ENHANCING EFFECTIVENESS OF AFAAN OROMO INFORMATION

RETRIEVAL USING LATENT SEMANTIC INDEXING AND DOCUMENT

CLUSTERING BASED SEARCHING

MSc. THESIS

BELETE BOGALE

SEPTEMBER 2020

HARAMAYA UNIVERSITY, HARAMAYA

Enhancing Effectiveness of Afaan Oromo Information Retrieval Using Latent
Semantic Indexing and Document Clustering Based Searching

A Thesis Submitted to Department of Information Science,

Post Graduate Program Directorate

HARAMAYA UNIVERSITY

In Partial Fulfillment of the Requirements for the Degree of

MASTER OF SCIENCE IN INFORMATION SCIENCE

Belete Bogale

September 2020

Haramaya University, Haramaya

i
 HARAMAYA UNIVERSITY

POSTGRADUATE PROGRAM DIRECTORATE

I hereby certify that I have evaluated this Thesis entitled Enhancing Effectiveness of Afaan
Oromo Information Retrieval Using Latent Semantic Indexing and Document Clustering
Based Searching that prepared by Belete Bogale under my advice and I recommend that it to be
submitted as fulfilling the thesis requirement.

Dr.Million Meshesha (PHD) ________________ ___________

Principal Advisor: signature date

As a members of the Board of Examiners of the MSc Thesis Open Defense Examination, I certify
that I have evaluated the Thesis prepared by Belete Bogale and examined the candidate. I
recommend that the thesis be accepted as fulfilling the Thesis requirements for the degree of
Master of Science in Information Science.

________________________ ____________________ _________________

Chairperson Signature Date

________________________ ____________________ _________________

Internal Examiner Signature Date

________________________ ____________________ _________________

External Examiner Signature Date

Final approval and acceptance of the Thesis is contingent upon the submission of its final copy to
the Council of Graduate Studies(CGS) through the candidate’s department or Postgraduate
committee (DGC or SGC).

ii
 DEDICATION
This work is dedicated to my parents. For their endless love, support and encouragement
throughout my life.

iii
 STATEMENT OF THE AUTHOR
By my signature below, I declare and affirm that this Thesis is my own work. I have followed all
ethical and technical principles of scholarship in the preparation, data collection, data analysis and
compilation of this Thesis. Any scholarly matter that is included in the Thesis has been given
recognition through citation. This Thesis is submitted in partial fulfillment of the requirement for
an MSc degree at Haramaya University. The Thesis is deposited in the Haramaya University
Library and is made available to borrowers under the rules of the library. I solemnly declare that
this Thesis has not been submitted to any other institution anywhere for the award of any academic
degree, diploma or certificate. Brief quotations from this Thesis may be used without special
permission provided that accurate and complete acknowledgement of the source is made. Requests
for permission for extended quotations from, or reproduction of this Thesis in whole or in part may
be granted by the Head of the School or Department when in his or her judgment the proposed use
of the material is in the interest of scholarship. In all other instances, however, permission must be
obtained from the author of the Thesis.

Name________________________ Signature_____________________

School/ Department______________ Date_________________________

iv
 BIOGRAPHICAL SKETCH
The author was born in Arsi Zone, Oromia region on Dec 21, 1994 from Ato Bogale Gurmu and
W/ro Kefene Guluma. He attended his elementary (1-8) at T/waji elementary school and his
secondary (9-10) Chilalo Terara secondary school and preparatory at Assella preparatory school.
He joined Haramaya University in 2012 and graduated with Bachelor degree in Information
Science in July 2014. After graduation, he was employed at Haramaya University as Graduate
Assistant and served there for about one year. Then in 2015, he has joined the Postgraduate
Program at Haramaya University for regular program and continued his studies in Information
Science.

v
 ACKNOWLEDGEMENT
First of all, I would like to thank one above all GOD for every success in my life, for giving me
wisdom and strength. Secondly, I would like to say as the great scientist Albert Einstein said “If
I have seen further than others, It is because I have stood upon the shoulder of giants.” Intended, I
would like to thank my advisor Dr. Million Meshesha, for his critical comments on my work, from
the beginning to the end of the study, and his patience in helping me to complete my work. Thirdly
it is my pleasure to thanks all my family, department staff, and other people for helping me when
I was working on this study. Finally, I offer my regards and blessings to all of those who supported
me in any aspect during the completion of this thesis as well as expressing my apology that I could
not mention all.

vi
 LIST OF ACRONYMS AND ABBREVIATIONS

AOIRS Afaan Oromo Information Retrieval System

CLIR Cross Language Information Retrieval

IR Information Retrieval

IRS Information Retrieval System

LSI Latent Semantic Indexing

LC Lexical Chain

MR Relevance Model

NLP Natural Language Processing

SVD Single Value Decomposition

VSM Vector Space Model

vii
 Table of Contents
DEDICATION ............................................................................................................................... iii
STATEMENT OF THE AUTHOR ............................................................................................... iv
BIOGRAPHICAL SKETCH .......................................................................................................... v
ACKNOWLEDGEMENT ............................................................................................................. vi
LIST OF ACRONYMS AND ABBREVIATIONS ..................................................................... vii
LIST OF TABLES ......................................................................................................................... xi
LISTS OF FIGURES .................................................................................................................... xii
ABSTRACT ................................................................................................................................. xiii
1. INTRODUCTION .................................................................................................................. 1
1.1. Background of the Study ..................................................................................................1

1.2. Statement of the Problem ..................................................................................................3

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

1.4. Significance of the Study ..................................................................................................5

1.5. Objective of the Study ......................................................................................................6

1.5.1. General Objective ......................................................................................................... 6

1.5.2. Specific Objectives ....................................................................................................... 6
2. LITERATURE REVIEW ........................................................................................................... 7
2.1. Overview ...........................................................................................................................7

2.2. History of Information Retrieval System..........................................................................7

2.3. Information Retrieval System Models ..............................................................................8

2.3.1. Vector Space Model ...................................................................................................... 8

2.3.2. Probabilistic Model ..................................................................................................... 10
2.3.3. Inference Network Model ........................................................................................... 11
2.4. Information Retrieval Process ........................................................................................12

2.4.1. Text operations............................................................................................................ 13

2.4.2. Term Weighting .......................................................................................................... 15

viii
 2.4.3. Indexing Approaches .................................................................................................. 16
2.4.4. Query Matching and Searching .................................................................................. 20
2.5. Document Clustering ......................................................................................................22

2.5.1. Approaches of Document Clustering .......................................................................... 22

2.6. Challenges of the Language in Information Retrieval System .......................................24

2.7. Overview of Afaan Oromo .............................................................................................26

2.7.1. Features of Afaan Oromo writing system ................................................................... 26

2.7.2. Morphology of Afaan Oromo: .................................................................................... 32
2.8. Related Works.................................................................................................................34

2.8.1. Foreign Related Works ............................................................................................... 34

2.8.2. Local Research Works ................................................................................................ 35
3. METHODS AND TECHNIQUES ....................................................................................... 42
3.1. Overview .....................................................................................................................42

3.2. Methodology of the study ...............................................................................................42

3.3. Research Design .............................................................................................................42

3.3.1. Document Corpus Preparation .................................................................................... 43

3.3.2. Implementation Tools ................................................................................................. 43
3.3.3. Evaluation procedure ................................................................................................. 43
3.4. The Architecture of AOIRs Using Semantic Indexing and Document Clustering Based
Searching. ..............................................................................................................................44

3.4.1. Dataset preprocessing ................................................................................................. 45

3.4.2. Indexing ...................................................................................................................... 49
3.4.3. TFIDF weighting ........................................................................................................ 51
3.4.4. Signle Value Decomposition (SVD) ........................................................................... 52
3.4.5. K-means Clustering .................................................................................................... 55
3.4.6. Searching..................................................................................................................... 57
4. RESULTS AND DISCUSSION ............................................................................................... 59
4.1. Overview .........................................................................................................................59

4.2. Document Corpus ...........................................................................................................59

ix
 4.3. Documents Preprocessing ...............................................................................................59

4.4. Single Value Decomposition with K-means Clustering .................................................61

4.5. Clustering Based Searching ............................................................................................63

4.6. Performance Measure .....................................................................................................64

4.6.1. Query Selection and Relevance Judgment .................................................................. 64

4.6.2. Experimental Result Using VSM................................................................................ 65
4.6.3. Experimental Result Using SVD ................................................................................ 65
4.6.4. Experimental Result Using Semantic Indexing and Document Clustering Based
Searching............................................................................................................................... 66
4.6.5. Performance Comparison............................................................................................ 68
4.7. Discussion of Result .......................................................................................................68

5. CONCLUSION AND RECOMMENDATION ........................................................................ 70

5.1. Conclusion ......................................................................................................................70

5.2. Recommendation and Future Direction ..........................................................................71

6. REFERENCES ..................................................................................................................... 72
7. APPENDICES ...................................................................................................................... 83

x
 LIST OF TABLES
1. The term document matric 13
2. The qubee in international phonetic writing (IPA) 29
3. Afaan Oromo Vowels 30
4. Gemination (doubling a consonant) in Afaan Oromo 31
5. Afaan Oromo personal pronouns 32
6. Adverbs in Afaan Oromo 33
7. Prepositions in Afaan Oromo 34
8. Related research works problem, approach, result and gap 42
9. List of documents collections 61
10. Proposed system relevance judgment 66
11. Effectiveness of VSM based Afaan Oromo IR 67
12. Retrieved result to query using SVD 67
13. Performance of SVD based K-means clustering for Afaan Oromo IR 68
14. Summarized result of overall performance of Afaan Oromo IR system 69

xi
 LISTS OF FIGURES
1. Basic inference network model 16

2 Architecture of a textual IR system 17

3. Singular Value Decomposition based Latent Semantic Indexing 24

4. Proposed system architecture 42

5. Inverted Indexing for Afaan Oromo Document 51

6. Singular Values Decomposition 53

xii
 ABSTRACT
This research work comes up with Latent Semantic Indexing and Document Clustering based
searching for Afaan Oromo documents. It intends to apply LSI and K-means clustering to handle the
semantic structure of words in documents. This mainly consists of three components; indexing,
clustering, and searching. Latent Semantic Indexing (LSI) model is a concept based retrieval method
that exploits the idea of a vector space model and singular value decomposition. On the other hand,
document clustering was investigated for improving the performance of information retrieval system.
Document clustering is an issue of measuring similarity between documents and grouping similar
documents together. K-means clustering was used to cluster the document using the Singular Value
Decomposition (SVD) matrix. Then, the retrieval process is further refined by making a similarity
measure between the query vector and cluster centroid vectors. IR pre-processing for tokenization,
normalization, stop word removal, and stems were used for selecting index and query terms. Finally,
a comparison is made between the SVD model with K-means clustering, VSM and SVD model. The
performance evaluation of the system was performed by using a selected set of documents and queries.
The experimental result showed that the proposed prototype registered on average 70% recall, 80%
precision, and 72% F-measure. Therefore, it indicated that the proposed method (SVD with Kmeans)
achieved significant improvement compared to the VSM and SVD model. Nevertheless, the
performance of the system is greatly affected by the statistical extraction of synonyms and polysemy,
mis-clustering, standard corpus, and stemming which need further research.

xiii
 1. INTRODUCTION
1.1. Background of the Study
Information Retrieval is defined as finding relevant documents that satisfy the information needs
of users from an unstructured large collection (Christopher et al., 2008). Information Retrieval is
one of the major branches of the Information Science discipline (Moukdad, 2003). The trend in
information storage and retrieval can be traced back to 2000 BC when people of Sumerians chosen
a special place to store clay tablets with cuneiform inscriptions (Christopher et al., 2008). After
they understand that their work is efficient in the use of information, they developed a special
categorization system that identifies every tablet and its content.
One of the major evolutions in Information Retrieval is the invention of the print machine in 1450
A.D (Judit and Gutman, 2003). A German goldsmith Johannes Gutenberg invented the first
movable printer, thousand years later after the Chinese invented paper which provides means for
disseminating and storing knowledge. Gutenberg‘s aim was allowing direct access to mass
information that was contained in the Bible and other scholarly works.
At the beginning of the 1990s, a single fact changed once and for all the perceptions towards
Information Retrieval was the introduction of the World Wide Web (Baeza-Yates and Ribeiro-
Neto, 2000). The Web is becoming a universal repository of human knowledge and culture which
has allowed unprecedented sharing of ideas and information on a scale never seen before (Baeza-
Yates and Ribeiro-Neto, 1999).
Information Retrieval has two main subsystems (Deerwester et al., 1989), Indexing, and
Searching. Indexing is an offline process of representing and organizing large document collection
using indexing structure such as Inverted file, sequential files, and signature file to save storage
memory space and speed up searching time. Searching is on the other hand an online process of
relating index terms to query terms and return relevant hits to the user's query.
Nowadays with the advent of digital databases and communication networks, huge repositories of
textual data have become available to a large public (Thomas, 1999). Although the use of elaborate
ergonomic elements like computer graphics and visualization has proven to be extremely fruitful
to facilitate and enhance information access (Thomas, 1999). The need to store and retrieve written
information became increasingly important over centuries, especially after the inventions of a
scientific paper (Christopher et al., 2008).
General search engines on the Web such as Google, Yahoo, and Microsoft Network are among the
popular tools to search for, locate, and retrieve information, and their use has been growing fast
(Moukdad, 2003). These engines handle English queries more or less in the same way, but their
handling of non-English queries is different from how these queries are handled by non-English
search engines that were designed for specific languages. Most general search engines enable
people to search using the English language. But people usually want to get the information they
need in the language that they can understand. In Ethiopia more than 80 languages are there, from
these languages Afaan Oromo is the most popular and widely spoken language by Oromo people
(Ibrahim, 2015).
Today, in Ethiopia, Afaan Oromo is among the major languages that are widely spoken and it is
considered to be one of the five most widely spoken languages from thousand languages of Africa
(Abera, 2009). Afaan Oromo is also relatively distributed within Ethiopia and some neighboring
countries like Kenya and Somalia (Ibrahim, 2015). Afaan Oromo is part of the Lowland East
Cushitic group within the Cushitic family of the Afro-Asiatic phylum (Melvyn, 2009), unlike
Amharic (an official working language of Ethiopia) which belongs to the Semitic language family.
Afaan Oromo has a very rich morphological variation like other African and Ethiopian languages
(Ibrahim, 2015). Therefore Afaan Oromo needs improved retrieval system.
In fact, there is no perfect Afaan Oromo Information Retrieval system that can take into account
semantically related words during searching (Kula et al., 2007). There are two sides’ issues
regarding the semantic relation of keywords; as broadly classified into synonymy and polysemy
(Thomas, 1999). Synonymy describes the fact that there are many ways to refer to the same object.
Users in different contexts or with different needs, knowledge, or linguistic habits will describe
the same information using different terms (Thomas, 1999). Polysemy refers to the general fact
that most words have more than one distinct meaning (Thomas, 1999). This can be when the same
term is used in different contexts or by different people.
There are various methodologies and techniques applied to come up with an effective and efficient
way of retrieving documents from a very large and unstructured corpus (Recardo, 1999). Some of
the scientific approaches to resolving the problems are ontology-based Information Retrieval
System, latent semantic indexing, query expansion, and reformulation.
Latent Semantic Indexing (LSI) is one method that helps to overcome the problems of lexical
matching (Christopher, 2009). It assumes that there is a latent structure in word usage that words

2
are partially obscured by variability in word choice. On the other hand, most of the current
document clustering algorithms are consider the semantic relationships which produce clustering
results (Neepa and Sunita, 2012). This clustering plays an important role in providing intuitive
navigation and browsing mechanisms by organizing such large amounts of information into a small
number of meaningful clusters (Andreas et al., 2003).
Here this study attempt has been done to design and develop an enhanced Afaan Oromo
Information Retrieval System; mainly consisted of three components; latent semantic indexing, K-
means clustering, and searching. Latent Semantic Indexing is an efficient vector space retrieval
approach, uses the Singular Value Decomposition (SVD) technique to reduce the rank of the
original term-document matrix TFIDF. The low-rank approximation yields a new representation
for each document in a collection. on the other hand document clustering was investigated for
improving the precision and/or recall in information retrieval systems by automatically grouping
documents that belong to the same topic in order to provide user’s browsing of retrieval results
(Adrian Kuhn et al., 2006). More recently, clustering has been proposed for use in browsing a
collection of documents or in organizing the results returned by a search engine in response to a
user’s query (Ramampiaro and Paulsen, 2009). From flat clustering techniques, K-means is widely
used in document clustering. K-means is based on the idea that a center point can represent a
cluster.
1.2. Statement of the Problem
Afaan Oromo is among the major languages that are widely spoken in Ethiopia and it has its own
writing system (Abera, 2009). With regard to the writing system, Qubee (Latin-based alphabet)
has been adopted and became the official script of Afaan Oromo since 1991 (Abera, 2009).
Currently, Afaan Oromo is an official language of Oromia Regional State and used as an
instructional media for primary and secondary schools of the region (Ibrahim, 2015). Furthermore,
literature, newspapers, magazines, educational resources, news, online education, books,
entertainment Media, videos, pictures, official documents, and religious writings are available on
the internet and offline. Therefore, these huge amounts of information available in an electronic
format both on the internet and offline need a potentially powerful Information Retrieval System
with good performance.

Accordingly, there were several research works done in Afaan Oromo Information Retrieval to
enhance the effectiveness of the system, especially on the text retrieval system. From those,

3
Wegari, (2017) conducted a rule-based root generation system for Afaan Oromo. In this
investigation, a rule-based method was used to develop a root generation system for Afaan Oromo
and is mainly used as a starting point to develop a complete morphological analysis and
information retrieval for the target language. Berhan, (2019) Explored the potential applications
of Latent Semantic indexing approach in Afaan Oromo text retrieval. The goal of the study was to
examine the benefits of LSI text retrieval approach for Afaan Oromoo text retrieval. LSI was used
to creat a concept space such that this concept space is much smaller than the word space
(mapping of terms against documents). Melkamu, (2017) integrated query expansion for
enhancing the effectiveness of the Afaan Oromo text retrieval system. Melkamu applied the
designed query expansion for the Afaan Oromo information retrieval system using WordNet that
was constructed as a reference for identifying the sense and meaning of the user’s query based on
word sense disambiguation by semantic similarity measure. Recently, Ashenafi, (2019) has made
attempts to extend the application of word sense disambiguation based query expansion. In this
study word sense disambiguation based query expansion approach was used to determine the
senses of words in queries by using Afaan Oromoo lexical resource.

However, these research works were more or less not consider the semantic nature of the language.
Having words in the information by itself does not meet the user’s information needs. It happened
because of different reasons but the major one is not properly addressed the semantic natures of
the language. The systems simply represent documents and queries as a “bag-of-words” with its
weight (Bo-Yeong, 2013), when users give some query the system tried to match those documents
which have terms in the query and the documents will be automatically retrieved but not otherwise.
Semantic and contextual understanding of the user information need and information in the
collection has to be well-managed to return an effective result to the user. There are several
problems tied to this simple “bag-of-words” representation of documents (Bo-Yeong, 2013). One
of them is semantic indexing (polysemy and synonymy) and the other is document clustering
(some documents are not retrieved because they do not share terms with the query).
Hence the aim of this study is to design enhanced Afaan Oromo Information Retrieval system
using semantic indexing and document clustering for organizing the Afaan Oromo document
corpus based on their similarity for effective searching to retrieve relevant documents.

To this end, this study attempts to explore and answer the following research questions:

4
  What is the suitable text preprocessing applied to document corpus for semantic indexing
document clustering based searching?
 What is the best approach for semantic indexing and document clustering based searching
to enhance Afaan Oromo Retrieval system?
 To what extent the proposed prototype improve the performance of the Afaan Oromo
Information Retrieval system?
1.3. Scope and Limitation of the Study
Semantic indexing and document clustering based searching was includes; semantic indexing,
document clustering, and searching Afaan Oromo textual document from the corpus. To do these,
the study was carried out using Latent Semantic Indexing (LSI) and K-means clustering. LSI is a
variant of the vector retrieval method or SVD that exploits dependencies or “semantic similarity”
between terms. LSI became famous as one of the first IR techniques exhibiting effectiveness in
dealing with the problems of synonymy, polysemy, and dimension reduction (Frakes and Baeza-
Yates, 1992). It helps with multiple meanings because the meaning of a word can be conditioned
not only by other words in the document but by appropriate words in the query not used by the
author of a particular relevant document. K-means clustering which basically included (dataset
preprocessing, labeling for clustering, and formation of clusters). The document that is related to
each other in a particular cluster is placed under that cluster. Thus, based on the relatedness among
the documents and the initial cluster of the clusters were formed.

The semantic document clustering based Afaan Oromo retrieval system has been developed and
tested using Afaan Oromo news article documents that were collected from Oromia Broadcast
Network (OBN). As a result of the time factor, it was difficult to prepare all relevant Afaan Oromo
documents. Therefore limited corpus were selected to developing and evaluating the performance
of this study. On the other hand, data types such as image, video, audio, and graphics are out of
focus of this study.

1.4. Significance of the Study

The semantic document clustering based Afaan Oromo retrieval system helps Afaan Oromo
information seekers to retrieve information needed and improve its performance. The system
retrieves as per the information needs of the user by retrieving documents which have similarity
to the query formulated by the user rather than exact terms match. The primary and target

5
beneficiaries of this are Afaan Oromo native speakers who can read, understand, and fluent enough
to produce queries to search for the information they need. Thus, users can able to search and
retrieve Afaan Oromo documents. In addition, the computational cost of the retrieval system was
also reduced. This happened during the system looking for a query result, it is not going through
the whole documents instead it searches only from a specific cluster which depends on specific
relation and dimension reduction. The system was also contributes to future researchers in the area
of Information Retrieval, especially the ontology-based Information Retrieval system. Generally,
the research outcome gave benefits to individuals, groups, and future researchers.

1.5. Objective of the Study

1.5.1. General Objective
The general objective of this study is to design enhanced Afaan Oromo information retrieval using
semantic indexing and document clustering based searching.

1.5.2. Specific Objectives

In order to accomplish the general objective the following specific objectives were formulated:

To perform text preprocessing and extract Afaan Oromo lexical texts.

To identify and experiment better semantic indexing and document cluster based searching
algorithm.
To design the architecture and develop the proposed information retrieval system for Afaan
Oromo.
To evaluate the performance of the proposed system using Information Retrieval
effectiveness measures (such as precision, recall, and F-measure).

6
 2. LITERATURE REVIEW
2.1. Overview
This section presents an overview of Information Retrieval System, indexing, document clustering,
searching and Afaan Oromo linguistic feature. This is broadly divided into three sections. The first
section discusses concepts related to IR text preprocessing, indexing, components of the IR system,
and document clustering. Concepts like extraction semantics of the term and term weighting are
also included. In the second section, the features of the Afaan Oromo writing system related to
Information Retrieval is briefly discussed. The Afaan Oromo alphabets, numbers, punctuation
marks, morpheme, and parts of speech of Afaan Oromo are also introduced. The third section
focuses on research works related to the Afaan Oromo Information Retrieval system.
2.2. History of Information Retrieval System
The need to store and retrieve written information became increasingly important over the
centuries, especially inventions like scientific paper (Christopher et al., 2008). Soon after
computers were invented, people realized that they could use the machine for storing and
mechanically retrieving large amounts of information. In 1945 Vinegar Bush published a
groundbreaking article titled “As We May Think” that gave birth to the idea of automatic access
to large amounts of stored knowledge. In the 1950s, this idea materialized into more concrete
descriptions of how to archives the text could be automatically searched (Christopher et al., 2009).
Several key developments in the field happened in the 1960s. The other most notable was the
works of Gerard Salton and his students; they were laid the foundation for the retrieval system.
They developed a Retrieval System and formulated a technique to evaluate the IR system, which
is using still today (Singhal, 2008).
On the other hand the perceptions towards Information Retrieval was greatly changed due to
introduction of the World Wide Web at the beginning of the 1990s and Information Retrieval
becoming the dominant form of information access, overtaking traditional database-style
searching (Christopher et al., 2008). It is defined as the process of looking for material (usually
documents) of an unstructured nature that satisfies an information need of the user from within
large collections (usually stored on computers) (Christopher et al., 2009). IR can also cover other
kinds of data and information beyond textual documents; it could be the image, multimedia, or
other data.

7
An Information Retrieval system does not inform (i.e. change the knowledge of) the user on the
subject of his inquiry (Rijsbergen, 1979). It merely informs on the existence (or non-existence)
and about documents relating to the user request. However, the computer is not likely to have
stored the complete text of each document in the natural language in which it was written
(Rijsbergen, 1979). It will have, instead, a document representative that may have been produced
from the documents either manually or automatically.
Information Retrieval applications concerning textual documents use automatically generated free
text index terms (post-coordinated), which are weighted by the statistical frequency of terms in
documents and collections (Marco and Kathleen, 2009; Baeza-Yates et al., 2004). For many
authors, the purpose of an Information Retrieval (IR) system is to process files of records and
requests for information and identify and retrieve from the files certain records in response to the
information requests (Marco and Kathleen, 2009).
2.3. Information Retrieval System Models
Potsdam, (2007) noted that, a retrieval model specifies representations used for documents and
queries, and how they are compared. An Information Retrieval model is a formalization of the way
of thinking about Information Retrieval (Potsdam, 2007). Such formalism can be defined in form
of algorithms, mathematical formulas, etc. A document model is a formal document representation
used by an Information Retrieval model to obtain document similarities (Potsdam, 2007). The
three most commonly known models are the vector space model, the probabilistic models, and the
inference network model (Recardo, 1999) as discussed below one after the other.
2.3.1. Vector Space Model
The Vector Space Model (VSM) has been a standard model of representing documents in
Information Retrieval system for almost three decades (Salton and McGill, 1983; BaezaYates and
Ribeiro-Neto, 1999). In this model, the text is represented by a vector of terms from the documents
and query (Recardo, 1999).
A document collection containing a total of n documents described by m terms is represented as a
term by document matrix in table 2.1. Each row of this matrix is a term vector and each column of
this matrix is a document vector. The element at row i, column j, is the weight of term j in the
document (Prabhakar et al., 2008; Baeza et al., 1982).

8
Table 2.1 term-document matrix (Prabhakar et al., 2008)
Term list Doc1 Doc2 Doc3 Doc4 … Doc n
Term1 w11 w12 w13 w14 … w1n
Term2 w21 w22 w23 w24 … w2n
Term3 w31 w32 w33 w34 … w3n
Term4 w41 w42 w43 w44 … w4n
Term5 w51 w52 w53 w54 … w5n
… … … … … … …
Term m wm1 wm2 wm3 wm4 wnm

Vector Space Model has been introducing a term-weight scheme known as tf-idf weighting. These
weights have a term frequency (tfi,j) factor measuring the frequency of occurrence of the terms in
the document or query texts and an inverse document frequency (idfi) factor measuring the inverse
of the number of documents that contain a query or document term (Sani et al., 2014; Baeza et al.,
1982);

𝑊𝑖, 𝑗 = 𝑡𝑓𝑖, 𝑗 𝑥 𝑖𝑑𝑓𝑖 (2.1)

𝑓𝑟𝑒𝑞𝑖,𝑗
Where 𝑡𝑓𝑖, 𝑗 = (2.2)
max
𝑓𝑟𝑒𝑞𝑖𝑗

𝑁
And 𝑖𝑑𝑓𝑖 = 𝑙𝑜𝑔 ( ) (2.3)
𝑛𝑖

freqi,j is the number of occurrences of terms ti in document dj, N is the number of documents in a
collection, and ni is the document frequency for term ti in the whole document collection.
To assign a numeric score to a document retrieved by a query, the model measures the similarity
between the query and vector created by vocabulary in the index versus documents in the corpus
(Singhal, 2008). Typically, the angle between the two vectors is used as a measure of divergence
between the vectors. The similarity can be measured using, such as cosine similarity (dot product)
(Singhal, 2008). When using a search engine the user wants to retrieve relevant documents (Rosell,
2006), he or she gives some keywords, a query, as input, this query, gets represented in the same
(or a similar) way as the texts, i.e. we get a vector q in the vector space representing the query. The
idea is that the relevant texts are those closest to the query in the vector space. The most common
measure of closeness or similarity is the cosine measure, the cosine of the angle between the query
and a text (Baeza et al., 1982):

9
 t
∑ Wqj ∗ W
dj.q j=1 dij
sim(dj, q ) = || dj|| ||q|| = (2.4)
t 2 t 2

√∑ (W ) ∗∑ (W )
qj dij
j=1 j=1

where documents and queries are represented as vectors: dj = (w1,j, w2j, … … … wt,j)

and q= (w1,j, w2j, … … … wt,q )

2.3.2. Probabilistic Model

Probabilistic IR models estimate the probability of relevance of documents for a user queries
(Bruce, 1995), this estimation is the essential part of the model, and this is a point where
probabilistic models different from others. It is based on the general principle that documents in a
collection should be ranked by decreasing the probability of their relevance to a query or user
information needs (Bruce, 1995). The probabilistic models work based on basic probability
notation, the probability of relevance for the document in the log (Fuhr and Buckley, 1991). The
parameters of the models relate to elements of the underlying representations (Fuhr and Buckley,
1991). The probabilities are estimated on the frequency of terms in each document, and they make
use of user-determined relevance judgments (Salinas, 2009). In probabilistic relevance, the index
term weight variables are all binary that a user query q is a subset of index terms (Sani et al., 2014).
Let R be the set of documents known to be relevant and Let R’ be the complement of R (that is,
the set of non-relevant documents) (Sani et al., 2014), P(R/d) be the probability that the document
d is relevant to query q and P (R’/d) be the probability that d is non-relevant to q. then the similarity
sim (d,q) of the documents d to the query q is defined by the ratio of relevant document to non-
relevant for a given query (Sani et al., 2014; Baeza et al., 1982).
𝑅
𝑃( )
𝑑
𝑠𝑖𝑚(𝑑, 𝑞) = 𝑅′
(2.5)
𝑃( )
𝑑

By using Bayes’ rule equation defined above can be expressed as follows;

𝑑
𝑃( )𝑃(𝑅)
𝑅
𝑠𝑖𝑚(𝑑, 𝑞) = 𝑅′
(2.6)
𝑃( )𝑃(𝑅 ′ )
𝑑

P(d/R) stands for the probability of randomly selecting the document d from the set R of relevant
documents. P(R) is a probability document that is randomly selected from the complete set of

10
documents that is relevant. The meaning attached to P(d/R’) and P(R’) is analogous and
complementary.

2.3.3. Inference Network Model

Inference Network Model attempts to model document retrieval as an inference process (Singhal,
2008). In the simplest implementation of this model, a document instantiates a term with a certain
strength, and the credit from multiple terms is accumulated given a query to compute the equivalent
of a numeric score for the document (Recardo, 1999). From an operational perspective, the strength
of the instantiation of a term for a document can be considered as the weight of the term in the
document, and document ranking based on the term strength of the documents. The basic network
model for information retrieval consists of two-component networks (Robert, 1991); a document
network, and a query network. The document network represents the document collection using a
variety of document representation schemes. The document network is built once for a given
collection and its structure does not change during query processing. The query network consists
of a single node that represents the user’s information need and one or more query representations
that express that information need. A query network is built for each information need and is
modified during query processing as existing queries are refined or new queries are added in an
attempt to better characterize the information need.
Document and query networks are joined by links between representation concepts and query
concepts (Robert, 1991), as shown in figure 2.1. All nodes in the inference network are binary
valued and take on values from the {false, true} (Bonald and Bruce, 2004)
Fig. 2.1 Basic inference network model (Robert, 1991)

11
The document network consists of document nodes (di’s), text representation nodes (ti’s), and
concept representation nodes (rk’s). Let D be the set of documents, T be the set of text
representations, and R be the set of representation concepts, where the cardinality of these sets in
nd, nt, and nr, respectively, then the event space represented by document network is Ed = DxTxR
(Donald and Bruce, 2004). Since all propositions are binary-valued, the size of the event space is
2nc. 2nt.2nr and each representation node contains a specific of the conditional probability
associated with the node given its set of parent text nodes (Robert, 1991). This specification
incorporates the effect of any indexing weights (for example, term frequency for each parent text)
or term weights (for example, inverse document frequency) associated with the representation
concepts.
The query network is an “inverted” directed acyclic graph with a single leaf corresponds to the
event that an information need is met and multiple roots that correspond to the concepts to the
concepts that express the information need (Singhal, 2008). A set of intermediate query nodes may
also be used in cases where multiple query representations are used to express the information
need. These nodes are a representation of convenience; it is always possible to eliminate them by
increasing the complexity of the distribution specification at the node representing the information
need (Singhal, 2008). If we let C represent the set of query concepts and Q represents the set of
queries where nc, and nq are the cardinalities of these sets, then the event space represented by the
query network is Eq = CxQxI. Since we can always eliminate query nodes |Eq| ≤ 2nc+1. The event
space represented by the entire inference network is then EdxEq.
2.4. Information Retrieval Process
Information Retrieval can be subdivided in many ways; it seems that there are three main areas of
research which between them make up a considerable portion of the subject (Rijsbergen, 1979).
They are content analysis, information structures, and evaluation. Briefly the first is concerned
with describing the contents of documents in a form suitable for computer processing; the second
is exploiting relationships between documents to improve the efficiency and effectiveness of
retrieval strategies; the third with the measurement of the effectiveness of retrieval. The term
information structure covers specifically a logical organization of information, such as document
representatives, for Information Retrieval (Rijsbergen, 1979). Figure 2.2 presents the generic
architecture of an information retrieval system with indexing and searching phases.

12
Fig. 2.2. Architecture of a textual IR system (Ceri et al., 2013).

2.4.1. Text operations

Textual IR exploits a sequence of text operations that translate the user’s need and the original
content of textual documents into a logical representation more amenable to indexing and querying
(Ceri et al., 2013). Textual operations translate the user’s need into a logical query and create a
logical view of documents (Ceri et al., 2013). Text operations are the task document preprocessing
which is a stage before a document model construction when data is analyzed, transformed, and
filtered from the document content (Potsdam, 2007).
During the preprocessing phase, some important text operations can be performed (Rijsbergen,
1979; Christopher et al., 2009). It is performed to control the size of the vocabulary. Some of the
major text operations are the following.
Lexical analysis: is a process of collecting tokenized terms that needs some other operation
(Singhal, 2008; Christopher et al., 2009). It is the process of converting a stream of characters (the
text of the documents) into a stream of words (the candidate words to adopt as index terms)
(Christopher et al., 2009). Thus, one of the major objectives of the lexical analysis phase is the
identification of the words in the text. It also incorporates a sort of text cleaning process, such as
the conversion of abbreviations and acronyms into their full text, conversion of cases, removal of
numbers and symbols (Matsumura et al., 2002), (example $, @, %, #, etc) which do not make up
good index terms.

13
Elimination of stop words: is a process of avoiding a word which has fewer relevancies to
determine the content of the document (Potsdam, 2007). Sometimes, some extremely common
words and non-content bearing words have little value in selecting documents matching a user
need are excluded from the vocabulary entirely. In English language, these words are most of the
time articles, prepositions, pronouns, conjunctions, and others. The elimination of stopping words
is a very critical task to control the size of the vocabulary (Christopher et al., 2009; Singhal, 2008).
Stop words usually refers to the most common words in a language, there is no single universal
list of stop words used by all-natural language processing tools, and indeed not all tools even used
by such a list (Singhal, 2008). However, the most common stop words in English language
includes; articles (a, an, the), prepositions (in, on of …), conjunctions (and, or but, if…), pronouns
(I, you, them, it) and possibly some verbs, adverbs, adjectives. Since stop words have no
discriminatory power, they can be removed from the list of index and query terms.
Stemming: It is the process of finding the root of the words given word variants (Christopher et
al., 2009). In any language, they observe the occurrence of variants of the same word into little
modification in its morpheme, such as prefix, suffix and circumfix (Unubi et al., 2017). Prefix is
attached to the beginning of a root word, i.e., in a word untouch, un is a prefix. Suffix is on the
other hand attached to the end of the root word i.e., in word touchable, able is the suffix. Also,
infix occurs when an affix is inserted within a root word, it is used rarely i.e., spoonful can be as
spoonsful and s is infix and circumfix is an affix attach to both the beginning and the end of a
word, i.e., il-legal-ity, in which il as prefix and ily as a suffix.
The goal of stemming is to reduce inflectional forms and derivational forms of a word to a common
base form or it is referring to a crude heuristic process that chops off the ends of words in the hope
of achieving this goal correctly most of the time, and often includes the removal of derivational
affixes (Cambridge, 2009) For instance:
 The word ‘nation’ can be written as nationalized, nationality, national, and others by
adding suffixes.
 am , are, is => be
 car, cars, car’s, cars’ => car, so this different form of the same word comes to its root term
by applying stemming algorithms.
The most common algorithm for stemming English words, and one that has repeatedly been shown
to be empirically very effective, is porter’s algorithm (Cambridge, 2009). Porter’s algorithm

14
consists of 5 phases of word reductions, applied sequentially (Cambridge, 2009). Within each
phase, there are various conventions to select rules, such as selecting the rule from each ruling
group that applies to the longest suffix. For instance (Cambridge, 2009):
Rule Example
SSES SS caresses caress
IES I ponies poni
SS SS caress caress
S cats cat
2.4.2. Term Weighting
Term weighting for the IR model is handled by statistics (Potsdam, 2007). There are three main
factors of term weighting: term frequency factor, term collection frequency factor, and document
vector length normalization factor. The end term weight might be constructed from all or a subset
of mentioned factors. For example, the inverse document frequency assumes that the importance
of a term is proportional with the number of documents the term appears in (Salton, 1983), a
document that mentions a query term more often has more to do with that query and therefore
should receive a higher score. Thereof, assign to each term in a document weight for term that
depends on the number of occurrences of the term in the document (Christopher et al., 2009). This
concept can be applied through assigning the weight; it is equal to the number of occurrences of
term t in document d. This weighting scheme is referred to as term frequency, denoted by tf, and
the document in its order (Christopher et al., 2009). The exact ordering of the terms in a document
is ignored but the number of occurrences of each term is material (in contrast to Boolean retrieval).
We only tried to capture information concerning the number of occurrences of each term
(Christopher et al., 2009).

Inverse Document Frequency: It will create a problem when we make all terms are equally
important; the impact would be explicitly viewed when it comes to assessing relevancy against a
query (Christopher et al., 2009). In fact certain terms have little or no discriminating power in
determining relevance. An immediate idea is to scale down the term weights with high collection
frequency, defined to be the total number of occurrences of a term in the collection. The idea would
be to reduce the tf weight of a term by a factor that grows with its collection frequency (Christopher
et al., 2009). Document frequency of the term, dft, defined to be the number of documents in the

15
collection that contain a term t aimed to obtain document level statistics, which deals with the
number of documents containing a term (Christopher et al., 2009). Using dft is difficult to measure
the discrimination power of the term among the documents. To overcome this problem, it is better
to use inverse document frequency; this is the logarithmic function of the total number of document
N over dft (Christopher et al., 2009). It is defined as (Baeza et al., 1982);

𝑖𝑑𝑓 = log 𝑁/𝑑𝑓 (2.5)

TFIDF weighting: Combining the definitions of term frequency and inverse document frequency
helps to produce a composite weight for each term in each document. The TFIDF weighting
scheme assigns to term t in document d given by (Baeza et al., 1982):

𝑇𝐹𝐼𝐷𝐹 = 𝑇𝐹 ∗ 𝐼𝐷𝐹 (2.6)

TFIDF t, d assigns to term t is a weight in document d that is;

 Highest when t occurs many times within a small number of documents (thus lending high
discriminating power to those documents);
 Lower when the term occurs fewer times in a document, or occurs in many documents
(thus offering a less pronounced relevance signal);
 Lowest when the term occurs in virtually all documents.
2.4.3. Indexing Approaches
An indexing language is a language used to describe documents and requests (Rijsbergen, 1979),
the elements of the index language are index terms, which may be derived from the text of the
document to be described, or maybe arrived at independently. The transformation from a document
text into a representation of text is known as indexing the documents (Sajendra and Ram-Kumar,
2012). Indexing can be defined as a process that automatically collects, parse, and stores data to
facilitate fast and accurate Information Retrieval (Clarck and Cormack, 1995).
Indexing techniques have been developed in order to make possible the identification of the
information content of documents (be they text documents, hypermedia, or multimedia ones) (Lei
Shi, 2005). It simply means pointing to or indicating the content, meaning, purpose, and features
of messages, texts, and documents (Xiao, 2010).
Automatic indexing is the ability of a computer to scan large volumes of documents against a
controlled vocabulary, taxonomy, thesaurus, or ontology and use those controlled terms for quickly

16
and effectively index large document collections (Rijsbergen, 1979). Typically the indexing of a
textual document is obtained through the identification of a set of terms or keywords which
characterize the document content that describe the topics dealt with in the document (LeiShi,
2005). The terms included in this set have not only to be representative of the topics covered in
the documents, but they also need to be distinguishing the possible way to discriminate one
document against the other documents in the collection covering the same or similar topic.
There are two major approaches for the automatic indexing of text documents (Lei Shi, 2005);
linguistic and statistical approaches.

2.4.3.1. Linguistic Approach

The linguistic approach involves syntactical analysis (Aurélie et al., 2008). It is based on the
knowledge base of the language it is working on (Latifur et al., 2013). Therefore, it needs ontology
of the language to determine the terms which describe a document. It can use semantics by
extracting content bearing words from the document that is going to be indexed (Latifur et al.,
2013). The use of ontology enables one to define concepts and relations representing knowledge
about a particular document in domain-specific terms (Paralic and Kostial, 1999). In order to
express the contents of a document explicitly, it is necessary to create links (associations) between
the document and relevant parts of a domain model, i.e. links to those elements of the domain
model, which are relevant to the contents of the document.
Ontology is a collection of concepts and their interrelationships that can collectively provide an
abstract view of an application domain (Latifur et al., 2013). The ontology (WordNet) lexical
database is now quite large and offers broad coverage of general lexical relations in the English
language (David et al., 2013). It describes the relationship between the words in different
situations. WordNet has been employed as a resource for many applications in natural language
processing (NLP) and Information Retrieval (IR). Word relationships in the database are useful
for NLP and IR applications, are not necessarily appropriate for a general, sometimes the WordNet
can be a domain-independent lexical database (Marti, 2008; Singhal, 2008). For information
retrieval purpose ontology enables the concept semantic indexing which is based language
knowledge base. Semantic indexing is used improve the performance of a traditional keyword-
based search, documents should be represented with their concept rather than bag-of-words‘(Bo-
Yeong, 2013). It assumes that there's some underlying or context structure in word usage that's
partially obscured by variability in word selection (Ding, 1999). However; most of the previous

17
works on indexing and Information Retrieval depend on lexical analysis and statistical methods.
Using these techniques is difficult to abstract the semantics of the documents, since it needs the
language expert (Bo-Yeong, 2013). A better approach would allow users to retrieve information
on the basis of a conceptual topic or meaning of a document (Barbara, 2000). One way of
expressing their commonality is to think of a searcher as having in mind a certain meaning, which
he or she expresses in words, and the system as trying to find a text with the same meaning
(Deerwester et al., 1989). Then, depends on the system representing query and text meaning in a
manner that correctly reflects their similarity for the human.

2.4.3.2. Statistical Approach

The statistical approach relies on various word counting techniques and vector space model. The
aim of statistical indexing is to capture content bearing words that have a good discriminating
ability and a good characterizing ability for the content of a document (Cambridge, 2010).
Latent Semantic Indexing (LSI) was one of an efficient vector space model retrieval approach,
uses the Singular Value Decomposition (SVD) technique to reduce the rank of the original term-
document matrix (Hofmann, 1999). Theoretically, SVD, a dimensionality reduction technique,
performs a term-to-concept mapping, and therefore, conceptual indexing and retrieval are made
possible. The main objective is not to describe the concepts verbally, but to be able to represent
the documents and terms in a unified way for exposing document-document, document-term, and
term-term similarities or semantic relationship which are otherwise hidden (Hofmann, 1999). It
tries to beat the problems of lexical matching by using statistically derived conceptual indices
instead of individual words for retrieval.
Latent semantic indexing tries to overcome the problems of diversity in the words people use to
describe the same object or concept (synonymy), and the same word often has more than one
meaning (polysemy) by automatically organizing objects into a semantic structure more
appropriate for information retrieval (Dumais, 1992). This is done by modeling the implicit higher-
order structure in the association of terms with objects. Deerwester et al., (1990) tried to overcome
the problems of lexical matching by using statistically derived conceptual indices instead of
individual words for retrieval. LSI assumes that there is some underlying or latent structure in
word usage that is partially obscured by variability in word choice (Rosario, 2000).
In the latent semantic space, a query and a document can have high cosine similarity even if they
do not share any terms - as long as their terms are semantically similar in a sense to be described

18
later (Rosario, 2000). This latent semantic space has fewer dimensions than the original space and
a method for dimensionality reduction. A dimensionality reduction technique takes a set of objects
that exist in a high-dimensional space and represents them in a low dimensional space, often in a
two-dimensional or three-dimensional space for the purpose of visualization (Rosario, 2000).
The goal of semantic indexing is to use semantic information (within the objects being indexed)
to improve the quality of information retrieval. Compared to traditional indexing methods, based
on keyword matching, the use of semantic indexes means that objects are indexed by the concepts
they contain rather than just the terms used to represent them (Marco and Kathleen, 2009);
 A semantic index is inherently multidimensional since any combination of properties cast
into a document concept can serve as an indexing element.
 As a structured concept, the indexing elements are not just attributed values but can be
based on complex descriptions of related objects.
 A semantic index as a whole is highly adaptable to patterns of usage. Indexing concepts
can be added or removed at will, making it very dense and precise with respect to
interesting sets of individuals, or very sparse in other less interesting areas.
 Since the index is actually a set of partial descriptions for the indexed objects, lots of
information can be drawn from the index alone without accessing individual descriptions
at all.
Latent semantic indexing is the application of a particular mathematical technique, called Singular
Value Decomposition (SVD), to a word-by-document matrix (Rosario, 2000). It is used to
estimate the structure in word usage across documents (Rosario, 2000). The truncated SVD, in one
sense, captures most of the important underlying structure in the association of terms and
documents, yet at the same time removes the noise or variability in word usage that plagues word-
based retrieval methods and it is represented as shown in figure 2.3 below.

19
Fig. 2.3 Singular Value Decomposition based Latent Semantic Indexing
A = U∑VT

Where: U and V are orthogonal matrices of m x n and m x m size respectively, and ∑ is diagonal
matrix of m x m. SVD finds the optimal projection to a low dimensional space is the key property
for exploiting word co-occurrence patterns.
For an m x n matrix A of rank r there is exists a factorization (Singular Value Decomposition =
SVD) as follows:

The columns of U are orthogonal eigenvectors of AAT

The columns of V are orthogonal eigenvectors of ATA
Eigenvalues λ1, λ2 … λn of AAT are the eigenvalues of ATA
∑ = diag (σ1, σ2 … σn), σi ≥ σi+1
where, σ1, σ2 … σn are singular values.
2.4.4. Query Matching and Searching
A query is a keyword that is consulted during search (Christopher et al., 2009). It is formulated by
users and compared to the index keywords to retrieve information needed. A comparison of query
and document representations is a very important task of the IR system. It is performed in order to
retrieve documents that are more similar to the specific query (Recardo, 1999). The ‘Search’ is a
systematic examination of information in a database, aiming in view to identify the items or
objects, which satisfy particular preset criteria (Juban and Falguni, 2007). In another way,
searching means the operation of locating a specific object in a given sequence of ‘n’ objects.

20
All search strategies are based on a comparison between the query and the stored documents
(Rijsbergen, 1979), sometimes this comparison is only achieved indirectly when the query is
compared with clusters (or more precisely with the profiles representing the clusters). The
distinctions made between different kinds of search strategies can sometimes be understood by
looking at the query language that is the language in which the information need is expressed and
the nature of the query language often dictates the nature of the search strategy. These strategies
are based on the common notion that the more often terms are found in both the document and the
query, the more "relevant" the document is deemed to be to the query (Grossman et al., 2004). A
retrieval strategy is an algorithm that takes a query Q and a set of documents D1, D2, ..., Dn and
identifies the Similarity Coefficient (Q, Di) for each of the documents 1≤ i ≤n and the retrieval
strategies identified are (Rijsbergen, 1979; Grossman et al., 2004):
Boolean Indexing: A score is assigned such that an initial Boolean query results in a ranking. This
is done by associating a weight with each query term so that this weight is used to compute the
similarity coefficient.
Vector Space Model: Both the query and each document are represented as vectors in the term
space. A measure of the similarity between the two vectors is computed.
Probabilistic Retrieval: A probability based on the likelihood that a term will appear in a relevant
document is computed for each term in the collection. For terms that match between a query and
a document, the similarity measure is computed as the combination of the probabilities of each of
the matching terms.
Language Models: A language model is built for each document, and the likelihood that the
document will generate the query is computed.
Inference Networks: A Bayesian network is used to infer the relevance of a document to a query.
This is based on the "evidence" in a document that allows an inference to be made about the
relevance of the document. The strength of this inference is used as the similarity coefficient.
Latent Semantic Indexing: The occurrence of terms in documents is represented with a term-
document matrix. The matrix is reduced via Singular Value Decomposition (SVD) to filter out the
noise found in a document so that two documents that have the same semantics are located close
to one another in a multidimensional space.
Neural Networks: A sequence of "neurons," or nodes in a network, that fire when activated by a
query triggering links to documents. The strength of each link in the network is transmitted to the

21
document and collected to form a similarity coefficient between the query and the document.
Networks are "trained" by adjusting the weights on links in response to predetermined relevant
and irrelevant documents.
Genetic Algorithms: An optimal query to find relevant documents can be generated by evolution.
An initial query is used with either random or estimated term weights. New queries are generated
by modifying these weights. A new query survives by being close to known relevant documents
and queries with less "fitness" are removed from subsequent generations.
Fuzzy Set Retrieval: A document is mapped to a fuzzy set (a set that contains not only the
elements but a number associated with each element that indicates the strength of membership).
Boolean queries are mapped into a fuzzy set intersection, union, and complement operations that
result in a strength of membership associated with each document that is relevant to the query.
This strength is used as a similarity coefficient
2.5. Document Clustering
Clustering is the process of partitioning a set of data objects into subsets (Recardo, 1999). It is a
sub-field in artificial intelligence and machine learning that refers to a group of algorithms that try
to find a natural grouping of objects based on some objective metric (Christoper, 2010). Clustering
is the most common form of unsupervised learning (Cambridge, 2009). No super-vision means
that there is no human expert who has assigned documents to classes.

Clustering is a powerful technique for large-scale topic discovery from the text (Bjornar and
Chinatsu, 1999). It involves two phases: feature extraction and applies clustering algorithms.
Feature extraction maps each document or record to a point in high-dimensional space, and then
clustering algorithms automatically group a set of documents based on their similarity and
dissimilarity.

2.5.1. Approaches of Document Clustering

The two main types of cluster analysis approaches are hierarchical and flat clustering. Hierarchical
clustering is another unsupervised learning algorithm that is used to group together the unlabeled
data points having similar characteristics (Recardo, 1999). There are two basic methods of
hierarchical clustering (Recardo, 1999): the first one is Agglomerative, which start with the points
as individual clusters and, at each step, merge the most similar or closest pair of clusters. This
requires a definition of cluster similarity or distance. The other one is Divisive, which starts with

22
one, all-inclusive cluster, and, at each step, split a cluster until only singleton clusters of individual
points remain. In this case, we need to decide, at each step, which cluster to split and how to
perform the split.

The other one is flat clustering. A flat/nonhierarchical clustering are creates a flat set of clusters
without any explicit structure that would relate clusters to each other (Cambridge, 2009). From flat
clustering K-means is perhaps the most widely used hard clustering algorithm due to its simplicity
and efficiency (Cambridge, 2009). The term K-means was first used by James Macqueen in 1967
(Ramakrishna at el., 2014). K-means is partitioning based clustering algorithm, it groups the
objects in continuous n-dimensional space, which uses centroid as mean of the group of objects
(Bader et al., 2009).

K-means divide a data set of N items into K clusters, and the hierarchical, which produce a nested
data set in which pairs of items or clusters are successively linked (Edie, 2015). Unlike hierarchical
clustering, K-means (partitioned) clustering directly divides data objects into K clusters, without
any corresponding hierarchical structure (Bader et al., 2009). Let P= {Pi}, i=1...n be the set of data
points, to be clustered into a set of K number of clusters without any prior knowledge of the input
objects. Predefined number of groups is indicated with K, where K is provided as an input
parameter. Assigning of each object to a cluster is based on the objects proximity to the mean of
the cluster. Then the mean of the cluster is in turn recomputed and the process of assigning objects
to cluster resumes. It is done by minimizing the sum of squares of distances between objects and
their corresponding cluster centroid. It groups the objects into K clusters and keeps iteratively
moving the cluster centers and re-assigning objects into clusters, based on minimum distance to
the closest cluster’s centroid. The process terminates when cluster centers are not moved anymore
and all objects have been assigned to their closest cluster center.

For K-means clustering, the cosine measure is used to compute which document centroid is closest
to a given document (Michael et al., 2009). While a median is sometimes used as the centroid for
K-means clustering. There are other distance measures (Ramakrishna at el., 2014): Manhattan,
Cosine similarity and Jaccard.Cosine and Jaccard measures are appropriated for document
clustering. Cosine similarity is mostly appropriate for document similarity. The cosine similarity
is measured in the following manner.

23
 t
∑ 𝑑1 ∗ 𝑑2
d1.d2 j=1
sim(d1, d2 ) = || d1|| ||d2|| = 2 2
(2.7)
t t
√∑ (d1 ) ∗∑ (𝑑1)
j=1 j=1

Here d1 and d2 are two representative points contain n number of values. K-means clustering is
very useful in exploratory data analysis and data mining in any field of research, and as the growth
in a computer, power has been followed by a growth in the occurrence of large data sets (Chartier,
2013). Its ease of implementation, computational efficiency, and low memory consumption have
kept the K-means clustering very popular, even compared to other clustering techniques.

In the Information Retrieval (IR) field, cluster analysis has been used to group similar documents
together with the goal of improving the efficiency and effectiveness of retrieval or to determine
the structure of the literature of a field (Edie, 2015). Cluster-based retrieval has as its foundation
the cluster hypothesis, which states that closely associated documents tend to be relevant to the
same requests (Rijsbergen, 1979). In other words, documents within a cluster should be as similar
as possible; and documents in one cluster should be as dissimilar as possible from documents in
other clusters (Cambridge, 2009).

2.6. Challenges of the Language in Information Retrieval System

Lemmatization and Stemming: Morphology is the study of the structure of a word and how it is
built (Norrby, 2016). (Martin and Jurafsky, 2019) describe it as small building blocks of a word
are called morphemes and they can be divided into two groups, stems, and affixes. A stem is the
main part of a word and affixes are morphemes that are added to a stem to give different meanings
to it. In the word dogs, for example, the dog is the stem and -s is the affix. Using affixes allows a
word to occur in different forms, it can give it different inflections and derivations. Lemmatization
and stemming are two methods that can be used to handle the challenges that the variations of
words impose. Christopher et al., (2008) explain that this is done by reducing a word to a simpler
form, and even though lemmatization and stemming aim for the same goal they achieve it in
different ways. Stemming cuts of the ends of a word, to get rid of affixes and get the simple form
of a word (Norbby, 2016). For example, the word writing would with a stemmer be reduced to
“writ”. This process is not flawless and can give incorrect results. According to Christopher et al.
(2008), such like incorrect results are improve the recall while it lowers the precision.
Lemmatization does it in a more proper way, by doing a morphological analysis using a vocabulary

24
to get the dictionary form of the given word (Christopher et al., 2008). The word writing will in
this case be normalized to “write” which is the dictionary word form.
Compound Splitting: split compound words seem particularly interesting when it comes to
information retrieval of languages with that characteristic (Norrby, 2016). When splitting
compound words there are different method. (Barbel, 2004) explain that a splitting method can be
regarded as an aggressive one, which means that it will split the compound into its smallest parts.
This means that some words such as “football” will be split into “foot”, “ball”. Since documents
with the word foot or ball may have nothing to do with football, this may be a disadvantage that
causes an IR to find irrelevant documents. Such like these also happened in Afaan Oromo (i.e.,
“abbaa gadaaa” will be split into “abbaa” and “gadaa”).
Irregular morphology: antennae are only a plural of the type of antenna that is associated with
an insect, not with a television antenna (Krovetz, 2000). Similarly, media is the plural of the
medium used in the sense of entertainment, not in the sense of a spiritualist.
Part-of-speech: intimation is derived from intimate (v), and intimately is derived from intimate
(adj). Because suffixes only attach to roots with particular parts-of-speech, this information can be
used to discriminate one homograph from another.
Run-ons: These are words that are morphologically related to the headword (the word that is being
defined in the dictionary) (Krovetz, 2000). They generally appear at the end of a homograph. Run-
ons have a predictable relationship to the root form (the headword) and are primarily a way for the
lexicographer to include additional entries without taking up much space. For example, the dancer
can appear as a run-on for dance, and craftsmanship as a run-on for the craftsman. Comparing run-
on entries with main entries can help us focus on the information that is needed to disambiguate
them (e.g., boxer as a type of dog versus a human).
Synonymy and polysemy: synonymy creates a problem when a document is indexed with one
term and the query contains a different term, and the two terms share a common meaning (Savoy,
2000). Most of human knowledge is coded in natural language. However, it is difficult to use
natural language as a knowledge representation language for computer systems. The current
retrieval models are based on either keyword for search or author. This keyword representation
creates a problem during retrieval due to polysemy, homonymy, and synonymy. Polysemy
involves the phenomenon of a lexeme with multiple meanings. Keyword matching may not always
include word sense matching (Savoy, 2000).

25
2.7. Overview of Afaan Oromo
Afaan Oromo is the second most widely spoken indigenous language in Africa next to Hausa in
Nigeria (Feyisa, 1996). It is one of the highly developed languages of the Cushitic languages
spoken in Ethiopian, Somalia, Sudan, Tanzania, and Kenya (Ibrahim, 2015). From forty different
Cushitic linguistic and cultural groups in Africa, the Afaan Oromo speakers are considered as one
of the largest in terms of population and their language (Feyisa, 1996). In Oromia state, in Ethiopia,
it is the official language used in courts, schools, and administration (Feyisa, 1996).
Currently, there are a growing number of publications in hard copies and a vast amount of
information in electronic formats for Afaan Oromo (Tilahun, 2008). This was made for the
convenience of the Latin script for the writing of Afaan Oromo from the linguistics, pedagogic,
and practical reasons. It is believed that many fold more texts were written in Afaan Oromo since
then than before. Afaan Oromo is a language that is used in a wide area in the country. According
to Gragg, (1976), quoted in (Wakshum, 2000), four major categories can be identified. These are
Western (Wellega, Iluababor, Kaffa, and parts of Gojjam), Eastern (Harar, Eastern showa, and
parts of Arsi and Bale), Central (Central Showa, Western Showa, and possibly Wollo), and
Southern (Parts of Arsi, Sidamo, and Borena).
2.7.1. Features of Afaan Oromo writing system
Afaan Oromo Writing Style: According to (Ladefoged, 1955), quoted in (Morka, 2001) some of
the finer anatomical feature involved in speech production include the vocal cords, velum, tongue,
teeth, palates, the alveolar ridge, the mouth, and lips. These anatomical components move to
different positions to produce various sounds and are referred to in articulators. Most of the
characters of a sound are determined by the position of these articulators in the oral tract.

The Oromo Alphabet and sounds (Qubeeleewwan fi Sagaleewwan): The alphabets of Afaan
Oromo are often called “Qubee Afaan Oromo (Ibrahim, 2015). Qubee has 33 characters
representing distinct sounds. It has both capital and small letters. Afaan Oromo has a considerable
amount of glottal stops. The major representatives of sources of the sound in a language are the
vowels and consonants. Afaan Oromo is a phonetic language, which means that is spoken in the
way it is written (Wakshum, 2000). The Afaan Oromo vowels represented by letters (a, e, o, u, and
i) are called “Dubbiftu/Dubbachiftu” in Afaan Oromo, and the consonants known as “dubbifamaa”
in Afaan Oromo. In the Afaan Oromo alphabet, a letter consists either of a single symbol or a

26
digraph (ch, dh, ny, ph, sh, ts zh). Gemination is not obligatorily marked for the digraphs
(Encyclopedia, 2018).

Table 2.2. The qubee in international phonetic writing (IPA) (Ibrahim, 2015)
Qubee IPA Qubee IPA Qubee IPA
A a /a/ L l /l/ W w /w/
B b /b/ M m /m/ X x /x/
C c /c/ N n /n/ Y y /y/
D d /d/ O o /o/ Z z /z/
E e /e/ P p /p/ CH ch /ch/
F f /f/ Q q /q/ DH dh /dh/
G g /g/ R r /r/ NY ny /ny/
H h /h/ S s /s/ PH ph /ph/
I i /i/ T t /t/ SH sh /sh/
J j /j/ U u /u/ TS ts /ts/
K k /k/ V v /v/ ZH zh /zh/

Consonants (Dubbifamaa) and Vowels (Dubbachiiftuu) phonemes: Afaan Oromo has the
typical Southern Cushitic set of five short (a, e, i, o, u) and five long vowels, indicated in the
orthography by doubling the five vowel letters (aa, ee, ii, oo, uu). The difference in length of
vowels results in change of meaning (Mewis, 2001). The Afaan Oromo vowels always are
pronounced in sharp and clear fashion which means each and every word is pronounced strongly,
for example;

Table 2.3 Afaan Oromo Vowels

Vowels Afaan Oromo English
A Hara lake
Haaraa new
E Eegee Waiting
ijoollee Children
I amajji May
Ilaali look
O oromoo Oromo
Haaloo revenge
U Utubuu, beekuu Support, knowing

27
Gemination (doubling a consonant) is also significant in Afaan Oromo. That is, consonant length
can distinguish words from one another (Tesfaye, 2010) Example:

Table 2.4 Gemination (doubling a consonant) in Afaan Oromo

Afaan Oromo English

Badaa bad
Baddaa highland
Hatuu cheat/steal
Hattuu thief

Punctuation Marks: Words in Afaan Oromo sentences are separated by white spaces the same
way as it is used in English. Different Afaan Oromo punctuation marks follow the same
punctuation pattern used in English and other languages that follow the Latin writing system
(Megersa, 2002). For example, comma (,) is used to separate listing of ideas, concepts, names,
items, etc and the full stop (.) in the statement, the question mark (?) in interrogative and the
exclamation mark (!) in command and exclamatory sentences mark the end of a sentence. All
punctuation marks that are used in English are also used for the same purpose in Afaan Oromoo
except the apostrophe. Unlike its use to show possession in English, it is used as a symbol to
represent a hiccup (called hudhaa) sound in Afaan Oromo writing. An apostrophe, and less
commonly a hyphen, is used “’” represent this sound in writing. Sometimes an H, which represents
the closest glottal sound, is also used in place of an apostrophe (Ibrahim, 2015), for a reason to be
apparent later, the apostrophe will be considered as a distinct symbol (say, as the 27th letter of the
alphabet).

Personal pronoun: Oromo pronouns include personal pronouns (refer to the persons speaking,
the persons spoken to, or the persons or things spoken about), indefinite pronouns, relative
pronouns (connect parts of sentences) and reciprocal or reflexive pronouns (in which the object of
a verb is being acted on by verb's subject).

28
Table 2.5 Afaan Oromo personal pronouns
English Base Subject Dative Instrumental Locative Ablative Possessive
adjectives
I ana, na ani, an naa, naaf naan natti narraa koo, kiyya
natti
you (sg.) si Ati sii, siif, siin sitti sirraa kee
sitti
he isa Inni isaa, isaattin isatti isarraa isaa
isaa(itti)f
,isaatti
she isii, isiin, etc ishii, ishiin ishiitti,etc ishiirraa isii
ishii, ishiif, ishii
ishee ishiitti
we nu nuti, nuu, nuun nutti nurraa keenna
nu’i, nuuf, keenya
nuy, nu nutti
you (pl.) isin Isini isinii, isiniin isinitti isinirraa keessan
isiniif,
isinitti
they isaan Isaani isaanii, isaaniitiin isaanitti isaanirraa isaani
isaaniif,

Adjectives: Adjectives are very important in Afaan Oromo because its structure is used in every
day conversation (Gezehagn, 2012). Afaan Oromo Adjectives are words that describe or modify
another person or thing in the sentence.

Adverbs (Ibsa Xumuraa): Afaan Oromo adverbs are part of speech or words that modify any
part of language other than a noun. Adverbs can modify verbs, adjectives (including numbers),
clauses, sentences and other adverbs (Mylanguage.org, 2011). The four categories of adverbs in
Afaan Oromo: Adverb of time, adverbs of place, adverbs of manner and adverbs of frequency.
Lists of adverbs are given in table 2.6. The lists on left side are in English while those on the right
side equivalent meaning of the terms in Afaan Oromo.

29
Table 2.6 Adverbs in Afaan Oromo
Adverb of time Adverb of manner
English Afaan Oromo English Afaan Oromo
Yesterday Kaleessa Very Baayyee
Today Harr’a Quite Baayyee
Tomorrow Bor Really Dhugumaan
Now Amma Fast Dafee
Then Gaafas Well Gaarii
Later Ager Hard Cimaa
Tonight Edana Quickly Dafee
Right now Amma is amma Slowly Suuta
Last night Eda Carefully Qalbiidhan
This morning Ganama kana Absolutely Matuma
Next week Turban dhufu Together Walii wajjin
Recently Dhiyeenya kana Alone qophaa
Soon Dhiyootti
Immediately hatattamaan
Adverbs of place Adverbs of frequency
Here As Always Yeroo hunda
There Achi Sometimes Gaaffii gaaf
Over there Gara sana Occasionally Gaaffii gaaf
Everywhere Iddoo hunda Seldom Darbee darbee
Nowhere Eessayyu Rarely Darbee darbee
Home Mana Never yoomiyyuu
Away Fagoo
Out Ala

Prepositions: Prepositions in Afaan Oromo are links nouns, pronouns and phrases to other words
in a sentence. The word or phrase that the preposition introduces is called the object of the
preposition (Mylanguage.org, 2011). Here are list of some prepositions:

30
Table 2.7 Prepositions in Afaan Oromo.
English Oromo prepositions English Oromo prepositions
prepositions prepositions
about waa’ee since ergii
above gubbaa/gararraa than manna
across gama through gidduu
after booddee/booda till hamma
against faallaa to tti
among jara giddu towards garas
around naannoo under jala/gajjallaa
As akka unlike faallaa
At Itti until hamma
about waa’ee since ergii
above gubbaa/gararraa than manna
across gama through gidduu
after booddee/booda till hamma
against faallaa to tti
among jara giddu towards garas
around naannoo under jala/gajjallaa
As akka unlike faallaa
At Itti until hamma
before Dura up gubbaa
behind dudduuba/dugda via karaa
duuba
below jala/gajjallaa with wajjin
beneath gajjallaa within keessatti
beside Bira without malee
between gidduu two words jechoota lama
beyond garas according to akka kanaatti
But garuu because of kanaaf
English Oromo prepositions English Oromo prepositions
prepositions prepositions
By ..dhaan close to bira
despite ta’uyyuu due to kanaaf
Down Lafa except for kana malee
during Utuu far froom Irraa, siqee/iraa fagaatee
except malee inside of keessa isaa
For F instead of kanaa manna
Table 2.7, Prepositions in Afaan Oromo Continued …

31
 From Irraa near to itti aanee
In keessa next to itti aanee
inside keessa outside of kanaa alatti
into keessatti prior to kanaan dura
near Bira three words jechoota sadii
next ittaanee as far as hamma
Of Kan as well as fi
On Irra in addition to datalatees
opposite fuullee/ fallaa in front of fullee isaa
out Ala in spite of ha ta’uyyu malee
outside Alla on behalf maqaa …
over irraan on top of kana irraan
per ..tti demonstrative agarsiisoo
prepositions
plus Fi this kana/tana
round naannoo that sana
these warra kana/ jara kana

2.7.2. Morphology of Afaan Oromo:

Morphology is a way of studying the language word structure (Algeo, 2010). It is about the way
words are put together, their internal structure. It is a level at which the structure of language is
analyzed. It deals with the analysis and examination of meaningful units of forms that make up
sentences. The smallest meaningful units of forms are called "morpheme", a meaningful form that
cannot be divided into smaller meaningful parts (Algeo, 2010; Ibrahim, 2015). Morphemes are
either "free" or "bound" (Richard, 2003). A free morpheme can occur on its own whereas bound
morphemes do not occur alone. Bound morphemes are of three types, these are, prefix attached to
the initial positions, infix inserted in the middle, and suffix attached to the final position of the
word. All three types of bound morphemes are called "affixes".

Every language has its own morphological structure that defines rules used for combining the
different components the language may have (Wakshum, 2000). The Afaan Oromo for instance is

32
different in its morphological structure from French, English, and other languages that use Latin
characters.
Morphologically there are many word-formation processes in Afaan Oromo (Wakshum, 2000),
affixation and compounding are among these word-formation processes. Affixation is generally
described as the addition of affixes at the beginning, in between, and/or at the end of a root/stem
depending on whether the affix is a prefix, infix, or suffix (Wakshum, 2000). In agglutinative
languages like Afaan Oromo, most of the grammatical information is conveyed through affixes,
(that is, prefixes and suffixes) attached to the root or stem of words.
Afaan Oromo word is composed of two parts (Ibrahim, 2015): (1) the root (base morpheme), which
generally consists of basic sound and provides the basic lexical meaning of the word, and (2) the
pattern, which consists of prefixes and/or suffixes and gives grammatical meaning to the word.
Thus, the root /Bar/ combined with the pattern /-e/ gives Bare ‘learned,’ whereas the same root
combined with the pattern /-te/gives barte ‘she learns’.
There are more than ten major and very common plural markers in Afaan Oromo, including -oota,
-oolii, -wwan, -lee, -an, een, -eeyyii, -oo, etc.). As an example, the Afaan Oromo singular noun
mana (house) can take the following different plural forms: manoota (mana + oota), manneen
(mana + een), and manawwan (mana + wwan). The construction and usages of such alternative
affixes and attachments are governed by the morphological and syntactic rules of the language
(Kekeba et al., 2007).
Compounding is the joining together of two linguistic forms, which functions independently suffix
(Wakshum, 2000). Examples compound nouns include; abbaa-buddenaa ‘step father’ from abba-
‘father’ and buddena ‘food’.
The extensive inflectional and derivational features of Afaan Oromo are presenting various
challenges for text processing and information retrieval tasks in the language (Tesfaye, 2010). In
information retrieval, the abundance of different word forms and lexical variability may result in
a greater likelihood of mismatch between the forms of a keyword in a query and its variant forms
found in the document index databases. Stemming, a technique that is used to bring morphological
variants of a word into the root or stem word, plays an important role in this regard.

33
2.8. Related Works
2.8.1. Foreign Related Works
Xiaoyong, (2010) proposed two new models for cluster-based retrieval and evaluate them. The
study used document clustering model, is viewed as a mixture model of three sources: the
document, the cluster/topic the document belongs to, and the collection. A relevant document
assumed being generated by this mixture model. Both partitioning and hierarchical agglomerative
clustering algorithms have been studied in the context of IR. They used a three-pass K-means
algorithm as an example of partitioning methods in their static clustering experiments. Two sets
of experiments are performed in this study. The first set of experiments investigates whether a
simple language model of clusters can be used to rank clusters. And the second set of experiments
examines the effectiveness of cluster-based retrieval in the context of query likelihood retrieval
and the relevance model (RM), for both static clustering and query-specific clustering. Both
experiments have given promising result to design and develop clustering-based retrieval which
were more effective than the traditional, cluster free search engine.

bharathi and venkatesan, (2012) investigated on improving information retrieval using document
clusters and semantic synonym extraction. Initially, document clustering was investigated for
improving the precision or recall in information retrieval systems and as an efficient way of finding
the nearest neighbors of a document. They proposed as clustering is important in browsing a
collection of documents or in organizing the results returned by a search engine in response to a
user’s query. This paper presents a new semantic synonym based correlation indexing method in
which documents are clustered based on nearest neighbors from the document collection and then
further refined by semantically relating the query term with the retrieved documents by making
use of a thesaurus or ontology model to improve the performance of Information Retrieval System
(IRS) by increasing the number of relevant documents retrieved. Results show that the proposed
method achieves significant improvement than the existing methods and may generate the more
relevant document in the top rank.

Maruf and Yuji, (2015) were conducted research on document clustering: Before and After the
Singular Value Decomposition. In this paper, they discussed the clustering of documents by
calculating pair-wise similarity between documents using the original term-document matrix and
the decomposed term-document matrix. They also reported an evaluation method based on the

34
clustering hypothesis and analyze the clustering results. They necessarily address issues involving
representation of documents and computation of similarities between a set of documents.

Zhang et al., (2016) proposed clustering by using SVD on clusters to improve precision of inter-
document similarity. They discussed as LSI (Latent Semantic Indexing) based on SVD (Singular
Value Decomposition) was proposed to overcome the problems of polysemy and homonym in
traditional lexical matching. However, it is usually criticized as with low discriminative power for
representing documents although it has been validated as with good representative quality. In this
paper, SVD on clusters is proposed to improve the discriminative power of LSI. The contribution
of this paper is three manifolds. Firstly, they made a survey of existing linear algebra methods for
LSI, including both SVD based methods and non-SVD based methods. Secondly, they proposed
SVD on clusters for LSI and theoretically explain that dimension expansion of document vectors
and dimension projection using SVD are the two manipulations involved in SVD on clusters.
Moreover, they developed updating processes to fold in new documents and terms in a
decomposed matrix by SVD on clusters. Thirdly, two corpora, a Chinese corpus and an English
corpus, are used to evaluate the performances of the proposed methods. Experiments demonstrate
that, to some extent, SVD on clusters can improve the precision of inter-document similarity
measure in comparison with other SVD based LSI methods.

2.8.2. Local Research Works

There were several research works done in Afaan Oromo Information Retrieval to enhance the
retrieval performance of the system. These research works are discussed below based on their
chronological order.
Tewodros, (2003) was developed Amharic text retrieval using latent semantic indexing with
singular value decomposition. In this thesis, the potential of LSI approach in Amharic text retrieval
was investigated and 206 Amharic documents and 25 queries were used to test the approach.
Automatic indexing of the documents resulted in 9256 unique terms which are not in the stop-
word list were used for the research. Finally, the performance of this study was compared with the
standard vector space and (0.80) precision of the LSI approach was indicated above that of the
standard vector space.

Tesfaye, (2010) was conducted research on IR, designed, and developed a search engine for Afaan
Oromo the performance evaluation using a selected set of documents and queries. Experiment on

35
some specific features of the language against the design requirements is also made. In this case,
the average precision for the top 10 results is 76%. This indicates that displaying only the first few
top results, the top 10 in this case, entails better user satisfaction as compared to displaying all the
results. When a query is posed to the search engine, the search was taking place using the default
OR operator of Lucene's search on the terms in the query. This causes more number of documents
to be retrieved. In effect, the precision was being negatively affected. The future directions stated
in the paper are: developing standard relevance judgment metrics, testing efficiency and
effectiveness by comparing other algorithms, the work in the paper supposed to be dynamic
whereas it is not; so it needs further work to make it dynamic, and the work was used a limited
number of documents since it better if it used large documents.
Gezehagn, (2012) further intended to make possible retrieval of Afaan Oromo text documents by
applying techniques of the modern Information Retrieval system. Information Retrieval is a
mechanism that enables finding relevant information material of unstructured nature that satisfies
the information need of users from the large collection. Afaan Oromo text retrieval developed in
this study has indexing and searching parts. The Vector Space Model of Information Retrieval
system was used to guide searching for the relevant document from the Afaan Oromo text corpus.
The model is selected since the Vector space model is the widely used classic model of the
Information Retrieval system. The index file structure used is an inverted index file structure. In
this study, the experimental result shows that the performance is on the average (57.5%) precision
and (62.64%) recall. The challenging tasks in the study are the absence of standard corpus,
handling synonymy and polysemy, the inability of the stemmer algorithm to all word variants, and
ambiguity of words in the language. The future direction stated by the author includes: stemmed
index terms might improve performance of the system and recommended to be used in the further
work, and the algorithm LSI may record better performance if it is used. In this study, VSM is
used and the indexing mechanism is used in here inverted index.
Eyob, (2013) tried to design and develop a corpus-based Afaan Oromo–Amharic cross-lingual
Information Retrieval system that able to retrieve Amharic information using Afaan Oromo
queries. This approach selected to be followed in the study is a corpus-based, particularly parallel
corpus. For this study parallel documents including news articles, bible, legal documents, and
proclamations from customs authority were used. The system is tested with 50 queries and 50
randomly selected documents. Two experiments were conducted, the first one by allowing only

36
one possible translation to each Afaan Oromo query term and the second by allowing all possible
translations. The retrieval effectiveness of the system is measured using recall and precision for
both monolingual and bilingual runs. Accordingly, the first experiment returned a maximum
average precision of 0.81 and 0.45 for monolingual (Afaan Oromo queries) and bilingual
(translated Amharic queries) run. The result of the second experiment showed a better result of
recall and precision than the first experiment. The result obtained in the second experiment is a
maximum average precision of 0.60 for the bilingual run and the result for the monolingual run
remained the same. From these results, he was concluded that cross-lingual Information Retrieval
for two local languages namely Afaan Oromo and Amharic could be developed and the
performance of the retrieval system could be increased with the use of larger and clean corpora.
Mulualem, (2013) explored semantic indexing and document clustering for Amharic information
retrieval using semantic indexing and document clustering. In order to solve the issues, integrating
of semantic indexing and document clustering techniques with generic IR system were improved
retrieval performance. Here the system comprises all processes exist in generic IR plus to the C-
value technique multi word term extraction, K-means algorithm document clustering base
searching strategy used. The system was tested using tagged Amharic news documents size of
650Kb and it registered F-measure of 66% accuracy. Nevertheless, the performance of the system
is greatly affected by synonyms and polysemy, incorrect clustering, cluster representative
problems, Amharic knowledge base.

Tewordros, (2014) Amharic text retrieval developed using thesaurus based semantic compression
for improving the performance. For this study Amharic text document corpus were prepared by
the researcher encompassing different news articles, books and Amharic websites. Also various
techniques of text preprocessing including tokenization, normalization, stop word removal and
stemming were used to identify content-bearing words. Once content bearing terms are identified,
then semantic compression technique is applied to identify semantic representation that is more
precise descriptor for each term. Tewordros was used thesaurus based inverted indexing scheme
for semantic compression of terms in indexing. In the same way, in the searching side also
thesaurus-based query expansion technique was employed. Experimental result on the average
70% precision and 77% recall respectively. The major challenges of this study were, the IR system
include lack of standard thesaurus for Amharic language and ineffectiveness of Amharic stemmer
to conflate Amharic inflectional words into their stem. Therefore, in order to improve the

37
performance of the system there is a need to develop an effective Amharic thesaurus as well as
Amharic stemmer.

Daniel, Ramesh, and Dereje, (2015) attempted to develop the Afaan Oromo-English CLIR system
which enables Afaan Oromo native speakers to access and retrieve the vast online information
sources that are available in English by writing queries using their own (native) language.
Evaluation of the system was conducted by both monolingual and bilingual retrievals. The
performance of the system was measured by recall and precision. As they were mentioned in their
study larger documents were retrieved for the monolingual run (i.e. for the retrieval of documents
by using baseline queries of Afaan Oromo) than for the bilingual run (i.e. for the retrieval of
English documents using Afaan Oromo queries after being translated into English). The
performance of the system propose was highly affected by the size, reliability, and correctness of
the corpus used for the study. Maximum average precision of 0.468 and 0.316 for Afaan Oromo
and English were obtained respectively. In this study, Afaan Oromo- English CLIR that is based
on a corpus-based approach was developed for Afaan Oromo users to specify their information
need in their native language and to retrieve documents in English. The performance of CLIR
systems using a corpus-based approach is highly affected by the size, reliability, and correctness
of the corpus used for the study. However, the size of the documents used for this research was
limited in size and not quite reliable and clear which affected the level of performance to be
achieved. Moreover, the domains of parallel documents used to carry out the research were limited.
The paper indicated future directions, encouraging developing Afaan Oromo-English CLIR by
using this approach. However, the low performance achieved and also affect the accuracy of
English documents are retrieved.
Workineh and Debela, (2017) was developed sense disambiguation which finds the sense of Afaan
Oromo words based on surrounding contexts. The idea behind this approach was to overcome the
problem of scarcity of training data by using unsupervised approach that exploits sense in a corpus
which is not labelled. The context of a given word is captured using term co-occurrences within a
defined window size of words. The similar contexts of target words are computed using vector
space model and then clustered by identifying semantically related Afaan Oromo words. In order
to develop semantic model for Afaan Oromo they followed three steps process which involve
corpus preprocessing which tokenize and remove stop words and perform normalization. Extract
context terms providing clue about the senses of the target word, and then clustering to group

38
similar context terms of the given target words, the number of clusters representing the number of
senses encoded by the target word. The result argued that the system yields an accuracy of 81%
which was encouraging result. Furthermore, the study has been pointed out how NLP plays a
significant role in enhancing the computer’s capability to process texts. To the end, they were
pointed out as a semantic 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.

Wegari, (2017) conducted a rule-based root generation system for Afaan Oromo. In their
investigation; they have been shown that a rule-based method can be used to develop a root
generation system for Afaan Oromo. The system is mainly used as a starting point to develop a
complete morphological analysis and information retrieval for the target language. The
experimental results show that the methodology proposed is effective in identifying root
boundaries. The main challenge in this paper is that, since the majority of the proposed rules are
based on suffixes, the most errors of the system are due to strings that are found both as parts of
roots and suffixes in the language. For instance, for the word rukuta, the system indicated ruk as
its root instead of rukut. The future directions of this paper show that the rule-based method can
be used to develop a root generation system for Afaan Oromo. It might be used as a starting point
to develop a complete morphological analysis and information retrieval for the target language.
Melkamu, (2017) integrate query expansion for enhancing the effectiveness of the Afaan Oromo
text retrieval system. this study was applied the designed query expansion for the Afaan Oromo
information retrieval system involves lexical resource WordNet that was constructed as a reference
for identifying the sense and meaning of the user’s query using word sense disambiguation by
semantic similarity measure. The experimental result of this paper showed that the integration of
query expansion registers 95% recall, 41% precision, and 56% F-measure. The challenges
indicated were the absence of standard well-crafted WordNet, an effective stemmer algorithm, and
an organized corpus. The future direction of this paper indicates applying query expansion module
with Afaan Oromo IR to see the extent to which it improves the performance of the Afaan Oromo
retrieval system. However precision of the system needs more improvement because of the
polysemy nature of Afaan Oromo.

39
Table 2.8 local related work approach
Author Problem Approach Result Gap
(year)
Tesfaye, Searching and Boolean average Better stemming algorithm,
(2010) ranking search and precision for the extracting semantic future.
technique Afaan phrasal search top 10 results is only the first few top results,
Oromo approaches 76% the top 10 limited number of
documents document are retrieved
Gezehagn, accessing statistical average (57.5%) stemming algorithm is not
(2012) information that retrieval precision and working well
satisfies approach(VS (62.64%) recall
information M) handling synonym and
needs of Afaan polysemy
Oromo users
(Berhanu, LSI in Afaan LSI 0.67(67%) LSI is used statistically related
2019) Oromo text precision and term extraction, while it is
retrieval 0.63(63%) better to use language
recall knowledge base to extract it
LSI is more concern to recall
but n prizes, but clustering
approach with LSI is better to
come up with good IR
performance
Melkamu, enhancing the Query 95% recall, 41% absence of standard well-
(2017) performance of expansion precision, and crafted WordNet, effective
Afaan Oromo involves 56% F-measure stemmer algorithm and
text retrieval lexical organized corpus, extracting
system resource semantically related words
WordNet
Ashenafi Word sense word sense 66% F-measure Used limited number of lexical
(2019) disambiguation disambiguatio resource because of the lack of
based query n based query resource, since similarity
expansion so as expansion measure and the use of query
to improve the expansion terms are limited
performance of based on the information
Afaan Oromoo available in the lexical
text retrieval resource.
system

Berhanu, (2019) Explored the potential application of Latent Semantic indexing approach in Afaan
Oromo text retrieval. The goal of the study was to examine the benefits of LSI text retrieval
approach for Afaan Oromoo text retrieval. LSI attempts to use this as a basis for creating a concept
space such that this concept space is much smaller than the word space (mapping of terms

40
against documents). Singular value decomposition (SVD) of the term by document matrix was
used for indexing and retrieval system. The performance of the system after User Relevance
Feedback is measured using recall and precision, the experiment shows that the performance
of the prototype is on the average 0.67(67%) precision and 0.63(63%) recall registered.
Recently, Ashenafi, (2019) has attempts to extend the application of word sense disambiguation
based query expansion. In this study word sense disambiguation based query expansion approach
was used to determine the senses of words in queries by using Afaan Oromoo lexical resource.
WordNet which is constructed by translating Princeton University’s WordNet to Afaan Oromoo
WordNet. WordNet is consists of all parts of speech such as nouns, verbs, adjectives and adverbs
is the first component that is used as knowledge base in this study and Word Sense Disambiguation
is second, which is used to identify the sense of the given query using semantic similarity measure
from the knowledge base. Then Query reformulation by adding terms with highest overlap with
each term in original query with reference to Afaan Oromoo WordNet. Finally, the query
expansion module is integrated with Information Retrieval system to show the enhancement of
Afaan Oromoo IR system performance. The experimental result showed an effective use of WSD
using semantic similarity for identifying the sense with 66% F-measure performance. The
indicated are lack of lexical resource which affect the similarity measure, since query expansion
terms are limited based on lexical resource available.
Table 2.8 Local Related Research Works Approach
Generally, the above researchers are enabling the Afaan Oromo retrieval system using different
methods, but they are still not considered document clustering and latent semantic indexing
(synonym and polysemy). As pointed out by, Melkamu, (2017) and Gezehagn, (2012) the
existence of synonym and polysemy Afaan Oromo words greatly affect the retrieval performance.
Berhanu, (2019) was used LSI to extract Afaan Oromo semantically related words. However, when
semantically related terms are used it increases recall and less precision. This is suppress the results
to query during searching. Therefore developing enhanced Afaan Oromo information retrieval is
takes into account; Latent Semantic indexing and clustering based searching to improve the
performance of the Afaan Oromo information retrieval system.

41
 3. METHODS AND TECHNIQUES
3.1. Overview
This chapter discusses the design and development of Semantic Indexing with the Document
Clustering based searching for the Afaan Oromo information retrieval system. In the next section
the architecture of the proposed IR system for Afaan Oromo is designed and presented. Further,
methodology, research design, implementation tools, data preparation tasks such as tokenization,
normalization, stop words removal and stemming of Afaan Oromo documents and query, selecting
index terms, weighting TFIDF, Latent Semantic Indexing, K-means clustering, cluster-based
searching, and evaluation procedures are discussed in detail.
3.2. Methodology of the study
The methodology describes actions to be taken to investigate a research problem and the rationale
for the application of specific procedures or techniques used to identify, select, process, and
analyze information applied to understand the problem, thereby, allowing the reader to critically
evaluate a study’s overall validity and reliability (Hevner and Chatterjee, 2015). A methodology
is “a system of principles, practices, and procedures applied to a specific branch of knowledge”
(Robertson, 2012). Such a methodology helps researchers to produce and present high-quality
scientific research.

3.3. Research Design

Research design is considered as the structure of research or the arrangement of conditions for
collection and analysis of data in a manner that aims to combine relevance to the research purpose
with the procedure (Klinger and Lee, 2000). It is the plan, structure, strategy, and investigation
conceived to obtain and ensure a logical basis for the decision and the blueprint for the collection,
preprocessing, and experimenting (Tanner, 2018). This study follows experimental research
design; thus, it is used to test theories or construct the theoretical explanations that require careful
attention to all details from setting up the developing procedures for implementing the experiment.

To conduct an experimental research, the researcher followed a three step approach: document
corpus preparation, implementation of a prototype and evaluation. Tasks performed and methods
used during each step are stated as follows.

42
3.3.1. Document Corpus Preparation
The source of documents corpus for this study was Afaan Oromo news articles that were collected
from Oromia Broadcast Network (OBN), www.obnoromia.com. Since it is difficult to prepare all
relevant documents, a limited document corpus was selected to develop a prototype and evaluate
the performance of the prototype. For experimenting and measuring the effectiveness of the
proposed approach 243 Afaan Oromo news articles were randomly selected from this corpus, the
articles are taken from agriculture, politics, agri-business, business, culture, education, health,
religious, sport, and social aspects. All these articles are in txt format and reorganized based on
their contents. The documents were preprocessed and prepared in an appropriate format before
experimentation. For the preprocessing Afaan Oromo collected news articles, text operation tasks
such as tokenization, normalization, stemming, and stop word removal techniques were applied
for both indexing and searching. The analysis begins with a weighted term-document matrix and
this weighted matrix was used for indexing and retrieval.

3.3.2. Implementation Tools

Python programming language is used to develop the intended prototype. Python was developed
by Guido Van Rossum in the late eighties and early nineties at the National Research Institute for
Mathematics and Computer Science in the Netherlands (Halterman, 2011) and it is derived from
many other languages, including ABC, Modula-3, C, C++, Algol-68, Small-Talk, UNIX shell, and
other scripting languages. Higher-level programming languages like Python allow are
programmers to express solutions to programming problems in terms that are much closer to a
natural language (Halterman, 2011). The reason to use Python programming language is; on one
hand, the exposure of the researcher to the language and on the other hand, since python is used
for developing clustering, indexing, and searching in IRS. Python is also a dynamic programming
language that is used in a wide variety of scientific application domains (Python Software
Foundation, 2012). It is simple, strong, involves natural expression of procedural code, modular,
dynamic data types, and embeddable within applications as a scripting interface.

3.3.3. Evaluation procedure

For evaluation purposes; Precision, Recall, and F-measure are used to measure the effectiveness
of IR systems. Based on the concept of relevance (i.e. the correspondence of documents with a
given query or information need), there are several techniques for measuring the effectiveness of

43
IR, such as precision and recall, F-measure, E-measure, MAP (Mean average precision), R-
measure (Christopher, 2011). To do this, appropriate queries, were systematically and subjectively
selected after reviewing the content of each article manually to test the performance of the system
either relevant or irrelevant to make relevance evaluation.

To show these metrics, assume the document collection be D. Let Rt is all retrieved documents
from the collection D and Rl a number of relevant documents in D. The joint of Rt and Rl is a set
of documents retrieved and relevant. Then, the recall and precision can be calculated using the
following equation. The recall is the percentage of relevant documents retrieved from the corpus
in response to users query and precision is the percentage of retrieved documents that are relevant
to the query (Christopher et al., 2009), thus;
|{RlՈRt}|
Precision(P) = (3.1)
|{Rt}|

|{RlՈRt}|
Recall(R) = (3.2)
|{Rl}|

A measure that combines precision and recall is F-measure or balanced F-score (Christopher et
al., 2009). F-measure is the harmonic mean of precision and recall, which is calculated as follows.

2(RP)
F − measure = (3.3)
R+P

Where, R and P are recall and precision respectively.

3.4. The Architecture of AOIRs Using Semantic Indexing and Document Clustering Based
Searching.
The architecture shows the different modules integrated with generic information retrieval model
to design Semantic Indexing and Document Clustering based Searching for Afaan Oromo
information retrieval. Here below figure 3.1 depicted the system architecture of the Afaan Oromo
information retrieval process implemented in this study.

44
Fig 3.1 Architecture for AOIRs

This Architecture was adopted from generic information retrieval system architecture (Ceri et al.,
2013) and modified by integrating SVD with K-means clustering methods. Below Tasks
performed and methods used during each step in the architecture are stated in detail.

3.4.1. Dataset preprocessing

In this phase the different data preprocessing and preparation tasks are performed. These tasks
include tokenization, normalization, stop word removal, and stemming of Afaan Oromo
documents and queries.
Tokenization: In a given character sequence and a defined document unit, tokenization is the task
of chopping it up into tokens, perhaps at the same time throwing away certain characters, such as
punctuation marks (Singhal, 2008). A token is an instance of a sequence of characters in some
particular documents that are grouped together as a useful unit for processing the document. In
this study tokenization algorithm, 3.1 is used for splitting documents into a sequence of words
based on word separators.

45
Algorithm 3.1: Tokenization algorithm (Gezehagn, 2012)

FUNCTION Tokenization ( )
//INPUT: Afaan Oromo documents
//OUTPUT: sequence of words
FOR file in corpus DO
DEFINE whitespace OR punctuation as word delimiter
READ each file in corpus
FOR file in read DO
IF there is whitespace OR punctuation THEN
Put each term as separate token
END IF
END FOR
END FOR
RETURN tokens
END Function

For tokenization purpose, whitespace and punctuation were used as a delimiter. Punctuations
marks in Afaan Oromo are almost similar to that of English except for apostrophe (“ ’ ”) which is
considered as part of the letter in Afaan Oromo. In Afaan Oromo apostrophe is used as a part of
latter and punctuation mark. Some of the punctuation marks are; ? ፡ " ! | / ? @ # * ~ $ % ^ & ( ) {
} < > [ ] _ + = - , ." ...\; - _ + and removal of punctuation marks helps to consider similar words
with different punctuations mark, in a similar way with no distinction of punctuation mark linked
to it. For example ‘mootummaa?’, ‘mootummaa!’, ‘mootummaa.’, ‘mootummaa, and etc. all are
represented as ‘mootummaa.
Normalization: after processing files into separated tokens, it should come to uniform writing
format, so that matches occur despite superficial, hyphen, abbreviation, and equivalent character
differences in the character sequences of the tokens (Mulualem, 2013). Primarily every term in the
document should be converted into lowercase. For instance ‘MOOTUMMAA’, ‘Mootummaa’,
‘mootummaa’ are all normalized to lowercase ‘mootummaa’ (government). Numbers are also
removed with punctuation marks.

46
Algorithm 3.2: Normalization algorithm (Gezehagn, 2012)

Function normalization ()
//INPUT: sequence of words/tokens
//OUTPUT: normalized sequence of words/tokens
READ each tokens in file
FOR word w in tokens DO
IF word w is not lowercase THEN
Convert to lowercase
IF word w in tokens have any punctuation marks THEN
Remove punctuation marks
IF word w in tokens is not digit THEN
RETURN tokens
END IF
END IF
END IF
END FOR
END Function

Removal of Stop Word: Some extremely common words which would appear to be of little value
or relevance in discrimination documents are removed from the token set (Zipf, 1932). These
words are called stop words. The general strategy for determining a stop list is to sort the terms by
collection frequency (the total number of times each term appears in the document collection), and
then to take the most frequent term to remove. In addition, the other method is using a stop word
list for the language itself (Singhal, 2008). Some of the design of modern Information Retrieval
Systems has focused precisely on how we can exploit the statistics of language so as to be able to
cope with common words in better ways (Tewdros, 2003). In Afaan Oromo; conjunctions (fi,
yokaan, moo…), pronouns (ani, naaf, narraa, koo, isaan…), prepositions (waa’ee, ergii, hamma,
garas, faallaa…), verbs (ture, jira, jedhamaa), adverbs (kaleessa, amma, dafee, cimaa, baayyee…)
and adjectives (bareedduu, dheeraa, midhagduu …) are treated as stop words and still extracted
manually from other words. In this study, the researcher was taken stopwords organized by
(Gezehagn, 2012) and also added missing stop-words from lists of documents.

47
Algorithm 3.3: Stop word removal algorithm (Gezehagn, 2012)

FUNCTION Stop word-removal ()

//INPUT: list of tokens, Stop words list
//OUTPUT: List of non-stop words
OPEN Stop words list file
READ tokens list
FOR each word w in tokens list
IF a word w is in stop words list THEN
REMOVE a word w
END IF
END FOR
RETURN non-stop word tokens
END Function

Stemming: For grammatical reasons, documents are written using different forms of a word,
though such morphological variant words have similar semantic interpretations (Nega and Peter,
2004). Morphological variation of words is also there in Afaan Oromo, for instance, ‘mootummaa’,
‘mootummaan’, ‘mootummotta’, ‘mootummollen’. As a result, such words have to be reduced to
their root, ‘mootummaa’ using stemming. In addition to that stemming is also used to reduce the
dictionary size (i.e. the number of distinct terms used in representing a set of documents). The
smaller the dictionary size results in the smaller storage space and reduces the processing time
required. Generally, the goal of stemming is to reduce inflectional forms and derivationally related
forms of a word to a common base form, while it is language-dependent. Therefore, every language
is using its own specific stemming technique. In the Afaan Oromo text, there are many word
variants/affixes. To conflate them into stem words HornMorpho 2.5 was used for this work.
HornMorpho was developed by a research group for human language technology and the
democratization of information (Gasser, 2012), it is part of the L3 project at Indiana University,
which USA dedicated to developing computational tools for under-resourced languages such as
Amharic, Afaan Oromo, and Tigrinya words into their constituent morphemes (meaningful parts)
and generate words, given root or stem and a representation of the word’s grammatical structure
and lists of tokens are returned by using the following algorithm 3.4.

48
Algorithm 3.4: stemming algorithm ( Nega and Willett, 2002)

FUNCTION stemmer ()
//INPUT: non-stop word tokens
//OUTPUT: stemmed words
IF word in a tokens matches with one of the rules
REMOVE the suffix and do the necessary adjustments
RETURN the word and RECORD it in stem dictionary
ELSE
IF there is no applicable condition and action exist
RETURN word as it is and RECORD it in stem
dictionary
END IF
END IF
END Function

3.4.2. Indexing
The function of any Information Retrieval system is to process a user request for information
needed and retrieve documents that could potentially satisfy the information need of the user
(Salton and Mcgill, 1986). Indexing is the process of analyzing text and deriving such short-form
descriptions for a document that together sum up the message of the document (Baeza et al., 1982).
Inverted index, or sometimes inverted file, has become the standard term in information retrieval
for organizing documents (Christopher, 2011).

An inverted index always maps back from terms to the parts document where they occur and the
basic idea of an inverted index is a dictionary of terms (sometimes also referred to as a vocabulary
or lexicon). Each item in the list appeared in a document is conventionally called a posting
(Christopher, 2011), the list is then called a posting list. Among the different representations of
index terms, an inverted file index term representation technique is used in this work. An inverted
file consists of a list of tokens where each token is followed by the identifier of every document
that contains the word along with their number of occurrences in the document is represented.
Using this information inverted file allows an information retrieval system to quickly determine
which documents contain a given set of words, and how often each word appears in the document
49
term (Baeza-Yates and Ribeiro-Neto, 1999). Inverted indexes associate each word in the text with
a list of pointers to the positions where the word appears in the document (Kumar et al., 2012).
For each term, the list of all documents which contained that term and additional information about
that term like frequency of the word is stored. For instance, suppose we have the following Afaan
Oromo documents to construct an inverted index.

TFij, number of occurrences of term tj in document di

DFj, number of documents containing tj
CFj,, collection frequency of tj in nj then index based on sorted list of terms, with each term having
links to the documents is described below in figure 3.2

Doc1: Teessoon galma abbaa gadaa.

Doc2. guyyaan itti heerri mootummaa itoophiyaa itti ragga'e sadaasni 29 waggaa
waggaan kabajamaa.
Doc3. sirni gadaa dimookiraasii ammayyaaf bu’uura kan ta’e dha.
Doc4. bulchiinsi sirna dimokraatawaa gama siyaasaatiin faayidaa guddaa qaba

50
Fig. 3.2. Inverted indexing for Afaan Oromo document

The word ‘abbaa’ is in document 1 and its frequency is only once, the word ‘gadaa’ is in documents
1 and 3, and its term frequency once in both documents. Also, the word ‘waggaa’ is in document
2 and its frequency is twice. Just like these, all words in the collections contain the positions of
each word within document or documents and its frequency.
3.4.3. TFIDF weighting
To compute index term weight, from the collection of documents, a term-document matrix was
created and TFIDF was calculated for each index. The TFIDF weighting mechanism is based on
term frequency, the number of occurrence of a term in a document, and the inverted document

51
frequency of the term. The final weight of a term which was made of the term frequency and the
inverse document frequency was calculated as follows.
Let freqi, j be the raw frequency of term ki in document dj (i.e the number of times the term ki is
mentioned in text of document dj). Then, the normalized frequency fi,j of term ki in document dj is
given by (Baeza et al., 1982);
𝑓𝑟𝑒𝑞𝑖𝑗
fi,j = 𝑚𝑎𝑥𝑓𝑟𝑒𝑞𝑖𝑗 (3.4)

where the maximum was computed over all terms which are mentioned in the text of the document
dj. Further, let df (document frequency) represent the number of documents in which the term ki
appears. This could assign a higher value to terms that appear in many documents and lower value
for terms that appear only in few documents. However, terms that appear in many different
documents are less indicative of the overall topic. Then, we can use idfi, (inversed document
frequency) to calculate the weight of a term and use the logarithm to dampen the effect of the idf
as follows (Baeza et al., 1982);
𝑁
idf = log(𝑑𝑓) (3.5)

Hence the idf of a rare term is high, whereas the idf of a frequent term is likely to be low. Now
combine the above definitions of term frequency and inverse document frequency equation, to
produce a composite weight for each term in each document, the term weighting equation (Baeza
et al., 1982);
Wi,j = fij * idf (3.6)
These term weights are used to compute the degree of similarity between each document stored in
the system and the user query. Moreover, the consideration of term weights can assist in ranking
documents by sorting retrieved documents in decreasing order of their degree of similarity.
However, in this study, the researcher used TFIDF for further process to apply LSI and clustering,
since the objective of this study is to design Enhanced Afaan Oromo Information Retrieval using
Semantic Indexing and Document based Clustering searching to improve performance.
3.4.4. Signle Value Decomposition (SVD)
Latent Semantic Indexing (LSI) is a method for discovering hidden concepts in document data
(Dumais, 1992). Each document and term (word) is then expressed as a vector with elements
corresponding to these concepts. Each element in a vector gives the degree of participation of the
document or term in the corresponding concept. The goal is not to describe the concepts verbally,

52
but to be able to represent the documents and terms in a unified way for exposing document-
document, document-term, and term-term similarities or semantic relationship which are otherwise
hidden. Latent Semantic Indexing, an efficient vector space retrieval approach, uses the Singular
Value Decomposition (SVD) technique to reduce the rank of the original term-document matrix.
Theoretically, SVD, a dimensionality reduction technique, performs a term-to-concept mapping,
and therefore, conceptual indexing and retrieval are made possible. In this study, the researcher
used SVD for the approximation of a term-document matrix by one of low rank using singular
value decomposition. The low-rank approximation yields a new representation for each document
in a collection. Here the original term-document matrix (TFIDF), X was therefore approximated
as Y (see fig. 3.3 below). From the collection of documents, a term-document matrix is computed
in which each entry consists of a weight corresponding to a specific term in a specific document.
Fig. 3.3: Singular Values Decomposition (Hasan and Matsumoto, 2012)

The SVD, X = To So DoT results, where t x m matrix, To, the orthonormal columns of which are
called left singular vectors, an m x m diagonal matrix, So of positive "singular values" sorted in
decreasing order, and an m x d matrix, DoT, the orthonormal columns of which are called right
singular vectors. The value m is the rank of the matrix, X. With the To, So, and Do matrices, X can
be reconstructed precisely.
The key advantage of LSI in this study is to retain only the r (reduced-rank) largest singular values
in the So matrix and set the others to zero. The value of r is a design parameter and small values
are typically chosen. The resulting singular vector and singular value matrices were used to map
term-based vectors for documents into a subspace in which semantic relationships from the term-
document matrix are preserved and the original matrix, X is approximated by:
Y = T S DT (3.7)

53
where T is a t x r matrix with orthonormal columns, S is a positive definite r x r diagonal matrix,
and DT is a d x r matrix with orthonormal columns.
Eigenvectors and eigenvalues are numbers and vectors associated to square matrices, and together
they provide the eigen-decomposition of a matrix that analyzes the structure of this SVD matrix.
Suppose that there exists a number λ ∈ R and a non-zero vector v ∈ Rn such Av = λv. Then we say
that λ is an eigenvalue of the matrix A, and that v is an eigenvector corresponding to the eigenvalue
λ. Then Av = λv = λIv, where I is the n x n identity matrix, so that (A – λIv = 0). Since v ∈ Rn is
non-zero, it follows that we must have det(A-λI)=0.
However, one of the limitations of LSI is determining the number of dimensions to be removed
(Hansaem and Kyunglag, 2015). There is no exact method to find the right dimensions. Deerwester
et al., (1989) suggest that this choice captures the underlying semantic structure (i.e., the concepts)
in the term-document matrix while rejecting the “noise” that results from term usage variations. In
other words, the elimination of the small singular values reduces the documents feature space into
a “document concept space”. Here in this study organized and decomposed term-document matrix
after singular value decomposition and cluster generation procedure operates on reduced vectors
or points of a t-dimensional space and the optimum dimension (10 for this research) for the
collection was chosen by the criteria of ‘best retrieval performance’.
The following pseudo-code describes how the Latent Semantic Indexing algorithm works (Phadnis
and Gadge, 2014);
1. Convert each document from test collection to a list of words (terms) and then into a vector
of words.
2. Scale each vector of terms so that every term reflects the frequency of its occurrence in the
document.
3. Combine these column vectors into a large term-document matrix. Rows represent terms
and columns represent documents. The cells in the matrix are filled with TFIDF weights.
4. Perform SVD on the term-document matrix. This will result in three matrices commonly
called To, So and DoT
5. Set all singular values to 0 except the r highest singular values.
6. Recombine the terms T, an eigenvector of matrix S and document (DT), to form reduced
matrix, Y =TSDT.

54
3.4.5. K-means Clustering
Document clustering has been investigated in different areas of text mining and information
retrieval. Initially, document clustering was investigated for improving the precision and/or recall
in information retrieval systems by automatically grouping documents that belong to the same
topic in order to provide user’s browsing of retrieval results (Adrian Kuhn et al., 2006). It has
always been used as a tool to improve the performance of retrieval and navigation of large data
(Chartier, 2013). More recently, clustering has been proposed for use in browsing a collection of
documents or in organizing the results returned by a search engine in response to a user’s query
(Ramampiaro and Paulsen, 2009).

From partition clustering techniques, K-means is widely used in document clustering. K-means is
based on the idea that a center point can represent a cluster. A proximity distance measure is needed
to assign data objects to the closest centroid. In the K-means algorithm, many different distance
measures are possible, the most widely used one is cosine similarity. This is defined as;
t
∑ 𝑑1 ∗ 𝑑2
d1.d2 j=1
sim(d1, d2 ) = || d1|| ||d2|| = 2 2
(3.8)
t t
√∑ (d1 ) ∗∑ (𝑑1)
j=1 j=1

d1 and 2 are two representative points contain n number of values respectively.

Since LSI increases recall and hurt precision, the researcher was applied clustering to improve the
precision of the information retrieval system. To do this, a reduced matrix Y= TSDT, which was
computed based on dimension reduction using r singular value was used. A cluster integration
with reduced SVD matrix procedure operates on vectors or points of a t-dimensional space, where
t is the number of terms (or concepts) with document indexed to applying K-means clustering. The
standard K-means algorithm works as, given a set of document D and a pre-specified number
clusters k, k data objects are randomly selected to initialize k clusters, each one being the centroid
of clusters (Adrian Kuhn et al., 2006). The remaining objects are then assigned to the cluster
represented by the nearest or most similar centroid. Next, new centroids are recomputed for each
cluster and run all documents are re-assigned based on the new centroids.

55
Algorithm: 3.5 K-means clustering algorithm (Ramampiaro and Paulsen , 2009).

FUNCTION kmeans ( )
// INPUT: The initial number of k clusters; reduced document collection matrix,
//OUTPUT: k clusters of documents
Repeat
For dj = 1 to n do //for each document in the collection
min sim := sim(dj,c0)
t := 0
For Ci = 1 to k do //for each cluster Ci
sim (dj, ci); //similarity of document dj and centroid ci
If sim (dj, ci) < minsim then
minsim := sim(dj,ci)
t := i
End If
End For
Ct = dj //Assign dj to the cluster Ct with the closest centroid
End For
For i = 1 to k do
Recalculate the positions of the centroid of Ci
End For
Until the global centroids is no longer move
Return k clusters of documents
End function

To determine the optimal number of clusters, the researcher has selected the value of k at the
“elbow”, i.e. the point after which the distortion/inertia start decreasing linearly. Thus for the given
data, to conclude that the optimal number of clusters the Sum Square Error elbow method was
used. In the elbow method, the variance (within-cluster sum of squared errors) is plotted against
the number of clusters (Wulandari, 2019). The first few clusters will introduce a lot of variance
and information, but at some point, the information gain will become low, thus imparting an
angular structure to the graph. The graph is available in Appendix I. The optimal number of clusters
was found out from this point; therefore, this is known as the “elbow criterion.” But this point
cannot always be determined without any sense of ambiguity and the drawback of the elbow
method is this ambiguity. The elbow method is easy to implement by looking at the ideal k value

56
graph with the position on the elbow along with the SSE (Sum of Square Error) (Syakur et al.,
2018).

Performance indicators use a number of squared errors (SSE) and clusters are said to be convergent
when obtaining a smaller value compared to others (Yuan and Yang, 2019). The Elbow algorithm
method in K-Means as shown below (El-Mandouh et al., 2019).

Algorithm 3.6 elbow algorithm (Syakur, et al., 2018)

FUNCTION ELBOW ()
//INPUT: range of cluster, document matrix
//OUTPUT: optimal k
Initialize k=1
REPEAT
Increment the value of k
Measure the cost of the optimal quality solution (using sum of
squares method)
IF at the same cost the solution drops dramatically then
RETURN k //that’s the true k
UNTIL the solution drops
End function

3.4.6. Searching
In Information Retrieval (IR) task, searching involves issues of comparing queries with a
collection of documents to locate a set of documents relevant to a particular query. In most cases,
a query is treated as a small document consisting of a few terms, and the similarities between the
query vectors and the document vectors are compared to find relevant documents. For searching
purpose similarity measure between the query vector and documents vector are computed, then
documents can be ranked (retrieved) in order of decreasing similarity to a query by computing
normalized inner products (cosine similarity) on the term-based vectors (Baeza-Yates and Ribeiro-
Neto, 1999);
t
∑ Wqj ∗ W
dj.q j=1 dij
sim(dj, q ) = || dj|| ||q|| = (3.9)
t 2 t 2

√∑ (W ) ∗∑ (W )
qj dij
j=1 j=1

57
Where, d is document, q is query from user, wqij is the weight of the jth term of the ith query and
wdij is the weight of the jth term of the ith documents.

Cosine similarity is a measure of similarity between two non-zero vectors of an inner product space
that measures the cosine of the angle between them (Singhal, 2008). Cosine similarity then gives
a useful measure of how similar two documents are likely to be in terms of their subject matter. In
the case of information retrieval, the cosine similarity of two documents will range from 0 to 1,
since the term frequencies (tfidf weights) cannot be negative. Vector space model (TFIDF, cosine
similarity) that help to index and rank the document according to query matching and as a result
gives the documents which most similar to the keyword.

3.4.6.1. Cluster based searching

Cluster-based retrieval uses the cluster hypothesis to retrieve a set of documents relevant to a query
(Liu and Croft, 2004). In the Cluster hypothesis documents in the same cluster behave similarly
concerning relevance to information needs (Voorhees, 2005). If a document in a cluster is relevant
to a query, then the rest of the documents in that cluster are potentially relevant to the same query.
According to Voorhees (2005), there are two approaches in cluster-based retrieval. The first
approach retrieves one or more clusters relevant to a query instead of retrieving documents relevant
to a query. In other words, this approach retrieves and ranks the relevant clusters instead of the
relevant documents. The second approach uses the generated clusters as a reference for improving
the retrieved relevant documents. In this approach, the given document collection is clustered
(static clustering) beforehand. When a set of documents is retrieved for a query, the generated
clusters (static clusters) of the collection are used as a reference to update the retrieved relevant
document list.
On the other hand Jardine and Rijsbergen (1979) suggested that document clustering could be used
to improve the effectiveness as well as the efficiency of retrieval, if the retrieval system were able
to find good clusters; retrieval performance can be improved over document-based retrieval.
In this study document clustering has been performed, on the reduced low rank from SVD of LSI,
independent of the user’s query and the task for the retrieval system is to match the query against
clusters centroid of documents instead of individual documents, and rank documents based on the
similarity of clusters centroid to the query. After the first ranked cluster is known the documents inside
the first ranked cluster are ranked based similarity to query and relevant to users as relevant documents.
These show that the clusters rather than first-ranked clusters are not relevant to the user query.
58
 4. RESULTS AND DISCUSSION
4.1. Overview
In this section, the source of document collections, the prototype of documents preprocessing,
latent semantic indexing, clustering-based searching, relevance judgment, an experimental result
(using VSM, SVD, SVD with K-means), performance comparison of proposed work and
challenges were discussed in detail.
4.2. Document Corpus
To test the proposed system 243 Afaan Oromo news articles collected from OBN were used. The
articles are taken from agriculture, politics, business, culture, education, health, religious, agri-
business sport, and social aspects (see table 3.1 below). All articles are in .txt format and before
applying experimentation data was prepared in an appropriate format.

Table 4.1. List of documents collections

Document Frequency Percentage (%)

Agriculture 29 12.0
Agri-business 16 6.6
Business 8 3.2
Culture 15 6.1
Education 27 11.1
Health 9 3.7
Religious 18 7.4
Sport 12 4.9
Social 28 11.6
Politics 81 33.4
Total 243 100

4.3. Documents Preprocessing

The prototype system development begins with preprocessing documents and then, constructing
an inverted file consisting of a vocabulary file and posting file, for indexing and organizing
documents to speed up searching. Preprocessing tasks include tokenization, normalization, stop
word removal, and stemming, then TFIDF term weighting has been employed. The system process
the index terms and queries by using python 3.5. For this purpose, code 4.1 were used to
preprocessing.

59
Code 4.1. A Program for Tokenization, Normalization and Stop Word Removal

There were challenges encountered by the researcher during tokenization, normalization, and stop
word removal. One of the challenges related to tokenization is differentiating single word from
compound word, for instance, AddisAbaba versus Addis Ababa, DirreeDawaa versus Dirree
Dawaa, AbbaaGadaa versus Abbaa Gadaa, waljaalachuun versus wal-jaalachuun. The other
challenge is identifying numerical values like dates, phone numbers, and IP addresses, and
apostrophes that are used in Afaan Oromo as a letter in the same words (i.e., ta'an si'a
mo'achuudhaaf, waa’ee); however, it is removed during tokenization. On the other hand, absence
of a standard stop-word list was a problem faced during stop word removal. This was forced the
researcher to use stop word used by previous research and added missing stop words in
consultation with language experts.
Stemming is language-dependent that reduces tokens to their root form of words to recognize
morphological variations. HornMorpho is stemmer based on context. Therefore the researcher was
used this stemmer for conflating words in Afaan Oromo documents, as shown below in code, 4.2.
Code 4.2. Stemming using HornMorpho

60
In the HornMorpho system for morphological processing linguistic consideration or irregular
morphological features of the language, query expansion, and the problem of the compound word
are not fully considered. There is limited number of available roots and stems, those when the root
is not known the verb guesser analyzer produces different analyses when in many cases only one
of these contains a root that actually exists in the language. However, the guesser analyzer itself is
a useful tool for extending the lexicon; when an unfamiliar root is found in multiple word-forms
and in multiple morphological environments, it can be safely added to the root lexicon. There is
also the problem of handling ambiguity. Even when a root or stem is known, there are often
multiple analyses, and the program provides no information about which analyses are more likely
than others.

4.4. Single Value Decomposition with K-means Clustering

Term weighting, Terms that were in the documents or in the query needed to get hold of weight
with respect to their documents. After inverted index file was created, the weight can be calculated
using TFIDF term weighting methods and SVD was applied on this matrix. The term weighting
was done by finding the frequency of terms and its inverse document frequency using code, 4.3
shown below.

Code 4.3. TF, IDF and TFIDF using numpy built in algorithm.

Singular Value Decomposition (SVD) was used to estimate the structure in word usage across
documents, and statistically derived vectors are more robust indicators of meaning than individual
terms. The weighted term-document matrix (TFIDF) is given to the SVD algorithm as an input
and a reduced dimensional representation of the matrix is generated. The SVD of this term-
document matrix was then computed and small singular values are eliminated from the singular
value matrix. The resulting singular vector and singular value matrices are used to map term-based
vectors for documents into a subspace in which semantic relationships from the term-document
matrix are preserved. The SVD of the matrix was computed for a range of dimensions and the
optimum dimension (10 for this research) for the collection was chosen by the criteria of ‘best
retrieval performance.’

61
Code 4.4 Single Value decomposition

K-means clustering algorithm was used for document clustering. To determine the number of
clusters, the researcher used the value of optimal-k found at elbow point after applied the method
in K-means. Based on this elbow method the number of clusters for documents used in this study
was 8 as indicated in Appendix I. The cluster was performed on reduced single value
decomposition (after LSI). Thereby documents which have higher cosine similarity score with
their centroid are under the same cluster. To perform this, K-means clustering algorithm from
sklearn library was used as shown in code below.

Code 4.5 K-means clustering using python sklearn library

Based on cosine similarity between query and clusters centroid relevant documents are retrieved.
However, there is a major problem encountered during experimentation. Thus, Re-computing the
SVD with larger data term-document matrix requires more computation time and, it may be
62
impossible due to memory constraints as well as massive datasets are inefficient in K-means
clustering. On the other hand, even if the researcher tried to use the elbow method to determine
the optimal number of cluster, there also challenges to define exact optimum number of clusters.
Here elbow is also needs to determine the range of clusters.

4.5. Clustering Based Searching

Searching compare query with a collection of documents representative to locate a set of relevant
documents that satisfies information need of user. The point is if a cluster was relevant to a query,
then the documents in that cluster were potentially relevant to the same query. Based on the cluster
hypothesis, documents from the top ranked cluster are relevant to a query than the documents from
the clusters with a least rank. For searching purpose similarity measure between the query vector
and clusters centroid vector was computed, then documents were ranked in order of decreasing
similarity to a query by computing normalized inner products (cosine similarity). After the
similarity between the centroids vector and the query vector is performed, then clusters are ranked
based on similarity to the query. Finally, documents in the first ranked cluster are relevant to users
query and returned in response to users.

Code 4.6 Clustering based searching

63
4.6. Performance Measure
In this research, the researcher attempting to design and implement Semantic Indexing with and
Document Clustering based searching for the Afaan Oromo information retrieval. As discussed in
previous chapters, the system aims to overcome the semantic problems exhibited in indexing and
similarity measurement. In this experiment 10 queries are used for evaluating the performance of
the system and experimented at three-level using; VSM, SVD, and SVD with K-means clustering.
Experimentation was done for identifying which documents were relevant for a given test query
and a result of each query was presented in the form of precision, recall, and F-measure.

4.6.1. Query Selection and Relevance Judgment

Ten (10) appropriate queries, which able to describe the document are systematically and
subjectively selected after reviewing the content of each article manually to test the performance
of the system. These queries were used with different techniques to measure performance and
queries are marked across each document as either relevant or irrelevant to make relevance
evaluation. All queries and documents should be stored in a form that is used for the term-
document matrix. TFIDF is assigned as a weight of the term-document matrix. Subjective
relevance judgment of each document was identified as a relevant or non-relevant to each selected
query for the experiment as listed below in table 4.2.

Table 4.2. Proposed system relevance judgment

No Queries Relevant Non-

relevant
1 Barnoota 32 211
2 Nageenya 38 205
3 sirna gadaa 21 222
4 hidha haaromsaa 130 113
5 balballomii ispoortii 12 231
6 rakkinoota hawaasa kessaa jiran 28 215
7 qonna bultooni beenya lafaa fudhatanii 30 213
8 hojii mootumman misooma, bulchinsa gaarii diriirsuu 23 220
9 Bulchiinsa Magaalaa kessaatti Qaamoleen hawaasaa 35 208
gahee olaanaa qaba
10 labsiin yeroo ariifachisaa labsame nageenya biyyaa 30 213
eegsisuu keessatti gahee olaanaa qaba jedhan

64
4.6.2. Experimental Result Using VSM
The researcher conduct an experiment to measure the effectiveness of vector space model (VSM)
in retrieving relevant Afaan Oromo documents for the given 10 queries. Detailed experimentation
result using VSM is depicted in table 4.3 below.

Table 4.3: Effectiveness of VSM based Afaan Oromo IR

No Queries R P F
1 Barnoota 0.40 0.30 0.30
2 Nageenya 0.40 0.20 0.26
3 sirna gadaa 0.43 0.30 0.20
4 hidha haaromsaa 0.7 0.32 0.20
5 balballomii ispoortii 0.20 0.13 0.15
6 rakkinoota hawaasa kessaa jiran 0.50 0.21 0.30
7 qonna bultooni beenya lafaa fudhatanii 0.23 0.24 0.20
8 hojii mootumman misooma, bulchinsa gaarii diriirsuu 0.70 0.40 0.20
9 Bulchiinsa Magaalaa kessaatti Qaamoleen hawaasaa gahee 0.80 0.20 0.30
olaanaa qaba
10 labsiin yeroo ariifachisaa labsame nageenya biyyaa 0.63 0.20 0.28
eegsisuu keessatti gahee olaanaa qaba jedhan
Average 0.50 0.25 0.24

As shown in table 4.3 the average result obtained using VSM is 50%, 25%, 24% recall, precision,
and F-measure respectively. The performance of the system was lowered for the reasons why VSM
only retrieve relevant documents based on exact matching between the index and query terms.
This means VSM has no mechanism to manage polysemy and synonymy. On the other side
problem of morphological inflection nature of the Afaan Oromo writing system for the purpose of
number, genders, possession, plural form, proposition, and conjunctions lowered VSM
performance. For this experiment, search results sample using the first query ‘barnoota’ (see
appendix III) was 46 documents.

4.6.3. Experimental Result Using SVD

The next experiment conducted in this study was SVD based latent semantic indexing.
Accordingly, this was approach evaluated using same 10 queries used for in VSM retrieval. Table
4.4 presents retrieval result using SVD.

65
Table 4.4: Retrieved result to query using SVD

No Queries R P F
1 Barnoota 0.75 0.20 0.30
2 Nageenya 0.95 0.20 0.30
3 sirna gadaa 0.98 0.18 0.20
4 hidha haaromsaa 0.86 0.50 0.64
5 balballomii ispoortii 0.58 0.40 0.38
6 rakkinoota hawaasa kessaa jiran 0.96 0.20 0.30
7 qonna bultooni beenya lafaa fudhatanii 0.8 0.20 0.25
8 hojii mootumman misooma, bulchinsa gaarii diriirsuu 0.95 0.20 0.30
9 Bulchiinsa Magaalaa kessaatti Qaamoleen hawaasaa 0.98 0.25 0.34
gahee olaanaa qaba
10 labsiin yeroo ariifachisaa labsame nageenya biyyaa 0.96 0.20 0.25
eegsisuu keessatti gahee olaanaa qaba jedhan
Average 0.88 0.25 0.34

As shown in Table 4.4 the proposed approach registers an average recall, precision, and F-measure
of 88%, 25%, and 34% respectively. In this method highest recall is registered as compared to
precision and F-measure. This result is registered because of the fact that the SVD method retrieves
as many documents as possible that are considered relevant for the given query. For example the
search result for sample query term ‘barnoota’ (see appendix III) was 158 documents. Obviously,
as the number of documents increases recall increases at the expense of precision. This means
some queries request an answer to specific issues and demand the returned documents to focus
exclusively on that particular issue and this advantage of the LSI method can sometimes result in
spurious answers. For instance; when the search for documents using ‘rakkinoota hawaasa kessaa
jiran’, and ‘balballomii ispoortii’ the system using SVD has retrieved 236 and 162 documents
respectively from 243 collections of documents used. This was provided the solution for the
conceptual matching problem, while it indicates the number of recall increases and affects the
precision.

4.6.4. Experimental Result Using Semantic Indexing and Document Clustering Based
Searching
To handle the weakness of VSM and SVD, the researcher integrate SVD and document clustering.
To do this document clustering has been applied, on the reduced low rank from SVD matrix of
LSI, independent of the user’s query and the task for the retrieval system is to match the query

66
against clusters centroid of documents instead of individual documents, and rank documents based
on the similarity of clusters centroid to the query. On the other hand, searching is being processed
by using a select query term. Finally, the cosine similarity measure identifies a relevant document
and the result obtained in the experiments indicates the IR system build based on semantic indexing
based document clustering registered better result than VSM and SVD (see appendix III). After
the first ranked cluster is known the documents inside the first ranked cluster are ranked based on
similarity to query and relevant to users as relevant documents. The performance of the proposed
approach using SVD based K-means clustering is presented in table 4.5.

Table 4.5: Performance of SVD based K-means clustering for Afaan Oromo IR

No Query R P F
1 Barnoota 0.72 0.88 0.79
2 Nageenya 0.60 0.80 0.70
3 sirna gadaa 0.70 0.90 0.80
4 hidha haaromsaa 0.80 0.60 0.70
5 balballomii ispoortii 0.83 0.70 0.74
6 rakkinoota hawaasa kessaa jiran 0.40 0.73 0.51
7 qonna bultooni beenya lafaa fudhatanii 0.6 0.90 0.72
8 hojii mootumman misooma, bulchinsa gaarii diriirsuu 0.70 0.80 0.74
9 Bulchiinsa Magaalaa kessaatti Qaamoleen hawaasaa gahee 0.50 0.70 0.60
olaanaa qaba
10 labsiin yeroo ariifachisaa labsame nageenya biyyaa eegsisuu 0.83 0.92 0.87
keessatti gahee olaanaa qaba jedhan
Average 0.70 0.80 0.72

Proposed prototype of Afaan Oromo information retrieval system was developed based on the
vector space of information retrieval model, while to improve system performance the researcher
was applied SVD and K-means clustering on the VSM matrix. The system has been recorded an
average 70% recall, 80% precision, and 72% F-measure. However, their major problem
encountered during experimentation were computation time for large collections, and it may be
impossible due to memory constraints as well as massive datasets are inefficient in K-means
clustering. The other problem of SVD was representing inherent semantics of the document, and
to define a similarity measure based on the semantic representation such that it assigns higher
numerical values to document pairs which have a higher semantic relationship.

67
4.6.5. Performance Comparison
After conducting a series of experiments using the selected 10 queries, the researcher selected the
best performing approach based on their F-Measure that enables us to see to what extent both recall
and precision of the retrieval result improved. Hence Table 4.6 below shows a comparison of the
three approaches (VSM, SVD, and SVD with K-means) for designing a prototype Afaan Oromo
information retrieval (AOIR).
Table 4.6: Summarized result of overall performance of Afaan Oromo IR system

IR effectiveness VSM based SVD based SVD with K-means based

measures AOIR AOIR AOIR
Recall 0.5 0.88 0.70
Precision 0.25 0.25 0.80
F-measure 0.24 0.34 0.72

The above table 4.6 showed that, while the proposed system score 72% F-measure, vector space
model (VSM) and single value decomposition (SVD) model registers of 24% and 34% F-measure
respectively. The proposed approach scores better performance since it attempts to measure
semantic relation of words, rather than matching words based on the shape similarity of words,
this further shows that the proposed approachable to control synonym and polysemy of words
thereby improving the precision and recall of the retrieval system. As a result, the researcher
proposed SVD with K-means clustering for designing the Afaan Oromo information retrieval
system.
4.7. Discussion of Result
Previously, different efforts have been put to fill or mitigate the gap between the individual
information needs and the provision of Afaan Oromo information retrieval. In addition to
comparison between three level results of VSM, SVD and SVD with K-means, the researcher also
tried compered this study with previous work done as follows. This research records better
performance with an improvement by 16% F-measure as compared to the work of Melkamu,
(2017), who integrate WordNet based query expansion for enhancing the effectiveness of Afaan
Oromo information retrieval. The approach followed by Melkamu may help to control synonym
words of Afaan Oromo, but not polysemy words. This study also registered better performance
than Berhanu, (2019) that explored the potential application of Latent Semantic indexing approach
68
in Afaan Oromo text retrieval by 7% recall and 13% precision respectively. Furthermore this
research scored better performance by 6% when compared to Ashenafi, (2019) work, attempts to
extend the application of word sense disambiguation based query expansion and registered 66%
of F-measure.

This research attempts further to integrate semantic indexing and cluster based searching to control
the effect of polysemy and synonym Afaan Oromo words, as a result of which a promising result
obtained for enhancing effectiveness of the IR system.

However, there were various challenges encountered during this study. The challenge started with
the complexity of the Afaan Oromo linguistic feature itself. Even if the researcher tried to keep
the semantics of the documents by extracting the statistical value of the term-document matrix
using SVD. Afaan Oromo linguistic computational level of the language was not matured and
getting the semantic meaning of the language was a bit difficult. If this problem is not addressed
well, it is difficult to get correct semantics of the Afaan Oromo documents, since it needs well
organized Afaan Oromo ontology to identify semantic relation (polysemy and/or synonymy) of
words correctly. The other challenge was that there is no standard and organized documents for
experimenting and measuring the effectiveness of the proposed approach in this study, here the
researcher used documents collected from OBN daily news and re-organized based on their
content, while it is not enough for better results. The third challenge was stemmer, there is no
suitable stemmer that is available for Afaan Oromo; however, though the researcher used
HornMorpho it does not still stem all words properly. There is also a problem related to the
clustering algorithm. When the documents were not clustered properly, the system was limited to
look for the relevant documents. Mis-clustering of documents occurred because a document might
have more shared terms with one cluster but it was talking about the other cluster. For instance, if
the document dealt with business and agriculture issues, agro-business has many terms of business
or agriculture whereas the document has to cluster in business may concern about agriculture and
visa verse.

69
 5. CONCLUSION AND RECOMMENDATION
5.1. Conclusion
The purpose of an Information Retrieval (IR) system is to process a collection of documents as per
requests for information, identify and retrieve from the corpus certain documents in response to
the information requests. Since main objective of this research study was to address the semantics
problem related to Afaan Oromo information retrieval and enhance the information retrieval
system through the latent semantic indexing and document clustering approach. The previous
research works in the Afaan Oromo information retrieval system were attempted to address
different problems such as word-sense disambiguation, query expansion, latent semantic indexing,
but not all are appropriately concerned with semantic relation and clustering to improving retrieval
system performance. Therefore in this study, an attempt has been to design and develop an
improved information retrieval system which is called enhanced Afaan Oromo Information
retrieval. As it has been discussed in the previous sections, the designed and prototyped system
mainly consisted of three components; latent semantic indexing, K-means clustering, and
searching.
The indexing part of the work involves pre-processing, such as tokenization, normalization,
stopwords removal, stemming, and weighting. The analysis begins with a weighted term-document
matrix and this weighted matrix was used for indexing and retrieval. The weighted matrix was
analyzed further by applying the SVD to derive the latent structure model which was used for
indexing by using the LSI model. Document clustering is a sub-component of the system which
aimed to cluster related documents into one category to facilitate searching. For documents
clustering purposes K-means clustering was employed and total document provided to the system
clustered in eight groups and these clusters centroids is used to retrieve documents based on a
similarity measure. Searching mainly consists of query pre-processing, similarity measure, and
retrieving the relevant document. Cosine similarity was measured between the query vector and
cluster centroid vectors that were used to retrieve relevant documents to each query.
To evaluate the performance of the system 243 document collections are randomly selected from
corpus collected from OBN with appropriate 10 queries. Based on experimentation the proposed
approach for developing Afaan Oromo information retrieval registered 70% recall, 80% precision,
and 72% F-measure. Finally, the experimentation indicated, the proposed system using Latent
Semantic Indexing and Document Clustering based searching was improved the performance of

70
the system in terms of precision, recall, F-measure, and reduce dimension when compared to VSM
and SVD model. This is a promising result to design an applicable information retrieval system.
5.2. Recommendation and Future Direction
The proposed Afaan Oromo retrieval system was attempted to enhance the performance of the
retrieval system by considering the semantic relation of words in documents. In this paper, the
researcher addressed Semantic document clustering of a document collection using inter-document
similarities derived from the term-document matrix. The need to investigate semantic document
clustering is that documents that are semantically related to each other are grouped into the same
cluster and documents that are semantically unrelated are grouped into another cluster.
Finally, based on the finding of the study, in order to obtain the optimum performance following
recommendations are recommended.
The main objective of information retrieval system is a well-organized corpus and pre-processing
based on the language knowledge. But there is no standard corpus, stemmer, tokenize, normalizer,
and stop word detector developed for Afaan Oromo and spelling checker also being important for
indexing and searching to write correct index terms or queries.
Advanced information retrieval system uses resources like a thesaurus, WordNet, machine-
readable dictionaries, and a better latent semantic indexing system which is based on polysemy
and synonyms of words rather than using statistically extracting semantic relation. Hence, the
researcher recommends future research to implement context-aware IR system using a well-
developed thesaurus or WordNet to aware of the context to which the terms in the text appear to
enhance precision and recall of the system.
The other recommendation is extending this study using another clustering algorithm such as
hierarchical and lexical chain clustering to handle the relationship between words used for
explaining different domains.
On the other hand, still, Afaan Oromo retrieval systems were developed for textual, but there is
also a need to further investigate the development of retrieval system for video, audio, and images.

71
 6. REFERENCES
Abara Nefa. 1988. Long Vowels in Afaan Oromo: A Generative Approach. MSc Thesis Addis
Ababa University, Ethiopia.

Abera Diriba. 2009. Automatic Classification of AFaan Oromo News Text:In Case of Radio Fana.
Msc Thesis Addis Ababa University, Ethiopia.

Abey Bruck. 2011. Semantic Based Query Expansion Technique for Amharic IR. MSc Thesis,
Addis Ababa, Ethiopia.

Adrian Kuhna, Stephane Ducasseb, Tudor Girba. 2006. Semantic Clustering: Identifying Topics
in Source Code. Journal of Information Systems and Technologie, France.

Amira M. El-Mandouh Laila A. Abd-Elmegid, Hamdi A. Mahmoud, Mohamed H. Haggag. 2019.

Optimized K-means clustering model based on gap statistic. International Journal of Adva
nced Computer Science Application, Canada , 10(1): 183-188.
Andreas Hotho, Steffen Staab, Gerd Stumme. 2003. Wordnet improves Text Document Clustering.
In proc. of the SIGIR, Semantic Web Workshop, Germany.

Artem Polyvyanyy, Dominik Kuropka. 2007. A Quantitative Evaluation of the Enhanced Topic-
Based Vector Space Model. Journal of Hasso planttner Institution, Potsdam,.

Ashenafi Tulu. 2019. Word Sense Disambiguation Based Query Expansion for Improving the
Performance of Afaan Oromoo Information Retrieval System, Msc. Thesis, Haramaya
University, Ethiopia

Aurélie Névéol, James G. Mork, Alan R. Aronson. 2008. Automatic Indexing of Specialized
Documents:Using Generic vs. Domain-Specific Document Representations. National
Library of Medicine,USA.

Bader Aljaber, Nicola Stokes, James Bailey Jian Pei. 2009. Document clustering of scientific texts
using citation contexts. Springer Science Business Media, LLC, Canada.

Baeza-Yates and Ribeiro-Neto. 1999. Modern Information Retrieval. Journal of the Association
for Computing Machinry. Cambridge University,, USA.

72
Balwinder Saini, Vikram Singh,Satish Kumar. 2014. Information Retrieval Models and Searching
Methodologies: Survey. International Journal of Advance Foundation and Research in
Science & Engineering ,India, 1(2).

Barbara Rosario. 2000. Latent Semantic Indexing: An overview. Springer, USA.

Berhanu Anbase. 2019. Applications of Information Retrieval for Afaan Oromo text based on
Semantic-based Indexing. Msc Thesis, Jimma University

Bethlehem Mengistu Hailemariam. 2002. N-gram-Based Automatic Indexing for Amharic Text.
Msc thesis school of information science Addis Ababa, Ethiopia.

Bjornar Larsen and Chinatsu Aone. (1999). Fast and Effective Text Mining Using Linear-time
Document Clustering. Association for Computing Machinry. Cambridge University,, USA.

Bo-Yeong Kang.(2013). A Novel Approach to Semantic Indexing Based on Concept, Department

of Computer Engineering. Kyungpook National University, Metropolitan.

C. J. van Rijsbergen. 1979. Information Retrieval. Elsevier Inc., USA.

Cambridge . 2009. The term vocabulary and posting lists Cambridge University, New-York.

Cambridge. 2009. Flat clustering. Cambridge University, New York.

Cambridge. 2010. Evaluation of information retrieval system. Cambridge University, New York.

Cambridge. 2010. Performance of Two Statistical Indexing Methods, with and without
Compound-word Analysis. Cambridge University Press, New York.

Catherine Griefenow-Mewis . 2001. A Grammatical Sketch of Written Oromo. Koln Rudiger,

Switzerland

Charles L.A. Clarcke, Gordon V. Cormack. 1995. Dynamic Inverted Index for a Distributed Full
text Retrieval System. Journal of Tech. Rep MT, University of Waterloo, Canada.

Choudhary and Bhattacharyya. 2002. Text clustering using semantics. in Proceeding of the 11th
International World Wide Web Conference, Bombay, India.

73
Chris Buckley and Alan.F. Lewit. 1998. Optimizations of inverted vector searches In Proceeding
of the 8th International ACM-SIGIR Conference on Research and Development in
information Retrieval USA: 97-110.

Chris H.Q. Ding. 1999. A Similarity-based Probability Model for Latent Semantic Indexing.
Proceedings of the 22nd International ACM SIGIR Conference on Research and
Development in Information Retrieval, USA: 59-65.

Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. Cambridge, MA. MIT
Press, USA

Christopher D. Manning, Prabhakar Raghavan, Hinrich Sch¨utze, D. 2011. an introduction to

information retrieval. Cambridge University press, New York.

Christopher D. Manning, Prabhakar Raghavan, Hinrich Sch¨utze. 2009. Introduction to

information retrieval.( second, Ed.) Cambridge University press, New York, .

Christopher D. Manning, Prabhakar Raghavan, Hinrich Sch¨utze. 2008. Introduction to

Information Retrieval. (Third, Ed.) Cambridge University Press, New York.

Christopher Issal Magnus Ebbessoni. 2010. Document Clustering. Master of Science Thesis,
University of Gothenburg, Sweden

ChunhuiYuan and Haito Yang. 2019. Research on K-Value Selection Method of KMeans
Clustering Algorithm. Journal of Multidisciplinay Science, China. 2(2), (pp. 226-235).

Daniel Bekele Ramesh Babu, Dereje Teferi. 2015. A Cross Lingual Information Retrieval (CLIR)
System for Afaan Oromo-English using a Corpus Based Approach. International Journal
of Engineering Research & Technology,Open Access, 4(5).

Daniel Bekele Ayana. 2011. Afaan Oromo-English Cross Lanaguage Information Retrieval
(CLIR): A Corpus Based Approach. Msc Thesis, Addis Ababa, Ethiopia.

Daphe Koller.1997. Hierarchically classifying documents using very few words. Proceedings of
the 14th International Conference on Machine Learning,ResearchGate. 170-178.

74
Deerwester Scott, Susan T. Dumais, Richard Harshman. 1989. Indexing by Latent Semantic
Analysis. Journal of the American Society for Information Science, Pennsylvania State
University.

Devika Deshmukh. Sandip Kamble Pranali Dandekar. 2013. Survey on Hierarchical Document
Clustering Techniques. International Journal of Advanced Research in Computer Science
and Software Engineering, 3(7).

Dinakar Jayarajan, Dipti Deodhare, B. Ravindran Sandipan Sarkar. 2000. Document Clustering
using lexical chains. Illinois Institute of Technology,India.

Edie, Rasmussen. 2015. Clustering Algorithm. Journal of American Society for Information
Science, University of Pittsburgh USA 35(1): 268-276).

Ellen M. Vdorhees. 2005. The cluster hypothesis revisited. Proceedings of the 8th Annual
International ACM SIGIR Conference on Research and Development in Information
Retrieval, USA. 85(1): 188-196.

Encyclopedia Britannica 2018. Oromo Language. Retrieved September 24, 2010, from:
DOI:https:// britannica.com/topic/Oromo

Eyob Nigussie. 2013. Afaan Oromo-Amharic cross lingual information retrieval: corpus based
approach. Msc. thesis Addis Ababa University, Ethiopia.

Feyisa Demie. 1996. Historical Challenges In The Development Of The Oromo Language. Journal
of oromo studies, Middle Tennessee State University, USA 3(2):18-27.

Fuhr and Buckley. 1991. A Probabilistic Learning Approach for Document Indexing. ACM
Transactions on Information Systems, 9(3): 224-248.

Gasser, M. (2011). HornMorpho: a system for morphological processing of Amharic, Oromo, and
Tigrinya: Indiana University, India.

Gerard Salton, Michael J. McGill. 1983,. Introduction to Modern Information Retrieval. Journal
of McGraw-HillComputer Science Series, McGraw-Hill Book Company, New York

75
Gerard Salton, Michael J. McGill. 1986. Introduction to modern information retrieval. McGraw-
Hill, Inc., USA

Gezehagn Gutema. 2012. Afaan Oromo Text Retrieval System. Msc Thesis, Addis ababa
Universiity, Ethioppia .

Giuliano Antoniol, Gerardo Canfora, Gerardo Casazza, Andrea De Lucia, Ettore Merlo. 2002.
Recovering traceability links between code and documentation. Transactions on Software
Engineering, IEEE 28(10): 970-983.

Greengrass, Ed. 2001. A survey of Information Retrieval. Survey, RsearchGate.

Grossman, David A., Frieder, Ophir. 2004. Information Retrieval Algorithms and Heuristics.
Springer Science+Business, Springer Netherlands.

Halterman, Richard. 2011. Learning to Program with Python. Southern Adventist University,
Collegedale, Tennessee

Hansaem Park, Kyunglag Kwon, Abdel-ilah Zakaria Khiati, Jeungmin Lee, and In-Jeong Chung.
2015. Agglomerative Hierarchical Clustering for Information Retrieval Using Latent
Semantic Index. International Conference on Smart City, IEEE.

Howard R.Turtle and W. Bruce Croft. (1992). A comparison of text retrieval models. Computer
Journal, Elsevier, USA. 35(3): 279-290.

Howard Robert Turtle, W. Bruce Croft. 2017. Inference Network For Document Retrieval. ACM
SIGIR, 51(1)

Ibrahim Bedane. 2015. The Origin of Afaan Oromo: Mother Language. Global Journal of Human
Social Science: Linguistics and Education,Global Journals Inc., USA 15(12).

James H. Martin, Daniel Jurafsky 2008. Discovery of WordNet Relations. MIT Press
Journals,USA.

Juban and Falguni. 2007. Search Algorithms an Aid to Information Retreival in Digital Libraries.
5th International CALIBER, Indian Statistical Institute, India: 401-414.

76
Judit Bar-Ilan and Tatyana Gutman. 2003. How do Search Engines Handle Non-English Queries?
- A case study. In Proceedings of the Alternate Papers Track of the 12th., Budapest,
Hungary.

K.Halliday Ma and Ruqaiya Hasan. 1976. Cohesion in English. English Language Series, USA.

kerlinger, F.N, and Lee, H.B. 2000. Foundation of Behavioral research. harcourt college publisher
Fort Worth Texas, USA:

Kerry Tanner . 2018. Research Methods. (second, Ed.) ScienceDirect, Elsevier (pp. 159-192).

Khaled Hammouda and Mohamed Kamel. 2010. Collaborative Document Clustering. Journal of
Machine Intelligence Research Group, University of Waterloo, Canada : 451-461.

Kothari, C. R. 2004. Research Methodology. (2nd Ed.). New Age International Publishers, New
Age International, New Delhi

Kowalski, Gerald J. 1997. Information Retrieval Systems: Theory and Implementation. springer

Kula Kekeba Tune, Vasudeva Varma, Prasad Pingali 2007. Oromo-English Cross Language
Information Retrieval. Journal of IJCAI, Springer, Berlin, Germany

Laurence Morissette and Sylvain Chartier. 2013. The k-means clustering technique: General
considerations and implementation in Mathematica . ResearchGate, Berlin, Germany.
9(1): 15-24.

Lei Shi and Rada Mihalcea. 2005. Putting pieces together:Combining FrameNet, VerbNet and
WordNet for robust semantic parsing. In Computational linguistics and intelligent text processing,
University of North Texas, Texas.100-111.

Lewis, Melvyn. 2009. Ethnologue: Languaes of the World. 16th Ed. SIL Internationl,
ResearchGate

Loulwah et al. 2007. Local Semantic Kernels for Text Document Clustering. In Workshop on Text
Mining, SIAM International Conference on Data Mining, George Mason University.

77
M,.A Syakur, B K . KHotimah, E M S Rochiman, B D Statoto. 2018. Integration K-means
clusteing Method and Elbow Method for Identification of the Best Customer Profile. IOP
Conference Seies Materials Science and Engineering, NASA, USA,

Magnus Rosell. 2006. Introduction to Informtion Retrieval and Text Clustering . Journal of KTH
CSC, New York Texas.

Marco Suárez Barón, Kathleen Salinas Valencia. 2009. An approach to semantic indexing and
information retrieval. Revista Facultad de Ingeniería Universidad de Antioquia, núm.
Universidad de Antioquia Medellín, Colombia. 174-187

Martin Brascheler and Barbel Ripplinger. 2004. . How Effective is Stemming and Decompounding
for German Text Retrieval. Proceedings of ACL-08: HLT, Short Papers , Columbus, Ohio,
USA. pp 253–256

Md Maruf Hasan Yuji Matsumoto. 2012. Document Clustering: Before and After the Singular
Value Decomposition. American Control Conference USA.

Megersa, Diriba. 2002. An Automatic Sentence Parser for Oromo Language using Supervised
Learning Technique, Msc. thesis, Addis Ababa University, Ethiopia.

Melkamu Abetu. 2017. Query expansion for Afaan oromo information retrieval based on wordnet.
Msc. Thesis, Haramaya University, Ethiopia

Michael Steinbach George Karypis Vipin Kumar. 2009. A Comparison of Document Clustering
Techniques. University of Minnesota, Minnesota, USA.

Morka Mekonnen. 2001. Text-to-Speech System for Afaan Oromo. Msc. thesis at school of
Information studies for Africa, Addis Ababa University, Ethiopia.

Moukdad Haidar. 2003. Lost In Cyberspace: How Do Search Engines Handle Arabic Queries?
The 12th International World Wide Web Conference.

Mulualem Wordofa. 2013. Semantic Indexing and Document Clustering for Ahmaric Information
Retrieval. Msc Thesis, Addis Ababa University, Ethiopia.

78
M. Ramakrishna Murthy , J.V.R. Murthy , P.V.G.D. Prasad Reddy and Suresh C. Sapathy. 2014.
A new approach for finding appropriate number of Clusters using SVD along with
determining Best Initial Centroids for K-means Algorithm. International Journal of
Enhanced Research in Science Technology & Engineering. 3(3) 430-437

Mylanguage.org. 2011. Oromo Numbers. Retrieved September 21, 2020, from Doi:
http://www.mylanguages.org/learn_oromo.php

Norbby Saiyad, Prajapati and Dabhi. 2016 "A survey of document clustering using semantic
approach," International Conference on Electrical, Electronics, and Optimization
Techniques (ICEEOT), Chennai, 2016. 2555-2562,

Naik, M.P., Prajapati, H.B., & Dabhi, V.K. (2015). A survey on semantic document
clustering. 2015 IEEE International Conference on Electrical, Computer and
Communication Technologies (ICECCT), India. 1-10.

Naohiro Matsumura, Yikio Ohsawa, Mitsuru Ishizuka. 2015. A Survey Paper On Different
Techniques Of Document Clustering. International Journal of Current Engineering and
Scientific Research (IJCESR), 2(1): 2393-8374.

Naohiro Matsumura, Yukio Ohsawa, Mitsuru Ishizuka. 2002. Priming Activation Indexing:
Automatic Indexing for Extracting Asserted Keywords from a Document. ResearchGate,
Berlin, German.

Neelam Phadnis and Jayant Gadge. (2014). Framework for Document Retrieval using Latent
Semantic Indexing. International Journal of Computer Applications, Berlin, German.
94(14)

Neepa Shah and Sunita Mahajan. 2012. Document Clustering: A Detailed Review. International
Journal of Applied Information Systems (IJAIS),Foundation of Computer Science, New
York, USA , 52(5): 42-52.

Nega Alemayehu and Peter Willett. 2002. Stemming of Amharic words for information retrieval,
in Literary Linguistic Computing, Literary and Linguistic Computing University of
Sheffield, Sheffield, UK. 17(1): 1-18.

79
Oren Zamir and Oren Etzioni. 1998. Web document clustering: a feasibility demonstration. In
Proceedings of the 21st annual international ACM SIGIR conference on Research and
development in information retrieval (SIGIR '98). Association for Computing Machinery,
New York, NY, USA, 46–54

Oren Zamir Oren Etzioni Omid Madani Richard M. Karp.1997. Fast and Intuitive Clustering of
Web Documents. In Proceedings of the 3rd International Conference on Knowledge
Discovery and Data Mining, AAAI Press. 287-290.

Paralic Jan and Ivan Kostial. 1999. Ontology-based Information Retrieval . journal Web
Technologies, ResearchGate, USA.

Paul Mcnamee. Charles Kenneth. Nicholas James. Clifton Mayfield. 2009. Addressing
Morphological Variation in Alphabetic Languages. In Proceedings of the 32nd.,.
international ACM SIGIR conference on Research and development in information
retrieval, Cambridge University Press, USA.

Ricardo Baeza-Yates, Carlos Castillo, Felipe Saint Jean. 2004. Web Dynamics, Structure and Page
Quality. Springer, Berlin, 93-109.

Richard, H. 2003. Encyclopedia of English. Retrieved October 20, 2017, from An encyclopedia of
English Grammar and Word Grammar: http://www.phon.ucl.ac.uk/home/dick/enc-
gen.htm

Robert Krovetz. 2000. Viewing morphology as an inference process. Elsevier Princeton, USA,
118, 277–294.

Sajendra Kumar, Ram Kumar Rana , Pawan Singh, 2012. Ontology based Semantic Indexing
Approach for Information Retrieval System. International Journal of Computer
Applications, Berlin, German, 49(12).

Sara Norrby. 2016. Using Morphological Analysis in an Information Retrieval System for
Résumés, Stockholm, Sweden.

Savoy Picard and Jacques Savoy. 2000. A logical information and retireval model based a
combination propositional logic and probability theory. ResearcGate, Berlin, German.

80
Septian Wulandari. 2020, January). Analyze K-Value Selected Method of K-Means Clustering
Algorithm to Clustering Province Based on Disease Case. International Journal of
Innovative Technology and Exploring Engineering (IJITEE), Blue Eyes Intelligence
Engineering & Sciences Publication, 9(3).

Singhal, A. 2008. Modern Information Retrieval. Retrieved October, from A Brief Overview,
Bulletin of the IEEE Computer Society Technical Committee on Data Engineering, IEEE.

Stefano Ceri, Alessandro Bozzon, Marco Brambilla Emanuele Della Valle,Piero Fraternali, Silvia
Quarteroni. 2013. Web Information Retrieval, Data-Centric Systems and Applications,
Proceedings of the IEEE-PIEEE.

Stephen E. Robertson. 2012. The methodology of information retrieval experiment, springer,

Berlin, German.

Susan T. Dumais. 2007. Enhancing Performance in Latent Semantic Indexing (LSI) Retrieval. 40th
Hawaii International Conference on Systems Science, Waikoloa, Big Island, HI, USA. 1-
19.

Tesfaye Guta. 2010. Afaan Oromo Search Engine. Msc Thesis, Addis Ababa University, Ethiopia.

Tewodros Hailemeskel. 2003. Amharic text retrieval: an experiment using Latent semantic
indexing (LSI) with singular value decomposing (SVD). Msc thesis, Addis Ababa
University, Ethiopia.

Tewodros Abebaw. 2014. Applying Thesaurus Based Semantic Compression For Improving The
Performance of Amharic Text Retrieval. Msc thesis, Addis Ababa University, Ethiopia.

Thomas Hofmann. 1999. Probabilistic latent semantic indexing. In Proceedings of the 22nd annual
international ACM SIGIR conference on Research and development in information retrieval
(SIGIR '99). Association for Computing Machinery, New York, NY, USA, 50–57.

Thomas Hofmann. 1999. Probabilistic Latent Semantic Indexing. Journal of ACM SIGIR, USA.
51(2)

81
Tilahun Gamta. 2000. The Politicization of My Oromo-English Dictionary. Journal of Oromoo
Studies. Addis Ababa University, Ethiopia, 7(2): 1-17.

Unubi, Abraham. 2017. Selected Derivational Morphological Processes in English, Hausa, Igala
and Some other Languages of the World. World Wide Journal of Multidisciplinary Research and
Development, Nigeria

Vallet David, Fernández Miriam, and Castells Pablo. 2013. An Ontology-Based Information
Retrieval Model,.The Semantic Web: Research and Applications, Endland. 455-470.

W. Bruce Croft. 1995. What Do People Want from Information Retrieval?, Doi:
hdl://cnri.dlib/november95-croft ‘.

Wakshum Mekonen. 2000. Development of a stemming algorithm for Afaan Oromo text, MSc.
thesis,: Addis Ababa University, Ethiopia.

Wegari Getachew. 2017. OroRoots: Rule-Based Root Generation System for Afaan Oromo. Addis
Ababa University Ethiopia.

Woldemariam, Assefa. 2005. Development of Morphological Analyzer for Afaan Oromo. MSc.
thesis. Addis Ababa University Ethiopia

Workineh Tesema and Debela Tesfaye. 2017. Word Sense Disambiguation and Semantics for
Afan Oromo Words using Vector Space Model. International Journal of Research Studies
in Science, Engineering and Technology 4(6), 10-15

Xiaoyong Liu and W. Bruce Croft. 2004. Cluster-based retrieval using language models. In
Proceedings of the 27th annual international ACM SIGIR conference on Research and
development in information retrieval (SIGIR '04). Association for Computing Machinery,
New York, NY, USA, 186–193

Yu Xiao. 2010. Survey of Document Clustering Techniques Comparison of LDA and moVMF.

Zipf, G. 1932. Selected Studies of the Principle of Relative Frequency in Language. Cambridge:
Harvard University press, USA.

82
 7. APPENDICES
Appendix I- Elbow Method Implementation Code and Graph

wcss =[]

for i in range(1,12):

kmeans = KMeans(i)

kmeans.fit(lowRankX)

wcss_iter = kmeans.inertia_

wcss.append(wcss_iter)

number_clusters = range(1,12)

plt.plot(number_clusters, wcss)

plt.title('the Elbow method')

plt.xlabel('number of clusters')

plt.ylabel('within cluster sum of square')

83
Appendix II-Implementation Code in Python

Indexing Code

import sys

import glob

import numpy

import string

from matplotlib import pyplot

from nltk.corpus import stopwords

import pandas as pd

from string import punctuation

import matplotlib.pyplot as plt

import l3

import numpy

import pandas as pd

# Structures holding documents and terms

document_list = []

document_ids = {}

token_list = []

token_ids = {}

# Repeat for every document in processed_pages

for filename in glob.glob('corpus/*'):

f = open (filename, 'r', encoding = 'utf-8')

84
 tokens = f.read()

tokens = tokens.lower()

tokens = tokens.split()

table = str.maketrans(' ', ' ', punctuation)

tokens = [w.translate(table) for w in tokens]

tokens =[word for word in tokens if word.isalpha()]

tokens =[word for word in tokens if len(word)>2 ]

stop_words = set(stopwords.words('oromo'))

tokens = [w for w in tokens if not w in stop_words]

f.close()

# Get the document name as last part of path

article_name = filename[filename.rfind('/')+1:]

doc_id = len(document_list)

# Insert ID in inverse list

document_ids[article_name] = doc_id

# Populate token structure for all tokens in document

for t in tokens:

if t not in token_ids:

token_ids[t] = len(token_list)

token_list.append(t)

# Transform the document's token list into the corresponding ID list

tids = [token_ids[t] for t in tokens]

85
# Store the document as both its token ID list and the corresponding set

document_list.append({

'name': article_name,

'tokens': tids,

'set': set(tids)

})

tokens = open('token_list.txt', 'w', encoding ='utf-8')

tokens.write(str(token_list))

tokens.close()

#stemming using hornmorph stemmer

l3.anal_file('om','token_list.txt', 'stemmed.txt' , citation =True, nbest =1)

stemmed_tokens = open('stemmed.txt', 'r', encoding = 'utf-8')

stem_list = stemmed_tokens.read()

number_of_documents = len(document_list)

number_of_tokens = len(stem_list)

# Building the TF-IDF matrix

sys.stderr.write ('Building the TF matrix and counting term occurrencies\n')

token_count = [0] * number_of_tokens # number of occurrences of word in document

TF = numpy.empty((number_of_tokens,number_of_documents), dtype=float)

# Scan the document list

for i,doc in enumerate(document_list):

# Initialize with zeros

86
 n_dt = [0] * number_of_tokens

# For all token IDs in document

for tid in doc['tokens']:

# if first occurrence, increase global count for IDF

if n_dt[tid] == 0:

token_count[tid] += 1

# increase local count

n_dt[tid] += 1

# Normalize local count by document length obtaining TF vector;

# store it as the i-th column of the TFIDF matrix.

TF[:,i] = numpy.array(n_dt, dtype=float) / len(doc['tokens'])

TF[] = numpy.array(n_dt, dtype=float) / len(doc['tokens'])

IDF = numpy.log10(number_of_documents / numpy.array(token_count, dtype=float))

TFIDF = numpy.diag(IDF).dot(TF)

TFIDF = pd.DataFrame(TFIDF)

#SVD for rank reduction

To, So, DoT = numpy.linalg.svd (TFIDF, full_matrices =False)

#r = number of low rank

#Y = TSD

r = 10

87
T = To[:, :r]

S = numpy.diag(So)[:r, :r]

DT = DoT[:r, :]

Y = numpy.dot(T, numpy.dot(S, DT))

#Kmeans clustering using reduced SVD

from sklearn.cluster import KMeans

n_clusters = 8

kmeans = KMeans(n_clusters, max_iter=100, n_init=1, verbose=1)

cluster = kmeans.fit_predict(Y) #Y=TSDT or reduced SVD using r number

label = kmeans.labels_

centroid = kmeans.cluster_centers_ #centroid of the clusters

sorted_centroid = centroid.argsort()[:, ::-1] #sorting centroid

centroid = pd.DataFrame(centroid)

centroid = kmeans.cluster_centers_ #centroid of the clusters

sorted_centroid = centroid.argsort()[:, ::-1] #sorting centroid

centroid = pd.DataFrame(centroid)

88
searching code

query = input('please enter query/Jechaa keessan galchaa!! ')

query = query.lower()

q_split = query.split()

q_split = [w.translate(table) for w in q_split]

q_split =[word for word in q_split if word.isalpha()]

q_split =[word for word in q_split if not word.isdigit()]

q_split =[word for word in q_split if len(word)>2 ]

stop_words = set(stopwords.words('oromo'))

q_split = [w for w in q_split if not w in stop_words]

q_tokens = open('q_split.txt', 'w', encoding ='utf-8')

q_tokens.write(str(q_split))

q_tokens.close()

l3.anal_file('om','q_split.txt', 'stemm_q.txt' , citation =True, nbest =1)

fq = open('stemm_q.txt', 'r', encoding = 'utf-8')

qr_list = fq.read()

q_tokens = set()

q_count = [0] * number_of_tokens

q_length = 0

for token in q_list:

try:

t_id = token_ids[token]

q_count[t_id] += 1

89
 q_length += 1

q_tokens.add(t_id)

except:

Pass

q_TFIDF = (numpy.array(q_count,dtype=float) / q_length) * IDF

sim = TFIDFM.T.dot(q_TFIDF)

#retrieve query result using VSM

N= len(document_list)

def print_top(sim,N,smallest=False):

sorted_sim = sorted(enumerate(sim),key=lambda t:t[1], reverse= not smallest)

for i,s in sorted_sim[:N]:

if s>0:

print( document_list[i]['name'])

#retrieve query result using SVD

# Compute the cosine similarity array given the SVD decomposition of the TFIDF matrix
(computed above), the normalized TFIDF query vector q and the desired rank r

def reduced_similarity (r, To, So, DoT, q):

T = To[:, :r]

S = numpy.diag(So)[:r, :r]

DT = DoT[:r, :]

q_r = S.dot(T.T).dot(q_TFIDF)

90
 return DT.T.dot(q_r)

sim_r = reduced_similarity (r, To, So, DoT, q_TFIDF)

print_top (sim_r,N)

#retrieve query result using SVD with K-means

#calculating similarity b/n cluster centroid and query vector

c_qsim =centroid.dot(q_TFIDF)

my_dict =c_qsim.to_dict()

def get_key(val):

for key, value in my_dict.items():

if val == value:

return key

# w is cluster with max similarity b/n query and centroid

w = get_key(c_qsim.max())

doc_dict = {i: numpy.where(label ==i)[0] for i in range(kmeans.n_clusters)}

doc_dict_key = list(doc_dict.keys())

doc_dict_val = list(doc_dict.values())

doc_list = []

for key in doc_dict.keys():

if w == key:

print ( 'doc:', list(doc_dict[w]))

91
Appendix III- Sample of Proposed work Search Results

Vector Space Model (VSM) using ‘barnoota’ query word result

92
Single Value Decomposition (SVD) using ‘barnoota’ query word result

93
Single Value Decomposition with K-means Clustering using ‘barnoota’ query word result

94