Addis Ababa University

College of Natural Sciences

Attention-based Amharic-to-Wolaita Neural Machine Translation

Workineh Wogaso Gaga

A Thesis Submitted to the Department of Computer Science in

Partial Fulfillment for the Degree of Master of Science in
Computer Science

October 2020
Addis Ababa, Ethiopia
 Addis Ababa University

College of Natural Sciences

Workineh Wogaso Gaga

Advisor: Yaregal Assabie (PhD)

This is to certify that the thesis prepared by Workineh Wogaso, titled: Attention-based Amharic-
to-Wolaita Neural Machine Translation and submitted in partial fulfilment of the requirements for
the Degree of Master of Science in Computer Science complies with the regulations of the
University and meets the accepted standards with respect to originality and quality.

Signed by the Examining Committee:

Name Signature Date

Advisor: Yaregal Assabie (PhD)________ __________________ _________________

Examiner: _Solomon Atinafu (PhD)_______ __________________ _________________

Examiner: _Mulugeta Libsie (PhD)_______ __________________ _________________

 Abstract

Natural language (NL) is one of the fundamental aspects of human behaviors and we can
communicate all over the world through it. Natural Language Processing (NLP) is a branch of
artificial intelligence that deals with the interaction between computers and humans using the NL.
In the NLP world, every NL should be well understood by the machine. Machine Translation (MT)
is a process by which computer software understand NL and an automatic translation of a text
from one NL (source language) to another (target language). Neural Machine Translation (NMT)
is a recently proposed approach to MT and has been able to gain State of The Art (SOTA)
translation quality for the last few years. Unlike the traditional MT approaches, NMT aims at
building a single neural network that can be jointly tuned to maximize translation performance. In
this thesis we proposed Attention-based Amharic-to-Wolaita NMT. We built our system based on
Encoder-Decoder architecture also called Sequence to Sequence (Seq2Seq) model by integrating
a Recurrent Neural Network (RNN) and Gated Recurrent Unit (GRU). For comparison of our
translation performance, we developed non-attention-based Amharic-to-Wolaita NMT. An
encoder in basic (non-attention-based) Encoder-Decoder architecture encodes the complete
information of the source (Amharic) sequence into a single real-valued vector called context vector
which is passed to the Decoder to produce an output (Wolaita) sequence. Here, a context vector
summarizes the entire input sequence into a single vector. As the length of the sentence increases,
the inter-dependency of words is loosely related and it is a major drawback. The second problem
of basic Encoder-Decoder model is handling of a large number of vocabulary sizes. As each word
in the sentence is visited, it must be assigned a new identity number. But when the length of the
corpus increases, the number used for word representation and dimension of word vector needed
becomes higher. These two basic issues are solved using an attention mechanism. However, either
attention-based or non-attention based NMT have not been developed for Amharic-to-Wolaita
language pair. Thus, we developed attention-based NMT for Amharic-to-Wolaita translation and
compare it against a non-attention-based NMT system. We used a BLEU score evaluation
technique for system evaluation and we got 0.5960 BLEU score for non-attention-based system
and 0.6258 BLEU score for attention-based system. Thus, attention-based system obtains up to
+0.02978 BLEU improvements over non-attention-based NMT system.
Keywords: Natural Language Processing, Machine Translation, Neural Machine Translation,
Recurrent Neural Network, Language Modelling, Attention-based Encoder-Decoder Model.

i|Page
 Dedication
It was dedicated to

1. My lovely family. Especially to My lovely Mother Workinesh Dana. Mom, I love you whole
my remaining life and I promise to live a life that will do justice to all the sacrifices you have
made to me.
2. An innocent people who lost their life in different Ethiopian regions.

ii | P a g e
 Acknowledgements

First and foremost, I would like to thank the Almighty God and St. Mary for giving me strength,
determination, endurance and wisdom to accomplish this study and for being always with me in
whole my journeys I passed.

Next, I would like to extend my deepest appreciation to my advisor, Dr. Yaregal Assabie, for his
excellent and enduring support. He has given me so much inspiration for my future academic
career. I taught more from him how to build up my academic brand, when to say no to distractors,
what to give to others, and many, many more great things. Without him, it will be uncertain if I
can make it all the way through my MSc.

I also wish to express my sincere gratitude to all whom through their supports contributed to the
successful completion of this work. First, I am highly indebted to Tewodros Abebe, currently PhD
Student of Addis Ababa University, for his expertise, generous time, and patience in helping me
to complete this thesis especially by informing some previous work of Wolaita language.

“Yoho ubbaa ba wodiyan puullayidi oottis; xoossay benippe doommidi alamiya

wurssettay gakkanaw oottido oosuwa asa na’ay pilggidi gakkenna mala meri
merinatettaa garssan immis”.

“ነገርን ሁሉ በጊዜው ውብ አድርጎ ሠራው፤ እግዚአብሔርም ከጥንት ጀምሮ እስከ ፍጻሜ ድረስ የሠራውን
ሥራ ሰው መርምሮ እንዳያገኝ ዘላለምነትን በልቡ ሰጠው”።

መጽሐፈ መክብብ 3:11.

iii | P a g e
 Table of Contents

Chapter 1: Introduction ..............................................................................................1

1.1 Background ...................................................................................................................... 1
1.2 Motivation ........................................................................................................................ 3
1.3 Statement of the Problem ................................................................................................. 4
1.4 Objectives ......................................................................................................................... 5
1.5 Methods ............................................................................................................................ 5
1.6 Scope and Limitations ...................................................................................................... 6
1.7 Application of Results ...................................................................................................... 6
1.8 Organization of the Rest of the Thesis ............................................................................. 6
Chapter 2: Literature Review .....................................................................................7
2.1 Introduction ...................................................................................................................... 7
2.2 Overview of Amharic Language ...................................................................................... 7
2.2.1 Amharic Morphology................................................................................................ 7
2.2.2 Amharic Phrases ..................................................................................................... 13
2.2.3 Amharic Sentences.................................................................................................. 15
2.2 Overview of Wolaita language ....................................................................................... 17
2.2.4 Wolaita Morphology ............................................................................................... 18
2.2.5 Wolaita Phrases ....................................................................................................... 25
2.2.6 Wolaita Sentences ................................................................................................... 26
2.3 Machine Translation ....................................................................................................... 27
2.4.1 History of Machine Translation .............................................................................. 27
2.4.2 Approaches to Machine Translation ....................................................................... 29
2.5 System Modelling and Language Modelling ................................................................. 35
2.5.1 System Modelling ................................................................................................... 35
2.5.2 Language Modelling ............................................................................................... 37
Chapter 3: Related Work .........................................................................................41
3.1 Introduction .................................................................................................................... 41
3.2 Machine Translation for non-Ethiopian Language Pairs ............................................... 41
3.3 Machine Translation involving Ethiopian languages ..................................................... 45
3.4 Summary ........................................................................................................................ 49
iv | P a g e
Chapter 4: Design of the Proposed System .............................................................50
4.1 Introduction .................................................................................................................... 50
4.2 System design................................................................................................................. 50
4.3 System Architecture ....................................................................................................... 54
4.4 Text Preprocessing ......................................................................................................... 56
4.5 Stemming ....................................................................................................................... 58
4.6 One-hot Representation .................................................................................................. 59
4.7 Word Embedding ........................................................................................................... 60
4.8 Padding ........................................................................................................................... 62
4.9 Encoding......................................................................................................................... 62
4.10 Decoding ........................................................................................................................ 64
4.11 Attention Mechanism ..................................................................................................... 66
Chapter 5: Experimentation and Evaluation ............................................................68
5.1 Introduction .................................................................................................................... 68
5.2 Data Collection and Preparation .................................................................................... 68
5.3 System environment/tools used for development .......................................................... 68
5.4 Parameter Optimization and Training the Experimental Systems ................................. 69
5.5 BLEU Evaluation Metrics .............................................................................................. 72
5.6 Experimental Results...................................................................................................... 72
5.7 Discussion on the Result of the Study............................................................................ 73
Chapter 6: Conclusion and Future work ..................................................................75
6.1 Conclusion...................................................................................................................... 75
6.2 Future Work ................................................................................................................... 75
References ................................................................................................................................. 77
Appendix I: Sample of parallel corpus ...................................................................................... 86
Appendix II: Common Amharic stop words. ............................................................................ 88
Appendix III: Each epoch’s loss level and time taken for training the system. ........................ 89
Appendix IV: The last epoch results with loss level and a time to taken for 116 batches. ....... 90
Appendix V: Sample output ...................................................................................................... 92

v|Page
 List of Tables

Table 2.1: Amharic Noun Plural Formation ................................................................................... 8

Table 2.2: Amharic Nouns Marked Definiteness ........................................................................... 9
Table 2.3: Amharic Nouns Marked for Gender .............................................................................. 9
Table 2.4: Amharic Adjectives Inflection ..................................................................................... 10
Table 2.5: Noun Formation in Wolaita Language ........................................................................ 19
Table 2.6: Wolaita Nouns Formation from Stems and Suffixes ................................................... 20
Table 2.7: The Third and the Fourth class of Wolaita Nouns ....................................................... 20
Table 2.8: Wolaita Nouns Case Marker ........................................................................................ 22
Table 2.9: Adjectives in Wolaita .................................................................................................. 22
Table 2.10: Wolaita Noun Derivation ........................................................................................... 24
Table 4.1: Stemmed Amharic Words............................................................................................ 58
Table 4.2: Word Representation in Indexing ................................................................................ 59
Table 4.3: Semantic Relationship of Words Representation in Word-embedding ....................... 61

vi | P a g e
 List of Figures

Figure 2.1: Placement of Affixes in Amharic Verbs .................................................................... 12

Figure 2.2: The Placement of Affixes in Amharic Nouns ............................................................ 12
Figure 2.3: Architecture of RBMT Approaches ........................................................................... 30
Figure 2.4: Major Tasks in Direct Machine Translation Approach .............................................. 31
Figure 2.5: System Modelling....................................................................................................... 36
Figure 2.6: Structure of Recurrent Neural Network ..................................................................... 38
Figure 2.7: Example of a RLM that Processes an Input Sentence ................................................ 39
Figure 2.8: Comparison of LSTM Vs GRU Structure .................................................................. 40
Figure 4.1: Language Model Testing in the Beam Search Graph ................................................. 60
Figure 4.2: System Architecture ................................................................................................... 60
Figure 4.3: Skip Gram Model ........................................................................................................ 60
Figure 4.4 Word-embedding Example in Matrices....................................................................... 61
Figure 4.5: Encoder Architecture .................................................................................................. 63
Figure 4.6: Decoder Architecture ................................................................................................. 65
Figure 4.7: Basic Encoder-Decoder Model without Attention .................................................... 66
Figure 4.8: Attention-based Encoder-Decoder Architecture ........................................................ 67
Figure 5.1: Loss Level Vs Number of Epochs .............................................................................. 70
Figure 5.2: Loss Level for each Batch Size and Embedding Dimension for Number of Epochs 71
Figure 5.3: Loss Level Vs Learning Rate for Training ................................................................. 71
Figure 5.4: Time Taken Vs Learning Rate for Training ............................................................... 72

vii | P a g e
 Acronyms

AI Artificial Intelligence
ANN Artificial Neural Network
BLEU Bi-Lingual Evaluation Understudy
BPTT Backpropagate Through Time
CNN Convolutional Neural Network
DL Deep Learning
DNN Deep Neural Network
EBMT Example Based Machine Translation
FDRE Federal Democratic Republic of Ethiopia
GLU Gated Linear Unit
GPU Graphical Processing Unit
GRU Gated Recurrent Unit
LM Language Modeling
LSTM Long Short Term Memory
MT Machine Translation
NL Natural Language
NLM Neural Language Modeling
NLP Natural Language Processing
NMT Neural Machine Translation
RBMT Rule Based Machine Translation
RLM Recurrent Language Model
RNN Recurrent Neural Network
SMT Statistical Machine Translation
SNNPRS Southern Nations Nationalities and Peoples Regional State
SOTA Start of The Art
SRLIM SR Luthra Institute of Management
WMT Conference on Machine Translation

viii | P a g e
 Chapter 1: Introduction
1.1 Background
Natural language (NL) is one of the fundamental aspects of human behaviours and is a crucial
component in our lives [1]. It is the method of communication in different ways which are by audio
(spoken), text (written) and signs to exchange ideas, emotions, and information [2]. Thus, we can
communicate all over the world through NL. The advancement of technology and the rise of the
Internet as a means of communication led to an ever-increasing demand for Natural Language
Processing [3]. Natural Language Processing (NLP), also called computational linguistics is
widely regarded as a promising and critically important endeavour in the field of computer research
[4]. It came into existence to ease the user’s work [4]. The general goal for most computational
linguists is to let the computer have the ability to understand and generate NL so that eventually
people can address their computers through text and speech as though they were addressing
another person [4]. NLP applications are useful in facilitating human-human, human-computer,
computer-human, and computer-computer communication via computing systems. In the NLP
world, every language should be well understood by the machine. The process that lets the machine
understand the different languages used all around the world is called machine translation [2].

Machine Translation (MT) is a process by which computer software is used to translate a text from
one NL to another [2]. It refers to an automatic translation of one language into one (bi-lingual) or
more languages (multi-lingual) through electronic devices that contain a dictionary along with the
programs needed to make logical choices as required for the new language [2]. MT is considered
to be the most substantial way in which machines could communicate to humans and vice versa
[4]. To process any translation, human or automated, the meaning of a text in the source language
must be fully restored in the target language. While on the surface this seems straightforward, it is
far more complex [33]. Translation is not a mere word-for-word substitution. A translator must
interpret and analyze all of the elements in the text and know how each word may influence another
[33]. This requires extensive expertise in grammar, syntax, semantics, etc., in the source and target
languages, as well as familiarity with each local region in which syntax and semantic mean
sentence structure and meanings respectively. The translation of NLs by machine was first dreamt
of in the 17th century [5].

1|Page
The advantages of MT over human translator is when time is a crucial factor; MT can make
translations quickly [34]. We don't have to spend hours poring over dictionaries to translate the
words and the machine translator can translate the content quickly. The next benefit of MT over
human translator is, MT is comparatively cheap. This is because if we use the professional
translator, s/he will charge us on a per-page basis which is going to be extremely costly.
Confidentiality is another matter which makes MT favourable. Giving sensitive data to a
professional translator might be risky. Universality is also another advantage of a machine
translator. A machine translator usually translates a text which is in any language while a
professional translator specializes in one particular field. Online translation and translation of web
page content is also a favourable advantage of a machine translator. Online translation services are
at hand and we can translate information quickly with this service. Furthermore, we can translate
any web page content and query of a search engine by the use of MT systems [1, 8].

A number of approaches are available for MT, such as Rule-Based Machine Translation (RBMT),
Corpus-based Machine Translation Approach (Corpus-based MT), Hybrid Machine Translation
(Hybrid MT) and Neural Machine Translation (NMT) [6]. NMT is a recently proposed framework
for MT based purely on neural networks. Neural networks are making in-roads into the MT
industry, providing major advances in translation quality over the existing industry standard
Statistical Machine Translation (SMT) technology. Because of how the technology functions,
neural networks better capture the context of full sentences [7]. This technique has begun to show
promising results when compared to other approaches [9, 10, 11, 12]. Unlike the traditional phrase-
based translation system, NMT being an end-to-end trained model attempts to build and train a
single large neural network [13]. The goal of neural network is to design a fully trainable model
of which every component is tuned based on training corpora to maximize its translation
performance [8]. People have been turning their heads towards NMT systems, which after being
introduced seriously in 2014 have seen many refinements.

There are three big wins of NMT. The first one is end-to-end training, where all parameters are
simultaneously optimized to minimize a loss function on a network’s output. The second is a
distributed representation of parameters. It is a way of mapping or encoding information to a neural
network. The last one is better exploitation of the word and phrase similarities [13]. Attention-
mechanism improves its performance when the length of words in a sentence becomes longer [8].

2|Page
1.2 Motivation
Wolaita language belongs to the Omotic language family among Ethiopian language families (such
as Semitic, Cushitic, Nilotic and Omotic). It is spoken by Wolaita people and some other parts of
the Southern Nations, Nationalities, and People's Region of Ethiopia such as Gamo, Gofa, Mello,
Kucha, and Dawro [14]. Wolaita language uses the Latin script for writing [14]. It is the working
language in the Wolaita Zone. Currently, primary, secondary and higher education institutions, as
well as different mass media, are using the Wolaita language in teaching and learning, and
information transformation processes [14].

Amharic is one of the languages in the Semitic family which is widely spoken in Ethiopia [15]. It
is the first largest spoken language in Ethiopia, the second most-spoken Semitic language in the
world (after Arabic), and one of the five largest languages on the African continent [25]. Amharic
is an official working language of the Federal Democratic Republic of Ethiopia (FDRE) [15]. Since
Southern Nations Nationalities and Peoples Regional State (SNNPRS) is a collection of different
Nation Nationalities having different languages and cultures, Amharic is also an official working
language in the region [15]. Thus, different types of official documents, newspapers, and vacancies
are written and produced in Amharic both at the federal and regional levels [15].

Speakers of Wolaita language who are unable to speak and understand Amharic cannot
communicate and interact with Amharic speakers in an easy way without finding translators. Thus,
non-Amharic speakers of Wolaita language speakers face problems of lack of information. The
constitution of FDRE [15] recommends it is better if every regional official document is translated
and documented in Amharic language parallel with the local language. Once Nelson Mandela
talked at public speech as “If you talk to a man in a language he understands, that goes to his head.
If you talk to him in his own language that goes to his heart” [16]. Thus, it is a good contribution
if there is a way to translate federal written Amharic documents and literal news into Wolaita
language which can prevent non-Amharic speakers of Wolaita peoples from lack of information
and face to unwanted expenses, such as time and cost. Amharic-to-Wolaita Machine Translation
can solve the aforementioned problems. This has motivated us to work on Attention-based
Amharic-to-Wolaita Neural MT.

3|Page
1.3 Statement of the Problem
Several studies and applications have been done for foreign languages MT using different
approaches. Most of the works have been done on language pair of English and other languages,
such as for Arabic-to-English Neural Machine Translation [17] and English-Japanese Machine
Translation [18], French to English Statistical Machine Translation system [19], etc. This is
because the English language is the dominant language spoken across the world [20]. However,
only a little work has been done on the MT system among English and Ethiopian languages. Some
of the studies are carried out on English-Amharic language pair [21, 22], English-Afaan Oromo
language pair [6, 23]. Some of the MTs done between Ethiopian languages are Amharic-to-
Tigrigna MT using a hybrid approach [24], bi-directional Ge’ez-Amharic MT [25] and so on.

Translations that have been made so far from Amharic-to-Wolaita and English-to-Wolaita are
religious books using human translator [26]. People use human translation and they tend to be
slower as compared to machines [27]. Sometimes it can be hard to get a precise translation that
reveals what the text is about without everything being translated word-to-word. Translation
software allows to translate entire text documents within seconds. Human translation takes much
longer, especially if specific meanings have to be looked up in a dictionary. Thus, MT helps to
save time [28]. This and the other issues are when MT comes in, which solves most of the problems
caused by a human translator.

Since Wolaita language is used as a means of communication in different government and non-
government institutions and serving as working language at zone, it benefits non-Amharic speakers
of Wolaita people if documents and news articles are automatically translated into Wolaita
language. In recent studies, NMT provides promising results than other MT approaches [9, 10, 11,
12]. Currently, the progress in technology is shifting from traditional MT approaches to deep
learning-based NMT approach. Despite the recent success of NMT in standard benchmarks, it
faces a problem when the length of words in a sentence becomes longer. This issue is solved when
using an attention mechanism with NMT. As the researcher's knowledge is concerned, there is no
prior study conducted on the development of Amharic-to-Wolaita Neural MT system. We found
that NMT is also a very important NLP task that has to be done for the Wolaita language [30].
Thus, we propose attention-based Amharic-to-Wolaita NMT.

4|Page
1.4 Objectives
General Objective
The general objective of this study is to design and develop an attention-based Amharic-to-Wolaita
Neural Machine Translation.
Specific Objectives
The specific objectives are:
1. Review related systems and literature.
2. Develop parallel bi-lingual corpus for Amharic and Wolaita languages.
3. Identify the linguistic behaviors of Amharic and Wolaita languages.
4. Design a general architecture for attention-based Amharic-to-Wolaita NMT.
5. Develop a prototype.
6. Test and evaluate the performance of the system.

1.5 Methods
Literature Review
For the purpose of finding up-to-date methodologies in the MT domain, a thorough literature
review will be conducted. For this study, secondary data sources, like books, articles, publications
and other resources related to the topic will be reviewed. This helps to have a better understanding
of the subject of the study. Studies related to this study will be compiled so as to know the pros
and cons of various NMT techniques. MT systems in different languages will be studied with
respect to the closeness and difference among the languages. The details of the approaches and
algorithms followed to build the translation system will also be reviewed. The linguistic behavior
of Amharic and Wolaita languages will also be investigated and identified.
Data Collection
To conduct NMT, a parallel corpus of source and target language is required. The translation
system we are going to develop tries to generate translations using the Amharic-Wolaita corpus
based on neural network methods. The sources for both languages are mostly from a religious
book, and other relevant materials.
Prototype Development
In order to develop a prototype for NMT, some approaches and techniques are needed. Word
alignment, reordering and language modeling can be performed with the help of a well-trained
5|Page
deep neural network. Word2vec generates the word-vectors that are used by recurrent auto-encoder
in reconstruction task. RNN has the capability to implement reordering rules on sentences.
Evaluation Mechanism
NMT system can be evaluated either using by human (manual) or automatic evaluation methods.
Manual evaluation is time-consuming and expensive to perform, BLEU score will be used to
evaluate the performance of the system, which is an automatic evaluation technique.

1.6 Scope and Limitations

Attention-based Amharic-to-Wolaita NMT is designed to perform translation of texts written in
Amharic to Wolaita.

Because of the unavailability of the standardized corpus (corpus ready for MT research purpose)
for the dataset, the dataset we are going to use for training and testing is mostly from religious
documents. Thus, words which are not in corpus is not translated in our system.

1.7 Application of Results

The results of this research work have many applications. The system can be used for translation
from Amharic-to-Wolaita texts, and the translation system can be used as a tool for the teaching-
learning process of the languages. This study can be used to simplify the barrier of language
difficulty among language users. It enables to access information and interaction easily and fills
the communication gap of speakers; Moreover, the study can be used as a component for other
NLP applications such as speech translation. Since the future of NMT focuses on a multi-task
learning, larger context, and mobile devices [1,30], this study may be used as input for the next
researches.

1.8 Organization of the Rest of the Thesis

This thesis is organized into six chapters. Chapter Two presents a literature review which includes
an overview of the languages and MT approaches especially NMT. Chapter Three presents
different related works in the MT domain. Chapter Four presents the design of attention-based
Amharic-to-Wolaita NMT system. The experiments and results are discussed in Chapter Five and
Chapter Six presents conclusions, contribution and recommendations.

6|Page
 Chapter 2: Literature Review
2.1 Introduction
In this Chapter, a brief overview of Amharic and Wolaita languages, and MT in general and NMT
in detail are discussed. Additionally, the advancement of NMT over other MT approaches:
Statistical Machine Translation (SMT), Rule-Based Machine Translation (RBMT), Example-
Based Machine Translation (EBMT), and Hybrid Machine Translation (HMT) are described.

2.2 Overview of Amharic Language

Amharic is one of the Semitic languages spoken in Ethiopia. Next to Arabic, it is the second most-
spoken Semitic language in the world and it is the official working language of the Federal
Democratic Republic of Ethiopia. It is also the first largest spoken language in Ethiopia and
possibly one of the five largest languages on the African continent. The Amharic alphabet is called
Fidel, which grew out of the Ge’ez abugida-called in Ethiopian Semitic language. The usual word
order of Amharic is Subject-Object-Verb (SOV) [25]. Modern written Amharic uses a unique
script called hohiyat (ሆህያት) which is conveniently written in a tabular format of seven columns
[35]. The first order is the basic form; the other orders are derived from it by more or less regular
modifications indicating the different vowels [36]. The alphabet is written from left to right in
contrast to some other Semitic languages such as Arabic and Hebrew. It consists of 34 consonants,
giving 7*34=238 syllable patterns, or fidels [37]. In addition to the 238 characters, there are other
non-standard alphabets which contain special features usually representing labialization. Each
alphabet represents a consonant together with its vowel. The vowels are fused to the consonant
form in the form of diacritic markings. The diacritic markings are strokes attached to the base
characters to change their order [38].
2.2.1 Amharic Morphology
Amharic morpheme can be free or bound, where a free morpheme can stand as a word on its own
whereas a bound morpheme cannot [39]. An Amharic root is a sequence of consonants and is the
basis for the derivation of verbs. On the other hand, a stem is a consonant or consonant-vowel
sequence which can be free or bound where a free stem can stand as a word on its own whereas a
bound stem has a bound morpheme affixed to it. An example of morpheme could be, the word
በልጅነት can be morphologically scrutinized into three separate morphemes: the prefix በ-, the root
7|Page
ልጅ, and the suffix -ነት. A word, which can be as simple as a single morpheme or can contain
several of them, is formed from a collection of phonemes or sounds [40, 41]. In Amharic, words
can be formed from morphemes in two ways. These are by inflection and by derivation.

A. Inflection

Inflectional Morphology is a morphology that deals with the combination of a word with a
morpheme, usually resulting in a word of the same class as the original stem, and serving the same
syntactic function. Inflection can be achieved by marking a word category for gender, number,
case, definiteness, aspect and politeness [24, 25]. Since the Amharic language is highly
inflectional, a given root of a language word can be found in different forms [24]. The author in
[42] states that Amharic word classes belong to five classes. These are noun (ስም), verb (ግስ),

adjective (ቅጽል), preposition (መስተዋደድ) and adverb (ተውሳከ ግስ). From these classes, highly
inflected parts in Amharic are discussed as follows.

Nouns (ስም): Noun is a name that represents a person, places, animal, thing, feeling and idea [25].
Amharic nouns are marked for gender, case, number, and definiteness, and results in an infected
word with affixes to the noun. It can be achieved by marking a word category for number as
singular and plural by affixation of morphemes (and vowel changes) or repetition of words as
shown in Table 2.1 [43, 44].
Table 2.1: Amharic Noun Plural Formation
Noun in Singular Form Morpheme Plural Form
ልብ -ኦች ልቦች
ባርያ -ዎች ባርያዎች
ጻድቅ -ን/-ኦች ጻድቃን/ጻድቃኖች
ገዳም -ት/ኦች ገዳማት/ገዳሞች
ዘበነ እነ- እነዘበነ
ቅጠል Plural formation by repetition ቅጠል-ኣ-ቅጠል[ቅጠላቅጠል]
Amharic nouns are marked for a word category of definiteness and it can be achieved by the
affixation of morphemes or vowels based on a number, gender, and/or ending of the noun as shown
in Table 2.2 [43, 44].

8|Page
 Table 2.2: Amharic Nouns Marked Definiteness

Indefinite Noun Number Gender Definite noun

ልጅ Singular Feminine ልጅ-ዋ[ልጅዋ]
Masculine ልጅ-ኡ [ልጁ]
Plural ልጆች-ኡ [ልጆቹ]
በግ Singular Feminine በግ-ዋ[በግዋ] በግ-ኢቱ [በጊቱ]
Masculine በግ-ኡ [በጉ]
Plural በጎች-ኡ[በጎቹ]
Amharic nouns are marked for a word category of gender as shown in Table 2.3 [43, 44] and it
can be achieved by the affixation of the morphemes -ኢት. For example ልጅ-ኢት[ልጂት] በግ-

ኢት[በጊት] አሮጌ-ኢት[አሮጊት]. Amharic nouns are marked for a word category of case and it can be
in both objective and possessive case. Objective case can be achieved by affixation of morpheme
-ን. For example, አህያ-ን [አህያን]. Possessive case can be achieved by the affixation of morphemes
or vowels based on person, number, gender, and/or ending of the noun (personal pronouns by
prefixing የ. የ-ለማ➔ [የለማ] ).

Table 2.3: Amharic Nouns Marked for Gender

Subjective case Person Number Gender Possessive case

ልጅ First Singular ልጅ-ኤ➔ልጄ
Plural ልጅ-ኣችን ➔ ልጃችን
Second Singular Masculine ልጅ-ህ➔ልጅህ
Feminine ልጅ-ሽ➔ልጅሽ
Plural ልጅ-ኣችሁ➔ልጃችሁ
Third Singular Masculine ልጅ-ኡ➔ልጁ
Feminine ልጅ-ዋ➔ልጅዋ
Plural ልጅ-ኣቸው➔ልጃቸው
በግ First Singular በግ-ኤ➔በጌ
Plural በግ-ኣችን ➔ በጋችን
Second Singular Masculine በግ-ህ➔በግህ
Feminine በግ-ሽ➔በግሽ
Plural በግ-ኣችሁ➔በጋችሁ
Third Singular Masculine በግ-ኡ➔በጉ
Feminine በግ-ዋ➔በግዋ
Plural በግ-ኣቸው➔በጋቸው

9|Page
Verb (ግስ) ፦ Generally, Amharic verbs are derived from roots and use a combination of prefixes
and suffixes to indicate the person, number, voice (active/passive), tense and gender [45]. A verb
possesses two behaviors that make it different from other word classes in Amharic. The first one
is, it is placed at the end of an Amharic sentence, and the second one is it has a suffix attached to
it indicating the subject of the sentence [42]. For example: - ኢትዮጵያ የመጀመሪያ ሳታላይቷን አ-
መጠቀ-ች /Ethiopia launched the first satellite/ the underlined word is the verb of the sentence
with the prefix አ- and the suffix -ች. እሷ ምሳዋን በላ-ች /she ate her lunch/ the underlined word is

a verb and the suffix “ች‟ indicates the subject of the sentence is she (እሷ) which is feminine gender.
Amharic verbs, in general, show a high degree of inflection since person, case, gender, number,
tense, aspect, mood and others are marked on the verb. For example, “አልሰበረንም" indicates: the

subject እሱ (third person, masculine, singular), the object እኛን (first person, plural), negation

አል.......ም, past tense -ሰበረ.

Adjective (ቅጽል): - An adjective is a word that describes, identifies, or further defines noun or
pronoun. Nouns tell about things nature, but adjectives tell about things behavior or characteristics,
like shape, size, colour, type, property. Adjectives of Amharic language are marked for gender and
number and results in an inflected word with affixes [43, 44]. In Amharic, some of the morphemes
that are used to inflect given adjectives are: -ት, -ኦ,-ኢት and ኦች as shown in Table 2.4 [43, 44].

Table 2.4: Amharic Adjectives Inflection

Singular form Plural form Prefix Suffix

ድንበር ድንበሮች ...ኦች
ሰማያዊ ሰማያዊያን ..ያን
ማን እነማን እነ...
እናት እናቶች ..ኦች

10 | P a g e
B. Derivation

Derivational morphology is a morphology concerned with the way in which words are derived
from morphemes through processes such as affixation or compounding.

Nouns: - Amharic nouns can be derived from verbal roots by infixing vowels between consonants
like ጥ-ቅ- ም➔ ጥቅም. It can be derived from adjectives by suffixing bound morphemes ደግ

(adjective)-ነት (morpheme)➔ ደግነት (derived noun). It can be derived from stems by prefixing or

suffixing bound morphemes. For example, ጠቀም-ኤታ➔ ጠቀሜታ. It can be derived from Stem-

like verbs by suffixing the bound morpheme -ታ. For example, ደስ-ታ➔ ደስታ. It can be derived

from nouns by suffixing bound morphemes. For example, ኃይልኧ-ኛ➔ ኃይለኛ. Compound Words

(sometimes by affixing the vowels ኧ and ኦ). For example, Noun + [ኧ] + Noun.

ቤት+[ኧ]+መንግስት➔ ቤተ-መንግስት.

Verbs: - Verbs are words which indicate action and they take place at the end of clause positions.
Amharic verbs take subject markers as a suffix like -ሁ /-hu/ for subject ‘I’, -ህ /-h/ for subject

‘You’, ች /-c/ for subject ‘She’ and so on, to agree with the subject of the sentence [46]. Amharic

verbs can be derived from Verbal Roots by affixing the vowel -ኧ- to produce CኧC1C1ኧ C-, e.g.

ስ-ብ-ር ➔ስኧብብኧር- [ሰበር] and repeating penultimate consonants and affixing the vowels -ኧ-

and -ኣ- to produce CኧC1ኣC1C1ኧC-, e.g, ፍ-ል-ግ➔ፍኧልኣልልኧግ [ፈላለግ]. Verbal Stems by

affixing morphemes አ-, ተ-, አስ e.g, መጠቅ-(Verbal Stem) + አ-(morpheme)➔አመጠቅ.

Adjectives: Amharic adjectives modify nouns or pronouns by describing, identifying, or

quantifying words [46]. Adjectives always come before nouns or pronouns which they modify.
But all the words that come before nouns cannot always be adjectives. For example: - ይህ ቤት “this

house” in this example ይህ “this” precedes the noun ቤት “house” but this doesn’t mean ይህ “this”
is an adjective, it is a pronoun. Amharic adjective can be derived from verbal roots by infixing
vowels between consonants [43]. For example, ድ-ር-ቅ by applying the vowel, ድኧርኧቅ it gives

the adjective ደረቅ. Nouns by suffixing bound morphemes (ኧኛ፣ ኣማ፣ኣም፣ አዊ). Stems by suffixing

bound morphemes (ኣ፣ ኡ፣ ኢታ ).

11 | P a g e
C. Affixation

Affix is a morpheme fastened to a stem or base form of a word, and modifies its meaning or creates
a new word [40]. In Amharic affixes can be prefix, suffix, and infix. Prefix is a morpheme added
at the beginning of a word whereas suffixes are added at the end to form derivatives. Infixes are
inserted in the body of a word causing a change in meaning, which can be easily observed in an
iterative and reciprocal aspect of a root word in the Amharic language [40, 42]. Amharic verbs can
have up to four prefixes and up to four suffixes as shown in Figure 2.1 [43, 44].

Prefix Suffix
prep/conj rel neg sbj ROOT sbj obj/def neg/aux/acc conj
Figure 2.1: Placement of Affixes in Amharic Verbs
Here the first, second, third, and fourth options of prefix represent preposition or conjunction,
relative, negation, and subject in terms of number, gender, person and definiteness respectively.
Relative verbs are marked using “የ” /ye-, “የሚ”/yemi-, “እሚ”/Imi- and negation is marked with

prefixes like “አይ”/ay-, “አል”/al-, etc. [22]. Similarly, the first and second options of suffix
represent subject and object, in terms of gender, number, person, and definiteness, respectively.
The third option represents negation or auxiliary or accusation, where negation can be marked with
“-ም”, auxiliary is usually marked with morpheme “አለ” , and accusative is marked with morpheme

“ን”/-n. The fourth option represents conjunction like “ም”, “-ስ” etc. [22, 47]. Amharic nouns have
up to two prefixes and up to four suffixes. Similarly, the prefix and suffix slots have two and four
sub-slots respectively. Figure 2.2 shows the placement of affixes in Amharic nouns [22, 47].

Prefix Suffix
prep/gen distrib STEM plur poss/def acc conj

Figure 2.2: The Placement of Affixes in Amharic Nouns

The prep/gen option of the prefix represents preposition or genitive. Genitive is marked using
morphemes -ye-/”የ-“. In the second option of prefix, distributive (distrib) is marked using

-Iy_e-/”እየ-“morpheme. In the case of suffix, option one represents the number of information.
Option two represents possessive or definiteness information. The third and fourth options
represent accusative and conjunction, respectively [47].
12 | P a g e
2.2.2 Amharic Phrases
Phrases are syntactic structures that consist of one or more than one word but lack the subject
predicate organization of a clause. These phrases are composed of either only head word or other
words or phrases with the head combination. The other words or phrases that are combined with
the head in phrase construction can be specifiers, modifiers and complements [46]. In Amharic,
phrases are categorized into five, namely noun phrase (NP), verb phrase (VP), adjectival phrase
(AdjP), adverbial phrase (AdvP) and prepositional phrase (PP) [43, 44].

A. Noun Phrases

A noun phrase (NP) is a phrase that has a noun as its head. In this phrase construction, the head
of the phrase is always found at the end of the phrase. This type of phrase can be made from a
single noun or combination of a noun with either other word classes including noun word class or
phrases. That means one noun can be a noun phrase. See the following example: አንበሳው ሁለት

ላሞችን ገደለ “the lion killed two cows” in this sentence, there are two parts: the subject አንበሳው”

the loin “and the object with the verb ሁለት ላሞችን ገደለ” killed two cows”. Thus, the first part (the
subject) is a noun phrase and the second one is a verb phrase. Therefore, the noun phrase in the
above example is only the noun አንበሳው” the lion” [46].

A Noun Phrase can be simple or complex [24]. The simple NP construction consists of a single
noun or pronoun, for instance, በግ “sheep”, መኪና “car”, እሱ “he”, እሷ “she” are simple NP and
do not consist subordinate clauses. A complex NP can consist of a noun with other constituents
(specifiers, modifiers and complements) but the phrase must contain at least one sentence [46].

Example: -In the ኢትዮጵያ ያመጠቀችው የመጀመርያው ሳታላይት (the satellite that Ethiopia

launched for the first time) here ኢትዮጵያ ያመጠቀችው (that Ethiopia launched) is a sentence which

is a modifier, whereas የመጀመርያው (the first) is the single word that is a complement.

B. Verb Phrases

A verb phrase (VP) is composed of a verb as a head, which is found at the end of the phrase, and
other constituents such as complements modifiers and specifiers [43, 44]. ካሳ [ወደ ቤተ ክርስቲያን]

ሄደ “Kassa went to church” [ወደ ቤተ ክርስቲያን] is a prepositional phrase modifying the verb ሄደ

13 | P a g e
‘went’ from place point of view. In general, the structural rule of verb phrases can be formulated
as VP => PP V|V|AdjP V|NP VP|NP PPVP|AdvP PP VP.

C. Adjectival Phrases

Amharic Adverbial phrases (AdvP) are made up of one adverb as head word and one or more
other lexical categories including adverbs itself as modifiers. The head of the AdvP is also found
at the end. Unlike other phrases, AdvPs do not take complements. Most of the time, the modifiers
of AdvPs are PPs that come always before adverbs [43, 44, 46].

Example: - ካሳ [እንደ አባቱ ክፉኛ] ታመመ (Kassa is severely sick like his father). Here the phrase in

the bracket is adverbial phrase and the head word is ክፉኛ (severely). The modifier that is found in

the AdvP is እንደ አባቱ (like his father) which is comparative PP. Generally, the structural rule for
Amharic languages can be formulated as follows: AdjP=>Adj|Spec Adv Adj|PP Adj|NP Adj.

D. Prepositional Phrase

Amharic prepositional phrase (PP) is made up of a preposition (Prep) head and other constituent
such as nouns, noun phrases, prepositional phrase, etc [6, 43, 44]. Unlike other phrase
constructions, prepositions cannot be taken as a phrase. Instead they should be combined with
other constituents and the constituents may come either previous to or subsequent to the
preposition which is the head of the phrase [46]. In a prepositional phrase, if the complements are
nouns or NPs, the position of prepositions is in front of the complements. For example, እንደ ትልቅ

ልጅ” like a big child” እንደ “like” is a preposition which is combined with the noun “ልጅ”/ child

and it came in front of complement” ትልቅ”/big. whereas if the complements are PPs, the position

will shift to the end of the phrase. For example, ከወንዙ አጠገብ” Next to the river”, አጠገብ is PP

which is combined with the noun “ወንዙ”/ river and it came at the end. In general, the structural
rule for Amharic can be written as: PP => PP PP |PP NP|PP NN |PPV|N PP

E. Adverbial Phrases

Amharic Adverbial phrases (AdvP) are made up of one adverb as head word and one or more
other lexical categories including adverbs itself as modifiers [4, 43, 44]. The head of the AdvP is

14 | P a g e
also found at the end. Unlike other phrases, AdvPs do not take complements. Most of the time, the
modifiers of AdvPs are PPs that come always before adverbs [46].

Example: - ካሳ [እንደ አባቱ ክፉኛ] ታመመ (Kassa is severely sick like his father). Here the phrase in

the bracket is adverbial phrase and the head word is ክፉኛ (severely). The modifier that is found in

the AdvP is እንደ አባቱ (like his father) which is comparative PP. The general structural rule for an
adverbial phrase is: AdvP => Adv|Adv Adv
2.2.3 Amharic Sentences
An Amharic sentence is formed from a noun phrase (NP) and a verb phrase (VP). Regarding their
order in a given sentence, NP comes first then VP follows [35]. The sentence structure for the
Amharic language is a Subject-Object-Verb (SOV), unlike English with a subject-verb-object
combination [4, 43, 44]. We can take the following as an Example, አበበ ምሳውን በላ/Abebe ate

his lunch, in this example, the sentences are composed of አበበ: subject, ምሳውን፡Object and በላ:
Verb which is different from English. Amharic sentences are constructed from simple or complex
NP and simple or complex VP but NP always comes first as a subject [46].

Example: - አብዛኛው ፖለቲካ ፓርቲዎች “many political parties”. ወደ ኢትዮጵያ ገብተዋል “went to

Ethiopia”. አብዛኛው ፖለቲካ ፓርቲዎች ወደ ኢትዮጵያ ገብተዋል “many political parties went to
Ethiopia”. The first two constructions do not express a full idea but the last one does. Because the
last one expresses full information such as who did go to Ethiopia? Where the political parties
went? etc. All these questions have been answered by the last word construction. In the last
construction there are NP and VP which build the sentence and these are NP አብዛኛው ፖለቲካ

ፓርቲዎች “many political parties” and VP ወደ ኢትዮጵያ ገብተዋል “went to Ethiopia”. The
remaining phrases (other than NP and VP) are being constructed in NPs or VPs that are found in
a sentence. Based on this construction, sentences can be simple or complex.

A. Simple Sentences

Simple sentences are sentences, which contain only one verb. A simple sentence can be
constructed from NP followed by VP which only contain single verb [4, 43, 44].

15 | P a g e
Example 1: -አስቴር ብርጭቆውን ሰበረችው “Aster broke the glass”. Here the sentence contains only

one verb ሰበረችው “she broke”. This sentence contains transitive verb ሰበረችው “broke” that takes

only one object ብርጭቆውን “the glass”.

Example 2: - አስቴር ለካሳ መጽሀፍ ሰጠችው “Aster gave Kassa a book”. Here also the sentence

contains only one verb ሰጠችው “gave” so it is a simple sentence. The difference in this example

from the previous one is the sentence here contains transitive verb ሰጠችው “gave” with two objects

ለካሳ “Kassa” and መጽሀፍ “book”. Generally, all the above stated examples are simple sentences
that contain different types of verbs. Simple sentences may contain intransitive verbs, transitive
verbs with one object and transitive verbs with two objects. As explained in [4, 43, 44] simple
sentences are classified into four namely: declarative sentences, interrogative sentences, negative
sentences and imperative sentences.

Declarative sentences are used to convey ideas and feelings that the speaker has about things,
happenings, etc., that could be physical, mental, real or imaginary. In Amharic, declarative
sentences always end with the Amharic punctuation mark “፡፡” which is equivalent of period (.) in

English. Example: - አስቴር ትምህርት ቤት ውስጥ ነች።/ “Aster is at school.” Here the sentence is
declarative because it describes where Aster is.

Interrogative sentence is a sentence that questions about the subject, the complement, or the
action that the verb specifies. Example: - አስቴር ምን ውስጥ ነች? In the above example, the question
is for the unknown thing just to get full information about it. This type of question or interrogative
sentences consist interrogative pronouns which are ማን “who”, መቼ “when”, ምን “what”, ስንት

“how many”, የት “where”, etc.

Negative sentences simply negate a declarative statement made about something. Example: - ካሳ

በግ አልገዛም “Kassa didn’t buy a sheep” In this example, the sentence is a negative declarative

sentence. The verb አልገዛም “did not buy” is negated by the prefix አል-“not”.

Imperative sentences convey instructions and mostly their subject is a second person pronoun
that is usually implied by the suffix on the verb. But when the command is passed for the third

16 | P a g e
person, the subject of the sentence can be third person pronoun or noun. Example: - ወጥ አምጪ

(bring wat) Here the subject is (you) second person feminine singular. Example: - ካሳ ልብስ ይጠብ
(Kassa, wash clothes) Here the command is for the third person that does not exist at the time the
order is transferred. So, the subject is (he) third person singular masculine.

B. Complex Sentences

Complex sentences are sentences that contain at least one complex NP or complex VP or both
complex NP and complex VP [4, 43, 44]. Complex NPs are phrases that contain at least one
embedded sentence in the phrase construction. The embedded sentence can be complemented.
Example: - [ኢትዮጵያ ያመጠቀችው የመጀመሪያው ሳተላይት] በጣም ዉጤታማ ነች “the first satellite

that Ethiopia launched is very effective”. Here the head of the noun phrase [ኢትዮጵያ ያመጠቀችው

የመጀመሪያው ሳተላይት “that” is ሳተላይት] “the first satellite that Ethiopia launched”. The head

with the complement የመጀመሪያው “the first” form simple noun phrase የመጀመሪያው ሳተላይት
“the first satellite” and this noun phrase has been combined with the embedded sentence or clause
ኢትዮጵያ ያመጠቀችው “that Ethiopia launched” to form a complex noun phrase.

2.2 Overview of Wolaita language

Wolaita language belongs to the Omotic language family among Ethiopian language families (such
as Semitic, Cushitic, Nilotic and Omotic). It is one of the main languages of the Ometo group of
the Omotic (named as ‘west Cushitic’ in [26]) family, which belongs to the Afro-Asiatic language
phylum. It is spoken by Wolaita people and some other parts of the Southern Nations, Nationalities,
and People's Region of Ethiopia such as Gamo, Gofa, Mello, Kucha, and Dawro [14, 50]. It is
because the people of Wolaytta are surrounded in the west and in the south by populations (such
as the Dawro, the K’uc’a, the Borodda and the Gamu) who speak another Ometo dialect and thus
an idiom very similar to Wolaytta [26]. The native people call the language “Wolaytta”
(wolaittáttuwa in local language). The language is also referred to as woláítta dóónaa (ወላይታ ኣፍ)

or woláítta Káálaa (ወላይታ ቃል). Wolaita is the working language of the Wolaita Zone. Currently,
primary, secondary and higher education institutions, as well as different mass media, are using
the Wolaita language in teaching and learning, and information transformation processes [14].

17 | P a g e
Wolaita language uses the Latin script for writing called WPXW which literally mean “The custom
in writing the Wolaytta letters”. [26, 34, 50]. WPXW was published in 1985 E.C by Wolaytta
Qaala Hiwote Maatteme Keettaa (WQHMK) [26, 34, 50]. WPXW’s writing system is as
convenient as Qubee, which is a system based on the Latin alphabet for writing the Afan-Oromo
language [34, 50]. According to [26], Wolaita was first written in the late 1940s by a team of
missionaries led by Dr. Bruce Adams. They translated the Bible into Wolaita in 2002. The
language is used in social, political and economic activity. At the primary school level, the
language is used as a medium of instruction and taught as a subject in secondary and high school.
Currently, the language is offered as a subject at Wolaita Soddo University and Arbaminch TVET
College. But it was planned for the future to be thought in all universities of Ethiopia. The
language was written as Wolaytta, Wolayta, Wolaita, Woleyta, Ometo by different writers in
different time [50]. But in our case, we use Wolaita for our study.
2.2.4 Wolaita Morphology
Morphology is a branch of linguistics that studies and describes how words are formed in language
[4, 43, 44]. Like Amharic, there are two categories of morphemes in Wolaita: free and bound
morphemes. A free morpheme can stand as a word on its own whereas bound morpheme does not
occur as a word on its own. Affixation and compounding are two basic word-formation processes
in Wolaita [14, 49]. Affixation is a process by which affixes are added/attached in some manner
to the root, which serves as a base. Affixes are morphemes that cannot occur independently. Prefix,
suffix, and infix are the three types of affixes. Wolaita language does not have prefix and infix.
Instead, Suffixation is the basic way of word formation in Wolaita. In forming a word, adding one
suffix to another is common in Wolaita. This process of adding one suffix to another suffix can
result in relatively long word, which often contains an amount of semantic information equivalent
to a whole Amharic phrase, clause or sentence [14, 50]. The second word-formation process in
Wolaita is compounding. According to [14], compounding is the joining together of two linguistic
forms, which function independently. Although Wolaita is very rich in compounds, compound
morphemes are rare in Wolaita and their formation process is irregular [49]. As a result, itis
difficult to determine the stem of compounds from which the words are made. As discussed in [14,
49], there are two kinds of morphology in Wolaita language: inflectional and derivational.

18 | P a g e
A. Inflectional Morphology in Wolaita

Inflectional morphology is concerned with the inflectional changes in words where word stems are
combined with grammatical markers for things like a person, gender, number, tense, case and
mode [14, 50]. It does not result in changes of parts of speech. Like Amharic language, Wolaita is
highly inflectional, a given root of a language word can be found in different forms [14, 49]. Highly
inflected word classes in Wolaita are discussed as follows.

Nouns:
Based on the syllable formulation of Lamberti and Sottile in [49] Wolaita nouns have the form
C1V1C2V2, where C and V stand for consonant and vowel respectively. However, C1 represents
a glottal stop while V1 seldom represents diphthong (diphthongs are sounds formed by the
combination of two vowels in a single syllable). In the same way, C2 may represent simple or
germinated constant and V2 either represents ending in the absolute case or ending required by the
syntactic function associated with noun stems. Most Wolaita words are bi-radical (the number of
consonants in the word). However, there are some pluri-radical words. See Table 2.5 [48, 49] for
Wolaita noun formation.

Table 2.5: Noun Formation in Wolaita Language

Character sequence Example Character sequence Example

V Aguntta/እሾክ/ thorn CVCV Kayisoi/ሌባ / Theaf
VV Aawaa/አባት/father CCVCCV Shappa/ወንዝ /river
VC Intarssa/ምላስ/tongue CVVCCVV Keettaa/ቤት/house
VCC oydda/አራት/four CVCCVV Mattaa/ንብ/bee
CVV Maataa/ሳር/grass CVC Zokkuwa/ጀርባ/back
As Demewoz Beldados discussed in [50], Wolaita nouns are formed from stems and suffixes in
which the stems never change while suffixes exhibit changes. Table 2.6 [48, 49] is an illustration
of his discussion.

19 | P a g e
 Table 2.6: Wolaita Nouns Formation from Stems and Suffixes

Nouns Stems Suffixes

keett-a/ቤት/house Keett- -a
keett-I/ቤት/house Keett- -i
keett-eta/ቤት/house Keett- -eta
na?-eti/ልጅ/ a boy na?- -eti

According to the ending they take in inflection Wolaita noun is classified into four major classes
[49]. First class nouns are nouns ending in –a in absolute case with the stress on the last syllable.
The second class nouns are nouns that have an absolute case ending in –iya and the stress is on the
penultimate. The third and the fourth class nouns are nouns ending in uwa and –iyu, respectively,
in the absolute case as shown in Table 2.7 [48, 49].

Table 2.7: The Third and the Fourth class of Wolaita Nouns

Noun Class Ending Examples

First -a Awa/ጸሃይ/sunrise፣ xuma/ጨለማ/darkness፣ ?aawa/አባት/father
Second -iya Bitanniya/ወንድ ያገባ/married man፣ayfiya/ዓይን/eye፣kusshiya/እጅ/hand
Third -uwa Kawuwa/መንግስት/government፣ metuwa/ችግር/'trouble፣
Fourth -iyu Kaniyu/ሴት ዉሻ/dog feminine፣bollotiyu/ሴት አማች/'mother-in-law

Wolaita nouns are marked for gender, number, and case [14, 49].

Gender: As indicated in [49] Wolaita, like other languages, exhibits two types of genders, i.e.,
masculine and feminine. According to the author, nouns that belong to the fourth class are feminine
and the rest (first, second and third class) belong to masculine. Feminine and masculine differ from
each other by their endings. While feminine ends in –iyo, masculine ends in –an in absolute case.

Example: Dorss-a /ወንድ በግ, masculine vs. dorss-iyo/በግ, feminine. Deeshsh-a /ወንድ ፍየል, masculine
vs. deessh-iyo ፍየል, feminine. Addiy-a/አዉራ ዶሮ, masculine indd-iyo/ ዶሮ, feminine.

Number: According to [49], Wolaita noun contains singular and plural. The plural noun is formed
by using suffixes and singular contains the basic form of the word. Example: Dorss-a በግ, singular
vs. dorssa-ta/በጎች, plural. Deeshsh-a /ወንድ ፍየል, singular vs. deessha-ta ፍየሎች, plural. Addiy-a/አዉራ
ዶሮ, singular adde-ta/ ዶሮዎች, plural.

20 | P a g e
Those of the 2nd class, instead, exhibit the ending –e-ta.
Noun (Singular) Noun (plural)
har-iya (donkey) ➔ har-e-ta 'donkeys'
?org-iya (he-goat) ➔?org-e-ta 'he-goats'
laagg-iya (friend) ➔ laagg-e-ta. 'friends

Those of the 3rd class are characterized by the ending –o-ta

Noun (Singular) Noun (plural)
Word-uwa (lie) ➔ word-o-ta 'lie pl.')
The nouns of the 4th class which consist of terms for female beings assume the plural form of heir
masculine counterpart.
Noun (Singular) Noun (plural)
?imatt-iyu 'female guest'➔ ?imatt-a-ta 'female guests'
boogaaanc-iyu 'female robber' ➔ boogaanc-a-ta 'female robbers'
laagg-iyu 'female friend' ➔ laagg-e-ta 'female friends'
If the feminine noun does not have any masculine counterpart, then it normally agrees with the
nouns of 2nd class and ends in –e-ta.
Noun (Singular) Noun (plural)
macc-iyu (wife) ➔ macc-e-ta 'wives'
misshir-iyu (married woman ) ➔ misshir-e-ta 'married women'

Case: The noun inflection takes place by the suffixation of case endings to the noun stem or to the
absolutive case form [50]. Accordingly, the absolutive case is characterized, as we have already
seen above, by the ending –a (1st class and plural), -iya (2nd class), -uwa (3rd class) and –(i)yu
(4th class) respectively, while the subject case ends in –y (first three classes), -i (plural), and –(i)ya
(4th class). The genitive is represented either by the noun stem alone or is more often characterized
by the lengthening of the final vowel of the absolutive form. The object case of the noun inflection
agrees with the respective absolute case [14, 49, 50]. Table 2.8 [48, 49] details Wolaita nouns case
marker.

21 | P a g e
 Table 2.8: Wolaita Nouns Case Marker

Case marker (morphemes) Function Example

-ssi, -w or –yoo dative and benefactive case Garssassissi/ለውስጠኛው
-kko or -mati directive case Garssakko/ወደ ዉስጥ
-ni locative Garsani/በዉስጥ በኩል
-ppe ablative Garsaappe/ከዉስጥ
-ra comitative case Garssara /ከታችኛው ጋር

Adjectives

According to [49], adjectives in Wolaita language are used to qualify nouns and they come before
the noun they qualify. They are also defined as words that modify nouns by expressing their
qualities, colour, size, etc. Adjectives in Wolaita language end in -a, -iy or –uw as shown in Table
2.9 [48, 49].
Table 2.9: Adjectives in Wolaita
Adjectives ending in -a: Adjectives ending in -iya: Adjectives ending in -uwa:

geessh-a /ንጹህ mal-iya /ጣፋጭ Lo7-uwa /ጥሩ

cinc-a /ብልጥ haankett-iya /ልዩ yuush-uwa /ዙሪያ
qant-a /አጭር iiti-iya /መጥፎ Im-uwa /መስጠት
Adjectives in Wolaita language precede the noun they modify and remain unchanged when
used in attributive position because they do not have to agree in Wolaita with their governing noun
either in gender or in number or in case [26, 49]. But most adjectives ending in –uwa and a few in
–iya are replaced by the endings -o and -e respectively.
Example: Lo7-uwa /ጥሩ➔_lo7-o asa /ጥሩ ሰው

haah-uwa /ሩቅ➔_ haah-o sohuwa /ሩቅ ቦታ

Lamberti and Sottile [49] also reported that when adjectives are used in predicative position, -uwa
will be changed to -o. For example: Lo7-uwa /ጥሩነት➔_ lo7-o/ጥሩ: he bitane lo7-o ➔ያ ሰውየው

ጥሩ ነው. But every word that comes before a noun is not necessarily an adjective. For example, in

the sentence Ha dorsai taga “ይህ በግ የኔ ነው “here the word Ha ‘ይህ’ has the role of an adjective
but it is demonstrative determiner [26].

22 | P a g e
Verbs
As pointed in [26] verbs of Wolaita, like most Ethiopian languages, are very complex. Wolaita
verbs usually have a consonant-vowel-consonant sequence. For example, uya-ጠጣ, gela/ ግባ.

Some of Wolaita verbs are borrowed from the Amharic language. Eamples: azzaza-እዘዝ, nabbaba-

ንባብ, kassas- ክሰስ.

Verbs in Wolaita language exhibit a very complex inflection system depending on mood, tense,
kind of action and aspect. For example, if we take verb 7imm-a (ስጥ), it has the following
inflection for past tense:
imm-a:si ➔ሰጥቸዋለሁ/ እኔ ሰጥቸዋለሁ, imm-adasa ➔ሰጥተሃል, imm-a:su ➔ሰጥተሻል

imm-i:si ➔ሰጥቱዋል, imm-ida ➔ሰጥተናል, imm-ideta ➔ሰጥታቹዋል, imm-idosona

➔ሰጥተዋሉ….etc.
We can produce many inflected verbs for future and present tense in the same way as we do for
past tense. For example imm-a:si➔ ሰጥቸዋለሁ/ እኔ ሰጥቸዋለሁ(past tense)

imm-a-is➔እየሰጠሁኝ ነው(present tense)

imm-a-na➔እሰጠዋለሁ(future tense)
Just like that of Amharic language, Wolaita verbs are found at the end of the sentence and suffix
bound morphemes which help to indicate the subject of the sentence in the sentences shown below.
Na7-ya mayuwa shama-su /ልጅቱዋ ልብስ ገዛች

neeni ne kawuwa ma-dasa /እራትክን በልተሃል

tani 7osuwa wursa-si /ስራዬን ጨርሸዋለሁ

In the above three sentences, shama-su, ma-dasa, wursa-si are verbs. The bound morphemes
{-su},{-dasa},{-si} show 3rd person, 2nd person and 1st person pronouns that are used as subjects
of the sentence. Verbs in Wolaita language change their shape for a person, gender, number and
time by attaching suffixes [14, 49, 50].

B. Derivational Morphology in Wolaita

Nouns: Wolaita nouns are formed by the class suffixes -a and -uwa can refer either to abstract
terms or to very concrete objects and they also serve to express action nouns as shown in Table
2.10 [14, 49, 50].
23 | P a g e
 Table 2.10: Wolaita Noun Derivation

Noun Suffix Derived noun

hassay- 'speak -a Hassaya/conversation
harg-‘sick’ -ya harg-iya ‘sickness’
gulba-‘knee’ -ta gulba-ta ‘knee’
wurse- 'end -tta wurse-tta 'end
eeyya- 'stupidity' -tetta eeyya-tetta 'stupidity'
kaawo-tetta 'kingdom' -tetta kaawo-tetta 'kingdom'

Verbs: Like other Cushitic and Semitic languages of Ethiopia, Wolaytta makes use of some
morphemes in order to derive a further stem from a root or a stem. This derivation procedure is
concretely applied in Wolaytta by suffixing one or more morphemes to verbal stem [14, 49, 50].
They further indicated that it is possible to form three different kinds of secondarily derived verbal
stems: iterative (or intensives), causatives and passives (or reflexives) [49, 50]. Iteratives and
intensives stems are expressed in Wolaytta by means of the same morpheme –erett-, which is
regularly suffixed to its verbal stem. Example: -ment- 'to break' ➔ ment-erett- 'to break many
times or in many pieces'. shissh- 'to collect' ➔ shissh-erett- 'to collect many times or in many
things'. The Wolaytta language possesses only productive causative morpheme, i.e –is-s.

gel- 'enter'➔ gel-iss- 'let someone enter/put into’.

Verbs which have their primary stem ending in -y- or -y-y- form their causative by replacing all
existence of -y- by -sh-.
Example: uy-y- 'drink' ➔ ush-sh- 'let someone drink.
yuuy-y- 'turn, intr'➔ yuush-sh- 'turn,tr

Adjectives: Wolaita adjectives can be derived from verbal roots by suffixing the morphemes like
–ta. For example, imma+ -ta ➔imota. It can be also derived from nouns, stems by suffixing bound
morphemes, and compound words.

C. Affixation

According to [49] there are three types of affixes (prefix, infix and suffix), Wolaita uses suffixation
to form words and prefixes and infixes are not used as word formation in Wolaita language. We
already discussed inflectional and derivational affixation of Wolaita language in the previous
sections.
24 | P a g e
2.2.5 Wolaita Phrases
In Wolaita, phrases are categorized into five categories, namely noun phrase (NP), verb phrase
(VP), adjectival phrase (AdjP), adverbial phrase (AdvP) and prepositional phrase (PP) [26, 49].

A. Noun Phrases

A noun phrase (NP) is a phrase that has a noun as its head. In this phrase construction, the head
of the phrase is always found at the end of the phrase. This type of phrase can be made from a
single noun or combination of a noun with either other word classes including noun word class or
phrases. A NP can be simple or complex [49]. The simple NP construction consists of a single
noun or pronoun. A complex NP can consist of a noun with other constituents (specifiers, modifiers
and complements) but the phrase must contain at least one sentence [49]. Example: - Itiyoophiya
yeddido koyro satalaytiya(the satellite that Ethiopia launched for first time) here Itiyoophiya
yeddido (that Ethiopia launched) is a sentence which is a modifier, whereas koyro satalaytiya (the
first satellite) is the single word that is a complement.

B. Verb Phrases

A verb phrase (VP) is composed of a verb as a head, which is found at the end of the phrase, and
other constituents such as complements modifiers and specifiers [4, 26, 49]. ካሳ [ወደ ቤተ

ክርስቲያን] ሄደ “kaasi woosa keettaa biisi” [ወደ ቤተ ክርስቲያን] is prepositional phrase modifying

the verb ሄደ ‘biisi from place point of view. In general, the structural rule of Wolaita verb phrases
can be formulated as VP => PP V|V|AdjP V|NP VP|NP PPVP|AdvP PP VP.

C. Adjectival Phrases

Wolaita Adverbial phrases (AdvP) are made up of one adverb as head word and one or more
other lexical categories including adverbs itself as modifiers. The head of the AdvP also found at
the end. Unlike other phrases, AdvPs do not take complements. Most of the time, the modifiers
of AdvPs are PPs that comes always before adverbs [26, 46, 49]. Example: - kaasi [ba aawagadan]
keehippe harggiis/ ካሳ [እንደ አባቱ ክፉኛ] ታመመ. Here the phrase in the bracket is adverbial phrase

and the head word is ክፉኛ (keehippe). The modifier that is found in the AdvP is እንደ አባቱ (ba
aawaadan) which is comparative PP. Generally, the structural rule for Wolaita languages can be
formulated as follows: AdjP=>Adj|Spec Adv Adj|PP Adj|NP Adj.
25 | P a g e
D. Prepositional Phrase

Wolaita prepositional phrase (PP) is made up of a preposition head and other constituents
such as nouns, noun phrases, prepositional phrase, etc[4, 26, 49]. In PP, if the complements are
nouns or NPs, the position of prepositions are in front of the complements. In general, the structural
rule for Amharic can be written as: PP => PP PP |PP NP|PP NN |PPV|N PP

E. Adverbial Phrases

Wolaita Adverbial phrases (AdvP) are made up of one adverb as head word and one or more
other lexical categories including adverbs itself as modifiers [4, 26, 49]. The head of the AdvP also
found at the end. Unlike other phrases, AdvPs do not take complements. Most of the time, the
modifiers of AdvPs are PPs that comes always before adverbs [4, 26, 49]. Example: - ካሳ [እንደ

አባቱ ክፉኛ] ታመመ (kaasi ba awwagadan keehippe harggiis). Here the phrase in the bracket is

adverbial phrase and the head word is ክፉኛ (keehippe). The modifier that is found in the AdvP is

እንደ አባቱ (ba awwagadan) which is comparative PP.

The general structural rule for adverbial phrase is: AdvP => Adv|Adv Adv
2.2.6 Wolaita Sentences
Like Amharic, Wolaita language has also two kinds of sentences: simple and complex sentences.

A. Simple Sentences

Simple sentences a contain only one verb. A simple sentence can be constructed from NP followed
by VP which only contain single verb [4, 43, 44]. For example: - Abebe kuwaasiya kaa7iis/Abebe
played football. Here the sentence contains only one verb kaa7iis /played.

B. Complex Sentences

Complex sentences are sentences that contain at least one complex NP or complex VP or both
complex NP and complex VP [4, 43, 44]. Complex NPs are phrases that contain at least one
embedded sentence in the phrase construction. The embedded sentence can be complements. See
the following examples. For example: - Abebe kuwaasiya kaa7i simmidi so oosuwa oottis/ Abebe
did his homework after playing football. Here the sentence contains two verbs kaa7i/ played and
oottis/ did.

26 | P a g e
2.3 Machine Translation
A language is used for conveying information or broadcasting the information. Stepping into the
modern digital age, language as the information carrier has become the most significant means for
a human to communicate. But it has been considered as the barrier of communications between
people from different countries and between peoples who speak different languages within the
same country. The problem of converting a language into another quickly and efficiently has
become a problem of common concern for humanity [51]. Due to the advent of the computer and
the Internet the world is becoming together to one [52]. Thus, the knowledge, culture, tradition,
history, religious, philosophy documents of one country language can be translated into another
language and the rest of the world via MT. To create a paperless working environment, translation
plays a great role and to make accessible the document of one language in another language.
Sharing of knowledge is also possible besides facilitating easy communication.

MT is an automatic translation of one language into one or more languages (in case of multi-
lingual) by means of a computer. High-quality translation requires a thorough understanding of
the source text and its intended function as well as good knowledge of the target language.
Translation itself is a challenging task for humans and is no less challenging for computers because
it deals with natural languages [53]. This section provides background information regarding the
field of MT, its history in general and NMT in detail, and various MT approaches.
2.4.1 History of Machine Translation
Although there are some disputes about who first had the idea of translating automatically
between human languages, the actual development of MT system can be traced back to the late
forties after World War II [45]. In this time, MT was constrained by several factors: limitation of
hardware particularly, inadequacy of memories and slow access and unavailability of high-level
programming language. The linguistic study was not correlated with MT research. So, researchers
relied on the dictionary-based approach and the application of statistical methods [54].

Researchers of that time were faced with a lot of technical constraints and realized that there
could be no perfect high-quality translation, and suggested the involvement of humans in the
process. They also proposed the development of controlled languages and restriction of systems
to specific domains. Criteria concerning the success and failure of MT were set in its first 50 years

27 | P a g e
of research and development. These criteria are the conceptual, engineering, operational,
commercial and communicative criteria [55].

An American mathematician and scientist named Warren Weaver in 1947 had a belief that a
computer is capable of translating one NL to another by using logic, cryptography, frequencies of
letter combinations, and linguistic patterns [52]. He published a memorandum that outlines his
belief. In the 1950s a research program at Georgetown University teamed up with IBM to perform
research on MT. Later in 1954, they demonstrated a system that translates a few phrases from
Russian to English. The research resulted in wide acceptance and interest in the field [52].

In 1966 the US sponsors of MT research committee called Automatic Language Processing

Advisory Committee (ALPAC) published an influential report which concluded that MT was
slower, less accurate and twice as expensive as human translation. However, in the following
decade, MT research took place largely outside the United States, in Canada and in Western
Europe and work continued to some extent [45].

In the 1970s and 80s researchers shifted their focus to assisting MT rather than replacing human
translators. It has three main strands: first, the development of advanced transfer systems building
upon experience with earlier Interlingua systems; secondly, the development of new kinds of
Interlingua systems; and thirdly, the investigation of AI techniques and approaches. That resulted
in the development of translation memory and many computer assisted tools (CAT) for MT. At
the end of the 1980s, MT entered a period of innovation in methodology which has changed the
framework of research. In 1981 came the first translation software for the newly introduced
personal computers, and gradually MT came into more widespread use [45].

During the late 1980s, MT advanced rapidly on many fronts. The dominance of the rule-based
approach waned in the late 1980s with the emergence of new methods called `corpus-based'
approaches, which did not require any syntactic or semantic rules in text analysis or selection of
lexical equivalents. The major reason for this change has been a paradigm shift away from
linguistic/rule-based methods towards empirical/data-driven methods in MT. This has been made
possible by the availability of large amounts of training data and large computational resources.

In the 1990s, the use of MT and translation aids by large corporations has grown rapidly.
Particularly impressive increase is seen in the area of software localization (i.e., the adaptation and

28 | P a g e
translation of equipment and documentation for new markets). On the research front, the principal
areas of growth are seen in example-based and statistical machine translation approaches, in the
development of speech translation for specific domains, and in the integration of translation with
other language technologies.

In the 2000s, the research state was moved toward combining the rule-based and SMT paradigms.
It was an approach developed by taking advantage of both statistical and rule-based approaches.
After the late 2000s the state of MT shifted to NMT. Recently, it has gained popularity in the field
of MT. NMT is starting to displace its corpus-based predecessor, SMT. It has proven to have better
efficiency in the MT domain with the goal of improving the quality of the output as well as the
performance of the system. With the emergence of the Internet and cheap and powerful computers
accelerated the progress of MT. Nowadays researches are focused on improving the quality and
performance of MT systems [52].
2.4.2 Approaches to Machine Translation
A machine translation system first analyses the source language input and creates an internal
representation. This representation is manipulated and transferred to a form suitable for the
target language. Then at last output is generated in the target language [56]. Based on the degree
of dependence of internal representation on the source and target languages, MT can be classified
into three approaches [6]: Rule-Based Machine Translation (RBMT), Corpus-based Machine
Translation Approach (Corpus based MT), and Hybrid Machine Translation (Hybrid MT).

A. Rule-Based Machine Translation (RBMT)

RBMT is also known as Knowledge-Based Machine Translation or Classical Approach of MT. It

is a general term that denotes machine translation systems based on linguistic information about
the source and target languages basically retrieved from (bilingual) dictionaries and a collection
of rules called grammar rules covering the main semantic, morphological, and syntactic
regularities of each language respectively [25]. In this approach, human experts specify a set of
rules to describe the translation process, so that an enormous amount of input from human experts
is required [25]. It consists of a bilingual or multilingual lexicon, and software programs to process
the rules. The rules play a major role in various stages of translation such as syntactic processing,
semantic interpretation, and contextual processing of language [56].

29 | P a g e
The basic principle of RBMT methodologies is to apply a set of linguistic rules in three different
phases [57]: analysis, transfer and generation. The core process (transfer) is mediated by bilingual
dictionaries. RBMT systems parse the source text and produce an intermediate representation as
shown in Figure 2.3 [24]. Rules for transforming source language structures into target language
structures is from dictionaries and rules for deriving (intermediary representations) from which
output can be produced. The preceding stage (analysis) interprets input source language strings
into a suitable translation unit. The succeeding stage of synthesis (generation) derives target
language output text from the target language structures or representations generated by the
transfer process [2, 56].

Figure 2.3: Architecture of RBMT Approaches

Based on the intermediate representation used this approach is further classified into the following
approaches [56, 58]: Direct, Interlingua and Transfer-Based MT approaches.
i. Direct Machine Translation

Direct MT approach is historically the earliest and known as the first generation of MT systems
employed around the 1950s to 60s when a need for MT was mounting. [45]. MT systems that use
this approach can translate a source language directly to target language. Words of the source
language are translated without passing through an additional/intermediary representation as
shown in Figure 2.4 [24]. No complex architecture will be involved. It carries out word by word
30 | P a g e
translation with the help of a bilingual dictionary usually followed by some syntactic
rearrangement. Due to this direct mapping, such systems are highly dependent on both the source
and target languages [59]. It needs only a little syntactic and semantic analysis and it is basically
bilingual and uni-directional.

Figure 2.4: Major Tasks in Direct Machine Translation Approach

ii. Interlingua Approach

Interlingua approach to MT mainly aims at transforming the texts in the source language to a
common representation which is applicable to many languages. Using this representation, the
translation of the text to the target language is performed and it should be possible to translate to
every language from the same Interlingua representation with the right rules [34]. In this approach,
the source language is transformed into the Interlingua language representation that is independent
of any language. The target language is then generated out of the Interlingua [57]. In short, the
translation in this approach is a two-stage process, i.e., analysis and synthesis [51]. The first stage
is particular to the source language and doesn’t require any knowledge about the target language
whereas the second stage is particular to the target language and doesn’t require any knowledge of
the source language. The main advantage of interlingua approach is that it creates an economical
multilingual environment that requires 2n translation systems to translate among n languages
wherein the other case, the direct approach requires n(n-1) translation systems [34].

iii. Transfer-based Machine Translation

The core idea of both transfer-based and interlingua-based MT is the same: to make a translation
it is necessary to have an intermediate representation that captures the "meaning" of the original
sentence in order to generate the correct translation. The difference is in interlingua-based MT this
intermediate representation must be independent of source and target languages, whereas, in
31 | P a g e
transfer-based MT, it has some dependence on the language pair involved [34]. Thus, the transfer-
based approach is preferred over interlingua-based approach on the basis when there is known
structural differences between the source and target languages.

Transfer based system can be broken down into three different stages: analysis, transfer and
generation. In the first stage, the source language parser is used to produce the syntactic
representation of the source language sentence (Internal representation). In the next stage, the
result of the first stage is converted into equivalent target language representation (another internal
representation). Finally, a target language morphological analyzer is used to generate the target
language text [60]. Transfer based need rules for syntactic transfer, semantic transfer, and lexical
transfer [57,61]. Syntactic transfer rules will tell us how to modify the source parse tree to resemble
the target parse tree. Semantic transfer uses semantic role labelling. Lexical transfer rules are based
on a bilingual dictionary. The dictionary can be used to deal with lexical ambiguity.

B. Corpus-based Machine Translation Approach

As of 1990s, MT research moved from classical rule-based approach to empirical or corpus-based

systems. Empirical systems are data-driven as opposed to rule-driven. This approach uses a large
amount of raw data in the form of parallel corpora. This raw data contains texts, dictionaries,
grammars, etc. and their translations. These corpora are used for acquiring translation knowledge
[58]. In recent years there is an increased interest in corpus-based MT systems because it needs
less effort from the language/linguistic experts and less human effort is required [59]. In recent
classification, Corpus-based approach has three varieties, namely, Example-Based Machine
Translation (EBMT), Statistical Machine Translation (SMT), and Neural Machine Translation.
The following subsections briefly explain these approaches.

i. Example-Based Machine Translation (EBMT)

EBMT is a translation method that retrieves similar examples (pairs of source phrases, sentences,
or texts and their translations) from a database of examples adapting the examples to translate new
input [2]. The system maintains an example-base (EB) consisting of translation examples. When
a source language sentence is given to the system, the system retrieves a similar source language
sentence from the EB with its translation. Then it adapts the example to generate the target
language sentence for the input sentence. The basic premise is that, if a previously translated phrase

32 | P a g e
occurs again, the same translation is likely to be correct again. Thus, the EBMT system rests on
the idea that similar sentences will have similar translations. The system has two main modules 1)
retrieval and 2) adaption [59]. There are three tasks in EBMT: Matching fragments against existing
examples, transferring (identifying the corresponding translation fragments), and recombining the
fragments to give the target text [2].

ii. Statistical Machine Translation (SMT Approach)

SMT is a method for translating text from one natural language to another based on the knowledge
and statistical models extracted from bilingual corpora. A supervised or unsupervised statistical
machine learning algorithm is used to build statistical tables from the corpora. This process is
called learning or training. The statistical tables consist of statistical information such as the
characteristics of well-formed sentences and the correlation between the languages. During
translation, the collected statistical information is used to find the best translation for the input
sentences. This translation step is called the decoding process [60].

In SMT, the core process (transfer) includes a translation model which takes as input source
language words or word sequences (phrases) and produces target language words or word
sequences as an output. The following stage includes a language model which synthesizes the sets
of source language words in meaningful strings which are meant to be equivalent to the input
sentences. The preceding (analysis) phase is represented by the conventional process of matching
individual words or word sequences of input source language text against entries in the translation
model [51]. The translation accuracy of these systems mainly depends on the parallel corpus
regarding its domain, quantity and quality. So, in order to have a good translation quality, the data
must be preprocessed consistently [24].
iii. Neural Machine Translation (NMT) approach

NMT is a new breed of corpus-based MT (also called data-driven or, less often, corpus-driven
machine translation), which is beginning to displace its corpus-based predecessor, SMT [9, 10, 11,
12]. It is a newly emerging approach to MT, recently proposed by Kalchbrenner and Blunsom
[11], Sutskever et al. [9] and Cho et al. [12]. Unlike the traditional phrase-based translation system,
which consists of many small sub-components that are tuned separately, NMT attempts to build
and train a single, large neural network that reads a sentence and outputs a correct translation. It

33 | P a g e
is trained on huge corpora of pairs of source language segments (usually sentences) and their
translations, that is, basically from huge translation memories containing hundreds of thousands
or even millions of translation units.

Deep neural networks (DNN) have shown great success in handwriting recognition [62, 63],
speech recognition [64, 65] and in natural language process such as language modelling [66],
paraphrase detection [67] and word embedding extraction [68]. Furthermore, in the field of MT,
DNN is a newly emerging approach and proved to achieve excellent performance [9, 10, 11, 12].
During the last wave of neural network research in the 1980s and 1990s, MT was in the sight of
researchers exploring these methods [69]. In fact, the models proposed by Forcada and Ñeco [70]
and Castaño et al. [71] are striking similar to the current dominant NMT approaches. However,
none of these models trained on data sizes large enough to produce reasonable results for anything.
The computational complexity involved by far exceeded the computational resources of that era,
and hence the idea was abandoned for almost two decades. During this hibernation period, data-
driven approaches such as phrase-based SMT rose from obscurity to dominance and made MT a
useful tool for many applications, from information gisting to increasing the productivity of
professional translators.

The modern resurrection of neural methods in MT started with the integration of NLMs into
traditional SMT systems. The pioneering work by Schwenk [72] showed large improvements in
public evaluation campaigns. However, these ideas were only slowly adopted, mainly due to
computational concerns. The use of Graphical Processing Unit (GPU) for training also posed a
challenge for many research groups that simply lacked such hardware or the experience to exploit
it. Moving beyond the use in language models, neural network methods crept into other
components of traditional SMT, such as providing additional scores or extending translation tables
by Schwenk [73]; Lu et al. [74], reordering by Kanouchi et al. [75]; Li et al. [76] and pre-ordering
models by de Gispert et al. [77], and so on. For instance, the joint translation and language model
by Devlin et al. [78] was influential since it showed large quality improvements on top of a very
competitive SMT system.

More ambitious efforts aimed at pure NMT, abandoning existing statistical approaches completely.
Early steps were the use of convolutional models proposed by Kalchbrenner and Blunsom and
sequence-to-sequence models by Sutskever et al. [9] and Cho et al. [79]. These were able to
34 | P a g e
produce reasonable translations for short sentences but fell apart with increasing sentence length.
The addition of the attention mechanism finally yielded competitive results [8, 80]. With a few
more refinements, such as byte pair encoding and back-translation of target-side monolingual data,
NMT became the new state of the art. Within a year or two, the entire research field of machine
translation went neural. To give some indication of the speed of change: At the shared task for MT
organized by the Conference on Machine Translation (WMT), only one pure NMT system was
submitted in 2015. It was competitive but outperformed by traditional statistical systems. A year
later, in 2016, a NMT system won in almost all language pairs. Since 2017, almost all submissions
were NMT systems.

Hybrid Machine Translation Approach (Hybrid MT)

Hybrid machine translation approach is developed taking advantage of both statistical and rule
based translation methodologies. It has proven to have better efficiency in the area of MT systems
at the time of its discoveries [23]. The hybrid approach can be used in many ways. In some cases,
translations are performed in the first stage using a rule-based approach followed by adjusting or
correcting the output using statistical information. In the other way, rules are used to preprocess
the input data as well as post-process the statistical output of a statistical-based translation system.
This technique has become better example-based MT and has more power, flexibility, and control
in translation since the emergence of current-state of Machine Translation called NMT.

2.5 System Modelling and Language Modelling

2.5.1 System Modelling
Artificial Intelligence (AI) and machine learning (ML) are the cornerstones of the next revolution
in computing [95]. These technologies hinge on the ability to recognize patterns then, based on
data observed in the past, predict future outcomes. Deep learning (DL) can be considered as a
subset of ML [88]. It is a field that is based on learning and improving on its own by examining
computer algorithms. It can be used to solve any pattern recognition problem without human
intervention. State of the art surveys on the data-driven methods and ML algorithms indicate that
DL, along with ML methods is the future of data science. Sequence to Sequence (Seq2Seq) models
are DL models that have achieved a lot of success in tasks like machine translation, text
summarization, and image captioning [89]. A Seq2Seq model is a model that takes a sequence of
items (words, letters, features of images, etc.) and outputs another sequence of items [90]. Its
35 | P a g e
models vary in terms of the exact architectures to use. A natural choice for sequential data is the
Recurrent Neural Network (RNN), used by most of the recent NMT work and for both the Encoder
and Decoder. The use of RNN models, however, differ in terms of (a) directionality –
unidirectional or bidirectional; (b) depth – single or deep multi-layer; and (c) type – often either a
vanilla RNN, an LSTM, or a Gated Recurrent Unit (GRU). In general, for the Encoder, almost any
architecture can be used since we have fully observed the source sentence. For example, the
researchers in [91] used a convolutional neural network (CNN) for encoding the source. Choices
on the Decoder side are more limited since we need to be able to generate a translation.

The Encoder-Decoder architecture with RNNs has become an effective and standard approach for
both NMT and Seq2Seq prediction in general [9]. In case of MT, a model takes a sequence of
source language texts (Amharic texts) as input and produces a sequence of target language texts
(Wolaita) as output as shown in Figure 2.5 [91]. Seq2Seq was first proposed by Cho et al. [12] to
model variable-length source input with temporal dependencies which is a drawback of N-gram
model [10]. The work in [12] is one of the frontier studies investigating NMT with sequences [91].
Google translate started using such a model in production in late 2016 [27]. These models are also
explained in the two pioneering papers [9, 12]. The standard Seq2Seq model is generally unable
to accurately process long input sequences since only the last hidden state of the Encoder RNN is
used as the context vector for the Decoder [81, 92]. On the other hand, the Attention Mechanism
directly addresses this issue as it retains and utilizes all the hidden states of the input sequence
during the decoding process. It does this by creating a unique mapping between each time step of
the Decoder output to all the Encoder hidden states. This means that for each output that the
Decoder makes, it has access to the entire input sequence and can selectively pick out specific
elements from that sequence to produce the output [93].

Figure 2.5: System Modelling

36 | P a g e
2.5.2 Language Modelling
After a few years, different authors outperformed SMT by using deep neural network (DNN) based
approaches. Nowadays, rather than using tedious steps like preparation of language modelling,
preparation of translation modelling, tuning and decoding steps of SMT, using Encoder-Decoder
based language modelling became preferred approaches [8]. Language modelling (LM) is the task
of assigning a probability to sentences in a language [95]. In recent time, the Encoder-Decoder
end-to-end LM is becoming attractive LM for MT task. It is LM which is based on a deep learning
algorithm. According to the work of different researchers, in the early emergence of NMT, the
RNN was used to generate word co-occurrence probabilities for MT tasks. Besides assigning a
probability to each sequence of words, LM also assigns a probability for the likelihood of a given
word (or a sequence of words) to follow a sequence of words [96]. So that one can judge if a
sequence of words is more likely or “fluent” than another [97].

To model these conditional probabilities, traditional n-gram LMs can only handle short contexts
of about 4 to 6 words and do not generalize well to unseen n-grams [98, 99]. Thus, the neural
language model is first proposed by [65] to solve the aforementioned problem in n-gram.
Developing better language models often results in models that perform better on their intended
NLP task. This is the basis for developing better and more accurate language models. Often
training better language models improves the underlying metrics of the downstream task (such as
BLEU score for translation), which makes the task of training better LMs valuable by itself [102].
Therefore, Neural Language Models (NLMs) became the better choice for the simplicity of the
steps of modelling. In NLM, the steps are interconnected.

A. Neural Language Models

The use of neural networks (NNs) in LM is often called Neural Language Modelling or NLM for
short. NN approaches are achieving better results than classical methods both on standalone LMs
and when models are incorporated into larger models on challenging tasks like machine translation
[97]. NN based LM involves the building of an end-to-end NN based LM that NN is trained to
map aligned bilingual texts from source sentence to target sentences without additional external
linguistic information [8, 9]. After the first proposal of NLMs in [10], it was enhanced by other
researchers [104, 105, 106]. As a natural development, subsequent MT systems in [72, 106, 107],
started adopting NLMs alongside with traditional n-gram LMs and generally obtain sizable
37 | P a g e
improvements in terms of translation quality [85]. To make NLMs even more powerful, recent
works in [73, 108, 109] proposed to condition on source words as well as the target context to
lower uncertainty in predicting next words. More recently, RNN and then networks with a long-
term memory like the Long Short-Term Memory network (LSTM), Gated Recurrent Unit (GRU)
and bidirectional RNN (bi-RNN) allow the models to learn the relevant context over much longer
input sequences than the simpler feed-forward networks.

B. Recurrent Neural Network (RNN) Language Model

For any of NLP task, a basic layered NN is used when there are a set of distinct inputs and we
expect that all inputs are independent of each other [111]. This is a preferred approach for
modelling languages for translation purpose because language is a sequence of words and each
next word is dependent on the words that come before it. If we want to predict the next word in a
sentence, we should know what the previous words were. This is analogues to the fact that the
human brain does not start thinking from scratch for every word we say. The heart of the Seq2Seq
modelling is RNN [9] and its structure is expressed in Figure 2.6 [16]. it is a powerful and
expressive architecture that can handle sequential data and has been extensively used for LM [16].

Figure 2.6: Structure of Recurrent Neural Network

An example of the RLM is illustrated in Figure 2.7 [16]. It can be noticed that the cell-state output
(Cn) at a current time-step is used as input for next time-step.

38 | P a g e
 Figure 2.7: Example of a RLM that Processes an Input Sentence
Here, xₜ is input to the network at time step t. x₁ is the input at index 1 in the sequence. Sₜ is a
hidden state of the network at time t. It is the memory of the network. This corresponds to the
weights in a normal NN which is learned at the time of training. Here is the advantage of using
NLM as language modelling. RNN LMs can be trained on large amounts of data, and outperform
n-gram models [110]. Here is an interesting example “I had a good time in Wolaita. I also learned
to speak some _________”. If we want to predict what word will go in the blank, we have to go
all the way to the word Wolaita and then conclude that the most likely word will be Wolaiteygna.
In other words, we have to have some memory of our previous outputs and calculate new outputs
based on our past outputs [110]. Another advantage of using RNNs for LM is that due to memory
constraints, they are limited to remembering only a few steps back [110]. This is ideal because the
context of the word can be captured in the five to ten words before it. We don’t have to remember
fifty to a hundred words of context. RNN can also be used for arbitrarily long sequences if we have
memory at our disposal. But it has also some drawbacks. As the context length increases, i.e., our
dependencies become longer; layers in the unrolled RNN also increase. As the network becomes
deeper, the gradients flowing back in the back-propagation step become smaller. As a result, the
learning rate becomes slow and makes it infeasible to expect long term dependencies of the
language. We used the GRU unit both in Encoder and Decoder part as the solution to this issue.

39 | P a g e
C. Gated Recurrent Unit (GRU)

Despite the advantage of RNN, it suffers from two major drawbacks: Exploding gradient and
vanishing gradient problem. The exploding gradient refers to the phenomenon when gradient
values become exponentially large as we perform backpropagate through time (BPTT). The
exploding gradient problem is solved by clipping the gradient after it reaches a certain threshold
value. The vanishing gradient, on the other hand, is a challenging task and it occurs when gradient
values start approaching zero as we perform BPTT. There are certain ways to solve the vanishing
gradient problem, such as specific leaky generators [46], regularization [113], and GRU [27], etc.
The GRU by [27] is the most widely used solution for gradient vanishing problem. GRU networks
are just an advanced version of plain RNNs that we discussed above. These networks are designed
to remember information for long periods without having to deal with the vanishing gradient
problem. In the GRU cell, the vanishing gradient problem is solved by writing the current state as
a memory of the network. This writing process is regulated reset gates. If we want to remember
something, we write it down. Similarly, GRUs also has the same intuition. The difference between
GRU and LSTM architectures is that GRU has no output gate, while LSTM has an output gate as
shown Figure 2.8. Therefore, GRU is LSTM that has no output gate, so writes the contents from
its memory cell to the larger network at each time step [27, 114]. LSTM cells add a large number
of additional parameters. For each gate alone, multiple weight matrices are added. More
parameters lead to longer training times and risk over-fitting [115].

Figure 2.8: Comparison of LSTM Vs GRU Structure

40 | P a g e
 Chapter 3: Related Work
3.1 Introduction
There are different studies conducted on MT approaches, strategies, techniques, and
implementations that have been documented. In Ethiopia, some MT systems have been tried to be
developed and documented as research work. This chapter discusses previous works related to our
study in MT domain. We classified the chapter into three sections as MT for Ethiopian language
pairs, MT for a foreign language to Ethiopian language pair and MT for foreign language pairs
and we will discuss all in detail in the next sub sections.

3.2 Machine Translation for non-Ethiopian Language Pairs

Choudhary et al. [81] conducted research in Neural Machine Translation for English-Tamil. In
this research, the basic objective of the researchers was to localize all the information available in
the English language to the local language by using efficient MT approach. They used the datasets
obtained from EnTam V2.05 and Opus. The sentences are taken from various domains like news,
bible, cinema, movie subtitles and combined to build their final parallel dataset. Their final dataset
contains 183,451 training corpus, 1,000 validation and 2,000 test corpus from English to Tamil.
The data used is encoded in UTF-8 format. They used a BLEU score method to evaluate the
performance of the system. Their model outperformed Google translator with a margin of 4.58
BLEU points.

Utiyama et al. [31] conducted research in Machine translation from Japanese and French to
Vietnamese. This study was done by using the SMT approach. They conducted the experiments
on parallel corpora collected from TED talks. They used phrase-based and tree-to-string models
and have shown that the SMT system trained on French to Vietnamese obtains better results than
the system of Japanese to Vietnamese because French and Vietnamese have more similarities in
the structures of sentences than between Japanese and Vietnamese.

Brour et al. [84] developed Arabic text language into Arabic sign language NMT. Their system is
based on previously developed two approaches with the same source and target language: rule-
based Interlingua and example-based approaches. They stated the limitations of previous studies
as the requirement of linguistic knowledge necessary to develop the rules. In addition to this, they

41 | P a g e
stated notable results achieved by neural machine translation by listing known companies:
including Google [27] who currently joined NMT. They implemented the system using a
feedforward back-propagation ANN model. The system is trained using a dataset containing about
9,715 sentences, and it is evaluated with 73 simple sentences. The system is evaluated using BLEU
score, and it is compared to the first version ATLASLang MTS, the 4-gram average BLEU score
obtained for ATLASLang NMT is 0.79.

Matsumura et al. [82] conducted research in English-Japanese NMT with Encoder-Decoder-

Reconstructor. In [8] researchers conducted research on Chinese-English NMT by using an
Encoder-Decoder-Reconstructor framework for back-translation. Encoder-Decoder Reconstructor
framework for back-translation is first time proposed by Tu et al. [83] to address the problem NMT
suffers, which is the repeating or missing words in the translation. In this method, they selected
the best forward translation model in the same manner as Bahdanau et al. [8] and then trained a
bi-directional translation model as fine-tuning. Their experiments show that it offers a significant
improvement in BLEU scores in Chinese-English translation task. The researchers in [82]
performed the same approaches and in English-Japanese task too in addition to evaluating the
effectiveness of pre-training by comparing it with a jointly-trained model of forward translation
and back-translation. They used two parallel corpora: Asian Scientific Paper Excerpt Corpus
(ASPEC) and NTCIR PatentMT Parallel Corpus for training and testing result. They used only the
first 1 million sentences sorted by sentence-alignment similarity but they excluded sentences with
more than 40 words from the training data. Their experiments had 512 hidden units, 512
embedding units, 30,000 vocabulary size and 64 batch size. They used Adagrad (initial learning
rate 0.01) for optimizing model parameters. They trained their model on GeForce GTX TITAN X
GPU. By using ASPEC corpus in Reconstructor (Jointly-Training) model, they obtained a 26.04
BLEU score in 174 hours for English-Japanese translation. By using NTCIR corpus in
Reconstructor (Jointly-Training) model, they obtained 29.04 BLEU score in 252 hours for English-
Japanese translation. By using ASPEC corpus in Reconstructor (Jointly-Training) model, they
obtained 16.29 BLEU score for Japanese- English translation. By using NTCIR corpus in
Reconstructor (Jointly-Training) model, they obtained 28.95 BLEU score for Japanese- English
translation. But finally, they concluded that the system does not significantly improve translation

42 | P a g e
accuracy in Japanese-English translation. In addition, it is proved that the encoder-decoder-
reconstructor without pre-training worsens rather than improves translation accuracy.

Almahairi et al. [17] proposed NMT for the task of Arabic translation in both directions (Arabic-
English and English-Arabic) and compared a Vanilla Attention-based NMT system against a
Vanilla Phrase-based system. Preprocessing Arabic texts can increase the performance of the
system, especially normalization, but the model consumes much time for training.

Sutskever et al. [9] conducted research in English to French Machine Translation using Sequence
to Sequence Learning with Neural Networks. The study was carried out by using Neural Machine
Translation approach. They applied Deep Neural Network (DNN) approach on the previous study
carried out in phrase-based SMT approach [87]. The model they used is the Recurrent Neural
Network language model. They used a multilayered Long Short-Term Memory (LSTM) to map
the input sequence to a vector of fixed dimensionality, and then another deep LSTM to decode the
target sequence from the vector. They used WMT’14 English to French dataset. They trained their
models on a subset of 12M sentences consisting of 348M French words and 304M English words,
which is a clean “selected” subset from [87]. They chose this translation task and this specific
training set subset because of the public availability of a tokenized training and for comparing the
performance from the baseline SMT [87]. They evaluated their models using the standard BLEU
score metric. On the WMT’14 English to French translation task, they obtained a BLEU score of
34.81 using a simple left-to-right beam-search decoder. For comparison, a phrase-based SMT
system achieves a BLEU score of 33.3 on the same dataset [87]. This result shows that a neural
network architecture outperforms a phrase-based SMT system. When they reverse the order of the
words in all source, the BLEU score increased to 36.5. Finally, they found that reversing the order
of the words in all source sentences (but not target sentences) improved the LSTM’s performance
markedly, because doing so introduced many short term dependencies between the source and the
target sentence which made the optimization problem easier. Additionally, they confirmed LSTM
did not have difficulty in long sentences.

Norouzi et al. [27] conducted research in Google’s Neural Machine Translation System: Bridging
the Gap between Human and Machine Translation. The study was done by using NMT approach
for the purpose of overcoming many of the weaknesses of previous Google’s conventional phrase-
based translation systems. In addition to this, they improved NMT robustness problem
43 | P a g e
(particularly when input sentences contain rare words) which is stated as a problem by many
previous researchers [9, 86]. Their model is sequence-to-sequence learning framework with
attention. It has three components: an encoder network, a decoder network, and an attention
network. It consists of a deep LSTM network with 8 encoder and 8 decoder layers. The encoder
transforms a source sentence into a list of vectors, one vector per input symbol. Given this list of
vectors, the decoder produces one symbol at a time, until the special end-of-sentence symbol
(EOS) is produced. A decoder is implemented as a combination of an RNN network and a softmax
layer. The encoder and decoder are connected through an attention module which allows the
decoder to focus on different regions of the source sentence during the course of decoding. To
improve parallelism and therefore decrease training time, their attention mechanism connects the
bottom layer of the decoder to the top layer of the encoder. To improve the handling of rare words,
they divided words into a limited set of common sub-word units (“word pieces”) for both input
and output. They followed a beam search technique for implementation purpose. For testing their
system, they used WMT’14 English-to-French and English-to-German benchmarks as a dataset.
They evaluated their models using the standard BLEU score metric. Specifically, on WMT’14
English-to-French, their single model scores 38.95 BLEU, an improvement of 7.5 BLEU from a
single model without an external alignment model reported in [85] and an improvement of 1.2
BLEU from a single model without an external alignment model reported in [86]. Likewise, on
WMT’14 English-to-German, their single model scores 24.17 BLEU, which is 3.4 BLEU better
than a previous competitive baseline [86]. They also reported as, on production data, their
implementation is even more effective. Finally, they reported as human evaluations show that their
system has reduced translation errors by 60% compared to their previous phrase-based system on
many pairs of languages: English↔French, English↔Spanish, and English↔Chinese.
Additionally, their experiments suggest the quality of the resulting translation system gets closer
to that of average human translators.

44 | P a g e
3.3 Machine Translation involving Ethiopian languages
Solomon Teferra et al. [3] conducted research in Parallel Corpora for bi-directional Statistical
Machine Translation for Seven Ethiopian Language Pairs. The researchers presented some
Ethiopian language researches conducted by graduate students and mainly raised the unavailability
of linguistic resources stated by students which in turn affects the results that they obtain. They
attempted towards the development of parallel corpora for English and Ethiopian Languages, such
as Amharic, Tigrigna and Ge’ez from the Semitic, Afan-Oromo from the Cushitic and Wolaytta
from Omotic language families. But they prepared a corpus by using religious resources other
domains. For example: - the Tigrigna-English and Afan Oromo-English corpora are in legal and
religious (both bible and other religious collections) domains. The Wolaytta-English and Ge’ez-
English language pairs are from the religious domain only. However, the Ge’ez-English corpus is
only from the Bible while the Wolaytta-English consists of the Bible and other religious
collections. They tried to study the nature of different language pairs. They used the corpora they
developed for conducting a bi-directional SMT experiment. In the experimental setup, they used
Moses with GIZA++ alignment tool for aligning words and phrases. SRILM toolkit was used to
develop language models using semi-automatically prepared corpora from the training and tuning
corpora of target languages. They achieved a BLEU score of 13.31 for English-Amharic translation
while the Amharic-English has a 22.68. Similarly, the English-Tigrigna and Tigrigna-English have
BLEU scores of 17.89 and 27.53, respectively. Likewise, English-Afaan Oromo has a 14.68 BLEU
and Afan Oromo-English has 18.88 BLEU score. In a similar way, the English-Wolaytta
translation has BLEU of 10.49 while Wolaytta-English has 17.39. Finally, The English-Ge’ez and
Ge’ez-English translation has a BLEU score of 6.67 and 18.01, respectively. Finally, they
concluded that the English-Ethiopian languages SMT systems have less BLEU scores than that of
Ethiopian languages-English ones. The reason they raised is the fact that when the Ethiopian
languages are used as a target language, the translation from English as a source language is
challenged by many-to-one alignment.

Dawit Mulugeta [45] conducted research in Ge’ez to Amharic automatic MT using SMT approach.
As a research methodology, the author used qualitative experimental method to investigate the
effect of variables such as normalization, corpus and test split options on the SMT result. The data
used for the research experiment were found from both online and manually prepared. Totally 12,
45 | P a g e
860 parallel sentences were used for both languages. Regarding the organization of the data, out
of the bilingual data, 90% for training and 10% for testing were used for the experiment. Moses
decoder, IRSTLM, GIZA++ and BLEU were used to build translation model, language model,
Word alignment and evaluation of the Ge’ez to Amharic MT system respectively. The parallel
corpus used for the experiment was sentence level aligned. The average translation result was 8.26
BLEU score.

Mulu Gebreegziabher and Besacier conducted preliminary experiments on English-Amharic SMT

[21]. The main objective of the research was the need to begin empirical researches towards
developing English-to-Amharic SMT. The major problem they stated was the rule-based approach
yet not recommended to be used for under-resourced languages like Amharic due to the different
linguistic knowledge, rules and resources required. To meet their goal, the total corpus size of 632
parliamentary corpora of which 115 had been used for the experiment. The experiment had been
conducted using 18,432 English-Amharic sentence pairs extracted from these corpora. Out of the
total 90% randomly selected sentence pairs had been used for training while the remaining 10%
sentence pairs were used for testing. There were different software resources used for the
experiment in general integrated with MOSES like SR Luthra Institute of Management (SRLIM)
to build the language model, Giza++ for building translation model and BLEU metric for
evaluating the performance of the MT system. When the researchers evaluate their MT system for
English-to-Amharic SMT the baseline phrase-based BLEU score results 35.32% translation had
been achieved. The preliminary experiment result shows that the EASMT can translate the basic
meaning of the English sentence when translating into Amharic sentence. However, there are some
strong as well as weak points in the performance of the EASMT. Keeping the storing side, to
address problems like non-translated words, wrongly translated, insertion, deletion, alignment
problem, preposition usage, and morphological errors they had used word segmentation on the
target side is vital. According to these results, more experimentation and research is required to
further improve the translation accuracy of the EASMT. The experiment done so far is encouraging
as the translation is done from less inflected English language to a morphologically rich language
Amharic.

46 | P a g e
Eleni Teshome developed bi-directional English-Amharic machine translation using a constrained
corpus [4]. The objective of this study was developing a bi-directional English-Amharic MT
system by using the SMT approach. In this paper, two different corpora prepared: the first corpus
(Corpus I) was made of 1,020 simple sentences that have been prepared manually. The second is
made of 1,951 complex sentences (Corpus II) from sources, one from the Bible and the other from
the public procurement directive of the Ministry of Finance and Economic Development. From
total of 2,971 sentences, 10% were taken for the testing process and the remaining 90% is used for
training. Since the translation is bi-directional, two language models were developed, one for
Amharic and the other for English and translation models were also built. Two different
experiments were conducted and the evaluation was performed by using two different
methodologies. The first experiment was performed using simple sentences and the accuracy of
Amharic to English translation is 94% and English to Amharic translation and 90.59% for Amharic
to English translation. The second experiment was performed by using the manual questionnaire
method, the accuracy of English to Amharic translation is 91% and the accuracy of Amharic to
English translation is 97%. For complex sentences, the first methodology, the accuracy of the
translation from English to Amharic was 73.38% and from Amharic to English translation was
84.12%. The second experiment was performed by using the manual questionnaire method was
87% for the English to Amharic translation and 89% for Amharic to English translation. The study
shows Amharic to English translation has better accuracy than English to Amharic translation.

Sisay Adugna developed English–Afaan Oromo machine translation using a statistical approach
[6]. This study had two main goals: the first is to test how far one can go with the available limited
parallel corpus for English-Afaan Oromo language pair and the applicability of existing
SMTsystems on these language pairs. The second one is to analyze the output of the system with
the objective of identifying the challenges that need to be addressed. In this study, the architecture
includes four basic components of statistical machine translation, which are language modelling,
translation modelling, decoding, and evaluation. By using a corpus of about 20, 000 bilingual
sentences, an author achieved the translation accuracy of 17.74%.

Jabesa Daba and Yaregal Assabie developed bi-directional English-Afaan Oromo MT using hybrid
approach [23]. The research work is implemented using a hybrid of rule-based and statistical
approaches. Corpus was collected from different domains. They collected 3,000 parallel sentences.

47 | P a g e
Out of those sentences 90% are used for training and the remaining 10% are used for testing. Since
the system is bidirectional, two language models are developed, one for English and the other for
Afaan Oromo. The study was carried out with two experiments that are conducted by using two
different approaches and their results are recorded. The first experiment is carried out by using a
statistical approach. The result obtained from the experiment has a BLEU score of 32.39% for
English to Afaan Oromo translation and 41.50% for Afaan Oromo to English translation. The
second experiment is carried out by using a hybrid approach and the result obtained has a BLEU
score of 37.41% for English to Afaan Oromo translation and 52.02% for Afaan Oromo to English
translation. From the result, we can see that the hybrid approach is better than the statistical
approach for the language pair and a better translation is acquired when Afaan Oromo is used as a
source language and English is used as a target language.

Yitayew Solomon conducted bidirectional English-Afaan Oromo MT Systems using SMT

approach [29]. For the researcher to have such an objective was, the research done by Sisay
Adugna [6] and Jabesa Daba [23] score poor performance with the BLEU score of 17% and 37%
respectively. This is due to the alignment quality of the prepared data and the unavailability of a
well-prepared corpus for the MT task for English to Afaan Oromo MT. To build the translation
model, 6400 parallel sentences and 19300 and 12200 sentences, to build a language model for both
English and Afaan Oromo languages were used respectively. Randomly, for training 90% and 10
% testing of corpus size were used. 700 simple and 5700 complex sentences with a total of 6400
sentences was used. Moses for Mere Mortal is used for SMT and integrates different toolkits which
are used for translation purpose such as IRSTLM for language model, Decoder for translation,
MGIZA++ for word alignment. Hunalign, Anymalign and MGIZA++ where software tools, used
for sentence, phrase and word level alignment, respectively. BLEU score was used to evaluate the
MT system. Preprocessing tasks sentence splitting, margining and true casing are used to make
ready the corpus for the experimentation. Six experiments were done by the researcher to select
the optimal alignment quality for English to Afaan Oromo where, Experiments I and II for word-
level alignment, Experiments III and IV for phase level alignment and Experiments V and VI for
sentence-level alignment. Word level alignment when the max phrase length is 4 and min is 1which
record 21% and 42% BLEU score from English-Afaan Oromo and Afaan Oromo-English
respectively. Phrase level alignment when the max phrase length is 16 and min is 4 which record

48 | P a g e
27% and 47% BLEU score from English-Afaan Oromo and Afaan Oromo-English respectively.
Sentence level alignment when the max phrase length is 30 and min is 20 which record 18% and
35% BLEU score from English-Afaan Oromo and Afaan Oromo-English respectively. Optimal
alignment is phrase-level alignment when the max phrase length is 16 and min is 4 which record
27% and 47% BLEU score from English-Afaan Oromo and Afaan Oromo-English respectively.
Finally, the researcher recommends, better results can be achieved by using the corpus with proper
alignment used for training the system. So, by increasing the size of the training data set that
properly aligned at phrase level one can develop a better bi-directional English-Afaan Oromo
machine translation.

3.4 Summary
DNN is a newly emerging approach and proved to achieve excellent performance. Unlike the
traditional SMT, the NMT aims at building a single neural network that can be jointly tuned to
maximize the translation performance. After a few pioneer works in exploring neural features in
SMT systems [28], NMT quickly become the dominant approach for MT. The researchers in [10]
and [11] first propose to use the encoder-decoder architecture to do sequence to sequence mapping.
At the same time, the other study in [8] applies end-to-end MT. In [9] the attention mechanism
proposed to dynamically attend to different source words when generating different target words,
which becomes the default component of current NMT systems. In general, as we have seen in
the above related works, most of the papers done for different language pairs in recent years are
based on attention-based NMT and they resulted in promising results when replacing SMT by
NMT [9, 10, 11, 12]. It is the current SOTA MT technology and also many researchers are joining
this approach today. Speakers of Wolaita language who are unable to speak and understand
Amharic cannot communicate and interact with Amharic speakers without finding translators.
Thus, non-Amharic speakers of Wolaita language speakers face problems of lack information. If
documents and news articles automatically translated into Wolaita language, from Amharic
language, it solves the problem. As the researcher knowledge, no prior study has carried out in
Amharic to Wolaita NMT. Thus, we apply the attention-based NMT approach for Amharic to
Wolaita neural machine translation.

49 | P a g e
 Chapter 4: Design of the Proposed System
4.1 Introduction
This Chapter covers the architectural design of Amharic-to-Wolaita NMT both in attention-based
and non-attention-based mechanism. The proposed system is based on Seq2Seq learning model
with Recurrent Neural Network based Encoder-Decoder architecture. The Seq2Seq model with
RNN based Encoder-Decoder architecture is the SOTA technique in NMT. Encoder and Decoder
help the model gain a deeper understanding of the two languages, i.e., source and destination
languages. The Encoder encodes the complete information of the source sequence, which is the
source language, Amharic, into a single real-valued vector using word-embedding, also known as
the context vector. This context vector is passed to the Decoder to produce an output sequence,
which is the target language, Wolaita. We used the GRU unit to train the system. The system based
on an attention mechanism solves the bottleneck of basic Encoder-Decoder non-attention-based
NMT.

4.2 System design

4.2.1 Language Model Training
To train our system model, we used a forward pass and back-propagation algorithms which uses a
deep multi-layer GRU architecture as we detailed in Algorithm 4.1 and Algorithm 4.2,
respectively. For each new input xi at the time i, GRU unit updates its memory to produce a hidden
state hi which one can think of as a representation for the partial sequence x1,x2, x3…….xi.
Mathematically it can be represented as:
ht = f (xt, ht−1) (1)
In the above formula, f is an abstract function that computes a new hidden state given the current
input xt and the previous hidden state ht−1. The starting state h0 is often set to 0 though it can take
any value. A popular choice of f is tanh as being a non-linear function.
ht = tanh(Wxhxt + Whhht−1) (2)
Since the Encoder and Decoder share many operations in common in forward pass algorithm, we
combine the Amharic sentence x (length mx), the Wolaita sentence y (length my), and the end of
sentence markers “<end>” together to form an input sequence s as shown in Line 1 of Algorithm
1. We first start with the Encoder weights and initial states set to zero (lines 2-3). The algorithm

50 | P a g e
switches to the Decoder mode at time mx+ 1 (line 5). The same GRU codebase (lines 8-11) is used
for both the Encoder and Decoder in which embedding is first looked up for the input st; after that,
hidden states as well as GRU cell memories are built from the bottom layer to the top one (the Lth
layer). In Line 10, GRU refers to the entire formulation in which anyone who interested to work
with other hidden units such as bi-RNN and LSTM can easily replace it. Lastly, on the Decoder
side, the top hidden state is used to predict the next symbol st+1 (line 13); then, a loss value lt and
a probability distribution pt computed according to Equation 3 and Equation 4.
St = Why.ht ) (3)
pt= softmax (st) ) (4)
Here, Why ∈ R|Y |×d, with d being the dimension of the GRU hidden state, to compute a score vector
St. Computing the total loss for (x, y) during the forward pass is written in Equation 5.
-1/n [log (y1 (word1) + log (y2 (word2) +log (yn ("<end>")) (5)
The back-propagation algorithm reveals many similarities compared to the forward pass algorithm
except that we have reversed the procedure. First, we initialize gradients of the GRU layers at the
final time step (line 1) as well the model weights on the Decoder size (line 2) to zero. At time mx,
we switch to the Encoder mode by saving the currently accumulated GRU and embedding
gradients for the Decoder (line 5) and starting to accumulate gradients for the Encoder weights
(line 6). The back-propagation procedure presented earlier for GRU can simplify the core NMT
gradient computation (lines 8-18) by making the following two referents: (a) Predict grad:- which
computes gradients for the target-side losses with respect to the hidden states at the top layer and
the softmax weights Why ; and (b) GRU grad:- which computes gradients for inputs to GRU and
the GRU weights per layer GRU(l). It is important to note that in Lines 10 and 15 of Algorithm 2,
we add the gradients (flowing vertically from either the loss or the upper GRU layer) to the gradient
of the below layer (which already contains the gradient back-propagated horizontally) instead of
overriding it. In Line 18, we perform sparse updates on the corresponding embedding matrix for
participating words only.

51 | P a g e
Algorithm 4.1: RNN language model training algorithm – forward pass.
Input: source sentence x of length mx, target sentence y of length my.
Parameters: encoder Wencoder, GRUencoder; decoder Wdecoder, GRUdecoder
Output: loss l and other intermediate variables for back-propagation.
1. s  [x,<end> , y,<end> ] ; // Length of s is mx + 1 + my + 1
2. We, GRUencoder (1…L)  Wencoder, GRUencoder; // Encoder weights
3. h0(1..L), c0(1..L)  0 ; // Zero initialization
4. for t = 1 ➔ (mx + 1 + my) do
// Decoder transition
5. if t == (mx + 1) then
We, GRUencoder (1…L)  Wdecoder, GRUdecoder;
6. end
// Multi-layer GRU
7. ht(0)  Emb LookUp(st, We) ;
8. for l = 1 ➔ L do
9. ht(l), ct(l)  GRU(ht-1(l), ct-1(l)), ht(l-1), GRU(l) // GRU hidden unit
10. end
// Target-side prediction
11. if t ≥ (mx + 1) then
12. lt, pt  Predict(st+1, ht(L), Why) ;
13. end
14. end
Algorithm 4.2: RNN language model training algorithm – back-propagation pass
1. dhmx+ 1 + my, (1..L), dcmx+ 1 + my, (1..L)  0; / / Cell and state gradients
2. dGRU(1…L), dWe, dWhy  0; // Model weight gradients
3. for t = (mx + 1 + my) ➔ 1 do // Encoder transition
4. if t == mx then
5. dWedecoder, dGRUdecoder  dWe, dGRU(1..L) ; // Save decoder gradients
6. dWe, dGRU(1..L)  0
7. end
8. if t ≥ (mx + 1) then // Target-side prediction
9. dh, dW  Predict_grad(st+1, pt, ht(L));
10. dht(L)  dht(L) + dh ; // Vertical gradients
11. dWhy  dWhy + dW;
12. end
// Multi-layer GRU
13. for l = L ➔ 1 do // Recurrent gradients
14. dht-1(l), dct-1(l), dx, dT  GRU_grad (dht(l), dct(l), ht-1(l), ct-1(l), ht(l-1))
15. dht(l-1)  dht(l-1) + dx; // vertical gradients
16. dGRU(l)  dGRU(l) + dT;
17. end
18. dWe  Emb_grad_update(st, dht(0), dWe) ;
19. end
20. dWeencoder, dGRUencoder  dWeencoder, dGRU(1..L) // Save encoder gradients

52 | P a g e
When we are training the Decoder, we don’t take the output of Decoder in previous step we
actually give the correct output from the previous step. That is called teacher forcing. In other word
we may want to produce the correct Amharic sentence “ሁሉ በአግባብና በሥርዓት ይሁን” equivalent in
Wolaita sentence like “ubbabaikka wogaaninne maaran hano”. But the system may produce
different word instead of “ubbabaikka” in Wolaita. Teacher forcing handles this issue by making
the system to produce correct translation. Thus, teacher forcing feeds the target as the next input.

4.2.2 Language Model Testing

After training our system model, we need to be able to use it to translate, or decode, unseen
Amharic sentences. The strategy we used to translate an Amharic sentence to Wolaita sentence is
performing beam searching mechanism which outperforms the greedy searching mechanism by
solving its garden-path problem [85, 104]. The idea is simple: we first encode preprocessed
Amharic sentence, "ሁሉ በአግባብና በሥርዓት ይሁን" and the decoding process is started as soon as an
end-of-sentence marker “<end>” for the Amharic sentence is fed as an input to the Encoder. For
performing beam searching mechanism we performed the following steps. (a) At each timestep on
the Decoder side, we keep track of the top K (the beam size) best translations together with their
corresponding hidden states. (b) Then we select the top K most likely words. (c) Given K previous
best translation ×K best words, we select a new set of K best translations for the current timestep
based on the combined scores (previous translation scores + current word translation scores).
When predicting the first word of the output sentence, we keep a beam of the top K most likely
word choices. They are scored by their probability. Then, we use each of these words in the beam
in the conditioning context for the next word. Due to this conditioning, we make different word
predictions for each. We now multiply the score for the partial translation (at this point just the
probability for the first word), and the probabilities from its word predictions. We select the
highest-scoring word pairs for the next beam. This process continues until the end of sentence
token <end> is produced. At each time step, we accumulate word translation probabilities, giving
us scores for each hypothesis. At this point, we remove the completed hypothesis from the beam
and reduce beam size by 1. The search terminates when no hypotheses are left in the beam.

The search produces a graph of hypotheses, as shown in Figure 4.1. It starts with the start of
sentence symbol <start> and its paths terminate with the end of sentence symbol <end>. Given the
complete graph; the resulting translations can be obtained by following the back-pointers. The
53 | P a g e
complete hypothesis (i.e., one that ended with a <end> symbol) with the highest score (i.e. the path
with bold lines in Figure 4.1) points to the best translation. When choosing among the best paths,
we score each with the product of its word prediction probabilities. During the search, all
translations in a beam have the same length, so the normalization would make no difference.

Figure 4.1: Language Model Testing in the Beam Search Graph

Here n = 5 best partial translations (called hypotheses) are selected. An output sentence is complete
when the end of sentence token <end> is predicted. We reduce the beam after that and terminate
when n full-sentence translations are completed. Following the back-pointers from the end of
sentence tokens allows us to read them off. Empty boxes represent hypotheses that are not part of
any complete path.

4.3 System Architecture

Attention-based Amharic-to-Wolaita NMT is a translation system where a given Amharic text is
translated into equivalent Wolaita sentence through three layers: Encoder, Attention, and Decoder.
An input Amharic sentence is preprocessed before performing one-hot representation. Word
embedding can be formed from one-ot representation and it is given to Encoder layer as input.
Encoder layer outputs attention weight and gives it to attention layer. Attention layer processes
attention weight and finally outputs the single representation for input sentence. This
representation is given to Decoder and it produces equivalent Wolaita translation based on input.
The architecture of the system is shown in Figure 4.2.

54 | P a g e
Figure 4.2: Architecture of Attention-based Amharic-to-Wolaita NMT
55 | P a g e
4.4 Text Preprocessing
For our study, the training and testing corpus is collected from different sources which contain
parallel text (Amharic and Wolaita) in religious domain, i.e., bible texts. Preprocessing is involved
in preparing the input sentence into a format that is suitable for the morphological analysis. The
preprocessing stage consists of steps such as tokenization, normalization and stop-word removal.
4.4.1 Tokenization
The first step in the preprocessing of the sentence is tokenization, which is also known as lexical
analysis. Tokenization is essentially splitting of a sentence into smaller units, such as individual
words. Each of these smaller units is called token. Most of the time, tokenization of words in
Amharic language is performed using Amharic punctuation marks as delimiter characters which
are white space, '፡'(hulet netib), ‘።'(arat netib), '፣'(netela serez),'፤'dereb serez), ‘!' (kaleagano) and
'?' (Question mark) [116]. We used white space (‘ ‘) delimiter for tokenization with the buit-in
function tf.keras.preprocessing.text.Tokenizer(filters=' ') library from a Tokenizer class. In order
to find the boundaries of the sentence, we added <end> tag as indicator of the end of a sentence so
that the model knows as it reached the end of words in input sentence. For example:- When we
perform tokenization in the input sentence ጳውሎስ እንዲህ በማለት ፅፏል ስለ ምንም ነገር አትጨነቁ ከዚህ ይልቅ
ስለ ሁሉም ነገር በፀሎትና በምልጃ ከምስጋና ጋር ልመናችሁን ለአምላክ አቅርቡ, it results with the tokens [‘<start>
‘,‘ጳውሎስ’, ‘እንዲህ’, ‘በማለት’, ፅፏል’, ’ስለ’,’ ምንም’, ’ነገር’,’ አትጨነቁ’, ‘ከዚህ’, ‘ይልቅ’, ‘ስለ’, ‘ሁሉም’, ‘ነገር’,
‘በፀሎትና’, ‘በምልጃ’, ‘ከምስጋና’, ‘ጋር’, ‘ልመናችሁን’, ‘ለአምላክ’, ‘አቅርቡ’, <end>’] as output. Tokenization
in Amharic and Wolaita presents a problem because of the rich and complex morphology of the
two languages. Thus, each token has to be normalized.
4.4.2 Normalization
Normalization is performed on the word tokens that result from text tokenization. We performed
text normalization for both languages based on the properties of the languages. For performing
Amharic sentence normalization, we used the algorithms adopted by Tessemma Mindaye
Mengistu in [116]. The author in [116] discussed the two types of normalization issues that arise
in the Amharic language. The first one is the identification and replacement of short hand
representation of a word that is written using forward slash “/” or period “.” as shown in Algorithm
4.1. An example is the replacement of “ት/ቤት” by “ትምህርት ቤት”/school/. The second normalization
issue is the identification and replacement of Amharic alphabets that have the same pronunciation
56 | P a g e
as shown in Algorithm 4.2. The replacement is made using a representative alphabet from a set of
similar alphabets. For example, identifying words like ውሃ, ውሀ, ውሓ, ውሐ, ውኃ, ውኀ/ wɨhä/ which
are pronounced the same to mean “Water” and replace by /wɨhä/. In addition to Tessema’s
normalization issues, we also performed normalizing of words with Labialized Amharic characters
such as በልቱዋል or በልቱአል to በልቷል. And also, we performed removing quotes, removing spaces;
adding a space between the word and the punctuations like “?”; cleaning digits; and removing of
special characters for both Amharic and Wolaita sentences. We also performed Converting of text
to lower case of the Wolaita sentence.

Algorithm 4.1: Amharic Word Expanding Algorithm from [116]

Read a character before ‘/’ or ‘.’
Search the location of the characters in corpus
If found
Return the corresponding expanded word
Else
Return the original word
Endif
Algorithm 4.2: Character Normalization Algorithm from [116]
Read a character
If the character is one of ኅሐ orኸ replace them with ሀ (The same applies for the orders,
i.e. the orders of ሐ and ኀ will be replaced by the corresponding orders of ሀ).
Return the replaced character
Else If the character is one of ኁሑ or ዅ replace them with ሁ.
Return the replaced character
Else If the character is one of ኂሒ orኺ replace them with ሂ.
Return the replaced character
Else If the character is one of ኃሓ or ኻ replace them with ሃ.
Return the replaced character
Else If the character is one of ኄሔ or ዄ replace them with ሄ.
Return the replaced character
Else If the character is one of ሕኅ or ኽ replace them with ህ.
Return the replaced character
Else If the character is one of ኆሖ orኾ replace them with ሆ.
Return the replaced character
Else If the character is ሠ replace it with ሰ (The same applies for ሡሢሣሤሦ).
Return the replaced character
Else If the character is ዐ replace it with አ(The same applies for ዑዒዓዔዕዖ)
Return the replaced character
Else If the character is ጸ replace it with ፀ (The same applies for ፁፂፃፄፅፆ)
Return the replaced character
Endif

57 | P a g e
4.4.3 Stop Word Removal
In Amharic, common words such as pronouns, prepositions and conjunctions occur so frequently
that they cannot give any useful information about the content and be discriminatory for a specific
class. These words are called stop words. Stop words are low information bearing words such as
“ነው” or “ና”, typically appearing with high frequency as listed in Appendix II [32]. Stop words
may be context dependent. There is not one definite list of stop words, which all tools use, and
such a filter is not always used. Some tools specifically avoid removing them to support phrase
search. Like other languages, Amharic has non-content bearing words. Usually words such as
articles (e.g. ‘ያኛው’, ‘ይሄ’), conjunctions (‘ና’, ‘ነገርግን’, ‘ወይም’) and prepositions (e.g. ‘ውስጥ’, ‘ላይ’)
do not have a significant discriminating power in the meaning of ambiguous words. In this thesis,
stop words like ‘ነው’, ‘እስከ’, ‘እንደ’, etc, are discarded from input texts as these words are
meaningless to derive the “sense” of the particular sentence. Then, the text containing meaningful
words (excluding the stop words) pass through stemming.

4.5 Stemming
Stemming is the process of removing affixes (i.e. prefixes, infixes and suffixes) that improve the
accuracy and performance MT systems. It is the process of reducing inflected and/ or sometimes
derived words to their word stem base or root form- generally a written word form. The stem need
not be identical to the morphological root of the word. It is very important to reduce the various
original forms of a word to a single root/ stem of these words as it reduces the dimensionality of
the feature space. It will in turn deliver a significant improvement in the accuracy of MT system.
In this thesis, we used Hornmorpho stemming algorithm developed by Gasser [47]. Table 4.1
shows an example of stemmed Amharic Words using Hornmorpho stemming algorithm.

Table 11: Stemmed Amharic Words

Amharic Root Attached negated? Is it Subj Obj Word

word word pp plural? pron pron tense
በሉ ብልእ 0 0 እነሱ ሀላፊ
እየበላ ብልእ 0 0 እሱ አሁን
ልትበላ ብልእ 0 0 እሷ ትንቢት
አልበላሁም ብልእ 1 0 እኔ
የበላሁት ብልእ የ 0 0 እኔ

58 | P a g e
4.6 One-hot Representation
In NMT the original sequential data is not suitable to read for the neural network. It needs some
form of representation in a suitable format. Words like "አትጨነቁ" in a corpus is a sequence of
ASCII character encodings. Since a neural network is a series of multiplication and addition
operations, the input data needs to be number(s). To change the data into this suitable format, each
word of the sentence in the data must be identified and represented by a unique index and the
process is called one-hot representation. It is sometimes called indexing. One-hot representation is
a way of converting individual unique words into unique numbers. In our model, we used a built-
in method index_word[index] methods to convert vocabulary words into a unique id
representation. Here the first word in a vocabulary is represented with the first index zero (0) and
the last word is represented with the index total number of unique words minus one. An
index_word attribute is a word-to-index dictionary where words are the keys and the corresponding
integers are the values. Thus, each word in a sentence corresponds to a vocabulary item in a
vocabulary. Here, we have a vocabulary item of 1,734,130 characters without blanks which is
from 254,328 words in 9280 lines. After normalizing, removing stop words and stemming, total
number of a unique vocabulary word length is 22,075 words for Amharic vocabulary and 18,987
words for Wolaita vocabulary and each component in that vector corresponds to a specific word
in the vocabulary. The neural network that operates on this vocabulary will not be able to do
anything with any other word except the words listed in vector representation. So, we represent
each word by simply putting 1(one) in the position corresponding to that words in the vector and
0(zero) everywhere else as shown in Table 4.2.

Table 12: Word Representation in Indexing

59 | P a g e
4.7 Word Embedding
From sequences of vector representation (one-hot representation), word embedding will be
formed. It is a way of representing words on a vector space where the words having the same
meaning have similar vector representations [117] as shown in Table 4.3 and Figure 4.3. Thus, it
focuses on the construction of a semantic representation of words based on the statistical
distribution of word co-occurrence in the text corpus. One of the basic advantages of neural word
embedding is the reduction of out-of-vocabulary impact. This is possible because words will not
be completely unknown as far as they have feature vectors even if they may not be seen in the
training dataset. Words can be encoded into limited vector spaces or neural word embeddings in
one of the two methods, the continuous bag of words (CBOW) method and the skip-gram method
(Word2Vec). CBOW takes the average of the possible contexts of a word in representing it in a
limited dimensional vector space. For example, ‘Apple’ can refer to either the name of the
company or the type of fruit. CBOW places the vector of the word ‘Apple’ to a medium position
of the two contexts.

Skip-gram model was introduced by Mikolov et al. [117] which is an efficient method for learning
high quality vector representations of words from large amounts of unstructured text data. The
authors of Skip-gram model stated as it is efficient and more accurate than CBOW in that it does
not involve dense matrix multiplications. Skip-gram method, on the other hand, can assign two
vector values for the above two contexts of ‘Apple’. Based on the above preferences, we used
Skip-gram Word2Vec algorithm to construct our semantic model in this thesis.

Figure 93: Skip Gram Model

60 | P a g e
 Table 13: Semantic Relationship of Words Representation in Word-embedding

Here we can represent it as, meat is food (ስጋ ምግብ ነው) highly in 90%, human is rogue (ሰው የዱር-
እንስሳ ነው) in 10%) and meat is neither carnivore nor rogue (ስጋ ስጋ-በዪም የዱር እንስሳም አይደለም). Thus,
we can represent the remaining in this way. As we can see in the representation, አንበሳ (lion) and
ሰው (human) are very similar. Thus, we can say both አንበሳ ስጋ በላ (lion eat meat) and ሰው ስጋ በላ
(human eat meat). But we cannot say በሬ ስጋ በላ (ox eat meat), because it is semantically incorrect.
Similar words will have a similar vector. The position of a word within the vector space is learned
from text and is based on the words that surround the word when it is used. Thus, it stores each
word close to the other word which is close to it in the meaning as shown in Figure 4.4 [117].

Figure 10 Word-embedding Example in Matrices

61 | P a g e
4.8 Padding
The next step after word embedding is applying padding to the sentences that are longer or shorter
than a certain length. When batching the sequence of word ids together, each sequence needs to
be the same length. Since sentences are dynamic in length, we can add padding to the end of the
sequences to make them the same length. The Encoder part receives this embedded form of data
and generates a contextual relation between words. The Decoder part receives the context
generated by Encoder and searches for matching words within Wolaita language. When the
sequence includes the longer sentence, the enhancement mechanism is needed to help to store of
the longer context within the sentence. This enhancement mechanism is called padding. Thus,
padding is a term that describes the act of taking a value and adding pad characters to the left or
right (or both) sides of a string in order to fix the length of the string. We performed it by using a
built-in function tf.keras.preprocessing.sequence.pad_sequences ([inputs]) and it results in all the
Amharic and Wolaita sequences have the same length by adding padding to the end of each
sequence.

4.9 Encoding
Encoding is the process of converting the sequences of inputs into a single-valued context vector.
In our Encoder-Decoder model, we have used a RNN based model to design Encoder for Seq2Seq
model. The Encoder of a Seq2Seq network is a RNN that encodes the complete information of the
source sequence which is the source language like Amharic into a single real-valued vector, also
known as the context vector, which is passed to the Decoder to produce an output sequence, which
is the target language like Wolaita in basic Encoder-Decoder model. To work on textual data by
Encoder part of Encoder-Decoder LM, the data must be given in the embedded form of sequential
data. If X={x1, x2, x3… xn} is the input sequence to the Encoder with n sequence length, the
Encoder will produce Y= {y1, y2, y3… ym} context in the sequence of input data. To drive this
context, the selection of neural network architecture is necessary.

In the RNN based Encoder-Decoder model, Encoder can use either of DNN architecture. Some of
these architectures are LSTM, GRU, gated linear unit (GLU) or bi-directional RNN [88]. From
these architectures, GRU and LSTM are most familiar for Encoder part of NMT on text-based
translation [117]. The difference between these two architectures is that GRU has no output gate,

62 | P a g e
while LSTM has both output gate and a backward gate for storing the internal state of the network.
Therefore, GRU is LSTM that has no output gate, so writes the contents from its memory cell to
the larger network at each time step [27, 117]. We preferred the GRU unit as both Encoder and
Decoder unit. However, we used single layer GRU in architecture Figure in 4.2, we have used
deep-multilayer GRU. The GRU uses gate unit to control the flow of information. It uses the
current input and its previous output, which can be considered as the current internal state of the
network, to give current output [110].

Figure 11: Encoder Architecture

GRU uses two parameters to drive the sentence context. These are the collection of currently
inputted data and the previous output of the network [88]. Therefore, when using RNN the previous
output is considered as the internal state of the neural network and if the current input is the starting
word, then the input is marked by the sign of the start of token symbol <start> token indicator and

63 | P a g e
given as input as we have explained in Equation. 4.2. Then Encoder outputs context vector as its
final output from current input data and previous output of the system. This context vector will be
fed into Decoder part either by using the enhancement mechanism which is attention-mechanism
(we will see the detail in section 4.11) or without using an attention mechanism. If the attention
mechanism is left, the last layer of Encoder network will be connected to the Decoder part of the
network [88]. If we use the attention mechanism, the last layer of Encoder network will be
connected to attention layer of the network [111].

4.10 Decoding
Decoding is the process of generating the equivalent meaning to what encoded in Encoder [113].
In our Encoder-Decoder model, we have also used RNN based model to design Decoder for
Seq2Seq model. There are several choices for the Decoder architecture that combines these inputs
to generate the next hidden state: linear transforms with activation function, GRUs, LSTMs, etc.
Typically, the choice here matches the Encoder. So, if we use GRUs for the Encoder, then we also
use GRUs for the Decoder too. Since we used GRU architecture for Encoder part, the Decoder
part of RNN based system is also designed by using GRU architecture. It takes some representation
of the input context and the previous hidden state. Once the Decoder is set up with its context, a
special token to signify the start of output generation (i.e., <end> token appended to the end of the
input; there’s also one at the end of the output Wolaita sentence) is passed to it. Then, all layers of
GRU run one after the other, following up with a softmax on the final layer’s output to generate
the first output word. Then, we pass that word into the first layer and repeat the generation. This
is how we get the GRUs to act as a language model. Until the end of text generation, Decoder
outputs word prediction and generates a new hidden Decoder state and a new output word
prediction [112]. The number of GRUs used is set as the maximum length of the sequence in the
target sentence data (i.e., 5 GRU unit is used for <start> ሁሉ በአግባብና በሥርዓት ይሁን tokens) and
when the shorter sentence entered into the network, the end of sentence marked with the end of
string indicator symbol <end> and the other units are fed with padding which is used to change
the variable-length vector into a fixed-length vector by completing the length of shorter sentence
by using zero value.

In order to find the matching word, the Decoder uses the beam searching mechanism of best search
type with the use of softmax function which searches for the best matching word. Therefore, the
64 | P a g e
Decoder generates one word at a time and repeats the searching until it encounters end of string
indicator <end> which shows the last word of the sentence. See Figure 4.6 for an example of a
Decoder network. Once we have the output sequence, we use the same learning strategy as usual.

Figure 12: Decoder Architecture

Here is a recap that a Seq2Seq model with RNN based GRU unit does in order to translate the
Amharic sentence “ሁሉ በአግባብና በሥርዓት ይሁን" into the Wolaita sentence "ubbabaikka wogaaninne
maaran hano". First, we start with 5 one-hot vectors (i.e. one for each word including <start> token)
for the input. Then, a GRU network reads the sequence and encodes it into a context vector. This
context vector is a vector space representation of the notion of translating the Amharic text. It is
used to initialize the first layer of another GRU. We run one step of each layer of this network,
perform softmax on the last layer’s output, and use that to select our first output word. This word
is fed back into the network as input, and the rest of the sentence “ubbabaikka wogaaninne maaran
hano" is decoded in this fashion. During back-propagation, the Encoder’s GRU weights are
updated so that it learns a better vector space representation for sentences, while the Decoder’s
GRU weights are trained to allow it to generate grammatically correct sentences that are relevant
to the context vector [113].

65 | P a g e
 Figure 13: Basic Encoder-Decoder Model without Attention

4.11 Attention Mechanism

Encoder in basic Encoder-Decoder based RNN architecture encodes the complete information of
the source sequence into a single real-valued vector, also known as the context vector, which is
passed to the Decoder to produce an output sequence. Here, a context vector has the responsibility
to summarize the entire input sequence into a single vector. Thus, the Decoder has access only to
the output of context vector of the Encoder as shown in Figure 4.8.

Therefore, the entire input sequence representation in a single vector is inefficient; and unable to
represent when the sentences are being long and the vocabulary size is becoming large. One
effective way to address such a problem is through the attention mechanism, which has gained
popularity recently in training neural networks. Unlike Encoder-Decoder architecture without
attention, in which the source representation is only used once to initialize the Decoder hidden
state; Encoder-Decoder architecture with Attention model predicts a target word based on the
context vectors associated with the source position and the previously generated target words. In
this way, we solved the basic Encoder-Decoder problem to handle long sentences by adding
66 | P a g e
attention mechanism to it. In addition to this, we handled a larger number of vocabulary size
problem of basic Encoder-Decoder based approach by adding attention mechanism to Encoder-
Decoder LM which minimizes higher dimension vector into lower dimension vector [114]. It helps
to pay attention to the most relevant information in the source sequence.

Figure 14: Attention-based Encoder-Decoder Architecture

67 | P a g e
 Chapter 5: Experimentation and Evaluation
5.1 Introduction
Under this Chapter, the conducted experiments and discussion is presented in a way that states the
evaluation and performance of the experiments. Section 5.2 details about data collection and
preparation, Section 5.3 about system environment/tools used for system development, Section
5.4details about parameter optimization and training of the experimental systems and Section 5.5
details about the experimental results of attention-based and non-attention-based system in
comparison. Finally, this chapter details about Evaluation metrics and discussion on the result of
the study.

5.2 Data Collection and Preparation

For our study, we have collected datasets of total 3000 sentences of parallel Amharic-Wolaita
corpus from different sources mostly from religious books. 6280 datasets are taken from [3]. Hee
the datasets are available separately and we make it to Amharic-Wolaita parallel dataset. The total
dataset we used for the system training and testing is 9280 parallel sentences. After we had finished
collecting the datasets, we uploaded it into Google drive to load into Google collaboratory [115]
since we have used GPU rather than CPU for the sake of fastening the training time. Then we
aligned this data of parallel corpus along with its translation with the use of space tab (‘ ‘)
separation. Therefore, this aligned data is stored in text format to feed into systems. Then, we
shuffled the data elements before classifying into a training set and testing set. After shuffling the
data element, we have classified the data elements into a training set and testing. For the
classification, we followed one of the best techniques called Pareto principle (80/20) in which 80
percent of total data is used for the training set, while 20 percent of the total dataset is left for
testing the system [41]. Thus, the training set consists of 7424 sentences of parallel corpus, while
the test set consists of 1856 sentences. We used a built-in function train_test_split(amharic_tensor,
wola_tensor, test_size=0.2) to split the total dataset into training and testing set.

5.3 System environment/tools used for development

To implement our system, choosing a programming language and preparing the required
environment is necessary. A programming language chosen to implement the system is a python

68 | P a g e
programming language. Python programming language supports a set of a freely available library
in the deep learning algorithm. We used Keras library, TensorFlow library and NumPy library
which is freely available. We have chosen colab as a baseline for implementation. Colab is the
only tool which supports graphical processing unit (GPU) which speeds up the training time
hundreds time faster than central processing unit (CPU).

5.4 Parameter Optimization and Training the Experimental

Systems
In our experimental study, we have trained and tested the proposed system on our parallel corpus
of 9280 sentences in two columns (one for Amharic and other for Wolaita) separated by space tab
as we provided a sample in Appendix I. To get the desired result, we have done different
experiments on different issues to adjust the parameters of the model. The first issue is the selection
of neuron units to the number of dense layers and batch sizes. For training and testing our dataset,
we can select one of 16, 32, 64, 128, 256, 512, 1024, 2048 neuron units with batch sizes based on
the number of data size we are using. Batch size is several samples processed before the model is
updated. Thus, the size of a batch must be more than or equal to one and less than or equal to the
number of samples in the training dataset. For our 7424 training samples and with a batch size of
64, the algorithm takes the first 64 samples from the training dataset and trains the network. Next,
it takes the second 64 samples and trains the network again. We can keep doing this procedure
until we have propagated all samples through the network. For each epoch, it iterates for 116 times
as detailed in Appendix IV. The number of batches in each epoch is inversely proportional to batch
size. Training a system by minimum batch size results in less accurate estimate of the gradient and
training a system by maximum batch sizes requires more memory and even networks train slower.
Thus, we preferred 1024 neuron units for 2 dense layers with 64 batch sizes for our training dataset
since our corpus is medium size.

The second issue after setting neuron units and batch size is the selection of word embedding size.
For this purpose, we trained 1024 neuron units with 64 batch size in three different number of
word embedding sizes 64, 128 and 256 respectively for 50 epochs. From these different word
embedding sizes, we selected 1024 neuron units with 64 batch size and 256 word embedding
dimension for 50 epochs because it minimizes the loss level when compared with 1024 neuron
units with 128 batch size and 128 embedding dimension, 1024 neuron units with 256 batch size
69 | P a g e
and 256 embedding dimension and 1024 neuron units with 64 batch size and 128 embedding
dimension with the same number of epochs as shown in Figure 5.2. The number of epochs is the
number of complete passes (number of looking back into a single dataset element) through the
training dataset. The bigger size shows a sign of faster convergence at the beginning but starts to
diverge very soon. Therefore, we choose the medium size of word embedding as 256.

After adjusting neuron units, batch size and word embedding, choosing a learning rate for training
is important. In order to select optimum learning rate, we trained our system in 0.01 learning rate
which takes 63, 844 secs and its loss level is 0.017, 0.001 learning rate which takes 6,612 secs and
its loss level is 0.015 and 0.0001 learning rate which takes 25,684 secs and its loss level is 0.025.
As shown in Figure 5.3 and Figure 5.4, Adam’s optimizer (which defaults learning rate to 0.001)
minimizes a training data’s training time highly in optimum loss level in comparison to 0.01
learning rate and 0.0001 learning rate. Thus, we have initialized parameter by using optimizer =
tf.keras. optimizers.Adam(), which defaults learning rate to 0.001 and we selected sigmoidal
optimizer to optimize parameter. Therefore, the training of our architecture took 6,612 seconds (1
hour 50 minute and 12 secs). To our model, we have used batch training with a batch size of 64
and trained for 50 epochs on GPU based core i5 with 4GB RAM computer. As we have seen in
Figure 5.1 the loss level is simultaneously decreased until reaching to the 30th epoch and then it
decreased in rare. If we can use more dense layers with big dataset it may minimize the loss level
with less than that of 50 epochs.
2.5
2.264
2
1.827
Loss level

1.5
1.257
1
0.555
0.5
0.213 0.067
0.024 0.017 0.016 0.015 0.015
0
1 6 11 16 21 26 31 36 41 46 50
Number of Epochs

Figure 15: Loss Level Vs Number of Epochs

70 | P a g e
 2.5

1.5
Loss level

128 x 128
1 256 x 256
64 x 256

0.5

0
1 2 3 4 5 6 7 8 9 10

Number of Epochs

Figure 16: Loss Level for each Batch Size and Embedding Dimension for Number of Epochs

0.0175
0.017
0.017

0.0165
Loss Level

0.016

0.0155
0.015 0.015
0.015

0.0145

0.014
0.1 0.01 0.0001
Leaning Rate

Figure 17: Loss Level Vs Learning Rate for Training

71 | P a g e
 Figure 18: Time Taken Vs Learning Rate for Training

5.5 BLEU Evaluation Metrics

To evaluate the performance of any MT in terms of translation accuracy, various methods have
been developed. One of such metrics is bi-lingual evaluation understudy (BLEU) that is used to
measure translation accuracy by comparing the system’s translation output against human
translated reference sentences. A high-quality translation is the one which is closer to a
professional human translation and BLEU’s main idea is the measurements of this closeness.
BLEU score value falls in the range between 0 and 1. Most of the time it is reported in percent
from 0-100. The higher the BLEU score the more the translation resembles with the translation of
a human translation.

5.6 Experimental Results

In our experimental study, we have implemented the two designed system of our study as Amharic
to Wolaita NMT without attention and attention-based. Rather than adding attention mechanism
for Attention-based system, we trained and tested the systems in a similar way to get a correct
comparison of their performance. After training and testing the two systems, we have measured
72 | P a g e
the result of testing the system by using BLEU score metrics to see the difference in their scores.
In order to report the result of our testing, we have used the average BLEU score result of 50
epochs. For training our system in this epoch, attention-based system has taken 6,612 secs and
non-attention-based system has taken 21,253 secs for the same training dataset in 0.01 learning
rate.

When translating input sentence to output sentence, we use human translated reference sentence
to get BLEU score of the translation. We calculated the average BLEU score from the translation
as some of the translation is detailed in Appendix V. Therefore, this average result of Amharic-to-
Wolaita NMT system without attention in BLEU score is 0.5960, and attention-based Amharic-
to-Wolaita NMT system is 0.6258. Thus, as we have discussed in section 4.11 attention-based
system gives us +0.02978 BLEU gain. In addition to BLEU score improvement, the training time
for attention-based system is efficient than non-attention-based system.

5.7 Discussion on the Result of the Study

To accomplish the objective of this thesis work, we focused on the design and implementation of
MT from Amharic to Wolaita language. We conducted two experiments using two different
mechanisms, i.e., non-attention-based and attention-based approach with an expectation to get a
better result in both BLEU score and time efficiency. As the length of the sentence increase, the
inter-dependency of words at the beginning of the sentence and at the end of the sentence is loosely
related. This architecture works quite well for short sentences and this may make it difficult for
the neural network to cope with long sentences since the context within a sentence is derived as
the inter-dependency of nearby words in a given sequences words in the sentence.

The second problem of Encoder-Decoder model is how to handle a larger number of vocabulary
sizes available within the data. As each word in the sentence is visited, it must be assigned a new
identity number in order to identify a word by a unique index at the time it encountered in the data.
But when the length of the dictionary increases, the number used for word representation becomes
higher and the dimension of word vector needed becomes higher. These two basic issues are solved
by using an attention mechanism with a basic Encoder-Decoder architecture. To the best of our
knowledge during the time of this work, there has not been any other work exploring the use of
attention-based architectures for NMT.

73 | P a g e
One can observe that the result recorded from BLEU score shows the attention-based approach is
better than the non-attention-based approach for Amharic-to-Wolaita translation. But the results
recorded from both experiments are from less size of the corpus. As the size of corpus increases
the accuracy also increases and similarly the BLEU score result can also increase proportionally.

According to the result of our experiments, non-attention-based Amharic-to-Wolaita NMT has

shown interesting translation result as a BLEU score of 0.5960 with respect to our small datasets.
Moreover, we have got great improvement by using attention-based system which has shown
0.02978 BLEU improvement when compared with non-attention-based system. Also, when we
compared attention-based system with non-attention-based system in case of training time,
attention-based system shows better time efficiency. Also, attention-based system can also store
longer contextual relevancies of words found in longer sentence of datasets where we compare it
to the non-attention-based system. Therefore, the comparison of BLEU score metrics for some
range of word length shows that both of the systems have shown good translation result for
translating shorter sentences than longer sentences. This feature has been described in Appendix-
V, which compares both systems on the translation of the top 30 sentences from the parallel corpus.
Therefore, as we can see from Appendix I when the length of the word in source sentence
increases, the BLEU score level is decreases with an increase of sentence length for both attention-
based and non-attention-based models. The results of the experiment confirmed that the attention-
based model works better than the non-attention-based system with an increase of sentence length.
Moreover, the other interesting feature seen on using attention-based translation with RNN model
is that the RNN based model keeps contextual relatedness in which related words are used for
translation rather than only using exact target word for translation. In that case, the BLEU score
can be affected as it is dependent on the exact matching word of reference sentence rather than
using related dictionary words. Therefore, the result of our model has been affected by this
property of BLEU score metric and would have better than what is reported in Appendix V.

74 | P a g e
 Chapter 6: Conclusion and Future work
This chapter discusses the conclusion driven from this research work and the Future work for any
person or organization that interested to work on the machine translation between Amharic and
Wolaita language pair (or any other language pair) or related task.

6.1 Conclusion
As the objective of the work is to improve the MT between the language pair by implementing
RNN based architecture, two experiments were conducted to check the accuracy of the system
using two different approaches. The first experiment is conducted by using the non-attention-based
approach and it has a BLEU score of 0.5960. The second experiment is carried out by using an
attention-based approach and it has a BLEU score of 0.6258. Therefore, the results of our
experiments have shown that our attention-based system has shown better BLEU score in
translations, which is 0.02978 improvements for translation and uses less training time in
comparison with our non-attention-based system. More importantly the attention-based system
shows the great result in case of translating longer sentence than the non-attention-based system.
Also, the attention-based system is faster by comparing against the non-attention-based system.
We have also seen as using attention-based translation with RNN model is that the RNN based
model keeps contextual relatedness in which related words are used for translation rather than only
using exact target word for translation.

6.2 Future Works

✓ Since our work is the first MT system developed from the Amharic language to Wolaita
language pair as the translation is from Amharic to Wolaita language, we recommend
anyone who interested in this work for making this system bi-directional in future work in
order to easily distribute resources written in both languages.
✓ As the objective of our study is to implement MT only at the level of text to text translation,
we recommend anyone to develop speech to text first and use our work for further study.

75 | P a g e
✓ Also, we recommend any interested one in our work to include a higher number of the
dataset to further increase the translation quality since neural network needs more dataset
in order to perform better and also, we recommend to try with more layers and more hidden
units.
✓ For future we recommend any interested one to train our system for another language pairs.
✓ For future, we recommend one to use an autoencoder for the corpus separated by a space
tab (like ሁሉ በአግባብና በስረዓት ይሁን \t ubbabaikka wogaaninne maaran hano ). To do this,
train as an auto-encoder, save only the Encoder network and train a new Decoder for
translation from there.
✓ In order to make a system reliable, we recommend one to replace the position of untrained
(unknown) words by some special references like UNK.

76 | P a g e
 References
[1] https://www.omniglot.com/language/articles/MachineTranslation.htm last visited Oct 4,2019.
[2] Vani K, “Example-Based Machine Translation.” Internet: http://dspace.cusat.ac.in/jspui/
bitstream/3623/1/EBMTorginal.htm Nov 10, 2017 [last visited: Oct 4, 2019].
[3] Solomon Teferra, Martha Yifru, Michael Melese, Million Meshesha, Solomon Atinafu,
Yaregal Assabie, Biniyam Ephrem, Wondimagegnhue Tsegaye, Tsegaye Andargie,
Wondwossen Mulugeta, Hafte Abera, Tewodros Abebe, Amanuel Lemma, Seifedin Shifaw
“Parallel Corpora for bi-Directional SMT for Seven Ethiopian Language Pairs”, Proceedings
of the First Workshop on Linguistic Resources for NLP, pp. 83–90, Santa Fe, New Mexico,
USA, August 20, 2018.
[4] Eleni Teshome, "Bidirectional English-Amharic Machine Translation: using constrained
corpus", unpublished masters thesis, Addis Ababa University, Ethiopia, 2013.
[5] Sloculn, Jonathan, "A survey of Machine Translation: its history, current status, and future
prospects", 1985.
[6] Sisay Adugna. "English-Afaan Oromo Machine Translation: An Experiment using Statistical
Approach", unpublished masters thesis, Addis Ababa University, Ethiopia, 2009.
[7] Microsoft.com,“Microsoft Translator Blog: Microsoft Translator launching Neural Network
based translations for all its speech languages,” Online: https://www.microsoft.com/en-
us/translator/blog/2016/11/15/microsoft-translator-launching neural-network-based-
translations- for-all-its-speech-languages/. [last accessed: Oct. 12, 2019]
[8] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. “Neural Machine Translation by
Jointly Learning to Align and Translate”. Proceedings of the 3rd International Conference on
Learning Representations (ICLR), pp. 1–15. 2015.
[9] Sutskever, I., Vinyals, O., & Le, Q. V. “Sequence to sequence learning with neural networks”,
In Advances in neural information processing systems pp. 3104-3112, 2014.
[10] H. Luong, K. Cho, and Ch. Manning, “Neural Machine Translation” - Tutorial ACL 2016.

77 | P a g e
[11] Nal Kalchbrenner and Phil Blunsom.. Recurrent continuous translation models. In
Proceedings of the 2013 Conference on Empirical Methods in NLP. Association for
Computational Linguistics, Seattle, Washington, USA, pp. 1700–1709, 2013.
[12] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk,
and Yoshua Bengio. “Learning Phrase Representations using RNN Encoder-Decoder for
SMT,” In Proceedings of the Conf. on EMNLP, pp. 1724-1734, 2014.
[13] “Neural Machine Translation - Tut ACL” Oct, 2019, https://sites.google.com/site/acl16nmt/
[14] Lemma Lessa, "Development of Stemming Algorithm for Wolaita Text", unpublished
masters, Addis Ababa University, Ethiopia, 2003.
[15] “Constitution of the Federal Democratic Republic of Ethiopia”, www.wipo.int/edocs/ laws/en/
et/et007en.pdf.
[16] “NMT using keras” https://www.analyticsvidhya.com/blog/2019/01/neural-machine-
translation-keras/ last accessed: oct 6, 2019.
[17] Almahairi, A., Cho, K., & Habash, N, “First Result on Arabic to English NMT”, Vol. 1, 2016.
[18] Takahashi S., Wada H., Tadenuma R., and Watanabe S. “English-Japanese Machine
Translation”. Readings in Machine Translation. 2003
[19] Holger Schwenk, Jean-Baptiste Fouet, and Jean Senellart, “First Steps towards a general
purpose French/English Statistical Machine Translation System”, In Proceedings of the 3rd
workshop on statistical Machine Translation, pp. 119-122, Ohio, 2008.
[20] “Top ten most spoken language in the world”, listverse.com, last visited Oct 2, 2019
[21] Mulu Gebreegziabher Teshome and Laurent Besacier. "Preliminary experiments on English
Amharic Statistical Machine Translation." In SLTU, pp. 36-41. 2012.
[22] Michael Gasser, “Toward a Rule-Based System for English-Amharic Translation”, In Proc of
the 8th Int. 4th workshop on African LT, 2012.
[23] Jabesa Daba and Yaregal Assabie “A Hybrid Approach to the Development of Bidirectional
English-Oromiffa MT”, In Proceedings of the 9th Int. Conference on NLP (PolTAL2014),
Springer Lecture Notes in Artificial Intelligence (LNAI), Vol. 8686, pp. 228-235, Warsaw,
Poland, 2014.
[24] Akubazgi Gebremariam, “Amharic-Tigrigna machine translation using hybrid approach”,
unpublished masters thesis, Addis Ababa University, Ethiopia, 2017.

78 | P a g e
[25] Tadesse Kassa, “Morpheme-Based Bi-directional Ge’ez Amharic Machine Translation”,
unpublished masters thesis, Addis Ababa University, Ethiopia, 2018.
[26] Wakasa Motomich. “A descriptive study of the modern Wolaytta language”. Unpublished
PhD thesis, University of Tokyo, 2008.
[27] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, and Mohammad Norouzi, “Google’s
Neural Machine Translation System: Bridging the Gap between Human and Machine
Translation” Oct 8 2016.
[28] J. Hirschberg and C. D. Manning. “Advances in NLP”, Retrieved October 4, 2019, from
Science Mag: http://science.sciencemag.org/, Sept 12 2019.
[29] Yitayew Solomon and Million Meshesha, " Bi-directional Afaan Oromo-English Statistical
Machine Translation", unpublished masters thesis, Addis Ababa University, Ethiopia, 2017.
[30] Rae Steinbach, “Neural machine translation: Now and into the future”
https://www.banklesstimes.com/2018/02/24/NMT-now-future/ last accesed: Oct 7 2019.
[31] M. Utiyama, D. Do, and E. Sumita, “Machine translation from Japanese and French to
Vietnamese, the difference among language families,” October 2015.
[32] Teshome Kassie.“Word Sense Disambiguation for Amharic Text Retrieval”, MSc thesis,
Addis Ababa University, Ethiopia, 2018.
[33] Mulubrhan Hailegebreal, “A Bidirectional Tigrigna – English SMT”, unpublished masters
thesis, Addis Ababa University, Ethiopia, 2017.
[34] Habtamu Fanta Alambo, “Speaker Dependent Speech Recognition for Wolaytta Language”,
unpublished masters thesis, Addis Ababa University, Ethiopia, 2010.
[35] Biruk Abel, “Geez to Amharic Machine Translation using Hybrid approach”, unpublished
masters thesis, Addis Ababa University, Ethiopia, 2018.
[36] Atelach Alemu Argaw, “Amharic Stemmer: Reducing Words to their Citation Forms”.,
Proceedings of the 5th Workshop on Important Unresolved Matters, pp. 104–110, 2007.
[37] Samuel Eyassu,. “Classifying Amharic News Text Using Self Organizing Maps”. Proceedings
of the ACL Workshop on Computational Approaches to Semitic Languages, pp. 71–78, Ann
Arbor. @2005 Association for Computational Linguistics. 2005.
[38] Bethelhem Mengistu. N-gram-Based Automatic Indexing for Amharic Text. unpublished
masters thesis, Addis Ababa University, Ethiopia. 2002.

79 | P a g e
[39] Gochel, Daniel. “An integrated approach to automatic complex sentence parsing or Amharic
text”. Diss. Addis Ababa University, 2003
[40] Tigist Tensou, “Word Sequence Prediction for Amharic Language” unpublished masters
thesis, Addis Ababa University, Ethiopia, 2014.
[41] Nega Alemayehu and Peter Willett, “Stemming of Amharic Words for Information
Retrieval”, Literary and Linguistic computing, 17(1): 1-17, 2002.
[42] ጌታሁን አማረ፣ የአማርኛ ሰዋስው በቀላል አቀራረብ, 1989.

[43] ባዬ ይማም፣ የአማርኛ ሰዋስው, Addis Ababa, Ethiopia: EMPDA Publications, 1995.

[44] ጌታሁን አማረ፣ ዘመናዊ የአማርኛ ሰዋስው በቀላል አቀራረብ, Addis Ababa, Ethiopia, 2010.
[45] Dawit Mulugeta, “Geez to Amharic Automatic Machine Translation: A Statistical Approach”,
unpublished masters thesis, Addis Ababa University, Ethiopia, 2015.
[46] Abeba Ibrahim, “A hybrid approach to Amharic base phrase chunking and parsing”,
unpublished masters thesis, Addis Ababa University, Ethiopia, 2013.
[47] Michael Gasser, Hornmorpho User's Guide, 2012.
[48] Berhanu Herano Ganta, “Part of Speech Tagging for Wolaita Language”, unpublished masters
thesis, Addis Ababa University, Ethiopia, June 2015.
[49] Lamberti, Marcello and Roberto Sottile, “The Wolaytta Language”, In Studia linguarum
Africae Orientalis, pp. 79–86. Cologne: Rüdiger Köppe, 1997.
[50] Demewoz Beldados, “Automatic Thesaurus Construction from Wolaytta Text”, unpublished
masters thesis, Addis Ababa University, Ethiopia, 2013.
[51] Li Peng, “A Survey of Machine Translation Methods”, 2013, http://www.iaesjournal.com/
online/index.php/telkomnika/article/viewFile/2780/ [ last accessed: Dec 30, 2019].
[52] M. D. Okpor , "Machine Translation Approaches: Issues and Challenges," IJCSI International
Journal of Computer Science Issues, Vol. 11, No. 5, pp. 159-165, 2014.
[53] Sainik Kumar Mahata, Dipankar Das, and Sivaji Bandyopadhyay “MTIL2017: Machine
Translation Using RNN on Statistical Machine Translation” Journal of Intelligent Systems ·
May 2018.
[54] Adam Lopez,“Statistical Machine Translation”, ACM Computing Surveys, Vol. 40, Aug 2008.
[55] Mark Y. Liberman and Kenneth W., “Text analysis and word pronunciation in text to speech
synthesis”, Advances in speech signal processing, pp. 791-831, Dekker, New York,1992.

80 | P a g e
[56] Suhad M. Kadhem, Yasir R. Nasir, “English to Arabic Example-based Machine Translation
System” IJCCCE Vol.15, No.3 January 2015, DOI: 10.13140/RG.2.2.10922.88006.
[57] Daniel Jurafsky and James Martin, “An introduction to natural language processing,
computational linguistics, and speech recognition”, USA: Prentice-Hall Inc, 2000.
[58] Harjinder Kaur, and Vijay Laxmi, “A Survey of Mt Approaches”, 2013.
[59] Jaganadh G, “Man to Machine: A tutorial on the art of Machine Translation”, 2010,
http://www.slideshare.net/jaganadhg/a-tutorialon-machine-translation.
[60] Antony P. J., “Machine Translation Approaches and Survey for Indian Languages”, 2013,
http://www.aclclp.org.tw/clclp/v18n1/v18n1a3.
[61] M. D. Okpor, "Machine Translation Approaches: Issues and Challenges," IJCSI International
Journal of Computer Science Issues, Vol. 11, No. 5, pp. 159-165, 2014.
[62] Leon Bottou, Yoshua Bengio, and Yann Le Cun. “Global training of document processing
systems using graph transformer networks”. In Computer Vision and Pattern Recognition,
Proceedings, IEEE Computer Society Conference on, pp. 489-494, 1997.
[63] Alex Graves, Marcus Liwicki, Santiago Fernandez, Roman Bertolami, Horst Bunke, ´ &
Jurgen Schmidhuber. 2009. “A novel connectionist system for unconstrained handwriting
recognition”. Pattern Analysis and Machine Intelligence, IEEE Transactions.pp:855–868.
[64] Alex Graves, Santiago Fernandez, and Jurgen Schmidhuber. “Connectionist temporal
classification: labelling unsegmented sequence data with RNNs”. In Proceedings of the 23rd
international conference on Machine Learning, pp. 369–376, 2006.
[65] Jan K Chorowski, Dzmitry Bahdanau, Dmitriy Serdyuk, Kyunghyun Cho, and Yoshua Bengio.
“Attention-based models for speech recognition”. In Advances in Neural Information
Processing Systems, pp. 577–585, 2015.
[66] Yoshua Bengio and Rejean Ducharme, Pascal Vincent, and Christian Janvin. “A neural
probabilistic language model”. Pp. 1137–1155, 2003.
[67] Richard Socher, Eric H. Huang, Jeffrey Pennington, Andrew Y. Ng, and Christopher D.
Manning. “Dynamic pooling and unfolding recursive autoencoders for paraphrase detection”.
In Advances in Neural Information Processing Systems,. 2011.
[68] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeff Dean. “Distributed
representations of words and phrases and their compositionality”. In Advances in Neural
Information Processing Systems 26, pp. 3111–3119, 2013.
81 | P a g e
[69] A. Waibel, A. N. Jain, A. E. McNair, H. Saito, A.G. Hauptmann, and J. Tebelskis. Janus: “A
speech-to-speech translation system using connectionist and symbolic processing strategies”.
In Proceedings of the 1991 International Conference on Acoustics, Speech and Signal
Processing (ICASSP), pp. 793–796, 1991.
[70] Mikel L Forcada and Ramón P Ñeco. “Recursive hetero-associative memories for translation”. In
Biological and Artificial Computation: From Neuroscience to Tech., Springer, pp 453–462, 1997.
[71] M. Asunción Castaño, Francisco Casacuberta, and Enrique Vidal. “MT using NNs and finite-
state models”. pp. 160–167, 1997.
[72] Holger Schwenk. “Continuous space LMs”. Computer Speech and Language 3(21):492 518.
https://wiki.inf.ed.ac.uk/twiki/pub/CSTR/ListenSemester2_2009_10/sdarticle.pdf, 2007.
[73] Holger Schwenk. “Continuous space translation models for phrase-based SMT”. In
Proceedings of COLING 2012: Posters. The COLING 2012 Organizing Committee, Mumbai,
India, pp. 1071–1080, 2012.
[74] Shixiang Lu, Zhenbiao Chen, and Bo Xu. “Learning new semi-supervised deep auto encoder
features for SMT”. In Proceedings of the 52nd Annual Meeting of the Association for
Computational Linguistics. Association for Computational Linguistics, Baltimore, Maryland,
pp. 122–132, 2014.
[75] Shin Kanouchi, Katsuhito Sudoh, and Mamoru Komachi. “Neural reordering model
considering phrase translation and word alignment for phrase-based translation”. In
Proceedings of the 3rd Workshop on Asian Translation. The COLING 2016 Organizing
Committee, Osaka, Japan, pp 94–103, 2016.
[76] Peng Li, Yang Liu, Maosong Sun, Tatsuya Izuha, and Dakun Zhang. “A neural reordering
model for phrase-based translation”. In Proceedings of COLING 2014, the 25th International
Conference on Computational Linguistics: Technical Papers. Dublin City University and
Association for Computational Linguistics, Dublin, Ireland, pp 1897-1907, 2014.
[77] Adrià de Gispert, Gonzalo Iglesias, and Bill Byrne. “Fast and accurate preordering for SMT
using neural networks”. In Proc of the 2015 Conference of the North American Computational
Linguistics: Human Language Technologies. Association for Computational Linguistics,
Denver, Colorado, pp 1012–1017, 2015.
[78] Jacob Devlin, Rabih Zbib, Zhongqiang Huang, Thomas Lamar, Richard Schwartz, and John
Makhoul. “Fast and robust neural network joint models for statistical machine translation”. In
82 | P a g e
 Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics.
Association for Computational Linguistics, Baltimore, Maryland, pp. 1370–1380, 2014.
[79] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. “On the
properties of NMT: Encoder–decoder approaches”. Association for Computational
Linguistics, Doha, Qatar, pp. 103–111. 2014.
[80] Sébastien Jean, Orhan Firat, Kyunghyun Cho, Roland Memisevic, and Yoshua Bengio. “NMT
systems for wmt. In ˇ Proc of the 10th Workshop on SMT. Association for Computational
Linguistics, Portugal, 2015.
[81] Himanshu Choudhary, Aditya Kumar Pathak, Rajiv Ratn Shah, and Ponnurangam
Kumaraguru, “Neural Machine Translation for English-Tamil” 2019.
[82] Yukio Matsumura, Takayuki Sato, and Mamoru Komachi, “English-Japanese Neural Machine
Translation with Encoder-Decoder-Reconstructor” Tokyo, Japan, 2018.
[83] Zhaopeng Tu, Yang Liu, Lifeng Shang, Xiaohua Liu, and Hang Li.” Neural Machine
Translation with Reconstruction”. Proceedings of the 31st AAAI Conference on Artificial
Intelligence (AAAI), pp. 3097–3103, 2017.
[84] Mourad Brour, and Abderrahim Benabbou, “Arabic text language into Arabic sign language
neural machine translation”, Journal of King Saud University, 2019.
[85] Luong M., Sutskever I., and Zaremba W. “Addressing the rare word problem in NMT”. In
Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and
the 7th International Joint Conference on NLP, 2015.
[86] Buck C., Heafield K., and Van Ooyen B. “N-gram counts and language models from the
common crawl”. In LREC, Vol 2, pp. 4, 2014.
[87] H. Schwenk. http://www-lium.univlemans.fr/˜schwenk/cslm_joint_paper/ 2014. last acc: Jan
4 2020].
[88] Kate Reyes. “What is Deep Learning and How Does Deep Learning Work?” simplilearn.com
https://www.simplilearn.com/tutorials/deep-learning-tutorial/what-is-deep learning. Last acc:
Aug 12 2020.
[89] Utkarsh Dixit, “Neural Machine translation and the need for Attention Mechanism”
medium.com https://medium.com/analytics-vidhya/neural-machine-translation-and-the-need-
for-attention-mechanism-60f9a39da9a. Last acc: Aug 2020.

83 | P a g e
 [90] Jason Brownlee. “Encoder-Decoder Recurrent Neural Network Models for Neural Machine
Translation”. https://machinelearningmastery.com/Encoder-Decoder-recurrent-neural-
network-models-neural-machine-translation/. Last acc: Aug 2020.
[91] Jason Brownlee. “Encoder-Decoder Long Short-Term Memory Networks”. https://machine
learningmastery.com/Encoder-Decoder-long-short-term-memory-networks/ [acc Aug 2020].
[92] Adrian Rosebrock, “Keras vs. TensorFlow – Which one is better and which one should I
learn?” https://www.pyimagesearch.com/2018/10/08/keras-vs-tensorflow-which-one-is-
better-and-which-one-should-i-learn/. Last acc: Aug 2020.
[93] Ronald Rosenfeld. “Two decades of statistical language modelling: Where do we go from
here?”, Volume 88, pp. 1270–1278. 2000.
[94] Yoav Goldberg and Bar Ilan. “Neural Network Methods in Natural Language Processing
[lecture notes]. Available: https://www.morganclaypoolpublishers.com/catalog_Orig/
samples/9781627052955_sample.pdf.
[95] Andreas Stolcke. “SRILM – an extensible language modelling toolkit”. In ICSLP, 2002.
[96] Yee Teh. “A hierarchical Bayesian language model based on Pitman-Yor processes”. 2006.
[97] Marcello Federico, Nicola Bertoldi, and Mauro Cettolo. “IRSTLM: an open source toolkit for
handling large scale language models”. In Interspeech, 2008.
[98] Kenneth Heafield.” KenLM: faster and smaller language model queries”. In WMT. 2011.
[99] Rafal Jozefowicz, Oriol Vinyals and Mike Schuster. “ Exploring the Limits of Language
Modelling”, 2016.
[100] Frederic Morin and Yoshua Bengio. “Hierarchical probabilistic neural network language
model”. In AISTATS, 2005.
[101] Andriy Mnih and Geoffrey Hinton.“A scalable hierarchical distributed language model”. 2009.
[102] Andriy Mnih and Yee Whye Teh. “A fast and simple algorithm for training neural probabilistic
language models”. In ICML. 2012.
[103] Minh-Thang Luong, Michael Kayser, and Christopher D. Manning. “Deep neural language
models for machine translation”. 2015a.
[104] Michael Auli, Michel Galley, Chris Quirk, and Geoffrey Zweig. “Joint language and
translation modelling with Recurrent Neural Networks”. 2013.
[105] Jacob Devlin, Richard Schwartz, and John Makhoul. “Fast and robust neural network joint
models for statistical machine translation”. 2014.
84 | P a g e
[106] Yoon Kim, Yacine Jernite, and Alexander M. Rush. “Character-Aware Neural Language
Model, 2015.
[107] T. Pawar, “Language modeling using Recurrent Neural Networks Part - 1” medium.com,
https://medium.com/praemineo/language-modelling-using-recurrent-neural-networks part-1-
427b165576c2. [Accessed Aug. 2, 2020].
[108] Tomas Mikolov, Stefan Kombrink, Lukas Burget, Jan Cernock, and Sanjeev Khudanpur..
“Extensions of Recurrent Neural Network language model”. 2011.
[109] Tom´aˇs Mikolov and Geoffrey Zweig. “Context dependent Recurrent Neural Network
language model”. 2012.
[110] H. Jaeger, M. Lukosevicius, D. Popovici and U. Siewert. “Optimization and applications of
echo state networks with leaky-integrator neurons”, pp 335–352. 2007.
[111] https://hackernoon.com/attention-mechanism
[112] Lecun, Y. Bengio, and G. Hinton, “Deep learning,” 2015.
[113] G. Genthial, L. Liu, B. Oshri, and K. Ranjan. Natural Language Processing with Deep
Learning [Lecture notes]. Available: https://web.stanford.edu/class/cs224n/ readings/cs224n-
2019-notes06-NMT_Seq2Seq_attention.pdf. 2019.
[114] Y. Liu, L. Ji, R. Huang, T. Ming, and C. Gao, “An attention-gated convolutional neural
network for sentence classification,” pp. 1–19.
[115] https://colab.research.google.com/drive
[116] Tessemma Mindaye Mengistu, "Design and Implementation of Amharic Search Engine" ,
AAU, Addis Ababa, 2007.
[117] An Intuitive Understanding of Word Embeddings: From Count Vectors to Word2Vec.
Retrieved from Analytics Vidhya: https://www.analyticsvidhya.com/blog/2017/06/word-
embeddings-count-word2veec/ Last Accessed: June 04 2020.

85 | P a g e
 Appendix I: Sample of parallel corpus
Amharic Wolaita
ነገር ግን ሁሉ በአግባብና በሥርዓት ይሁን። SHin ubbabaikka wogaaninne maaran hano.
በዓለም ምናልባት ቁጥር የሌለው የቋንቋ ዓይነት Ha sa7an daro qaalai de7ennan aggenna; qassi
ይኖራል ቋንቋም የሌለው ሕዝብ የለም፤ birshshetti bainna qaali issoinne baawa.
ምድርም ሁሉ በአንድ ቋንቋና በአንድ ንግግር kase biittiya hanna issi qaalaninne issi haasayan daasu.
ነበረች።
በበሩ የሚገባ ግን የበጎች እረኛ ነው። SHin penggeera geliyaagee dorssaa heemmiya asa.
አንተማ መልካም ታመሰግናለህ፥ ሌላው ግን Neeni Xoossaa lo77o galataasa; shin hegee hinkko
አይታነጽበትም። bitaniyaa maaddenna.
የኋለኛው ጠላት የሚሻረው ሞት ነው፤ Ubbaappe wurssettan xayana morkkee haiquwaa.
ንቁ፥ በሃይማኖት ቁሙ፥ ጎልምሱ ጠንክሩ። Tishshi giite. Ammanuwan minnite. Xalite. Minnite.
በእናንተ ዘንድ ሁሉ በፍቅር ይሁን። Ooso ubbaa siiquwan oottite.
ማንም የማያውቅ ቢኖር ግን አይወቅ። SHin ooninne akeekana xayikko, akeekoppo, aggi bayo.
እኔ ጳውሎስ ይህን ሰላምታ በገዛ እጄ ጽፌአለሁ። Taani PHauloosi ha sarotaa ta kushiyan xaafaas.
እንግዲህ እኔን የምትመስሉ ሁኑ ብዬ Simmi tanadan hanite yaagada inttena woossais.
እለምናችኋለሁ።
በዋጋ ተገዝታችኋል፤ የሰው ባሪያዎች አትሁኑ። Xoossai inttena waagan shammiis. Asa aille gidoppite.
ጌታም አለ። ዓመፀኛው ዳኛ ያለውን ስሙ። Qassikka Godai, “He makkala daannai giidoogaa siyite.
ይህን ብታውቁ፥ ብታደርጉትም ብፁዓን ናችሁ። Intte hegaa eridi oottiyaabaa gidikko anjjettidaageeta.
ወደ ቤትም ስትገቡ ሰላምታ ስጡ፤ Qassi soo geliiddi sarotite.
ሁሉንም ተወ፤ ተነሥቶም ተከተለው። Yaagin Leewi ubbabaa aggi bayidi, denddi eqqidi, a
kaalliis.
ከሁሉም በኋላ እንደ ጭንጋፍ ለምሆን ለእኔ Qassi ubbaappe wurssettan, wodee gakkennan yelettida
ደግሞ ታየኝ። taassikka qoncciis.
ወንድሞች ሆይ፥ በአእምሮ ሕፃናት አትሁኑ፤ Ta ishatoo, intte iita yohuwan yiira mala gidanaappe
ለክፋት ነገር ሕፃናት ሁኑ እንጂ በአእምሮ የበሰሉ attin, qofan na7a mala gidoppite; intte wozanaa qofan
ሁኑ። wozannaamata gidite.
እንዲያማ ካልሆነ፥ አንተ በመንፈስ ብትባርክ Hegaa gidana xayikko, neeni ne ayyaanaa xalaalan
ባልተማሩት ስፍራ የተቀመጠው የምትለውን Xoossaa galatiyaabaa gidikko, he sohuwan neeni
ካላወቀ እንዴት አድርጎ ለምስጋናህ አሜን giyoogaa erenna asi neeni Xoossaa galatiyoogaa
ይላል? erennan, neeni galatidoogan woigidi, “Amin77i”
gaanee?
በዚች ሕይወት ብቻ ክርስቶስን ተስፋ ያደረግን Nuuni ha de7uwaa xalaalaassi Kiristtoosa yainnidabaa
ከሆነ፥ ከሰው ሁሉ ይልቅ ምስኪኖች ነን። gidikko, ha sa7an de7iya asa ubbaappekka nubai pala.
ፍቅርን ተከታተሉ፥ መንፈሳዊ ስጦታንም Siiquwaa kaallite. Ayyaanaa imuwaakka minttidi
ይልቁንም ትንቢት መናገርን በብርቱ ፈልጉ። koyite. SHin ubbaappe aattidi, hananabaa
yootiyoogaakka koyite.
መጽሐፍም እንደሚል በሦስተኛው ቀን ተነሣ፥ Haiqqidi moogettiis. Xoossaa maxaafai
giyoogaadankka, heezzantto gallassi haiquwaappe
denddiis.
ለኬፋም ታየ በኋላም ለአሥራ ሁለቱ፤ Denddidi, PHeexiroosassi qoncciis. Hegaappe
guyyiyan, ba kiittido tammanne naa77atussi qoncciis.

86 | P a g e
ከዚያም በኋላ ከአምስት መቶ ለሚበዙ Hegaappe guyyiyan, ichchashu xeetaappe dariya
ወንድሞች በአንድ ጊዜ ታየ፤ ከእነርሱም ishatussi issitoo qoncciis. Etappe dariya baggai hanno
የሚበዙቱ እስከ አሁን አሉ አንዳንዶች ግን gakkanaassikka paxa de7ees; shin amaridaageeti
አንቀላፍተዋል፤ haiqqidosona.
ከዚያም በኋላ ለያዕቆብ ኋላም ለሐዋርያት ሁሉ Hegaappe guyyiyan, Yaaqoobassi qoncciis; guyyeppe
ታየ፤ ba kiittido ubbatussi qoncciis.
እኔ ከሐዋርያት ሁሉ የማንስ ነኝና፥ Aissi giikko, Yesuusi kiittido ubbaappe taani laafa; taani
የእግዚአብሔርን ቤተ ክርስቲያን ስላሳደድሁ Xoossaa woosa keettaa waissido gishshau, Yesuusi
ሐዋርያ ተብዬ ልጠራ የማይገባኝ፤ kiittidoogaa geetettada xeesettanaukka bessikke.
ነገር ግን በእግዚአብሔር ጸጋ የሆንሁ እኔ ነኝ፤ SHin Xoossaa aaro kehatettan taani tanakka. Qassi
ለእኔም የተሰጠኝ ጸጋው ከንቱ አልነበረም taassi i immido aaro kehatettai hada gidibeenna; shin
ከሁላቸው ይልቅ ግን ደከምሁ፥ ዳሩ ግን ከእኔ taani eta ubbaappe aattada oottaas. SHin tanaara de7iya
ጋር ያለው የእግዚአብሔር ጸጋ ነው እንጂ እኔ Xoossaa aaro kehatettai oottiisippe attin, tana gidikke.
አይደለሁም።
እንግዲህስ እኔ ብሆን እነርሱም ቢሆኑ እንዲሁ Simmi tana gidikkokka, woikko eta gidikkokka, nuuni
እንሰብካለን እንዲሁም አመናችሁ። ubbai yootiyoogee hagaa; inttekka hagaa ammanideta.
ክርስቶስ ከሙታን እንደ ተነሣ የሚሰበክ ከሆነ Kiristtoosi haiquwaappe denddidoogaa yootiyoobaa
ግን ከእናንተ አንዳንዶቹ። ትንሣኤ ሙታን የለምgidikko, intteppe issi issi asati yaatin woigidi,
እንዴት ይላሉ? “Haiqqidabai haiquwaappe denddenna” yaagiyoonaa?
አሁን ግን፥ ወንድሞች ሆይ፥ ወደ እናንተ መጥቼSHin ta ishatoo, ha77i taani inttekko baada, dumma
በልሳኖች ብናገር፥ በመግለጥ ወይም በእውቀት dumma qaalan haasayaidda, ajjuutan woi eratettan woi
ወይም በትንቢት ወይም በትምህርት hananabaa yootiyoogan woi tamaarissiyoogan inttessi
ካልነገርኋችሁ ምን እጠቅማችኋለሁ? yootana xayikko, taani inttena ai go77iyaanaa?
ነፍስ የሌለበት ነገር እንኳ ዋሽንትም ክራርም Harai atto, shemppoi bainnabai pulaale gidin woikko
ቢሆን ድምፅ ሲሰጥ የድምፁን ልዩነት ባይገልጥ diitta gidin giiriiddi, ba giirettaa dummatettaa erissana
በዋሽንት የሚነፋው ወይስ በክራር የሚመታው xayikko, pulaale punniyaakkonne, woi diitta
መዝሙር እንዴት ይታወቃል? diixxiyaakko, asi waatidi eranee?
ደግሞም መለከት የማይገለጥን ድምፅ ቢሰጥ Qassi malkkataa waasoi geeyidi erettana xayikko,
ለጦርነት ማን ይዘጋጃል? olaassi oonee giigettanai?
እንዲሁ እናንተ ደግሞ የተገለጠውን ቃል Hegaadankka, qassi dumma dumma qaalan intte
በአንደበት ባትናገሩ ሰዎች የምትናገሩትን እንዴት
haasayiyoogee qoncce gidennan xayikko, intte
አድርገው ያስተውሉታል? ለነፋስ የምትናገሩ haasayiyoogaa asi waatidi akeekanee? Aissi giikko,
ትሆናላችሁና። carkkuwaassi intte haasayeeta.
እንግዲህ የቋንቋውን ፍች ባላውቅ ለሚናገረው SHin issi asi haasayiyo qaalaa taani erana xayikko, he
እንግዳ እሆናለሁ፥ የሚናገረውም ለእኔ እንግዳ haasayiyaagee taassi imatta gidees; taanikka assi imatta
ይሆናል። gidais.
እንዲሁ ደግሞ እናንተ መንፈሳዊ ስጦታን በብርቱHegaadankka, qassi Geeshsha Ayyaanaa imuwaa intte
የምትፈልጉ ከሆናችሁ ቤተ ክርስቲያንን ለማነጽ ekkanau minttidi koyiyaageeta gidiyo gishshau, woosa
እንዲበዛላችሁ ፈልጉ። keettaa dichchanau maaddiyaabaa ubbabaappe aattidi
koyite.
ነገር ግን ሌሎችን ደግሞ አስተምር ዘንድ SHin taani woosa keettan harata tamaarissanau dumma
በማኅበር እልፍ ቃላት በልሳን ከመናገር ይልቅ dumma qaalan tammu sha7u qaalata haasayiyoogaappe
አምስት ቃላት በአእምሮዬ ልናገር እወዳለሁ። ichchashu qaalata ta wozanaa qofan haasayanau dosais.
ነቢያትም ሁለት ወይም ሦስት ሆነው ይናገሩ Hananabaa yootiyaageetikka naa77u gididi woi heezzu
ሌሎችም ይለዩአቸው፤ gididi haasayona. Qassi harati eti giyoobaa pirddona.

87 | P a g e
 Appendix II: Common Amharic stop words.
በኋላ እኔ ታወሰ ይፋ ተፈጸመ እንግዲህ ወጣ ምን
ናት እኛ ዛሬ መሰረት ነገሰ ጥቃት ጊዜ ፈጠረ
ይናገራል እነሱ ልዩ ታወቀ የሚባሉ አስተዋወ ተበረከተ እርግጥ
ያንጻል እሱ ይሆናሉ አስታወቀ ንዴት አስፈጸመ እንዲያ ሄደ
እንዲሁ እሷ ተናገረ ተስማማ አነጋገር ልብ ሰነባበተ ተካሄደ
ግን አንተ ተባለ ወሰነ ተወሰነ ተቆጠበ ላከ ይሄን
ደግሞ እናንተ ርእይ ጠየቀ ተነጋገረ ሆይ ናደ አቀፈፈ
ሁሉም እና ዋለ ተጠየቀ አረጋገጠ ታመነ አወሰተ ንዴት
ይበል ወይ አመነ ተጠናቀ ሰበር ዘለለ አይስተዋ ተጫወተ
ማንም በላይ ተሳተፈ አገኘ ጠበቀ ተመለሰ ል አላግባብ
ነገርግን ወዘተ ተጋለጸ ምንም አሸነፈ አክዋያ ተጻፈ ወይ
ዘንድ አቶ ተናገርኸ ወይንም ከለከለ ተቋረጠ ወዲያው ጠቈመ
እላለሁ ተለያየ ተለመደ አይደለም ወሰደ አስተዋወ ን ተጠቈመ
ነው እስከ ቀረ ያህል ረገድ ተለየ ተሰማ ጠቀሰ
ፊት ግን ተደረገ ነኝ ተጫወተ አነሰ ኤፍቢሲ ተካተተ
አለ መጣ ተጠቀመ ተጠቃ አሁን ነውና ነን ውስጥ
ሆነ አመጣ ተያያዘ ተሻለ አሳሰበ አኳኋን ፖሊሲ ማረተ
በሇላ ወይም ለውጥ ደረገ ቀጠለ ወጥ ተገቢ ጀመረ
ስለ እንደ ተቆጠረ ተረጋገጠ ተመረጠ ምክንያት ራስ አስፈጻ
እንጂ ሁኔታው አደረ ሁልጊዜ ሀላፊነት አሳሰበ ገባ ሚ
ደገመ የት በአል ተናግሮብ ወዲ ተደነገገ አጋለጠ ደነገገ
ከ መቼ ተነገረ ታየ ምንድነ በአል ጨመረ ነች
በ ኋላ ተፈጠረ ሞከረ ምላሽ ቻለ ቀደም እንኳን
ናቸው ሁኔታ እነ ተረዳ ራስ ተቻለ ሳልቫ አወጣ
ጥቂት ሁል ተገመተ መጠን ተለየ አስቻለ ተደመጠ ወይስ
በርካታ ቢቢሲ ለም ቅርብ ቆየ ሰራ አካል እንግዲህ
ብቻ ብዛት ዋደደ እዚህ ተወገደ እንኳ ምንድን ወዲያ
ሌላ ቦታ ነገረ ነህ አጠቃላ ናችሁ ያሉት ይልዋል
ሁሉ በጣም ተወሰደ ተከላከለ ይ ቀረበ አስመዘገ ዳረገ
አንዳንድ በተለይ ተከሰተ አነሳ ወደ አቀረበ ተቀመጠ ፈለገ
ማን ተመለከ ጠቀመ አስቸኳይ ወደቀ ተናጋረ እንዴት አን
ባክዎ ተመሳሰለ አሳየ ደረሰ አስቸገረ ጭምር በለጠ በኩል
ባክ ተገልጹል ወጣ ዘነጋ ጠቀሰ ቴአትር አልያ ከበደ
ተጨማሪ ችግር ገናማ አስከተለ አስቀመ አስመለከ ተጠቀሰ ሙሉ
ሰአት አስታወሰ ድረግ መልክ ጠ ድረ ሄዱከ ደላ
ዉጪ አሳሳበ ይሄ ፈጸመ ድጋሚ አደረገ ግልጽ ፓርቲ
ናት ስፈላጊ ይህ ያዘ ተከናወነ ዋነኛ ተደመጠ ደነገገ
ያ አስገነዘበ ይህንን አመጥ ኢጋድ አንድ ተጠረጠ መከረ
ወይዘሮ አበራራ ይኸ እነሆ ገለጸ አስወገደ ርእሰ ፌስ
ወይዘሪት አስረዳ ስነ ተከራከረ ተገለጸ ተሰጠ አወቀ ተገቢ
ታች አንጻር ተናገር ዘንድ ይልቅ ሰጠ ቶሎ ቃለ
ከተተ እንኳ አመለከተ እንዲ አከናወነ ፈለገ ላይ አቃጠለ
መካከል ገና እጅግ ውድ አከለ ከፊል ነበር ጠፋ
ሰሞን ወቅት ከዛ ላቀ አካተተ ማንኛው ጠራ አጠናቀቀ
ትናንት ዋና ተጠበቀ ተገኘ ተደጋጋ ኖረ አስቈጠረ ተነሳ
ትናንትና ወጭ ተከተለ ተራዘመ ተመከረ መካነ ወጣ ጋር
ሆነ ጋራ ቀነሰ አይ አመመ ተጠቀሰ አገለገለ መሰለ

88 | P a g e
 Appendix III: Each epoch’s loss level and time taken for
training the system.

Epoch 1 Loss 2.264173 Time 153.21986 sec Epoch 26 Loss 0.067203 Time 91.1177 sec
Epoch 2 Loss 2.111294 Time 90.3907 sec Epoch 27 Loss 0.051981 Time 91.0183 sec
Epoch 3 Loss 2.054009 Time 90.9183 sec Epoch 28 Loss 0.041225 Time 90.9578 sec
Epoch 4 Loss 1.997920 Time 90.9472 sec Epoch 29 Loss 0.033504 Time 91.0428 sec
Epoch 5 Loss 1.920717 Time 93.0655 sec Epoch 30 Loss 0.028677 Time 92.7335 sec
Epoch 6 Loss 1.827197 Time 91.3056 sec Epoch 31 Loss 0.024998 Time 91.2618 sec
Epoch 7 Loss 1.735980 Time 91.2038 sec Epoch 32 Loss 0.021703 Time 90.8723 sec
Epoch 8 Loss 1.633517 Time 91.3208 sec Epoch 33 Loss 0.019390 Time 91.0475 sec
Epoch 9 Loss 1.520991 Time 91.4875 sec Epoch 34 Loss 0.017768 Time 91.0090 sec
Epoch 10 Loss 1.395341 Time 93.1666 sec Epoch 35 Loss 0.016498 Time 92.5059 sec
Epoch 11 Loss 1.257370 Time 91.3188 sec Epoch 36 Loss 0.015730 Time 90.9708 sec
Epoch 12 Loss 1.109244 Time 91.3559 sec Epoch 37 Loss 0.014732 Time 91.0016 sec
Epoch 13 Loss 0.954131 Time 91.0642 sec Epoch 38 Loss 0.014117 Time 91.0047 sec
Epoch 14 Loss 0.803845 Time 91.2477 sec Epoch 39 Loss 0.014345 Time 90.9625 sec
Epoch 15 Loss 0.668424 Time 92.7495 sec Epoch 40 Loss 0.014517 Time 92.8075 sec
Epoch 16 Loss 0.555440 Time 91.3822 sec Epoch 41 Loss 0.014911 Time 91.0141 sec
Epoch 17 Loss 0.462380 Time 91.2049 sec Epoch 42 Loss 0.015411 Time 90.8811 sec
Epoch 18 Loss 0.383793 Time 91.1040 sec Epoch 43 Loss 0.016670 Time 90.8737 sec
Epoch 19 Loss 0.316827 Time 91.0404 sec Epoch 44 Loss 0.016608 Time 91.0245 sec
Epoch 20 Loss 0.260945 Time 92.8468 sec Epoch 45 Loss 0.016925 Time 92.8748 sec
Epoch 21 Loss 0.213393 Time 91.2266 sec Epoch 46 Loss 0.017245 Time 91.0994 sec
Epoch 22 Loss 0.171829 Time 90.9922 sec Epoch 47 Loss 0.017937 Time 91.0599 sec
Epoch 23 Loss 0.136427 Time 91.1438 sec Epoch 48 Loss 0.017741 Time 91.0709 sec
Epoch 24 Loss 0.108045 Time 91.2057 sec Epoch 49 Loss 0.016927 Time 90.8503 sec
Epoch 25 Loss 0.085063 Time 93.0783 sec Epoch 50 Loss 0.016383 Time 92.7557 sec

89 | P a g e
 Appendix IV: The last epoch results with loss level and a time
to taken for 116 batches.
Epoch 50 Batch 0 Loss 0.013544 Epoch 50 Batch 28 Loss 0.012309
Epoch 50 Batch 1 Loss 0.017937 Epoch 50 Batch 29 Loss 0.022938
Epoch 50 Batch 2 Loss 0.025752 Epoch 50 Batch 30 Loss 0.023369
Epoch 50 Batch 3 Loss 0.009433 Epoch 50 Batch 31 Loss 0.017606
Epoch 50 Batch 4 Loss 0.017541 Epoch 50 Batch 32 Loss 0.017034
Epoch 50 Batch 5 Loss 0.015115 Epoch 50 Batch 33 Loss 0.020166
Epoch 50 Batch 6 Loss 0.010091 Epoch 50 Batch 34 Loss 0.022465
Epoch 50 Batch 7 Loss 0.027269 Epoch 50 Batch 35 Loss 0.015764
Epoch 50 Batch 8 Loss 0.011607 Epoch 50 Batch 36 Loss 0.013880
Epoch 50 Batch 9 Loss 0.011041 Epoch 50 Batch 37 Loss 0.014561
Epoch 50 Batch 10 Loss 0.014244 Epoch 50 Batch 38 Loss 0.011305
Epoch 50 Batch 11 Loss 0.015845 Epoch 50 Batch 39 Loss 0.015690
Epoch 50 Batch 12 Loss 0.021646 Epoch 50 Batch 40 Loss 0.018793
Epoch 50 Batch 13 Loss 0.023589 Epoch 50 Batch 41 Loss 0.015226
Epoch 50 Batch 14 Loss 0.010957 Epoch 50 Batch 42 Loss 0.019245
Epoch 50 Batch 15 Loss 0.021193 Epoch 50 Batch 43 Loss 0.013860
Epoch 50 Batch 16 Loss 0.023150 Epoch 50 Batch 44 Loss 0.015987
Epoch 50 Batch 17 Loss 0.022084 Epoch 50 Batch 45 Loss 0.010803
Epoch 50 Batch 18 Loss 0.014431 Epoch 50 Batch 46 Loss 0.016218
Epoch 50 Batch 19 Loss 0.013112 Epoch 50 Batch 47 Loss 0.009233
Epoch 50 Batch 20 Loss 0.021191 Epoch 50 Batch 48 Loss 0.022518
Epoch 50 Batch 21 Loss 0.012547 Epoch 50 Batch 49 Loss 0.017496
Epoch 50 Batch 22 Loss 0.025012 Epoch 50 Batch 50 Loss 0.010595
Epoch 50 Batch 23 Loss 0.016545 Epoch 50 Batch 51 Loss 0.013958
Epoch 50 Batch 24 Loss 0.013147 Epoch 50 Batch 52 Loss 0.017162
Epoch 50 Batch 25 Loss 0.010812 Epoch 50 Batch 53 Loss 0.017755
Epoch 50 Batch 26 Loss 0.014996 Epoch 50 Batch 54 Loss 0.023228
Epoch 50 Batch 27 Loss 0.015847 Epoch 50 Batch 55 Loss 0.012986

90 | P a g e
Epoch 50 Batch 56 Loss 0.011647 Epoch 50 Batch 87 Loss 0.015663
Epoch 50 Batch 57 Loss 0.009399 Epoch 50 Batch 88 Loss 0.019641
Epoch 50 Batch 58 Loss 0.014679 Epoch 50 Batch 89 Loss 0.020596
Epoch 50 Batch 59 Loss 0.021417 Epoch 50 Batch 90 Loss 0.022625
Epoch 50 Batch 60 Loss 0.018727 Epoch 50 Batch 91 Loss 0.012869
Epoch 50 Batch 61 Loss 0.025010 Epoch 50 Batch 92 Loss 0.022783
Epoch 50 Batch 62 Loss 0.011620 Epoch 50 Batch 93 Loss 0.017098
Epoch 50 Batch 63 Loss 0.025125 Epoch 50 Batch 94 Loss 0.011240
Epoch 50 Batch 64 Loss 0.010695 Epoch 50 Batch 95 Loss 0.026242
Epoch 50 Batch 65 Loss 0.011838 Epoch 50 Batch 96 Loss 0.016545
Epoch 50 Batch 66 Loss 0.019401 Epoch 50 Batch 97 Loss 0.017733
Epoch 50 Batch 67 Loss 0.013548 Epoch 50 Batch 98 Loss 0.013144
Epoch 50 Batch 68 Loss 0.021070 Epoch 50 Batch 99 Loss 0.017378
Epoch 50 Batch 69 Loss 0.011381 Epoch 50 Batch 100 Loss 0.011456
Epoch 50 Batch 70 Loss 0.013061 Epoch 50 Batch 101 Loss 0.015737
Epoch 50 Batch 71 Loss 0.019269 Epoch 50 Batch 102 Loss 0.025300
Epoch 50 Batch 72 Loss 0.013836 Epoch 50 Batch 103 Loss 0.014221
Epoch 50 Batch 73 Loss 0.013661 Epoch 50 Batch 104 Loss 0.020778
Epoch 50 Batch 74 Loss 0.016355 Epoch 50 Batch 105 Loss 0.022121
Epoch 50 Batch 75 Loss 0.010690 Epoch 50 Batch 106 Loss 0.018971
Epoch 50 Batch 76 Loss 0.013935 Epoch 50 Batch 107 Loss 0.013128
Epoch 50 Batch 77 Loss 0.011561 Epoch 50 Batch 108 Loss 0.013268
Epoch 50 Batch 78 Loss 0.017078 Epoch 50 Batch 109 Loss 0.010633
Epoch 50 Batch 79 Loss 0.022206 Epoch 50 Batch 110 Loss 0.016226
Epoch 50 Batch 80 Loss 0.016161 Epoch 50 Batch 111 Loss 0.014788
Epoch 50 Batch 81 Loss 0.013388 Epoch 50 Batch 112 Loss 0.015013
Epoch 50 Batch 82 Loss 0.017339 Epoch 50 Batch 113 Loss 0.015009
Epoch 50 Batch 83 Loss 0.012918 Epoch 50 Batch 114 Loss 0.014251
Epoch 50 Batch 84 Loss 0.012886 Epoch 50 Batch 115 Loss 0.017589
Epoch 50 Batch 85 Loss 0.015998 Epoch 50 Loss 0.016383
Epoch 50 Batch 86 Loss 0.010652 Time 92.7558 sec
91 | P a g e
 Appendix V: Sample output
Translating the Input ነገር ግን ሁሉ በአግባብና በሥርዓት ይሁን
same sentence Reference SHin ubbabaikka wogaaninne maaran hano.
A2W NMT Translation shin ubbabaikka wogaaninne maaran hano.
BLEU score 0.665076
Attention-based Translation shin ubbabaikka wogaaninne maaran hano.
A2W NMT BLEU score 0.665076
Translating the Input በዓለም ምናልባት ቁጥር የሌለው የቋንቋ ዓይነት ይኖራል ቋንቋም የሌለው ሕዝብ
same sentence የለም፤'
Reference Ha sa7an daro qaalai de7ennan aggenna; qassi birshshetti
bainna qaali issoinne baawa.
A2W NMT Translation ha sa7an daro qaalai de7ennan aggenna qassi birshshetti bainna
qaali issoinne baawa.
BLEU score 0.578965
Attention-based Translation ha sa7an daro qaalai de7ennan aggenna qassi birshshetti bainna
A2W NMT qaali issoinne baawa.
BLEU score 0.575757
Translating the Input ምድርም ሁሉ በአንድ ቋንቋና በአንድ ንግግር ነበረች።
same sentence Reference kase biittiya hanna issi qaalaninne issi haasayan daasu.
A2W NMT Translation kase biittiya hanna issi qaalaninne issi haasayan daasu.
BLEU score 0.614788
Attention-based Translation kase biittiya hanna issi qaalaninne issi haasayan daasu.
A2W NMT BLEU score 0.614788
Translating the Input በበሩ የሚገባ ግን የበጎች እረኛ ነው።
same sentence Reference SHin penggeera geliyaagee dorssaa heemmiya asa.
A2W NMT Translation shin penggeera geliyaagee dorssaa heemmiya asa.
BLEU score 0.638943
Attention-based Translation shin penggeera geliyaagee dorssaa heemmiya asa.
A2W NMT BLEU score 0.638943
Translating the Input አንተማ መልካም ታመሰግናለህ፥ ሌላው ግን አይታነጽበትም።
same sentence Reference Neeni Xoossaa lo77o galataasa; shin hegee hinkko bitaniyaa
maaddenna.
A2W NMT Translation aissi giikko, eti ainne hanennan de7iiddi, he urai issi asati a bolli
misimaariyan xishettidi i haa yeddido gishshau, a bolli
pirddanau yiidoogaappe attin, issi asati a bolli misimaariyan
xishettidi i haa yeddido gishshau, a bolli pirddanau
yiidoogaappe attin, issi asati a bolli
BLEU score 0.379671
Attention-based Translation neeni xoossaa lo77o galataasa shin hegee hinkko bitaniyaa
A2W NMT maaddenna.
BLEU score 0.588566
Translating the Input የኋለኛው ጠላት የሚሻረው ሞት ነው፤
same sentence Reference Ubbaappe wurssettan xayana morkkee haiquwaa.
A2W NMT Translation ubbaappe wurssettan xayana morkkee haiquwaa.
92 | P a g e
 BLEU score 0.665438
Attention-based Translation ubbaappe wurssettan xayana morkkee haiquwaa.
A2W NMT BLEU score 0.665438
Translating the Input ንቁ፥ በሃይማኖት ቁሙ፥ ጎልምሱ ጠንክሩ።
same sentence Reference Tishshi giite. Ammanuwan minnite. Xalite. Minnite.
A2W NMT Translation tishshi giite. ammanuwan minnite.
BLEU score 0.688725
Attention-based Translation tishshi giite. ammanuwan minnite. xalite. minnite.
A2W NMT BLEU score 0.630365
Translating the Input በእናንተ ዘንድ ሁሉ በፍቅር ይሁን።
same sentence Reference Ooso ubbaa siiquwan oottite.
A2W NMT Translation ooso ubbaa siiquwan oottite.
BLEU score 0.712104
Attention-based Translation ooso ubbaa siiquwan oottite.
A2W NMT BLEU score 0.712104
Translating the Input ማንም የማያውቅ ቢኖር ግን አይወቅ።'
same sentence Reference SHin ooninne akeekana xayikko, akeekoppo, aggi bayo.
A2W NMT Translation shin ooninne akeekana xayikko, akeekoppo, aggi bayo.
BLEU score 0.606819
Attention-based Translation shin ooninne akeekana xayikko, akeekoppo, aggi bayo.
A2W NMT BLEU score 0.606819
Translating the Input እኔ ጳውሎስ ይህን ሰላምታ በገዛ እጄ ጽፌአለሁ።
same sentence Reference Taani PHauloosi ha sarotaa ta kushiyan xaafaas.
A2W NMT Translation taani phauloosi ha sarotaa ta kushiyan xaafaas.
BLEU score 0.655997
Attention-based Translation taani phauloosi ha sarotaa ta kushiyan xaafaas.
A2W NMT BLEU score 0.655997
Translating the Input እንግዲህ እኔን የምትመስሉ ሁኑ ብዬ እለምናችኋለሁ።
same sentence Reference Simmi tanadan hanite yaagada inttena woossais.
A2W NMT Translation taani intteyyo yootidoogee tuma intte nagaran haiqqanaagaa
intteyyo yootidoogee tuma intte nagaran haiqqanaagaa intteyyo
yootidoogee tuma intte nagaran haiqqanaagaa intteyyo
yootidoogee tuma intte nagaran haiqqanaagaa intteyyo
yootidoogee
BLEU score 0.343295
Attention-based Translation simmi tanadan hanite yaagada inttena woossais.
A2W NMT BLEU score 0.641936
Translating the Input በዋጋ ተገዝታችኋል፤ የሰው ባሪያዎች አትሁኑ።
same sentence Reference Xoossai inttena waagan shammiis. Asa aille gidoppite.
A2W NMT Translation xoossai inttena waagan shammiis. asa aille gidoppite.
BLEU score 0.622333
Attention-based Translation xoossai inttena waagan shammiis. asa aille gidoppite.
A2W NMT BLEU score 0.622333
Input ጌታም አለ። ዓመፀኛው ዳኛ ያለውን ስሙ።

93 | P a g e
Translating the Reference Qassikka Godai, “He makkala daannai giidoogaa siyite.
same sentence
A2W NMT Translation qassikka godai, he makkala daannai giidoogaa siyite.
BLEU score 0.624953
Attention-based Translation qassikka godai, he makkala daannai giidoogaa siyite.
A2W NMT BLEU score 0.624953
Translating the Input ይህን ብታውቁ፥ ብታደርጉትም ብፁዓን ናችሁ።
same sentence Reference Intte hegaa eridi oottiyaabaa gidikko anjjettidaageeta.
A2W NMT Translation intte hegaa eridi oottiyaabaa gidikko anjjettidaageeta.
BLEU score 0.617252
Attention-based Translation intte hegaa eridi oottiyaabaa gidikko anjjettidaageeta.
A2W NMT BLEU score 0.617252
Translating the Input ወደ ቤትም ስትገቡ ሰላምታ ስጡ፤
same sentence Reference Qassi soo geliiddi sarotite.
A2W NMT Translation qassi soo geliiddi sarotite.
BLEU score 0.731111
Attention-based Translation qassi soo geliiddi sarotite.
A2W NMT BLEU score 0.731111
Translating the Input ሁሉንም ተወ፤ ተነሥቶም ተከተለው።
same sentence Reference Yaagin Leewi ubbabaa aggi bayidi, denddi eqqidi, a kaalliis.
A2W NMT Translation yaagin leewi ubbabaa aggi bayidi, denddi eqqidi, a kaalliis.
BLEU score 0.587832
Attention-based Translation yaagin leewi ubbabaa aggi bayidi, denddi eqqidi, a kaalliis.
A2W NMT BLEU score 0.587832
Translating the Input ከሁሉም በኋላ እንደ ጭንጋፍ ለምሆን ለእኔ ደግሞ ታየኝ።
same sentence Reference Qassi ubbaappe wurssettan, wodee gakkennan yelettida
taassikka qoncciis.
A2W NMT Translation qassi ubbaappe wurssettan, wodee gakkennan yelettida
taassikka qoncciis.
BLEU score 0.596476
Attention-based Translation qassi ubbaappe wurssettan, wodee gakkennan yelettida
A2W NMT taassikka qoncciis.
BLEU score 0.596476
Translating the Input ወንድሞች ሆይ በአእምሮ ሕፃናት አትሁኑ ለክፋት ነገር ሕፃናት ሁኑ እንጂ
same sentence በአእምሮ የበሰሉ ሁኑ'
Reference Ta ishatoo, intte iita yohuwan yiira mala gidanaappe attin,
qofan na7a mala gidoppite; intte wozanaa qofan wozannaamata
gidite.
A2W NMT Translation ta ishatoo intte iita yohuwan yiira mala gidanaappe attin, qofan
naa mala gidoppite intte wozanaa qofan wozannaamata gidite
BLEU score 0.524634
Attention-based Translation ta ishatoo intte iita yohuwan yiira mala gidanaappe attin, qofan
A2W NMT na7a mala gidoppite intte wozanaa qofan wozannaamata gidite.
BLEU score 0.523645

94 | P a g e
Translating the Input እንዲያማ ካልሆነ፥ አንተ በመንፈስ ብትባርክ ባልተማሩት ስፍራ የተቀመጠው
same sentence የምትለውን ካላወቀ እንዴት አድርጎ ለምስጋናህ አሜን ይላል?
Reference Hegaa gidana xayikko, neeni ne ayyaanaa xalaalan Xoossaa
galatiyaabaa gidikko, he sohuwan neeni giyoogaa erenna asi
neeni Xoossaa galatiyoogaa erennan, neeni galatidoogan
woigidi, “Amin77i” gaanee?
A2W NMT Translation aissi giikko, xoossaa maxaafai attin, cashsha asa woi asaa, ai
oottiyoonaa? haiqqidaageeti a kiitanchchaa gelidi, ubbai
xoossaa kiitanchchaa gelidi, ubbai xoossaa kiitanchchaa gelidi,
ubbai xoossaa kiitanchchaa gelidi, ubbai xoossaa kiitanchchaa
gelidi, ubbai xoossaa kiitanchchaa gelidi, ubbai xoossaa
BLEU score 0.347871
Attention-based Translation hegaa gidana xayikko, neeni galatidoogan woigidi, amin77i
A2W NMT gaanee? <end>
BLEU score 0.577351
Translating the Input በዚች ሕይወት ብቻ ክርስቶስን ተስፋ ያደረግን ከሆነ፥ ከሰው ሁሉ ይልቅ
same sentence ምስኪኖች ነን።
Reference Nuuni ha de7uwaa xalaalaassi Kiristtoosa yainnidabaa gidikko,
ha sa7an de7iya asa ubbaappekka nubai pala.
A2W NMT Translation nuuni kiristtoosaara nuuyyo ashshiya xoossai koyiyoogaadan
gidenna.
BLEU score 0.566641
Attention-based Translation Nuuni ha de7uwaa xalaalaassi Kiristtoosa yainnidabaa gidikko,
A2W NMT ha sa7an de7iya asa ubbaappekka nubai pala.
BLEU score 0.606306
Translating the Input ፍቅርን ተከታተሉ፥ መንፈሳዊ ስጦታንም ይልቁንም ትንቢት መናገርን በብርቱ
same sentence ፈልጉ።'
Reference Siiquwaa kaallite. Ayyaanaa imuwaakka minttidi koyite. SHin
ubbaappe aattidi, hananabaa yootiyoogaakka koyite.
A2W NMT Translation xoossaa sunttaa gishshau, intte huuphiyan intte godaa yesuus
kiristtoosa sunttaa gishshau, intte huuphiyan intte godaa yesuus
kiristtoosa sunttaa gishshau, intte huuphiyan intte godaa yesuus
kiristtoosa sunttaa gishshau, intte huuphiyan intte godaa yesuus
kiristtoosa
BLEU score 0.362558
Attention-based Translation siiquwaa kaallite. ayyaanaa imuwaakka minttidi koyite. shin
A2W NMT ubbaappe aattidi, hananabaa yootiyoogaakka koyite.
BLEU score 0.526641
Translating the Input መጽሐፍም እንደሚል በሦስተኛው ቀን ተነሣ፥
same sentence Reference Haiqqidi moogettiis. Xoossaa maxaafai giyoogaadankka,
heezzantto gallassi haiquwaappe denddiis.
A2W NMT Translation haiqqidi moogettiis. xoossaa maxaafai giyoogaadankka,
heezzantto gallassi haiquwaappe denddiis.
BLEU score 0.545018

95 | P a g e
Attention-based Translation haiqqidi moogettiis. xoossaa maxaafai giyoogaadankka,
A2W NMT heezzantto gallassi haiquwaappe denddiis.
BLEU score 0.545018
Translating the Input ከዚያም በኋላ ከአምስት መቶ ለሚበዙ ወንድሞች በአንድ ጊዜ ታየ፤ ከእነርሱም
same sentence የሚበዙቱ እስከ አሁን አሉ አንዳንዶች ግን አንቀላፍተዋል፤
Reference Hegaappe guyyiyan, ichchashu xeetaappe dariya ishatussi
issitoo qoncciis. Etappe dariya baggai hanno gakkanaassikka
paxa de7ees; shin amaridaageeti haiqqidosona.
A2W NMT Translation hegaappe guyyiyan, ichchashu xeetaappe dariya ishatussi
issitoo qoncciis. etappe dariya baggai hanno gakkanaassikka
paxa deees shin amaridaageeti haiqqidosona.
BLEU score 0.495419
Attention-based Translation hegaappe guyyiyan, ichchashu xeetaappe dariya ishatussi
A2W NMT issitoo qoncciis. etappe dariya baggai hanno gakkanaassikka
paxa de7ees shin amaridaageeti haiqqidosona.
BLEU score 0.494676
Translating the Input ከዚያም በኋላ ለያዕቆብ ኋላም ለሐዋርያት ሁሉ ታየ፤
same sentence Reference Hegaappe guyyiyan, Yaaqoobassi qoncciis; guyyeppe ba
kiittido ubbatussi qoncciis.
A2W NMT Translation hegaappe guyyiyan, yaaqoobassi qoncciis guyyeppe ba kiittido
ubbatussi qoncciis.
BLEU score 0.567128
Attention-based Translation hegaappe guyyiyan, yaaqoobassi qoncciis guyyeppe ba kiittido
A2W NMT ubbatussi qoncciis.
BLEU score 0.567128
Translating the Input ነገር ግን በእግዚአብሔር ጸጋ የሆንሁ እኔ ነኝ፤ ለእኔም የተሰጠኝ ጸጋው ከንቱ
same sentence አልነበረም ከሁላቸው ይልቅ ግን ደከምሁ፥ ዳሩ ግን ከእኔ ጋር ያለው
የእግዚአብሔር ጸጋ ነው እንጂ እኔ አይደለሁም።
Reference SHin Xoossaa aaro kehatettan taani tanakka. Qassi taassi i
immido aaro kehatettai hada gidibeenna; shin taani eta
ubbaappe aattada oottaas. SHin tanaara de7iya Xoossaa aaro
kehatettai oottiisippe attin, tana gidikke.
A2W NMT Translation shin taani inttenaara de7iyo gishshau, taani ha77i taani ha77i
taani ha77i taani ha77i taani ha77i taani ha77i taani ha77i taani
ha77i taani ha77i taani ha77i taani ha77i
BLEU score 0.388549
Attention-based Translation shin xoossaa aaro kehatettan taani tanakka. qassi taassi i
A2W NMT immido aaro kehatettai oottiisippe attin, tana gidikke. <end>
BLEU score 0.536172
Translating the Input እንግዲህስ እኔ ብሆን እነርሱም ቢሆኑ እንዲሁ እንሰብካለን እንዲሁም አመናችሁ
same sentence Reference Simmi tana gidikkokka, woikko eta gidikkokka, nuuni ubbai
yootiyoogee hagaa; inttekka hagaa ammanideta.
A2W NMT Translation simmi tana gidikkokka woikko eta gidikkokka nuuni ubbai
yootiyoogee hagaa ammanideta.
BLEU score 0.488549

96 | P a g e
Attention-based Translation simmi tana gidikkokka woikko eta gidikkokka nuuni ubbai
A2W NMT yootiyoogee hagaa inttekka hagaa ammanideta.
BLEU score 0.531831
Translating the Input ክርስቶስ ከሙታን እንደ ተነሳ የሚሰበክ ከሆነ ግን ከእናንተ አንዳንዶቹ ትንሳኤ
same sentence ሙታን የለም እንዴት ይላሉ
Reference Kiristtoosi haiquwaappe denddidoogaa yootiyoobaa gidikko,
intteppe issi issi asati yaatin woigidi, “Haiqqidabai
haiquwaappe denddenna” yaagiyoonaa?
A2W NMT Translation nuuni eroos pirddettenna shin taani intteyyo odikke yaagiis.
BLEU score 0.421257
Attention-based Translation kiristtoosi haiquwaappe denddidoogaa yootiyoobaa gidikko
A2W NMT intteppe issi issi asati yaatin woigidi haiqqidabai haiquwaappe
denddenna yaagiyoonaa
BLEU score 0.621557
Translating the Input በትህትና ሁሉና በየዋህነት በትእግስትም እርስ በርሳችሁ በፍቅር ታገሱ
same sentence Reference leemisuwawu, nuuni intteyyo aayye7ana.
A2W NMT Translation leemisuwawu nuuni intteyyo aayye7ana.
BLEU score 0.668791
Attention-based Translation leemisuwawu nuuni intteyyo aayye7ana.
A2W NMT BLEU score 0.668791
Translating the Input አንድ ጌታ አንድ ሃይማኖት አንዲት ጥምቀት
same sentence Reference Issi Godai, issi ammanoinne issi xinqqatee de7ees.
A2W NMT Translation issi godai issi ammanoinne issi xinqqatee de7ees.
BLEU score 0.788549
Attention-based Translation issi godai issi ammanoinne issi xinqqatee de7ees.
A2W NMT BLEU score 0.788549
Translating the Input ከሁሉ በላይ የሚሆን በሁሉም የሚሠራ በሁሉም የሚኖር አንድ አምላክ የሁሉም
same sentence አባት አለ።
Reference Ubbaa Godai, ubbaa baggaara oottiyaagee, ubbaa giddon
de7iyaagee, issi Xoossai, ubbaa Aawai de7ees.
A2W NMT Translation ubbaa godai ubbaa giddon de7iyaagee issi xoossai ubbaa aawai
de7ees.
BLEU score 0.468549
Attention-based Translation ubbaa godai ubbaa baggaara oottiyaagee ubbaa giddon
A2W NMT de7iyaagee issi xoossai ubbaa aawai de7ees.
BLEU score 0.636172
Average BLEU score of A2W NMT is 0.5960
Average BLEU score of Attention-based A2W NMT is 0.6258

97 | P a g e
 Declaration
I, the undersigned, declare that this thesis is my original work and has not been presented for a
degree in any other university, and that all source of materials used for the thesis have been duly
acknowledged.

Declared by:

Name: _____Workineh Wogaso Gaga ______________________________________

Signature: ____________________________________________________________

Date: _______October 27, 2020 ___________________________________________

Confirmed by advisor:

Name: _____Yaregal Assabie (PhD) _______________________________________

Signature: ____________________________________________________________

Date: ________________________________________________________________

98 | P a g e