotion triggers
contained so far in the lexicon based on the relations DefinedAs, LocationOf,
CapableOf, PropertyOf and UsedFor. For Spanish, we add the cultural context by
using the Larousse Ideologic Dictionary of the Spanish Language.

6.2.4. ASSIGNMENT OF POLARITY AND EMOTION TO THE


EMOTION TRIGGERS
The next step consists in assigning polarity and emotion to the terms in the
database. This is done with the following rules, both for the terms in Maslows
pyramid as well as for those in Neefs matrix:
1. The primary emotion triggers are assigned a positive value.
2. The terms (also emotion triggers in the final lexical database) synonyms
and hyponyms of the primary emotion triggers, as well as the entailed verbs
are assigned a positive value.
3. The terms opposed and antonym of those from 1. and 2. are assigned a
negative valence.
4. Emotion triggers added further on inherit the valence from the emotion
trigger they are related to in case of synonyms, hyponyms and entailment
and change their valence from positive to negative or negative to positive in
the case of antonyms.
5. Value of all emotion triggers is modified according to the valence shifters
they are determined by.
Further on, we assign an emotion triggers a value on each of the 6 categories of
emotion proposed for classification in the SemEval Task No. 14 joy, sadness,
anger, fear, disgust and surprise, using the following rules:
1. The emotion triggers found in the levels of Maslows pyramid of needs and
those found in the components of Neefs matrix of fundamental human
needs are manually annotated with scores for each of the 6 categories.
56

http://en.wikipedia.org/wiki/EuroWordNet
http://web.media.mit.edu/~hugo/conceptnet/

57

186

2. The primary emotion triggers are assigned a positive value.


3. The terms (also emotion triggers in the final lexical database) synonym and
hyponym of the primary emotion triggers, as well as the entailed verbs are
assigned a positive value.
4. The terms opposed and antonym of those from 1. and 2. are assigned a
negative valence.
5. Emotion triggers added further on inherit the valence from the emotion
trigger they are related to in case of synonyms, hyponyms and entailment
and change their valence from positive to negative or negative to positive
in the case of antonyms with opposed values.
6. Value of all emotions of an emotion triggers is modified according to the
valence shifters they are determined by.
7. If any of the values calculated in 6 is higher than 100, it is set to 100; if it is
lower than -100, it is set to -100.

6.2.5. A METHOD FOR EMOTION DETECTION IN TEXT BASED ON


EMOTION TRIGGERS
In order to be able to recognize the change in meaning of emotion triggers due to
modifiers, we have defined a set of valence shifters words that negate the emotion
triggers, intensify or diminish their sense. The set contains:
! Words that introduce negation (no, never, not, doesnt, dont and negated
modal verbs)
! A set of adjectives that intensify the meaning of the nouns they modify
big, more, better etc.
! A set of adjectives that diminish the meaning of the nouns they modify
small, less, worse etc.
! The set of modal verbs and conditional of modal verbs that introduce
uncertainty to the active verb they determine- can, could, might, should,
would
! The set of modal verbs that stress on the meaning of the verb they
determine - must
! A set of adverbs that stress the overall valence and intensify emotion of the
context surely, definitely etc
! A set of adverbs that shift the valence and diminish emotion of the context
maybe, possibly etc
For each of the valence shifters, we define a weight of 1.5 for the meaning
intensifiers and 0.5 for the meaning diminishers. These are coefficients that will be
multiplied with the weight assigned to the emotion trigger level and emotions- level
association ratio corresponding to the given emotion trigger in the case of emotion
187

triggers built from Maslows pyramid. In the case of emotion triggers stemming
from Neef`s matrix of fundamental human needs, the weights of the valence shifters
are multiplied with the emotion-category association ratio, computed for each
emotion trigger and each of the four existential categories. In order to determine
the importance of the concepts to a specific domain, we will employ the association
ratio formula.
The association ratio score provides a significance score information of the most
relevant and common domain of a word. The formula for calculating it is:
AR ( w; D) " Pr(w, D) log 2

Pr(w, D) , where:
Pr(w) Pr(D)

! Pr(w,D) is the probability of the word in the given domain


! Pr(w) is the probability of the word
! Pr(D) is the probability of the domain
In our approach, besides quantifying the importance of each emotion trigger in a
manner appropriate to the level and emotion it conveys, we propose to use a variant
of the association ratio that we call emotion association level. This score will
provide the significance information of the most relevant emotion to each level. The
corresponding formula is therefore:
Pr(e, L) , where:
AR (e; L) " Pr(e, L) log 2
Pr(e) Pr(L)

!
!
!

Pr(e,L) is the probability of the emotion in the given level


Pr(e) is the probability of the emotion
Pr(L) is the probability of the level

The Construction-Integration Model is a psychological model of text


comprehension (Kintsch, 1999), based on the idea that while reading a text, a
person will activate the features of words that are appropriate to the context and
inhibit those that are not.
In this model, the process of text comprehension consists of two phases. The
first one the construction - uses rules in the form a production system to generate,
from the linguistic representation of words, a propositional network of related
mental elements. Further, adding the knowledge experience of the reader, a more
elaborated propositional network is created.
The second phase integration- takes as input the crude representation of text in
the form of the elaborated propositional network, with nodes linked with positive
and negative connections meant to represent the relations between them, and tunes
it using connectionist relaxation techniques.

188

6.2.6. IMPLEMENTATION OF THE EMOTION DETECTION METHOD


BASED ON EMOTION TRIGGERS
The final system built to classify text according to the polarity of the sentiment
expressed and the emotion it contained is depicted in Figure 6.2:
TEXT
Parser
Emotion triggers
and modifiers
identification
Valence calculation
formula
Valence

CI module

Emotion

Figure 6.2. System valence and emotion classification-components


First, the input text is parsed with Minipar to obtain for each word the
grammatical category, the lemma and its modifiers. Further on, the emotion triggers
in the text are identified, together with their corresponding modifiers.
We calculate the valence of the text on the basis of the identified emotion
triggers and their modifiers, using the formulas described in what follows.
In the case of emotion triggers obtained from Maslows pyramid, we calculate a
score called weighted valence of emotion trigger (wv) using the following formula:
wv(et ij ) " w(m) * w(l j ) * v(et i ) , where

! w(m) is the weight of modifier


! w(lj )is the weight of level
! v(eti)is the emotion trigger valence
! i is the index of the emotion trigger
! j is the number of the level
In the case of emotion triggers obtained from Neefs matrix, we calculate a score
called weighted valence of emotion trigger (wv) using the following formula:
wv(et i ) " w(m) * v(et i ) , where

!
!
!

w(m) is the weight of modifier


v(eti)is the emotion trigger valence
i is the index of the level
189

The total valence of text equals the sum of all weighted valences of all emotion
triggers. The obtained value is rounded to the closest of the two possible values : 0
and 1.
Further on, we calculate the emotions present in the text, by the following
method:
! for each emotion trigger stemming from Maslows pyramid, we compute the
emotion to level association ratio
! for each emotion trigger stemming from Neefs matrix, we the emotion to
category association ratio
We then apply the Construction Integration Model in a manner similar to that
described by Lemaire (2005) and construct a spreading activation network. We
consider the working memory as being composed of the set of emotion triggers and
their emotion association ratio value which is considered as activation value. The
semantic memory is set up of the modifiers and the top 5 synonyms and antonyms
of emotion triggers with their AR value. We set the value of each emotion trigger to
1. We create a link between all concepts in the semantic memory with all the
emotion triggers. We consider the strength of link the higher of the two Emotion
trigger Association Ratio scores.
The text is processed in the order in which emotion triggers appear and finally
obtain the activation value for each emotion trigger.
The output for the values of the emotions in text is obtained by multiplying the
activation values with 100 and adding the scores obtained for the same emotion
from different emotion triggers when it is the case.

6.2.7. EVALUATION OF THE EMOTION DETECTION METHOD


BASED ON EMOTION TRIGGERS
The evaluation of the system presented was done using the test data provided within
the SemEval Task No. 14: Affective Text test set (Strapparava and Mihalcea, 2007)
and its Spanish translation. In the task proposed in SemEval, the objective was to
assign valence positive or negative - and classify emotion of 1000 news headlines
provided as test set according to 6 given emotions: joy, fear, sadness, anger,
surprise and disgust. In order to test our emotion trigger method, we employed this
test set and its translation to Spanish. The results we obtained are presented in Table
6.1 for valence classification and in Table 6.2 for one of the 6 emotions fear,
which is the predominant emotion in the emotion triggers lexicon (for the F
measure we considered alpha 0.5):

190

Precision Recall F-measure


English
75.23
65.01
69.74
Spanish
71.1
66.13
68.52
Table 6.1. System results for valence annotation
Precision Recall F-measure
English
47.21
45.37
46.27
Spanish
46.01
43.84
44.89
Table 6.2. System results for emotion annotation for fear
Motivated by the low results and the low coverage of the resource, we
subsequently proposed a more thorough method to detect emotion from text using
commonsense knowledge.

6.3. EMOTION DETECTION FROM TEXT USING APPRAISAL


CRITERIA BASED ON COMMONSENSE KNOWLEDGE
6.3.1. INTRODUCTION
The different emotion theories proposed in psychology give various explanations as
to why certain episodes lead to a specific affective state: models viewing emotions
as expressions believe that there are specific action tendencies given by stimuli;
models of emotions as embodiments view them as physiological changes; models
of emotions as social constructs view emotions as built through experience and
language; cognitive approaches to emotion the so-called appraisal theories
(Scherer, 1999) state that an emotion can only be experienced by a person if it is
elicited by an appraisal of an object that directly affects them and that the result is
based on the persons experience, goals and opportunities for action (Calvo and
DMello, 2010).
The latter model, besides being the leading one in cognitive psychology, is also
the one that can be best applied to detect emotions expressed in text, as most of the
factors that explain affect in other theories cannot be extracted from natural
language only (the physiological changes factor, for example, cannot be detected
from text unless they are explicitly stated therein).
The aim of this research is to:
1. Propose a method for modeling affective reaction to real-life situations
described in text, based on the psychological model of the appraisal theory.
Practically, we propose the modeling of situations presented in text as
action chains (changes produced by or occurring to an agent related with
191

the state of a physical or emotional object) and the context in which they
take place, using ontological representations. In this way, we abstract from
the treatment of texts as mere sequences of words to a conceptual
representation, able to capture the semantics of the situations described, in
the same manner as psychological models claim that humans do.
2. Design and populate a knowledge base of action chains called EmotiNet,
based on the proposed model. We will show the manner in which using the
EmotiNet ontologies, we can describe the elements of the situation (the
actor, the action, the object etc.) and their properties - corresponding to
appraisal criteria. Moreover, we demonstrate that the resource can be
extended to include all such defined criteria, either by automatic extraction,
extension with knowledge from other sources, such as ConceptNet (Liu and
Singh, 2004) or VerbOcean (Chklovski and Pantel, 2004), inference or by
manual input.
Motivated by the fact that most of the research in psychology has been
made on self-reported affect, the core of the proposed resource is built from
a subset of situations and their corresponding emotion labels taken from the
International Survey on Emotion Antecedents and Reactions (ISEAR)
(Scherer and Walbott, 1997).
3. Propose and validate a method to detect emotion in text based on EmotiNet
using new examples from ISEAR. We thus evaluate the usability of the
resource and demonstrate the appropriateness of the proposed model.

6.3.2. CONTEXT OF RESEARCH


Affect-related phenomena have traditionally been studied in depth by disciplines
such as philosophy or psychology. However, due to the advances in computing and
the growing role of technology in everyday life, the past decades have shown an
increasing interest in building software systems that automatically process affect. In
order for such systems to benefit from the knowledge acquired in social sciences,
interdisciplinary methods have been proposed, which use the existing theoretical
models as basis for engineering computational ones.
This section explores the state of the art in the three domains our present
research is closely related to: approaches to emotion detection in artificial
intelligence, appraisal models in psychology and knowledge bases in NLP
applications.

192

EMOTION DETECTION SYSTEMS IN ARTIFICIAL INTELLIGENCE


In Artificial Intelligence (AI), the term affective computing was first introduced by
Picard (1995). Although there were previous approaches in the 80s and 90s, in the
field of NLP, the task of emotion detection has grown in importance together with
the exponential increase in the volume of subjective data on the Web in blogs,
forums, reviews, etc. Previous approaches to spot affect in text include the use of
models simulating human reactions according to their needs and desires (Dyer,
1987), fuzzy logic (Subasic, 2000), lexical affinity based on similarity of contexts
the basis for the construction of WordNet Affect (Strapparava and Valitutti, 2004) or
SentiWord-Net (Esuli and Sebastiani, 2005), detection of affective keywords (Riloff
et al., 2003) and machine learning using term frequency (Pang et al., 2002; Wiebe
and Riloff, 2006). The two latter approaches are the most widely used in emotion
detection systems implemented for NLP, because they are easily adaptable across
domains and languages. Other proposed methods include the creation of syntactic
patterns and rules for cause-effect modelling (Mei Lee et al., 2009). Significantly
different proposals for emotion detection in text are given in the work by (Liu et al,
2003) and the recently proposed framework of sentic computing (Cambria et al.,
2009), whose scope is to model affective reaction based on commonsense
knowledge. Danisman and Alpkocak (2008) proposed an approach based on
vectorial representations. The authors compute the set of words that is
discriminatory for 5 of the 7 emotions in the ISEAR corpus and represent the
examples using measures computed on the basis of these terms.
Finally, an up-to-date survey on the models of affect and their AC applications is
presented by Calvo and DMello (2010).
APPRAISAL THEORIES
The set of models in psychology known as the appraisal theories claim that
emotions are elicited and differentiated on the basis of the subjective evaluation of
the personal significance of a situation, object or event (De Rivera, 1977; Frijda,
1986; Ortony, Clore and Collins, 1988; Johnson-Laird and Oatley, 1989). Thus, the
nature of the emotional reaction can be best predicted on the basis of the
individuals appraisal of an antecedent situation, object or event. In consequence,
there is a need to contextualize emotional response, due to which the same situation
can lead to different affective reactions and similar reactions can be obtained
through different stimuli.
There are different explanations for the elements considered in the appraisal
process (see Scherer, 1999), which are called appraisal criteria. Currently, there is
no common set of such criteria, as different versions of the appraisal theory have
defined their own list of such factors. However, Scherer (1988) shows that the
193

appraisal criteria proposed in the different theories do converge and cover the same
type of appraisals.
Examples of such criteria are the ones proposed and empirically evaluated by
Lazarus and Smith (1988), organized into a four categories:
i.
Intrinsic characteristics of objects and events;
ii.
Significance of events to individual needs and goals;
iii.
Individuals ability to cope with the consequences of the event;
iv.
Compatibility of event with social or personal standards, norms and
values.
Scherer (1988) proposed five different categories of appraisal (novelty, intrinsic
pleasantness, goal significance, coping potential, compatibility standard),
containing a list of 16 appraisal criteria (suddenness, familiarity, predictability,
intrinsic pleasantness, concern relevance, outcome probability, expectation,
conduciveness, urgency, cause: agent, cause: motive, control, power, adjustment,
external compatibility standards, internal compatibility standards). He later used the
values of these criteria in self-reported affect-eliciting situations to construct the
vectorial model in the expert system GENESIS (Scherer, 1993). The system
maintains a database of 14 emotion vectors (corresponding to 14 emotions), with
each vector component representing the quantitative measure associated to the
value of an appraisal component. The values for the new situations are obtained by
asking the subject a series of 15 questions, from which the values for the appraisal
factors considered (components of the vector representing the situation) are
extracted. Subsequetnly, the label assigned to the emotional experience is computed
by calculating the most similar vector in the database of emotion-eliciting
situations.
The appraisal models defined in psychology have also been employed in
linguistics. The Appraisal framework (Martin and White, 2005) is a development of
work in Systemic Functional Linguistics (Halliday, 1994) and is concerned with
interpersonal meaning in text the negotiation of social relationships by
communicating emotion, judgement and appreciation.
KNOWLEDGE BASES FOR NLP APPLICATIONS
As far as knowledge bases are concerned, many NLP applications have been
developed using manually created knowledge repositories such as WordNet
(Fellbaum, 1998), CYC 58 , ConceptNet or SUMO 59 . Some authors tried to learn
ontologies and relations automatically, using sources that evolve in time - e.g. Yago

58
59

http://cyc.com/cyc/opencyc/overview
http://www.ontologyportal.org/index.html

194

(Suchanek et al., 2007) which employs Wikipedia to extract concepts, using rules
and heuristics based on the Wikipedia categories.
Other approaches to knowledge base population were by (Pantel et al., 2004),
and for relation learning (Berland and Charniak, 1999). DIPRE (Brin, 1998) and
Snowball (Agichtein and Gravano, 2000) label a small set of instances and create
hand-crafted patterns to extract ontology concepts.

6.3.3. ISSUES IN PRESENT APPROACHES


As seen from the previous section, an important body of research already exists in
NLP, dealing with emotion detection in text. Thus, it is important to understand why
a new approach is needed and in what way it differs and improves the existing ones.
We illustrate the need to build a more robust model, starting from a series of
examples.
To start from a simple case, a sentence such as (1) I am happy should be
labeled by an automatic system with joy.
Given this sentence, a system working at a lexical level would be able to detect
the word happy (for example using WordNet Affect) and would correctly identify
the emotion expressed as joy. But already a slightly more complicated example
(2) I am not happy would require the definition of inverse emotions and the
approach would no longer be straightforward. In the second example, although
emotion words are present in the text, additional rules have to be added in order to
account for the negation.
Now let us consider the example: (3) Im going to a party, which should be
labeled with joy as well. A system working at a lexical level would already be
overwhelmed, as no word that is directly related to this emotion is present in the
text. A method to overcome this issue is proposed in by sentic computing (Cambria
et al., 2009) and by (Liu et al, 2003), whose main idea is acquiring knowledge on
the emotional effect of different concepts. In this manner, the system would know
that going to a party is something that produces joy. This approach solves the
problem of indirectly mentioning an emotion by using the concepts that are relating
to it instead. However, it only spots the emotion contained in separated concepts
and does not integrate their interaction or context (cognitive or affective) in which
they appear. If the example we considered is extended as in (4) Im going to a
party, although I should study for my exam., the emotion expressed is no longer
joy, but most probably guilt (or a mixture of joy and guilt, but from which
guilt prevails). As it can be noticed, even if there are concepts that according to our
general knowledge express a certain emotion (e.g. going to a party- joy), their
presence in the text cannot be considered as a mark that the respective sentence
directly contains that emotion (e.g. (5) Going to a party is not always fun., which
can be a view expressed in a text). Also, their meaning can be completely changed
195

depending on the context in which they appear (e.g. (6) I must go to this party, in
which the obligation aspect completely changes the emotion label of the situation).
As we can see from examples (3) to (6), even systems employing world knowledge
(concepts instead of words) would fail in most of these cases.
Similarly, we can show that while the fuzzy models of emotion perform well for
a series of cases that fit the described patterns, they remain weak at the time of
acquiring, combining and using new information.
The most widely used methods of affect detection in NLP are the based on
machine learning models built from corpora. Even such models, while possibly
strong in determining lexical or syntactic patterns of emotion, are limited as far as
the extraction of the text meaning is concerned, even when deep text understanding
features are used. They are also limited at the time of, for example, semantically
combining the meaning of different situations that taken separately lead to no
affective reaction, but their interaction does (i.e. when world knowledge is required
to infer the overall emotional meaning of the situation).
Besides the identified shortcomings, which can be overcome by using existing
methods, there are also other issues, which none of the present approaches consider.
Even if we follow only our own intuition, without regarding any scientific model of
emotion, we can say that, for example, the fact that the development of emotional
states is also highly dependent on the affect at the current moment; also the context
in which the action takes place, the characteristics of the agent performing it, or of
the object of the action, and all the peculiarities concerning these elements can
influence the emotion felt in a specific situation.
Given the identified pitfalls of the current systems and their impossibility to take
into account such factors as context and characteristics of the elements in it, we
propose a new framework for modeling affect, that is robust and flexible and that is
based on the most widely used model of emotion that of the appraisal theories.
The implementation of such models showed promising results (Scherer, 1993).
However, they simply represented in a quantitative manner the appraisal criteria in
a self-reported affective situation, using multiple choice questionnaires. The
problem becomes much more complex, if not impossible, when such factors have to
be automatically extracted from text. If the appraisal criteria for the
actor/action/object of the situation are not presented in the text, they cannot be
extracted from it.
Given all these considerations, our contribution relies in proposing and
implementing a resource for modeling affect based on the appraisal theory, that can
support:
a) The automatic processing of texts to extract:
! The components of the situation presented (which we denote by action
chains) and their relation (temporal, causal etc.)
196

The elements on which the appraisal is done in each action of the chain
(agent, action, object);
! The appraisal criteria that can automatically be determined from the text
(modifiers of the action, actor, object in each action chain);
b) The inference on the value of the appraisal criteria, extracted from external
knowledge sources (characteristics of the actor, action, object or their modifiers that
are inferable from text based on common-sense knowledge);
c) The manual input of appraisal criteria of a specific situation.
!

6.3.4 A METHOD FOR MODELING AFFECTIVE REACTION USING


APPRAISAL MODELS
As we have seen, different criteria have been defined for the appraisal process.
They can easily be extracted, as in the case of the GENESIS system, when specific
questions are asked about them. However, automatically determining all the
appraisal criteria from a text is not a trivial issue. Sometimes, this is impossible
when these factors are not present in the text, either direcly or inferable from
common-sense knowledge - e.g. familiarity, concern relevance, outcome
probability, predictability, expectation. To illustrate this case, we will employ an
example of self-reported affective situation from the ISEAR databank and try to
answer the questions asked by the GENESIS system based on the information
present in the text. Should we be able to extract the answers from the text manually,
it would also be possible to perform this extraction process automatically. Table 6.3
presents a self-reported situation when the emotion joy was felt (more
information on the person recalling this situation can be obtained from the ISEAR
database e.g. age, sex, religion, etc., but we will only try to answer the questions
using the information present in the text). The questions are the ones asked in the
GENESIS system, and are taken from (Scherer, 1993). In the third column of the
table, we annotate whether or not we can answer the question, using YES (we can
answer it based on the information in the text), YES/I (we can answer the question
based on inference on the information presented in the text), WK (we can answer it
based on our world knowledge), FD (we can answer it based on factual data in the
ISEAR database, on the subject describing the experience) and NO (we cannot
answer this question).

197

Situation
I went to buy a bicycle with my father. When I wanted to pay, my father took his purse
and payed.
Q.
Question
Can
we
No.
answer?
1.
Did the situation that elicited your emotion happen very suddenly or YES
abruptly?
2.
Did the situation concern an event or action that had happened in the YES
past, that had just happened or that was expected in the future?
3.
This type of event, independent of your personal evaluation, would it WK
be generally considered as pleasant or unpleasant?
4.
Was the event relevant for your general well-being, for urgent needs WK
you felt, or for specific goals you were pursuing at the time?
5.
Did you expect the event and its consequences before the situation YES
actually happened?
6.
Did the event help you or hinder you in satisfying your needs, in YES
pursuing your plans or in attaining your goals?
7.
Did you feel that action on your part was urgently required to cope WK
with the event and its consequences?
8.
Was the event caused by your own actions in other words, were YES/I
you partially or fully responsible for what happened?
9.

Was the event caused by one or several other persons in other YES
words, were other people fully or partially responsible for what
happened?
10. Was the event mainly due to chance?
NO
11. Can the occurrence and the consequences of this type of event NO
generally be controlled or modified by human action?
12. Did you feel that you had enough power to cope with the event i.e. NO
being able to influence what was happening or to modify the
consequences?
13. Did you feel that, after you used all your means of intervention, you WK
could live with the situation and adapt to the consequences?
14. Would the large majority of people consider what happened to be WK/FD
quite in accordance with social norms and morally acceptable?
15. If you were personally responsible for what happened, did your NO
action correspond to your self image?
Table 6.3. Analysis of the possibility to extract answers concerning appraisal criteria
from a self-reported affective situation (the questions are reproduced from Scherer, 1993)

198

As we can see from Table 6.3, the majority of appraisal criteria cannot be
extracted automatically from the text, as there is no information on them directly
mentioned therein. Some criteria can only be inferred from what is said in the text,
others depend on our use of the world knowledge and there are even questions to
which we cannot answer, since those details are specific to the person living the
reported situation.
Nonetheless, this analysis can offer us a very good insight on the phenomena
that is involved in the appraisal process, from which we can extract a simpler
representation. Viewed in a simpler manner, a situation is presented as a chain of
actions, each with an author and an object; the appraisal depends on the temporal
and causal relationship between them, on the characteristics of the actors involved
in the action and on the object of the action.
Given this insight, the general idea behind our approach is to model situations as
chains of actions and their corresponding emotional effect using an ontological
representation. According to the definition provided by Studer et al. (1998), an
ontology captures knowledge shared by a community that can be easily sharable
with other communities. These two characteristics are especially relevant if we
want the recall of our approach to be increased. Knowledge managed in our
approach has to be shared by a large community and it also needs to be fed by
heterogeneous sources of common knowledge to avoid uncertainties. However,
specific assertions can be introduced to account for the specificities of individuals
or contexts.
In this manner, we can model the interaction of different events in the context in
which they take place and add inference mechanisms to extract knowledge that is
not explicitly present in the text. We can also include knowledge on the appraisal
criteria relating to different concepts found in other ontologies and knowledge bases
(e.g. The man killed the mosquito. does not produce the same emotional effect as
The man killed his wife. or The man killed the burglar in self-defence., because
the criteria used to describe them are very different).
At the same time, we can define the properties of emotions and how they
combine. Such an approach can account for the differences in interpretation, as the
specific knowledge on the individual beliefs or preferences can be easily added as
action chains or affective appraisals (properties) of concepts.

6.3.5 BUILDING THE EMOTINET KNOWLEDGE BASE


Based on the model we proposed, we aim at representing chains of actions and their
corresponding emotional labels from several situations in such a way that we will
be able to extract general patterns of appraisal.
The approach we propose defines a new knowledge base to store action chains,
called EmotiNet, which aims to be a resource for detecting emotions in text, and a
199

(semi)automatic, iterative process to build it, which is based on existing knowledge


from different sources, mainly text corpora. This process principally aims at
extracting the action chains from a document and adding them to the knowledge
base.
From a more practical viewpoint, our approach defines an action chain as a
sequence of action links, or simply actions that trigger an emotion on an actor. Each
specific action link can be described with a tuple (actor, action type, patient,
emotional reaction). Specifically, the process proposed was divided into nine steps
(between brackets, we specify the manner in which the step was performed and the
tools used):
1. Selection of initial set of examples of situations corresponding to the seven
emotions from the ISEAR databank corresponding to the core of the
knowledge base (automatic selection from the database, using as filtering
condition the mention of a family member);
2. Clustering of examples based on their similarity used to represent classes
of examples that are similar (automatically, using the Lesk similarity
implemented in Pedersens Statistics package and performing clustering
with K-Means using Weka60);
3. Selection of cluster representatives to be modelled (randomly);
4. Semantic role identification in the examples to extract triples subjectaction-object to be added to the KB (using the SRL system in (Moreda et
al., 2007));
5. Modeling of the situations based on the subject action object emotional reaction model (core examples are manually represented under
this tuple form);
6. Modeling of emotions and their interaction based on psychological theories
(manually);
7. Evaluation of the obtained ontology (automatically, through consistence
verification);
8. Extension of the ontology using the VerbOcean resource (automatically).
ISEAR SELF-REPORTED AFFECT
Self-reported affect is the most commonly used paradigm in psychology to study
the relationship between the emotional reaction and the appraisal preceding it
(Scherer, 1999). The affective response to events depends on the context in which
they take place. Clues on the affective state of a person, as well as their personal
traits influence the emotional reaction to a situation.

60

http://www.cs.waikato.ac.nz/ml/weka/

200

In the International Survey of Emotional Antecedents and Reactions


(ISEAR) 61 (Scherer and Wallbott, 1997), the student respondents, both
psychologists and non-psychologists, were asked to report situations in
which they had experienced all of 7 major emotions (joy, fear, anger,
sadness, disgust, shame, and guilt). In each case, the questions covered the
way they had appraised the situation and how they reacted. Some examples
of entries in the ISEAR databank are: I felt anger when I had been
obviously unjustly treated and had no possibility to prove they were wrong.
A bus drove over my right leg. The event itself was not very frightening but
when I had to wait in the emergency ward for three hours and then my leg
began to swell, I was frightened. Each example is attached to one single
emotion.
In order to have a homogenous starting base, we selected from the 7667
examples in the ISEAR database only the ones that contained descriptions of
situations between family members. This resulted in a total of 174 examples of
situations where anger was the emotion felt, 87 examples for disgust, 110 examples
for fear, 223 for guilt, 76 for joy, 292 for sadness and 119 for shame.
Subsequently, the examples were POS-Tagged using TreeTagger. Within each
category, we then computed the similarity of the examples with one another, using
the implementation of the Lesk distance in Ted Pedersens Similarity Package. This
score is used to split the examples in each emotion class into six clusters using the
Simple K-Means implementation in Weka. The idea behind this approach is to
group examples that are similar, in vocabulary and structure. This fact was
confirmed by the output of the clusters.
SEMANTIC ROLE LABELING OF SITUATIONS
The next step performed was extracting, from each of the examples, the actions that
are described. In order to do this, we employ the semantic role labeling (SRL)
system introduced by Moreda et al. (2007). From the output of this system, we
manually extract the agent, the verb and the patient (the surface object of the verb).
For example, if we use the input sentence Im going to a family party because my
mother obliges me to, the system extracts two triples with the main actors of the
sentences: (I, go, family party) and (mother, oblige, me), related by the causal
adverb because.
Moreover, the subjects of these sentences might include anaphoric expressions
or, even, self-references (references to the speaker, e.g. I, me, myself) that
have to be solved before creating the instances. The resolution of anaphoric
61

http://www.unige.ch/fapse/emotion/databanks/isear.html

201

expressions (not self-references) was accomplished automatically, using a heuristic


selection of the family member mentioned in the text that is closest to the anaphoric
reference and whose properties (gender, number) are compatible with the ones of
the reference. Self-references were also solved (replacement of the mentions of I
with the speaker), by taking into consideration the entities mentioned in the
sentence and deducing the possible relations. In case of ambiguity, we chose the
youngest, female (if any) member. Following the last example, the subject of the
action would be assigned to the daughter of the family because the sentence is
talking about her mother and these triples would be updated: (daughter, go,
family_party), (mother, oblige, daughter) and (daughter, feel, angry).
Finally, the action links (triples) were grouped and sorted in action chains. This
process of sorting was determined by the adverbial expressions that appear within
the sentence, which actually specify the position of each action on a temporal line
(e.g. although because, when). We defined rules according to which the
actions introduced by these modifiers happen prior to or after the current context.
MODELS OF EMOTION
Representing emotions is a challenging task. It has been argued that emotional
representations cannot be separated from the experiences they correspond to, as
there are no real labels that can be assigned to emotions. Other issues in emotion
representation are that they cannot be studied individually, they do not happen
instantaneously (but there is a continuum of emotions). In order to describe the
emotions and the way they relate and compose, we employ Robert Plutchiks wheel
of emotion (Plutchik, 2001) and Parrots tree-structured list of emotions (Parrot,
2001). These models are the ones that best overlap with the emotions comprised in
the ISEAR databank. Moreover, they contain an explicit modeling of the relations
between the different emotions. Plutchiks wheel of emotions contains 8 basic
emotions and a set of advanced, composed emotions. The model described by
Parrot comprises primary, secondary and tertiary emotions. The primary ones are
love, joy, surprise, anger and fear.
Our approach combines both models by adding the primary emotions missing in
the first model and adding the secondary and tertiary emotions as combinations of
the basic ones. Using this combined model as a reference, we manually assigned
one of the seven most basic emotions (anger, fear, disgust, shame, sadness, joy or
guilt) or the neutral value to all the action links obtained after the SRL of examples,
thus generating 4-tuples (subject, action, object, emotion), e.g. (daughter, go, family
party, neutral) or (mother, oblige, daughter, disgust), that have the appropriate
structure to be integrated in the core of EmotiNet.

202

ONTOLOGY DESIGN
The process of building the core of the EmotiNet knowledge base (KB) of action
chains started with the design of the core of knowledge, in our case ontology.
Specifically, the design process was divided in three stages:
1. Establishing the scope and purpose of the ontology. The ontology we
propose has to be capable of defining the concepts required in a general
manner, which will allow it to be expanded and specialized by external
knowledge sources. Specifically, the EmotiNet ontology needs to capture
and manage knowledge from three domains: kinship membership, emotions
(and their relations) and actions (characteristics and relations between
them).
2. Reusing knowledge from existing ontologies. In a second stage, we
searched for other ontologies on the Web that contained concepts related to
the knowledge cores we needed. At the end of the process, we located two
ontologies that would be the basis of our ontological representation: the
ReiAction ontology62, which represents actions between entities in a general
manner and whose RDF (Resource Description Framework) graph is
depicted in Figure 6.3, and the family relations ontology63, which contained
knowledge about family members and the relations between them.
3. Building our own knowledge core from the ontologies imported. This
third stage involved the design of the last remaining core, i.e. emotion, and
the combination of the different knowledge sources into a single ontology:
EmotiNet. In this case, we designed a new knowledge core from scratch
based on a combination of the models of emotion presented (see Figure
6.4). This knowledge core includes different types of relations between
emotions and a collection of specific instances of emotion (e.g. anger, fear,
joy). In the last step, these three cores were combined using new classes and
relations between the existing members of these ontologies.
ONTOLOGY EXTENSION AND POPULATION
The next process extended EmotiNet with new types of action and instances of
action chains using real examples from the ISEAR corpus. We began the process by
manually sellecting a subset of 175 documents from the collection after applying
the SRL system proposed by Moreda et al. (2007), with expressions related to all
the emotions: anger (25), disgust (25), guilt (25), fear (25), sadness (25), joy (25)
and shame (25). The criteria for selecting this subset were the simplicity of the
sentences and the variety of actions described.
62
63

www.cs.umbc.edu/~lkagal1/rei/ontologies/ReiAction.owl
www.dlsi.ua.es/~jesusmhc/EmotiNet/family.owl

203

Figure 6.3. Partial RDF graph of ReiAction ontology

Figure 6.4. Partial RDF graph of the Emotion Ontology.

204

Figure 6.5. Main concepts of EmotiNet.


Once we extracted the actors, actions and objects from the sentences in the 175
situations chosen to be modeled, we ordered them and assigned each of the actions
an emotion. Thus, we obtained 175 action chains (ordered lists of tuples). In order
to be included in the EmotiNet knowledge base, all their elements needed to be
mapped to existing concepts or instances within the KB. When these did not exist,
they were added to it.
In EmotiNet, each action link was modeled as a 4-tuple of ontology instances
(actor, action, patient, emotion) that could be affected by a set of modifiers (e.g.
opposite, degree of intensity) that directly affects each action. We would like to
highlight that each tuple extracted from the process of semantic role labelling and
emotion assignation has its own representation in EmotiNet as an instance of the
subclasses of Action. Each instance of Action is related to an instance of the class
Feel, which represent the emotion felt in this action. Subsequently, these instances
(action links) were grouped in sequences of actions (class Sequence) ended by an
instance of the class Feel, which, as just mentioned, would determine the final
emotion felt by the main actor(s) of the chain.
In our example, we created two new classes Go and Oblige (subclasses of
DomainAction) and two new instances from them: instance act1: (class Go,
actor daughter, target family_party); and instance act2: (class Oblige, actor
mother, target daughter). The last action link already existed within EmotiNet
from another chain so we reused it: instance act3: (class Feel, actor daughter
emotionFelt anger).
The next step consisted in sorting and grouping these instances into sequences
by means of instances of the class Sequence. This is a subclass of Action that can
establish the temporal order between two actions (which one occurred first).
Considering a set of default rules that modelled the semantics of the temporal
adverbial expressions, we group and order the instances of action in sequences,
205

which must end with an instance of Feel. Figure 6.6 shows an example of a RDF
graph, previously simplified, with the action chain of our example.
Action Chain
anger

emotionFelt

mother_f1

feel_anger_1
actor

actor
hasChild

disgust

oblige_1

implies

target

daughter_f1

actor

second

go_1

first

target

sequence_2
second

party_1
first

sequence_1

Figure 6.6. RDF graph of an action chain.


Following this strategy, we finally obtained a tight net of ontology instances that
express different emotions and how actions triggered them. We used Jena 64 and
MySQL for managing and storing the knowledge of EmotiNet on a database in
order to carry out our experiments.
ONTOLOGY EXPANSION
Although our knowledge core was favorably assessed, the number of action types
was small taking into consideration the number of verbs that appear, for instance, in
an English dictionary. In order to extend the coverage of the resource and include
certain types of interactions between actions, we expanded the ontology with the
actions and relations from VerbOcean (Chklovski and Pantel, 2004). In particular,
299 new actions were automatically included as subclasses of DomainAction,
which were directly related to any of the actions of our ontology through three new
relations: can-result-in, happens-before and similar. This knowledge extracted from
VerbOcean is the basis of inferences when the information extracted from new texts
does not appear in our initial set of instances. Thus, this process of expansion is
essential for EmotiNet, since it adds new types of action and relations between
actions, which might not have analyzed before. This reduced the degree of
dependency between the resource and the initial set of examples. The more external
64

http://jena.sourceforge.net/

206

sources of general knowledge we add the more flexible will be EmotiNet, thus
increasing the possibilities of processing unseen action chains.

6.3.6 EXPERIMENTS AND EVALUATION


The evaluation of our approach consists in testing if by employing the model we
built, we are able to detect the emotion expressed in other examples pertaining to
the categories in ISEAR.
EXPERIMENTAL SETTINGS
In order to assess the new examples from ISEAR, we followed the same process we
used for building the core of EmotiNet, with the exception that the manual
modeling of examples into tuples was replaced with the automatic extraction of
(actor, verb, patient) triples from the output given by the (Moreda et al., 2007) SRL
system. Subsequently, we eliminated the stopwords in the phrases contained in
these three roles and performed a simple coreference resolution. Next, we order the
actions presented in the phrase, using the adverbs that connect the sentences,
through the use of patters (temporal, causal etc.). The resulted action chains
represent the test set which will be used in carrying different experiments.
During the development of EmotiNet, we stored two partial versions in order to
evaluate the manner in which the number of examples manually included in the
EmotiNet core and the number of emotions considered influence the results of the
emotion detection task.
The first version of the test set (marked with A) contains as core knowledge on 4
emotions (the number of examples included is specified in brackets): anger (26),
disgust (10), guilt (24) and fear (16). The second test set (marked with B) contains
knowledge on the same 4 emotions as in A, but with an equal number of examples
(25). The third test set (marked with C) comprises the final version of the EmotiNet
core knowledge base containing all 7 emotions in ISEAR, and 25 examples for
each.
On all these test sets, we perform the following series of experiments:
(1). In the first experiment, for each of the situations in the test sets (represented
as action chains), we search the EmotiNet KB to encounter the sequences in which
these actions in the chains are involved and their corresponding subjects. As a result
of the search process, we obtain the emotion label corresponding to the new
situation and the subject of the emotion based on a weighting function. This
function takes into consideration the number of actions and the position in which
they appear in the sequence contained in EmotiNet. The issue in this first approach
is that many of the examples cannot be classified, as the knowledge they contain is
207

not present in the ontology. The experiment was applied to each of the three test
sets (A, B, C), and the corresponding results are marked with A1, B1 and C1,
respectively.
(2). A subsequent approach aimed at surpassing the issues raised by the missing
knowledge in EmotiNet. In a first approximation, we aimed at introducing extra
knowledge from VerbOcean, by adding the verbs that were similar to the ones in the
core examples (represented in VerbOcean through the similar relation).
Subsequently, each of the actions in the examples to be classified that was not
already contained in EmotiNet, was sought in VerbOcean. In case one of the similar
actions was already contained in the KB, the actions were considered equivalent.
Further on, each action was associated with an emotion, using ConceptNet relations
and concepts. Action chains were represented as chains of actions with their
associated emotion. Finally, new examples were matched against chains of actions
containing the same emotions, in the same order. While more complete than the
first approximation, this approach was also affected by lack of knowledge about the
emotional content of actions. To overcome this issue, we proposed two heuristics:
(2a) In the first one, actions on which no affect information was available,
were sought in within the examples already introduced in the EmotiNet and were
assigned the most frequent class of emotion labeling them. The experiment was
applied to each of the three test sets (A, B, C), and the corresponding results are
marked with A2a, B2a and C2a, respectively.
(2b) In the second approximation, we used the most frequent emotion
associated to the known links of a chain, whose individual emotions were
obtained from SentiWordNet. In this case, the core of action chains is not
involved in the process. The experiment was applied to each of the three test sets
(A, B, C), and the corresponding results are marked with A2b, B2b and C2b,
respectively.
EVALUATION RESULTS
a) Test set A:
We performed the steps described on the 516 examples (ISEAR phrases
corresponding to the four emotions modelled, from which the examples used as
core of EmotiNet were removed). For the first approach, the queries led to a result
only in the case of 199, for the second approach, approximations (A2a) and (A2b) 409. For the remaining ones, the knowledge stored in the KB is not sufficient, so
that the appropriate action chain can be extracted. Table 6.4 presents the results of
the evaluations on the subset of examples, whose corresponding query returned a
result.

208

Table 6.5 reports on the recall obtained when testing on all examples. The baseline
is random, computed as average of 10 random generations of classes for all
classified examples.
b) Test set B:
We performed the steps described on the 487 examples (ISEAR phrases
corresponding to the four emotions modelled, from which the examples used as
core of EmotiNet were removed). For the first approach, the queries led to a result
only in the case of 90, for the second approach, approximations (B2a)- 165 and
(B2b) - 171. For the remaining ones, the knowledge stored in the KB is not
sufficient, so that the appropriate action chain can be extracted. Table 6.6 presents
the results of the evaluations on the subset of examples, whose corresponding query
returned a result. Table 6.7 reports on the recall obtained when testing on all
examples.
The baseline is random, computed as average of 10 random generations of
classes for all classified examples.
c) Test set C:
We performed the steps described on the 895 examples (ISEAR phrases
corresponding to the seven emotions modelled, from which the examples used as
core of EmotiNet were removed). For the first approach, the queries led to a result
only in the case of 571, for the second approach, approximations (C2a) 617 and
(C2b) - 625. For the remaining ones, the knowledge stored in the KB is not
sufficient, so that the appropriate action chain can be extracted. Table 6.8 presents
the results of the evaluations on the subset of examples, whose corresponding query
returned a result. Table 6.9 reports on the recall obtained when testing on all
examples.
The baseline is random, computed as average of 10 random generations of
classes for all classified examples.
Table 6.10 reproduces the results reported in Danisman and Alpkocak (2008)
which we will mark as DA-, on the ISEAR corpus, using ten-fold cross-validation.
We compare them to the results we obtained in the presented experiments. As the
authors only present the mean accuracy obtained for 5 of the emotions in ISEAR
(anger, disgust, fear, joy, sadness) and they perform ten-fold cross-validation on all
examples in the abovementioned corpus, the results are not directly comparable. In
fact, a ten-fold cross validation means that they have used 90% of the cases to train
the classifier and only tested on the rest of 10% of the cases. As a proof of the
significance of the results obtained, we also include the evaluation outcome of the
same system on another corpus (marked as DA- SemEval). Again, we cannot

209

directly compare the results, but we can notice that our system performs much
better in terms of accuracy and recall when tested on new data.
In any case, we believe that such comparisons can give a clearer idea of the task
difficulty and the measure of the success of our approach. On the last line, we
include the results reported for the GENESIS system (Scherer, 1993). However, it
should be noted that this expert system does not directly detect and classify emotion
from text; it only represents answers to a set of questions aimed at determining the
values of the appraisal factors included in the emotional episode, after which it
computes the similarity to previously computed vectors of situations.
Correct

Total
Accuracy
Emotion
A
A1
A2a
A2b
A1
A2b
A1
A2a
A2b
2a
disgust
11
25
25
26
59
63
42.3
42.4
39.7
anger
38
27
26
62
113
113
61.3
23.9
23
fear
4
5
7
29
71
73
16
7.1
9.6
guilt
13
30
26
86
166
160
15.1
18.1
16.3
Total
66
87
84
199
409
409
31.2
21.3
20.5
Baseline
61
84
84
229
409
409
21.9
20.5
20.5
Table 6.4. Results of the emotion detection using EmotiNet on classified examples in
test set A
Correct
Total
Recall
A1
A2a
A2b
A1
A1
A2a
A2b
disgust
11
25
25
76
16.3
32.9
32.9
anger
38
27
26
148
27
18.3
17.6
fear
4
5
7
94
4.5
5.3
7.5
guilt
13
30
26
198
7.7
15.2
13.1
Total
66
87
84
516
14
16.9
16.2
Baseline
112
112
112
516
21.7
21.7
21.7
Table 6.5. Results of the emotion detection using EmotiNet on all test examples in
test set A
Emotion

Correct
Emotion
disgust
anger
fear
210

B1 B2a B2b

B1

10
16
37

41
102
55

28
39
43

29
39
44

Total
B
2a
52
114
74

Accuracy
B2b
67
119
76

B1

B2a

B2b

24.39 53.85 43.28


15.69 34.21 32.77
67.27 58.11 57.89

Correct

Total
Accuracy
Emotion
B
B1 B2a B2b B1
B2b B1
B2a
B2b
2a
guilt
27 55
59 146 157 165 18.49 35.03 35.76
Total
90 165 171 344 397 427 26.16 41.56 40.05
Table 6.6. Results of the emotion detection using EmotiNet on classified examples in
test set B
Correct
Total
Recall
B1
B2a
B2b
B1
B1
B2a
B2b
disgust
10
28
29
59
16.95 47.46 49.15
anger
16
39
39
145
11.03 26.90 26.90
fear
37
43
44
85
43.53 50.59 51.76
guilt
27
55
59
198
13.64 27.78 29.80
Total
90
165
171
487
18.48 33.88 35.11
Baseline
124
124
124
487
0.25
0.25
0.25
Table 6.7. Results of the emotion detection using EmotiNet on all test examples in
test set B
Emotion

Correct

Total
Accuracy
Emotion
C
C1 C2a C2b C1
C2b C1
C2a C2b
2a
disgust
16
16
21
44 42
40 36.36 38.09 52.50
shame
25
25
26
70 78
73 35.71 32.05 35.62
anger
31
47
57 105 115 121 29.52 40.86 47.11
fear
35
34
37
58 65
60 60.34 52.30 61.67
sadness
46
45
41 111 123 125 41.44 36.58 32.80
joy
13
16
18
25 29
35
52
55.17 51.43
guilt
59
68
64 158 165 171 37.34 41.21 37.43
Total
225 251 264 571 617 625 39.40 40.68 42.24
Table 6.8. Results of the emotion detection using EmotiNet on classified examples in
test set C
Emotion
disgust
shame
anger
fear

C1
16
25
31
35

Correct
C2a
C2b
16
21
25
26
47
57
34
37

Total
C1
59
91
145
85

C1
27.11
27.47
21.37
60.34

Recall
C2a
27.11
27.47
32.41
52.30

C2b
35.59
28.57
39.31
61.67
211

Correct
Total
Recall
C1
C2a
C2b
C1
C1
C2a
C2b
sadness
46
45
41
267
17.22
16.85
15.36
joy
13
16
18
50
26
32
36.00
guilt
59
68
64
198
29.79
34.34
32.32
Total
225
251
264
895
25.13
28.04
29.50
Baseline
126
126
126
895
14.0.7
14.07
14.07
Table 6.9. Results of the emotion detection using EmotiNet on all test examples in
test set C
Emotion

Method
Mean Accuracy
DA: NB (with stemming)
67.2
DA: NB (without stemming)
67.4
DA: SVM (with stemming)
67.4
DA: SVM (without stemming)
66.9
DA SemEval: NB (with stemming)
F1= 27.9
DA SemEval: NB (without stemming)
F1= 28.5
DA SemEval: SVM (with stemming)
F1= 28.6
DA SemEval: SVM (without stemming)
F1= 27.8
DA SemEval: Vector Space Model (with stemming)
F1= 31.5
DA SemEval: Vector Space Model (without stemming)
F1= 32.2
EmotiNet A1
31.2
EmotiNet A2a
21.3
EmotiNet A2b
20.5
EmotiNet B1
26.16
EmotiNet B2a
41.56
EmotiNet B2b
40.05
EmotiNet C1
39.4
EmotiNet C2a
40.68
EmotiNet C2b
42.24
GENESIS
77.9
Table 6.10. Comparison of different systems for affect detection using ISEAR or
self-reported affect in general
DISCUSSION AND CONCLUSIONS
From the results in Tables 6.4 to 6.9 we can conclude that the approach is valid,
although much remains to be done to fully exploit the capabilities of EmotiNet.
Given the number of core examples and the results obtained, we can see that the
212

number of chains corresponding to one emotion in the core do influence the final
result directly. However, the systems performs significantly better when an equal
number of core examples is modeled, although when more emotions are evaluated
(the difference between test sets B and C), the noise introduced leads to a drop in
performance.
The comparative results shown in Table 6.10 show, on the one hand, that the task
of detecting affect in text is very difficult. Thus, even if the appraisal criteria are
directly given to a system (as in the case of GENESIS), its accuracy level only
reaches up to 80%. If a system is trained on 90% of the data in one corpus using
lexical information, its performance reaches up to around 68%. However, the results
drop significantly when the approach is used on different data, showing that it is
highly dependent on the vocabulary it uses. As opposed to this, the model we
proposed based on appraisal theories proved to be flexible, its level of performance
improving either by percentual increase, or by the fact that the results for different
emotional categories become more balanced. We showed that introducing new
information can be easily done from existing common-sense knowledge bases and
that the approach is robust in the face of the noise introduced.
From the error analysis we performed, we could determine some of the causes of
error in the system. The first important finding is that extracting only the action,
verb and patient semantic roles is not sufficient. There are other roles, such as the
modifiers, which change the overall emotion in the text (e.g. I had a fight with my
sister sadness, versus I had a fight with my stupid sister anger). Therefore,
such modifiers should be included as attributes of the concepts identified in the
roles, and, additionally, added to the tuples, as they can account for other appraisal
criteria. This can also be a method to account for negation. Given that just 3 roles
were extracted, there were also many examples that did not make sense when input
into the system. Further on, we tried to assigned emotion to all the actions contained
in the chains. However, some actions have no emotional effect. Therefore, an
accurate source of knowledge on the affect associated to concepts has to be added.
Another issue we detected was that certain emotions tend to be classified most
of the times as another emotion (e.g. fear is mostly classified as anger). This is due
to the fact that emotions are subjective (one and the same situation can be a cause
for anger for a person and a cause of fear to another or a mixture of the two); also,
in certain situations, there are very subtle nuances that distinguish one emotion
from the other.
A further source of errors was that lack of knowledge on specific actions. As we
have seen, this knowledge can be imported from external knowledge bases and
integrated in the core. This extension using larger common-sense knowledge bases,
may lead to problems related to knowledge consistency and redundancy, with which
we have not dealt with yet. VerbOcean extended the knowledge, in the sense that
213

more examples could be classified. However, given the ambiguity of the resource
and the fact that it is not perfectly accurate also introduced many errors.
Finally, other errors were produced by NLP processes and propagated at various
steps of the processing chain (e.g. SRL, co-reference resolution). Some of these
errors cannot be eliminated; however, others can be partially solved by using
alternative NLP tools. A thorough analysis of the errors produced at each of the
stages involved in the application and extension of EmotiNet must be made in order
to obtain a clear idea of the importance/noise of each component.

6.4. CONCLUSIONS ON METHODS FOR IMPLICIT


SENTIMENT DETECTION
In this second part of the chapter, we presented our contribution concerning three
major topics. The first one was the proposal of a method to model real-life
situations described in text based on the model of the appraisal theory. Based on the
proposed model, the second contribution was the design and population of
EmotiNet, a knowledge base of action chains representing and storing affective
reaction to real-life contexts and situations described in text. We started our
approach by modelling the actions presented in a set of phrases, describing affective
situations, together with their affective value. We subsequently extended our model
using VerbOcean. Finally, the third contribution lied in proposing and evaluating a
method to detect emotion in text based on EmotiNet, using other examples in the
ISEAR corpus.
We conclude that our approach is appropriate for detecting emotion in text,
although additional elements should be included in the model and extra knowledge
is required. Moreover, we found that the process of automatic evaluation was
influenced by the low performance of the NLP tools used. Thus, alternative tools
must be tested in order to improve the output. We must also test our approach on
corpora where more than one emotion is assigned per context (such as the one in
SemEval 2007 Task14).
All in all, we have seen that the issue of sentiment analysis cannot be resolved
easily, by remaining at a word level or requiring the explicit mention of emotion or
affect-related words. With this chapter, we completed our vision on the needs of
sentiment analysis system and opened the door to a new direction for research,
which is given by the necessity to develop methods that detect implicitly-expressed
emotion.

214

CHAPTER 7. CONTRIBUTIONS
Motto: The value of the sentiment is the value of the sacrifice you are prepared to
make for it. (John Galsworthy)
This thesis focused on the resolution of different problems related to the task of
sentiment analysis. Specifically, we concentrated on:
1. Defining the general task and related concepts, by presenting an overview
of the present definition and clarifying the inconsistencies found among the
ones that were previously given in the literature;
2. Proposing and evaluating methods to define and tackle sentiment analysis
from a variety of textual genres, in different languages;
3. Redefining the task and proposing methods to annotate specific corpora for
sentiment analysis in the corresponding text type, in different languages in
case the task of sentiment analysis was not clearly defined for a specific
textual genre and/or no specific corpora was available for it. These
resources are publicly available for the use of the research community;
4. Applying opinion mining techniques in the context of end-to-end systems
that involve other NLP tasks as well. To this aim, we concentrated on
performing sentiment analysis in the context of question answering and
summarization.
5. Carrying out experiments using existing question answering and
summarization systems, designed to deal with factual data only.
6. Proposing and evaluating a new framework for what we called opinion
question answering and new methods for opinion summarization,
subsequent to experiments showing that systems performing question
answering and summarization over factual texts were not entirely suited in
the context of opinions;
7. Presenting a general method for the detection of implicitly-expressed
emotion from text. First, we presented the method to build a lexicon of
terms that in themselves contain no emotion, but that trigger emotion in a
reader. Subsequently, we abstracted from the analysis of sentiment
expressed in text based on linguistic cues and proposed and evaluated a
method to represent text as action chains. The emotion elicited by the
situation presented in the text was subsequently judged using commonsense
knowledge on the emotional effect of each action in the chain;
8. Evaluating our approaches in international competitions, in order to
compare our approaches to others and validate them.

215

Further on, we present in detail the contributions we have made to the research
in the field of sentiment analysis throughout this thesis and show how the methods
and resources we proposed filled important gaps in the existing research. The main
contributions answer five research questions:
1. How can sentiment analysis and, in a broader perspective, opinion mining
be defined in a correct way? What are the main concepts to be treated in
order to do that?
In Chapter 2, we presented a variety of definitions that were given to concepts
related and involved in the task of sentiment analysis subjectivity, objectivity,
opinion, sentiment, emotion, attitude and appraisal. Our contribution in this chapter
resided in clearly showing that sentiment analysis and opinion mining are not
synonymous, although in the literature they are usually employed interchangeably.
Additionally, we have shown that opinion, as it is defined by the Webster
dictionary, is not synonymous to sentiment. Whereas sentiments are types of
opinions, reflecting the feelings (i.e. conscious part of emotions), all opinions are
not sentiments (i.e there are types of opinions that are not reflective of emotions).
We have also shown that subjectivity analysis is not directly linked to sentiment
analysis as it is considered by many of the researchers in the field. In other words,
detecting subjective sentences does not imply directly obtaining the sentences that
contain sentiment. The latter, as expressions of evaluations based on emotion, are
not necessarily indicated in subjective sentences, but can also be expressed in
objective sentences. Subjective sentences can or cannot contain expressions of
emotions. The idea is summarized in Chapter 2, Figure 2.1:
Finally, we have shown that there is a clear connection between the work done
under the umbrella of sentiment analysis/opinion mining and the one in
appraisal/attitude analysis. Although all these areas are usually considered to refer
to the same type of work, the wider aim of attitude or appraisal analysis can capture
much better the research that has been done in sentiment analysis, including all
classes of evaluation (affective, cognitive, behavioral) and the connection between
author, reader and text meaning. Based on this observation and in view of the
Appraisal Theory, in Chapter 6 we proposed a model of emotion detection based on
commonsense knowledge.
Clearly defining these concepts has also helped in defining in an appropriate
manner the sentiment analysis task in the context of the different textual genres we
employed in our research. Subsequently, the correct definition has made it possible
to define annotation schemes and create resources for sentiment analysis in all the
textual genres we performed research with. All these resources were consequently
employed in the evaluation of the specific methods we created for sentiment
analysis from different textual genres. Both the resources created, through the high
inter-annotator agreement we obtained, as well as the methods we proposed,
216

through the performance of the systems implementing them, have shown our efforts
to give the clear definition were indeed an important contribution to this field.
2. Can sentiment analysis be performed using the same methods, for all text
types? What are the peculiarities of the different text types and how do they
influence the methods to be used to tackle it? Do we need special resources
for different text types?
3. Can the same language resources be used in other languages (through
translation)? How can resources be extended to other languages?
In Chapter 4, we showed the peculiarities of different text types (reviews,
newspaper articles, blogs, political debates), analyzed them and proposed adequate
techniques to address them at the time of performing sentiment analysis. We
evaluated our approaches correspondingly and showed that they perform at the
level of state-of-the-art systems and in many cases outperform them. In this chapter,
we presented different methods and resources we built for the task of sentiment
analysis in different text types. We started by presenting methods to tackle the task
of feature-based opinion mining and summarization, applied to product reviews.
We have analyzed the peculiarities of this task and identified the weak points of
existing research. We proposed and evaluated different methods to overcome the
identified pitfalls, among which the most important were the discovery of indirectly
mentioned features and computing the polarity of opinions in a manner that is
feature-dependent, using the Normalized Google Distance and Latent Semantic
Analysis as measure of term association. Subsequently, we proposed a unified
model for sentiment annotation for this type of text, able to capture the important
phenomena that we had identified different types of sentiment expressions
direct, indirect, implicit, feature mentioning and span of text expressing a specific
opinion. Such a distinction had not been proposed in the literature. Its contribution
is not only given by the annotation process, but also by the fact that considering
larger spans of text as representing one single opinion has lead us to research on
opinion question answering (see Chapter 5) using retrieval of 3-sentences-long text
snippets, highly improving the performance of the opinion question answering
system. Following this annotation scheme, we also proposed a method to detect and
classify opinion stated on the most important features of a product, on which stars
were given in the review, based on textual entailment. Apart from this contribution,
the performance obtained applying this method is indicative of the fact that it is
possible to obtain, apart from a 2-way classification of opinions on product features,
a summary of the most important sentences referring to them. These latter can be
employed to offer support snippets to the feature-based opinion mining and
summarization process, which normally only offers a percent-based summary of the
opinions expressed on the product in question.
217

Further on, we explored different methods to tackle sentiment analysis from


newspaper articles. After the initial experiments, we analyzed the reasons for the
low performance obtained and redefined the task, taking into account the
peculiarities of this textual genre. We created an annotation model and labeled two
different corpora of newspaper article quotations, in English and German. After
redefining the task and delimiting the scope of the sentiment analysis process to
quotations small text snippets containing direct speech, whose source and target
are previously known-, the annotation agreement rose significantly.
Additionally, improving the definition of the task made it possible to implement
automatic processing methods that are appropriate for the task and significantly
improve the performance of the sentiment analysis system we had designed. In the
view of applying sentiment analysis to different types of texts, in which sentimentbearing content is highly mixed with non-opinionated one and where the sources
and targets of opinions are multiple, we have proposed different general methods
for sentiment analysis, which we applied to political debates. The results of this
latter experiment motivated us to analyze the requirements of a general labeling
scheme for the task of sentiment analysis, which can be used to capture all relevant
phenomena in sentiment expression. To this aim, in (Boldrini et al., 2009), we
defined EmotiBlog, an annotation scheme that is able to capture, at a fine-grained
level, all linguistic phenomena related to sentiment expression in text. The
subsequent experiments have shown that this model is appropriate for the training
of machine learning models for the task of sentiment analysis in different textual
genres, in both languages in which experiments have been carried out using it
English and Spanish.
Finally, we have shown that the corpus annotated in this manner can be used to
extract features for machine learning models that can be employed to tackle
sentiment analysis in other languages, through translation. The good results
obtained in the SemEval 2010 Task 18 Disambiguation of Sentiment Ambiguous
Adjectives competition, where we translated the texts from Traditional Chinese to
English and applied an machine learning model trained on the EmotiBlog English
data have proven that the annotation model is robust enough and useful even in the
context of noisy data.
4. How can we deal with opinion in the context of traditional tasks? How can
we adapt traditional tasks (Information Retrieval, Question Answering,
Text Summarization) in the context of opinionated content? What are the
new challenges in this context?
In Chapter 4, we have only concentrated on the task of sentiment analysis as a
standalone challenge, omitting the steps required in order to obtain the texts on
which the sentiment analysis methods were applied or eliminating redundancy in
218

the information obtained.


In a real-world application scenario, however,
automatically detecting the opinion expressed in a text is often not the first, neither
the last task to be performed. Prior to the analysis of the sentiment contained within,
the relevant texts in which opinions are contained on required targets must be
retrieved. Additionally, in many of the cases, the results obtained after
automatically processing texts to determine the sentiment they contain still pose
many problems in terms of volume. Thus, even if the sentiment is determined
automatically, one may still require a summarization component, in order to further
reduce the quantity of information, so that it is can be read and used by a person.
Bearing in mind these necessities, in Chapter 5 we researched on methods to
combine opinion mining with question answering and summarization. Here, we
have shown that performing traditional tasks in the context of opinionated text has
many challenges and that systems that were designed to work exclusively with
factual data are not able to cope with opinion questions. Our contribution resides in
demonstrating, through evaluation, that in the case of such queries, for the question
treatment, new elements have to be defined. We proposed the inclusion of the next
elements: Expected Polarity Type, Expected Source, Expected Target, and have
defined methods to detect them both from the question, as well as the candidate
answers, using opinion mining techniques and employing Semantic Role Labeling.
For the retrieval process, as we have shown before and confirmed in the OQA
scenario, larger spans of texts are more appropriate when dealing with opinionated
content. Specifically, we have shown that retrieving 3-sentences long snippets leads
to better results in the case of OQA systems. By proposing and evaluating these
new elements and techniques, we have shown the manner in which OQA can be
tackled in a robust manner, consistent with the characteristics of opinionated texts.
Lastly, we evaluated the impact of using different tools and resources in this task,
for performing anaphora resolution and expanding the question through
paraphrasing. Studying the manner in which these two components can be
optimally added to the system is an important line for future work that we have
opened.
In the case of opinion summarization, our contribution resides in: a) studying the
order in which the opinion mining and the summarization systems have to be
employed; b) studying the manner in which opinion mining and summarization
systems can be used in tandem; c) studying the effect of topic detection in opinion
mining in the context of opinion summarization; d) proposing a method to
summarize opinion based on the intensity of the sentiment expressed.
In this part of the research, our contribution resides in showing that the
sentiment analysis system must be employed prior to the summarization system and
that the sentiment analysis component must be enhanced with topic-detection
mechanisms. In other cases, although the sentiment analysis systems performs the
219

classification correctly (according to the polarity of the sentiment), the fact that
topic-relatedness is not contemplated leads to the introduction of irrelevant data in
the final summaries. Finally, we have shown that in the case of opinionated text,
relevance is given not only by the information contained, but also by the polarity of
the opinion and its intensity. Although initial results have shown that there is no
correlation between the Gold Standard annotations and the intensity level of
sentences, as output by the sentiment analysis system, given the fact that using this
method, we obtained high results as far as F-measure is concerned in TAC 2008, we
believe that more mechanisms for opinion intensity should be studied, so that the
clear connection between sentence relevance and the opinion it contains, as well as
the intensity it has, can be established.
5. Can we propose a model to detect emotion from text, in the cases where it
is expressed implicitly, needing world knowledge?
In the first chapters of this thesis, we explored the task of sentiment analysis in
different text types and languages, proposing a variety of methods that were
appropriate for tackling the issues in each particular text type. Most of the times,
however, the approaches we took were limited to discovering only the situations
where sentiment was expressed explicitly (i.e. where linguistic cues could be found
in the text to indicate it contained subjective elements or sentiment). Nevertheless,
in many cases, the emotion underlying the sentiment is not explicitly present in text,
but is inferable based on commonsense knowledge (i.e. emotion is not explicitly,
but implicitly expressed by the author, by presenting situations which most people,
based on commonsense knowledge, associate with an emotion, like going to a
party, seeing your child taking his/her first step etc.).
In Chapter 6 of the thesis, we presented our contribution to the issue of
automatically detecting emotion expressed in text in an implicit manner. The initial
approach is based on the idea that emotion is triggered by specific concepts,
according to their relevance, seen in relation to the basic needs and motivations,
underpinning our idea on the Relevance Theory. The second approach we propose
is based on the Appraisal Theory models. The general idea behind it is that
emotions are most of the times not explicitly stated in texts, but results from the
interpretation (appraisal) of the actions contained in the situation described, as well
as the properties of their actors and objects. Our contribution in this last part of the
research resides in setting up a framework for representing situations described in
text as chains of actions (with their corresponding actors and objects), and their
corresponding properties (including the affective ones), as commonsense
knowledge. We show the manner in which the so-called appraisal criteria can be
automatically detected from text and how additional knowledge on the properties of
the concepts involved in such situations can be imported from external sources.
220

Finally, we demonstrate through an extensive evaluation that such a representation


is useful to obtain an accurate label of the emotion expressed in text, without any
linguistic clue being present therein.
As sentiments are directly related to the presence of emotion, detecting implicit
expressions of emotion can increase the performance of sentiment analysis systems,
a fact that we have proven in the experiments we presented in Chapter 3 in the case
of sentiment analysis in product reviews.

221

CHAPTER 8. FUTURE WORK


Motto: Nothing exists except atoms and empty space. Everything else is an
opinion. (Democritus)
In this final chapter, we present the possible directions for the further development
of the research performed in this thesis. In order to be coherent with the structure of
the thesis, these directions will be presented in relation to the research questions we
aimed at answering and the points that remain to be addressed:
1. How can sentiment analysis and, in a broader perspective, opinion mining
be defined in a correct way? What are the main concepts to be treated in
order to create a good definition that can be used to appropriately define
the task and subsequently propose correct methods to tackle it?
In Chapter 2 of the thesis, we presented an overview of the definitions given in
the NLP literature to the related tasks of subjectivity analysis, sentiment analysis,
opinion mining, appraisal/attitude analysis, emotion detection, as well as the
concepts involved in these tasks. We subsequently showed that there is a high
inconsistency between the definitions given both to the tasks, as well. Finally, we
proposed an operational definition that was consistent with the manner in which the
different terms related to sentiment analysis were defined in well-established
sources and in a manner that was coherent with the approaches we took in the
research. Subsequently, in Chapter 4, we showed that sentiment analysis must be
tackled in a different manner, depending on the types of text considered. From these
efforts, a future line of work is the definition of a unified framework for sentiment
analysis, that describes the task in a general, yet consistent manner across genres
and applications. In this sense, as we have shown in the case of newspaper articles,
such a framework should be built by taking into account not only the textual
content, but also the elements that relate to the author and reader.
The subsequent research question we addressed in this thesis were:
2. Can sentiment analysis be performed using the same methods, for all text
types? What are the peculiarities of the different text types and how do they
influence the methods to be used to tackle it? Do we need special resources
for different text types?
3. Can the same language resources be used in other languages (through
translation)? How can resources be extended to other languages?
In Chapter 4, we presented different methods and resources for the task of
sentiment analysis in different text types (reviews, newspaper articles, blogs,
political debates), in different languages (English, Spanish, German). A future line
223

of work is the extension of the proposed resources for other languages, either
through translation, in which case the resources should be refined and evaluated on
different types of texts in their original language, or by direct annotation in the
target language. In the latter case, it would be interesting to study the differences
given by the peculiarities of sentiment expression in a manner that is dependent on
the culture and language differences. In direct relation to the work developed in this
thesis, in the different text types and languages, future lines of research could be:
1. In the case of review texts:
a. The automatic extraction of taxonomies for product features
b. The extension of the proposed framework for review annotation
and feature-based opinion mining and summarization for languages
other than English and Spanish
2. In the context of newspaper quotations and, in a more general manner,
newspaper articles:
a. The study of the impact of news source (i.e. in terms of bias,
reputation, trust) on the sentiment analysis process, within sources
from the same country/culture and across countries and cultures
b. The study of the influence that the reader background has on the
manner in which sentiment is perceived from newspaper articles
c. The automatic and semi-automatic extension of the resources we
built to other languages (i.e. in addition to the collection of
quotations we annotated for English and German)
d. The development of a framework for sentiment analysis that takes
into account the 3 proposed components of text the author, reader
and text and the manner in which they interact in the text meaning
negotiation, according to the Speech-Act and the Appraisal
Theories
e. The study of the influence of news content (i.e. what we denoted as
good versus bad news) on the manner in which sentiment is
expressed and subsequently on the performance of the automatic
sentiment analysis task
3. In the context of political debates and general texts:
a. Study the methods to represent the dialogue structure and the
influence of different methods for discourse analysis, as well as
tools (e.g. for anaphora resolution) on the proposed sentiment
analysis methods.
b. The study and use of topic modeling techniques in order to
accurately detect the target of the sentiment expressed, independent
of its nature being known beforehand

224

c. The study of affect-based argumentation techniques in order to


detect the use of emotion triggers in relation to the discussion topic
4. In the context of blogs:
a. The extension of the proposed method of annotation and labeled
corpora to other languages and similar text types (e.g. forums,
discussion boards)
b. The study and use of topic modeling techniques, as well as of the
impact of co-reference resolution at an inter-textual level
4. How can we deal with opinion in the context of traditional tasks? How can
we adapt traditional tasks (Information Retrieval, Question Answering, Text
Summarization) in the context of opinionated content? What are the new
challenges in this context?
Bearing in mind these necessities of real-world applications, which need, apart
from the sentiment analysis component, also text retrieval and text summarization
components, in Chapter 5 we proposed methods to combine opinion mining with
question answering and summarization. We showed that performing traditional
tasks in the context of opinionated text has many challenges and that systems that
were designed to work exclusively with factual data are not able to cope with
opinion questions. In this thesis, we proposed new methods and techniques to adapt
question answering and summarization systems to deal with opinionated content.
In the case of opinion question answering systems, future work includes the
development of a benchmark for opinion questions classification and the proposal
of adequate methods to tackle each type of opinion queries, in a monolingual,
multilingual and cross-lingual setting. Additionally, the framework for opinion
question answering should be extended with appropriate resources to other
languages. Further on, as we have seen from our experiments in the NTCIR 8
MOAT competition, there is an immediate need to include high-performing
methods for temporal expression resolution and anaphora resolution. Unfortunately,
due to the low performance of systems resolving these aspects, at this point the
influence they have on the opinion question answering systems performance is
negative. Another line of future work is the study of query expansion techniques
that are appropriate for opinionated content. From what we have seen in our
experiments, the use of a paraphrase collection that is not specifically designed for
the sentiment-bearing textual content leads to a drop in performance of the final
system.
In the case of sentiment summarization approaches, the lines for future work
include the development of appropriate topic-sentiment relevance detection
techniques and the study of qualitative measures for the evaluation of the opinion
summarization approaches.
225

The last research question we addressed in this thesis was:


5. Can we propose a model to detect emotion (as a component of sentiment)
from text, in the cases where it is expressed implicitly, requiring world
knowledge for its detection?
In Chapter 6 of the thesis, we proposed methods to detect implicit expressions of
emotion, from situations where the affective label is only inferable based on
commonsense knowledge (i.e. emotion is not explicitly, but implicitly expressed by
the author, by presenting situations which most people, based on commonsense
knowledge, their needs and motivations, associate with an emotion, like war,
terrorism, going to a party, seeing your child taking his/her first step etc.).
The initial approach we proposed is based on the idea that emotion is triggered by
specific concepts, according to their relevance, seen in relation to the basic needs
and motivations. Subsequently, we propose a framework for affect detection in text
based on the Appraisal Theory, modeling the possible interpretation (appraisal) of
the actions contained in the situation described, according to the commonsense
knowledge on the actions and the properties of their actors and objects. Future work
includes the extension of the created resources to other languages and the expansion
of the common-sense knowledge base of emotion eliciting situation with additional
external sources of knowledge. In this context, interesting lines for future work
could be:
! The development techniques for the personalization and adaptation of the
information content depending on the user context, through the
exploitation of the subjective content generated by the user, as well as the
analysis of the preferences expressed explicitly, through options in user
profiles and implicitly, through the connections in social networks,
contributions to forums, reviews, blogs, microblogs.
! The design, implementation, population and extension of a knowledge
base for user preference modeling, based on which opinion mining can be
performed in a personalized manner (i.e. depending on the users goals,
beliefs, opinions, views, feelings). In this context, the semantic search
capabilities can be improved and extended in the context of subjective
information, in a user-centric manner.

226

REFERENCES
1. [Agichtein and Gravano, 2000] Agichtein, D. and Gravano, L. (2000).
Snowball: Extracting Relations from Large Plain-Text Collections. In
Proceedings of the 5th ACM International Conference on Digital
Libraries (ACM DL) 2000: 85-94.
2. [Atkinson and Van der Goot, 2009] Atkinson, M. and Van der Goot, E.
(2009). Near real time information mining in multilingual news. In
Proceedings of the 18th International World Wide Web Conference,
2009: 1153-1154.
3. [Austin, 1976] Austin, J.L. (1976). How to do things with words.
Oxford et. al.: Oxford University Press.
4. [Balahur and Montoyo, 2008] Balahur, A. and Montoyo, A. Applying a
Culture Dependent Emotion Triggers Database for Text Valence and
Emotion Classification. In Proceedings of the AISB 2008 Convention
Communication, Interaction and Social Intelligence, 2008.
5. [Balahur and Montoyo, 2008a] Balahur, A. and Montoyo, A. (2008a).
Applying a culture dependent emotion triggers database for text valence
and emotion classification. In Procesamiento del Lenguaje Natural,
40(40).
6. [Balahur and Montoyo, 2008b] Balahur, A. and Montoyo, A. (2008b).
Building a recommender system using community level social filtering.
In 5th International Workshop on Natural Language and Cognitive
Science (NLPCS), 2008:32-41.
7. [Balahur and Montoyo, 2008c] Balahur, A. and Montoyo, A. (2008c).
Determining the semantic orientation of opinions on products - a
comparative analysis. In Procesamiento del Lenguaje Natural, 41(41).
8. [Balahur and Montoyo, 2008d] Balahur, A. and Montoyo, A. (2008d). A
feature-driven approach to opinion mining and classification. In
Proceedings of the NLPKE 2008.
9. [Balahur and Montoyo, 2008e] Balahur, A. and Montoyo, A. (2008e).
An incremental multilingual approach to forming a culture dependent
emotion triggers lexical database. In Proceedings of the Conference of
Terminology and Knowledge Engineering (TKE 2008).
10. [Balahur and Montoyo, 2008f] Balahur, A. and Montoyo, A. (2008f).
Multilingual feature-driven opinion extraction and summarization from
customer reviews .In Lecture Notes in Computer Science, 5039, NLDB,
2008:345-346.
11. [Balahur and Montoyo, 2009] Balahur, A. and Montoyo, A. (2009).
Semantic approaches to fine and coarse-grained feature-based opinion
227

mining. In Proceedings of the International Conference on Application


of Natural Language to Information Systems , NLDB 2009:142-153
12. [Balahur and Montoyo, 2010] Balahur, A.; Montoyo, A. OpAL:
Applying Opinion Mining Techniques for the Disambiguation of
Sentiment Ambiguous Adjectives in SemEval-2 Task 18. In
Proceedings of SemEval-2, the 5th International Workshop on Semantic
Evaluation, satellite workshop to ACL 2010.
13. [Balahur and Steinberger, 2009] Balahur, A. and Steinberger, R. (2009).
Rethinking Opinion Mining in Newspaper Articles: from Theory to
Practice and Back. In Proceedings of the first workshop on Opinion
Mining and Sentiment Analysis (WOMSA 2009).
14. [Balahur et al, 2010] Balahur, A., Boldrini, E., Montoyo, A., MartnezBarco, P. (2010). OpAL: a System for Mining Opinion from Text for
Business Applications. In Marta E. Zorrilla, Jose-Norberto Mazn,
scar Ferrndez, Irene Garrigs, Florian Daniel and Juan Trujillo (Eds.).
Business Intelligence Applications and the Web: Models, Systems and
Technologies, IGI Global Press.
15. [Balahur et al., 2008] Balahur, A., Lloret, E., Ferrndez, O., Montoyo,
A., Palomar, M., and Muoz, R. (2008). The DLSIUAES Teams
Participation in the TAC 2008 Tracks. In Proceedings of the Text
Analysis Conference .
16. [Balahur et al., 2009a] Balahur, A., Boldrini, E., Montoyo, A., and
Martnez-Barco, P. (2009). A comparative study of open domain and
opinion question answering systems for factual and opinionated queries.
In Proceedings of the Conference on Recent Advances in Natural
Language Processing 2009.
17. [Balahur et al., 2009b] Balahur, A., Boldrini, E., Montoyo, A., and
Martnez-Barco, P. (2009). Cross-topic opinion mining for real-time
human-computer interaction. In Proceedings of the Workshop on
Natural Language and Cognitive Science , NLPCS 2009: 13-22.
18. [Balahur et al., 2009c] Balahur, A., Boldrini, E., Montoyo, A., and
Martnez-Barco, P. (2009). Fact versus opinion question classification
and answering: Challenges and keys. In Proceedings of the International
Conference on Artificial Intelligence , ICAI 2009: 750-755.
19. [Balahur et al., 2009d] Balahur, A., Boldrini, E., Montoyo, A., and
Martnez-Barco, P. (2009). Opinion and generic question answering
systems: a performance analysis. In Proceedings of the ACL-IJCNLP
2009: 157-160.
20. [Balahur et al., 2009e] Balahur, A., Kozareva, Z., and Montoyo, A.
(2009). Determining the polarity and source of opinions expressed in
228

political debates. Lecture Notes in Computer Science, 5449,CICLing


2009:468-480.
21. [Balahur et al., 2009f] Balahur, A., Steinberger, R., Van der Goot, E.,
and Pouliquen, B. (2009). Opinion mining from newspaper quotations.
In Proceedings of the Workshop on Intelligent Analysis and Processing
of Web News Content, IAT Workshops,2009: 523-526.
22. [Balahur et al., 2009g] Balahur, A., Kabadjov, M., Steinberger, J.,
Steinberger, R., Montoyo, A. (2009). Summarizing Opinions in Blog
Threads. Proceedings of the 23rd Pacific Asia Conference on Language,
Information and Computation (PACLIC), pp. 606-613, Hong Kong.
23. [Balahur et al., 2009h] Balahur, A., Boldrini, E., Montoyo, A., MartnezBarco, P. (2009). Evaluating Question Answering Systems on Opinion
Queries. Proceedings of TSA 2009, 1st CIKM Workshop on TopicSentiment Analysis for Mass Opinion Measurement.
24. [Balahur et al., 2009i] Balahur, A., Lloret, E., Boldrini, E., Montoyo, A.,
Palomar, M., Martnez-Barco, P. (2009). Summarizing Threads in Blogs
Using Opinion Polarity. Proceedings of the EETTs 2009 Workshop,
satellite to the RANLP 2009 Conference.
25. [Balahur et al., 2010a] Balahur, A., Boldrini, E., Montoyo, A., MartnezBarco, P. (2010). Going Beyond Traditional QA Systems: Challenges
and Keys in Opinion Question Answering. In Proceedings of the 23rd
International Conference on Computational Linguistics (COLING
2010).
26. [Balahur et al., 2010b] Balahur, A., Boldrini, E., Montoyo, A., MartnezBarco, P. (2010). The OpAL System at NTCIR 8 MOAT. In
Proceedings of the 8th NTCIR Workshop Meeting on Evaluation of
Information Access Technologies: Information Retrieval, Question
Answering, and Cross-Lingual Information Access (NTCIR 8).
27. [Balahur et al., 2010c] Balahur, A., Boldrini, E., Montoyo, A., MartnezBarco, P. (2010). Opinion Question Answering: Towards a Unified
Approach. In Proceedings of the 19th European Conference on
Artificial Intelligence (ECAI 2010), 511-516.
28. [Balahur et al., 2010d] Balahur, A., Steinberger, R., Kabadjov, M.,
Zavarella, V., Van der Goot, E., Halkia, M., Pouliquen, B., Belyaeva,
J. (2010). Sentiment Analysis in the News. In Proceedings of the
7th International Conference on Language Resources and Evaluation
(LREC'2010), pp. 2216-2220. Valletta, Malta.
29. [Balahur et al., 2010e] Balahur, A., Kabadjov, M., Steinberger,
J. (2010). Exploiting Higher-level Semantic Information for the
Opinion-oriented Summarization of Blogs. International Journal of
229

Computational Linguistics and Applications, ISSN: 0976-0962, Vol. 1,


No. 1-2, pp. 45-59.
30. [Balahur et al., 2011a] Balahur, A., Hermida, J.M., Montoyo, A. (2011).
Detecting Emotions in Social Affective Situations Using the EmotiNet
Knowledge Base. In Lecture Notes in Computer Science Nr. 6677,
proceedings of the 8th International Symposium on Neural Networks
ISNN 2011.
31. [Balahur et al., 2011b] Balahur, A., Hermida, J.M., Montoyo, A. (2011).
EmotiNet: a Knowledge Base for Emotion Detection in Text Build on
the Appraisal Theories. In Lecture Notes in Computer Science (in press),
proceedings of NLDB 2011.
32. [Banea et al., 2008a] Banea, C., Mihalcea, R., and Wiebe, J. (2008a). A
bootstrapping method for building subjectivity lexicons for languages
with scarce resources. In Proceedings of the Conference on Language
Resources and Evaluations (LREC 2008), Maraakesh, Marocco.
33. [Banea et al., 2008b] Banea, C., Mihalcea, R., Wiebe, J., and Hassan, S.
(2008b). Multilingual subjectivity analysis using machine translation. In
Proceedings of the Conference on Empirical Methods in Natural
Language Processing (EMNLP 2008), 127-135, Honolulu,Hawaii.
34. [Banea et al., 2010] Banea, C., Mihalcea, R. and Wiebe, J. (2010).
Multilingual subjectivity: are more languages better? In Proceedings of
the International Conference on Computational Linguistics (COLING
2010), p. 28-36, Beijing, China,
35. [Banfield, 1982] Banfield, A. (1982). Unspeakable sentences: Narration
and Representation in the Language of Fiction. Routledge and Kegan
Paul.
36. [Bansal et al., 2008] Bansal, M., Cardie, C. and Lee, L. (2008). The
power of negative thinking: Exploiting label disagreement in the min-cut
classification framework. In Proceedings of the International
Conference on Computational Linguistics (COLING), 2008. Poster
paper, pp.15-18.
37. [Berland and Charniak, 1999] Berland, M. and Charniak, E. (1999).
Finding parts in very large corpora. In Proceedings of ACL, 1999,
College Park, Md, pp.57-64.
38. [Boldrini et al., 2009] Boldrini, E., Balahur, A., Martnez-Barco, P., and
Montoyo, (2009). Emotiblog: an annotation scheme for emotion
detection and analysis in non-traditional textual genres. In Proceedings
of the 5th International Conference on Data Mining (DMIN 2009), 491497.

230

39. [Boldrini et al., 2010] Boldrini, E., Balahur, A., Martnez-Barco, P.,
Montoyo, A. (2010). EmotiBlog: a finer-grained and more precise
learning of subjectivity expression models. In Proceedings of the
4th Linguistic Annotation Workshop (LAW IV), satellite workshop to
ACL 2010, pp. 1-10, Uppsala Sweden.
40. [Bossard et al., 2008] Bossard, A., Gnreux, M., and Poibeau, T.
(2008). Description of the LIPN systems at TAC 2008: Summarizing
information and opinions. In Proceedings of the Text Analysis
Conference of the National Institute for Standards and Technology,
Geithersbury, Maryland, USA.
41. [Breckler and Wiggins, 1992] Breckler, S. J. and Wiggins, E. C. (1992).
On defining attitude and attitude theory: Once more with feeling. In A.
R. Pratkanis, S. J. Breckler, and A. C. Greenwald (Eds.), Attitude
structure and function. Hillsdale, NJ, Erlbaum. pp. 407427.
42. [Brin, 1998] Brin, S. (1998). Extracting patterns and relations from the
World-Wide Web. In Proceedings of the 1998 International Workshop
on Web and Databases (WebDB98), 1998: 172-183.
43. [Calvo and DMello, 2010] Calvo, R. A. and DMello, S. (2010). Affect
Detection: An Interdisciplinary Review of Models, Methods and Their
Applications. IEEE Transactions on Affective Computing, Vol. 1, No. 1,
Jan.-Jun 2010, pp.18-37.
44. [Cambria et al., 2009] Cambria, E., Hussain, A., Havasi, C. and Eckl, C.
(2009). Affective Space: Blending Common Sense and Affective
Knowledge to Perform Emotive Reasoning. In Proceedings of the 1st
Workshop on Opinion Mining and Sentiment Analysis (WOMSA) 2009:
32-41.
45. [Cardie et al., 2004] Cardie, C., Wiebe, J., Wilson, T., Litman, D.
(2004). Low-Level Annotations and Summary Representations of
Opinions for Multiperspective QA. In Mark Maybury (ed), New
Directions in Question Answering , AAAI Press/MIT Press, pp.17-98.
46. [Cerini et al., 2007] Cerini, S., Compagnoni, V., Demontis, A.,
Formentelli, M., and Gandini, G. (2007). Micro-wnop: A gold standard
for the evaluation of auto-matically compiled lexical resources for
opinion mining, Milano, IT.
47. [Chaovalit and Zhou, 2005] Chaovalit, P. and Zhou, L. (2005). Movie
Review Mining: a Comparison between Supervised and Unsupervised
Classification Approaches. In Proceedings of HICSS-05, the 38th
Hawaii International Conference on System Sciences, IEEE Computer
Society.

231

48. [Chklovski and Pantel, 2004] Chklovski, T. and Pantel, P. (2004).


VerbOcean: Mining the Web for Fine-Grained Semantic Verb Relations.
In Proceedings of EMNLP, 2004: 33-40.
49. [Choi et al., 2005] Choi, Y., Cardie, C., Rilloff, E., Padwardhan, S.
(2005). Identifying Sources of Opinions with Conditional Random
Fields and Extraction Patterns. In Proceedings of the HLT/EMNLP.
50. [Cilibrasi and Vitanyi, 2006a] Cilibrasi, D. and Vitanyi, P. (2006a).
Automatic Meaning Discovery Using Google. IEEE Journal of
Transactions on Knowledge and Data Engineering.
51. [Conroy and Schlesinger, 2008] Conroy, J. and Schlesinger, S. (2008).
Classy at TAC 2008 metrics. In Proceedings of the Text Analysis
Conference of the National Institute for Standards and Technology.
52. [Cruz et al., 2008] Cruz, F., Troyano, J., Ortega, J., and Enrquez, F.
(2008). The Italica system at TAC 2008 opinion summarization task. In
Proceedings of the Text Analysis Conference of the National Institute for
Standards and Technology.
53. [Cui et al., 2006] Cui, H., Mittal, V., and Datar, M. (2006). Comparative
Experiments on Sentiment Classification for Online Product Reviews. In
Proceedings of the 21st National Conference on Artificial Intelligence
AAAI 2006, pp.1265-1270.
54. [Cytowic, 1989] Cytowic, R. (1989). Synesthesia. A Union of the
Senses, New York, Springer Verlag.
55. [Dagan et al., 2006] Dagan, I., Glickman, O. and Magnini, B. (2006).
The PASCAL Recognising Textual Entailment Challenge. In
Quionero-Candela, J.; Dagan, I.; Magnini, B.; d'Alch-Buc, F. (Eds.)
Machine Learning Challenges. Lecture Notes in Computer Science ,
Vol. 3944, pp. 177-190, Springer.
56. [Danisman and Alpkocak, 2008] Danisman, T. and Alpkocak, A.
(2008). Feeler: Emotion Classification of Text Using Vector Space
Model. In Proceedings of the AISB 2008 Convention, Communication,
Interaction and Social Intelligence, vol.2, pp.53-59, Aberdeen,UK.
57. [Das and Chen, 2001] Das, S. and Chen, M. (2001). Yahoo! for
Amazon: Extracting market sentiment from stock message boards. In
Proceedings of the Asia Pacific Finance Association Annual Conference
(APFA).
58. [Das and Chen, 2007] Das, S.R. and Chen, M. Y. (2007). Yahoo! for
Amazon: Sentiment extraction from small talk on the Web. Management
Science, 53(9), pp.13751388.
59. [Dave et al., 2003] Dave, K., Lawrence, S., and Pennock, D. (2003).
Mining the Peanut Gallery: Opinion Extraction and Semantic
232

Classification of Product Reviews. In Proceedings of WWW-2003, 519528.


60. [De Rivera, 1977] De Rivera, J. (1977). A structural theory of the
emotions. Psychological Issues, 10 (4), Monograph 40, New York,
International Universities Press.
61. [Deerwester et al., 1990] Deerwester, S., Dumais, S., Furnas, G. W.,
Landauer, T. K., and Harshman, R. (1990). Indexing by latent semantic
analysis. Journal of the American Society for Information Science,
3(41).
62. [Denecke, 2008] Denecke, K. (2008). Using SentiWordNet for
multilingual sentiment analysis. In Proceedings of the International
Conference on Data Engineering (ICDE 2008), Workshop on Data
Engineering for Blogs, Social Media, and Web 2.0.,2008: 430-435.
63. [Devitt and Ahmad, 2007] Devitt, A. and Ahmad, K. (2007). A lexicon
for polarity: Affective content in financial news text. In Proceedings of
the Conference on Language for Special Purposes, ACL 2007.
64. [Ding et al., 2008] Ding, X., Liu, B., and Yu, P. (2008). A holistic
lexicon-based approach to opinion mining. In Proceedings of WSDM
2008, New York, USA.
65. [Dyer, 1987] Dyer, M. (1987). Emotions and their computations: three
computer models. Cognition and Emotion, 1, 323-347.
66. [Edmonds and Hirst, 2002] Edmonds, P. and Graeme, H. (2002). Nearsynonymy and lexical choice. Computational Linguistics, 28(2), June
2002, 105144.
67. [Ekman, 1999] Ekman, P. (1999). Basic Emotions. In T. Dalgleish and
M. Power (Eds.). Handbook of Cognition and Emotion. Sussex, U.K.:
John Wiley & Sons, Ltd., 1999.
68. [Erkan and Radev, 2004] Erkan, G. and Radev, D. R. (2004). LexRank:
Graph-based centrality as salience in text summarization. Journal of
Artificial Intelligence Research (JAIR), 22,457-479.
69. [Esuli and Sebastiani, 2005] Esuli, A. and Sebastiani, F. (2005).
Determining the semantic orientation of terms through gloss analysis. In
Proceedings of CIKM 2005: 617-624, Bremen, Germany.
70. [Esuli and Sebastiani, 2006] Esuli, A. and Sebastiani, F. (2006).
Sentiwordnet: A publicly available resource for opinion mining. In
Proceedings of the 6th International Conference on Language Resources
and Evaluation, pp.417-422.
71. [Evans, 2001] Evans, D. (2001). Emotions. The science of sentiment.
Oxford: Oxford University Press.

233

72. [Fellbaum, 1999] Fellbaum, C. (1999). WordNet: An Electronic Lexical


Database (Language, Speech, and Communication). The MIT Press.
73. [Ferrndez et al., 1999] Ferrndez, A., Palomar, M., and Moreno, L.
(1999). An Empirical Approach to Spanish Anaphora Resolution.
Machine Translation. Special Issue on Anaphora Resolution In Machine
Translation, pages 191216.
74. [Frijda, 1986] Frijda, N. (1986). The emotions. Cambridge University
Press.
75. [Gamon et al., 2005] Gamon, M., Aue, S., Corston-Oliver, S., and
Ringger, E. (2005). Mining Customer Opinions from Free Text. Lecture
Notes in Computer Science, IDA 2005:121-132.
76. [Gamon, 2004] Gamon, M. (2004). Sentiment classification on customer
feedback data: Noisy data, large feature vectors, and the role of
linguistic analysis. In Proceedings of the International Conference on
Computational Linguistics (COLING) 2004, pp.841-847.
77. [Goldberg and Zhu, 2006] Goldberg, A. B. and Zhu, J. (2006). Seeing
stars when there arent many stars: Graph-based semi-supervised
learning for sentiment categorization. In HLT-NAACL 2006 Workshop
on Textgraphs: Graph-based Algorithms for Natural Language
Processing.
78. [Goldie, 2000] Goldie, P. (2000) Emotions. A Philosophical
Exploration. Oxford University Press.
79. [Goleman, 1995] Goleman, D. (1995). Emotional Intelligence. Bantam
Books, New York, USA.
80. [Gong and Liu, 2002] Gong, Y. and Liu, X. (2002). Generic text
summarization using relevance measure and latent semantic analysis. In
Proceedings of ACM SIGIR, 2002: 19-25, New Orleans, USA.
81. [Grninger and Fox, 1995] Grninger, M. and Fox, M. S. (1995).
Formal ontology in information systems. In Proceedings of the IJCAI
Workshop on Basic Ontological Issues in Knowledge Sharing, Montreal.
82. [Halliday, 1994] Halliday, M. A. K. (1994). An Introduction to
Functional Grammar. Edward Arnold, London,UK.
83. [Hatzivassiloglou and McKeown, 1997] Hatzivassiloglou, V. and
McKeown, K. (1997). Predicting the semantic orientation of adjectives.
In Proceedings of the 35th Annual Meeting of the ACL and the 8th
Conference of the European Chapter of the ACL, ACL 1997: 174-181,
Madrid, Spain.
84. [Hatzivassiloglou and Wiebe, 2000] Hatzivassiloglou, V. and Wiebe, J.
(2000). Effects of adjective orientation and gradability on sentence
subjectivity. In Proceedings of COLING 2000: 299-305..
234

85. [He et al., 2008] He, T., Chen, J., Gui, Z., and Li, F. (2008). CCNU at
TAC 2008: Proceeding on using semantic method for automated
summarization yield. In Proceedings of the Text Analysis Conference of
the National Institute for Standards and Technology.
86. [Hovy, 2005] Hovy, E. H. (2005). Automated text summarization. In
Mitkov, R., editor, The Oxford Handbook of Computational Linguistics,
pages 583598. Oxford University Press, Oxford, UK.
87. [Hu and Liu, 2004] Hu, M. and Liu, B. (2004). Mining Opinion Features
in Customer Reviews. In Proceedings of Nineteenth National
Conference on Artificial Intellgience AAAI-2004, San Jose, USA.
88. [Iftene and Balahur-Dobrescu, 2007] Iftene, A. and Balahur-Dobrescu,
A. (2007). Hypothesis transformation and semantic variability rules used
in recognizing textual entailment. In Proceedings of the ACL-PASCAL
Workshop on Textual Entailment and Paraphrasing 2007.
89. [Jiang and Vidal, 2006] Jiang, H. and Vidal, J. M. (2006). From rational
to emotional agents. In Proceedings of the AAAI Workshop on Cognitive
Modeling and Agent-based Social Simulation.
90. [Jijkoun et al., 2010] Jijkoun, V., de Rijke, M., Weerkamp, W. (2010).
Generating Focused Topic-Specific Sentiment Lexicons. ACL 2010, pp.
585-594.
91. [Johnson-Laird and Oatley, 1989] Johnson-Laird, P. N. and Oatley, K.
(1989). The language of emotions: An analysis of a semantic field.
Cognition and Emotion, 3, 81-123.
92. [Kabadjov et al, 2009a] Kabadjov, M. A., Steinberger, J., Pouliquen, B.,
Steinberger, R., and Poesio, M. (2009). Multilingual statistical news
summarisation: Preliminary experimentswith english. In Proceedings of
theWorkshop on Intelligent Analysis and Processing of Web News
Content at the IEEE / WIC / ACM International Conferenceson Web
Intelligence and Intelligent Agent Technology(WI-IAT), 2009:519-522.
93. [Kabajov et al., 2009] Kabadjov, M., Balahur, A., Boldrini, E. (2009).
Sentiment Intensity: Is It a Good Summary Indicator? Proceedings of
the 4th Language Technology Conference LTC, 2009: 203-212.
94. [Kim and Hovy, 2004] Kim, S.-M. and Hovy, E. (2004). Determining
the Sentiment of Opinions. In Proceedings of COLING 2004, pp.13671373, Geneva, Swizerland.
95. [Kim and Hovy, 2005] Kim, S.-M. and Hovy, E. (2005). Automatic
detection of opinion bearing words and sentences. In Companion
Volume to the Proceedings of the International Joint Conference on
Natural Language Processing (IJCNLP), Jeju Island, Korea.

235

96. [Kim and Hovy, 2006] Kim, S.-M. and Hovy, E. (2006). Automatic
identification of pro and con reasons in online reviews. In Proceedings
of the COLING/ACL Main Conference Poster Sessions, pp. 483490.
97. [Kim et al., 2010] Kim, J., Li, J.-J. and Lee, J.-H. (2010). Evaluating
Multilanguage-Comparability of Subjectivity Analysis Systems. In
Proceedings of the 48th Annual Meeting of the Association for
Computational Linguistics, pages 595603, Uppsala, Sweden, 11-16
July 2010.
98. [Kintsch, 1999] Kintsch, W. (1999). Comprehension: A paradigm for
cognition. New York, Cambridge University Press.
99. [Kintsch, 2000] Kintsch, W. (2000). Metaphor comprehension: A
computational theory. Psychonomic Bulletin and Review, 7, 527-566.
100. [Koppel and Shtrimberg, 2004] Koppel, M. and Shtrimberg, I.
(2004). Good or bad news? Let the market decide. In Proceedings of the
AAAI Spring Symposium on Exploring Attitude and Affect in Text:
Theories and Applicationns.
101. [Kouleykov and Magnini, 2006] Kouleykov, M. and Magnini, B.
(2006). Tree Edit Distance for Recognizing Textual Entailment:
Estimating the Cost of Insertion. In Proceedings of the Second PASCAL
Challenges
Workshop
on
Recognising
Textual
Entailment, Venice, Italy.
102. [Kozareva et al., 2007] Kozareva, Z., Vzquez, S., and Montoyo, A.
(2007). Discovering the Underlying Meanings and Categories of a
Name through Domain and Semantic Information. In Proceedings of the
Conference on Recent Advances in Natural Language Processing
(RANLP 2007), Borovetz, Bulgaria.
103. [Ku et al., 2005] Ku, L.-W., Li, L.-Y., Wu, T.-H., and Chen, H.-H.
(2005). Major topic detection and its application to opinion
summarization. In Proceedings of the ACM Special Interest Group on
Information Retrieval (SIGIR), Salvador, Brasil.
104. [Ku et al., 2006] Ku, L.-W., Liang, Y.-T., and Chen, H.-H. (2006).
Opinion extraction, summarization and tracking in news and blog
corpora. In AAAI Symposium on Computational Approaches to
Analysing Weblogs (AAAI CAAW 2006), AAAI Technical Report, SS06-03, Standford University, 100-107.
105. [Kudo and Matsumoto, 2004] Kudo, T. and Matsumoto, Y. (2004).
A boosting algorithm for classification of semistructured text. In
Proceedings of the Conference on Empirical Methods in Natural
Language Processing (EMNLP) 2004.

236

106. [Lappin and Leass, 1994] Lappin, S. and Leass, H.J. (1994). An
algorithm for pronominal anaphora resolution. In Journal of
Computational Linguistics, 20(4):535561.
107. [Larousse, 1995] Larousse (1995). Diccionario Ideolgico de la
Lengua Espaola. Larousse Editorial, RBA Promociones Editoriales,
S.L.
108. [Laver et al., 2003] Laver, M., Benoit, K., and Garry, J. (2003).
Extracting policy positions from political texts using words as data.
American Political Science Review , 97(2, May): 311-331.
109. [Lazarus and Smith, 1988] Lazarus, R. S. and Smith, C. A. (1988).
Knowledge and appraisal in the cognition-emotion relationship.
Cognition and Emotion, 2, 281-300.
110. [Lee, 2004] Lee, L. (2004). I'm sorry Dave, I'm afraid I can't do
that: Linguistics, Statistics, and Natural Language Processing circa
2001. In Computer Science: Reflections on the Field, Reflections from
the Field (Report of the National Academies' Study on the Fundamentals
of Computer Science), pp. 111-118, 2004.
111. [Li et al., 2008] Li, W., Ouyang, Y., Hu, Y., Wei, F. (2008). PolyU
at TAC 2008. In Proceedings of the Text Analysis Conference 2008,
National Institute for Science and Technology, Gaithersburg, Maryland,
USA .
112. [Li et al., 2008a] Li, F., Zheng, Z., Yang, T., Bu, F., Ge, R., Yan
Zhu, X., Zhang, X., and Huang, M. (2008a). THU QUANTA at TAC
2008 QA and RTE track. In Proceedings of the Text Analysis
Conference (TAC 2008).
113. [Li et al., 2008b] Li, W., Ouyang, Y., Hu, Y., and Wei, F. (2008b).
PolyU at TAC 2008. In Proceedings of the Text Analysis Conference
2008.
114. [Lin and Pantel, 2001] Lin, D. and Pantel, P. (2001). Discovery of
Inference Rules for Question Answering. In Journal of Natural
Language Engineering 7(4):343-360.
115. [Lin et al., 2006] Lin, W., Wilson, T., Wiebe, J., and Hauptman, A.
(2006). Which Side are You on? Identifying Perspectives at the
Document and Sentence Levels. In Proceedings of the Tenth Conference
on Natural Language Learning CoNLL06.
116. [Lin, 1998] Lin, D. (1998). Dependency-based Evaluation of
MINIPAR. In Workshop on the Evaluation of Parsing Systems,Granada,
Spain, May,1998.

237

117. [Lita et al., 2005] Lita, L., Schlaikjer, A., Hong, W., and Nyberg, E.
(2005). Qualitative dimensions in question answering: Extending the
definitional QA task. In Proceedings of AAAI 2005.
118. [Liu and Maes, 2007] Liu, H. and Maes, P. (2007). Introduction to
the semantics of people and culture, International Journal on Semantic
Web and Information Systems, Special Issue on Semantics of People and
Culture (Eds. H. Liu & P. Maes), 3(1), Hersey, PA: Idea Publishing
Group.
119. [Liu and Singh, 2004] Liu, H. and Singh, P. (2004). ConceptNet: A
Practical Commonsense Reasoning Toolkit. BT Technology
Journal, Volume 22, nr.4, pp.211-226, Kluwer Academic Publishers.
120. [Liu et al., 2003] Liu, H., Lieberman, H. and Selker, T. (2003). A
Model of Textual Affect Sensing Using Real-World Knowledge. In
Proceedings of IUI 2003.
121. [Liu, 2007] Liu, B. (2007). Web Data Mining. Exploring
Hyperlinks, Contents and Usage Data. Springer, first edition.
122. [Liu, 2010] Liu, B. (2010). Sentiment Analysis and
Subjectivity. In Handbook of Natural Language Processing, eds.
N.Indurkhya and F.J.Damenan, 2010.
123. [Lloret et al., 2008] Lloret, E., Ferrndez, O., Muoz, R., and
Palomar, M. (2008). A text summarization approach under the influence
of textual entailment. In Proceedings of the 5th International Workshop
on Natural Language Processing and Cognitive Science (NLPCS 2008).
124. [Lloret et al., 2009] Lloret, E., Balahur, A., Palomar, M., and
Montoyo, A. (2009). Towards building a competitive opinion
summarization system: Challenges and keys. In Proceedings of the
NAACL-HLT 2009 Student Research Workshop, 2009: 72-77.
125. [Macleod et al., 1994] Macleod, C., Grishman, R., and Meyers, A.
(1994). Creating a common syntactic dictionary of english. In
Proceedings of the International Workshop on Sharable Natural
Language Resources.
126. [Macleod et al., 1998] Macleod, C., Grishman, R., Meyers, A.,
Barrett, L., and Reeves, R. (1998). Nomlex: A lexicon of
nominalizations. In Proceedings of EURALEX 1998, Liege, Belgium.
127. [Martin and Vanberg, 2007] Martin, L. W. and Vanberg,G. (2007).
Coalition Government and Political Communication. In Political
Research Quarterly, September 2008, Vol. 61, No. 3, pp. 502-516.
128. [Martin and Vanberg, 2008] Martin, L. W. and Vanberg, G. (2008).
A robust transformation procedure for interpreting political text.
Political Analysis Advance Access, Oxford University Press, 16(1).
238

129. [Martin and White, 2005] Martin, J. R. and White, P. R. (2005).


Language of Evaluation: Appraisal in English. Palgrave Macmillan.
130. [Martn et al., 2010] Martn, T., Balahur, A., Montoyo, A., Pons,
A. (2010). Word Sense Disambiguation in Opinion Mining: Pros and
Cons. In Journal Research in Computing Science, Special Issue: Natural
Language Processing and its Applications, ISSN: 1870-4069, Vol. 46,
pp. 119-130.
131. [Martn-Wanton et al., 2010] Martn-Wanton, T., Pons-Porrata, A.,
Montoyo-Guijarro, A., Balahur, A. (2010). Opinion Polarity Detection
Using Word Sense Disambiguation to Determine the Polarity of
Opinions. In Proceedings of the 2ndInternational Conference on Agents
and Artificial Intelligence (ICAART 2010).
132. [Maslow, 1943] Maslow, A. (1943). A theory of human motivation.
Psychological Review, 1943, Vol. 50, pp. 370396.
133. [Masum Shaikh et al., 2007] Masum Shaikh, A., Prendinger, M.,
and Mitsuru, (2007). Assessing sentiment of text by semantic
dependency and contextual valence analysis. Lecture Notes in Computer
Science Nr. 4738, ACII, 2007: 737-738.
134. [Matsumoto et al., 2005] Matsumoto, S., Takamura, H., and
Okumura, M. (2005). Sentiment classification using word sub-sequences
and dependency sub-trees. In Proceedings of PAKDD05, the 9th
Pacific-Asia Conference on Advances in Knowledge Discovery and
Data Mining, PAKDD, 2005: 301-311.
135. [Matthiassen, 1995] Matthiassen, C. (1995). Lexico-grammatical
cartography: English systems. International Language Science
Publishers.
136. [Max-Neef, 1991] Max-Neef, M. A. (1991). Human scale
development: conception, application and further reflections. The Apex
Press, New York.
137. [McCarthy, 1959] McCarthy, J. Programs with Common Sense.
In Mechanisation of Thought Processes, Proceedings of the Symposium
of the National Physics Laboratory, pages 77-84, London, U.K., 1959.
Her Majesty's Stationery Office. Reprinted in John McCarthy.
Formalizing Common Sense: Papers by John McCarthy. Ablex
Publishing Corporation, 1990.
138. [McDonald et al., 2007] McDonald, R., Hannan, K., Neylon, T.,
Wells, M., Reynar, J. C. (2007). Structured Models for Fine-to-Coarse
Sentiment Analysis. In Proceedings of ACL 2007.

239

139. [Mei Lee et al., 2009] Mei Lee, S. Y., Chen, Y. and Huang, C.-R.
(2009). Cause Event Representations of Happiness and Surprise. In
Proceedings of PACLIC 2009, Hong Kong.
140. [Mihalcea et al., 2007] Mihalcea, R., Banea, C., and Wiebe, J.
(2007). Learning multilingual subjective language via cross-lingual
projections. In Proceedings of the Conference of the Annual Meeting of
the Association for Computational Linguistics 2007, pp.976-983,
Prague, Czech Republic.
141. [Moreda et al., 2007] Moreda, P., Navarro, B. and Palomar, M.
(2007). Corpus-based semantic role approach in information
retrieval, Data Knowledge English. (DKE) 61(3): 467-483.
142. [Moreda et al., 2008a] Moreda, P., Llorens, H., Saquete, E., and
Palomar, M. (2008a). Automatic generalization of a qa answer
extraction module based on semantic roles. In Proceedings of AAI IBERAMIA , vol. 5290, 2008: 233-242.
143. [Moreda et al., 2008b] Moreda, P., Llorens, H., Saquete, E., and
Palomar, M. (2008b). The influence of semantic roles in qa: a
comparative analysis. In Proceedings of the SEPLN 2008, vol.41, pp.5562, Madrid, Spain.
144. [Moreda, 2008] Moreda, P. (2008). Los Roles Semnticos en la
Tecnologa del Lenguaje Humano: Anotacin y Aplicacin. Doctoral
Thesis. University of Alicante.
145. [Mullen and Collier, 2004] Mullen, T. and Collier, M. (2004).
Sentiment Analysis Using Support Vector Machines with Diverse
Information Sources. In Proceedings of EMNLP 2004.
146. [Mullen and Malouf, 2006] Mullen, T. and Malouf, R. (2006). A
preliminary investigation into sentiment analysis of informal political
discourse. In Proceedings of the AAAI Symposium on Computational
Approaches to Analysing Weblogs (AAAI-CAAW), 2006: 159-162.
147. [Nasukawa and Yi, 2003] Nasukawa, T. and Yi, J. (2003).
Sentiment analysis: Capturing favorability using natural language
processing. In Proceedings of the Conference on Knowledge Capture
(K-CAP), 2003: 70-77.
148. [Neviarouskaya et al., 2010] Neviarouskaya, A., Prendinger, H. and
Ishizuka, M. (2010). EmoHeart: Conveying Emotions in Second Life
Based on Affect Sensing from Text. Advances in Human-Computer
Interaction, vol. 2010.
149. [Ng et al., 2006] Ng, V., Dasgupta, S., and Arifin, S.M. N. (2006).
Examining the Role of Linguistic Knowledge Sources in the Automatic
Identification and Classification of Reviews. In Proceedings 40th
240

Annual Meeting of the Association for Computational Linguistics, 2006:


611-618.
150. [Niu et al., 2005] Niu, Y., Zhu, X., Li, J. and Hirst, G. (2005).
Analysis of polarity information in medical text. In Proceedings of the
American Medical Informatics Association 2005 Annual Symposium.
151. [Oatley et al., 2006] Oatley, K., Keltner, D. and Jenkins, J. M.
(2006). Understanding Emotions. Wiley-Blackwell, 2nd edition.
152.
[Oatley, 2004] Oatley, K. (2004). Emotions. A Brief History.
Oxford: Wiley Blackwell.
153. [Ortony et al., 2005] Ortony, A., Norman, D., and Revelle, W.
(2005). Affect and pro-affect in effective functioning. In J.M. Fellous &
M.A. Arbib, Who needs emotions: The brain meets the machine. New
York: Oxford University Press.
154. [Ortony, 1997] Ortony, A. (1997). Metaphor and Thought.
Cambridge Press, 2nd edition.
155. [Ounis et al., 2007] Ounis, I., Macdonald, C., de Rijke, M., Mishne,
G., and Soboroff, I. (2007). Overview of the TREC 2006 blog track. In
The Fifteenth Text REtrieval Conference (TREC 2006), National
Institute for Science and Technogies Publication, NIST, Gaisthersbury,
MD, USA.
156. [Pang and Lee, 2003] Pang, B. and Lee, L. (2003). Seeing stars:
Exploiting class relationships for sentiment categorization with respect
to rating scales. In Proceedings of the 43rd Annual Meeting of the ACL,
2003: 115124.
157. [Pang and Lee, 2004] Pang, B. and Lee, L. (2004). A sentimental
education: Sentiment analysis using subjectivity summarization based
on minimum cuts. In Proceedings of the ACL 2004, pp.271-278,
Barcelona, Spain.
158. [Pang and Lee, 2008] Pang, B. and Lee, L. (2008). Opinion mining
and sentiment analysis. Foundations and Trends in Information
Retrieval, Vol 2, Nr. 1-2, 2008.
159. [Pang et al., 2002] Pang, B., Lee, L. and Vaithyanathan, S. (2002).
Thumbs up? Sentiment classification using machine learning techniques.
In Proceedings of EMNLP-02, 2002: 79-86.
160. [Pang et al., 2002] Pang, B., Lee, L., and Vaithyanathan, S. (2002).
Thumbs up? Sentiment classification using machine learning techniques.
In Proceedings of EMNLP-02, the Conference on Empirical Methods in
Natural Language Processing, 2002: 79-86, Philadelphia, PA, USA.

241

161. [Pantel and Ravichandran, 2004] Pantel, P. and Ravichandran, D.


(2004). Automatically Labeling Semantic Classes. In Proceedings of
HLT/NAACL-04, pp. 321-328. Boston.
162. [Parrot, 2001] Parrott, W. (2001). Emotions in Social Psychology.
Psychology Press, Philadelphia.
163. [Pennebaker et al., 2003] Pennebaker, J. W., Mehl, M. R. and
Niederhoffer, K. (2003). Psychological aspects of natural language use:
Our words, our selves. In Annual Review of Psychology 54, 547-577.
164. [Piao et al., 2007] Piao, S., Ananiadon, S., Tsuruoka, Y., Sasaki, Y.,
and McNaught, J. (2007). Mining opinion polarity of citations. In
Proceedings of the International Workshop on Computational Semantics
IWCS 2007.
165. [Picard, 1995] Picard, R. (1995). Affective computing. Technical
report, MIT Media Laboratory.
166. [Picard, 1997] Picard, R. (1997). Affective Computing. MIT Press.
167. [Platt, 1998] Platt, J. (1998). Sequential minimal optimization: A
fast algorithm for training support vector machines. Microsoft Research
Technical Report MSR-TR-98-14.
168. [Plutchik, 2001] Plutchik, R. (2001). The Nature of Emotions.
American Scientist. 89, 344.
169. [Polanyi and Zaenen, 2004] Polanyi, L. and Zaenen, A. (2004).
Exploring attitude and affect in text: Theories and applications.
Technical Report SS-04-07.
170. [Popescu and Etzioni, 2005] Popescu, A. M. and Etzioni, O. (2005).
Extracting product features and opinions from reviews. In In
Proceedings of HLT-EMNLP 2005, Vancouver, Canada.
171. [Pouliquen et al., 2007] Pouliquen, B., Steinberger, R., and Best, C.
(2007). Automatic detection of quotations in multilingual news. In
Proceedings of the International Conference Recent Advances in
Natural Language Processing (RANLP 2007), pp.25-32, Borovets,
Bulgaria.
172. [Pustejovsky and Wiebe, 2005] Pustejovsky, J. and Wiebe, J.
(2005). Introduction to special issue on advances in question answering.
Language Resources and Evaluation (2005), (39).
173. [Quirk, 1985] Quirk, R. (1985). A Comprehensive Grammar of the
English Language. Longman, London, UK.
174. [Ramchurn, 2004] Ramchurn, S. (2004). Multi-Agent Negotiation
Using Trust and Persuasion. Ph.D Thesis, University of Southhampton,
UK.

242

175. [Ratner, 2000] Ratner, C. (2000). A cultural-psychological analysis


of emotions. Culture and Psychology, vol.(6), pp.5-39.
176. [Riloff and Wiebe, 2003] Riloff, E. and Wiebe, J. (2003). Learning
extraction patterns for subjective expressions. In Proceedings of the
2003 Conference on Empirical Methods in Natural Language
Processing, EMNLP 2003, pp.105-112.
177. [Riloff et al., 2003] Riloff, E., Wiebe, J., and Wilson, T. (2003).
Learning subjective nouns using extraction pattern bootstrapping. In
Proceedings of the Conference on Natural Language Learning (CoNLL)
2003, pp.25-32, Edmonton, Canada.
178. [Riloff et al., 2005] Riloff, E., Wiebe, J., and Phillips, W. (2005).
Exploiting subjectivity classification to improve information extraction.
In Proceedings of the AAAI 2005.
179. [Rumbell et al., 2008] Rumbell, T., Barnden, J., Lee, M., and
Wallington, A. (2008). Affect in metaphor: Developments with wordnet.
In Proceedings of AISB Symposium of Affect in Human and Machine.
180. [Saggion and Funk, 2010] Saggion, H., Funk, A. (2010).
Interpreting SentiWordNet for opinion classification. In Proceedings of
LREC 2010, RTF, Tagged XML, Bib Tex Google Scholar.
181.
[Saggion et al., 2010] Saggion, H., Lloret, E., Palomar, M. (2010).
Using text summaries for predicting rating scales. In: Proceedings of the
1st Workshop on Subjectivity and Sentiment Analysis (WASSA
2010),pp.44-51.
182. [Saquete et al., 2006] Saquete, E., Ferrndez, O., Martnez-Barco,
P. and Muoz, R. (2006). Reconocimiento temporal para el italiano
combinando tcnicas de aprendizaje automtico y adquisicin
automtica de conocimiento. In Proceedings of the 22nd International
Conference of the Spanish Society for the Natural Language Processing
(SEPLN 2006), pp. 161-168.
183. [Scherer and Walbott, 1997] Scherer, K. and Wallbott, H. (1997).
The ISEAR Questionnaire and Codebook. Geneva Emotion Research
Group.
184. [Scherer, 1987] Scherer, K. (1987). Toward a dynamic theory of
emotion. The component process of affective states. Cognition and
Emotion, 1(1), 1-98.
185. [Scherer, 1989] Scherer, K. (1989). Appraisal Theory. Handbook of
Cognition and Emotion. John Wiley & Sons Ltd.
186. [Scherer, 1993] Scherer, K. R. (1993). Studying the EmotionAntecedent Appraisal Process: An Expert System Approach. Cognition
and Emotion, 7 (3/4), 323-355.
243

187. [Scherer, 2001] Scherer, K. (2001). Emotional expression is subject


to social and technological change: Extrapolating to the future. Social
Science Information, 40(1), 125-151.
188. [Scherer, 2005] Scherer, K. (2005). What are emotions? and how
can they be measured? Social Science Information, 3(44), 695-729.
189. [Shen et al., 2007] Shen, D., Wiegand, M., Merkel, A., Kazalski, S.,
Hunsicker, S., Leidner, J. L., and Klakow, D. (2007). The Alyssa
System at TREC QA 2007: Do we need blog06? In Proceedings of
TREC 2007.
190. [Solomon, 2005] Solomon, R. C. (2005). Subjectivity. In
Honderich, Ted. Oxford Companion to Philosophy, Oxford University
Press, p.900.
191. [Somasundaran et al., 2007] Somasundaran, S., Wilson, T., Wiebe,
J., and Stoyanov, V. (2007). Qa with attitude: Exploiting opinion type
analysis for improving question answering in on-line discussions and the
news. In Proceedings of the International Conference on Weblogs and
Social Media (ICWSM 2007), Boulder< Colorado, USA.
192. [Sperber and Wilson, 1995] Sperber, D. and Wilson, D. (1995).
Relevance: Communication and cognition (2nd ed.) Oxford: Blackwell.
Also: Sperber, D. and Wilson, D. (2004) "Relevance Theory" in G.
Ward and L. Horn (eds) Handbook of Pragmatics. Oxford: Blackwell,
607-632.
193. [Steinberger and Jeek, 2004] Steinberger, J. and Jeek, K.(2004).
Text summarization and singular value decomposition. In Lecture Notes
in Computer Science, Nr. 3261, pp. 245-249, Proceedings of the 3rd
Advances in Information Systems (ADVIS) conference, Izmir, Turkey.
194. [Steinberger and Jeek, 2009] Steinberger, J. and Jeek, K.(2009).
Update summarization based on novel topic distribution. In Proceedings
of the 9th ACM DocEng, Munich, Germany.
195.
[Steinberger et al., 2007] Steinberger, J., Poesio, M., Kabadjov, M.
A., and Jeek, K. (2007). Two uses of anaphora resolution in
Summmarization. Information Processing and Management,
43(6):16631680. Special Issue on Text Summarisation (Donna
Harman, ed.).
196. [Stevenson, 1963] Stevenson, C. (1963). Facts and Values: Studies
in Ethical Analysis. Yale University Press, New Haven, USA.
197. [Stone et al., 1966] Stone, P., Dumphy, D. C., Smith, M. S., and
Ogilvie, D. M. (1966). The General Inquirer: A Computer Approach to
Content Analysis. The MIT Press, Cambridge, USA.

244

198. [Stoyanov and Cardie, 2006] Stoyanov, V. and Cardie, C. (2006).


Toward opinion summarization: Linking the sources. In Proceedingns of
the COLING-ACL 2006 Workshop on Sentiment and Subjectivity in
Text.
199. [Stoyanov et al., 2004] Stoyanov, V., Cardie, C., Lawrence, D. and
Wiebe, J. (2004). Evaluating an opinion annotation scheme using a new
multi-perspective question and answer corpus. In AAAI Spring
Symposium on Exploring Attitude and Affect in Text: Theories and
Applications, AAIT Press.
200. [Stoyanov et al., 2005] Stoyanov, V., Cardie, C., and Wiebe, J.
(2005). Multiperspective question answering using the opqa corpus. In
Proceedings of the Human Language Technology Conference and the
Conference on Empirical Methods in Natural Language Processing
(HLT/EMNLP 2005), pp.923-930, Vancouver, Canada.
201. [Strapparava and Mihalcea, 2007] Strapparava, C. and Mihalcea, R.
(2007). Semeval 2007 task 14: Affective text. In Proceedings of ACL
2007, Prague, Czech Republic.
202.
[Strapparava and Valitutti, 2004] Strapparava, C. and Valitutti, A.
(2004). Wordnet-affect: an affective extension of wordnet. In
Proceedings of the 4th International Conference on Language
Resources and Evaluation (LREC 2004), pp.1083-1086, Lisbon,
Portugal.
203. [Studer et al., 1998] Studer, R., Benjamins, R. V. and Fensel, D.
(1998). Knowledge engineering: Principles and methods. Data and
Knowledge Engineering, 25(1-2):161197.
204. [Subasic and Huettner, 2000] Subasic, P. and Huettner, A. (2000).
Affect Analysis of text using fuzzy semantic typing. IEEE Transactions
on Fuzzy System, 9, 483-496.
205. [Suchanek et al., 2007] Suchanek, F., Kasnei, G. and Weikum, G.
(2007). YAGO: A Core of Semantic Knowledge Unifying WordNet and
Wikipedia. In Proceedings of WWW 2007, pp.697-706, Banff, Alberto,
USA.
206. [Taboada and Grieve, 2004] Taboada, M. and Grieve, J. (2004)
Analyzing Appraisal Automatically. American Association for Artificial
Intelligence Spring Symposium on Exploring Attitude and Affect in Text.
Stanford. March 2004. AAAI Technical Report SS-04-07, pp.158-161.
207. [Tan and Zhang, 2007] Tan, S. and Zhang, J. (2007). An empirical
study of sentiment analysis for chinese documents. In Expert Systems
with Applications.

245

208. [Tanev et al., 2009] Tanev, H., Pouliquen, B., Zavarella, B., and
Steinberger, R. (2009). Automatic expansion of a social network using
sentiment analysis, Annals of Information Systems, 2009, Springer,
Verlag.
209. [Tanev, 2007] Tanev, H. (2007). Unsupervised learning of social
networks from a multiple-source news corpus. In Proceedings of the
Workshop Multi-source Multilingual Information Extraction and
Summarization (MMIES 2007), held at RANLP 2007, Borovets,
Bulgaria.
210. [Terveen et al., 1997] Terveen, L., Hill, W., Amento, B.,
McDonald, D. and Creter, J. (1997). PHOAKS: A system for sharing
recommendations. Communications of the Association for Computing
Machinery (CACM), 40(3):5962, 1997.
211. [Teufel and Moens, 2000] Teufel, S. and Moens, M. (2000). Whats
yours and whats mine: Determining intellectual attribution in scientific
texts. In Proceedings of the 2000 Joint SIGDAT conference on
Empirical methods in natural language processing and very large
corpora: held in conjunction with the 38th Annual Meeting of the
Association for Computational Linguistics, Hong Kong.
212. [Thayer, 2001] Thayer, R. E. (2001). Calm Energy: How People
Regulate Mood With Food and Exercise. Oxford University Press, N.Y.
213. [Thomas et al., 2006] Thomas, M., Pang, B., and Lee, L. (2006).
Get out the vote: Determining support or opposition from congressional
floor-debate transcripts. In Proceedings of EMNLP 2006, pp,327-335.
214. [Tong, 2001] Tong, R.M. (2001). An operational system for
detecting and tracking opinions in on-line discussion. In Proceedings of
the Workshop on Operational Text Classification (OTC), 2001, New
Orleans, USA.
215. [Toprak et al., 2010] Toprak, C., Jakob, N. and Gurevych, I. (2010).
Sentence and Expression Level Annotation of Opinions in UserGenerated Discourse. In Proceedings of ACL 2010: 575-584.
216. [Turing, 1950] Turing, A. (1950). Computing machinery and
intelligence. Mind, 1950.
217. [Turney and Littman, 2003] Turney, P. and Littman, M. (2003).
Measuring praise and criticism: Inference of semantic orientation from
association. ACM Transactions on Information Systems, 21(4), 315-346.
218. [Turney, 2002] Turney, P. (2002). Thumbs up or thumbs down?
Semantic orientation applied to unsupervised classification of reviews.
In Proceedings 40th Annual Meeting of the Association for
Computational Linguistics, ACL 2002, 417-424, Philadelphia, USA.
246

219. [Uspensky, 1973] Uspensky, B. (1973). A Poetics of Composition.


University of California Press, Berkeley, California.
220. [Van den Bos, 2006] Van den Bos, G. (2006). APA Dictionary of
Psychology. Washington, DC: American Psychological Association.
221. [Varma et al., 2008] Varma, V., Pingali, P., Katragadda, R., Krisha,
S., Ganesh, S., Sarvabhotla, K., Garapati, H., Gopisetty, H., Reddy, V.,
Bysani, P., and Bharadwaj, R. (2008). Iit hyderabad at tac 2008. In
Proceedings of the Text Analysis Conference (TAC) 2008.
222. [Vzquez et al., 2004] Vzquez, S., Montoyo, A., and Rigau, G.
(2004). Using relevant domains resource for word sense disambiguation.
In Proceedings of the International Conference on Artificial
Intelligence, IC-AI, 2004: 784-789.
223. [Whitelaw et al., 2005] Whitelaw, C., Garg, N., and Argamon, S.
(2005). Using appraisal groups for sentiment analysis. In Proceedings of
the CIKM 2005: 625-631, Bremen, Germany.
224. [Wiberg, 2004] Wiberg, M.(ed) (2004). The Interaction Society:
Theories, Practice and Supportive Technologies. Information Science
Publishing, Idea Group Inc.
225. [Wiebe and Mihalcea, 2006] Wiebe, J. and Mihalcea, R. (2006).
Word sense and subjectivity. In Proceedings of ACL 2006.
226. [Wiebe and Riloff, 2005] Wiebe, J. and Riloff, E. (2005). Creating
Subjective and Objective Sentence Classifiers from Unannotated Texts.
In Proceedings of the 6th International Conference on Computational
Linguistics and Intelligent Text Processing (CICLing-05), 73-99.
227. [Wiebe and Wilson, 2005] Wiebe, J. and Wilson, T. (2005).
Annotating attribution and private states. In Proceedings of the ACL
Workshop on Frontiers in Corpus Annotation II: Pie in the Sky.
228. [Wiebe et al., 2001] Wiebe, J., Bruce, R. , Bell, M., Martin, M. and
Wilson, T. (2001). A Corpus Study of Evaluative and Speculative
Language. In Proceedings of the 2nd ACL SIGdial Workshop on
Discourse and Dialogue (SIGdial-2001). Aalborg, Denmark, pp.186
195.
229. [Wiebe et al., 2004] Wiebe, J., Wilson, T., Bruce, R., Bell, M., and
Martin, M. (2004). Learning subjective language, Computational
Linguistics, 30(3), 277-308.
230. [Wiebe et al., 2005] Wiebe, J., Wilson, T., and Cardie, C. (2005).
Annotating expressions of opinions and emotions in language. In
Language Resources and Evaluation, vol. 39(2-3),119-122.
231. [Wiebe, 1994] Wiebe, J. (1994). Tracking point of view in
narrative. Computational Linguistics, 20(2), 233-287..
247

232. [Wiegand et al., 2010] Wiegand, M., Balahur, A., Roth, B.,
Klakow, D., Montoyo, A. (2010). A Survey on the Role of Negation in
Sentiment Analysis. In Proceedings of the Workshop on Negation and
Speculation in Natural Language Processing, NeSp-NLP 2010.
233. [Wilson and Wiebe, 2003] Wilson, T. and Wiebe, J. (2003).
Annotating opinions in the world press. In Proceedings of SIGdial 2003.
234. [Wilson et al., 2004] Wilson, T., Wiebe, J., and Hwa, R. (2004).
Just how mad are you? Finding strong and weak opinion clauses. In
Proceedings of AAAI 2004,761-769, San Jose, USA.
235. [Wilson et al., 2005] Wilson, T., Wiebe, J., and Hoffmann, P.
(2005). Recognizing contextual polarity in phrase-level sentiment
analysis. In Proceedings of HLT-EMNLP 2005, pp.347-354, Vancouver,
Canada.
236. [Wu and Jin, 2010] Wu, Y., Jin, P. (2010). SemEval-2010 Task 18:
Disambiguating Sentiment Ambiguous Adjectives. In Proceedings of
the 5th International Workshop on Semantic Evaluation, ACL 2010,
pages 8185, Uppsala, Sweden, 15-16 July 2010.
237. [Yi et al., 2003] Yi, J., Nasukawa, T., Bunescu, R. and Niblack, W.
(2003). Sentiment analyzer: Extracting sentiments about a given topic
using natural language processing techniques. In Proceedings of the
IEEE International Conference on Data Mining (ICDM). 2003: 427434.
238. [Yu and Hatzivassiloglou, 2003] Yu, D. and Hatzivassiloglou, V.
(2003). Towards answering opinion questions: Separating facts from
opinions and identifying the polarity of opinion sentences. In
Proceedings of the Conference on Empirical Methods in Natural
Language Processing (EMNLP), 2003: 129-136, Shapporo, Japan.

248

ANNEX A. SCIENTIFIC CONTRIBUTIONS AND


MERITS DURING THE PHD STUDI ES
1. SCIENTIFIC PUBLICATIONS:
BOOK CHAPTERS:
1. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. OpAL: a System
for Mining Opinion from Text for Business Applications. In Marta E.
Zorrilla, Jose-Norberto Mazn, scar Ferrndez, Irene Garrigs, Florian
Daniel and Juan Trujillo (Eds.). Business Intelligence Applications and the
Web: Models, Systems and Technologies, IGI Global Press, 2010 (in press).
2. Balahur A.; Hermida, J. Lost in Translation: The Impact of New
Technologies on the Education Gap. In Balahur, D., Qvarsel, B. (Eds.)
Childrens Rights to Education and Information in a Globalized World.
Alexandru Ioan Cuza University Press, pp. 103-116, ISBN 978-973-703374-1, 2008.

JOURNAL ARTICLES:
1. Balahur, A.; Hermida J.M.; Montoyo, A. Detecting Implicit Expressions of
Emotion Using Commonsense Knowledge and Appraisal Models. IEEE
Transactions on Affective Computing. Special Issue on Naturalistic Affect
Resources for System Building and Evaluation, 2011 (accepted).
2. Balahur, A.; Hermida J.M.; Montoyo, A. Building and Exploiting EmotiNet:
a Knowledge Base for Emotion Detection Based on the Appraisal Theory
Model. In Lecture Notes in Computer Science Nr. 6677, 2011 - Proceedings
of the 9th International Symposium on Neural Networks (in press).
3. Balahur, A.; Hermida J.M.; Montoyo, A. EmotiNet: a Knowledge Base for
Emotion Detection in Text Built on the Appraisal Theories. In Lecture Notes
in Computer Science Proceedings of the 16th International Conference on
the Applications of Natural Language to Information Systems, 2011 (in
press).
4. Balahur, A.; Kabadjov, M.; Steinberger, J.; Steinberger, R.; Montoyo, A.
Opinion Summarization for Mass Opinion Estimation. Journal of Intelligent
Information Systems, 2011 (accepted).
5. Kabadjov, M.; Balahur, A.; Boldrini, E. Sentimtent Intensity: Is It a Good
Summary Indicator? In Lecture Notes in Artificial Intelligence, Nr. 6562
249

6.

7.

8.

9.

10.

11.

12.

Selected Papers from the 4th Language Technology Conference 2009 (in
press).
Balahur, A., Kabadjov, M. Steinberger, J. Exploiting Higher-level Semantic
Information for the Opinion-oriented Summarization of Blogs. International
Journal of Computational Linguistics and Applications, ISSN: 0976-0962,
Vol. 1, No. 1-2, pp. 45-59, 2010.
Martn, T.; Balahur, A.; Montoyo, A.; Pons, A. Word Sense Disambiguation
in Opinion Mining: Pros and Cons. In Journal Research in Computing
Science, Special Issue: Natural Language Processing and its Applications,
ISSN: 1870-4069, Vol. 46, pp. 119-130.
Balahur, A.; Kozareva, Z.; Montoyo, A. Determining the Polarity and
Source of Opinions Expressed in Political Debates. In Proceedings of the
10th International Conference on Intelligent Text Processing and
Computational Linguistics (CICLing) 2009 (Lecture Notes in Computer
Science No. 5449, 2009).
Balahur, A.; Montoyo, A. A Semantic Relatedness Approach to Classifying
Opinion from Web Reviews. In Procesamiento del Lenguaje Natural No. 42,
2009.
Balahur, A; Balahur, P. What Does the World Think About You? Opinion
Mining and Sentiment Analysis in the Social Web. In Analele Stiintifice ale
Universitatii Al.I. Cuza Iasi, Nr.2, pp. 101-110, 2009, ISSN: 2065-2917.
Balahur, A.; Montoyo, A. Definicin de disparador de emocin asociado a
la cultura y aplicacin a la clasificacin de la valencia y de la emocin en
textos. Procesamiento del Lenguaje Natural, Vol: 40, Num: 40, 2008.
Balahur, A.; Montoyo, A. Multilingual Feature-Driven Opinion Extraction
and Summarization from Customer Reviews. Journal Lecture Notes in
Computer Science, 5029, pp. 345-346 Proceedings of the 13th International
Conference on Applications of Natural Language to Information Systems
(NLDB 2008).

ARTICLES IN CONFERENCE PROCEEDINGS:


1. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. Going Beyond
Traditional QA Systems: Challenges and Keys in Opinion Question
Answering. In Proceedings of the 23rd International Conference on
Computational Linguistics (COLING 2010).
2. Wiegand, M.; Balahur, A., Roth, B.; Klakow, D., Montoyo, A. A Survey on
the Role of Negation in Sentiment Analysis. In Proceedings of the Workshop
on Negation and Speculation in Natural Language Processing, NeSp-NLP
2010.
250

3. Balahur, A.; Montoyo, A. OpAL: Applying Opinion Mining Techniques for


the Disambiguation of Sentiment Ambiguous Adjectives in SemEval-2 Task
18. In Proceedings of SemEval-2, the 5th International Workshop on
Semantic Evaluation, satellite workshop to ACL 2010.
4. Boldrini, E.; Balahur, A., Martnez-Barco, P.; Montoyo, A. EmotiBlog: a
finer-grained and more precise learning of subjectivity expression models.
In Proceedings of the 4th Linguistic Annotation Workshop (LAW IV),
satellite workshop to ACL 2010.
5. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. The OpAL
System at NTCIR 8 MOAT. In Proceedings of the 8th NTCIR Workshop
Meeting on Evaluation of Information Access Technologies: Information
Retrieval, Question Answering, and Cross-Lingual Information Access
(NTCIR 8), 2010.
6. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. Opinion
Question Answering: Towards a Unified Approach. In Proceedings of the
19th European Conference on Artificial Intelligence (ECAI 2010).
7. Balahur, A.; Steinberger, R.; Kabadjov, M.; Zavarella, V.; Van der Goot, E.;
Halkia, M.; Pouliquen, B.; Belyaeva, J. Sentiment Analysis in the News. In:
Proceedings of the 7th International Conference on Language Resources and
Evaluation (LREC'2010), pp. 2216-2220. Valletta, Malta, 19-21 May 2010.
8. Martn-Wanton, T.; Pons-Porrata, A.; Montoyo-Guijarro, A.; Balahur,
A. Opinion Polarity Detection Using Word Sense Disambiguation to
Determine the Polarity of Opinions. In Proceedings of the 2ndInternational
Conference on Agents and Artificial Intelligence (ICAART 2010).
9. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. Opinion and
Generic Question Answering systems: a Performance Analysis. In
Proceedings of the ACL-IJCNLP 2009 Conference.
10. Balahur, A.; Kabadjov, M.;Steinberger, J.; Steinberger, R.; Montoyo,
A. Summarizing Opinions in Blog Threads. Proceedings of the 23rd Pacific
Asia Conference on Language, Information and Computation (PACLIC), pp.
606-613, Hong Kong, 3-5 December 2009.
11. Kabadjov, M.; Balahur, A.; Boldrini, E. Sentiment Intensity: Is It a Good
Summary Indicator?.Proceedings of the 4th Language Technology
Conference LTC, pp. 380-384. Poznan, Poland, 6-8.11.2009.
12. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. Evaluating
Question Answering Systems on Opinin Queries. Proceedings of TSA 2009
(The First CIKM Workshop on Topic-Sentiment Analysis for Mass Opinion
Measurement).

251

13. Balahur, A.; Steinberger, R. Rethinking Opinion Mining in Newspaper


Articles: from Theory to Practice and Back. Proceedings of the first
workshop on Opinion Mining and Sentiment Analysis (WOMSA 2009).
14. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. A comparative
study of Open Domain and Opinion Question Answering Systems for
Factual and Opinionated Queries.RANLP 2009.
15. Balahur, A.; Lloret, E.; Boldrini, E.; Montoyo, A.; Palomar, M; MartnezBarco,
P. Summarizing
Threads
in
Blogs
Using
Opinion
Polarity. Proceedings of the EETTs 2009 Workshop, satellite to the RANLP
2009 Conference.
16. Balahur, A.; Steinberger, R.; van der Goot, E.; Pouliquen, B.; Kabadjov,
M. Opinion Mining from Newspaper Quotations. In Proceedings of the
Workshop on Intelligent Analysis and Processing of Web News Content,
2009 IEEE/WIC/ACM International Conference on Web Intelligence held in
conjunction with IAT09, September 2009, Milan, Italy.
17. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. A Comparative
Study of Open Domain and Opinion Question Answering Systems for
Factual and Opinionated Queries. In Proceedings of the Conference on
Recent Advances in Natural Language Processing (RANLP 2009).
18. Balahur, A.; Montoyo, A. Semantic Approaches to Fine and CoarseGrained Feature-Based Opinion Mining. In Proceedings of the
14th International Conference on Applications of Natural Language to
Information Systems (NLDB 2009).
19. Boldrini, E.; Balahur, A., Martnez-Barco, P.; Montoyo, A. EmotiBlog: an
Annotation Scheme for Emotion Detection and Analysis in Non-traditional
Textual Genres. In Proceedings of the 5th International Conference on Data
Mining (DMIN 2009).
20. Lloret, E.; Balahur, A.; Palomar, M.; Montoyo, A. Towards Building a
Competitive Opinion Summarization System: Challenges and Keys. In
Proceedings of the NAACL-HLT 2009 Student Research Workshop, 2009.
21. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. Fact Versus
Opinion Questions Classification and Answering: Challenges and Keys. In
Proceedings of the International Conference on Artificial Intelligence (ICAI
2009).
22. Balahur, A.; Boldrini, E.; Montoyo, A.; Martnez-Barco, P. Cross Topic
Opinion Mining for Real-time Human-Computer Interaction (Application of
EmotiBlog an Annotation Schema for Emotion Detection). In Proceedings
of the Workshop on Natural Language Processing and Cognitive Science
(NLPCS 2009).

252

23. Balahur, A.; Lloret, E.; Ferrndez, O.;Montoyo, A., Palomar, M.; Muoz,
R. The DLSIUAES Teams Participation in the TAC 2008 Tracks
Proceedings of the Text Analysis Conference 2008 Workshop, Washington,
USA, 2008.
24. Iftene, A.; Balahur, A. A Language Independent Approach for Recognizing
Textual Entailment. Research in Computing Science , Vol: 334, 2008.
25. Balahur, A.; Montoyo, A. Applying a Culture Dependent Emotion Triggers
Database for Text Valence and Emotion Classification. In Proceedings of
the AISB 2008 Symposium on Affective Language in Human and Machine,
Aberdeen, Scotland, 2008.
26. Balahur, A.; Montoyo, A. Building a Recommender System Using
Community Level Social Filtering. In Proceedings of the 5th International
Workshop on Natural Language and Cognitive Science, Barcelona, Spain,
2008.

2. PARTICIPATION IN SCIENTIFIC COMPETITIONS:


1. Participation in the competition SemEval 2010, task 18 Disambiguation
of sentiment ambiguous adjectives (in Traditional Chinese). ( Rank: 5/16)
2. Participation in the competition NTCIR 8 MOAT (Multilingual Opinion
Analysis Task), English mono-lingual task and English-Chinese crosslingual task. (Rank: 1/6 cross-lingual)
3. Participation in the competition Opinion Pilot in the Text Analysis
Conference (TAC) 2008. (5/18 in F-measure)
4. Participation in the competition Answer Validation Exercise (AVE) 2008 in
Romanian and English, at CLEF 2008 (Cross-Language Evaluation Forum).
(Rank: 1/8 Romanian)
5. Participation in the competition Answer Validation Exercise (AVE) 2007 in
English, at CLEF 2007. (Rank: 1/13)
6. Participation in the competition Recognizing Textual Entailment 3 (RTE 3),
organized by the PASCAL Excellence Network. (Rank: 3/26)

3. PARTICIPATION IN PROJECTS:
1. Title of the project: Living in Surveillance Societies (LiSS) COST
(European Cooperation in the field of Scientific and Technical Research)
Action IS0807 (European Commission FP7)

253

2. Title of the project: Las tecnologas del lenguaje humano ante los nuevos
retos de la comunicacin digital ACOMP/2010/286 Consellera de
Educacin
3. Title of the project: Procesamiento del lenguaje y sistemas de informacin
(GPLSI) VIGROB-100 Universidad de Alicante
4. Title of the project: Las Tecnologas del Lenguaje Humano ante los
nuevos retos de la comunicacin digital
(TIN2009-13391-C04-01) Ministerio de Ciencia e Innovacin (Ministry of Science and Innovation)
5. Title of the project: Desarollo conjunto de un grupo de herramientas y
recursos de tecnologas del lenguaje humano (Joint Development of Tools
and Resources for Human Language Technologies)
6. Title of the project: Desarollo y licencia de software para la extraccin de
informacin de documentos notariales (Software Development and
Licencing for the Extraction of Information from Notary Documents)
7.
P ometeo /2009/119 - Desarrollo de tcnicas inteligentes
era de textos.

4. PARTICIPATION IN PROGRAM COMMITTEES:


I was a Program Committee member of the following events:
! eETTs 2009 (Workshop on Events in Emerging Text Types), with RANLP
2009;
! WOMSA 2009 (Workshop on Opinion Mining and Sentiment Analysis),
with CAEPIA 2009;
! SMUC 2010 (2nd International Workshop on Search and Mining Usergenerated Contents), with CIKM 2010;
! WASSA 2010 (1st Workshop on Computational Approaches to Subjectivity
and Sentiment Analysis), with ECAI 2010;
! WASSA 2.011 (2nd Workshop on Computational Approaches to
Subjectivity and Sentiment Analysis), with ACL-HLT 2011;
! SNMART 2011 (Social Network Mining, Analysis and Research Trends:
Techniques and Applications);
! SAAIP 2011 (Sentiment Analysis where AI Meets Psychology), with
IJCNLP 2011;
! SENTIRE 2011 (Sentiment Elicitation from Natural Text for Information
Retrieval and Extraction), with ICMD 2011.
! ICL 2011: Workshop on Iberian Cross-Language NLP tasks at SEPLN
2011.
! RANLP 2011 Conference on Recent Advances in Natural Language
Processing.
254

I was also a reviewer for the Language Resources and Evaluation journal - Special
Issue on Short Texts, IEEE Transactions on Affective Computing - Special Issue on
Naturalistic Affect Resources for System Building and Evaluation, ACM
Transactions on Itelligent Systems and Technology, RANLP 2009, NLDB 2009,
SEPLN 2009, IceTAL 2010, NLDB 2011.

5. ORGANIZATION OF SCIENTIFIC EVENTS:


Together with Prof. Andrs Montoyo, Prof. Patricio Martnez-Barco and Ester
Boldrini, I was an organizer (co-chair) of the WASSA 2010 (1st Workshop on
Computational Approaches to Subjectivity and Sentiment Analysis), celebrated
with the 19th European Conference on Artificial Intelligence (ECAI 2010).
This year, we are organizing the second edition of the workshop - WASSA 2.011 at
ACL-HLT 2011, in Portland, Oregon, on the 24th of June.
Additionally, I am one of the organizers of NLDB 2011 (16th International
Conference on the Applications of Natural Language to Information Systems).

6. AWARDS AND SCHOLARSHIPS OBTAINED:


1. Scholarship from the Ministry of Education for a 3-month research traineeship at
a foreign centre, with the aim of obtaining the European Mention on the PhD title,
2009-2010, at the University of Saarland, Saarbrcken, Germany.
2. PhD Scholarship from the University of Alicante (2007-2011).
3. Scholarship from CLEF for the participation in the summer school entitled
TrebleCLEF Summer School on Multilingual Information Access, celebrated
between the 15th and 19th of June 2009 in Santa Croce de Fossabanda, Pisa, Italy.
(summer school director - Prof. Carol Peters) .
4. 6-month traineeship with the European Commission, Joint Research Centre, Ispra,
Italy. Institute for the Protection and Security of the Citizens, Global Security and
Crisis Management Unit, Open Source Text Information Mining and Analysis
Action.
5. Scholarship from the Duques de Soria Foundation (Fundacin Duques de Soria)
for the participation in the language technologies summer courses, entitled
"Tcnicas de extraccin y visualizacin de informacin: aplicacin en la
construccin de portales especializados", 7-11 July 2008, Soria, Spain.

255

ANNEX B. RESUMEN EN CASTELLANO


INTRODUCCIN Y MOTIVACIN
La era en la que vivimos ha sido denominada de muchas maneras. "Aldea global",
"era tecnotrnica", "sociedad postindustrial", "sociedad de la informacin", "era de
la informacin", y "sociedad del conocimiento" son slo algunos de los trminos
utilizados en un intento de describir los profundos cambios que han ocurrido en la
vida de las personas y las sociedades en todo el mundo como resultado del rpido
desarrollo de las tecnologas de la informacin y la comunicacin (TIC), el acceso a
Internet y su transformacin gradual en una Web social. En este nuevo contexto,
tener acceso a grandes cantidades de informacin ya no es un problema. Todos los
das cualquier persona con una conexin a Internet tiene acceso a terabytes de
nueva informacin que se producen en la Web. A diferencia de otros tiempos,
cuando la bsqueda de fuentes de informacin era el problema principal de los
usuarios, la sociedad de la informacin reta a las compaas y los individuos a crear
y utilizar mecanismos para la bsqueda, recuperacin y procesamiento de datos
relevantes entre la gran cantidad de informacin disponible, para obtener
conocimiento que puedan utilizar en su beneficio. En contraste con unos aos atrs,
cuando esta ventaja dependa de la capacidad de encontrar fuentes de informacin,
en la sociedad actual, que est inundada por un flujo de informacin que cambia a
un ritmo alto, la ventaja viene dada por la calidad (exactitud, fiabilidad) del
conocimiento extrado y su concrecin. En la poca en que vivimos, la informacin
se ha convertido en el principal objeto de comercio. Tener a mano informacin
oportuna de alta calidad es crucial en todos los mbitos de actividad humana: social,
poltico y econmico, por mencionar slo algunos.
Sin embargo, en muchos casos, la informacin pertinente no se encuentra en
fuentes estructuradas (es decir, tablas o bases de datos), sino en documentos no
estructurados, escritos en lenguaje humano. La elevada cantidad en la que estos
datos se encuentran en la web requiere el uso de tcnicas de procesamiento
automtico para poder tratarlos.
La disciplina que se ocupa del tratamiento automtico del lenguaje natural (o
lenguaje humano) en textos o el habla se llama Procesamiento del Lenguaje Natural
(PLN). El PLN es parte del rea de investigacin de la Inteligencia Artificial (IA),
que se define como "la ciencia y la ingeniera de hacer mquinas inteligentes"
(McCarthy, 1959) mediante la simulacin de los mecanismos de la inteligencia
humana. Dentro del PLN existen diversas reas de investigacin. Cada una de estas
reas se ha establecido para resolver problemas difciles que se pueden encontrar en
257

cualquier mbito del PLN (por ejemplo, la Desambiguacin del Sentido de las
Palabras, la Resolucin de la Co-referencia, la Resolucin de las Expresiones
Temporales, el Etiquetado de Roles Semnticos), o en dependencia de una
aplicacin final especfica (por ejemplo, la Recuperacin de Informacin, la
Extraccin de Informacin, la Bsqueda de Respuestas, los Resmenes de Texto, la
Traduccin Automtica).
Tradicionalmente, estas reas de aplicacin de la PLN fueron diseadas para el
tratamiento de los textos que describen datos factuales (hechos que se pueden
observar y comprobar en la realidad). No obstante, hoy en da, la informacin sobre
hechos ya no es la fuente principal de donde se extrae el conocimiento fundamental
o ms bsico.
El presente est marcado por la creciente influencia de la web social (la web de
la interaccin y la comunicacin) en las vidas de las personas en todo el mundo.
Ahora, ms que nunca, la gente est totalmente dispuesta y feliz por compartir sus
vidas, conocimientos, experiencias y pensamientos con el mundo entero, a travs de
blogs, microblogs, foros, wikis o, incluso, sitios de comercio electrnico que dan la
opcin de compartir reseas sobre los productos que venden. La gente est
participando activamente en los principales acontecimientos que tienen lugar en
todas las esferas de la sociedad, expresando su opinin sobre ellos y comentando las
noticias que aparecen. El gran volumen de datos que contienen opiniones
disponibles en Internet, en reseas, foros, blogs, microblogs y redes sociales ha
producido un importante cambio en la forma en que las personas se comunican,
comparten conocimientos y emociones e influyen en el comportamiento social,
poltico y econmico. En consecuencia, esta nueva realidad ha dado lugar a
importantes transformaciones en la forma, extensin y rapidez de circulacin de las
noticias y sus opiniones asociadas, dando lugar a fenmenos sociales, econmicos y
psicolgicos nuevos y desafiantes.
Para estudiar estos fenmenos y abordar la cuestin de extraer el conocimiento
fundamental que en la actualidad figura en los textos que contienen expresiones con
sentimientos, han nacido nuevos campos de investigacin dentro del PLN, que
tienen el objetivo de detectar la subjetividad en el texto y/o extraer y clasificar los
sentimientos de las opiniones expresadas en distintas categoras (por lo general en
positivos, negativos y neutrales). Las nuevas tareas que se abordan en el PLN son
principalmente el anlisis de la subjetividad (que trata sobre los "estados privados"
(Banfield, 1982), un trmino que contiene sentimientos, opiniones, emociones,
evaluaciones, creencias y especulaciones), el anlisis de sentimientos y la minera
de opiniones. Esta no ha sido la nica forma de referirse a los enfoques adoptados.
Tambin se han utilizado tambin otras terminologas que utilizaban otros trminos
para denominar a las tareas, por ejemplo, minera de reseas o la extraccin de
valoracin. Asimismo, los trminos de anlisis de sentimientos y minera de
258

opiniones se han utilizado indistintamente, ya que algunos autores consideran que


hacen referencia a la misma tarea (Pang and Lee, 2008). Uno de los problemas del
PLN estrechamente relacionado tambin con las tareas mencionadas es la
deteccin de la emocin, que tiene como objetivo la clasificacin de los textos de
acuerdo con la emocin expresada en ellos. Todas estas reas de investigacin son
parte de una esfera ms amplia de la Inteligencia Artificial denominada
computacin del afecto (Picard, 1995).
La presente tesis doctoral se ocupa de las cuestiones y los desafos en el
desarrollo de mtodos y recursos para la tarea del PLN denominada anlisis de
sentimientos. Definido de forma general, el objetivo de esta tarea es la deteccin
automtica de los sentimientos expresados en textos (normalmente por una fuente,
sobre un objeto, que puede ser una persona, un evento, un producto, una
organizacin etc.) y su clasificacin segn la polaridad/orientacin que tienen
(normalmente positiva, negativa o neutra, aunque distintos autores han propuesto
escalas ms finas de sentimientos, incluyendo por ejemplo las clases muy positivo o
muy negativo).

PROBLEMAS DEL DOMINIO Y OBJETIVOS DE LA


INVESTIGACIN
La investigacin en este campo, la minera de opiniones (anlisis de sentimientos),
ha demostrado que el anlisis de los sentimientos es un problema difcil, que se
tiene que abordar desde diferentes perspectivas y en diferentes niveles, dependiendo
de una serie de factores. Estos factores incluyen: nivel de de inters (general o
especfico, dependiendo de si la opinin general sobre el objeto en cuestin es
suficiente o se necesita conocimiento detallado de los sentimientos expresados
sobre distintos componentes del objeto), la frmula de consulta (Nokia E65 /
Por qu la gente compra el Nokia E65?), el tipo de texto (revisin en un
foro/blog/dilogo/artculo de peridico), y la forma de expresar la opinin directamente (mediante declaraciones opinin, por ejemplo, Me parece que este
producto es maravilloso! o Esta es una iniciativa brillante!), de forma indirecta
(utilizando vocabulario relacionado con la expresin del afecto, por ejemplo, Me
encantan las fotos tomadas con esta cmara! o Personalmente, estoy
conmocionado por cmo se puede proponer una ley as!) o implcitamente (con
adjetivos que expresan una evaluacin, cuyo objeto se sobrentiende por ejemplo,
"Es ligero como una pluma y cabe perfectamente en mi bolsillo" o presentando una
situacin factual de la que se puede inferir, utilizando conocimiento comn, una
emocin positiva o negativa por ejemplo Se rompi en dos das.). Otros
factores que hacen la tarea de minera de sentimientos difcil es la aplicacin final y
el tipo de texto que se utiliza (reseas, que contienen solo opiniones sobre un
259

producto, escritas por una solo fuente, en comparacin con blogs o debates, que
tienen una estructura de dialogo, en el que se expresan opiniones sobre distintos
objetos, por distintas fuentes). Finalmente, para las aplicaciones finales, el anlisis
de sentimientos no es la primera ni la ltima tarea que se debe realizar. Para extraer
el sentimiento de textos, primero es necesario recuperar un conjunto de documentos
relevantes. El resultado del procesamiento de un texto con un sistema de anlisis de
sentimientos puede tener mucha informacin redundante e incluso puede no
resolver totalmente el problema, debido a la gran cantidad de datos existentes.
Los sistemas implementados para la tarea de anlisis de sentimientos se basan en
reglas, bolsas de palabras, utilizando un lxico de palabras que tienen una
orientacin del sentimiento (positivo o negativo), mtodos estadsticos o
aprendizaje automtico.
Analizando los sistemas existentes, hemos identificado los siguientes problemas:
! La tarea de anlisis de sentimientos y los conceptos relacionados no son
definidos de forma nica en los diferentes trabajos de investigacin. Por
tanto, no est claro siempre si los distintos investigadores que trabajan en el
anlisis de sentimientos pueden comparar el rendimiento de sus sistemas,
ya que los textos sobre los que evalan pueden tener diferentes elementos
anotados.
! La tarea de anlisis de sentimientos se resuelve de la misma manera,
independientemente del tipo de texto que se procesa y del objetivo de la
aplicacin final.
! No existen recursos anotados para la tarea de anlisis de sentimientos en
todos los gneros textuales.
! No existen lxicos de palabras que expresen sentimientos para otros
idiomas distintos al ingls.
! La mayora de sistemas trabajan a nivel lxico, utilizando reglas, lxicos,
mtodos estadsticos o aprendizaje automtico. La investigacin que se ha
hecho hasta ahora no toma en cuenta otros niveles de anlisis, como el
sintctico o semntico. Por tanto, el asegurar que la fuente de la opinin
expresada es la requerida o sobre qu objeto se expresa la opinin en un
texto son aspectos que no se toman en consideracin. Estos aspectos
pueden tener un alto impacto sobre el rendimiento y la utilidad de los
sistemas de anlisis de opiniones.
! La mayor parte de la investigacin no distingue sobre los distintos
componentes de un texto, en especial sobre el autor, el texto y el lector. La
tarea de anlisis de sentimientos puede tener diferentes objetivos,
dependiendo de la perspectiva que se requiere analizar (por ejemplo, si el
autor tiene preferencia sobre un cierto objeto descrito, si el texto contiene

260

informacin que es buena o mala en s, si el lector confa en la fuente de la


informacin).
! Las tareas tradicionales (bsqueda de informacin, bsqueda de respuestas,
resmenes automticos) se enfrentan con problemas adicionales en el caso
de que la informacin buscada o resumida es de tipo opinin, dadas por las
caractersticas del lenguaje afectivo. Por tanto, para poder adecuar los
sistemas de este tipo para tratar informacin que contiene expresiones de
afecto, las peculiaridades de este lenguaje tienen que ser estudiadas y se
tienen que proponer mtodos adecuados para resolver los problemas
encontrados de forma eficaz.
El objetivo de nuestro trabajo ha sido crear, explotar y evaluar mtodos y
recursos tanto nuevos como consagrados para la deteccin y posterior clasificacin
de acuerdo a su polaridad (positiva / negativa/ neutro) de los sentimientos
expresados en textos.
En concreto, el primer objetivo es desarrollar tcnicas adecuadas para la
deteccin y clasificacin automtica de los sentimientos expresados de forma
directa, indirecta o implcita en los textos de diferentes tipos (reseas, artculos de
peridicos, dilogos/debates y blogs) en diferentes idiomas. El segundo objetivo es
aplicar los mtodos de anlisis de sentimientos que se proponen en el contexto o
conjuntamente con otras tareas de PLN (bsqueda de respuestas y resmenes
automticos) y proponer tcnicas adecuadas para hacer frente a las cuestiones
planteadas en estas tareas por las peculiaridades de la expresin del afecto.
En concreto, nos centramos en:
Definir la tarea y conceptos generales relacionados, a partir del estudio de
las definiciones existentes en la literatura y la clarificacin de las
inconsistencias detectadas;
Proponer y evaluar mtodos para definir y abordar el anlisis de los
sentimientos de diversos gneros textuales en diferentes idiomas;
Redefinir la tarea y proponer mtodos para anotar corpus especficos para
el anlisis de sentimientos en para un tipo de texto en diferentes idiomas, en
el caso de que la tarea de anlisis de sentimientos no hubiera sido
claramente definida para el gnero textual en cuestin y/o ningn corpus
estuviera disponible para el mismo. Estos recursos estn disponibles al
pblico para el uso de la comunidad cientfica;
Aplicacin de tcnicas de minera de opinin en el contexto de los sistemas
end-to-end y tambin en conjunto con otras tareas del PLN. Para ello, nos
hemos concentrado en realizar anlisis de emociones en las tareas de
bsqueda de respuesta y resumen automtico;

261

Llevar a cabo experimentos con sistemas de bsqueda de respuesta y


sistemas de resmenes automticos, diseados para hacer frente a datos
factuales solamente;
Proponer y evaluar un nuevo marco para lo que llamamos "bsqueda de
respuestas a preguntas de opinin" (en ingls - Opinion Question
Answering) y los nuevos mtodos para "elaborar resmenes de opiniones
de forma automtica" (en ingls - Opinion Summarization), tras realizar un
conjunto de experimentos que mostraron que los sistemas de bsqueda de
respuestas y de resumen automtico sobre textos presentando hechos no
funcionaban correctamente para analizar textos que contenan opiniones;
Presentacin de un mtodo general para la deteccin de la emocin
expresada de manera implcita en texto. En primer lugar, presentamos el
mtodo para construir un lxico de trminos que en s mismos no contienen
la emocin, pero que disparan la emocin en un lector. Posteriormente, se
propuso un mtodo para resumir textos a partir del anlisis de los
sentimientos expresados basado en claves lingsticas, as como se propuso
y evalu un mtodo para representar el texto como las cadenas de accin.
La emocin provocada por la situacin que se presenta en el texto se juzga
posteriormente en base a conocimiento de sentido comn sobre el efecto
emocional de cada accin en la cadena;
La evaluacin de nuestros enfoques en las competiciones internacionales, a
fin de comparar nuestros enfoques con los dems y validarlos.

Con el fin de alcanzar los objetivos propuestos, el trabajo que se presenta ha


sido estructurado en torno a responder a cinco preguntas de investigacin. Cada
uno de los captulos de esta tesis presenta mtodos y evaluaciones hechas con el
fin de responder a estas preguntas.

CONTRIBUCIN
Este apartado presenta en detalle las contribuciones realizadas a la investigacin en
el campo de anlisis de sentimientos a lo largo de esta tesis y muestra cmo los
mtodos y recursos propuestos llenan vacos importantes en la investigacin
existente. Las principales contribuciones responder a cinco preguntas de
investigacin:
1. Cmo se puede definir la tarea de anlisis de sentimientos y, en una
perspectiva ms amplia, la minera de opiniones de una manera correcta?
Cules son los conceptos principales que se deberan definir antes de
afrontar la tarea?
262

En el captulo 2, presentamos un conjunto de definiciones asociadas a cada uno


de los conceptos involucrados en la tarea de anlisis de sentimientos la
subjetividad, la objetividad, la opinin, el sentimiento, la emocin la actitud, y la
evaluacin.
Nuestra aportacin en este captulo reside en mostrar con claridad que el anlisis
de los sentimientos y la minera opinin no son sinnimos, aunque en la literatura
por lo general son empleados indistintamente. Adems, demostramos que la
opinin, como es definida por el diccionario Webster, no es sinnimo de
sentimiento. Considerando que los sentimientos son tipos de opiniones, los que
reflejan las vivencias (es decir, la parte consciente de las emociones), pero todas las
opiniones no son los sentimientos (es decir, hay tipos de opiniones que no son el
reflejo de las emociones). Tambin demostramos que el anlisis de la subjetividad
no est directamente relacionado con el anlisis de los sentimientos, a pesar que as
est considerado por muchos de los investigadores en el campo. En otras palabras,
la deteccin de frases subjetivas no implica directamente la obtencin de las frases
que contienen sentimiento. Este ltimo, como expresiones de las evaluaciones
basadas en la emocin, no son necesariamente indica en frases subjetivas, pero
tambin puede expresarse en frases objetivas. Las frases subjetivas pueden o no
contener expresiones de las emociones. Las relaciones entre los diferentes
conceptos se resumen en la siguiente figura, que contiene un diagrama de
conjuntos:
Subjectivo

Objectivo

Opinin

Sentimiento
Emocin
Por ltimo, demostramos la existencia de una clara conexin entre el trabajo
realizado en el marco de anlisis de sentimientos/minera opinin y la de
evaluacin/anlisis de la actitud. A pesar de todas estas reas se consideran
generalmente para referirse al mismo tipo de trabajo, el objetivo ms amplio de la
actitud o el anlisis de evaluacin pueden captar mucho mejor la investigacin que
se ha hecho en el anlisis de los sentimientos, incluyendo todas las clases de
evaluacin (afectivo, cognitivo-conductual) y la relacin entre autor, lector y el
significado del texto. En base a esta observacin y en vista de la teora de la
valoracin, en el captulo 6 se propone un modelo de deteccin de emociones
263

basado en el conocimiento de sentido comn. La definicin clara de estos conceptos


tambin ha ayudado a definir de manera adecuada la tarea de anlisis de
sentimientos en el contexto de los diferentes gneros textuales que empleamos en
nuestra investigacin. Posteriormente, esta correcta definicin ha hecho posible la
definicin de esquemas de anotacin y la creacin de recursos para el anlisis de los
sentimientos en textos de todos los gneros tratados en la investigacin. Todos estos
recursos fueron empleados tanto en la evaluacin de los mtodos especficos que
hemos creado para el anlisis de los sentimientos de los diferentes gneros
textuales. Tanto los recursos creados, a travs de la alta concordancia entre
anotadores obtenida, as como los mtodos propuestos, a travs del desempeo de
los sistemas de aplicacin, demostraron que nuestros esfuerzos para dar una
definicin clara eran de hecho una necesaria contribucin a este campo.
2. El anlisis de sentimiento puede ser realizado usando los mismos mtodos
para todos los tipos de texto? Cules son las peculiaridades de los
diferentes tipos de textos y cmo influyen en los mtodos que se utilizan
para hacerle frente? Necesitamos recursos especiales para diferentes
tipos de textos?
3. Puede utilizarse un recurso en un idioma dado para resolver problemas
en textos escritos en otros idiomas (a travs de la traduccin)? Cmo
pueden ampliarse los recursos a otros idiomas?
En el captulo 4, mostramos las particularidades de los diferentes tipos de textos
(reseas, artculos de peridicos, blogs, debates polticos), se analizaron y
propusieron tcnicas adecuadas para analizar los sentimientos contenidos en cada
uno de ellos. Los trabajos presentados en este captulo fueron publicados en varias
conferencias y revistas (incluidas en la seccin de referencias bibliogrficas y en el
listado de contribuciones cientficas del Anexo A). Se evalu cada una de las
aproximaciones utilizadas y se demostr que al menos funcionan al nivel de los
sistemas ms avanzados y, en algunos casos, incluso superan sus resultados.
En este captulo, presentamos los diferentes mtodos y recursos que hemos
construido para la tarea de anlisis de sentimientos en varios tipos de textos. Se
comenz por explorar los mtodos para afrontar la tarea de la minera de opiniones
y su resumen basada en rasgos, aplicado en reseas de productos. Se analizaron las
caractersticas de esta tarea y se identificaron las carencias en la investigacin
existente. Propusimos y evaluamos diferentes mtodos para superar los obstculos
identificados, entre los cuales los ms importantes fueron el descubrimiento de las
caractersticas mencionadas y de computacin indirecta de la polaridad de
opiniones de una manera que es una caracterstica dependiente, con la Distancia
Normalizada de Google (Cilibrasi and Vitanyi, 2006) y el Anlisis Semntico
Latente (Deerwester et al., 1990) como medida de asociacin semantica. Con
264

posterioridad, se propuso un modelo unificado para la anotacin del sentimiento en


este tipo de textos (Balahur and Montoyo, 2009), capaz de capturar los fenmenos
importantes que hemos identificado los diferentes tipos de expresiones de
sentimiento: directos, indirectos, implcitos; la caracterstica que citan y el
fragmento de texto que expresa una opinin especfica. Esta distincin no se haba
propuesto hasta el momento en la literatura. La contribucin no slo se da por el
proceso de anotacin, sino tambin por el hecho de que la consideracin de grandes
tramos de texto en la representacin de una sola opinin nos ha llevado a la
investigacin en bsqueda de respuestas para preguntas de opinin (vase el
captulo 5). La recuperacin de texto utilizando fragmentos de tres oraciones puede
mejorar de forma considerable el rendimiento del sistema de bsqueda de respuesta
en estas preguntas. Conjuntamente con este esquema de anotacin, se propuso un
mtodo para detectar y clasificar la opinin que aparece sobre las caractersticas
ms importantes de un producto, asignando un nmero de estrellas en base,
utilizando como base un sistema de implicacin textual. Este mtodo permite
obtener, adems de una clasificacin de dos vas de opinin sobre las caractersticas
del producto, un resumen de las frases ms importantes que hacen referencia a
ellos. Este resumen puede ser empleado como apoyo en la tarea de minera de
opiniones basado en funciones y en el proceso de compresin, que ofrece un
resumen basado en las opiniones expresadas acerca del producto en cuestin.
Ms adelante, se exploraron diferentes mtodos para abordar el anlisis de los
sentimientos de artculos periodsticos. Tras los experimentos iniciales, se
analizaron las razones del bajo rendimiento obtenido y se redefini la tarea,
teniendo en cuenta las caractersticas de este gnero textual. Creamos un modelo de
anotacin y anotamos dos corpus diferentes sobre citas en artculo de prensa, en
ingls y alemn. Una vez redefinida la tarea y delimitado el alcance del proceso de
anlisis de sentimientos a las citas pequeos fragmentos de texto que contiene el
discurso directo, cuyo origen y destino son previamente conocidos, el acuerdo de
anotacin aument significativamente. Asimismo, la redefinicin de la tarea hizo
posible la aplicacin de mtodos de procesamiento automtico ms apropiados que
consiguieron mejorar significativamente el rendimiento del sistema de anlisis de
sentimientos que se haba diseado.
En cuanto a la aplicacin del anlisis de sentimientos a los diferentes tipos de
textos, que contienen una mezcla de informacin afectiva con informacin factual y
donde las fuentes y los objetivos de las opiniones son mltiples, hemos propuesto
diferentes mtodos generales para solventar la tarea, en este caso aplicado sobre
textos con debates polticos. Los resultados de este ltimo experimento nos
motivaron para analizar los requisitos de un esquema general de etiquetado para la
tarea de anlisis de sentimientos, que pudiera ser usado para capturar todos los
fenmenos relevantes en la expresin del sentimiento.
265

Para ello, en Boldrini et al. (2009), definimos EmotiBlog, un esquema de


anotacin que permite capturar, con una alta granularidad, todos los fenmenos
lingsticos relacionados con la expresin sentimiento en el texto. Los experimentos
posteriores han demostrado que este modelo es apropiado para la formacin de
modelos de aprendizaje automtico para esta tarea en textos de diferentes gneros
en los dos idiomas en los que se han realizado los experimentos ingls y espaol.
Por ltimo, hemos demostrado que el corpus anotado con EmotiBlog se puede
utilizar para extraer las caractersticas de los modelos de aprendizaje automtico
que permitan abordar el anlisis de los sentimientos en otros idiomas, a travs de un
proceso de traduccin. Los buenos resultados obtenidos en la tarea 18 de SemEval
2010 Desambiguacin del sentimiento ambiguo competencia adjetivos, donde
tradujimos textos en ingls a chino tradicional y aplicamos un modelo de
aprendizaje automtico entrenado con los datos de EmotiBlog en ingls,
demostraron que el modelo de anotacin suficientemente robusto y til incluso con
datos que contengan ruido.
4. Cmo podemos tratar las opiniones en el contexto de las tareas
tradicionales del PLN? Cmo podemos adaptar las tareas tradicionales
(Recuperacin de Informacin, de Respuestas, resumen de texto) en el
contexto de los textos que contienen opiniones? Cules son los "nuevos"
desafos en este contexto?
En el captulo 4, nicamente nos concentramos en la tarea de anlisis de
sentimientos como un reto independiente, omitiendo los pasos necesarios para
obtener los textos sobre los que se aplican los mtodos de anlisis de opinin o para
eliminar la redundancia en la informacin obtenida. En un escenario del mundo
real, sin embargo, la deteccin automtica de la opinin expresada en un texto a
menudo no es la primera ni la ltima tarea por realizar. Previo al anlisis de los
sentimientos contenidos en los textos, las opiniones sobre ciertos objetivos deberan
ser recuperadas. En un gran nmero de casos, los resultados obtenidos tras
determinar automticamente el sentimiento contenido en los textos siguen
planteando muchos problemas en trminos de volumen. As, aunque se extrae el
sentimiento de forma automtica, todava puede ser necesario un componente de
compresin, con el fin de reducir an ms la cantidad de informacin, de modo que
se puede ser leda y utilizada por una persona.
Teniendo en cuenta estas necesidades, en el captulo 5 se investig sobre los
mtodos para combinar la minera de opiniones con bsqueda de respuestas y
generacin de resmenes automticos. En este captulo, demostramos que la
realizacin de las tareas tradicionales en el contexto del texto que contiene
opiniones tiene muchos retos y que los sistemas que fueron diseados para trabajar
exclusivamente con datos factuales no son capaces de contestar a preguntas de
266

opinin. Nuestra contribucin reside en la demostracin, mediante un proceso de


evaluacin, que para poder tratar consultas de este tipo, nuevos elementos tienen
que ser definidos. Por ello, propusimos la inclusin de los siguientes elementos:
tipo de polaridad, fuente previsto, objeto de la opinin previsto; y definimos los
mtodos para detectar la fuente y el objeto de los sentimientos expresados, as como
las respuestas candidatas, utilizando tcnicas de minera de opiniones y el empleo
de anotaciones semnticas. Para el proceso de recuperacin, como hemos
demostrado antes y confirmado en el escenario Bsqueda de Respuestas en
Opiniones (en ingls Opinion Question Answering OQA), utilizar grandes
fragmentos de textos es ms apropiado cuando se trata con contenido que incluye
expresiones de emocin. En concreto, demostramos que la recuperacin de
fragmentos de tres oraciones conduce a mejores resultados en el caso de los
sistemas de OQA. Al proponer y evaluar estos nuevos elementos y tcnicas, que
han mostrado la forma de abordar OQA de una manera slida, coherente con las
caractersticas de los textos con opiniones. Por ltimo, se evalu el impacto del uso
de diferentes herramientas y recursos en esta tarea, como las para la resolucin de la
anfora y la ampliacin de la pregunta por medio de parfrasis. Estudiar la forma
ptima en que estos dos componentes son combinados en un sistema es una
importante lnea de trabajo futuro abierta.
En el caso de los resmenes de opinin, nuestra aportacin reside en: a) estudiar
el orden en que los sistemas de minera de opiniones y de compresin tienen que ser
empleados, b) estudiar la manera en que los sistemas de minera de opinin y de
resumen automtico puede ser utilizados en conjunto, c) estudiar el efecto de la
deteccin del tema para la minera de opiniones en el contexto de resmen de
opinin; d) proponer un mtodo para resumir opiniones basado en la intensidad del
sentimiento expresado.
En esta parte de la investigacin, nuestra aportacin reside en mostrar que el
sistema de anlisis de sentimientos debe ser empleado antes en el sistema de
compresin y que el componente de anlisis de sentimientos debe ser mejorado con
mecanismos de deteccin del tema. En otros casos, aunque los sistemas de anlisis
de emociones realicen una clasificacin correcta (de acuerdo a la polaridad del
sentimiento), el hecho de que la relacin con el tema no est contemplada conduce
a la introduccin de los datos no relevantes en los resmenes finales. Por ltimo,
demostramos que en el caso del texto que contiene expresiones de opinin, la
relevancia est dada no slo por la informacin contenida, pero tambin por la
polaridad de la opinin y su intensidad. Aunque los resultados iniciales han
demostrado que no existe una correlacin entre las anotaciones del estndar de
anotacin y el nivel de intensidad de los sentimientos detectados, el sistema de
anlisis de sentimientos que us este mtodo en la competicin TAC 2008, obtuvo
resultados altos en cuanto a la medida-F. Por tanto, creemos que se deben estudiar
267

otros mecanismos para la medicin de la intensidad de opinin, de modo que sea


posible descubrir la conexin entre la importancia del contenido con opinin en
cuanto a la informacin que aporta, y la intensidad que tiene.
5. Podemos proponer un modelo para detectar las emociones de un texto, en
los casos en que se expresa de forma implcita, que requieren conocimiento
del mundo para su deteccin?
En los primeros captulos de esta tesis, exploramos la tarea de anlisis de
sentimientos en diferentes tipos de textos e idiomas y proponemos una amplia
variedad de mtodos apropiados para afrontar los problemas que aparecieron para
cada tipo de texto. La mayora de las veces, sin embargo, los diferentes enfoques se
han limitado a conocer slo las situaciones donde el sentimiento se expres de
forma explcita, es decir, cuando se pueden encontrar en el texto seales lingsticas
que indiquen que contiene elementos subjetivos o sentimiento. No obstante, en
muchos casos, la emocin en la que se basa el sentimiento no est explcitamente
presente en el texto, pero se puede inferir en base al conocimiento de sentido comn
(es decir, la emocin no est explcitamente, pero implcitamente expresada por el
autor, mediante la presentacin de situaciones que la mayora de las personas,
gracias a su sentido comn, asocian con una emocin, como "ir a una fiesta", "ver a
su hijo dar sus primeros pasos", etc).
En el captulo 6 de la tesis, presentamos nuestra contribucin a la cuestin de la
deteccin automtica de emocin expresada en el texto de manera implcita. La
aproximacin inicial, sustentada sobre la Teora de la Relevancia (Sperber and
Wilson, 2000), se basa en la idea de que la emocin es provocada por conceptos
especficos, de acuerdo a su importancia, y debe considerarse en relacin con las
necesidades y motivaciones bsicas.
El segundo enfoque que proponemos se basa en la Teora de los modelos de
evaluacin (presentada en detalle por Scherer, 1999). La idea general que subyace
es que la mayora de veces las emociones no se declaran explcitamente en los
textos, sino que resultan de la interpretacin (evaluacin) de las acciones contenidas
en la situacin descrita, as como las propiedades de sus actores y objetos. Nuestra
contribucin en esta ltima parte de la investigacin reside en la creacin de un
marco para la representacin de situaciones que se describen en el texto como
cadenas de accin (con sus correspondientes actores y objetos), y sus propiedades
correspondientes (incluyendo las afectivas), como el conocimiento de sentido
comn. Se muestra la forma de detectar automticamente a partir del texto los
llamados "criterios de evaluacin" y la forma de extender el conocimiento sobre las
propiedades de los conceptos involucrados en cada situacin utilizando fuentes
externas. Por ltimo, se demuestra a travs de una extensa evaluacin de que tal
representacin es til para obtener una etiqueta exacta de la emocin expresada en
268

el texto, sin ninguna pista lingstica est presente en l. Debido a la relacin


directa entre los sentimientos y la presencia de emociones, la deteccin de
expresiones implcitas de emocin puede aumentar el rendimiento de los sistemas
de anlisis de sentimientos, un hecho que hemos demostrado en los experimentos
que presentamos en el captulo 3 en el caso de anlisis de sentimientos de
comentarios.

CONCLUSIONES Y TRABAJOS FUTUROS


Finalmente, se presentan las posibles direcciones de desarrollo futuro de la
investigacin realizada en esta tesis. Con el fin de ser coherente con la estructura de
la tesis, estas direcciones se presentan de la misma forma, en relacin con las
preguntas de investigacin que nos propusimos a responder y los puntos que siguen
pendientes de resolucin:
i.

Cmo se puede definir la tarea de anlisis de sentimientos y, en una


perspectiva ms amplia, la minera de opiniones, de una manera correcta?
Cules son los conceptos principales que se deberan definir antes de
afrontar la tarea?
En el captulo 2 de la tesis, presentamos un resumen de las definiciones que se
dan en la literatura de PLN para las tareas relacionadas con el anlisis de la
subjetividad, el anlisis de los sentimientos, la minera opinin, evaluacin/anlisis
de la actitud, la deteccin de las emociones, as como los conceptos que participan
en dichas tareas. Demostramos posteriormente que hay una alta inconsistencia entre
las definiciones de las tareas. Finalmente, propusimos un conjunto de definiciones
operativas coherentes con la manera en que los diferentes trminos relacionados
con el anlisis de los sentimientos se definieron en fuentes bien establecidas y los
planteamientos que hicimos en la investigacin. En el captulo 4, mostramos que el
anlisis de los sentimientos debe de ser abordado de una manera diferente,
dependiendo de los tipos de texto considerado. A partir de estos esfuerzos, una
futura lnea de trabajo es la definicin de un marco unificado para el anlisis de los
sentimientos, que describa la tarea de una manera general, aunque uniforme a travs
de gneros y aplicaciones. En este sentido, demostramos en los experimentos con
artculos de prensa, dicho marco debe ser construido tomando en cuenta no slo el
contenido textual, sino tambin los elementos que se relacionan con el autor y el
lector.
Las siguientes preguntas de investigacin abordadas en esta tesis fueron:
2. El anlisis de sentimiento puede ser realizado usando los mismos mtodos
para todos los tipos de texto? Cules son las peculiaridades de los
269

diferentes tipos de textos y cmo influyen en los mtodos que se utilizan


para hacerle frente? Necesitamos recursos especiales para diferentes
tipos de textos?
3. Puede utilizarse un recurso en un idioma dado para resolver problemas
en textos escritos en otros idiomas (a travs de la traduccin)? Cmo
pueden ampliarse los recursos a otros idiomas?
En el Captulo 4, presentamos los diferentes mtodos y recursos para la tarea de
anlisis de sentimientos en diferentes tipos de textos (reseas, artculos de
peridicos, blogs y debates polticos) en diferentes idiomas (ingls, espaol y
alemn). Una futura lnea de trabajo es la extensin de los recursos propuestos para
otros idiomas, ya sea a travs de la traduccin, en cuyo caso, los recursos deberan
ser refinados y evaluados para los diferentes tipos de textos en su idioma original, o
por anotacin directa en el idioma de destino. En este ltimo caso, sera interesante
estudiar las diferentes formas de expresar los sentimientos en dependencia de la
cultura y el idioma de los textos. En relacin directa con el trabajo desarrollado en
esta tesis, en diferentes tipos de textos e idiomas, las futuras lneas de investigacin
podran ser:
i.
En el caso de los textos de revisiones:
a) La extraccin automtica de taxonomas de las caractersticas del
producto.
b. La extensin del marco de trabajo para la anotacin de revisiones y
la minera de opiniones basada en rasgos y resmenes en otros idiomas
diferentes del ingls y espaol.
ii.
En el contexto de las citas de peridicos y, de manera ms general, artculos
de prensa:
a) El estudio del impacto de la fuente de noticias (es decir, en trminos de
sesgo, la reputacin, la confianza) en el proceso de anlisis de
sentimientos, dentro de las fuentes del mismo pas/cultura y entre pases
y culturas diferentes;
b) El estudio de la influencia del lector sobre la forma en que el
sentimiento es percibido de los artculos periodsticos;
c) La extensin automtica y semiautomtica de los recursos que
construidos a otros idiomas (es decir, en adicin a la coleccin de citas
se anotaron para ingls y alemn);
d) El desarrollo de un marco para el anlisis de los sentimientos que se
tienen en cuenta los tres componentes propuestos en el texto autor,
lector y texto - y la manera en que interactan en la negociacin el
significado del texto, de acuerdo con el acto de habla y las teoras de
evaluacin;

270

iii.

iv.

e) El estudio de la influencia del contenido de las noticias (es decir, lo que


denota como "una buena noticia frente a lo malo") en la forma en que el
sentimiento se expresa y, posteriormente, en el desempeo de la tarea
sentimiento anlisis automtico.
En el contexto de los debates polticos y textos de carcter general:
a) Estudio de los mtodos para representar la estructura de dilogo y la
influencia de los diferentes mtodos de anlisis del discurso, as como
las herramientas (por ejemplo, para la resolucin de la anfora) sobre
los mtodos de anlisis propuestos sentimiento;
b) El estudio y el uso de tcnicas de modelado tema a fin de detectar con
precisin el objetivo del sentimiento expresado, independientemente de
su naturaleza que se conoce de antemano;
c) El estudio de cmo afectan las tcnicas de argumentacin con el fin de
detectar el uso de disparadores de emocin en relacin con el tema de
discusin.
En el contexto de los blogs:
a) La extensin del mtodo propuesto de anotacin y etiquetado de corpus
a otros idiomas y tipos similares de texto (por ejemplo, foros, paneles
de discusin).
b) El estudio y el uso de tcnicas de modelado del tema, as como de los
efectos de la resolucin co-referencia en un nivel intertextual.

4. Cmo podemos tratar las opiniones en el contexto de las tareas


tradicionales del PLN? Cmo podemos adaptar las tareas tradicionales
(Recuperacin de Informacin, de Respuestas, texto de resumen) en el
contexto del contenido que expresa opinin? Cules son los "nuevos"
desafos en este contexto?
Teniendo en cuenta estas necesidades de aplicaciones del mundo real, que
necesitan, aparte del componente de anlisis de sentimientos, tambin la
recuperacin de texto y los componentes de texto de resumen, en el captulo 5 se
propone mtodos para combinar la minera opinin con bsqueda de respuestas y
generacin de resmenes automticos. Demostramos que la realizacin de las tareas
tradicionales en el contexto del texto de tipo opinin tiene muchos retos y que los
sistemas que fueron diseados para trabajar exclusivamente con datos factuales no
son capaces de contestar a preguntas de opinin. En esta tesis se propone nuevos
mtodos y tcnicas para adaptarse y responder la pregunta de los sistemas de
compresin para tratar el contenido dogmtico.
En el caso de los sistemas de bsqueda de respuestas a preguntas de opinin, el
trabajo futuro incluye el desarrollo de un punto de referencia para la clasificacin
de preguntas de opinin y la propuesta de mtodos adecuados para hacer frente a
271

cada tipo de consultas de opinin, en un entorno monolinge, multilinge y entre


idiomas. Adems, el marco para responder a preguntas de opinin debe ampliarse
con los recursos adecuados a otros idiomas. Como apreciamos en nuestros
experimentos en el concurso MOAT NTCIR 8, existe una necesidad inmediata para
incluir mtodos de alto rendimiento para la resolucin de expresiones temporales y
la resolucin de la anfora. Lamentablemente, debido al bajo rendimiento de los
sistemas de resolucin de estos aspectos, la influencia que tienen sobre los sistemas
de bsqueda de respuestas a preguntas de opinin es negativa. Otra lnea de trabajo
futuro es el estudio de las tcnicas de ampliacin de consultas que sean apropiadas
para el contenido de tipo opinin. Por lo visto en nuestros experimentos, el uso de
una coleccin de parfrasis que no est especficamente diseado para el contenido
textual que contiene sentimientos conduce a una cada en el rendimiento del sistema
final. En el caso de los enfoques resumen automtico de texto con sentimientos, las
lneas de trabajo futuro incluyen el desarrollo de tcnicas adecuadas para la
deteccin de la relevancia y el estudio de medidas cualitativas para la evaluacin de
los mtodos de compresin de opinin.
La ltima pregunta de investigacin que aborda en esta tesis fue:
5. Podemos proponer un modelo para detectar las emociones de un texto, en
los casos en que se expresa de forma implcita, que requieren conocimiento
del mundo para su deteccin?
En el captulo 6 de la tesis, propusimos dos mtodos para detectar expresiones
implcitas de la emocin, de situaciones donde la etiqueta afectiva puede inferirse
basada en el conocimiento de sentido comn (es decir, la emocin no est
explcitamente, sino implcitamente expresada por el autor, mediante la
presentacin de situaciones que la mayora de la gente, basndose en el sentido
comn, sus necesidades y motivaciones, asocian con una emocin, como "guerra",
"terrorismo", "ir a una fiesta", "ver a su hijo dar sus primeros pasos", etc.). El
planteamiento inicial que se propone se basa en la idea de que la emocin es
provocada por conceptos especficos, de acuerdo a su tema, se ve en relacin con
las necesidades y motivaciones bsicas. Posteriormente, se propone un marco para
la deteccin de emocin en texto basado en la teora de la valoracin (Appraisal
Theory), el modelado de la posible interpretacin (evaluacin) de las acciones
contenidas en la situacin descrita, de acuerdo con el conocimiento de sentido
comn sobre las acciones y las propiedades de sus actores y objetos. El trabajo
futuro incluye la ampliacin de los recursos creados para otros idiomas y la
ampliacin de la base de conocimiento de sentido comn de situaciones que puedan
suscitar una emocin, con fuentes externas de conocimiento. En este contexto, las
lneas de inters para la labor futura podran ser:

272

Las tcnicas de desarrollo para la personalizacin y adaptacin del


contenido de la informacin en funcin del contexto de usuario, a travs de
la explotacin de los contenidos subjetivos generados por el usuario, as
como el anlisis de las preferencias expresadas de forma explcita, a travs
de opciones en los perfiles de usuario y de forma implcita, a travs de las
conexiones en las redes sociales, las contribuciones a los foros,
comentarios, blogs o microblogs;
El diseo, la implementacin, la poblacin y la ampliacin de una base de
conocimiento para el modelado de preferencias de usuarios, para poder
realizar minera de opiniones de forma personalizada (es decir, en funcin
de los objetivos del usuario, las creencias, opiniones, observaciones,
sentimientos). En este contexto, las capacidades de bsqueda semntica
podran ser mejoradas y ampliadas de una manera ms centrada en el
usuario.

273















5HXQLGRHO 7ULEXQDOTX HVXVFULE HHQ HOGtDGHODIH FKD DFRUGyR WRUJDUSRU       DOD7H VLV
'RFWRUDOGH'RQ'xDODFDOLILFDFLyQGH     

$OLFDQWH     GH     GH     




(O
6HFUHWDULR

(O3UHVLGHQWH














UNIVERSIDAD DE ALICANTE
CEDIP
/DSUH VHQWH7HVL VGH' BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBK DVL GR
UHJLVWUDGDFRQHOQBBBBBBBBBBBBGHOUHJLVWURGHHQWUDGDFRUUHVSRQGLHQWH
$OLFDQWHBBBGHBBBBBBBBBBGHBBBBB

(O(QFDUJDGRGHO5HJLVWUR

La defensa de la tesis doctoral realizada por D/D


se ha realizado en las siguientes
lenguas:
y
, lo que unido al cumplimiento del resto de requisitos establecidos en la
Normativa propia de la UA le otorga la mencin de Doctor Europeo.
Alicante,
EL SECRETARIO

de

de
EL PRESIDENTE

You might also like