JIMMA UNIVERSITY

COLLEGE OF NATURAL SCIENCE

DEPARTMENT OF INFORMATION SCIENCE

TOWARDS THE SENSE DISAMBIGUATION OF AFAAN OROMOO

WORDS USING HYBRID APPROACH (UNSUPERVISED MACHINE

LEARNING AND RULE BASED)

By: WORKINEH TESEMA GUDISA

A Thesis Submitted to Jimma University Institute of Technology(JiT) in Partial Fulfillment

of the Requirements for the Degree of Master of Science in Information Technology

Advisor: Debela Tesfaye (Ass.Prof)

Co-Advisor: Teferi Kebebew (MSc)

October 3, 2015
 JIMMA UNIVERSITY

COLLEGE OF NATURAL SCIENCE

DEPARTMENT OF INFORMATION SCIENCE

TOWARDS THE SENSE DISAMBIGUATION OF AFAAN OROMOO

WORDS USING HYBRID APPROACH (UNSUPERVISED MACHINE

LEARNING AND RULE BASED)

Jimma, Ethiopia
October 3, 2015
Advisor Name: Debela Tesfaye(Assistant Professor) and Teferi Kebebew(Lecturer)
Rank: Ass. Pro
School/Department: Computing
Institute / University: Jimma University
City: Jimma
______________________________________________________________________________

Declaration and Certification

This is to certify that the thesis entitled “Towards the Sense Disambiguation of Afaan Oromoo
Words Using Hybrid Approach (Unsupervised Machine Learning and Rule Based)” submitted
by Workineh Tesema, the M.Sc (IT) student of School of Computing, Jimma Institute of
Technology, Jimma University, for the award of Master of Science in Information Technology, is a
record of original work carried out by him under my supervision and guidance. The thesis has
fulfilled all requirements as per the regulations of the University and in my opinion the thesis has
reached the standard needed for submission. The results embodied in the thesis have not been
submitted to any other University or Institute for the award of any degree or diploma. I hereby
declare that the student has incorporated the comments given during the mock defense to improve the work
substantially.

Debela Tesfaye (Assistance Professor) _______________ ______________

Advisor Signature Date
Teferi Kebebew (MSc) _______________ ______________
Co-Advisor Signature Date
______________________________ _______________ ______________
External Examiner Signature Date
______________________________ _______________ ______________
External Examiner Signature Date
______________________________ _______________ ______________
Institute coordinator for Research, P, GS&C Signature Date

Date: 12/01/2015
Place: Jimma University _________________________

_________________________

i
 Dedication

This thesis work is dedicated to my father Tesema Gudisa, my mother Askala Jifar, all my brothers

and sisters who were able to reap the fruit of their own.

Workineh Tesema Gudisa

October, 2015

ii
 Declaration

This study is my original work and has not been submitted as a partial requirement for a

degree in any University and that all sources of material used for the study have been duly

acknowledged.

Workineh Tesema Gudisa

October, 2015

iii
 Acknowledgment

First of all I would like to thank Jesus Christ who helped me to succeed in all my life long learn.
Next, I would like to thank my advisor Debela Tesfaye for his valuable assistance in providing his
genuine, professional advice and encouragement goes even beyond the accomplishment of this
study. Some noteworthy aspects of this study are entirely because of him; all faults are in spite of
him. He initiated me to do by giving precious comments on necessary points. My thanks go to him
again, since it is difficult to mention his contribution to my achievements in words, it is better to say
my heart has recorded it forever. And also I would like to acknowledge my co-advisor Teferi
Kebebew for all his guidance at every step of the thesis work, for patiently listening to me even
during uncountable hours, for advising me to do my study, for imparting so much valuable
knowledge and for all his encouragement and words of kindness. Next, Words cannot express my
deepest thanks to my Father Mr. Tesema Gudisa and mother Askala Jifar and my brothers Hirpa
Tesema, Muleta Tesema, Naol Tesema and Ibsa Tesema for their moral and financial support I
received from them throughout my life of learning. My honest thanks equally go to my sisters Sorse
Tesema, Lamme Tesema and Shonkore Tesema for their help in providing me concerning
background information. I would like to thanks Jimma University, Department of Afan Oromo
Linguistic professional Lecturers Tefera, Waktola, Chimdi, Lenin and all my colleagues who have
helped me on my progress; Chala Diriba and Worku Jimma.

Lastly but not least, I would like to thanks to ORTO and Oromia Folklore Bureau for their
permission to data collection and professional contribution and Jimma University for financial
support.

iv
 Table Content

Table of Contents Page

Declaration and Certification ............................................................................................................... i

Dedication ...........................................................................................................................................ii

Acknowledgement .............................................................................................................................. iv

List of Tables ...................................................................................................................................... ix

List of Figures ...................................................................................................................................... x

List of Appendix………………………………………………………………………………………………………………………………………xi

List of Acronyms ...............................................................................................................................xii

Abstract…………………………………………………………………………………………………………………………………………………..xiii

Chapter One ......................................................................................................................................... 1

Introduction .......................................................................................................................................... 1

1.1. Background .............................................................................................................................. 1

1.2. Statement Problem ................................................................................................................... 3

1.3. Objective of the Study ............................................................................................................. 5

1.3.1. General Objective .......................................................................................................... 5

1.3.2. Specific Objectives: ....................................................................................................... 5

1.4. Scope and Limitation of the Study........................................................................................... 5

1.5. Significance of the Study ......................................................................................................... 6

1.6. Organization of the Thesis Report ........................................................................................... 6

Chapter Two......................................................................................................................................... 7

2. Literature Review............................................................................................................................. 7

2.1. Introduction .............................................................................................................................. 7

2.2. Basic Concepts of Word Sense Disambiguation ..................................................................... 7

v
 2.3. Approaches to Word Sense Disambiguation ........................................................................... 8

2.3.1. Supervised Word Sense Disambiguation ....................................................................... 8

2.3.2. Unsupervised Word Sense Disambiguation................................................................... 9

2.3.3. Bootstrapping Machine Learning Approach ................................................................ 10

2.3.4. Corpus-Based Approaches ........................................................................................... 10

2.3.5. Knowledge-Based Approaches .................................................................................... 11

2.3.6. Hybrid Machine Learning Approach ........................................................................... 12

2.4. Context Window Size .............................................................................................................. 13

2.5. Vector Space Model ................................................................................................................. 13

2.6. Clustering Algorithms.............................................................................................................. 14

2.6.1. Hierarchical Agglomerative Algorithms....................................................................... 14

2.6.2. Partitional Algorithms ................................................................................................... 17

2.7. Divisive Clustering ................................................................................................................. 19

2.8. Rule Based Approach ............................................................................................................. 20

2.9. History of Word Sense Disambiguation (WSD) ..................................................................... 21

2.10. Related Works ....................................................................................................................... 23

Chapter Three..................................................................................................................................... 30

3. Overview of Afan Oromo ............................................................................................................. 30

3.1. Introduction ............................................................................................................................. 30

3.2. Language Structure ................................................................................................................. 31

3.3. Alphabet (Qubee Afaan Oromoo) ........................................................................................... 31

3.4. Sentence Structure .................................................................................................................. 31

3.5. Adjectives ............................................................................................................................... 32

3.6. Articles .................................................................................................................................... 32

3.7. Punctuation Marks .................................................................................................................. 32

vi
 3.8. Conjunctions ........................................................................................................................... 33

3.9. Word and Sentence Boundaries .............................................................................................. 33

3.10. Word Segmentation .............................................................................................................. 33

3.11. Homonymy ........................................................................................................................... 33

3.12. Synonym ............................................................................................................................... 34

3.13. Polysemy ............................................................................................................................... 34

Chapter Four ……………………………………………………………………………………... 36

4. Methodology .................................................................................................................................. 36

4.1. Introduction .......................................................................................................................... 36

4.2. Method ................................................................................................................................. 37

4.3. Corpus .................................................................................................................................. 37

4.3.1. Corpus Acquisitions and Preparations ....................................................................... 37

4.3.2. Corpus Preprocessing................................................................................................. 38

4.4. Implementation Tools .......................................................................................................... 40

4.5. Testing.................................................................................................................................. 40

4.6. Algorithms ........................................................................................................................... 42

4.6.1. Machine Learning Approach ..................................................................................... 42

4.6.1.1. Extracting the Context Words ....................................................................... 43

4.6.1.2. Clustering the Contexts ................................................................................. 43

4.6.1.3. Vector Space Model ...................................................................................... 44

4.6.1.4. Clustering Algorithms................................................................................... 45

4.7. Hybrid Approach ................................................................................................................. 45

4.7.1. Constructing Rule for Extracting Contexts ................................................................ 46

4.7.1.1. Modifiers ...................................................................................................... 46

4.7.1.2. Modifiers Rule ............................................................................................. 47

vii
 4.8. Evaluation Method ............................................................................................................... 48

Chapter Five ....................................................................................................................................... 50

5. Experimentation Result, Discussion and Evaluation ..................................................................... 50

5.1. Introduction ............................................................................................................................. 50

5.1.1. Data ............................................................................................................................... 51

5.2. Experiment on the Machine Learning Approach .................................................................... 51

5.2.1. Experiment on the Context Window Sizes ................................................................... 51

5.2.2. Experiment on the Clustering ....................................................................................... 53

5.2.3. Experimental Results .................................................................................................... 54

5.2.4. Summary of the Machine Learning Approach.............................................................. 56

5.3. Experiment on the Hybrid Approach ...................................................................................... 57

5.3.1. Effect of Hybrid Approach on Accuracy of Disambiguation ....................................... 58

5.3.2. Summary of the Hybrid Approach ................................................................................ 59

5.4. Discussion .............................................................................................................................. 61

Chapter Six......................................................................................................................................... 73

6. Conclusion and Recommendation ................................................................................................. 73

6.1. Conclusion .............................................................................................................................. 73

6.2. Recommendation .................................................................................................................... 75

References .......................................................................................................................................... 77

Appendix A ........................................................................................................................................ 84

viii
 List of Tables

Table 2.1.Summary of Word Sense Disambiguation Approaches………………………………….21

Table 4.1. Stop Words ……………………………………………………………………………...39

Table 4.2. List of Ambiguous Words for testing..…………………………………………………..41

Table 5.1. Determining Optimal Window Size…………….……………………………………….52

Table 5.2. Disambiguation using evidence from Machine Learning….…………………………....54

Table 5.3. The Accuracy of the Machine Learning…………………………………………………55

Table 5.4. Summary of Machine Learning Clustering Algorithm.………………………...…….…56

Table 5.5. Disambiguation using evidence from Hybrid Approach………………………………...59

Table 5.6. The Accuracy of the Hybrid Approach………………………………………………….60

Table 5.7. Summary of Hybrid Approach Clustering Algorithms………………………………….60

Table 5.8. Evaluation of WSD ………………….………….………………………………………63

Table 5.9. Extracted Context Using Window Size………………………………………………….65

Table 5.10. Cosine Similarity Measure Representations……………………………………….......66

Table 5.11. Extracted Modifiers for ambiguity word bahe …………………...……………………70

Table 5.12. Summary of the Test………………………………………………………..………….72

ix
 List of Figures

Figure 2.1. Surrounding Contexts of the term Afaan…………….…………………………………..8

Figure 2.2. a. Single-Link, b. Complete-Link and c) Average Link….………………………….....16

Figure 2.3. Hierarchical Clustering and Divisive Clustering………………………………….........20

Figure 4.1. The Architecture (pipe line) of the System……………………….…………………….36

Figure 5.1. User Interface of the WSD……………………………………………………………...50

Figure 5.2. Single-Link(left), Complete-Link(Middle) and Average-Link(right) Dendrogram……68

Figure 5.3. Expectation Maximization and K-Means Clustering…………………………………...69

Figure 5.4. EM (left side) and K-Means (right side)………..………………………….…………..71

x
 List of Appendix

Appendix A. Question For Collection of Ambiguous Words………………………………….84

xi
 List of Acronyms

AI: Artificial Intelligence

CSV: Comma-Separated Values

EM: Expect Maximization

HAC: Hierarchical Agglomerative Clustering

IR: Information Retrieval

KBD: Knowledge-Based Disambiguation

ML: Machine Learning

MT: Machine Translation

MAX: Maximum

MIN: Minimum

NLP: Natural Language Processing

ORTO: Oromia Radio and Television Organization (ORTO)

POS: Part-of-Speech

QA: Question Answering

VSM: Vector Space Model

WSD: Word Sense Disambiguation

xii
 Abstract
Word Sense Disambiguation is a technique in the field of Natural Language Processing where the
main task is to find the appropriate sense in which ambiguous word occurs in a particular context.
It is a fundamental problem for many natural language technology applications(Machine
Translation, Text Summarization, Question and Answering, Information extraction and text mining
and Information Retrieval). A word may have multiple senses and the problem is to find out which
particular sense is appropriate in a given context. Ambiguity is a cause of poor performance in
searching and retrieval system. The objective of this work is to develop hybrid word sense
disambiguation which finds the sense of words based on surrounding contexts. Hence, this study
presents a Word Sense Disambiguation strategy which combines an unsupervised approach that
exploits sense in a corpus and manually crafted rule. The idea behind the approach is to overcome
the problem a bottleneck for the machine learning approaches, while hybrid method can improve
the accuracy and suitable when there is scarcity of training data. This makes our approach suitable
for disambiguation when there is lack of resource and sense definitions. In this study, the context of
a given word is captured using term co-occurrences within a defined window size of words. The
optimal window sizes for extracting semantic contexts is window +1 and +2 words to the right and
left of the ambiguous word. The similar contexts of a given senses of ambiguous word are clustered
using hierarchical and partitional clustering. Each cluster representing a unique sense. Some
ambiguous words have two senses to the five senses. The result argued that WSD yields an
accuracy of 70% in Unsupervised Machine learning and 81.1% in Hybrid Approach. The machine
learning were a useful information source for disambiguation but that it not as robust as a
linguistic(rule based) [89]. Based on this, the integration of deep linguistic knowledge with
machine learning improves disambiguation accuracy. Therefore, for Afan Oromo semantic has
come to the conclusion that the sense of words are closely connected to the statistics of word usage.
The achieved result was encouraging, despite it is less resource requirement. Yet; further
experiments using different approaches that extend this work are needed for a better performance.

Key words: Word Sense Disambiguation, Afan Oromo, Ambiguous Word, Disambiguation, Rule Based, Hybrid.

xiii
 Chapter One
Introduction
1.1. Background

In today’s world, where World Wide Web technology is keeping on growing very fast, many users
go to the web to search for information, for entertainment, to read documents and electronic books.
Sometimes it is observed that the result of a search is not appropriate. The reason behind is, there is
an ambiguous word in the query. According to [1] in information retrieval, the stored information
items and the incoming search requests are normally represented by sets of content identifiers as
keywords, index terms, or simply terms. Information Retrieval (IR) can potentially benefit from the
correct meanings of words provided by Word Sense Disambiguation (WSD). Ambiguity is a cause
of poor performance in IR systems. In the application to IR, WSD can bring different benefits. The
queries may contain ambiguous words (terms), which have multiple meanings. The ambiguities of
these query words negatively affect the number of retrieved documents. Identifying the correct
meaning of the ambiguous words in both queries and documents can help improve retrieval of the
documents. The other one is, query words may have tightly related meanings with other words not in
the query. Making use of these relations between words can improve the number of retrieved
relevant documents. IR is one of the Natural Language Processing (NLP) applications that paybacks
from WSD which most of the words used to execute queries in IR systems have more than one
meaning [2].

In natural language processing, WSD is the manner of determining which sense (meaning) of an
ambiguous word is activated by the use of the ambiguous word in a particular context. It is a natural
language problem, a given ambiguous word and its possible senses as defined by an occurrence of
the ambiguous word in context into one or more of its sense [3]. Natural languages have ambiguous
words which need to be disambiguated and thus the appropriate sense of an ambiguous word in a
given context can be identified. An ambiguous word can take several senses depending on the
context in which it appears. The same form and pronunciation can take different meanings in
different contexts [4].

_______________________________
In this paper, ambiguous word and target word are used interchangeably

1
The ambiguity of human language is a greatly debated problem in many research areas. Human
language is ambiguous, because the ambiguous words can be interpreted in multiple ways depending
on the context in which it occurs [2].

It is considered as an open problem, that is a task whose solution is as hard as the most difficult
problems in natural language [5]. Natural languages are ambiguous, hence an ambiguous word can
take several senses depending on the context in which it appears. Consider the following simple
sentences:

 Afaan koo faaya kooti.

 Siifaan afaan bal’atti.

In the first context, the meaning of the word afaan means ‘the method of human communication or
humans language’. In the second context, it means ‘mouth or the opening and cavity in the lower
part of the face, Surrounded by the lips’. In this case, the word form (afaan) is the same in both
sentences. The word (afaan) sharing the same spelling and pronunciation, but have different senses.
This multiple sense can be totally unrelated to each other. The word (afaan) has more than one
senses, so it is too difficult to say the sense of (afaan) is language or mouth at a given moment. This
is what makes word sense disambiguation difficult because of the given ambiguous word has more
senses. However, the sense of any word was based on the contexts around the target word in
sentences, paragraph and text. It is the context which decides the sense of the ambiguous words
according to it appearing in the sentences [6].

The assignment of sense to word is accomplished by using major sources of information contained
within the context in which the word appears [2]. All disambiguation work uses the context of the
instance of the word to be disambiguated with either ruled based or information about the contexts of
previously disambiguated instances of the ambiguous word derived from corpora (corpus based
WSD). People decide the meaning of an ambiguous word based on the characteristic senses of a
discussion or situation using their own evidences. But machines have no ability to decide such an
ambiguous situation unless some procedures have been planted into the machines’ memory.

The task of building computer programs that understand natural language is not straightforward [7]
because, there are no decisive ways of identifying where one sense of an ambiguous word ends and
the next begins. In computational linguistics, WSD is an open problem of natural language

2
processing which governs the process of identifying which sense of an ambiguous word
(i.e. meaning) is used in a sentence, when the ambiguous word has multiple meanings [8]. This
problem impacts on other computer-related processes, such as improving relevance of search and
MT (Machine Translation). WSD as a heart of NLP because it is one of the most fundamental tasks
in many applications in NLP directly or indirectly rely on WSD [65].

The purpose of this study is to combine both the rule based and machine learning approaches into a
hybrid approach and to investigate word sense disambiguation for Afan Oromo words. In this case
the approach is that the combination of both types of strategies, namely unsupervised machine
learning and rule based can improve WSD effectiveness, because it overcomes the limitation of
algorithms by rule based (linguistic knowledge). The motivation behind WSD is Afan Oromo to
allow the users to make ample use of the available technologies because ambiguities present in any
language provide great difficulty in the use of information technology as words in human language
that occur in a particular context can be interpreted in more than one way depending on the context.

1.2. Statement Problem

Afan Oromo is an official language of Oromia Regional State (which is the largest regional State
among the current Federal States in Ethiopia) [9]. It is one of the major languages that are widely
used as a working language of the Regional State of Oromia, that is used in Ethiopia and Africa [10].
Afan Oromo has a large number of ethnic groups when compared with the rest of Ethiopian
languages [11]. In Afan Oromo, like other natural languages, there exists same form of words, that
has more than one different meanings. These words are homonyms which has multiple distinct
meanings. For words having multiple senses, the challenge is to find out which particular sense is
appropriate in a given context. In natural language, WSD is the problem of determining which sense
(meaning) of a word is active in a particular context [12]. However, the task proved to be difficult
for computer and believed that it is an open problem [14] and the problem is worse for under
resourced languages like Afan Oromo.

Identifying the correct senses of the ambiguous words is easy for human being basically, sometimes
it is difficult. However it is too tough for the machine to identify the correct sense of these words.
For instance, applications like machine translation (English to Afan Oromo, Amharic to Afan
Oromo) the role of the word sense disambiguation is great because it helps to find the meaning of

3
words. In such cases, it is required to have words that have translation for different senses and that
are potentially ambiguous within a given domain [13].

In order to have a clear understanding of ambiguous words in the language, WSD for Afan Oromo
language is also needed to be developed. Because, now a days as the development of technology is
increasing rapidly, like any other language Afan Oromo has also started to use the technology for
different purposes. Many different Afan Oromo documents are available in different sites which is
an opportunity, yet creating a problem for information retrieval.

As already mentioned, human readers have the extraordinary capability to infer the meaning of a
word from the context, even if they have never heard of the word before, and there are no visual cues
present to support its interpretation and they can still make reasonable assumptions about the
meaning. The surrounding contexts allow guessing the meaning of ambiguous words. However, this
situation becomes more complicated if we want to use a computer to infer the meaning of words.
Unlike humans, computer does not have any a priori knowledge of ambiguous words in natural
languages like Afan Oromo. The words sharing the same spelling and pronunciation, but have
different senses, which has multiple senses can be totally unrelated to each other. This is a
fundamental problem for many established human language technology applications.

The intention of this study is to device an algorithm by adopting and integrating the already existing
once for word sense disambiguation for Afan Oromo words. Currently, few researches in word sense
disambiguation have been commenced for Ethiopian languages, particularly for Amharic in different
domains by adopting different techniques. Our work presents a contribution towards developing
natural language processing applications for Ethiopian languages exhibiting similar patterns with
Afan Oromo. Specifically, it increases the scope of the word sense disambiguation research by
investigating its applicability for Afan Oromo language. The other techniques used in this study are
vector space model for similarity measure and clustering. The contexts are captured via extracting
co-occurrences in a defined window of words from a corpus.

4
1.3. Objective of the Study

1.3.1. General Objective:

The general objective of this study is to develop the hybrid word sense disambiguation prototype to
Afan Oromo ambiguous words.

1.3.2. Specific Objective:

In line with the general objective, the research is aimed at addressing the following specific
objectives:

 To review related works so as to have a conceptual understanding of the state-of-the-art in

Afan Oromo word sense disambiguation;
 To acquire the required corpus from different sources to use it for training and testing and to
contribute as the language resource;
 To build a model using Unsupervised Machine Learning and Hybrid approach;
 To develop a prototype that disambiguate multiple sense words in Afan Oromo;
 To construct the rule that support the machine learning disambiguation model;
 To forward the direction of the WSD research;

1.4. Scope and Limitation of the Study

The scope of the study is limited to investigating word sense disambiguation for Afan Oromo words
particularly at word level, which study what is the actual meaning of the ambiguous word in a given
context. The technique used was a hybrid approach which combine machine learning (which is
unsupervised) with detail background knowledge of the language (rule based). Due to the absence of
standard train and test corpus, we had to prepare corpus for the experimentation which is
unannotated machine readable free text. However, the amount of corpus prepared for this study is
relatively small and requires further development. In this case, we have studied ambiguous words at
word level in sentences based on surrounding contexts and modifiers. The main idea of this work is
to explore word sense disambiguation using corpus and information from rule in Afan Oromo. The
study investigated the meaning of words by looking in an unstructured corpus prepared from

5
different sources and official websites. The lack of annotated corpus and senses definition are among
the constraints of this study.

1.5. Significance of the Study

WSD is considered as an input for many Natural Language Processing (NLP) applications. The
researchers in the area of NLP of Afan Oromo Language could use the output of this thesis.
Particularly, researchers interested in area of MT (Machine Translation), Question Answering (QA),
Text Summarization are among the top beneficiaries. Ambiguity resolution has been pursued as a
way to improve retrieval systems, and generally get better information access. Moreover, for the
students who are always practicing to learn word meaning, the system is very helpful. The output of
this study fixes the problem of word meaning and complexity in the society. Similarly, this work has
an input for translation of English and Amharic to Afan Oromo. Interested users who want to browse
web page by Afan Oromo could also use the system. The finding of this study is expected to support
Afan Oromo researchers who deal with NLP applications. It also contributes to future researches and
development in the area of NLP specifically in machine learning, Information retrieval as those areas
require word sense disambiguation as complements.

1.6. Organization of the Thesis Report

The thesis is organized into five chapters comprising Introduction, Literature review, overview of
Afan Oromo, Methodology, Experimentation and Evaluation, and Conclusion & Recommendations.
The first chapter gives the general introduction of the study. The second chapter presents review
made on different literatures regarding Word Sense Disambiguation together with its approaches and
different machine learning techniques. The third chapter discusses the overview of the grammatical
categories in Afan Oromo. The fourth chapter discusses the methodology and presents the process of
data preparation for the system to be developed for Afan Oromo. It present the algorithms and rule
preparation based on the analysis of rule categories as reviewed. The fifth chapter discusses the
experimentation and discusses of the findings. Chapter sixth presents the conclusion and the
recommendations as well as some directions to future works.

6
 Chapter Two
2. Literature Review
2.1. Introduction

In this chapter, related literature review in the Word Sense Disambiguation (WSD) is presented. This
part of study enlightens briefly on some of the work done by those researchers.

2.2. Basic Concepts of Word Sense Disambiguation

In natural languages, [15] each word may have more than one meaning that is why a single word
sometimes can have many senses. Disambiguation means to remove all ambiguity. WSD is a
technique used in finding the meaning of a word in a sentence. A word can have multiple meanings
and the exact meaning of the word is decided based upon context by humans. Computers also can
follow similar technique i.e. decide meaning of an ambiguous word in a sentence using context
information. Likewise, it is the process of identification to decide the appropriate meaning of an
ambiguous word in a particular context. People decide the meaning of a word based on the
characteristic senses of a discussion or situation using their own merits. Machines have no ability to
decide such an ambiguous situation unless some practices have been planted into the machines’
memory [16]. The word sense is extracted by human by connecting it to that specific context. But for
corpus this function is performed by natural language running by the qualities of the term. It
disambiguates ambiguity of an individual word that can be used (in different contexts) to express
multiple meanings. The context is the only real way to recognize the significance of the homonym
term. Generally, context can be used in the bag of words and relational information. Here the bag of
words is the context that considered as words in some window surrounding the ambiguous word in
terms of distance. The surrounding word of the ambiguous word decides correct choice of word.
Assume, the ambiguity word afaan has surrounded with the following contexts. So, these contexts
give us the clue of information what the meaning of this ambiguity word is. It seems that this
ambiguity word identify its senses with the help of its contexts as shown the in Figure below:

7
 Figure 2.1. Surrounding Contexts of the term Afaan

2.3. Approaches to Word Sense Disambiguation

Regardless of the approach, WSD systems extract contextual information of an ambiguous word (in
the corpus) and compare against the sense differentiation information stored in that word [13]. Many
approaches have been used through the evolution of WSD research.

Based on the level of supervision the machine learning approaches are categorized as Supervised
machine learning, Unsupervised machine learning and Bootstrapping machine learning approaches.

2.3.1. Supervised Word Sense Disambiguation

Supervised Word Sense Disambiguation use machine-learning techniques to learn a classifier from
labeled training sets, that is, sets of examples encoded in terms of a number of features together with
their appropriate sense label (or class) [3]. The systems in the supervised learning approach category
are trained to develop a classifier that can be used to assign a yet unseen example to one of a fixed
number of senses. That means, there is trained corpus, where the system learns to classify and a test
corpus which the system must annotate. So, supervised learning can be considered as a classification
task.

8
Supervised learning requires labeled training data where every instance in the training data is
associated with an output value or label that can be thought of as a special attribute or feature for
each instance. For WSD, every instance in the training data should be assigned a label that
corresponds to the correct sense of the ambiguous word that the instance contains or represents.
Machine learning algorithms make use of the instance attributes or features of the training data and
generate a model to predict the label of any given instance. This model can be applied to unseen
instances to predict their labels. Algorithms that can learn to predict discrete valued labels are
classification algorithms or classifiers, whereas the algorithms that can learn to predict continuous
valued labels are called regression algorithms. As the task of WSD only involves discrete valued
labels for word senses, only classification algorithms used in this method.

The main problem associated with supervised approach is the need for a large sense tagged training
set. Despite the availability of large corpora in some language, manual sense-tagging of a corpus is
very difficult, time consuming, limiting the number of sense tagged words to be used and very few
sense-tagged data are available.

2.3.2. Unsupervised Word Sense Disambiguation

Unsupervised Word Sense Disambiguation methods are based on unlabeled corpora, and do not
exploit any manually sense-tagged corpus to provide a sense choice for a word in context unlike
supervised method [3]. Unsupervised methods have the potential to overcome the knowledge
acquisition bottleneck [17], that is, the lack of large-scale resources manually annotated with word
senses. These approaches to WSD are based on the idea that the same sense of a word has similar
neighboring words. They are able to induce word senses from input text by clustering word
occurrences, and then classifying new occurrences into the induced clusters. They do not rely on
labeled training text and, in their purest version, do not make use of any machine-readable resources
like dictionaries, thesauri, ontology. However, the main disadvantage of fully unsupervised systems
is that, as they do not exploit any dictionary, they cannot rely on a shared reference inventory of
senses [18]. Most of the time, supervised approaches are superior to unsupervised in terms of
accuracy of automatic disambiguation when used on the same type of texts that the systems were
trained on [2].

Nevertheless, there is another issue connected with the problem of the definition of a meaning, i.e.,
an issue of the creation of other resources used for automatic system performing WSD. This is

9
especially evident in the creation of corpora that is manually annotated (tagged) with the senses,
which are used for training machine learning classifiers in a supervised setting. There are two
important problems during manual sense tagging of a corpus: low inter annotator agreement (IA) and
high cost of the annotation process. IA is a way of measuring how much an annotation assigned by
one annotator differs from annotations assigned by another annotator. IA is used for the estimation
of an upper bound on performance on automatic WSD but there is also another measure [19].

2.3.3. Bootstrapping Machine Learning Approach

The bootstrapping approach is situated between the supervised and unsupervised approach of WSD.
The aim of bootstrapping is to build a sense classifier with little training data, and thus overcome the
main problems of supervision: the data scarcity problem specially lack of annotated data. The
bootstrapping methods use a small number of contexts labeled with senses having a high degree of
confidence. This could be accomplished by hand tagging with senses the contexts of an ambiguous
word w for which the sense of w is clear because some seed collocations [2] occur in these contexts.
These labeled contexts are used as seeds to train an initial classifier. This is then used to extract a
larger training set from the remaining untagged contexts. Repeating this process the number of
training contexts grows and the number of untagged contexts reduces.

In addition to the above approaches, another criterion for categorization of approaches are use of
data, knowledge (dictionary) [5].

2.3.4. Corpus - Based Approaches

A major challenging face WSD research is the ability to acquire a large number of words with their
different contexts. Corpus-based approaches came up with an alternate solution to the challenge by
obtaining information necessary for WSD directly from textual data which is a corpus. A corpus
provides a bank of samples which enable the development of numerical language models, and thus
the use of corpora goes hand-in-hand with empirical methods.

10
2.3.5. Knowledge - Based Approaches

Knowledge-based WSD is based on lexical resources like dictionaries, thesauri, and corpora where
Machine-readable dictionaries (MRDs) are the primary source of acquisition of data. There are
various knowledge based approaches such as WSD uses Selectional Preferences, Lesk’s algorithm,
Walker’s algorithm, WSD uses conceptual density and WSD uses Random Walk Algorithm. Rely on
external knowledge resources (example, WordNet, Thesaurus), may use grammar rules for
disambiguation and may use hand coded rules for disambiguation. Under this approach
disambiguation is carried out using information from an explicit lexicon or knowledge base. Since
corpus based approaches require a considerable amount of work to create a classifier for each word
in a language. Knowledge-based approaches use an explicit lexon like, MRD, thesauri,
computational lexicons such as WordNet or (hand-crafted) knowledge bases as information source to
resolve lexical ambiguities for many words.

The created knowledge bases [20] which associate each sense in a dictionary with a signature
composed of the list of words appearing in the definition of that sense. Disambiguation will be
accomplished by selecting the sense of the target word whose signature contained the greatest
number of overlaps with the signatures of neighboring words in its context. Because of the fact that
dictionaries are created for human use, not for computers, there are some inconsistencies. Although
they provide detailed information at the lexical level, they lack pragmatic information used for sense
determination.

Thesauri provide information about relationships among words, most notably synonymy [18].
Thesaurus based disambiguation makes use of the semantic categorization provided by a thesaurus
or a dictionary with subject categories. The basic inference in thesaurus based disambiguation is that
semantic categories of the words in a context determine the semantic category of that context as a
whole. And this category, then determines the correct senses that are used. Similar to machine
readable dictionaries, a thesaurus is a resource for humans, so there is not enough information about
word relations.

Computational Lexicons are large electronic database containing useful lexical relations in linguistic
psycholinguistic and computational [3]. Lexicon like WordNet is used for sense evaluation and for
similarity measure in WSD. For example [21] created a knowledge base from WordNet's hierarchy

11
and apply a semantic similarity function to accomplish disambiguation, also for the purposes of
information retrieval.

Finally, we can categorize WSD approaches as token based and type-based. Token based approaches
associate a specific meaning with each occurrence of a word depending on the context in which it
appears. In contrast, type-based disambiguation is based on the assumption that a word is
consensually referred with the same sense within a single text. Consequently, these methods tend to
infer a sense (called the predominant sense) for a word from the analysis of the entire text and
possibly assign it to each occurrence within the text [5].

2.3.6. Hybrid Machine Learning Approach

The combinations of machine learning and rule based approaches have been used to make hybrid
methods (in this study). A major weakness with unsupervised method is the accuracy and lack of
ability to place a label on each discriminated word group. Combining an unsupervised method with a
rule based method could overcome this weakness.

Since they obtain disambiguation information from both corpora and explicit linguistic knowledge
bases, hybrid approaches do not fall into either knowledge or corpus-based. Hybrid systems aim to
use the strengths of the both conquering specific limitations associated with a particular approach, to
improve WSD accuracy. They base both on a ‘knowledge driven, corpus-supported’ theme, utilizing
as much information as possible from different sources. [22] used Bootstrapping approaches where
initial data come from an explicit knowledge source which is then improved with information
derived from corpora. He defines a small number of seed definitions for each of the senses of a word
(the seeds can also be derived from dictionary definitions or lexicons such WordNet). Then the seed
definitions are used to classify the obvious cases in a corpus.

Although above three approaches are very popular for WSD, they have weakness respectively.
Semantic relations only cannot disambiguate any homonym word in any context. However,
WordNet does not define the noun-verb relations. The supervised machine learning approach needs a
great deal of training data that are expensive and sometimes hard to obtain. Therefore, a hybrid
approach that uses both the machine and rule based methods to solve WSD problems. It is an
approach that provides better results and make the results satisfy users’ preferences.

12
2.4. Context Window Size

Context is the only means to identify the sense of an ambiguous word in sense disambiguation. All
algorithms make use of the contexts to provide information for sense disambiguation. The context
window size is defines the size of the window of context. A window size of N means that there will
be a total of N words in the context window. If N is a number (positive), then there will be N word
on the left and right side of the target word, where N is for example: 1, 2, 3, 4,5, etc. For example, if
the window size is 2, then there will be 2 words on the left and right side of the target word. In order
to disambiguate a given word, a small and wider context should be considered in the performance of
the system to rise overall. However, a wider context implies more data and thus further features. The
different methods and algorithms can benefit from the choice of the context size in a distinct way,
which means that the optimal size of the context can depend on the used method and that the
selection of the size of the context leads to a variation of the WSD output. [22] also claimed that the
ambiguity word occurring in the data can be captured by a different size of the context.

2.5. Vector Space Model

Vector space model is the most important model to obtain a vector that represents each word in the
sentences. This vector is usually chosen to capture the contexts in which words can appear in. The
representation for a word is a point in a dimensional space. The dimensions stand for context items
(for example, co-occurring words), and the coordinates depend on the co-occurrence counts. Each
dimension corresponds to a separate word. If ambiguous word occurs with contexts word within the
collected contexts, it has similarity on the vector. Similarly, all words in the contexts have its weight
(the occurrence of word in the corpus); zero means the word has dissimilarity with the contexts.
Several different ways of computing these values as weights (term) have been used. One of the best
known schemes is weighting (frequency occurrence of the word with the contexts) [23].

The dimensionality of the vector is the number of occurrence of word in the contexts. As all vectors
under consideration, a cosine value of zero means that the ambiguous word and contexts are
dissimilar and have no match (i.e. the ambiguous word does not exist with the contexts). The context
is formally a text that surrounds a language unit (e.g. a word) and helps to determine its
interpretation. It represent the occurrences of target words as word vectors. From these vectors,
vectors are formed and found that is a function of cosine similarity between the contexts [23].

13
The computation of meaning similarity as operationalized by vector-based models has found
widespread use in many tasks within NLP. The popularity of vector-based models lies in
unsupervised nature and ease of computation. The model represent the meaning of each word as a
point in a dimensional space, where each component corresponds to some co-occurring contextual
[24]. The advantage of taking such a geometric approach is that the similarity of word meanings can
be easily quantified frequency by measuring the cosine similarity of the angle between them [25].

2.6. Clustering Algorithms

Clustering algorithms are generally categorized as hierarchical and partitional. The next section
describes some common clustering algorithms. Here are general properties that characterize
clustering algorithms [26].

Agglomerative vs Divisive: In agglomerative algorithms (bottom-up approach), each element is

initially its own cluster and then the most similar clusters are iteratively merged until we are left with
one large cluster containing all elements or until a stopping condition is met. Conversely, divisive
algorithms (top-down approach) initially begin with a single all-encompassing cluster and iteratively
split the clusters until each element belongs to its own cluster or until a stopping condition is met.

2.6.1. Hierarchical Agglomerative Algorithms

Hierarchical algorithms produce a nested partitioning of the data elements by merging clusters.
Agglomerative algorithms iteratively merge clusters until all-encompassing cluster is formed [27],
while divisive algorithms iteratively split clusters until each element belongs to its own cluster. The
merge and split decisions are based on the similarity metric. The resulting decomposition (tree of
clusters) is called a dendrogram.

The different versions of agglomerative clustering differ in how they compute cluster similarity. The
most common versions of the agglomerative clustering algorithm are single-link, complete-link and
average-link clustering [22].

14
A. Single Link Clustering: The single link algorithm is a MIN version of the hierarchical
agglomerative clustering method which is a bottom-up strategy, compare each point with
each point.

Each context is placed in a separate cluster, and at each step merge the closest pair of
clusters, until certain termination conditions are satisfied. For the single link, the distance of
two clusters is defined as the minimum of the distance between any two points in the
clusters. In single-link clustering the similarity between two clusters is the similarity between
their most similar members (e.g. using the Euclidean distance) [28]. It is capable of
discovering clusters of varying shapes like the clusters of Figure 2.2 a. However, single-link
is not practical because it suffers from the chaining effect [27]. For example, in Figure 2.2 a,
single-link clustering generates an elongated cluster because of a bridge of elements
connecting two clusters.

B. Complete Link Clustering: The complete linkage algorithm is the MAX version of the
hierarchical agglomerative clustering method which is a bottom-up strategy: compare each
point with each point. Each context is placed in a separate cluster, and at each step merge the
farthest pair of clusters, until certain termination conditions are satisfied.

In complete-link clustering, the similarity between two clusters is the similarity between their
maximum similar members (e.g. using the Euclidean distance) [29]. Although complete-link
clustering is not capable of discovering clusters like the two in Figure 2.2 b, it does not suffer
from the chaining effect. Rather than producing straggly elongated clusters like single-link,
complete-link generates compact clusters. Complete-link generates better clustering’s than
single-link in many applications[30]. Figure 2.2 a illustrates the cluster similarity of
complete-link.

15
 Figure 2.2. a) Single Link, b) Complete Link and c) Average Link

C. Average Link Clustering: Average-link clustering produces similar clusters to

complete link clustering except that it is less susceptible to outliers [31]. It computes the
similarity between two clusters, as the average similarity between all pairs of contexts across
clusters (e.g. using the Euclidean distance). Figure 2.2 c shows merging decisions average
linkage algorithms.

Generally, to use a hierarchical clustering procedure, need to express these distances in Weka tool.
As an example, easily compute the Euclidean distance between cluster B and cluster C (generally
referred to as d (B, C)) with regard to the two variables x and y by using the following formula,
Weka by default has Euclidean distance.

16
2.6.2. Partitional Algorithms

Partitional algorithms do not produce a nested series of partitions. Instead, they generate a single
partitioning, often of predefined size k, by optimizing some criterion. A combined search of all
possible clustering’s to find the optimal solution is clearly intractable. The algorithms are then
typically run multiple times with different starting points. Partitional algorithms are not as versatile
as hierarchical algorithms, but they often offer more efficient running time [31].

A. K-means
This algorithm has the objective of classifying a set of n contexts into k clusters, based on the
closeness to the cluster centers. The closeness to cluster centers is measured by the use of a
Euclidean distance algorithm. K-means is an iterative clustering algorithm in which items are moved
among sets of clusters until the desired set is reached. A high degree of similarity among senses in
clusters is obtained, while a high degree of dissimilarity among senses in different clusters achieved
simultaneously [31].
Although the k-means algorithm often produces good results, it is not time-efficient and does not
scale well. By saving distance information from one iteration to the next, the actual number of
distance calculations that must be made can be reduced. Some k-means variations examine ways to
improve the chances of finding the global optimum. This often involves careful selection of initial
clusters and means. Another variation is to allow clusters to be split and merged. The variance within
a cluster is examined and if it is too large, a cluster is split. Similarly, if the distance between two
cluster centroids is less than a predefined threshold, they will be combined.
k-means clustering [32] is a method of cluster analysis which aims to partition n observations into k
clusters in which each observation belongs to the cluster with the nearest mean. K-means [33] is one
of the simplest unsupervised learning algorithms that solve the well-known clustering problem. The
procedure follows a simple and easy way to classify a given data set through a certain number of
clusters (assume k clusters) fixed a priori. The main idea is to define k centroids, one for each
cluster. These centroids should be placed in a cunning way because of different location causes
different result. So, the better choice is to place them as much as possible far away from each other.
The next step is to take each point belonging to a given data set and associate it to the nearest
centroid.

17
When no point is pending, the first step is completed and an early group age is done. At this point we
need to recalculate k new centroids as bar centers of the clusters resulting from the previous step.
After we have these k new centroids, a new binding has to be done between the same data set points
and the nearest new centroid. A loop has been generated. As a result of this loop we may notice that
the k centroids change their location step by step until no more changes are done. In other words
centroids do not move any more.
The most commonly used family of partitional algorithms is based on the K-means algorithm [31].
K-means clustering is often used on large data sets since its complexity is linear in k, the number of
elements to be clustered. It creates a partitioning such that the intra-cluster similarity is high and the
inter-cluster similarity is low. K-means using the concept of a centroid where a centroid represents
the center of a cluster. A centroid is usually not an element of the cluster. Rather, it is a pseudo-
element that represents the center of all other elements. Often the mean of the feature vectors of the
elements within a cluster is used as that cluster centroid. It is often difficult to define a centroid for
categorical features.

K-means iteratively assigns each element to one of k clusters according to the centroid closest to it
and recomputed the centroid of each cluster as the average of the cluster’s elements. The following
steps outline the algorithm for generating a set of k clusters:

1. Randomly select K elements as the initial centroids of the clusters;

2. Assign each element to a cluster according to the centroid closest to it;
3. Recomputed the centroid of each cluster as the average of the cluster’s elements;
4. Repeat Steps 2-3 for T iterations or until a criterion converges, where T is a
predetermined constant.

B. Expectation Maximization

Expectation Maximization (EM) algorithm [34] is also an important algorithm of data mining. We
used this algorithm when we are satisfied the result of k-means methods. An expectation
maximization (EM) algorithm is an iterative method for finding maximum likelihood estimates of
parameters in statistical models, where the model depends on unobserved latent variables. The EM
[35] iteration alternates between performing an expectation which computes the expectation of the
log-likelihood evaluated using the current estimate for the parameters, and maximization which

18
computes parameters maximizing the expected log-likelihood found. These parameter-estimates are
then used to determine the distribution of the latent variables.
Expectation maximization is a clustering algorithm that works based on partitioning methods. This
algorithm is a memory efficient and easy to implement algorithm, with a profound probabilistic
background. EM is widely used iterative algorithms for estimating model parameters [30].

Expectation maximization (EM) is a well-known algorithm used for clustering in the context of
mixture models. This method estimates missing parameters of probabilistic models. Generally, this
is an optimization approach, which had given some initial approximation of the cluster parameters,
iteratively performs two steps: first, the expectation step computes the values expected for the cluster
probabilities, and second, the maximization step computes the distribution parameters and their
likelihood given the data. It iterates until the parameters being optimized reach a fix point or until the
log-likelihood function, which measures the quality of clustering, reaches its maximum [30].

There are two potential problems when using the EM algorithm. First, it is computationally
expensive and convergence can be slow for problems with large numbers of model parameters. To
solve the above problem we used small data set for this study. Second, if the likelihood function is
very unbalanced it may always converge to a local maximum and not find the global maximum.

To simplify the discussion, we first briefly describe the EM algorithm. The algorithm is similar to
the K-means procedure to that a set of parameters are re-computed until a desired convergence value
is achieved. The parameters are re-computed until a desired convergence value is achieved. The
finite mixture model assumes all attributes to be independent random variables [30].
A mixture is a set of N probability distributions where each distribution represents a cluster. An
individual instance is assigned a probability that it would have a certain set of attribute values given
it was a member of a specific cluster.

2.7. Divisive Clustering

Although it is not as common as agglomerative clustering. Divisive clustering algorithms start with a
single cluster containing all elements. Considering all possible splits of the cluster into two clusters
(2n-1)
gives 2 – 1 possibilities. Using a splitting heuristic to iteratively split the largest cluster,
Divisive clustering algorithms has worst case time complexity O (n2logn).

19
 Figure 2.3. Hierarchical Clustering and Divisive Clustering [28]

2.8. Rule Based Approach

Rule-based disambiguation is the popular approach that uses hand-written rule for disambiguating.
Hand-written rule are used to identify the correct meaning when a word has more than one possible
meaning. Disambiguation is done by analyzing the linguistic features of the word, based on such
factors as taking its preceding word, its following word and other aspects. This information is coded
in the form of rules. The fact that a rule-based disambiguate that automatically learns its rule can
perform so well should offer encouragement for researcher to further explore rule-based
disambiguating, searching for a better and more expressive set of rule.

20
The advantages of this approach are that it is easy to incorporate domain knowledge into the
linguistic knowledge which provides highly accurate results. Furthermore, the linguistic knowledge
acquired for one natural language processing system may be reused to build knowledge required for
a similar task in another system [36].

However, rule based is the most method used in NLP of disambiguation. When there is lack of
available resources and their limitations, the rule based approach was used which rely on hand-
constructed linguistic rule in systems and resources. The use of rule based transformations are based
on a core of solid linguistic knowledge [36].

Obviously, the manual creation of rule is an expensive and time consuming effort, which must be
repeated every time the disambiguation scenario changes. Knowledge of the linguistic used for WSD
are either lexical knowledge released to the public or world knowledge learned from a training
corpus [37]. The detail description of WSD approaches are in the following table 2.1.

Algorithm Category Strengths Weaknesses

Rule Based Accuracy It takes time, needs expert and

also it is very expensive

Hybrid Approach Accuracy and it mix from others -

Unsupervised Approach No pre-training necessary Works Merely discriminates between

on multiple languages with no word senses; not disambiguate
modification to the algorithm word senses

Supervised Approach It needs pre-training (labeled Dependent on pre-annotated

data) corpora for training data

Rely on precompiled
Knowledge Based Accuracy lexical knowledge resources

Table 2.1.Summary of Word Sense Disambiguation Approaches

21
2.9. History of Word Sense Disambiguation (WSD)

WSD is one of the open and oldest problems in computational linguistics. It was in the 1940s that
WSD was first formulated as a separate task [8]. In the 1950's, ideas for machine translation
encouraged the first major research push in WSD, and although there was relatively little computing
power around at the time, researchers thinking about WSD created algorithms and ideas that are still
in use today. [8] discussed a 'window size' around the target word, taking every word within a
distance N before and after the target word, He found that using a window size of only 2 words
either side of the target word offered no substantial difference in disambiguation accuracy than using
the whole sentence.

The inherent difficulty in the task of WSD was well appreciated further in the 1960s, at that time;
researchers had already in mind essential ingredients of WSD, such as the context in which a target
word occurs, statistical information about words and senses, knowledge resources. Very soon it
became clear that WSD was a very difficult problem, also given the limited means available for
computation. Indeed, its acknowledged hardness [38] was one of the main obstacles to the
development of MT in the 1960s. WSD was then resurrected in the 1970s, within the research in
Artificial Intelligence on complete natural language understanding. Wilks's preference semantics, as
mentioned in [39], was one of the first systems to explicitly account for WSD.

The 1980s witnessed a turning point in WSD research, and large scale corpora and other lexical
resources became available. Before the 1980s, much of WSD research depended on handcrafting of
rules. Now it became possible to use knowledge extracted automatically from the resources[20].

The research by [29] came up with a simple yet seminal algorithm that used dictionary definitions
from the Oxford Advanced Learner's Dictionary (OALD), and this marked the beginning of
dictionary-based WSD. [40] combined the information in Roget's thesaurus with co-occurrence data
from large corpora in order to learn disambiguation rules for Roget's classes.

The 1990s saw further improvements in the field, which can be categorized in three major groups
WordNet, statistical NLP, and SensEval (later SemEval). WordNet made it possible for all the
researchers to have easy and free access to a standardized inventory using which to compare their
work. Its hierarchical structure, synsets, and other such features made it the most used general sense
inventory in WSD research.

22
The typical approach in WSD so far has been supervised learning, where systems are trained on
manually tagged corpora. Statistical approaches in supervised learning were used by [39] and several
others which was a foresight to the so-called statistical revolution" in the 1990s. [41] was the first to
use corpus-based WSD in statistical MT. SensEval (later SemEval) made it possible for researchers
to compare different systems with each other because of the fixed set of test words, annotators, sense
inventories, and corpora. Before SensEval, the only common ground that WSD researchers had were
a lower bound (calculated by either picking a random sense, or taking into account the most frequent
senses) and an upper bound (derived from inter-tagger agreement). Now it became possible to
develop different systems and evaluate them on the data sets provided by SensEval, thereby
introducing scientific rigor and uniformity. SensEval eventually became the primary forum for all
WSD evaluations. In SensEval-3, a conclusion was reached that WSD in itself has reached a
performance level, and no significant rise in the results obtained already is possible. It has been since
then, that people started thinking about new directions in which WSD research can go. In particular,
in recent years there has been considerable growth in the areas of parallel bilingual corpora, and
unsupervised corpus-based WSD. This study employed hybrid WSD and attempts to draw upon the
idea that hybrid WSD is the way to go in future.

2.10. Related Works

The work from various books, papers, articles, journals and web pages has been referred to this of
the study. Some of them are reviewed at the following:

According to [42] unsupervised machine learning approach for Amharic using five selected
algorithms were used; these are Simple k-means, EM (Expect Maximization) and agglomerative
single, average and complete link clustering algorithms. The tested unsupervised machine learning
method that deals with clustering of contexts for a given word that express the same meaning. The
above mentioned work concluded that simple k means and EM clustering algorithms achieved
higher accuracy on the task of WSD for selected ambiguous word, provided with balanced sense
distribution in the corpus.

According to [43] the meaning of words are extracted based on live contexts using supervised and
unsupervised approaches. Unsupervised approaches use online dictionary for learning, and
supervised approaches use manual learning sets. Hand tagged data are populated which might not be

23
effective and sufficient for learning procedure. This limitation of information is the main flaw of the
supervised approach. They developed approach focuses to overcome the limitation using a learning
set which is enriched in a dynamic way of maintaining new data. The trivial filtering method is
utilized to achieve appropriate training data. They introduce a mixed methodology having “Modified
Lesk” approach and “Bag-of-Words” having enriched bags using learning methods. The approach
establishes the superiority over individual “Modified Lesk” and “Bag-of-Words” approaches based
on experimentation.

The research by [44] developed a prototype for Amharic language which is semi-supervised method
of WSD. Their finding showed that Semi-supervised learning using a bootstrapping algorithm
performs better. It is more adaptive on WSD for Amharic language. Specifically, Adtree, Adaboost
and bagging are potential algorithms to be applied to Amharic WSD systems using semi-supervised
learning methods. The authors concluded that a window size of 3-3 can be a standard window size of
Amharic WSD systems development.

According to [45] developed a supervised learning approach for Amharic WSD. However,
supervised machine learning approach of WSD performs better by human intervention, but it has
knowledge acquisition bottleneck i.e. it requires manually labelled sense examples which take much
time, very laborious and therefore very expensive to create when the corpus size increases.
Development of these resources requires huge amount of human efforts and typically takes years for
a building. As these resources are not available for Afan Oromo, supervised techniques cannot be
immediately applied. Few attempts have been made on unsupervised WSD like [22], which seeks
minimum human involvement, in the form of providing a few seed words that occur with each sense
of the ambiguous word for bootstrapping the algorithm. The algorithm then classifies each
occurrence of the ambiguous word in a corpus into several clusters so that all the occurrences are in
the same sense within a cluster. Additional co-occurrences are collected in this process, which are
then used for disambiguating unseen texts from the held-out corpus.

One research by [46] explored Word Sense Disambiguation using supervised, unsupervised, and
knowledge-based approaches for WSD. The finding showed that supervised methods, undoubtedly
perform better than the other approaches. However, relying on the availability of large training
corpora for different domains, languages, and tasks is not a realistic assumption. Their finding also

24
furnishes an idea of a few of the WSD algorithms and their performances, which compares and
assesses the need of the word sense disambiguation.

According to [47] unsupervised WSD does not require much time in creating high quality resources
and perform great in terms of accuracy. The other reason not to go for supervised approach is that,
even if all possible resources are available to build a great supervised approach, it cannot be
implemented in other languages easily. The resources have to be replicated for all possible
languages. Another disadvantage of using the supervised approaches is, by using fixed sense
repositories, the fixed number of senses present in that repository. It cannot discover new senses of
words, which are not present in the sense corpus. Hence, only considering the accuracy of the
approach is not a good idea, but considering its versatility and portability to other languages and
domains is also equally important. These are the reasons that many unsupervised approaches were
being tried by many researchers in WSD.

The research by [2] conducted the overall survey and early work on word sense extraction in
context. The author started with a word sense discussion which was done by [48], which describes
the need for word sense extraction in machine translation. The very first efforts at automatic sense
disambiguation were created in the framework of machine [48] traces the foundation of a strategy to
word sense removal in a word that is the following. The useful issue is: “What minimum value of N
(Neighboring word) will, at least in a tolerable fraction of cases, lead to the correct choice of
meaning for the central word?” Then the authors concentrated on early test with preceding query
done by [49] which extracts sense of the main word by getting a couple of phrases on both sides of a
central word. The impressive truth about it is that early function on word sense removal is the
amount to which the strategies and essential difficulties towards the difficulty were foreseen and
created in those days. It gives overall earlier work on word sense extraction. It provides idea about
selecting values of neighboring word for extracting word sense in a particular context.

According to [3] surveyed the field of word sense disambiguation. The author asserted that human
language is ambiguous, so that many words can be interpreted in multiple ways depending on the
context in which they occur. Most of the time humans do not even think about the ambiguities of
language, machines need to process unstructured textual information and transform them into data
structures which must be analyzed in order to determine the underlying meaning. The computational
identification of the meaning of words in context is WSD. The author, distinguished two main

25
approaches to WSD namely: i) Supervised WSD: These strategies use machine-learning methods to
understand a classifier from tagged training sets, that's models of examples secured in conditions of
a quantity of attributes as well as their proper sense tag (or course). ii) Unsupervised WSD: These
procedures are centered on unlabeled corpora and do not utilize any manually sense-tagged corpus to
supply a sense alternative for a word in context. It gives detailed survey of word sense extraction
with approaches. It is useful to select an approach for sensing word in context.

The research by [50] stated that lexical meaning of words using vector spaces and linear algebra is
paramount. He argued that the meanings of words will be represented using vectors, as part of a
high-dimensional “semantic space”. The fine-grained structure of this space is provided by
considering the contexts in which words occur in large corpora of text. Words can easily be
compared for similarity in the vector space, using any of the standard similarity or distance measures
available from linear algebra, for example the cosine of the angle between two vectors. Other
authors said that the hypothesis underlying distributional models of word meanings is the so-called
distributional hypothesis: the idea that “Words that occur in similar contexts tend to have similar
meanings” [51]. According to [50] now the basis vectors correspond to whole sequences of
grammatical relations, relating the ambiguous word and contexts. Which paths to choose is a
parameter of the approach, with the idea that some paths will be more informative than others.

The research stated by [52] addressed that the task of computing vector space representations for the
meaning of word occurrences, which can vary widely according to context. This task is a crucial step
towards a more robust, vector-based compositional account of sentence meaning. The authors argued
that existing models for this task do not take syntactic structure sufficiently into account. In the
paper, it was concluded that structured vector space models of word meaning relying on this
procedure of vector composition are limited both in their scope and scalability. The best-known
work in this category is that of [25]. He first computes “first-order” vector representations for word
meaning by collecting co-occurrence counts from the entire corpus. Then, he determines “second-
order ” vectors for individual word instances in their context, which is taken to be a simple surface
window, by summing up all first-order vectors of the words in this context. The resulting vectors
form sense clusters. In addition to a vector representing a word’s lexical meaning, it contains vectors
representing the word’s selectional preferences. These selectional preferences play a central role in
the computation of meaning in context.

26
The other research by [53] stated the unsupervised learning algorithm for sense disambiguation that,
when trained on unannotated English text, rivals the performance of supervised techniques that
require time-consuming hand annotations. The algorithm is based on two powerful constraints that
words tend to have one sense per discourse and one sense per collocation exploited in an iterative
bootstrapping procedure.

According to [54] has a pioneered work in the hierarchical clustering of word senses. In his
disambiguation experiments, Schutze used post-hoc alignment of clusters to word senses. Because
the top-level cluster partitions based purely on distributional information do not necessarily align
with standard sense distinctions, he generated up to 10 sense clusters and manually assigned each to
a fixed sense label (based on the hand-inspection of 10-20 sentences per cluster).

The research by [55] published a work on early application of bootstrapping to augment training sets
for a supervised sense tagger. She trained her fully supervised algorithm in hand-labelled sentences,
applied the result to new data and added the most confidently tagged examples to the training set.
Regrettably, this algorithm was only described in two sentences and was not developed further.
Other unsupervised methods have shown great promise. [56] developed a method using co-
occurrence statistics in independent monolingual corpora of two languages to guide lexical choice in
machine translation. Translation of a Hebrew verb-object pair such as lahtom (sign or seal) and hoze
(contract or treaty) is determined using the most probable combination of words in an English
monolingual corpus. This work showed that leveraging bilingual lexicons and monolingual language
models can overcome the need for aligning bilingual corpora.

One research by [57] presented a model that represents word meaning in context by vectors which
are modified according to the words in the target’s syntactic context. Contextualization of a vector is
realized by reweighting its components, based on distributional information about the contexts.
Vector-space models of meaning lend themselves as a basis for determining a soft and gradual
concept of semantic similarity (for example, through the cosine measure), which does not rely on a
fixed set of dictionary senses with their well-known problems [58].

Inspired by earlier work of [59], who developed a network algorithm to extract context specific
vector representations for words in context, [60] investigated the systematic combination of
distributional representations of word meaning along syntactic structure. They used to represent the
meaning of a complex expression that consists of two syntactically related words w and w’ by a

27
vector obtained by combining the word vectors of w and w’, and find that component-wise
multiplication performs best for the task under consideration. They consider their proposal primarily
under the aspect of compositionality, but it can also be taken to be a method to contextualize a target
word through its dependents.

A different approach has been taken by [61] and Reisinger and [62]. Instead of “refining” vector
representations ranging over all words in a corpus by means of vector composition, they start out
from “token” vectors for individual instances of words in context and then group these token vectors
into different sense specific clusters. Once features are selected, SenseClusters creates a vector for
each test instance to be discriminated where each selected feature is represented by an entry/index.
Each vector shows if the feature represented by the corresponding index occurs or not in the context
of the instance (binary vectors), or how often the feature occurs in the context (frequency vectors).
This is referred to as a first order context vector, since this representation directly indicates which
features make up the contexts. Here we are following [63], who likewise took this approach to
feature representation. [25] utilized second order context vectors that represent the context of a target
word to be discriminated by taking the average of the first order vectors associated with the
unigrams that occur in that context. In SenseClusters we have extended this idea such that these first
order vectors can also be based on co–occurrence or bigram features from the training corpus.

Rule based systems exploit the hand crafter rule for WSD task. The rule based systems require
extensive work of expert linguists and thus can result in near human accuracy. The research done by
[64] Afan Oromo rule based Afan Oromo Grammar Checker, shows a promising result. The finding
of the study shows that rule based is an approach used in the morphologically rich language like
Afan Oromo. This rule based approach for languages, such as Afan Oromo, advanced tools has been
lacking and are still in the early stages. The rule based method is advisable for under resourced
languages.

One research by [65] the result of the study shows that the limited availability of resources in the
form of digital corpora and annotated, the rule based method is applied. All senses are discovered
using a set of rules and knowledge base for later use in the disambiguation process. The finding
shows an improvement in assigning correctly the corresponding disambiguation over the baseline
method. All the generated information is stored in the Sense Knowledge base. Since this information
can affect the accuracy of the disambiguation process.

28
According to [66], there are three important characteristics of an ambiguous word: grammatical
information about the ambiguous word to be disambiguated, words that are syntactically related, and
words that are topically related to the ambiguous word. The semantic and syntactic information to
disambiguate an ambiguous word, consideration is taken in the first two types. For each entry of the
Sense Knowledge base, it consists of the following: ambiguous and related words, sense of the
ambiguous and related words, and part of speech of the related word. Moreover, this information can
be converted into an understandable rule format that best describes the relationship between, the
ambiguous word and the related word. The rules can be easily constructed to describe the
relationship between, the ambiguous and related words.

According to [67] the automatic generation and the evaluation of sets of rules for word sense
disambiguation (WSD). The ultimate aim is to identify high-quality rules that can be used as
knowledge sources in a relational WSD model. The knowledge-based approaches which depend on
the manual encoding of accurate linguistic knowledge and disambiguation rules.

29
 Chapter Three

3. Overview of Afan Oromo

3.1. Introduction

Afan Oromo, also called Oromiffaa or Afaan Oromoo, is a member of the Cushitic branch of
the Afro-Asiatic language family [68]. It is the third most widely spoken language in Africa,
after Hausa and Arabic. Its original homeland is an area that includes much of what is
today Ethiopia, Somalia, Sudan and northern Kenya and some parts of other East African countries
[69]. Currently, it is an official language of Oromia Regional State (which is the biggest region
among the current Federal States in Ethiopia). It is used by Oromo people, who are the largest ethnic
group in Ethiopia, which amounts to 50 % of the total population in 2007[70]. With regard to the
writing system, Qubee (a Latin-based alphabet) has been adopted and become the official script of
Afan Oromo from 1991 [10].

Among the major languages that are widely spoken and used in Ethiopia, Afan Oromo has the
largest speakers [69]. It is considered to be one of the five most widely spoken languages among the
roughly one thousand languages of Africa [71]. Afan Oromo, although relatively widely distributed
within Ethiopia and some neighboring countries like Kenya, Tanzania and Somalia, is one of the
most resource scarce languages [72]. It is part of the Lowland East Cushitic group within the
Cushitic family of the Afro-Asiatic phylum [69], unlike Amharic (an official language of Ethiopia)
which belongs to Semitic language family. Although it is difficult to identify the actual number of
Afan Oromo speaking societies (as a mother tongue), due to lack of appropriate and current
information sources, according to the census taken in 2007 it was estimated that 50 % of Ethiopians
are ethnic Oromo [73].

It is widely used as both written and spoken language in Ethiopia and neighboring countries [73].
Currently it is used as an instructional media for primary, junior secondary schools and Ethiopian
Universities in the region and out of the region like Mekele, Dilla and Arbamich Universities. It is
also given as a subject starting from grade one to high school throughout the schools in Oromia
region. It is also broadcasted via televisions and radio stations locally and internationally.
Furthermore, few literature works, newspapers (Bariisaa, Kallacha Oromiyaa and Oromiyaa),

30
magazines, educational resources, official documents and religious writings are written and
published in this language [10].

3.2. Language Structure

Afan Oromo has a very rich morphology like other African and Ethiopian languages [72]. With
regard to the writing system, Qubee (Latin-based alphabet) has been adopted and became the official
script of Afan Oromo since 1842 [72]. The writing system of the language is straightforward, which
is designed based on the Latin script. Thus letters in the English language are also in Afan Oromo
except the way it is spelled. A detailed description of Afan Oromo Writing System can be found in
any text related to the language, but [74] discussed writing system of the language.

3.3. Alphabet (Qubee Afaan Oromoo)

Afan Oromo uses Qubee (Latin based alphabet) that consists of thirty three basic letters, of which
five are vowels, twenty-four are consonants, out of which seven are paired letters and fall together (a
combination of two consonant characters such as ‘ch’). The Afan Oromo alphabet characterized by
capital and small letters as in the case of the English alphabet. In Afan Oromo language, as in
English language, the vowels are sound makers and are sound by themselves. Vowels in Afan
Oromo are characterized as short and long vowels. The complete list of the Afan Oromo alphabets is
found on the manuscript by [68] and [75]. The basic alphabet in Afan Oromo does not contain ‘p’,
’v’ and ‘z’, because there are no native words in Afan Oromo that formed from these characters.
However, in writing Afan Oromo language, they are used to refer to foreign words such as “polisii”
(“police”).

3.4. Sentence Structure

Afan Oromo and English are different in sentence structuring. Afan Oromo uses subject-object-verb
(SOV) language. SOV is the type of language in which the subject, object and verb appear in that
order. Subject-verb-object (SVO) is a sentence structure where the subject comes first, the verb
second and the third object. For instance, in the Afan Oromo sentence “Mooneeraan bilisa bahe”.
“Mooneeraa “is a subject, “bilisa” is an object and “bahe” is a verb. Therefore, it has SOV structure.
The translation of the sentence in English is “Mooneeraa has got freedom” which has SVO structure.
There is also a difference in the formation of adjectives in Afan Oromo and English. In Afan Oromo

31
adjectives follow a noun or pronoun; their normal position is close to the noun they modify while in
English adjectives usually precede the noun. For instance, namicha gaarii (good man), gaarii (adj.)
follows namicha (noun).

3.5. Adjectives

An adjective is a word which describes or modifies a noun or pronoun. A modifier is a word that
limits, changes, or alters the meaning of another word. Unlike English, adjectives are usually placed
after the noun in Afan Oromo. For instance, in Qotiyyoo diimaa gate “ He lost red Ox” the adjective
diimaa comes after the noun Qotiyyoo [10].

3.6. Articles

Afan Oromo does not require articles that appeared before nouns, unlike that of English. In English
there are three main semantic choices for article insertion: definite article (the), indefinite article (a,
an, some, any) and no article. In Afan Oromo, however, the last vowel of the noun is dropped and
suffixes (-icha,-ittii,-attii) are added to show definiteness instead of using definite article. For
example, “the man” is “namticha” to indicate certainty [68].

3.7. Punctuation Marks

Punctuation marks used in both Afan Oromo and English languages are the same and used for the
same purpose with the exception of the apostrophe. Apostrophe mark (‘) in English shows
possession, but in Afan Oromo it is used in writing to represent a glitch (called hudhaa) sound. It
plays an important role in Afan Oromo reading and writing system. For example, it is used to write
the word in which most of the time two vowels appeared together like “ba’e” to mean (“get out”)
with the exception of some words like “ja’a” to mean “six” which is identified from the sound
created. Sometimes apostrophe mark (‘) in Afan Oromo interchangeable with the spelling “h”. For
instance, “ba’e”, “ja’a” can be interchanged by the spelling “h” like “bahe”, “jaha” respectively still
the senses of the words is not changed.

32
3.8. Conjunctions

Conjunctions are used to connect words, phrases or clauses. In Afan Oromo there are different words
that are used as a conjunction. Conjunctions in Afan Oromo are “fi”, “haa ta’u malee”, “garuu”,
“akkasumas”. Example: Amboo, Finfinnee fi Jimmaan magaalota Oromiyaati.(Ambo, Finfine and
Jimma are Oromian cities.)

3.9. Word and Sentence Boundaries

In Afan Oromo, like in other languages, the blank character (space) shows the end of one word.
Moreover, parenthesis, brackets, quotes are being used to show a word boundary. Furthermore,
sentence boundaries punctuations are almost similar to English language i.e. a sentence may end
with a period (.), a question mark (?), or an exclamation point (!) [76].

3.10. Word Segmentation

The word, in Afan Oromo “jecha” is the smallest unit of a language. There are different methods for
separating words from each other. This method might vary from one language to another. In some
languages, the written or textual script does not have whitespace characters between the words.
However, in most Latin languages a word is separated from other words, by white space characters
[77]. Afan Oromo is one of Cushitic family that uses Latin script for textual purpose and it uses
white space character to separate words from each other’s. For example, “Bilisummaan Finfinnee
deeme”. In this sentence the word “Bilisummaan”,” Finfinnee” and” deeme” are separated from each
other by white space character. Therefore, the task of taking an input sentence and inserting
legitimate word boundaries, called word segmentation, is performed using the white space
characters.

3.11. Homonymy

A homonym is a group of words sharing same spelling and pronunciation, but have different senses.
These multiple senses can be totally unrelated to each other. Homonymy is defined as a relation that
holds between words that have the same form with unrelated meanings. The items taking part in
such a relation are homonyms [16]. Homonyms are those lexical items with the same phonological

33
form but with different meanings which will cause ambiguity. It can be illustrated with the following
example:

Afaan koo faaya kooti My language is My beautifulness, My mouth is my beautifulness

In the example the underlined “Afaan” is an ambiguous word having the same pronunciation and
spelling but different meaning. This notation indicates that these are two separate lexemes, with
distinct and unrelated meanings, that happen to share an orthographic form.

3.12. Synonym

Polysemy is the special case of homonymy where multiple senses of the word are related to each
other [79]. The phenomenon of synonymy is sufficiently widespread to account for the popularity.
The notion of synonymy, has a deceptively simple definition different lexeme with the same
meaning. The synonyms can substitute for one another in a sentence without changing either the
meaning or the acceptability of the sentence [16].

As an example: Seenaan gargaarsa naaf godhe. “Sena helped me”.

The underlined word gargaarsa can be substituted by the word “deeggarsa” which is not used in this
sentence. These two words can each other without changing the meaning of the sentence. By this,
two lexemes share a central core meaning.

3.13. Polysemy

The phenomenon of a single lexeme with multiple related meanings is polysemy. The notion of
polysemy allows to state that sense is related to, and possibly derived from, without asserting that it
is a distinct lexeme [16]. For example: Qorachuun mala dhayiiti. The study is the way of putting
solution.

34
Here the word underlined Qorachuun which mean Iyyaafachuu “call out” and gaafachuu “ask” are
two related meanings according the contexts of the sentence. As one suspect, the task of
distinguishing homonymy from polysemy is not quite a straightforward. There are two criteria that
are typically invoked to determine whether or not the meanings of two lexemes are related or not:
the history, or etymology, of the lexemes and how the words are conceived by native speakers [16].
In the absence of detailed etymological evidence, a useful intuition to use in distinguishing
homonymy from polysemy is the notion of coincidence. On the other hand, it is far more difficult to
accept cases of polysemy as coincidences.

35
 Chapter Four
4. Methodology
4.1. Introduction

This chapter describes the methods employed in this research. In order to develop disambiguation
model for Afan Oromo we followed three steps process which involve: (a) text preprocessing which
take input and corpus, tokenize to remove stop words and perform normalization (b) extract context
terms providing clue about the senses of the ambiguous term using two techniques (window size and
rule based), (c) clustering to group similar context terms of the given ambiguous terms, the number
of clusters representing the number of senses encoded by the ambiguous term. In order to cluster
similar context terms we computed the degree of similarity using the vectors constructed from co-
occurrence information. The architecture (pipe line) of the system with the underlying steps is
presented in figure 4.1 below:

Figure 4.1. The Architecture (pipe line) of the System

36
The description of the corpus preparation procedures followed by the various algorithms employed
are described in Section 4.3 and Section 4.6 respectively.

4.2. Method

To achieve the objectives stated in Section 1.3, this study relies on the patterns learned from the
corpus (unsupervised approach) in combination with the manually crafted rules for clustering similar
contexts of ambiguous word and extracting the contexts respectively. The motivation behind the use
of hybrid is mainly aroused form the fundamental problem of corpus-based approach in relation to
the sparseness of the training contexts. The idea of this research is therefore to combine both the rule
based and unsupervised machine learning approaches into a hybrid approach. Such method of word
sense disambiguation as also employed in this work, combine the advantages from machine learning
and rule based, potentially yielding better results. Hence, the reason why the hybrid approach was
used is taking the availability and reliability of linguistic knowledge on the top of the semantic
techniques and training methods learned from corpus for learning the role of the words in its context.
We have described the techniques employed under each task in the following sections.

4.3. Corpus

This section deals with the collection of data and preprocessing the collected data. For this study, we
have prepared new corpus for the WSD of Afan Oromo. The size of the corpus is an important factor
in the quality of the disambiguated senses, with the general message that more data is better data
[78]. However, in such a work, it is very difficult to obtain standardized dataset for under resourced
language. The procedure for collecting and preprocessing corpus is described here below.

4.3.1. Corpus Acquisitions and Preparations

The corpus acquisition and preparation are set of techniques required for gathering and compiling
data for training and testing algorithms. The lack of resource has led researcher to explore the use of
unannotated raw corpus to perform hybrid disambiguation. It should be noted that unsupervised
disambiguation cannot actually label specific terms as a referring to a specific concept that would
require more information than is available.

37
The mechanism to acquire resource is to use the data from various sources and hence it is not
previously used in any research on Afan Oromo. The collected data is machine readable free text. It
is collected from newspapers (Bariisaa, Kallacha Oromiyaa and Oromiyaa. Bariisaa is a weekly
newspaper, whereas the rest two come out once in two weeks), bulletins, news (ORTO), government
and official websites. Moreover, ORTO found in Adama releases daily news through radio and
television broadcast and on its official website [10]. To reduce the data sparsity we have used data
from these sources since they are believed to represent texts addressing various issues of the
language. Actually, the collected data was not directly used for the purpose. We have applied
preprocessing (tokenize, stop word and normalize) techniques before it used. Hence, the acquired
corpus is a free sentence of the corpus, we have prepared and filtered to make as it ready for the
study in the following Sections.

4.3.2. Corpus Preprocessing

The preprocessing step includes tokenizing, stop-word removal and normalization (change
characters to the same form). For stop word removal, we have used the Afan Oromo stop-word
compiled from different text in addition to the stop-word list prepared by [79]. In case of the
normalization, we have simply convert the corpus into smaller case (the same case), hence it has no
effect on the meaning of the words.

A. Tokenization

The corpus which is a set of sentences first tokenized into words. Since, Afan Oromo uses Latin
alphabet the sentences can split using similar word boundary detection techniques like the use white
space in English. In this work tokenization is needed for some purposes that are to select contexts,
stop word removal and for further steps. For instance if we have the sentence akkanatti hin haftu ni
dabarti in our corpus. We have tokenized into set of words on the white space, like akkanatti, hin,
haftu, ni and dabarti. After tokenization we will apply our algorithm to extract contexts among the
tokens as discussed in the coming sections (Section 4.6.1.1).

38
 B. Stop Word Removal

After tokenization take place, we have removed Afan Oromo stop words, hence it has no effect on
meaning of the words. In this work, Stop word removal is used to remove stop words from the
selected contexts because the absence or presence of these words has no contribution to identify
appropriate sense. Not all tokenized words are necessary for this work hence, one word carry the
meaning than other words and other words that have no own meaning. For instance, words such as
(‘as’, ‘achi’), conjunctions(“’fi’, ‘akkasumas’, ‘kana malees’). Since stop words do not have
significant discriminating powers in the meaning of ambiguous words; we filtered stop-words list to
ensure only content bearing words are included. Few list of stop words in Afan Oromo is shown in
the table 4.2 below:

Stop Words Meaning (English)

Kana This

Achi There

Yookan Or

Haa ta’u malee However

Irra On

Jala Under

As Here

Kanaaf So

Jedhama Called

Ammo However, but

Garuu But

Bira Beside, at, near of

Ala Outside, out

Akka Such as, like, according to

Table 4.2. Stop Words (source: unpublished [80])

39
 C. Normalization

In this work, some characters of the same words are sometimes represented in uppercase or
lowercase in the corpus as well as in the user input and hence we have normalized them into
lowercase. The purpose of normalization in this case is to make similar the words in different cases
in our corpus. This is the observation that many different strings of characters often convey
essentially identical meanings. Given that to get at the meaning that underlies the words, it seems
reasonable to normalize superficial variations by converting them to the same form. The most
common types of normalization are case folding (converting all words to lower case). Case folding is
easy in Afan Oromo for example Qabsoo similar to qabsoo.

4.4. Implementation Tools

To perform this experiment, the algorithms were implemented in Java, NetBeans 8.0.2 which run on
the prepared corpus and the clustering were performed in Weka 3.6.5 tool. Java package is a general-
purpose and open source programming language. Moreover, it is optimized for software quality,
developer productivity, program portability, and component integration. Nowadays Java is
commonly used around the world for Internet scripting, systems programming, user interfaces,
product customization, numeric programming, and more. It is generally considered to be among the
top four or five most widely-used programming languages in the world.

The Weka 3.6.5 is the other tool used to implement WSD specifically context clustering, using the
hierarchical clustering (which are single link, complete link and average link) and partitional
clustering (EM, K-means) algorithms built into the tool. The tool has by default Euclidian distance
which is used to measure the distance between clusters.

4.5. Testing

The WSD system was trained on unannotated corpus, which constitutes ambiguous words. The work
was tested by using most frequent 15 ambiguous words collected from the public via questions (in an
Appendix A) from random native speakers of the language. To this end, we collected the words from
20 individuals and selected the top 15 frequent ambiguous terms. The senses encoded by the words
in the test set are determined from the information obtained from the language experts. The selected
ambiguous words are listed below:

40
No Ambiguous Possible Senses Defined Number
Word of Senses

1 Afaan Language / Mouth 2

2 Bahe Freedom / Highland / Cloth / Witness / Dead(pass) 5

3 Boqote Break (rest) / Died 2

4 Darbe Cross/ Pass from class to class / Died / broadcast 4

5 Diige Fence / Absence on Meeting / cancel to start new 3

6 Dubbatate Struggle / Wedding 2

7 Tume Make / Hit / Contraceptive 3

8 Haare Sad / Burn 2

9 Handhuura Center / Gift / Navel 3

10 Ija Eye/ Tree Fruit / Vengeance/ Wide-eyed/ a little 5

11 Ji’a Stars / Month 2

12 Lookoo Pretty / Rope 2

13 Dhahe Follow / Hit /Fail 3

14 Mirga Direction / Human right / Brave 3

15 Waraabuu Hyena / Fetch / Record 3

Table 4.1. List of Ambiguous Words for testing

41
4.6. Algorithms

In our approach, two important features need to be extracted: the first one is determining all possible
contexts (the candidate sense words) of the ambiguous words and the other one is to group these
various contexts (senses) of the word, each group representing a specific sense of the ambiguous
word. To this end, we developed two kinds of approaches towards the word sense disambiguation.
The first approach is completely machine learning in its nature. Unsupervised machine learning
approach extracts the two important features (the various contexts of the ambiguous words and their
clustering) without supplying linguistic rule (Section 4.6.1.2). The second approach combined
unsupervised machine learning algorithm and rule based (background linguistic). Such algorithm is
called hybrid approach, which is a mixture of rule based (background knowledge) and corpus-based
approaches applied. In this latter approach we combined the rule learned from a linguistic feature of
Afan Oromo with the semantic feature learned from corpus using unsupervised machine learning
approaches. First, the manually crafted linguistic rule extract the various contexts assumed by the
ambiguous word used. We then relied on the unsupervised machine learning algorithm to cluster
these contexts as described in Section 4.6.1.4. Hence, this latter approach extracts all possible
contexts assumed by the ambiguous word employing rule-based approach, which make use of
linguistic knowledge, whereas the corpus-based approaches use information acquired from the
corpus; and the hybrid approach merges characteristics from both approaches.

The machine learning approach followed towards the WSD is basically unsupervised and mainly
used for clustering the possible contexts for giving words that express the same meaning. Hence we
didn’t provide explicit sense labels for each group as the machine learning approach is unsupervised.
Yet, small list of ambiguous words are required to test the algorithm.

4.6.1. Machine Learning Approach

As discussed in the above (Section 4.6), two important features need to be extracted. We have
described the procedures followed to extract the features in the subsequent Sections.

42
 4.6.1.1. Extracting the Context Words

In the machine learning, the surrounding contexts were extracted by sliding window of n words. The
reason why the contexts extracted is that, the words that occur in similar contexts tend to have
similar meanings. This idea, known as the ‘distributional hypothesis’, has been proposed by various
scholars. In order to extract the contexts from a set of sentences the role of window is great. Because
a word's meaning cannot be fully grasped unless one takes the context into account. Meaning and
context can be captured in terms of (more or less direct) neighborhood, i.e. words co-occurring.

Most of WSD algorithms make use of the contexts to provide information for sense disambiguation.
As already mentioned in this approach, we followed automatic procedure to extract all possible
contexts assumed by the ambiguous word. This is the idea of the study, which is context based
meaning of words. We have determined a set of contexts based on the frequency co-occurrence of
context words in the corpus with target words, by determining a window size contexts to the left and
to the right.

N – size (left side) N + size (right side)

[Sabni cunqurfame fi dararame afaan dheebuun bilisuummaa eega].

Once the context words are extracted, the next step cluster similar contexts based on their inherent
semantics, the number of the cluster representing the number of senses assumed by the ambiguous
word. The procedure employed for clustering the context is described below.

4.6.1.2. Clustering the Contexts

For each context extracted in Section 4.6.1.1 above, we constructed a vector space matrix from co-
occurrences. After the co-occurrence matrix, the cosine similarity was computed based on the angle
between vectors of the contexts.

43
These cosine similarity values were used to cluster similar contexts by using Weka 3.6.5 tool. The
cosine similarity result which is in .CSV (Comma Separated Value) file format was entered into the
Weka tool. The Weka tool cluster the contexts based on cosine similarity result. Each cluster
represents a unique sense. To this end, we used hierarchical (single, complete, and average link)
clustering, and (K-means, EM) algorithms from partitional clustering to merge the contexts based on
their cosine values. The vector space model and the clustering algorithms are described in Sections
4.6.1.3 and 4.6.1.4 respectively.

4.6.1.3. Vector Space Model

The vector space model is used to represent the meaning of a word, hence it used to measure
semantic similarity. The rationale behind the vector space model is that, words that occur in similar
contexts tend to have similar meanings (as it described earlier). We construct vectors for each
extracted context terms of the ambiguous word for measuring their similarity. For the given pair of
context terms the algorithm extract words co-occurring in a predefined window. After extracting the
term co-occurrences we computed weight based on their frequency. We did the same for the other
pair of the context term. In order to identify the similarity between the context words the cosine
similarity of the vectors were computed. The cosine similarity of the vectors is then considered to
determine the semantic similarity between the terms. It is preferable because it takes into account
variability of data and features' relative frequencies as well as useful to intend with cluster as it
captures similarity overall.

To illustrate this with a simple example: consider a co-occurrence matrix populated by simple
frequency counting: if the term i co-occurs 5 times with term j in the corpus, we have used 5 in the

fij(frequency of term i and j) to compute the cosine similarity. The co-occurrences are normally

counted within a context window spanning some number of words. In our case, the vectors are
extracted by taking words regardless of their passion, hence, it doesn’t consider the ordering of the
words in the corpus. It uses the angle between vectors instead of distance in order to measure the
similarity between the context terms.

44
 ……………………………..1

Where SimCos – is cosine similarity

x – is the dimension of the vector one
y – is the dimension of the second vector

4.6.1.4. Clustering Algorithms

As already mentioned, we cluster context terms of the ambiguous words using their similarity values
produced in Section 4.6.1.3. The clustering algorithms used in this study are hierarchical
agglomerative clustering, which include single link, complete link, average link and EM and K-
means clustering from partitional clustering (see details in Section 2.6). These clustering algorithms
have their own unique nature. The hierarchical clustering begin by assuming that each context of an
ambiguous word forms its own cluster (and therefore represents a unique sense). Then, it merges the
contexts that have the minimum dissimilarity between them (and are therefore most alike). The
partitional clustering algorithms started by partitioning into predefined k sizes [82]. It found the one
which is the nearest to initial centroid. A centroid is usually not an element of the cluster. Rather, it
represents the center of all other elements. Often the mean of the feature vectors of the elements
within a cluster is used as that cluster centroid. The minimum specified cutoff which determines the
number of clusters is taken. In this case, the minimum specified cutoff of the number of clusters is
two hence one ambiguity word has at least two senses.

4.7. Hybrid Approach

According to [22] WSD heavily relies on knowledge of machine learning and linguistics. The hybrid
approach in this work, constitute the unsupervised machine learning approach to cluster the contexts
followed by hand crafted rule to extract the modifiers of the ambiguous word as described in the
following Sections.

45
4.7.1. Constructing Rule for Extracting Contexts

The linguistic knowledge of the language plays an important role to create the rule. The linguistic
knowledge required for the natural language can be obtained in different ways. In this study, the
rules were created based on the inherent structure of Afan Oromo in forming meaning of words.
However, an effort has been to develop the rule of the language (more details in Section 4.7.1.2).

The most common way of representing language is via rules [68]. The rule underlies many linguistic
theories of the language, which turn into a set of rules [82]. The modifiers and contextual
information were the basis of the linguistic properties of Afan Oromo word meaning. In this study,
the modifier of the ambiguous word is used in order to get the semantic clues of the particular sense
in which the ambiguous word is used. To achieve this we analyzed the structure of Afan Oromo
sentence formation with respect to modifier patterns to develop the rule. The constructed rule used
for extracting all modifiers modifying the ambiguous term. The modifiers are word or phrase which
provide information about a word and give more description about the words it modify. The
modifiers (can be single word or phrase) establish understanding of the ambiguous word.

In Afan Oromo, the meaning of words fundamentally based on the words preceded by the word
(modifies) [83] [84]. Hence, the words are described (modified) by the Noun or Verb preceding
them. However, since there is no annotated corpus of this language we didn’t identify whether this
assertion is holds true or not. The ambiguous word may appear in the middle or the end of a
sentence, but modifiers always becomes before the word it modifies. As an example, “Seenaan
kaleessa daara bahe.” [Seenaa got cloth yesterday]. From the construction of this sentence, the word
“bahe” is modified by the Noun “daara”. According to Afan Oromo structure, the Noun and Verb
always appear before the word they modify [66].

4.7.1.1. Modifiers

In Afan Oromo like other languages, the word meaning has own its rule. However, the correct sense
cannot be only found by choosing the one that is related to another. Promising techniques relied on
linguistic knowledge also for extracting semantic features, in our case to mine context of the
ambiguous term via of modifiers specializing its meaning [72].

46
The modifiers have a great role to decide on the word meaning according to its role in the sentence.
The modifiers can appear before the target word (the word it modify or describe). Like English, the
sentences would be pretty boring without modifiers to provide excitement and intrigue. A modifier
adds detail or limits or changes the meaning of the other word or phrase.

In Afan Oromo, the words preceding a specific word are more likely to influence the meaning of a
word.

For example, [Finfinneen magaala Oromiyaati eenyumatti hin kenninu].

Finfine is Oromian city nobody will take it from Oromia.

In this example, the core of the sentence is “[Finfinneen magaala Oromiyaati

eenyumatti hin kenninu] Finfine is Oromian city nobody will take it from Oromia”.
The word “[Finfine]” is a modifier; it gives an extra information that is part of the
sentence. In this case, it is a Noun modifier, because it is modifying the ambiguity
word “magaala”. A modifier should be placed next to the word it describes.

4.7.1.2. Modifiers Rule

Disambiguation is done by analyzing the linguistic features of the word and its preceding word. The
rule-based section of our approach disambiguate word automatically using rule in order to
complement the features learned from training data. This information is coded in the form of rules.
As it discussed in the Section 4.7.1.1, the modifiers always precede the target word in Afan Oromo.
Based on this notion, the rule developed was as follows:

 IF ambiguous word preceded by Modifiers THEN collect the modifiers to disambiguate.

 IF ambiguous word is NOUN, the modifiers immediately following ambiguous word
 IF ambiguous word was VERB, the modifiers immediately preceding ambiguous word

47
4.8. Evaluation Method

To our knowledge, there were no previous standard Afan Oromo word sense disambiguation dataset
for evaluation as presented in Section 4.5. For this reason we did not evaluate against the other
systems. In this work, the evaluation were undertaken on the basis of precision and recall. Precision
is defined as the percentage of correctly disambiguated words out the total of disambiguated words.
Recall is defined as the percentage of correctly disambiguated words out of the total number of
ambiguous words [85].

# correctly disambiguated words

Precision(%) = ___________________________

# disambiguated words

# correctly disambiguated words

Recall(%) = ___________________________

# total number of ambiguous words

On the other hand, the clustering algorithms were evaluated comparing the result produced by the
clustering algorithm with the manually grouped similar contexts of the ambiguous words in the test
set by experts. The evaluation constitutes the following two points:

1. To evaluate how much the produced clusters are comply with the clusters prepared by human
experts as a benchmark..

In order to achieve this we used the following criteria:

 How many of the clustered contexts are correct, i.e. to evaluate if all the similar contexts
of the ambiguous words are placed in the same group.

2. Given the number of senses assumed by the ambiguous words in the test, judge the system on
the basis of the number of senses identified by the system.

48
Similarly, in order to achieve this the following steps performed:

a. Start with a small list of ambiguous words in the test with known number of senses N.
b. Run the algorithm on the test to identify the possible senses based on it’s the number of
clusters of the context as extracted from the big corpus
c. Count the number of clusters
d. Compare it against the already prepared sense clusters by experts

49
 Chapter Five

5. Experimentation Result, Discussion and Evaluation

5.1. Introduction

In this chapter we have described the training data (Section 5.1.1), the experimental procedure and
the result of the unsupervised machine learning and the hybrid WSD approach in Section 5.2, 5.3
respectively.

The following figure shows the user interface of the sense disambiguation used during the
experimentation.

Figure 5.1. User Interface of the WSD

To disambiguate the given ambiguous word, it takes one ambiguous word at a time in the interface
provided and produce the cosine similarity to cluster.

50
5.1.1. Data

In this work, two types of data were used: (a) the small list of ambiguous words to test the
algorithms. Fifteen (15 ) highly frequent ambiguous words were selected from the language speakers
using the procedure described in Section 4.5, (b) the big corpus containing thousands of sentences to
extract contexts and its vectors to represent the group of contexts. Hence, the compiled corpus for
the experiment is a free text. It is composed of a set of 57,062 lines of sentences and 412,366 words.
This corpus is prepared for the purpose of this study as there is no standard corpus for Afan Oromo
language. However, it is unlabeled data and never used in any researcher and it needs to be
developed further (as discussed in Section 4.3.1).

5.2. Experiment on the Machine Learning Approach

As mentioned in chapter four, the machine learning approach uses a window of N words surrounding
the ambiguous word to extract contexts. The extracted context words are then clustered together
based on their similarity values using clustering algorithms and vector space model respectively.

In the following section, we have presented the experimental procedure and result on various
windows size and suggested the optimal window size for Afan Oromo.

5.2.1. Experiment on the Context Window Sizes

Defining an optimal window size plays a crucial role to identify the information surrounding the
ambiguous word. In English a standard window of two-words on either side of the ambiguous word
is found to be enough for disambiguation [66]. Similarly, for Afan Oromo WSD, two window size
on both sides for the ambiguous word is found to be enough. Therefore, this has been established for
Afan Oromo WSD. The investigated experiment shows that the window sizes around the ambiguous
word influences WSD performance with various window sizes.

The system provides context terms to the left and right side of ambiguous word (excluding stop
words) based on the provided window size. In this experiment, window size of up to ±10 words on
both sides of ambiguous word have been tried for Afan Oromo. Starting with the first term in the test
set, extract words appearing N-words left and right of the ambiguous words. In order to measure the
performance of the window size for the disambiguation, we run the disambiguator using all the

51
window sizes (starting from 1 to 10) and observe the differences in the performance of the
disambiguator.

As evidenced by the experiment, differences in the window sizes yield different results. The result
has proved that window of two-words to the left and right of the ambiguous word achieved the best
performance than other windows. This result converges with the result obtained by [66]. Table 5.1
below shows the performance of the disambiguate using different window sizes.

Window Size (N) Accuracy (%)

1 73.34

2 66.67

3 60.0

4 55.5

5 51.6

6 46.3

7 41.1

8 36.0

9 31.8

10 28.6
Table 5.1. Determining Optimal Window Size

As can be seen from table 5.1, a narrow window of context, one and two words to either side
(73.34% and 66.67% respectively), was found to perform better than wider windows (28.6%). The
accuracy is conducted by measuring the performance of WSD with varying the window size (the
tested ambiguous word found in sentences). It is very likely that smaller window sizes have yield
significantly higher accuracy than other windows and different windows gave different results. From
this experiment, we conclude that smaller window sizes usually lead to accuracy while bigger
window sizes relatively low accuracy.

52
5.2.2. Experiment on the Clustering

The underlying idea of using a clustering algorithm is to cluster contexts which are providing a
useful way to discover semantically related senses and the same sense of a word has similar
neighboring words. In this case, the clustering algorithms directly use the vector representations
(cosine similarity measure) of the contexts of the ambiguous word (extracted using rules or the
window of words) as input. The basic point of sense clustering is that related word senses must be
associated and grouped in the same cluster. Hence, it identifies groups of senses, which are assumed
to represent different meanings. The Weka clustering tool is used to group similar contexts together
assuming, the induced clusters represent the different senses of the ambiguous word.

The contexts of the ambiguous words are clustered using three hierarchical clustering algorithms and
two partitional clustering algorithms. Each and every of clustering algorithm has its own measures of
similarity so as the cluster formed by the different clustering algorithms were produce different
result. One algorithm may make 5 clusters the other may define only 2 cluster groups. The number
of clusters formed by the clustering algorithms depends upon the uniqueness and the dissimilarity
present in the cosine similarity.

An important point here is how to decide which constitutes good clustering, since it is commonly
acknowledged that there is no absolute best criterion which would be independent of the final aim of
the clustering. Consequently, it is the researcher who supply the criterion that best suits their
particular needs and the result of the clustering algorithm can be interpreted in different ways (see
detail in Section 4.6.1.4). One approach is to group data in an exclusive way, so that if a certain item
of data belongs to a definite cluster, then it could not be included in another cluster. Another
approach, so-called overlapping clustering, uses unclear sets of cluster data in such a way that each
item of data may belong to two or more clusters with different degrees of membership.

Given a set of context words of the ambiguous word, with the cosine similarity values corresponding
to their similarity, the hierarchical clustering begins by setting each item of cosine as a cluster and
proceeds by uniting the two nearest clusters. After a few iterations it reaches the final clusters
wanted. The partitional clustering is partitioning the contexts into k clusters based on the closeness to
the cluster centers and finding the maximum probabilistic.

53
5.2.3. Experimental Results

As already mentioned, the evaluation of the clustering algorithms is from two dimensions (refer
Section 4.8 for more detail):

a. Evaluate the accuracy of the clustering on the basis of the number of clusters produced
b. On the bases of the correctness of the produced clusters

We tested the algorithm on fifteen natural ambiguous words as shown on Table 5.2 below.

Accuracy (%)

Ambiguous Single Complete Average K-Means EM

Word
Link Link Link

Afaan 50% 50% 50% 50% 50%

Bahe 60% 60% 60% 60% 80%

Boqote 50% 50% 50% 50% 50%

Darbe 50% 50% 50% 50% 50%

Diige 50% 50% 50% 66.67% 66.67%

Dubbatate 50% 50% 50% 50% 50%

Dhahe 66.67% 66.67% 66.67% 66.67% 66.67%

Haare 50% 50% 50% 50% 50%

Handhuura 66.67% 66.67% 66.67% 50% 66.67%

Ji’a 50% 50% 50% 50% 50%

Tume 66.67% 66.67% 66.67% 66.67 66.67%

Lookoo 50% 50% 50% 50% 50%

Ija 40% 40% 40% 60% 80%

Mirga 50% 50% 50% 66.67% 66.67%

Waraabuu 66.67% 66.67% 66.67% 66.67% 66.67%

Table 5.2. Disambiguation using evidence from Machine Learning

54
The average accuracy for test terms were 56.2% for the machine learning approach. Thus, in total
there are 15 ambiguous words to be discriminated, 6 words with 2 senses, 6 words with 3 senses, 2
words with 5 senses, and 1 word with 4 senses. Five different configurations of clustering are run for
each word, leading to a total of 75 experiments. The results are shown on Table 5.2. As indicated on
table 5.2, when the contexts of the word senses that are not similar are clustered together resulting in
low performance and vice versa.

No Clustering Algorithms Accuracy (%)

1 Single Link 54.4%

2 Complete Link 54.4%

3 Average Link 54.4%

4 K-Means 56.9%

5 EM 60.7%

Table 5.3. The Accuracy of the Machine Learning

The test by unsupervised machine learning method, that deals with clustering of contexts for a given
word that express the same meaning. The simple K-Means and EM clustering algorithms achieved
much accuracy on the task of WSD for selected ambiguous word. The partitional clustering which
include K-means and EM resulted 56.9 % and 60.7 % respectively achieved performance in
clustering.

55
5.2.4. Summary of the Machine Learning Approach

Within Time
Name of the No of Cluster No of clusters taken Log Unclustered
Algorithms Clusters Instances Iterations sum of to likelihood Instances
squared build
errors model

Single Link 24 0 - (50%) - - 0 - -

1 - (50%)

Complete 24 0 - (70%) - - 0 - -
Link 1 - (30%)

Average Link 24 0 - (70%) - - 0 - -

1 - (30%)

K-Means 26 0 - (40%) 2 7.66736 0 - -

1 - (60%)

EM 28 0 - (70%) - - 0.4 26.37927 -

1 - (30%)

Table 5.4. Summary of Machine Learning Clustering Algorithms

As shown on table 5.4 the Weka experiment, the result of clustering algorithms varies with the
distance they used. The single link algorithm uses the shortest is the most similar senses. Contrary,
the complete link algorithm uses the maximum distance is the most similar senses and average
clustering use the average of both algorithms.

A good clustering algorithm was automatically discovered an approximately same number of

clusters as senses of that word. As can be seen from table 5.2, most of the ambiguous words have 2
to 5 senses. The experimental evaluation showed that K-Means and EM always lead to significantly
higher accuracy than hierarchical algorithms which suggests that partitional algorithms are well-
suited for clustering cosine similarity due to not only their relatively low computational
requirements, but also comparable or even better clustering performance.

We verify that our cluster estimate the number of clusters in a dataset by observing the experimental
results with varying number of k clusters for the parameter. As shown on table 5.3, K-means and EM
relatively considerable than other clustering algorithms. In the case of K-means, the results are

56
entirely dependent on the value of k i.e. the number of clusters. There is no way of knowing how
many clusters exist. The same algorithm can be applied to the same cosine value, which can may
produce two or three clusters. There is no general theoretical solution to find the optimal number of
clusters for any given data. For this reason, the result produced by algorithms cross checked with the
result by experts. EM is the analytical methods available to achieve probability distribution
parameters, in all probability the value of the variable is given.

As can be seen from table 5.3 above, out of the total clusters grouped by experts, the possible
clusters are produced by different algorithms that correspond to different senses. An important issue
in clustering is how to decide what is the best set of clusters for a given dataset, in terms of both the
number of clusters and the membership of those clusters. For this reason we compare the number of
clusters produced with the number of senses chosen by a group of human subjects (experts).
However, counting the number of clusters only doesn’t report about the accuracy of the system if
one doesn’t take the quality of the clusters into account. Hence, we have also evaluated how good
the clusters are by comparing them against the clusters produced by experts. Put, differently, we
measured the number of contexts assigned to correct and incorrect clusters (senses).

Based on the experiment, few of the clusters wrongly clustered with different contexts which has a
different sense, for instance the pair of contexts dhugaa and ragaa are appearing in different clusters
when clustering as shown in Figure 5.2 and Figure 5.3 below. The hierarchical clustering algorithms
are not much satisfying result because of they are on a limited amount of information (single link),
or they assume that all the cosine value in the cluster is very similar to each other at distance
(complete link). The average link measuring the similarity of both these clustering algorithms, as the
average of the pairwise similarity of the cosine value from each cluster.

5.3. Experiment on the Hybrid Approach

The second part of the experiment in this work is combining unsupervised machine learning and rule
based approaches. Unlike unsupervised machine learning in Section 5.2, the hybrid approach relies
on both machine learning and rule based, which is hand-constructed rules that are acquired from the
structure of language rather than automatically trained from data. The rule based WSD benefit from
various sources of linguistic knowledge [86]. This approach is used when there is a scarcity of data.
The rule depend on domain knowledge to bridges the gap caused by data scarcity.

57
Contrary, unsupervised machine learning which uses window size, the hybrid approach used rules to
extract modifiers of the ambiguous word and consider them as contexts. These modifiers are
therefore identified according to the developed rule planted. In the hybrid approach, the context
extraction process involves finding the distribution of all possible modifiers of the ambiguous words.
Differently, the window which varies window sizes to find the contexts, in the case of rule based it is
limited to the modifiers preceding ambiguous word. In this case, the modifiers aren’t varying the
sizes as window size.

The advantages of this approach is that it is easy to incorporate the linguistic knowledge [40]. In this
research, we try to answer “Does the addition of linguistic knowledge, improve a word sense
disambiguation performance for Afan Oromo ambiguous word?”. As the conducted experiment
shows the obtained result was more robust than unsupervised machine learning. In this case, with
worldly knowledge and reasoning, it is necessary to assist disambiguation. [22] certain knowledge
sources provide evidence as to the word sense used or disallowed by a particular context, but most of
them use linguistic knowledge. [87]the hybrid define a contextual representation is a characterization
of the linguistic contexts in which a word appears.

However, in this work the rule based approach does not independently used to disambiguate, rather
it is combined with unsupervised machine learning, just to mine the contexts of the ambiguous word.
The modifiers (Section 4.7.1.1) are identified accordingly the rule planted to it. Similar to
unsupervised machine learning, we have used the same test set in the hybrid approach. The
following sections show the experiment taken place by hybrid approach.

5.3.1. Effect of Hybrid Approach on Accuracy of Disambiguation

In this hybrid experiment, most of the tested ambiguous words have relatively higher performance
when compared with machine learning. As indicated in table 5.5 below, the hybrid approach result
the accuracy of nearly 73.9% on the nine of the test set and achieved poor performance on the
remaining of the test set. The reason behind the enhanced accuracy might be that, due to the nature
of unsupervised machine learning trained from free text corpus and hybrid method mix the
advantages from both methods especially from the rules. From defined number of senses by experts
for each term, the accuracy of disambiguation was very encouraging.

58
 Accuracy (%)
Ambiguous Single Complete Average
Word Link Link Link K-Means EM

Afaan 50% 50% 50% 50% 50%

Bahe 80% 60% 80% 80% 80%

Boqote 50% 50% 50% 100% 100%

Darbe 75% 75% 75% 75% 75%

Diige 66.67% 66.67% 66.67% 66.67% 66.67%

Dubbatate 50% 50% 50% 100% 100%

Tume 66.67% 66.67% 66.67% 66.67 66.67%

Haare 50% 50% 50% 50% 50%

Handhuura 66.67% 66.67% 66.67% 50% 100%

Ija 60% 60% 60% 80% 80%

Ji’a 50% 50% 50% 100% 50%

Lookoo 50% 50% 50% 50% 100%

Dhahe 66.67% 66.67% 66.67% 66.67% 66.67%

Mirga 66.67% 66.67% 66.67% 66.67% 66.67%

Waraabuu 66.67% 66.67% 66.67% 66.67% 66.67%

Table 5.5. Disambiguation using evidence from Hybrid Approach

As can be shown on table 5.5, an average accuracy of test set was 65.5% which is encouraging than
the unsupervised machine learning work. Table 5.5, shows the accuracies for each senses. We
calculated the accuracy for all such tests, which yielded disambiguation accuracy. From this

59
experiment table 5.5, hybrid approach finding the right number of clusters is equivalent to finding
the number of senses.

No Clustering Algorithms Accuracy (%)

1 Single Link 61%

2 Complete Link 59.7%

3 Average Link 61%

4 K-Means 71.2%

5 EM 74.6%
Table 5.6. The Accuracy of the Hybrid Approach
The accuracy of any clustering algorithm depends on the quality of the produced clusters. A good
clustering algorithm produces high-quality clusters such that inter-cluster similarity is minimized
and intra-cluster similarity is maximized as shown on table 5.6. A perfect clustering solution was the
one that leads to clusters that contain right contexts.

5.3.2. Summary of the Hybrid Approach

Within Time
clusters taken
Name of No of Cluster No of sum of to Log Unclustered
the Clusters Instances Iterations squared build likelihood Instances
Algorithm errors model
0: 50%
Single Link 28 1: 50% - - 0 - 0
Complete 0: 50%
Link 27 1: 50% - - 0 - 0
Average 0: 50%
Link 28 1: 50% - - 0 - 0
0: 75%
K-Means 31 1: 25% 2 2.0 0 - 0
0: 85%
EM 32 1: 15% - - 0.04 50.1999 0

Table 5.7. Summary of Hybrid Approach Clustering Algorithms

60
The way of using clustering algorithms in this hybrid approach also similar to the algorithms that are
used in unsupervised machine learning. Likewise, the manner of these algorithms used, determined
and evaluated also similar but their performance is different. To evaluate the accuracy, we consider
the basis of number of clusters produced and correctness of the produced clusters. Additionally, the
time (rate of convergence) and the number of iterations were also the other evaluations (as shown on
table 5.7).

Based on the experiment on table 5.5 and table 5.6, from the total five clusters, the EM and K-Means
clusters which yield significantly higher accuracy than other clusters when it compared with
machine learning result (as shown on table 5.2 and table 5.3). EM is the best-suited algorithm for the
given datasets, since it depends on the probability distributions where each distribution represents a
cluster. As presented in the above Section, if the log likelihood probability was maximum shows that
high accuracy. Similarly, K-Means clustering performance was also better performance. However;
Single, complete and average link clusters result not much surprise when compared with EM and K-
Means clusters. Therefore; for WSD to Afan Oromo, EM and K-means enhance the accuracy of
sense clustering than hierarchical single, complete and average clustering algorithms.

5.4. Discussion

In this paper, we reported unsupervised machine learning and hybrid approach for Afan Oromo
WSD. We described the experiment we conducted to compare the performance of the two
approaches. These approaches were evaluated with a benchmark test set (by experts), to facilitate the
comparison with the results of both approaches. The accuracy results of our method, when compare
machine learning and hybrid results (table 5.2 and table 5.5) are promising and proving method
respectively[22].

In the machine learning experiment, the system yields an accuracy of 56.2% for all senses as well as
in the hybrid experiment, the system yields an accuracy of 65.5%. Looking at the smallest dataset
available to both machine learning and hybrid, they found an average figure of different yielded
results. Initially, we evaluated our WSD method with all the 15 words. This lead, to a total of 15
words tested in this evaluation, and these ambiguous words have two senses to five senses. Six terms
have two senses (the terms with two senses are afaan, boqote, dubbatate, haare, ji’a, lookoo ), and
six terms have three senses (the terms with three senses are diige, tume, handhuura, dhahe, mirga,

61
waraabuu ) and two terms have five senses (the terms with five senses are bahe, ija) and the left one
has four senses (the term with four senses is darbe ) out of 15 ambiguous terms [94]. The table 5.2
and table 5.5 show the number of senses in detail.

The hybrid approach is promising result as compared with the performance of independent machine
learning approach or independent rule based approaches to WSD task (as shown in table 5.2 and
table 5.5). The rules were helpful to extract contexts overcoming the shortage of training data. This
rule is beneficial for WSD, especially the integration of deep linguistic knowledge with algorithms
markedly improves disambiguation accuracy. We argue that using hybrid approach to find senses,
significantly increase disambiguation performance. This result also seems to support the result
obtained by [88]. In this study we claim that using either a machine learning or hybrid approach (as
discussed in Section 4.6 and Section 4.7 respectively) possible to find the sense of any ambiguous
word in a given contexts but the hybrid approach was recommended when there are scarcity of
training data and the accuracy needed [36].

The result found that using a window size of ±2 words either side of the target word offered the
accuracy of disambiguation than using the whole sentence. Therefore; smaller value of the window
size, which leads to the proper choice of sense for the target word. Based on this result, we conclude
that for Afan Oromo window 2 was recommended unlike Amharic language [45] which window size
of 3 is recommended.

The conducted experiment shows that, the semantic has come to the conclusion that the meaning of
words are closely connected to the statistics of word usage, which are working with window size and
vectors value derived from event frequencies; that is, we are dealing with Vector Space Model
(cosine similarity) and clustering (Euclidean distance) [89]. By using cosine similarity we include
important semantic information in the purely statistical process of selecting the appropriate sense for
a particular word. This benefits both unsupervised, hybrid approaches to WSD by increasing the
chances of matching a particular contexts.

As shown in table 5.8, the result obtained by unsupervised machine learning and hybrid approach
were different as the semantic information extracted by the algorithms is distinct from the rule.
However, the inclusion of unsupervised machine learning only as features does not always improve
performance. This is [89] that machine learning algorithms were a useful information source for
disambiguation but that it not as robust as a linguistic (modifiers in this case). The most likely reason

62
for this is that our approach relies on automatically assigned immediately preceding words while
machine learning are needs to left and right of unannotated data set. On the other hand, the machine
learning is noisy while the rule is more reliable and prove to be a most useful linguistic knowledge
for WSD.

As the conducted experiment showed, each clusters have a context group, where the sense of these
context groups are hopefully different. The underlying assumption is that the senses found in similar
contexts are similar meanings. Then, new occurrences of the context can be classified into the
closest induced clusters (senses). All contexts of related senses are included in the clustering and
thus performed over all the contexts in the sentences. The underlying hypothesis is that ambiguous
word contexts clustering captures the reflected unity among the contexts and each cluster reveal
possible relationships existing among these contexts (as seen in table 5.3 and table 5.6) [82].

As the evaluation of the system indicates that, unsupervised machine learning usually 70% on the
test, while the hybrid approach accuracy 81%. The below table 5.8 contains the evaluation
performance of the WSD:

Unsupervised Machine Hybrid Approach

No Learning

Precision Recall Precision Recall

Correctly Disambiguated Senses 70% 46% 81% 60%

Table 5.8. Evaluation of WSD

From Table 5.8, the hybrid approach WSD is an encouraging result than unsupervised machine
learning approach. This is due to unsupervised approach is not as hybrid approach, especially
hierarchical clustering result was noisy. As already discussed before, the obtained result in both
approach were different. Therefore, the linguistic knowledge (hybrid approach) the best approach to
solve WSD than machine learning algorithms in Afan Oromo [64] as shown in the experiment

63
(Table 5.8). However, the overall system performance gained thus far is not surprusing since this
language was under resourced materials and tools.

From the finding of this experiment the addition of deep linguistic knowledge to a WSD system is a
significant rise in disambiguation accuracy and coverage as compared with results discussed so far.
It is especially interesting that using the preceding modifiers of the ambiguous word perform better
result. We can conclude that modifiers contain a lot of valuable clues for disambiguation [72].

The WSD developed for Afan Oromo has its own strength and weakeness sides. As the result
showed that, the experiment attempts to disambiguate any ambiguous words if it running in corpus
rather than limiting itself to treating a restricted ambiguous words. This is one of the strongest sides
of this WSD. It is argued that this approach is more likely to assist the creation of practical systems.
This system has the first work, that integerated different algorithms to find the appropriate sense of
ambiguous word in Afan Oromo.

However, there is some ambiguous word on which the performance of our approach is actually low
(as shown in Table 5.2 and Table 5.5). The system reported that the vector space model was affected
by the data sparsity. The frequency of co-occurrence of most context words are zero due to the
limited corpus size of the language. This result affects our cosine similarity values (as shown in
Table 5.10 below). The other weakness of the system is context clustering in hierarchical clustering
which was a noisy and yielded low performance as compared to K-Means and EM clustering (as
shown in Table 5.3 and Table 5.6).

64
 Walk-through Using an Example

A. Machine Learning

In this section, a walk-through of the experiment is described using an example in the following
sections. As described in Section 5.2.1, the contexts are identified using window sizes. For
instance, assuming the ambiguity word is bahe the unsupervised machine learning WSD, extracts
the following contexts words.

No Contexts

1 Bilisa

2 Ragaa

3 Dhugaa

4 Qabsoo

5 Tabba

6 Gaara

7 Uccuu

8 Duute

9 Daara

10 Lubbuu

Table 5.9. Extracted Context Using Window Size

As already mentioned, we take a property of vector space model that the values of the contexts in a
vector space derived from event frequencies. We constructed a vector for each context and calculate
the angle between these contexts using cosine similarity measure. The entry in the table 5.10 below
shows the cosine similarity of the contexts.

65
 Bilisa Qabsoo Gaara Tabba Daara Uccuu Duute Lubbuu Ragaa Dhugaa

Bilisa 1 0.98 0.09 0.05 0.01 0.15 0.04 0.17 0.07 0.19

Qabsoo 0.99 1 0.09 0.015 0.063 0.14 0.02 0.246 0.111 0.25

Gaara 0.07 0.089 1 0.98 0.062 0.027 0.004 0.20 0.11 0.10

Tabba 0.09 0.07 0.97 1 0.05 0.06 0.04 0.09 0.06 0.12

Daara 0.40 0.12 0.22 0.46 1 0.98 0.0 0.0 0.014 0.016

Uccuu 0.52 0.14 0.34 0.63 0.98 1 0.0 0.0 0.031 0.013

Duute 0.41 0.08 0.22 0.41 0.0 0.0 1 0.91 0.0 0.038

Lubbuu 0.248 0.247 0.11 0.20 0.033 0.08 0.94 1 0.066 0.27

Ragaa 0.43 0.14 0.287 0.49 0.083 0.009 0.0 0.12 1 0.98

Dhugaa 0.41 0.20 0.22 0.36 0.07 0.08 0.05 0.28 0.95 1

Table 5.10. Cosine Similarity Measure Representations

For the clustering result, we have loaded this cosine values in CSV format into the Weka tool that is
shown in the Table 5.10. If the cosine value is not in CSV format we need to be converting it to
.CSV (Comma-Separated Values) file. The reason for the format change is the incompatibility of the
tool and to use in the further steps. Weka tool present with visual dendrogram graphically and non-
hierarchical clustering algorithms.

From the total contexts, the Weka provided five clusters for the ambiguity word bahe. Out of the
total, it correctly clustered three senses. While the rest two pairs of cluster were incorrectly clustered
with different senses. Based on the experiment shows that the ambiguity word bahe with given
contexts has the following senses; the first cluster include bilisa and qabsoo at dissimilarity 1.15, the
second cluster include gaara and tabba at dissimilarity 1.23 and the third cluster include daara and
uccuu at dissimilarity 1.42, the two wrongly clustered contexts are dhugaa with lubbuu, and ragaa
with duute. But it should cluster dhugaa with ragaa, lubbuu with duute to give the senses of witness
and death/pass respectively as experts evaluated.

66
One thing that is clear from the experiment is that the senses are clustered in 0 - 4 clusters (5
clusters) where cluster 0 is the pair of contexts which are bilisa and qabsoo clustered and make the
sense of got freedom, cluster 1 is the pair of contexts gaara and tabba are grouped to make the sense
of highland and cluster 2 is daara and uccuu are merged to make the sense of cloth, the other two
clusters: cluster 3 and cluster 4 are incorrectly clustered and cannot make a sense.

The above experiment was summarized as the following:

a. Cluster 0: this set of senses comprises the contexts which have highly similar. The two
contexts merge together under this cluster are the context words bilisa and qabsoo, this
context has similar sense which shows freedom or political releases.
b. Cluster 1: this set of senses comprises the context words, namely gaara and tabba which are
clustered together to make the sense of highland or topography
c. Cluster 2: this set of senses comprises the context words, namely daara and uccuu which are
clustered to show the similar sense which is the cloth
d. Cluster 3: this set of senses comprises the context words, namely dhugaa and lubbuu, which
are incorrectly clustered together to make inappropriate sense.
e. Cluster 4: this set of senses comprises the context words, namely ragaa and duute, which are
incorrectly clustered together to make the sense.

The figure 5.2 below dendrogram shows the more description of these experiments:

67
 Figure 5.2. Single-Link(left), Complete-Link(Middle) and Average-Link(right) Dendrogram

The result obtained by the three hierarchical clustering are similar from the above dendrogram (as
shown in Figure 5.2). The first step in the clustering process is to look for the pair of contexts that
are the most similar, that are the closest in the sense of having the lowest dissimilarity this is the pair
assume bilisa and qabsoo. These two samples are then joined at a level in the first step of the
dendrogram, or clustering tree.
As figure 5.2 indicates that, in the hierarchical clustering experiment individual senses are leaf nodes
in a binary tree of clusters, and are treated as singleton clusters. Any intermediate node is the
centroid of its two children,which are more similar to each other than any other cluster as well as the
root of the tree contains all clusters and therefore contains all senses.

It appears from this dendrogram that, the data can be represented by three clusters. However, as the
number of cases increases, it may not be obvious. Indeed, one of the biggest problems with this
cluster analysis is identifying the optimum number of clusters. As the fusion process continues
increasingly dissimilar clusters must be fused. Deciding upon the optimum number of clusters is

68
largely subjective, although looking at a graph of the level of similarity at fusion versus a number of
clusters may help.

As evident from the visualization Figure 5.3 below, the output has been classified into 5 clusters
using K-Means and EM clustering. The algorithm took 2 iterations to complete and reach the result.
The distortion (average within the cluster sum of squared errors) amongst the clusters themselves is
7.66 units. EM is not nested (it means not hierarchical result). The result of this cluster result is
written to a class indices. The values indicate the class indices, where a value '0' refers to the first
cluster; a value of '1' refers to the second cluster. As the experiment shows that EM and K-Means
produce better result, as compared with hierarchical clusters. In case of the quality of the clusters,
EM and K-Means are relatively good quality as compared with others. Our result demonstrate that
our criteria match the predictability presented by experts, indicating there is no single perfect cluster
criterion. Instead, it is necessary to select the promising criterion to match a human subject's
generalization needs.

Figure 5.3. Expectation Maximization and K-Means Clustering

69
 B. Hybrid Approach

No The Modifiers

1 Bilisa

2 Qabsoo

3 Tabba

4 Gaara

5 Dhugaa

6 Ragaa

Table 5.11. Extracted Modifiers for ambiguity word bahe

As can be seen from table 5.11, the modifiers identified accordingly the rule developed. Similar to
machine learning, the identified modifiers with its cosine similarity in .CSV file loaded to Weka
tool. The process of using clustering algorithms was similar to in unsupervised machine learning
experiment. As the experiment shows that senses are similar between each other and are dissimilar to
the senses belonging to different clusters. The clustering is which two or more senses are considered
belong to the same cluster if it defines a concept common to all these senses.

70
 Figure 5.4. EM (left side) and K-Means (right side)

The other interesting way to examine the data in these cluster is inspect it visually. One of the
options from this pop-up menu is visualizer cluster assignments. A window pops up that the results
and see them visually. This is shown in a chart how the clusters are grouped in terms of instances.
Also, turn up the “Jitter” which is artificially scatter the plot points to allow to see more easily.

As a general, the WSD performance was promising result as it is hybrid approach which compiled
unsupervised machine learning and rule based approaches. The hybrid approach was very important
for WSD system especially for under resourced language like Afan Oromo. In this work, it has
essential role in this study in terms of accuracy and data scarcity. Similarly, it has a strict sense of
well-formedness in mind and imposes linguistic constraints to satisfy well formedness. It relies on
hand-constructed rules that are to be acquired from language specialists rather than automatically
trained from data [64]. Based on the experiment, the possible senses of the test ambiguous word
according to the given contexts has the following senses on table 5.12.

71
No Ambiguous Words No of Senses List of Senses Sense
Sense 1 Language
1 Afaan 2 Sense 2 Mouth
Sense 1 Freedom
Sense 2 Highland
2 Bahe 5 Sense 3 Cloth
Sense 4 Witness
Sense 5 Dead / Pass
Sense 1 Break / Rest
3 Boqote 2 Sense 2 Died
Sense 1 Cross
4 Darbe 4 Sense 2 Class to class pass
Sense 3 Died
Sense 4 Broadcast
Sense 1 Absence on Meeting
5 Diige 3 Sense 2 Fence
Sense 3 cancel to start new
Sense 1 Struggle
6 Dubbatate 2 Sense 2 Wedding
Sense 1 Hit
7 Tume 3 Sense 2 Make
Sense 3 Contraceptive
Sense 1 Sad
8 Haare 2 Sense 2 Burn
Sense 1 Center
9 Handhuura 3 Sense 2 Gift
Sense 3 Navel
Sense 1 Eye
10 Ija 5 Sense 2 Fruit of Tree
Sense 1 Stars
11 Ji’a 2 Sense 2 Month
Sense 1 Pretty
12 Lookoo 2 Sense 2 Rope
Sense 1 Hit
13 Dhahe 3 Sense 2 Follow
Sense 3 Fail
Sense 1 Direction
14 Mirga 3 Sense 2 Human right
Sense 3 Brave
Sense 1 Fetch
15 Waraabuu 3 Sense 2 Hyena
Sense 3 Record
Table 5.12. Summary of the Test

72
 Chapter Six

6. Conclusion and Recommendation

6.1. Conclusion

The study deals with Afan Oromo word sense disambiguation. As discussed in the previous chapters,
in this research, hybrid word sense disambiguation was recommended for Afan Oromo and for under
resourced languages in general. To this end, we have forwarded the conclusion and recommendation
as presented in the following sections.

The goal of Information Retrieval is striving to answer users’ requests correctly as intended by the
user. The contribution of NLP in achieving such goal of Information Retrieval Systems has been
clearly pointed out. Furthermore, it has been pointed out how NLP plays a significant role in
enhancing the computer’s capability to process texts. To that end, word sense disambiguation is one
component of NLP contributing a lot to the effort of solving the problem of Information Retrieval
Systems in answering users’ requests by introducing semantics of a query term and index terms.

The overall focus of this research is to investigate Word Sense Disambiguation which addresses the
problem of deciding the correct sense of an ambiguous word based on its surrounding context’s and
the modifiers. In this study, a hybrid machine learning approach which combines unsupervised
machine learning and rule based methods. To this end, we relied on several techniques which
include: N-gram models (N- windows of words) to extract context words, vector space model for
measuring similarity between the context words and clustering algorithms (Single, Complete and
Average clustering, EM and K-means clustering algorithms) to group related context words. This
hybrid method, avoid the problem of the knowledge acquisition bottleneck, that is, lack of large-
scale resources and annotated corpora. Our approach to WSD has been based on the idea that the
semantics of the context words belonging in the same sense of a word will have similar neighboring
words. The context is hence a source of information and is the only means to identify the meaning of
an ambiguous word. In order to achieve this, the approach does not rely on labeled training text and
doesn’t make use of any expensive resources like dictionaries, thesauri, and WordNet [22].

73
For under resourced Ethiopian language like Afan Oromo the hybrid approach is recommended.
Since here is no annotated corpus (even difficult to obtain electronic materials for such language),
hybrid approach plays a great role to disambiguate. The hybrid approach relies on hand-constructed
rules that are acquired from language specialists rather than automatically trained from data.

Prior to this research, for Afan Oromo, there is no work recommending standard optimal context
window size, which refers to the number of surrounding words that are sufficient for extracting
useful contexts. Based on this study, the window size of +2 on both sides of the ambiguous word
was found to be enough for disambiguation. Similar to English, window of +2 is the standard
window applicable for disambiguation [92] in Afan Oromo. The nearest words surrounding the
ambiguous word give more information than words far from the ambiguous word and consequent
surrounding words to the left and to the right provide any information for the purpose of
disambiguation.

There are several types of clustering algorithms. In this paper we relied on hierarchical and
probabilistic algorithms. A hierarchical clustering is based on the union between the two nearest
clusters. The beginning condition is realized by setting every datum as a cluster. After a few
iterations, it reaches the final clusters wanted. The probabilistic algorithms use around model
matching using probabilities as opposed to distances to decide clusters. EM and K-Means is an
example of this type of clustering algorithm [95].

We did experiments on five different clustering algorithms namely K-Means, EM, single, complete
and average link. Based on the result of the experiment out of the five algorithms simple K-Means
and EM algorithms are the best of all to identify the meaning of ambiguous word in a context. We
believe that the observed poor performance of hierarchical agglomerative algorithms is because of
the errors they make during early agglomeration. The superiority of partitional algorithm also
suggests that partitional clustering algorithms are well-suited for obtaining clustering solutions due
to not only their relatively low computational requirements, but also comparable or better
performance.

We also presented an evaluation methodology for measuring the precision and recall of discovered
senses. In our experiments, we proved that the hybrid approach outperformed the machine learning
approach.

74
In this work the hybrid approach and unsupervised machine learning achieved an accuracy of 81%
and 70% respectively. We found that better results were achieved using a combination of rule and
unsupervised machine learning features commonly used for WSD. The results obtained were
encouraging as there is lack of resource of the language because of shortage of dictionary, labelled
corpus, thesaurus.

6.2. Recommendation

The following recommendations are forwarded based on our findings with regard to the
developments of resources and future research directions to WSD for Afan Oromo:

 Researches in WSD for other languages use knowledge resources like WordNet, lexicon,
machine readable dictionaries and machine translation software. In this study, we faced a
significant challenge as Afan Oromo lacks those resources. Taking into account their
contribution to WSD and other research concerned institutions should develop these
resources.
 For other language a standard sense annotated data are available for WSD research and for
testing a standardized WSD system. The researcher doesn’t have such data for Afan Oromo
language. So, there need to be an initiative to prepare the data for WSD research.
 For the under resourced Ethiopian language like Afan Oromo, the linguistic knowledge was
made better result, it needs linguistic background knowledge of the language.
 All the result showed that the techniques are fairly successful and effective in the
disambiguation task. Thus, more research work should be exerted to carry out further
improvements on the performance of these techniques.
 The hybrid method has been used. Thus, more research work should be employed to carry
out further improvements on the performance using Supervised, Semi-Supervised
approaches.
 The size of the corpus used for this research was limited and doesn’t has good quality. These
limitations affected the accuracy of word sense disambiguation because if resources used are
small, we may not be able to find all possible senses. Therefore, some work should be done
with large and high quality corpus to minimize these problems.

75
Future research directions for WSD in Afan Oromo include:

 In addition to corpus based approach, there are also knowledge source based and hybrid
approach (combination of knowledge source based and corpus based approach) which are
used for WSD in another language and found a good result [93] like WordNet, annotate data.
These approaches need to be investigated for Afan Oromo as well.
 As unsupervised WSD are based on unlabeled corpora and do not exploit any manually
sense-tagged corpus to provide a sense choice for a word in context, these results less
accuracy than other approaches. This approach overcomes the main problems of supervision
and the data scarcity problem, especially the lack of annotated data like Afan Oromo. For
example, an evaluation of Yarowsky’s bootstrapping algorithm leads to very high
performance over 90% accuracy on a small-scale data set [22]. This approach overcomes the
main problems of supervision and the data scarcity problem, especially the lack of annotated
data like Afan Oromo.

76
References

1. Salton G(1981) The Measurement of Term Importance in Automatic Indexing: In Journal of

the American Society for Information Science, vol.32.
2. Ide, Nancy and Jean Véronis (1998)Word sense disambiguation: The state of the art,
Computational Linguistics, vol.24.
3. Roberto Navigli (2009) Word Sense Disambiguation: A Survey, ACM Computing Surveys,
USA, Washington Dc.
4. David Yarowsky and Radu Florian (2002)Evaluating sense disambiguation across diverse
parameter spaces. Natural Language Engineering.
5. Mohd Shahid Husain and Mohd Rizwan Beg (2013) Word Sense Ambiguity: A Survey,
Lucknow, India, Volume 02.
6. Harris(1998) Distributional structure: Word, Cambridge University Press, Cambridge, UK.
7. Getachew M.(2011) Parts of Speech Tagging for Afaan Oromo, M.Sc. Thesis, Addis Ababa
University, Addis Ababa, Ethiopia.
8. Karov, Y. and Shimon, E(1998), Similarity-based word sense disambiguation. computational
Linguistics, Vol.24.
9. Abebe Keno(2002) Case Systems in Oromo, MA Thesis. Ababa University, Ethiopia.
10. Girma Debele (2014)Afan Oromo News Text Summarizer, Master’s thesis, Pohang
University of Science and Technology, Pohang, Korea.
11. Daniel Bekele(2011) Afaan Oromo Information Retrieval (CLIR): A Corpus Based
Approach, M.Sc. Thesis, Addis Ababa University, Addis Ababa, Ethiopia.
12. Agirre, E. and Martinez, D.(2000) Exploring automatic word sense disambiguation with
decision lists and the web, Toronto, Ontario.
13. Nancy Ide and Jean Véronis, (2007)Word Sense Disambiguation Algorithms and
Applications, Ney work, Spring, vol.33.
14. Bar-Hillel, Yehoshua (1960)The present status of automatic translation of languages.
Advances in Computers, ed. by Franz Alt, New York: Academic Press.
15. Ravi Mante, Mahesh Kshirsagar and Dr. Prashant Chatur(2014) A Review Of Literature On
Word Sense Disambiguation, PhD, Government college of engineering Amravati,
Maharashtra.

77
16. Jurafsky, D., Martin, J.(2009) Speech and Language Processing (2nd Edition) Pearson
Education.
17. Gale, W., Church, K., & Yarowsky, D. (1993)A method for disambiguating word
senses in a large corpus, Computers and the Humanities, vol.26.
18. Doina Tatar, Gabriela Serba(2001) A New Algorithm for Word Sense Disambiguation,
Studia Universitatis "Babes-Bolyai", Seria Informatica, Vol.16.
19. Artstein, R. and Poesio, M.(2008) Inter-coder agreement for computational linguistics.
Computational Linguistics, Vol 34.
20. Lesk, M. (1986) Automatic sense disambiguation using machine readable dictionaries,
Toronto, Ontario.
21. Smeaton, A.F(1995) Linguistic Approaches to Text Management: An Appraisal of
Progress, Journal of Document & Text Management, Vol.2.
22. David Yarowsky(1994)Decision lists for lexical ambiguity resolution: application to accent
restoration in Spanish and French, USA, Stroudsburg.
23. Marco Baroni and Gemma Boleda(2009) Distributional Semantics: Natural Language
Processing.
24. Scott McDonald(2000)Environmental Determinants of Lexical Processing Effort. Ph.D.
thesis, University of Edinburgh.
25. H. Schutze(1998)Automatic word sense discrimination, Computational Linguistics, vol.24(1).
26. Jain, A. K., Murty, M. N., and Flynn, P. J. (1999), Data clustering: a review, ACM
Computing Surveys, Vol. 31.
27. Nagy, G(1968) State of the art in pattern recognition, Cambridge University Press,
Cambridge, UK.
28. Sneath, P. H. A. and Sokal, R. R(1973) Numerical Taxonomy: The Principles and Practice of
Numerical Classification, London, UK: Freeman.
29. King, B(1967) Step-wise clustering procedures, Journal of the American Statistical
Association, Vol.69.
30. Jain, A. K. and Dubes (1988), R. C. Algorithms for Clustering Data, Prentice-Hall.
31. Han, J. and Kamber, M.(2001) Data Mining–Concepts and Techniques: Morgan Kaufmann.
32. Juanying Xie, Shuai Jiang, Weixin Xie, Xinbo Gao (2009) An Efficient Global K-means
Clustering Algorithm, Vol.6.

78
33. M. N. Murty and A. K. Jain,(1999) “Data clustering: a review,” ACM Computing Surveys,
vol. 31.
34. A. Dempster, N. Laird, and D. Rubin(1977) Maximum likelihood from incomplete data via
the EM algorithm, Journal of the Royal Statistical Society, Vol. 39.
35. Celeux, G. and Govaert, G. (1992) A classification EM algorithm for clustering and two
stochastic versions. Computational statistics and data analysis, vol.14.
36. Abdel Monem, K. Shaalan, A. Rafea, H. Baraka (2008) Generating Arabic Text in
Multilingual Speech-to-Speech Machine Translation Framework, Machine
Translation,Springer,vol.20(4).
37. Amaal Saleh Hasan (2010) Word Sense Disambiguation and Semantics techniques, MSc
thesis, Sultan Qaboos University.
38. Gang Li, Guangzeng Kou, Ercui Zhou, and Ling Zhang(2009) Symmetric Trends: Optimal
Local Context Window in Chinese Word Sense Disambiguation, USA, Washington DC.
39. Yorick Wilks(1975) Formal semantics of natural language: Cambridge University Press,
Cambridge, UK.
40. Gale, W., Church, K., & Yarowsky, D. (1993)A method for disambiguating word
senses in a large corpus. Computers and the Humanities, vol.26(5).
41. Brown, P., S. Della Pietra, V. Della Pietra, and R. Mercer(1991) A statistical approach
to sense disambiguation in machine translation, Pacific Grove CA.
42. Solomon A.(2011) Unsupervised Machine Learning Approach For Word Sense
Disambiguation To Amharic Words, M.Sc. Thesis, Addis Ababa University, Addis Ababa,
Ethiopia.
43. Alok Ranjan Pal, Anirban Kundu, Abhay Singh, Raj Shekhar, Kunal Sinha(2013) A Hybrid
Approach To Word Sense Disambiguation Combining Supervised And Unsupervised
Learning, West Bengal, India.
44. Getahun Wassie (2014) A Word Sense Disambiguation Model for Amharic Words using
Semi-Supervised Learning Paradigm, School of graduate studies, Addis Ababa University,
Ethiopia.
45. Solomon A.(2011) Unsupervised Machine Learning Approach For Word Sense
Disambiguation To Amharic Words, M.Sc. Thesis, Addis Ababa University, Addis Ababa,
Ethiopia.

79
46. J. Sreedhar, S. Viswanadha Raju, A. Vinaya Babu, Amjan Shaik, P. Pavan Kumar(2012)
Word Sense Disambiguation: An Empirical Survey, Hyderabad, India.
47. Shaikh Samiulla Zakirhussain(2013)Unsupervised Word Sense Disambiguation, Indian
Institute of Technology, Bombay.
48. Warren Weaver(1955) Machine Translation of Languages, MIT Press, Cambridge.
49. Kaplan A(1955)An experimental study of ambiguity and context, Mechanical Translation,
Vol.2.
50. Stephen Clark (2014)Vector Space Models of Lexical Meaning, Handbook of Contemporary
Semantics, 2nd ed, edited by Shalom Lappin and Chris Fox.
51. Turney, Peter D. & Patrick Pantel (2010) From frequency to meaning: Vector space models
of semantics, Journal of Artificial Intelligence Research, vol.37
52. Katrin Erk and Sebastian Padó(2008) A structured vector space model for word meaning in
context. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language
Processing, Honolulu, HI, USA.
53. Yarowsky, D.(2007) Unsupervised word sense disambiguation rivaling supervised methods,
In Proceedings of the 33rd Annual Meeting of the Association for Computational
Linguistics, Cambridge, M.A.
54. Schutze, H. (1992)Dimensions of meaning: In Supercomputing ’92: Proceedings of the 1992
ACM/IEEE Conference on Supercomputing. IEEE Computer Society Press, Los Alamitos,
CA.
55. Hearst, Marti(1991) "Noun Homograph Disambiguation Using Local Context in Large Text
Corpora," in Using Corpora, University of Waterloo, Waterloo, Ontario.
56. Dagan, Ido and Alon Itai(1994) Word Sense Disambiguation Using a Second Language
Monolingual Corpus, Computational Linguistics, vol.20.
57. Stefan Thater, Hagen Fürstenau, and Manfred Pinkal(2011)Contextualizing semantic
representations using syntactically enriched vector models: In Proceedings of the 48th
Annual Meeting of the Association for Computational Linguistics, Uppsala, Sweden.
58. Adam Kilgarriff(1997) I don’t believe in word senses. Computers and Humanities, vol.31(2)
59. Honolulu(2008) A Structured Vector Space Model for Word Meaning in Context
Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing.
60. Jeff Mitchell and Mirella Lapata(2008) Vector-based models of semantic composition: In
Proceedings of ACL-08: HLT, Columbus, USA.

80
61. Katrin Erk and Sebastian Padó(2008) A structured vector space model for word meaning in
context. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language
Processing, Honolulu, HI, USA.
62. Joseph Reisinger and Raymond J Mooney(2010)Multi-prototype vector-space models of
word meaning, In Proceedings of HLT-NAACL.
63. Pedersen B. S.(1997) Lexical ambiguity in machine translation: expressing regularities in
the polysemy of Danish Motion Verbs. PhD Thesis,Center for Sprogteknologi,Copenhagen.
64. Debela Tesfaye(2011)A Rule-based Afaan Oromo Grammar Checker, Addis Ababa
University, Addis Ababa, Ethiopia.
65. Francisco Oliveira, Fai Wong, Yiping Li, Jie Zheng(2005) Unsupervised Word Sense
Disambiguation and Rules Extraction using non-aligned bilingual corpus, MSc thesis,
Tsinghua University, Beijing.
66. Ide, N., & Veronis, J. (1998) Introduction to the special issue on word sense
disambiguation: the state of the art. Comput. Linguist., vol.24(1).
67. Specia L.,Oliveira-Neto S. F., Nunes M.G.V.,Stevenson, M. (2005) An automatic approach to
create a sense tagged corpus for word sense disambiguation in machine translation:To be
published in Proceedings of the 2nd Workshop of the Meaning Project, Trento.
68. Tullu Guya(2003) CaasLuga Afaan Oromoo: Jildii-1,Gumii Qormaata Afaan Oromootiin
Komishinii “Aadaa fi Turizimii Oromiyaa”, Finfinnee.
69. Abera N. (1988) Long vowels in Afan Oromo: A generic approach, Master’s thesis, School of
graduate studies, Addis Ababa University, Ethiopia.
70. Debela Tesfaye (2010) Designing a Stemmer for Afan Oromo Text: A hybrid approach,
Master’s thesis, School of graduate studies, Addis Ababa University, Ethiopia.
https://om.wikipedia.org/wiki/Oromoo Accessed at 1/1/2015
71. Gragg and Gene B.(1996)Oromo of Wollega: non-semetic languages of Ethiopia , East
lansing, Michigan state university press.
72. Tilahun Gamta(1995) Seera Afaan Oromoo, Finfinnee, Boolee Press.
73. Kula Kekeba Tune, Vasudeva Varma and Prasad Pingali, (2007) “Evaluation of Oromo
English Cross-Language Information Retrieval”, IJCAI 2007 Workshop on CLIA,
Hyderabad, India.
74. Wakshum Mekonnen (2000) Development of Stemming Algorithm for Oromo Texts. MA
Thesis.

81
75. Tilahun Gamta(1989)Oromo-English Dictionary: Addis Ababa. University Press.
76. Getachew Rabirra(2014) Furtuu: Seerluga Afaan Oromoo, Finfinnee Oromiyaa press.
77. Atelach Alemu Argaw and Lars Asker(2010) An Amharic Stemmer: Reducing Words to their
Citation Forms, M.S Thesis, Department of Computer and Systems Sciences, Stockholm
University, Sweden.
78. Rada R., Mili H., Bicknell E. and Blettner M.(2009) Development an Application of a Metric
on Semantic Nets, in IEEE Transactions on Systems,Man and Cybernetics, vol. 19.
79. Geoffrey Leech(1991)The state of the art in corpus linguistics, in English Corpus
80. Diriba Megersa(2002)An Automatic Sentence Parser For Oromo Language Using
Supervised Learning Technique, M.Sc. Thesis, Addis Ababa University, Ethiopia.
81. Curran, James R. (2004) From Distributional to Semantic Similarity, Ph.D. thesis,
University of Edinburgh.
82. Fei Shao, Yanjiao Cao(2005) A New Real-time Clustering Algorithm, Department of
Computer Science and Technology, Chongqing University of Technology Chongqing 400050,
China.Linguistics:Linguistic Studies in Honour of Jan Svartvik, London, Longman.
83. Malmkjaar, Kirsten(1995)The linguistics Encyclopedia, New York: Routledge.
84. McCarthy D. and Carroll J.(2003) Disambiguating nouns, verbs and adjectives using
automatically acquired selectional preferences, Computational Linguistics, Vol. 29.
85. Flickinger, D.(2015)Natural Language Engineering-Efficient Processing with HPSG:
Methods, Systems, Evaluation. At: http://www.coli.uni-sb.de/nlesi/ Accessed at 3/15/2015.
86. Agirre, E. and Martinez, D.(2001)Exploring automatic word sense disambiguation with
decision lists and the web: In Proceedings of the COLING 2000 Workshop on Semantic
Annotation and Intelligent Content, London, Longman.
87. Jung-Wei Fan, MS1 and Carol Friedman(2008)Word Sense Disambiguation via Semantic
Type Classification,http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2655949/ Accessed at 1/8/2015
88. Rada Mihalcea, Ted Pedersen(2005) Advances in Word Sense Disambiguation, Tutorial at
ACL Conference.
89. Kaplan,A.(1955) An experimental study of ambiguity and context, Mechanical Translation,
Vol 2.
90. Forster, R.(2006) Document clustering in large German corpora using natural language
processing, Ph.D. dissertation, University of Zurich.
91. Slava Kisilevich, Florian Mansmann, Daniel Keim(2001) P-DBSCAN: A density based

82
 clustering algorithm for exploration and analysis of attractive areas using collections of geo-
tagged photos, University of Konstanz.
92. Tesmegen Negassa (1995) word formation: the structure of independent and dependent
clauses in oromo, Addis Abeba, Bole.
93. George Miller.(1990) Wordnet: An on-line lexical database. International Journal of
Lexicography, vol.3(4).
94. Ted Pedersen and Rebecca Bruce(1997)Distinguishing word senses in untagged text: In
Proceedings of the 2nd Conference on Empirical Methods in Natural Language Processing,
New York, Routledge.
95. Rui Xu and Donald Wunsch (2005) Survey of Clustering Algorithms and Transactions On
Neural Networks, Cambridge University Press, Cambridge, UK. Vol. 16.

83
Appendix A: Question For Collection of Ambiguous Words

The questions are prepared and converted to Afaan Oromoo language to make suitable to collect the
ambiguous words test from public.
Saala: _______ Umurii:___________________
Sadarka Barumsa: kutaa 1-5:_________________
kutaa 6-12:________________
Barataa Kolleejjii:__________
Barataa Yuuniversitii:_______
Barsiisaa:________________
Q/Bulaa:_________________
Kan biraa:_______________

1. Jechoota hiika lamaa ol qaban (ambiguous words) kan naannawaa keessanitti fayyadamtu naaf
tarreessi?
Lakk Jechoota Baay’ina Hiikaa Hiika
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

84