Computers in Human Behavior 64 (2016) 782e792

Contents lists available at ScienceDirect

Computers in Human Behavior

journal homepage: www.elsevier.com/locate/comphumbeh

Full length article

Towards computational discourse analysis: A methodology for mining

Twitter backchanneling conversations
Carlo Lipizzi b, Dante Gama Dessavre b, Luca Iandoli a, b, *,
Jose Emmanuel Ramirez Marquez b
a
Dept. of Industrial Engineering, University of Naples Federico II, Piazzale Tecchio 80, 80125 Napoli, NA, Italy
b
School of Systems and Enterprises, Stevens Institute of Technology, Castel Point on Hudson 1, 07030 Hoboken, NJ, USA

a r t i c l e i n f o a b s t r a c t

Article history: In this paper we present a methodology to analyze and visualize streams of Social Media messages and
Received 19 March 2016 apply it to a case in which Twitter is used as a backchannel, i.e. as a communication medium through
Received in revised form which participants follow an event in the real world as it unfolds. Unlike other methods based on social
3 July 2016
networks or theories of information diffusion, we do not assume proximity or a pre-existing social
Accepted 22 July 2016
structure to model content generation and diffusion by distributed users; instead we refer to concepts
Available online 8 August 2016
and theories from discourse psychology and conversational analysis to track online interaction and
discover how people collectively make sense of novel events through micro-blogging. In particular, the
Keywords:
Online discourse analysis
proposed methodology extracts concept maps from twitter streams and uses a mix of sentiment and
Conversational analysis topological metrics computed over the extracted concept maps to build visual devices and display the
Social media mining conversational flow represented as a trajectory through time of automatically extracted topics. We
Social representations evaluated the proposed method through data collected from the analysis of Twitter users’ reactions to
Big data the March 2015 Apple Keynote during which the company announced the official launch of several new
Online communities products.
Online conversation © 2016 Elsevier Ltd. All rights reserved.
Computer-mediated communication
Collective intelligence
Semantic analysis
Sentiment analysis
New product launch

1. Introduction breadth in which a large number of users generates and shares

contents through hyper-connected social networks while trying to
The large-scale adoption of Social Media (SM) is one of the most accomplish various types of tasks such as voicing their opinion,
relevant technological and social trends in the history of the providing or asking for help, sharing information, contributing to a
Internet. According to the Pew Research Center (Duggan, 2015), cause, reaching out to friends and acquaintances, or applying for
about three quarters of adult Internet users in the US spend memberships to a community.
considerable time on Social Network sites such as Facebook, Much of this interaction happens through online conversations
Twitter, Instagram, Pininterest, and LinkedIn, with 65% of online that can be tracked and mined to extract online analytics for
adults using social media sites (Perrin, 2015), with a tenfold in- different applications. A broad practical question for media analysts
crease in ten years (this percentage goes up to 70% and 59%, is the availability of reliable tools to summarize large conversa-
respectively, for the adults who use daily Facebook and Instagram). tional flows into effective representations to answer a number of
SM are thus creating interaction spaces of unprecedented size and “what” and “who” questions such as: what the users are talking
about? What matters most to them? What is trending right now?
Who is talking about what? Who is talking to whom? Etc. The
* Corresponding author. Dept. of Industrial Engineering, University of Naples answers to such questions are of immediate use in multiple ap-
Federico II, Piazzale Tecchio 80, 80125 Napoli, NA, Italy. plications in fields as diverse as marketing (Hoffman & Fodor, 2010,
E-mail addresses: [email protected] (C. Lipizzi), [email protected] pp. 1e11; Smith, Rainie, Shneiderman, & Himelboim, 2014), politics
(D.G. Dessavre), [email protected], [email protected] (L. Iandoli), (Bartlett, Froio, Littler, & McDonnell, 2013; Parker, 2014), or
[email protected] (J.E. Ramirez Marquez).

http://dx.doi.org/10.1016/j.chb.2016.07.030
0747-5632/© 2016 Elsevier Ltd. All rights reserved.
 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792 783

national security (Brachman, 2014; Klausen, 2015). Moreover, inspired to ideas and theories from studies in discourse analysis.
thanks to the possibility to track online interaction in an objective The proposed methodology is applied to extract social and concept
fashion, online conversations offer plenty of interesting empirical maps from micro-blogs streams and to compute a set of conver-
materials for the analysis and modelization of learning in virtual sational analytics for the design of visual devices able to display the
social networks, although predominantly in research on use of SM conversational flows.
in education (Greenhow & Greenhow, 2011; Yu, Tian, Vogel, & Since the maps are generated as the conversation unfolds, our
Kwok, 2010). approach helps to analyze how the structure of these graphs
In this paper our approach to answer “what” and “who” ques- changes in time and reflect not only what the participants are
tions is based on tracking and visualizing the process through talking about but also how they are talking about the event.
which large-scale interaction favors the emergence of shared Structural analysis is then used to visualize the conversational flow
meaning through a process of participatory sense-making (Jaegher and identify emergent conversational patterns and the process that
& Paolo, 2007). As we show later in the paper (Section 2), the generates them. The combination of social, semantic, and more
analysis of this process has received limited attention in the liter- traditional traffic metrics helps to observe conversational interac-
ature on online conversations, which instead could benefit from the tion and its outcome in a dynamic fashion, possibly offering in-
application of computational methods based on natural language sights on the role of social cognition on meaning generation and
processing, automated semantic analysis, and social network transformation (Jaegher & Paolo, 2007) and empirical grounding to
analysis as well as from theoretical and methodological insights better justify the use collaborative/social technologies to support
from studies in conversational analysis and discourse psychology informal and continuous learning (Siemens, 2005).
(Section 3). We report empirical findings obtained from one case study
Through the development of a methodology that combines related to the launch of new products, the Apple Keynote presen-
ideas and tools from conversation analysis, online analytics, and tation that took place in March 2015, a 100 min event during which
data visualization (Section 4), we provide a methodology and a tool the company launched the new MacBook and the much-awaited
to better handle the trade-off between in depth, small-scale, AppleWatch. We selected this event as an example of a social
qualitative analysis and processing of large amount of data gener- happening that is highly focused, that is object of wide attention
ated in conversations. We then show that SM-enabled conversa- and ample coverage from other media, and that is scheduled
tions can support sense-making about new events by favoring the regularly by Apple twice per year. The application to a marketing
accumulation of common ground generated by multiple partici- case is also used to offer an example of practical applications, but
pants when following a same event in real time (Clark & Brennan, the proposed method can be applied to other events that exhibit
1991) (Section 5). similar characteristics in terms of focus, popularity, and
Among the several available SM, we focus on Twitter for a expectations.
number of reasons. Twitter is popular and widespread among
Internet users (Duggan, 2015) and the access to its data is relatively 2. Current approaches to twitter streams analysis
easy and low cost with most users making their content public.
Tweets lack the complexity and richness of more structured mes- 2.1. Overview
sages, such as those that can be found in online forums, thanks to
the well-known 140 characters constraint. Finally, Twitter is The availability of reliable and efficient methods for the collec-
increasingly used as a backchannel to follow up with events tion and analysis of SM data is an increasing concern for companies
unfolding in the real world (Dork, Gruen, Williamson, & and policy makers (Hoffman & Fodor, 2010, pp. 1e11; Smith et al.,
Carpendale, 2010). Twitter backchannel feeds are also frequently 2014). In particular, without the ability to listen to and measure
used to support virtual participation to a real event (McNely, 2009) SM flows, media analysts miss a valuable opportunity to exploit the
by providing participants with a parallel communication channel to abundant and informative data that is incessantly generated by
exchange comments about what is going on. millions of online users and transform this information into
Backchanneling represents an interesting case for the empirical actionable analytics to support more responsive and data-driven
analysis of SM because the short duration of the events being fol- decision-making (Culnan, McHugh, & Zubillaga, 2010).
lowed helps to create a strong focus in the discussion, unlike the These needs bring new online analytic challenges due to the
more unstructured and chaotic feature of uneventful micro- growing variety of online media to be monitored and the conse-
blogging. More importantly, backchanneling is akin to virtual quent production of massive quantity of data to be analyzed. For
participation to a real world event, which participants are trying to instance, different types of SM enable different types of conversa-
make sense of and to which they are attracted for a variety of in- tional tasks and allow users to assume multiple online identities,
dividual motivations. We suggest that this sense-making process is especially among younger users (Bolton, Aksoy, van Riel, &
easier to observe in those cases in which the participants are Kandampully, 2013; Williams, Crittenden, Keo, & McCarty, 2012).
generally ignorant about the ways the event is going to unfold, Facebook, for instance, can be used for sharing personal contacts,
although they may have some pre-existing knowledge and expec- news and experiences with friends, family and other acquaintances.
tations about its development and the subject in general. This is the Twitter is born with the idea of giving people a simpler tool to
case for events such as a TV political debate (Shamma, Kennedy, & answer the question: what are you doing now? Lately, it has
Churchill, 2009), a new episode of a TV show (Harrington, evolved into a public space in which people broadcast their opin-
Highfield, & Bruns, 2012), and the use of SM to coordinate in ions or whatever they think is worthwhile of public attention. SM
emergency situation (Hughes & Palen, 2009; Mills, Chen, Lee, & conversations are also very different from online forums, which are
Raghav Rao, 2009). often directed towards a common topic and support longer lasting
In the next section we show that the dominant approaches to conversations that are attended by a smaller number of
SM mining are biased towards the analysis of user-generated con- participants.
tent rather than of the process through which the same content is Consequently, alternative ways to visualize these new types of
generated. Our challenge is instead to dig deeper into Twitter social streams have been proposed recently while methods devel-
streams to observe the emergence of shared meaning in online, oped to analyze more “traditional” types of online conversations
large-scale conversations by using a quantitative methodology such as in online forum (Melville, Sindhwani, & Lawrence, 2009) or
784 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792

email exchanges (Vie !gas, Golder, & Donath, 2006) are not appro- “listening platforms” (Gartner, 2013, pp. 1e98), aimed at displaying
priate to the always-on, large scale, fragmented conversations that social data in visual dashboards to support analysts and decision
take place on the new media. makers (Cui et al., 2011; Dork et al., 2010; Pousman, Stasko, &
In this paper specifically we focus on an application of Twitter Mateas, 2007; Sack, 2000). Companies adopt listening platforms
that is achieving increasing popularity known as backchanneling to exploit social intelligence and turn SM data into actionable
(Dork et al., 2010). Backchanneling is the practice of using com- marketing and business knowledge. SM metrics may be based on
puters to sustain a real-time, online conversation alongside a pri- raw indicators, such as number of comments posted, friends
mary event that is unfolding offline or on another medium, such as reached, and the like; however, the creation of added value metrics
TV shows and conferences (Wikipedia definition). Twitter back- usually requires meaningful and smart aggregation of raw counters,
channeling data provide a more focused opportunity for SM min- such as sentiment, buzz, rating, influence, exposure, and virality
ing. Examples include monitoring TV audience reactions (Highfield into compound indicators. Opinion mining and sentiment analysis
& Bruns, 2013; Nielsen, 2013; Wakamiya, Lee, & Sumiya, 2011), are part of the standard features of many commercial platforms for
understanding ad predicting traders’ reactions in the stock market monitoring SM, such as Metavana, Sysomos, or Crimson Hexagon.
(Evangelopoulos, Magro, & Sidorova, 2012), or measuring the Sentiment analysis is applied to automatically identify and assess
diffusion of ideas in online discussions (Tsur & Rappoport, 2012). In value judgments expressed by online users with respect to a spe-
all of these cases the objective is to extract meaningful and repre- cific object (e.g., a brand) (Chamlertwat, Bhattarakosol, &
sentative feedback from online conversations and exploit this Rungkasiri, 2012).
feedback to help decision makers to identify effective courses of
actions. 2.2. Limitations of current methods
Because the amount of data to track and analyze in back-
channeling conversations can easily become huge, the empirical None of the methods and approaches outlined above assumes
analysis of user-generated content typically requires the adoption conversational interaction as the primary unit of analysis but they
of quantitative and automated data mining tools. The most rather focus on the output generated by the users or on the social
commonly used approaches in social web mining are based on infrastructure that enables information exchange.
theories of Information Diffusion in Networks. Another important Information diffusion and Viral models are content-neutral i.e.,
category of methods is based on computer-based semantic analysis they typically assume that meaning is not problematic and that it is
and includes applications such as sentiment and reputation anal- preserved when information is passed along social links while little
ysis or text mining of the content published on SM site through attention is paid to the recursive and cyclic nature of the interaction
natural language processing algorithms. taking place between users. This view of information as a neutral
Diffusion theory provides conceptual and methodological tools token to be passed between an active sender and a passive receiver
to assess the rate of information circulation throughout a social is due to the influential channel metaphor proposed by Claude
network and has been applied to off-line as well as to online studies Shannon (Sloane & Wyner, 1948, pp. 5e83). One consequence is
such as the adoption of new ideas or practices in a community that SNA, or viral models for that matter, work better when the
(Rogers, 1995), the spreading of computer viruses (Vespignani, semantic alteration of content produced by users is not problem-
2005) or the implementation of viral marketing in online net- atic, e.g. when diffusion is about transferring packaged information
works (Leskovec, Adamic, & Huberman, 2007). The unit of analysis (e.g. a computer virus) or very short pieces of content, e.g. Internet
is the structural organization of connections (“topology”) between memes (Bauckhage, 2011).
users in a social network and the ways through which it favors or Approaches based on semantic analyses are characterized by a
hinders the diffusion of a message. Social graphs can be analyzed similar bias: the focus is on the automated extraction of meaning
through social network analysis models (SNA) aimed at repre- from the messages propagating in a network, but little attention is
senting structural properties of networks such as centrality, links paid to the social dynamics through which messages are created
distribution (Barab! asi, Albert, & Jeong, 2000), embeddedness (Uzzi, and modified while diffusing in the network. Semantic analysis
1997), ties strength (Granovetter, 1973), and structural holes (Burt, methods also have technical limitations when applied to SM min-
2004). ing because most of these methodologies have been developed for
Other and more recent computer-enabled methods are based on more coherent and traditional texts rather than for analyzing
semantic analysis and natural language processing techniques for conversations. While td-idf and its variations are used in several
the automated extraction of meaning and sentiment from the applications such as in search engines (Croft, Metzler, & Strohman,
messages that are exchanged in a network (Brown, 2012; He, Zha, & 2010), their applicability to SM conversation has been questioned
Li, 2013; Saif, He, & Alani, 2012). Statistical methods to extract in- (Hu & Liu, 2012; Qian, Zhou, Zhang, & Zhao, 2013). Similarly, co-
sights from the text are also widely used. For example, frequency occurrence cannot be easily applied to dispersed online conversa-
weighting (such as term frequency - inverse document frequency, tions e such as those happening on Twitter - because the text,
td-idf) is a popular technique used to extract important terms out of typically a sparse collection of short messages produced by multi-
a corpus of documents according to their relevance (Manning, ple authors in a dynamic way, often lacks the coherence of a written
Raghavan, & Shutze, 2009, pp. 1e569). Another commonly used document or of conversations directed at discussing specific topic
method to analyze text is co-occurrence. According to this method, (Naveed, Gottron, Kunegis, & Alhadi, 2011). Because of these limi-
two or more words are considered semantically related when they tations, sentiment/opinion analysis is often chosen as the main
frequently appear within a given range of proximity in the text option. Sentiment analysis, however, is just concerned with the
(degree of separation). Using co-occurrence, it is possible to create a polarity of messages and does not provide any means to analyze
network of words, with nodes being the words and edges being their content and discursive function.
their co-occurrences. Creating topological representations from Table 1 summarizes the most common SM mining approaches in
words in text based on word co-occurrence has been applied to terms of key assumption, dominant metaphor and modelization
extract information on the structure of the text in online chats approach.
(Introne & Drescher, 2013). Diffusion theories, including information exchange in social
The design of effective tools to represent SM content is a priority networks and viral models, assume that social connections preexist
for the development of commercial products, generally known as to the actual exchanges. The structure of the relational or proximity
 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792 785

Table 1
Dominant approaches to SM analysis.

Approach Key assumption Dominant Modeling

metaphor

Information exchange in Information percolates throughout social links Communication Communication networks modeled
social networks channel through sociographs
Viral models Information is diffused through contagion mechanism that requires Contagion Proximity Networks, infection likelihood
physical, virtual or cognitive proximity and threshold
Semantic and Sentiment Meaning can be detected from messages via lexical, statistical, or semantic Information Statistical techniques
Analysis analysis processing Automated classifiers of the polarity of
Social mood individual judgments
Listening Platforms Data visualization supports decision Undisclosed makers' insights Traffic, flow Data visualization techniques

network is thus given and any feedback between the content being from their peers to accomplish a variety of tasks such as: a) un-
exchanged and the formation of the network is typically derstand what's new and exciting with the new product; b) who is
disregarded. buying in and why; c) find reasons, motives, and evidence to justify
Semantic and Sentiment analyses assumes that conversational purchasing intentions.
interaction can produce order and structure, for instance in terms of Following Clark and Brennan (1991) we assume that an online
positive or negative collective feeling; however, neither approach conversation is a form of joint action requiring participants to co-
investigates how these structures emerge from or are influenced by ordinate on content and on process. Coordination on content re-
interaction. quires contributors to achieve mutual understanding and build
Finally, Listening Platforms rely heavily on data visualization common ground by asking for and providing evidence to confirm
devices aimed at facilitating interpretation by analysts and decision they understand each other and by adding validated new infor-
makers, but these representations are not typically driven by the- mation to their stock of mutual knowledge. Feedback comes
ory; moreover, since visual dashboard are developed as commercial through verbal requests (“what do you mean by … ?”) or nonverbal
tools, the assumptions and the algorithms behind the tools are acts like nodding, frowning, pointing, humming, etc. Coordination
often undisclosed and there is no way to assess them in terms of on process is based on the reply structure; turn-taking serves the
their theoretical grounding. purpose of achieving consensus that a new piece of information
In conclusion, most of the existing tools and approaches are not adds up to the stock of mutual knowledge.
aimed at detecting conversational patterns in SM streams. In the The narrow focus on a specific event that characterizes back-
next section we will argue instead that the identification of channeling streams helps participants to achieve coordination on
conversational patterns can help to assess not only how people feel content, while the use of other Twitter functions supports a mini-
about a new product (sentiment) but to develop a better under- mal level of reply structure. These functions include “mentions”
standing of what people say and how people talk to each other. (mentioning the Twitter id of another user in the tweet, @user_id),
Following some key tenets of Discursive Psychology, we explore the “retweets” (reposting a tweet initially posted by another users,
constructive and pragmatic nature of online conversations in SM RT:<tweet〉), and “replies” (reply directly to another user, Reply
backchanneling and try to incorporate some of the ideas and to@user_id).
principle used in conversational analysis in a computational DP and conversational analysis offer a set of theoretical con-
methodology for the analysis of online social streams in Twitter. structs and mechanisms to model conversational content gener-
ated on micro-blogging platforms.
3. A discursive psychology approach to online conversational First, micro-blogs conversations may lead to the construction of
analysis of Twitter streams shared representations; however, since process coordination is
constrained by the limited affordances that Twitter offers in terms
In this paper we follow a theoretical perspective based on of reply structure, we expect these representations to be more
Discourse Psychology (DP) (Potter & Edwards, 1999), Social Rep- fleeting than they would be in a real, longer term discourse. Sec-
resentation theory (Moscovici, 2001), and conversational analysis ond, the concept of adjacency pairs (Clark & Schaefer, 1989) can be
(Clark & Brennan, 1991) to model micro-blog streams generated in exploited to map the content of tweets over a collective concept
backchanneling applications as loosely-coupled conversations. DP map that is updated as the conversation proceeds. In ordinary
questions the cognitivist view according to which language is a conversations adjacency pairs are two ordered utterances pro-
means to access what is in people mind and assumes instead a duced by speakers, as in a question/answer pair (e.g.: A: “Where is
pragmatic view in which utterances are linguistic acts people Connie?”, B: “At the store”). A conversation can then be repre-
perform to accomplish specific tasks (Clark, 1996). A fundamental sented as a stream of adjacency pairs uttered within a turn-taking
meta-task is the need to understand each other in conversations structure.
through the accumulation of shared knowledge. According to In a Twitter conversation adjacency pairs are infrequent because
Potter and Edwards (1999) in DP shared knowledge representations most of the time participants do not reply to each other (Sysomos,
are discursive objects that people construct in talk and texts and 2010) and, even when this happens, the conversational flow is
the objective of conversational analysis is to investigate the way dispersed in a linear, time-based visualization of posts. It is then
these representations are constructed and used to perform action more likely that users follow the stream by visualizing local por-
(e.g. assigning blame, eliciting invitations, and so on). tions of it by proximity to their own or others’ contribution of in-
The kinds of questions that discursive psychologists might ask terest. This local focus helps to sustain a conversational scope that
regard how particular descriptions of a concept are used to do is situated in a specific time and screen space. In our methodology
particular things. In the cases analyzed in this paper, the official we model this aspect by reframing adjacency in terms of socio-
launch of a new, untested product around which there is a lot of semantic proximity: two contributions are related because they
expectation, but whose features, uses and even market positioning overlap spatially; two users are related because they are talking
are unclear, is a case in which potential customers look for feedback about a same entity. As we show in the next section, this proximity
786 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792

perspective is used to extract the content generated by users in a ! Stop-wording: we filter the words from the previous step using
twitter stream and reassemble it into a concept map that is updated a customizable list of words we want to remove from the con-
as the conversation proceeds. versation (e.g. stop words, search keyword, etc.)
A concept map is a network of keywords that can be analyzed ! Lemmatization: inflected forms of a same word are grouped
through topological and semantic metrics. For instance, it may together so they can be reduced to a single base lemma (e.g.
exhibit a certain arrangement of nodes and links (topology) and work is the base lemma for worker, worked, working etc.).
some degree of internal lexical diversity. Keywords may be orga- ! Bi-gramming: we extract most frequent bigrams from the
nized in emerging topical clusters. Finally, a number of network stream of tweets and process them as a single word (e.g.
metrics can be used to measure the structural properties of the Apple_watch).
network and how they evolve in time. In the following section we
show how we implemented these ideas through a methodology The dataset is partitioned in time windows that can be identi-
that is able to dynamically and automatically extract and visualize fied either establishing a fixed duration or a given number of
concept maps from twitter streams; we then present semantic tweets.
metrics and data visualization formats to monitor the discursive
interaction taking place in Twitter. 4.2. Step 2: Semantic structure extraction

4. Methodology The output of the pre-processing step is a dataset composed by

people who sent the messages and keywords they used. The con-
In this section we describe a methodology and its related soft- nections between the set of users U and the set of keywords W can
ware tools we have developed to extract metrics and contents from be represented through a 2-mode (“bipartite”) network G ¼ (U, W,
micro-blog streams. We then use metrics and contents as inputs for A) such that if ui is a Twitter user and wj is a keyword, there is an
visual tools designed to display how a backchanneling conversation edge aij ¼ ðui ; wj Þ2A if and only if ui sent a tweet containing the
unfolds in time. In the following we provide a step-by-step keyword vj . The W-projection of G % GW- is computed to extract the
description of the procedure we developed to build such network of keywords mentioned in the conversation

GW ¼ G & GT ¼ ðW; Aw Þ where Aw ¼ fðp; qÞ; d x2U : ðp; xÞ2A and ðq; xÞ2Ag

visualization. where GT is the transpose of G. GW represents a network of words

The core step of the method is the extraction of the semantic that are connected because they are “shared” by users directly or
and social structure. In particular in step 2 we operationalize the through some degree of separation. Those structures are “socially
idea of logic adjacency pairs and, in step 3, the concept of reply generated” and they include logic adjacency pairs obtained
structure, as described in the previous section. We then represent through the concatenation of keywords linked together by
common ground accumulation through dynamic concepts maps different users. The process is cumulative because it supports at-
that are built and modified as the conversation proceeds and from tempts of expansion or strengthening of shared content: such
which our algorithm is capable of extracting clusters of keywords content is expanded as users create links to new keywords and
corresponding to emerging topics. A metric of semantic continuity strengthened when associations between keywords are re-
is then proposed in step 4 to detect and display changes in the asserted (e.g., through retweets), in which case the weight asso-
common ground accumulation process as signaled by semantic ciated to an existing link is increased as the associations becomes
switches occurring during the conversation. more frequent.
After calculating basic topological metrics for the keywords
4.1. Step 1: Data collection and pre-processing network (such as density or centrality), we apply clustering analysis
to extract words clusters that potentially corresponds to topics in
In this step, an event around which a backchannel conversation the conversation.
is likely to be started by interested users must be identified. Back- To identify words clusters each we use a method based on based
channeling events have to be popular enough to attract the atten- on k-cores. A k-core H of a graph G is a maximal connected sub-
tion of a reasonably large number of users for a short time. graph of G in which all vertices have degree at least equal to k. It can
Examples of such events can be the launch of a new product by a be proved that H is the maximum subgraph with this property
famous company, a live sport, or a TV show. (Alvarez-Hamelin, DallAsta, Barrat, & Vespignani, 2008). The value
The tweets are downloaded via Twitter search Application of k is determined as follows:
Program Interface (API), through a Python scripts, using keywords
related to the event. In addition to their 140 characters text, tweets a) the k-core algorithm generates a list of “core values” per each
contain a large amount of meta-information, which includes items node and we select as k value the dth largest value in the list
such as user id, timestamp to location, language, and additional of “core values”, where d is the number of topics to be deter-
information on the users. For our analysis we extract from each mined for each partition. This parameter is proportional to the
tweet only a triple < u; W; t > where u2U is a user sending the total number of keywords in the partition and it can be varied
tweet, W is a vector of words wp 2W composing the text of the heuristically to fine tune the granularity of the topic
tweet, and t is the time the tweet has been posted. The word vector representation;
is obtained as follows: b) we then look for clusters into H through the Louvain modularity
community detection method based on greedy optimization
! Tokenization: the raw text in each tweet is mapped onto a word (Black, 2005; Blondel, Guillaime, Lambiotte, & Lefebvre, 2008),
vector; and extract d clusters in each partition;
 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792 787

4.3. Step 3: Reply structure extraction most relevant keyword in each circle is determined by computing
the similarity between all the words in the cluster through pairwise
In Twitter Users may reply to each other using the reply func- comparison; then the keywords are ranked in terms of their
tion, “mentions” of other users in their tweet through special average, intra-cluster similarity. The top keywords are then used to
characters (@ or #) and sometimes the retweet functionality. So it is label the circles and the semantic distance among the representa-
possible to trace back the conversational links between any two tive labels of two contiguous circles is computed and reported on
users and leverage this information to capture and visualize the the vertical axis in Fig. 1. The Natural Language ToolKit (NLTK) in
reply structure in tweets stream. The reply structure is a social Python was used to perform the above calculations.
network from which a number of social metrics can be extracted. In
particular, we computed the clustering coefficient to check for the 4.5. Summary
emergence of subgroup of users that are replying to or mentioning
each other in a given time slice. We refer to this metric as relational The proposed methodology supports the visualization of
clustering coefficient to distinguish it from the clustering coeffi- conversational patterns in Twitter backchanneling through the
cient that is computed over the keywords network for the identi- combination of several analytics that are mined from Twitter
fication of topics. streams. More specifically, we extract topological representations
from the content and the reply structure that is generated by the
4.4. Step 4: Visualization users and use a set of topological, traffic and sentiment metrics to
observe how these topologies evolve in order to track what users
The visualization proposed in this paper displays the content are talking about and how they are talking. The proposed visuali-
generated by the users, processed through the procedure described zations help to summarize this evolution to support media analysts
in step 2, as the conversation unfolds along the timeline of the to develop a better understanding about how users are collectively
event of interest (Fig. 1). In particular we are interested in identi- making sense of the event. In the specific case study we present in
fying whether the conversation is able to achieve some level of the next section, the official presentation of new products, the
semantic convergence around a specific topic or related set of analysis of the backchanneling behavior can help analysts to un-
topics of interest for the participants. derstand on which particular aspects the users focus, if there are
Each chunk of tweets that are posted in a given time slice is dominant topics or themes, if some topics generate more conver-
represented by a bubble whose diameter is proportional to the sational interaction than others, and how is the sentiment about
number of messages posted in that time slice. In this way the size of specific attributes or features of the new product.
the bubble highlights the moments in which the conversation is
more crowded. Each bubble contains a white dot at the center, the 5. Results
larger is this dot is, the higher the number of topical clusters that
are identified in the time slice as described in step 2. In other words, This section presents the results obtained by applying the pro-
the white dot is a visual proxy to provide an idea of how much posed methodology to a corpus of 46,804 tweets in English only,
focused the conversation is in a particular time slice: for instance, a downloaded in a time window of 1h40’ during the Apple Keynote
small dot indicates that participants are very much focused on a Event in March 2015. The tweets were obtained via search through
few topics in particular time window. Twitter API using as keyword “apple”. In order to perform the
On the vertical axis we report the semantic similarity among search, we created a Python script interfacing Twitter API, down-
pieces of the conversation that are contiguous in time. The idea is to loading the tweets over time and storing them in a non-structured
observe if there is any semantic continuity in the stream of tweets database (MongoDB). We also used our own Python scripts to fully
to identify parts of the conversation in which there is potentially automate the methodology described in the previous paragraph.
some incremental accumulation of meaning or to visualize points Using those scripts, we processed the tweets we downloaded,
in which the conversation shifts abruptly to a different subject. In creating the graph in Fig. 2.
both cases the analysis of the event as well as of the tweets can be Fig. 2 shows the first visualization for the Apple Keynote event
later used to validate if proximity or semantic jumps are actually over time. The time axis is segmented into moments corresponding
present in the data. The semantic similarity between any two to the various parts of the presentation as derived from the
contiguous bubble is computed as the semantic distance between verbatim transcript of the event.
the most relevant keywords in each bubble by using Lin similarity A quick visual inspection of the whole graph shows that the
computed over an existing lexical database (Wordnet). First, the backchanneling conversation taking place on Twitter during the

Fig. 1. Example of variable visualizations. First figure shows constant number of tweets with focus of the conversation increasing. Second figure shows the number of tweets
increasing.
788 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792

Fig. 2. Visualization of the MacBook and AppleWatch release event e The timeline of the event is labeled from the presentation agenda and the transcripts of the keynote speech.

keynote presentation is highly fragmented and scattered; the sanitize the corpus manually before the analysis. The retrieval of
variance of the semantic distance is rather high, which implies that this sequence shows that our algorithm is able to isolate what may
there is, on average, little semantic continuity between adjacent appear to be an unusual sequence to an interpreter that is knowl-
chunks of tweets. This result seems to reflect the nature of the edgeable about the event.
event being followed, a keynote in which the speaker is in full It is interesting to observe that, overall, the average semantic
control of the presentation agenda and the speech touches upon a distance is lower in the MacBook window than for the Apple
large variety of topics. The scattered aspect of the conversation, Watch; we interpret this finding on the base of the higher famil-
characterized by abrupt semantic switches between a bunch of iarity of the users with the MacBook, a new product that, in fact,
different topics whose content mirrors quite well the keynote exhibited only incremental improvements compared to the previ-
speech content, shows that the conversation is highly reactive and ous model (e.g. better screen resolution, more computational po-
the users tend to process the event quickly, by sticking to the wer, improved battery life, and other minor design changes).
speaker's agenda and following the rhythm imposed by the event Because of the higher familiarity with the product and the marginal
timeline. novelties in the new release, the users were more likely to focus on
Fig. 2 also shows that the width of the semantic switches be- a few, related aspects and sustain a less hectic conversation.
comes quite high each time the keynote speaker shifted to a new Conversely, during the AppleWatch time window, this conti-
item in the agenda, notably when the speech turned from the Apple nuity is less apparent; we argue that, being the watch a truly new
TV to the iPhone, then from the iPhone to the MacBook, and finally product, the participants were mainly interested in the “here and
from the MacBook to the AppleWatch. now” and more likely to zigzag from one subject to another to
Despite the scattered character of the overall conversation, it is explore the novel functions and characteristics of the new product,
possible to notice that the variability of the semantic distance is more or less as they were introduced in the presentation.
different within different macro-topics and that there are some Variations in the semantic distance might actually signal two
streams in which the variation between consecutive bubbles is very different types of conversational dynamics: one characterized by
small. Here are some examples of “local” smooth conversations: lower variation that is more oriented to exploitation (e.g. more
focus, more in depth discussion, more analysis) and another char-
! launch-major refers to the introduction of a major novelty of the acterize by higher semantic volatility in which users are exploring,
new the iOS (Car Play, 10:16 am); perhaps in a chaotic and unstructured fashion, the meaning of what
! the sequence love-magic represents a first excited reaction to the is being said or presented.
introduction of the new MacBook at 10:30 am; Fig. 3 also shows that the bubbles in the AppleWatch portion of
! The sequence why-pushes refers to Apple “obsession” to make the keynote are bigger than those generated by the updates on the
its products thinner and thinner 10:35; iPhone and the presentation of the new MacBook. The data shows
! announcement-want-wants the first reaction of excitement to that the iWatch attracted more attention, being the real novelty in
the announcement of the Apple watch at 10:53 am; the keynote presentation and a product that had been waited for a
! sales-pay is related to the possibility to buy items through the long time after several announcements, some deferments, and
watch with Apple Pay (11:13 am); well-crafted, anticipated online buzz. The high level of excitement
! answer-leave refers to Apple Watch functionality that allows its about the AppleWatch is well represented by the big bubble labeled
users to make calls through the watch (11:15 am). “announcement” followed by “want” that appears when the new
device is introduced for the first time.
It is interesting to note that the sequence taylor-pro-korean re- The users’ focus, which in the proposed visual is given by the
fers to some tweets that we found, upon later analysis, to be ratio between the area in the white circle and the width of the
generated by advertisement messages unrelated to the event and bubble, is instead higher in the MacBook case, with the most
that our platform processed anyway because we decided to not focused bubble being love in time slice 174. The difference in focus
 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792 789

Fig. 3. Alternative visualization of the MacBook and AppleWatch release event.

can be interpreted again on the base of the different degree of MacBook and the AppleWatch were introduced.
novelty of these two products: the MacBook was a new release of a We assessed the trustworthiness of our method by triangulating
popular laptop, whose functions and characteristics were sub- the information extracted through the maps with the transcript of
stantially known to the public, so the interest was polarized just on the keynote speech and a set of external sources including blogs
a few new features and improvements. The AppleWatch was a and other media that covered the event and that we assumed as a
totally new product whose applications and target users were sort of baseline (ground truth e Aiello et al., 2013). We found out
relatively unknown, thus participants were less able to focus on that the content extracted through our algorithm was strongly
specific features and uses; on the contrary they seemed more keen aligned with the official agenda of the speaker and with what was
to explore different attributes and potential issues of the new echoed by other independent observers of the same event.
product.
The highest level of focus during the event was achieved during 6. Discussion and conclusions
the time window from 10:29 a.m.-10:30 a.m US west coast time,
when the MacBook's presentation started. The second highest focus 6.1. Contribution and implications
peak took place during the time window 10:53 a.m.-10:54 a.m.,
when the new smartwatch was introduced to the world. So, these In this paper we have presented a methodology to extract and
results suggest that during live events, Twitter users attention is visualize in an automated way the content generated by Twitter
highly reactive and deeply focused on what they are following. users in backchanneling conversations. We have presented
In Fig. 3 we represent the same event, but this time the size of empirical findings from a study in which we applied our method to
the bubbles is proportional to the relational coefficient, a metric the analysis of a backchanneling case related to a well-known
that indicates the presence of more dense conversational clusters in event. The method has been developed using as theoretical
the tweets stream generated in a given time slice as determined by perspective based on discourse psychology and conversational
the use of reply, mention, and retweet. In other words, the bigger analysis studies to increase our understanding of the process
the bubble, the higher the amount of conversational interaction. through which participants use online conversations to make sense
Using this visual it is possible to identify moments and topics in the and generate shared representations of the event they are
event in which the participants engaged more intensively in attending in a virtual fashion. Our findings show that the proposed
conversational exchanges. method is able to capture and quantify several interesting features
Our data in Fig. 3 show that the highest level of conversational of backchanneling conversational interaction such as: how atten-
interaction were achieved at the beginning of the event, during the tion and focus change across the event, identification and mea-
opening and the phase in which some updates about upcoming surement of semantic switches in the conversation, and
new products were presented (Apple TV, new iPhone). Instead in identification of moments and topics characterized by the presence
the time windows of the MacBook and AppleWatch, the number of denser conversational clusters.
and size of conversational clusters drop significantly. By later in- By identifying measurable structural properties in the massive
spection of the tweets' content we found that this behavior can be and continuous flow of data that typically characterize SM stream,
explained by the high pace and strong focus of the keynote event: our method allows analysts to identify underlying conversational
during the non-core phases such as the opening and the updates, patterns that could be used to profile conversations and investigate
users were less interested in the event content and had more time whether certain configurations are associated to outcomes of in-
to interact with other participants; during the core phases instead, terest; for instance, in marketing conversational patterns may be
the users attention was mainly directed at commenting on the informative about or even predictive of customers' satisfaction,
products’ features. Another factor that may have contributed to purchase intention, or actual consumers' behavior. The analysis of
limited conversational interaction could be the information over- the “local” conversations can give highly detailed insights about
load due to sensible increase in traffic that occurred when the customers’ concerns or sources of excitement or just point out to
790 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792

moments and topics that induced more conversational interaction. term, self-organized patterns that might be generated in SM that
More broadly, conversational patterns may be predictive of in- are supportive of more sophisticated communication, such as
dividual or collective behavior, in terms of belief formation, inten- Facebook.
tion, and individual or collective action (Ajzen, 1991). For instance, One inconvenience of using SM data is that the sample of users
in another study we have found empirical evidence that back- may not be statistically representative. This problem could be
channeling conversational patterns were predictive of box office overcome at least in part by extracting demographics from users’
revenues for newly released movies (ref. omitted for anonymity online profile or through the set-up of ad hoc, closed discussions in
sake). While there are many studies on the impact of peer to peer which participants are recruited in an unbiased way. However, even
influence on persuasion and behavior in many diverse fields demographic controls could not overcome the fact that the audi-
ranging from marketing (Leskovec et al., 2007) to health behavior ence of the event is self-selected: the Apple event is a case in point
(Centola, 2010), interaction is generally modeled as exchange of because it is reasonable to assume that Apple fans were over-
information packages enabled by social links and little empirical represented among the participants who followed the keynote.
analysis is offered to keep into account the participative, iterative, This may explain the high level of excitement for the AppleWatch
and constructive nature of those exchanges. that was probably behind the record early sales of the product, but
In particular we think that that more empirical research that proved to not be enduring as the subsequent sales level
assuming conversations as unit of analysis in addition to social dropped significantly and were later judged as disappointing from
network topology is needed to improve our understanding of how market analysts. A possible way to escape the self-selection bias is
information and knowledge is created and gained through online to triangulate the findings obtained through SM mining with those
interaction. For instance some conversational patterns may be obtained through more neutral and traditional data collection
associated to more informative interaction that could possibly lead methods.
to individual and collective learning. The study of learning in virtual Another critical issue in this respect is the increasing use of e-
social networks based on conversational analytics could provide bots publishing tweets that are generated by machines. The
deeper insight on how online conversation and other collaborative detection of these manipulation attempts is not easy to perform,
tool can be used and also improved to facilitate learning and was outside the scope of this work and we are aware that several
improve quality of online debate. researchers are developing algorithms to automatically detect
From the practical point of view, the proposed methodology spamming and other artificial ways to bias the conversation (Zhao
could support marketing and opinion analysts in better managing et al., 2014). However it was comforting for us to find out that
the reach-richness tradeoff that is typical of marketing or social despite the lack of countermeasure to spam attempts, our algo-
research. Traditionally, marketing analysts use qualitative research rithm was able to isolate one local conversation that was originated
methods, such as focus groups, applied to small samples of pro- by unrelated advertisement. Thus, a future development of our
spective or actual customers to gain an in-depth understanding of work may concern its application to the detection of possible spam
the factors influencing customers' preferences and of the causal based on semantic analysis, very much in the same way as human
relationships between product attributes, product performances, analysts could spot incoherent segments of conversation.
and customers’ needs. However traditional qualitative methods Our initial validation allowed us to pass a face validity test and
have two major drawbacks: they require subjective interpretation achieve some degree of external validity through the comparison of
of the findings and cannot be applied to large sample because the maps generated by our tool with content on the same event
qualitative analysis is very costly and time consuming. On the other available through independent sources. A more structured and
hand, large-scale, quantitative traditional market analysis, based on rigorous topic validation method is of course required, but given
standard surveys tends to be shallow, especially when it comes to the complexity and methodological nuances involved in this type of
the identification of emerging/unexpected trends. assessment and the space constraint, it was not possible to include
More broadly the proposed method could be used in other ap- a detailed account of the validation methodology in this paper.
plications in which analysts or decision makers want to learn from These materials are however available to interested readers on
an event for which backchannel data are accessible almost in real request.
time, such as in emergency management (Hughes & Palen, 2009), We collected tweets with no language restriction but we only
political mobilization (Theocharis, Lowe, van Deth, & García- considered English posts in the analysis. This is a limitation for
Albacete, 2015) or live events (Ross, Terras, Warwick, & Welsh, events that have an international audience. On the other hand,
2011). because we were interested in spotting conversational patterns and
because people need to speak the same language to have a con-
6.2. Limitations and future developments versation, we think that the non inclusion of tweets in other lan-
guage was not relevant for our purposes. Besides, our method can
We decided to use Twitter because of its popularity and con- be applied to other languages by using the appropriate dictionary.
venience. The proposed method could be applied to other more Since our method is fully scalable, further replications of the
popular SM through the appropriate changes to the data ingestion proposed analysis over a large sample of comparable events could
step; with more conversational venues such as online forums the be desirable to identify statistical correlation between some of the
linguistic and semantic complexity of conversations in which users metrics we propose and provide stronger empirical validation of
are not constrained to use very short messages would require more the method; however, the data collection for a large sample of
sophisticated semantic preprocessing than the relatively simple, similar events is time consuming and very expensive, especially if
lexical approach that we present in this paper. Moreover, user- the analysis requires the acquisition of past data from professional
generated content in SM such as Facebook does not have the data providers, whose costs may prove to be unaffordable to most
same level of accessibility because users tend to not share their academic budgets. For this reason we preferred to focus on single
content publicly and Facebook endorses a less liberal data access case study to show the potentiality and usability of our method-
policy. ology and provide some validation based on empirical findings
A limitation of the present study is that the observation interval from real event. In a parallel line of work, we are applying our
is limited to an event of less than 2 h. While the short duration analysis to a larger sample of conversations to find out if the
helps to create focus, it may prevent the observation of more long- conversational patterns elicited through our method can help to
 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792 791

predict consumers behavior in the purchasing of newly launched Highfield, T., & Bruns, A. (2013). More than a backchannel: Twitter and television.
Journal of Audience Reception Studies, 1e5.
products.
Hoffman, D., & Fodor, M. (2010). Can you measure the ROI of your social media
marketing? (pp. 1e11).
Acknowledgements Hughes, A. L., & Palen, L. (2009). Twitter adoption and use in mass convergence and
emergency events. International Journal of Emergency Management, 6(3e4),
248e260.
The research leading to these results has received funding from Hu, X., & Liu, H. (2012). Text analytics in social media. In Mining text data (pp.
the Strategic Research Counsel at the Academy of Finland under 385e414). Boston, MA: Springer US.
Introne, J. E., & Drescher, M. (2013). Analyzing the flow of knowledge in computer
grant agreement n. 293446 e Platform Value Now: Value capturing mediated teams (p. 341). Presented at the CSCW '13: Proceedings of the 2013
in the fast emerging platform ecosystems. conference on Computer supported cooperative work, New York, New York,
USA.
Jaegher, H. D., & Paolo, E. D. (2007). Participatory sense-making. Phenomenology and
References the Cognitive Sciences, 6(4), 485e507. http://doi.org/10.1007/s11097-007-9076-
9.
Aiello, L. M., Petkos, G., Martin, C., Corney, D., Papadopoulos, S., Skraba, R., et al. Klausen, J. (2015). Tweeting the Jihad: Social media networks of Western foreign
(2013). Sensing trending topics in Twitter. IEEE Transactions on Multimedia, fighters in Syria and Iraq. Studies in Conflict & Terrorism, 38(1), 1e22.
15(6), 1268e1282. Leskovec, J., Adamic, L. A., & Huberman, B. A. (2007). The dynamics of viral mar-
Ajzen, I. (1991). The theory of planned behavior. Organizational Behavior and Human keting. ACM Transactions on the Web (TWEB), 1(1), 5.
Decision Processes, 50(2), 179e211. Manning, C., Raghavan, P., & Shutze, H. (2009). An introduction to information
Alvarez-Hamelin, J. I., DallAsta, L., Barrat, A., & Vespignani, A. (2008). K-Core retrieval (pp. 1e569). Cambridge: Cambridge University Press.
decomposition of Internet graphs. Networks and Heterogeneous Media, 3(2), McNely, B. (2009). Backchannel persistence and collaborative meaning-making. In
1e23. Proceedings of the 27th ACM international conference on Design of communication
Baraba!si, A. L., Albert, R., & Jeong, H. (2000). Scale-free characteristics of random (pp. 297e304). ACM.
networks: The topology of the world-wide web. Physica A: Statistical Mechanics Melville, P., Sindhwani, V., & Lawrence, R. (2009). Social media analytics: Chan-
and its Applications, 281(1), 69e77. neling the power of the blogosphere for marketing insight. Proc. of the WIN,
Bartlett, J., Froio, C., Littler, M., & McDonnell, D. (2013). New political actors in Europe: 1(1), 1e5.
Beppe Grillo and the M5S. Mills, A., Chen, R., Lee, J., & Raghav Rao, H. (2009). Web 2.0 emergency applications:
Bauckhage, C. (2011). Insights into Internet memes. In International AAAI conference How useful can Twitter be for emergency response? Journal of Information
on web and social media (pp. 42e49). Privacy and Security, 5(3), 3e26.
Black, P. E. (2005). Dictionary of algorithms and data structures at the National Moscovici. (2001). Social Representations: Explorations in social psychology. New
Institute of standards and Technology (NIST). York: NYU Press.
Blondel, V., Guillaime, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of Naveed, N., Gottron, T., Kunegis, J., & Alhadi, A. C. (2011). Searching microblogs:
communities in large networks. Journal of Statistical Mechanics: Theory and coping with sparsity and document quality (pp. 183e188). Presented at the
Experiment, 2008(10), 1e12. CIKM '11: Proceedings of the 20th ACM international conference on Informa-
Bolton, R. N., Aksoy, L., van Riel, A., & Kandampully, J. (2013). Understanding gen- tion and knowledge management, New York, New York, USA.
eration Y and their use of social media: A review and research agenda. Journal of Nielsen. (2013, August 6). New Nielsen research indicates two-way causal influence
Service Management, 24(3), 245e267. between Twitter activity and TV viewership. Retrieved February 14, 2014, from
Brachman, J. (2014). Transcending Organization: Individuals and 'the Islamic state. http://wp-s3.socialguide.com.s3.amazonaws.com/wp-content/uploads/2013/
College Park, Maryland: START Analytical Brief. 08/080613.NielsenTwitterCausation.FINAL_.pdf.
Brown, E. D. (2012). Will twitter make you a better investor? a look at sentiment, Parker, E. (2014). Social media and the Hong Kong protest. The New Yorker magazine,
user reputation and their effect on the stock market. In Conference proceedings October 1st.
for the southern association for information systems SAIS conference (pp. 36e42). Perrin, A. (2015). Social media usage: 2005-2015. Retrieved from http://www.
Burt, R. S. (2004). Structural holes and good ideas. American Journal of Sociology, pewinternet.org/2015/10/08/social-networking-usage-2005-2015/.
110(2), 1e51. Potter, J., & Edwards, D. (1999). Social representations and discursive psychology:
Centola, D. (2010). The spread of behavior in an online social network experiment. From cognition to action. Culture & Psychology, 5(4), 447e458.
Science, 329(5996), 1194e1197. Pousman, Z., Stasko, J. T., & Mateas, M. (2007). Casual information Visualization:
Chamlertwat, W., Bhattarakosol, P., & Rungkasiri, T. (2012). Discovering consumer Depictions of data in everyday life. IEEE Transactions on Visualization and
insight from twitter via sentiment analysis. Journal of Universal Computer Sci- Computer Graphics, 13(6), 1145e1152.
ence, 18(8), 973e992. Qian, W., Zhou, A., Zhang, Z., & Zhao, B. (2013). Identification of collective view-
Clark, H. H. (1996). Using language (vol. 4). Cambridge: Cambridge University Press. points on microblogs. Datak, 87(C), 374e393.
Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. Perspectives on Rogers, E. M. (1995). Diffusion of innovations. Free Pr.
Socially Shared Cognition, 13(1991), 127e149. Ross, C., Terras, M., Warwick, C., & Welsh, A. (2011). Enabled backchannel: Con-
Clark, H. H., & Schaefer, E. F. (1989). Contributing to discourse. Cognitive Science, ference Twitter use by digital humanists. Journal of Documentation, 67(2),
13(2), 259e294. 214e237.
Croft, W. B., Metzler, D., & Strohman, T. (2010). Search engines: Information retrieval Sack, W. (2000). Conversation Map: An Interface for very large-scale conversations.
in practice (vol. 283). Addison-Wesley, Ed. Journal of Management Information Systems, 17(3).
Cui, W., Liu, S., Tan, L., Shi, C., Song, Y., & Gao, Z. (2011). TextFlow: Towards better Saif, H., He, Y., & Alani, H. (2012). Semantic sentiment analysis of twitter (Vol. Part I,
understanding of evolving topics in text. IEEE Transactions on Visualization and Volume Part I). In Presented at the ISWC'12: Proceedings of the 11th international
Computer Graphics, 17(12), 2412e2421. conference on the semantic web. Springer-Verlag.
Culnan, M., McHugh, P., & Zubillaga, J. (2010). How large US companies can use Shamma, D. A., Kennedy, L., & Churchill, E. F. (2009). Tweet the debates: Under-
Twitter and other social media to gain business value. MIS Quarterly Executive, standing community annotation of uncollected sources. In Proceedings of the
243e259. first SIGMM workshop on Social media (pp. 3e10). Retrieved from http://dl.acm.
Dork, M., Gruen, D., Williamson, C., & Carpendale, S. (2010). A visual backchannel for org/citation.cfm?id¼1631148.
large-scale events. Visualization and Computer Graphics. IEEE Transactions on, Siemens, G. (2005). Connectivism: A learning theory for the digital Age. Interna-
16(6), 1129e1138. tional Journal of Instructional Technology and Distance Learning. Retrieved from
Duggan, M. (2015, August 19). The demographics of social media users. Retrieved http://er.dut.ac.za/handle/123456789/69.
from http://www.pewinternet.org/2015/08/19/the-demographics-of-social- Sloane, N., & Wyner, A. (1948). A mathematical theory of communication. Claude E.
media-users/. Shannon: Collected Papers (pp. 5e83).
Evangelopoulos, N., Magro, M. J., & Sidorova, A. (2012). The dual micro/macro Smith, M. A., Rainie, L., Shneiderman, B., & Himelboim, I. (2014). Mapping twitter
informing role of social network Sites: Can twitter macro messages help predict topic networks: From polarized crowds to community clusters. Retrieved from
stock prices? Informing Science: the International Journal of an Emerging Trans- Pew Research Internet Project http://www.pewinternet.org/2014/02/20/part-2-
discipline, 15, 247e267. conversational-archetypes-six-conversation-and-group-network-structures-in-
Gartner. (2013). Hype cycle for social software, 2013. Retrieved from http://www. twitter/.
gartner.com/newsroom/id/2575515. Sysomos. (2010). Replies and retweets on twitter. Retrieved from https://sysomos.
Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, com/inside-twitter/twitter-retweet-stats.
78(6), 1360e1380. Theocharis, Y., Lowe, W., van Deth, J. W., & García-Albacete, G. (2015). Using twitter
Greenhow, C., & Greenhow, C. (2011). Online social networks and learning. On the to mobilize protest action: Online mobilization patterns and action repertoires
Horizon (vol. 19(1), pp. 4e12). in the occupy wall street, indignados, and Aganaktismenoi movements. Infor-
Harrington, S., Highfield, T., & Bruns, A. (2012). More than a backchannel: Twitter mation, Communication & Society, 18(2), 202e220.
and television. Audience Interactivity and Participation, 13e17. Tsur, O., & Rappoport, A. (2012). What's in a hashtag?: content based prediction of
He, W., Zha, S., & Li, L. (2013). Social media competitive analysis and text mining: A the spread of ideas in microblogging communities (p. 643). Presented at the
case study in the pizza industry. International Journal of Information Manage- WSDM '12: Proceedings of the fifth ACM international conference on Web
ment, 33(3), 464e472. search and data mining, New York, New York, USA.
792 C. Lipizzi et al. / Computers in Human Behavior 64 (2016) 782e792

Uzzi, B. (1997). Social structure and competition in interim networks. Administrative the ICUIMC '11: Proceedings of the 5th International Conference on Ubiquitous
Science Quarterly, 1e34. Information Management and Communication, New York, New York, USA.
Vespignani, A. (2005). Complex networks: Behind enemy lines. Nature Physics, 1(3), Williams, D. L., Crittenden, V. L., Keo, T., & McCarty, P. (2012). The use of social
135e136. media: An exploratory study of usage among digital natives. Journal of Public
Vie
!gas, F. B., Golder, S., & Donath, J. (2006). Visualizing email content: Portraying Affairs, 12(2), 127e136.
relationships from conversational histories. In Proceedings of the SIGCHI con- Yu, A. Y., Tian, S. W., Vogel, D., & Kwok, R. C.-W. (2010). Can learning be virtually
ference on Human Factors in computing systems (pp. 979e988). boosted? an investigation of online social networking impacts. Computers &
Wakamiya, S., Lee, R., & Sumiya, K. (2011). Towards better TV viewing rates: Education, 55(4), 1494e1503.
exploiting crowd's media life logs over Twitter for TV rating (p. 1). Presented at