ADDIS ABABA UNIVERSITY

SCHOOL OF GRADUATE STUDIES

INFORMATION EXTRACTION MODEL FROM AMHARIC

NEWS TEXTS

BY: Getasew Tsedalu

A THESIS SUBMITTED TO THE SCHOOL OF GRADUATE STUDIES OF

ADDIS ABABA UNIVERSITY IN PARTIAL FULFILMENT FOR THE
DEGREE OF MASTERS OF SCIENCE IN COMPUTER SCIENCE

November, 2010
 ADDIS ABABA UNIVERSITY
SCHOOL OF GRADUATE STUDIES
FACULTY OF COMPUTER and MATHEMATICAL SCIENCE
DEPARTMENT OF COMPUTER SCIENCE

INFORMATION EXTRACTION MODEL FROM AMHARIC NEWS

TEXTS

BY: Getasew Tsedalu

ADVISOR:

Solomon Atnafu (PhD)

APPROVED BY
EXAMINING BOARD:
1. Dr. Solomon Atnafu, Advisor ______________________
2. _______________________ ______________________
3. _______________________ ______________________
 Dedication

For dad: You have been working so hard to raise me, my two brothers and my sister. You are
the hero in my life the things you did to me and our family will always stay in my heart. May
God bless you and give you a long life and good health.

For mom: you are always positive and supportive of things that I do. I thank you for being the
light of my life.

II
 Acknowledgment

All the praise goes to God, who gives me the strength from beginning to the end of this research
work. Doing something new has a lot of obstacles which sometimes make someone to quit
instead of striving for its successful completion. In those times to overcome the struggle that I
face different individuals has their own contribution for the successfully completion of the work.
The first person I would like to thank is my advisor Dr. Solomon Atnafu for his suggestive and
constructive comments which strengthen this research work. I would also like to acknowledge
Tesfaye Guta, who has been working day in day out with me in the lab. I extend my thanks to the
officials in ENA for providing me the necessary data that is important for training and testing of
the prototype developed. Finally my gratitude goes to Abebe Abeshu, Teklay G/Egiziabher,
Mequannt Munye, Mandefro Kejela, Esmael Kedir, Desta Berihu, Michael Shiferaw, Moges
Ahmed and all my classmates for the discussion we have and for the ideas we share which was
very helpful for the successful completion of this work.

III
 Table of Contents
List of Tables .............................................................................................................................. VIII

List if Figures ................................................................................................................................ IX

Acronyms & Abbreviations ........................................................................................................... X

Abstract ......................................................................................................................................... XI

CHAPTER ONE ............................................................................................................................. 1

1. INTRODUCTION....................................................................................................................... 1

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

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

1.3 Objective ............................................................................................................................................. 3

1.4 Scope and Limitation .......................................................................................................................... 4

1.5. Methodology ...................................................................................................................................... 4

1.5.1 Literature Review ......................................................................................................................... 4

1.5.2 Analysis of the Amharic Language .............................................................................................. 4

1.5.3 Design and Implementation of ATIE System .............................................................................. 5

1.5.4 Development Tools ...................................................................................................................... 5

1.6 Application of Results ......................................................................................................................... 5

1.7 Thesis Organization............................................................................................................................. 6

CHAPTER TWO............................................................................................................................. 7

2. LITERATURE REVIEW............................................................................................................ 7

2.1 Information Extraction ........................................................................................................................ 7

2.2 Related NLP Fields to Information Extraction.................................................................................... 8

2.2.1 Information Retrieval ................................................................................................................... 8

2.2.1 Text Summarization ..................................................................................................................... 8

2.2.3 Question and Answering System ................................................................................................. 9

IV
 2.3 Architecture of IE System .................................................................................................................10

2.3.1 Preprocessing of Input Texts ......................................................................................................15

2.3.2 Learning and Application of the Extraction Model ....................................................................16

2.3.3 Post Processing of Output ..........................................................................................................17

2.4 Approaches for Information Extraction.............................................................................................17

2.4.1 Knowledge Engineering Approach ............................................................................................18

2.4.2 Machine Learning Approach ......................................................................................................18

2.5 Information Extraction as a Text categorization Task ......................................................................21

2.6 Evaluation Metrics ............................................................................................................................22

3. RELATED WORKS ................................................................................................................. 25

3.1 Information Extraction from English Text ........................................................................................25

3.2 Information Extraction from Portuguese Text ..................................................................................27

3.3 Information Extraction from Thai Text .............................................................................................28

3.4 Information Extraction from Chinese Text .......................................................................................30

3.5 Information Extraction from Spanish Text........................................................................................32

3.6 Information Extraction and Related Works on Amharic Text ..........................................................33

3.6.1 Text Categorization from Amharic Text ....................................................................................33

3.6.2 Information Extraction from Amharic Text ...............................................................................35

Summary .................................................................................................................................................35

CHAPTER FOUR ......................................................................................................................... 37

4. THE AMHARIC LANGUAGE ................................................................................................ 37

4.1 The Amharic Language .....................................................................................................................37

4.2 Grammatical Structure of Amharic ...................................................................................................37

4.3 Amharic Punctuation Marks and Numerals ......................................................................................38

4.4 Characteristics of Amharic writing ...................................................................................................40

V
 4.5 Sentences in Amharic ........................................................................................................................42

Summary .................................................................................................................................................43

CHAPTER FIVE ........................................................................................................................... 44

5. THE AMHARIC TEXT INFORMATION EXTRACTION MODEL ..................................... 44

5.1 Components of ATIE ........................................................................................................................44

5.2 Document Preprocessing ...................................................................................................................46

5.3 The Text Categorization Component ................................................................................................47

5.4 Learning and Extraction Component ................................................................................................49

5.5 Post Processing..................................................................................................................................51

Summary .................................................................................................................................................51

CHAPTER 6.................................................................................................................................. 53

6. IMPLEMENTATION ............................................................................................................... 53

6.1 Data Source .......................................................................................................................................53

6.2 Document Preprocessing ...................................................................................................................54

6.2.1 Tokenization ...............................................................................................................................54

6.2.2 Character Normalization ............................................................................................................55

6.2.3 Number Prefix Separator ............................................................................................................56

6.2.4 Number Normalization ...............................................................................................................56

6.3 Text Categorization ...........................................................................................................................59

6.3.1 Training Data Preparation ..........................................................................................................59

6.3.2 Training a Classifier Model ........................................................................................................62

6.3.3 Using the Trained Classifier Model ...........................................................................................63

6.4 Learning and Extraction ....................................................................................................................63

6.4.1 Training Data Preparation ..........................................................................................................63

6.4.2 Processing of the News Texts for Feature Extraction ................................................................64

VI
 6.4.3 Feature Extractor ........................................................................................................................66

6.4.4 Preprocessing the Data Using Weka ..........................................................................................69

6.4.5 Training a Classifier Model ........................................................................................................69

6.4.6 Using Trained Classifier for Information Extraction .................................................................69

6.5 Post Processing of the Data ...............................................................................................................70

Summary .................................................................................................................................................72

CHAPTER 7.................................................................................................................................. 73

7. EXPERIMENT.......................................................................................................................... 73

7.1 Experimental Procedures...................................................................................................................73

7.1.1 Data Collection ...........................................................................................................................73

7.2 Performance Evaluation ....................................................................................................................74

7.2.1 Evaluation of Text Categorization Component ..........................................................................74

7.2.2 Evaluation of IE component .......................................................................................................79

7.2.2.1 Discussion on the Experimental Results of IE ............................................................................86

Summary ....................................................................................................................................... 87

CHAPTER 8................................................................................................................................ 878

CONCLUSION, CONTRIBUTION AND RECOMMENDATION ............................................ 88

8.1 Conclusion.........................................................................................................................................88

8.2 Contribution of the Work ..................................................................................................................89

8.3 Recommendation ...............................................................................................................................90

Reference ....................................................................................................................................... 91

Appendices .................................................................................................................................... 95

Appendix A: List of Stop Words.................................................................................................................95

Appendix B: List of Titles............................................................................................................. 95

Appendix C: List of Infrastructure Names .................................................................................... 97

VII
List of Tables

Table 2.1: Two Class Confusion Matrix Example ........................................................................ 23

Table 4.1: Number Representations in Amharic ........................................................................... 39

Table 4.2: Amharic Fraction and Ordinal Representation ............................................................ 40

Table 4.3 Amharic Characters with Same Sound ......................................................................... 41

Table 4.4 Word Spelling Variations .............................................................................................. 41

Table 4.5 Word Variations due to Pronunciation .......................................................................... 42

Table 6.1 Amharic Characters which have different symbols but similar sound ......................... 55

Table 7.1 Confusion matrix for text categorizationcomponent using decision tree...................... 75

Table 7.2 Detailed accuracy by class for text categorizationcomponent using decision tree ....... 76

Table 7.3 Confusion matrix for text categorizationcomponent using Naïve Bayes...................... 77

Table 7.4 Detailed accuracy by class for text categorizationcomponent using Naïve Bayes ....... 78

Table 7.5 Confusion matrix for text categorizationcomponent using SMO ................................. 78

Table 7.6 Detail accuracy by class for text categorizationcomponent using SMO....................... 79

Table 7. 7 Confusion matrix for IE component for scenario 1 using SMO .................................. 81

Table 7. 8 Detailed accuracy by class for IE component for scenario 1 using SMO .................... 81

Table 7. 9 Confusion matrix for IE component for scenario 2 using Naïve Bayes ...................... 82

Table 7. 10 Detailed accuracy by class for IE component for scenario 2 using Naïve Bayes ...... 83

Table 7. 11 Confusion matrix for IE component for scenario 3 using SMO ................................ 84

Table 7. 12 Detail accuracy by class for IE component for scenario 3 using SMO ..................... 84

Table 7. 13 Confusion matrix for IE component for scenario 4 using SMO ................................ 85

Table 7. 14 Confusion matrix for IE component for scenario 4 using SMO ................................ 86

VIII
List of Figures

Figure 5.1 The Amharic Text IE Model ........................................................................................ 45

Figure 5.2 The Information Extraction Subcomponent ................................................................ 51

Figure 6.1 Algorithm for Number Normalization ......................................................................... 59

Figure 6.2 Algorithm for Name Prefix Separator ......................................................................... 66

Figure 6.3 Algorithm for Feature Extractor .................................................................................. 68

Figure 6.4 Sample prototype system after the data is preprocessed............................................. 71

IX
Acronyms & Abbreviations
IE: Information Extraction

IR: Information Retrieval

ATIE: Amharic Text Information Extraction

NLP: Natural Language Processing

MUC: Message Understanding Conference

ARFF: Attribute Relation File Format

GATE: General Architecture for Text Engineering

NLQA: Natural Language Question Answering

ST: Scenario Template Production

TR: Template Relation Construction

TE: Template Element Construction

NER: Named Entity Recognition

POS: Part of Speech Tagger

ENA: Ethiopian News Agency

ROC: Receiver Operating Characteristic

X
Abstract
As the growth of unstructured documents in the web and intranet is increasing from time to time,
a tool that can extract relevant data to facilitate decision making is becoming crucial. IE is
concerned with extraction of relevant information from text and stores them in a database for
easy use and management of the data. As the first comprehensive work on IE from Amharic text
we designed a model that is genuine enough to deal with different domains in the Amharic
language. The proposed model has document preprocessing, text categorization, learning and
extraction and post processing as its main components. The document preprocessing component
handles the normalization of the document while text categorization and learning and extraction
handle the categorization of the news text and extracting the predefined relevant information
from the categorized text respectively. The post processing component format and save the
extracted data to the database.

Various evaluation techniques, which are used to evaluate the performance of the classifier
machine learning algorithms, are used for IE and text categorization. Among the different
classifier machine learning algorithms used for text categorization component, the Naïve Bayes
algorithm performs by correctly classifying 92.83% of the 1200 news texts used as a dataset. On
the other hand, 1422 instances are used for training and testing the Information Extraction
component. Different scenarios are used to evaluate the role of the different features in
predicting the category for the candidate texts. Among the different scenarios we considered and
the different machine learning algorithms we employed the SMO algorithm correctly classified
94.58% of the instances correctly, when all the features are considered which yields higher
precision and recall rate for the different attributes considered for extraction.

Key words: Amharic Text Information extraction, Machine Learning Approach to Information
Extraction, Amharic Text Categorization, Information Extraction

XI
 CHAPTER ONE

1. INTRODUCTION

1.1 General Background

The explosive growth of documents on the web and intranet is highly increasing the amount of
information available in unstructured and semi structured format from time to time. Every year a
large amount of books, journals and other materials are published which will take more than a
life time for human beings to read and use the important information contained in them [1]. The
numbers of Amharic documents on the web and in other machine readable forms are also
increasing from time to time. As a result of this growth, the huge amount of text which contains
different valuable information which can be used in education, business, security and other many
areas are hidden under the unstructured representation of the textual data. This shows that getting
the right information for decision making from existing abundant unstructured text is a big
challenge. The non availability of tools for extracting and exploiting the valuable information
which is effective enough to satisfy the users need have been a major problem for years [2].

Information analysts, who have been working in information extraction for the database input,
have to do the extraction manually. They have to go through all the text written in the document
and extract the information that will be used as an input for the database, categorize them
according to their properties and identify their relationship manually. Extracting the information
manually from a large amount of unstructured text is a very tiresome and time consuming job
[3]. On the other hand, users who want to use the available unstructured information for different
purposes have to read all the relevant texts that are related to their need and have to manually
extract the information they want from the abundant unstructured text which takes very long time
and which in turn highly minimizes the users efficiency in decision making.

To solve this problem a lot of researches have been conducted in the area of information
extraction, information retrieval and NLP. Information Extraction (IE) is the automatic
extraction of facts from text, which includes detection of named entities, entity relations and
1
events used to extract facts from unstructured text. It is based on analyzing natural language in
order to extract information. The process takes texts (and sometimes speech) as an input and
produces structured data as an output. This data may be used directly by users, or may be stored
in a database or spreadsheet for later analysis, or may be used for indexing purposes in
information retrieval (IR) applications such as web search engines [4, 5].

Information extraction is quite different from information retrieval. An IR system finds relevant
texts that is based on a query and presents them to the user while an IE application analyzes texts
and presents only the specific information from it that the user is interested in. For example, a
user of an IR system wanting information on trade group formations in agricultural commodities
markets would enter query statements and receive in return a set of documents (e.g., newspaper
articles) that contain likely matches. The User then would have to read the documents and extract
the necessary information from the retrieved data. In contrast, an IE system would automatically
populate a spreadsheet directly with the names of relevant information instead of the text which
contain the information. There are advantages and disadvantages with IE in comparison to IR. IE
systems are more difficult and knowledge-intensive to build, and are to varying degrees tied to
particular domains and scenarios. IE is more computationally intensive than IR. However, in
applications where there are large text volumes IE is potentially much more efficient than IR
because of the possibility of dramatically reducing the amount of time people spend reading texts
[4].

Information extraction has three different components regardless of the language and domain on
which it is developed for, which are linguistic preprocessing, the learning and application stage
and semantic post processing to do the extraction of data from a given text. The linguistic
preprocessing uses different tools to make the natural language texts ready for extraction. The
learning and the application component learns a model and extract the required information from
the preprocessed text and the last component which is the semantic post processing assign the
extracted information into their predefined attribute category and manages the normalization and
duplication problem with the extracted data [6].

Information extraction unlike the other research domains is language and domain dependent [7].
Different researchers tried to develop information extraction systems for different languages and

2
domains. The IE system developed for English and any specific domain in English cannot work
for Amharic language as there are different language and domain specific issues which cannot be
handled by the system developed for English. This makes it there is a need to do a research in
extracting data from the Amharic texts as most of the information that exist in Amharic language
is unstructured. This initiates us to engage in research to develop a model for extracting facts
from the Amharic news texts and then develop a prototype to realize the efficiency of the model
in extracting facts which will be used as an input for the database system.

1.2. Statement of the Problem

The IE tools developed so far are language and domain dependent. The information extraction
system developed for English or any other language and for some specific domain cannot work
for other languages of the same domain. This is due to the reason that IE system has to be trained
about the different nature of the language and the domain for which they are developed for. If
someone try to use an information extraction system which is developed for English and use it to
work on the same domain for Amharic the result will be very poor. Thus, in this work we will
design an effective information extraction model for Amharic language as to the best knowledge
of the researcher there is no information extraction system which is developed for Amharic
language so far.

1.3 Objective
General Objective

The general objective of this research work is to design a generic model of information
extraction system from Amharic news text using a classification machine learning approach and
evaluate its efficiency by developing a prototype Amharic text IE System.

Specific Objective

To achieve the general objective, the following specific objectives are identified:

Analyse the general grammatical structure of Amharic language for the purpose of
identifying its characteristics for IE.

3

Identify the representation of different entities on Amharic news text and its grammatical
structure.

Develop a model for IE from Amharic news text

Use the existing algorithms or developing a new algorithm for IE from Amharic news
text

Identify and collect a corpus of the Amharic news text.

Develop a prototype system based on the model

Evaluate the effectiveness and usability of the algorithm developed for Amharic IE

1.4 Scope and Limitation

Information extraction is a very complex and rigorous task which needs the understanding of the
natural language and the specific domain in need. A full-fledged information extraction system
will require a number of Natural language processing tools such as Sentence Parser, Part of
Speech tagger (POS), Named Entity Recognizer (NER), Co-reference Resolution and others.
Even though some of the NLP systems for Amharic language have been done by some
researchers, they are not publicly available. Having these limitations in mind, the scope of the
research work is limited in extracting numeric and name data from the Amharic news text and
assign them to the predefined database slot. The extraction of relationship between entities is out
of the scope of this research work. News texts that are about a single issue are only considered.

1.5. Methodology
To achieve the general and specific objectives of the study, the following methods are employed

1.5.1 Literature Review

Literature review have been conducted on books and research works which are related to
information extraction and related fields to have understanding of the problem domain and the
related researches that are done so far, which helps for the better understanding to design a new
model for Amharic language.

1.5.2 Analysis of the Amharic Language

The different facts about Amharic language like the grammatical structure the number
representation and other language specific issues that are important for the research work have
4
been reviewed and presented. It helps to understand the nature of the language with regard to
information extraction.

1.5.3 Design and Implementation of ATIE System

A classification machine learning approach is used to develop ATIE (Amharic Text Information
Extraction) system. It contains the document preprocessing, text categorization, extraction and
learning and post processing as a four main components.

1.5.4 Development Tools

In order to develop a prototype system, different appropriate tools have been selected and used.
The java programming language is used to implement the different language specific algorithms
developed. Weka is used to handle the classification of the candidate text and news text in case
of information extraction and text categorization respectively. The POS which is used as one of
the features in IE component is developed using HMM (Hidden Markov Model) available in the
Ling Pipe open source NLP software.

1.6 Application of Results

IE system for Amharic news text can be used to extract information that exists in the news text in
the way that it can be used as an input for database or for other applications. Nowadays most of
the available information is in an unstructured textual format. Since we cannot query it in simple
ways, automatic systems cannot use the contained information and human beings cannot easily
access and manage it. The IE system for Amharic news text can be used for better use of the
contained information in the Amharic news texts.

Implementing the IE system for the Amharic news text will have the following uses:

I. The unstructured information which exist in the Amharic news text will be structured,
which make it easy for the different users to access it within a short period of time. For
example, if somebody wants to know the infrastructures that are developed in different
part of the country, he/she doesnot have to read all the news papers. The IE system for
the Amharic news will extract the summarized information related to infrastructure and
store it in the database for easy use and management. The users can then easily access the
data by using a database query language.
5
 II. The extracted structured information from unstructured text can be used as an input for
the other applications. Since that data is structured, applications can share the information
and use it for different purposes.

1.7 Thesis Organization

The rest of this thesis report is organized as follows. Chapter 2 discusses the different issues in
IE and the related subject areas as literature review. This Chapter lays the foundation in
understanding what an IE system comprises of, what approaches are used, and the different
components which are needed by IE system. Chapter 3 is devoted to discuss related works done
on IE systems in different languages and on different domains. Chapter 4 discusses the Amharic
language with regard to IE. Many language specific issues such as the writing system have been
presented. The architectural and design issues of our system are discussed in Chapter 5. The
main components of our system, their functional operation and the specific sub-component of
each component are briefly discussed in this Chapter. Chapter 6 is devoted to discussing the
main implementation issues of our IE system. The algorithms, techniques and methods used in
how the system has been successfully developed are discussed in this Chapter. Chapter 7
presents the evaluation of the system. Chapter 8 concludes the thesis by outlining the
contribution of the research work. It also shows some research directions that can be used in the
future to improve IE system for Amharic.

6
 CHAPTER TWO

2. LITERATURE REVIEW
In this Chapter, we provide a brief overview of the field of information extraction. The different
components of IE and the approaches used to develop information extraction systems are
reviewed and presented. The related NLP fields to information extraction are also reviewed and
presented in order to see their similarity and difference with IE. Evaluation standards for the
performance of IE system which are used for the evaluation purpose are also presented in this
chapter.

2.1 Information Extraction

Now a days, digitally stored information is available in abundance and in a numerous forms to an
extent of making it near impossible to manually search, filter and choose which information one
should use for his/her own purpose. This abundance of information must instead be filtered and
extracted in order to avoid drowning in it [9]. Different scholars in the area tried to develop
different information management systems so that the drowning of summarized and relevant
information from an ocean of information can be facilitated and the right information for
decision making can be acquired. Among the different solution to the problem which solves it in
different ways are Information Retrieval (IR), Information Extraction (IE), Question and
Answering, Text Summarization, Text Categorization, etc.

Information Extraction (IE) addresses the problem of information overload by locating the target
phrases from document and transforms them in to structured representation. As it is defined by
MUC [10], IE is the “task of extraction of information from a text in the form of text string
which are placed in to slots labeled to indicate the kind of information that can fill them”. Given
a free text an IE system will extract the specific information from the text and put them in the
database so that they can easily be retrieved and managed [11].

Un Yong Nahm [12] also defines IE as “a form of shallow text understanding that locates
specific pieces of data in natural language documents, transforming unstructured text into a
structured database”. IE is all about extracting structural factual data mostly from unstructured
7
text (web pages, text documents, office documents, presentations, and so on). IE usually uses
NLP tools, lexical resources and semantic constraints for better efficiency. The General
Architecture for Text Engineering (GATE) [13], which is the widely known open source
software system for computations related to natural language, defines Information Extraction
(IE) as a system which analyses unstructured text in order to extract information about pre-
specified types of events, entities or relationships.

2.2 Related NLP Fields to Information Extraction

2.2.1 Information Retrieval

Information Retrieval (IR), which is a relatively matured field of study, provides a method that is
being widely used to acquire a set of relevant material from a huge amount of data collection
especially in unstructured form. It is defined as “finding material (usually documents) of an
unstructured nature (usually text) that satisfies an information need from within large collections
(usually stored on computers)” [3]. IR system interface takes a natural language text from the
user and convert it to equivalent query and retrieve the document from abundance of documents
which contain the data that the user wants [3]. The number of resulting pages returned by the IR
system could be very large and not all information contained in a single page is interesting,
which requires further refinement by the user [11]. Even if the information retrieval highly
minimizes the abundance of documents to small number of documents that are more appropriate
to the users need they are still large enough for the user to use them for decision making.

2.2.1 Text Summarization

Text summarization (or rather, automatic text summarization) is the technique where a computer
automatically creates an abstract, or summary, of one or more texts. The initial interest in
automatic text summarization started during the nineteen sixties in American research libraries.
A large amount of scientific papers and books were to be digitally stored and made searchable.
However, the storage capacity was very limited and full papers and books could not be fit into
databases those days. Therefore, summaries were stored, indexed and made searchable [9].
Radev [14] defines a summary as “a text that is produced from one or more texts, that convey
important information in the original text(s), and that is no longer than half of the original text(s)
and usually significantly less than that". The summary can be created from a single or multiple
8
documents by applying different methodologies. This simple definition captures three important
aspects that characterize research on automatic summarization
• Summaries may be produced from a single document or multiple documents,
• Summaries should preserve important information and
• Summaries should be short
Even if the text summarization minimize the large amount of text by extracting the most relevant
sentence from larger text it is still requires the user involvement to read the summary and extract
the specific information she/he needs and the data is not yet easily used by other computer
application directly.

2.2.3 Question and Answering System

Natural Language Question answering (NLQA) retrieves a textual fragment from a document
collection that represents the answer to a question. NLQA extract the specific information from
the given document which is retrieved by using IR system based on the given question by the
user. It might return one or a set of textual fragments from a document collection that represents
the answer to the question. The output of NLQA is unstructured information based on the given
question by the user [15]. NLQA is profoundly different from Information Retrieval (IR) or
Information Extraction (IE). IR systems locate relevant documents that relate to a query, but do
not specify exactly where the answers for the users request are or where the specific information
the user required is located.

Even if the question and answering system extract answers from a document based on the input
question from the user it still returns unstructured and small specific information from the
document. The information that the NLQA system returns can’t be managed by the computer as
it is more specific information to the question. Therefore the question answering system is very
helpful in extracting a specific answer for a specific question which reduces the time of the users
greatly but it is limited in extracting all the information the user wants as it is much bound to the
question formation in the language.

The different solutions that are listed above to solve the information overload issues have their
own advantage and disadvantage in solving the problem. IE unlike text summarization, IR and

9
question and answering system its output is in structured form which can be easily managed and
accessed as it is stored in the database. Different students at the postgraduate level try to develop
a question and answering, text summarization system for Amharic language which will help to
easily manage the information processing and use of the Amharic language in the computer
system. IE system will rather simplify the problem more by supporting the extraction of the
specific information the user wants from the given set of Amharic news document and fill the
predefined slots attributes and the output can be used by other application for different purpose
directly unlike NLQA, text summarization and information retrieval.

2.2.4 Text Categorization

Text categorization is a process involving the assignment of predefined categories to free
(unlabeled) text data. With the rapid growth of online data and information, text categorization
has become one of the key techniques for processing and organizing text documents. Text
categorization techniques are used to classify new stories, to find interesting information on the
World Wide Web and to guide a user's search through hypertext. The output of text
categorization is a category of the text [33].

2.3 Architecture of IE System

Different IE systems for different languages and different domains using different approaches are
developed so far and are still on development but they all use the different task breakdown for
IE. The research work in [38] categorizes IE in to five different tasks.

Named Entity Recognition

Co reference resolution

Template Element Construction

Template Relation Construction

Scenario Template Production

Named Entity Recognition (NER)

Named entities are one of the most often extracted types of tokens during extracting information
from documents. Named entity recognition is classification of every word in a document as
being a person-name, organization, location, date, time, monetary value, percentage, or “none of

10
the above”. Some approaches use a simple lookup in predefined lists of geographic locations,
company names, person names and name of animals and other things from the gazetteers, while
some others utilize trainable Hidden Markov Models to identify named entities and their type
[16].

For example the NER takes the following Amharic news text obtained from ENA and recognize
the named entities and numbers which will be used as attributes for the predefined database slot
ባሌ ዞን በመገባደድ ላይ ባለው የበጀት ዓመት 25 ሚሊዮን ብር በላይ ካፒታል ላስመዝግቡ ባለሃብቶች የኢንቨስተመንት
ፈቃድ መሰጠቱን የዞኑ ኢንቨስትመንት ጽህፈት ቤት አስታወቀ፡፡ የጽህፈት ቤቱ ተወካይ አቶ ሁሴን መሐመድ ዛሬ
እንደገለጹት በዓመቱ ፈቃዱ የተሰጠው በግብርናው ዘርፍ በፍራፍሬና በሰብል ልማት እንዲሁም በሆቴልና ቱሪዝም መስኮች
ለተሰማሩ 15 ባለሃብቶች ነው፡፡ ባለሃብቶቹ ከዞኑ አስተዳደር ከ1 ሺህ 150 ሄክታር በላይ ቦታ መረከባቸውን ገልጸው፣
ከመካከላቸው አስሩ አስፈላጊውን ዝግጅት አጠናቀው ወደ ሥራ መግባታቸውንም ተናግረዋል፡፡ እንደ ተወካዩ ማብራሪያ
እስካሁን ወደ ስራ የገቡት እነዚሁ ባለሃብቶች 1ሺህ ለሚሆኑ ወገኖች ቋሚና ጊዜአዊ የሥራ ዕድል ፈጥረዋል፡፡

The words which are names and numbers that represent different thing will be extracted like ባሌ,
25, ኢንቨስትመንት ጽህፈት ቤት, ሁሴን መሐመድ, 15, 1 ሺህ 150, 1ሺህ which are the named entity attributed
in the text which represent different things.

Co-Reference Resolution
It is a process of finding multiple references to the same object in a text. It refers to the task of
identifying noun phrases that refer to the same extra linguistic entity in a text. This is especially
important since the same thing about a single entity is expressed in different sentences using
pronouns [39].

Template Element Construction (TE)

It is one of the IE task which associate descriptive information for the NER results. The different
recognized name entities in association with their co reference are the input of template element
construction [37]. The different recognized named entities will have different attributes For
example, if we consider one of the entity organizations we might have the attribute like

Type: Which define the type of the organization whether it is government, commercial, non-
governmental, etc
11
Name: name of the organization
Aliases: another form of representation of the organization name

From the example in the NER, the template element construction assign attributes to the
recognized named entities the attributes might be represented as follows
Place: ባ ሌ
Amount of money for investment: 25 ሚሊዮን
Organization name: ኢን ቨ ስ ትመን ት ጽህ ፈት ቤት
Manager: ሁሴን መሐመድ
No of investor: 15
Land in hectare: 1 ሺህ 150
No of employees: 1ሺህ

Template Relation Construction (TR)

Template relation task is the identification of possible relationship between template elements
identified during the template element construction task. This might be, for example, an
employee relationship between a person and a company, a family relationship between two
persons, or a subsidiary relationship between two companies and other different relationship that
exist in natural language text. Extraction of relations among entities is a central feature of almost
any information extraction task, although the possibilities in real-world extraction tasks are
endless [38].

Scenario Template Production (ST)

It is a combined output of template element construction and template relation construction and
creates a scenario. Scenario templates (STs) are the prototypical outputs of IE systems, being the
original task for which the term was coined. They tie together TE entities and TR relations into
event descriptions. For example, TE may have identified Isabelle, Dominique and Françoise as
People entities present in the Robert edition of Napoleon’s love letters. ST might then identify
Facts such as that Isabelle moved to Paris in August 1802 from Lyon to be nearer to the little
chap, that Dominique then burnt down Isabelle’s apartment block and that Françoise ran off with
one of Gerard Depardieu’s ancestors [38].
12
In another work survey conducted on information extraction the tasks are categorized in to five
different and independent components [39]. As IE activity can be a very complex task
decomposing it into different task is advantageous. The advantages of decomposition are that:

It will be easy to choose the techniques and algorithms that suit each task,

It will be easy to locally debug an Information Extraction program since the module
responsible for each task is completely independent from the others and

Information Extraction can be a customized activity according to an application's needs,
by reordering, selecting and composing some of the tasks.
The considered tasks in the survey are:

Segmentation,

Classification,

Association,

Normalization and

Co-Reference Resolution.

Segmentation
The Segmentation task divides the text into atomic elements, called segments or tokens. Even
though this task is simplified for Western languages due to the existence of whitespaces
separating words, there are some cases in which simple whitespace separation may not be
enough. Usually, segmentation for these cases is performed using rules that show how to handle
each case. The major problems related to this task can be found in oriental languages. For
example, the Chinese doesn’t have whitespaces between words. For this reason, solving the
problems described above is not enough in this language. In these cases, it is typically necessary
to use external resources. Lexicons and grammars can also be used in order to accomplish the
task of segmentation using syntactic or lexical analysis. Another approach for segmentation of
Chinese texts uses techniques based on statistics [39].

Classifications
The Classification task determines the type of each segment obtained in the segmentation task. In
other words, it determines the classified output data structure where the inputs are segments. The
13
result of this task is the classification of a set of segments as entities, which are elements of a
given class potentially relevant for the extraction domain. The rule-based techniques used in the
classification task are usually based on linguistic resources, such as lexicons and grammars. One
of the most popular approaches to undertake classification is machine learning. Machine learning
techniques used in this task are usually supervised, which means that an annotated corpus is
needed [39].

Associations
The association task seeks to identify how the different entities found in the classification task
are related. The systems that perform extraction of relationships are less common than the ones
that perform the classification task. This happens due to the difficulty in achieving good results
in this task. Many techniques in the association task are based on rules. The simplest approach
uses patterns to extract a limited set of relationships. A more generic rule-based approach for
association is based on syntactic analysis. Often, the relationships that we want to extract are
grammatical relationships. For example, a verb may indicate a relationship between two entities.
The association task can also use machine learning techniques [39].

Normalization and Co-reference Resolution

Normalization and Co-reference resolution are the less generic tasks of the Information
Extraction process since they use heuristics and rules that are specific to the data domain. The
normalization task is required because some information types do not conform to a standard
format. This task is typically achieved through the use of conversion rules that produce a
standard format previously chosen. Co reference arises whenever the same real world entity is
referred in different ways in a text fragment. This problem may arise due to the use of
I. Different names describing the same entity (e.g., the entity Bill Gates" can be found
in the text as William Gates"),
II. Classification expressions (e.g., a few years ago, Bill Gates" was referred as the world's
richest man"),
III. Pronouns (e.g., in the sequence of sentences Bill Gates is the world's richest man. He was
a founder of Microsoft", the pronoun He" refers to Bill Gates").

14
Rule-based approaches for co reference usually take into account semantic information about
entities. A machine leaning approach for co reference resolution is based on clustering
algorithms for grouping similar entities [39].

The description of IE components presented in the above two works are almost the same even if
they are presented using different names. The idea is that as IE is a complex task dividing it in to
sub task will minimize the complexity of developing the system for IE. All the above listed
components of IE can be categorized in to Linguistic preprocessing, learning and application
phase and semantic post processing as the three main blocks regardless of the language, the
domain and the approach on which the IE is developed for as presented in [2]. Each of the three
blocks comprises tasks that are relevant for the selected information extraction approach. Even if
the extraction algorithm and procedure are different in different approaches the basic issue
behind all of them is to take text corpus and target structure as an input and produce structured
data as an output by filling the predefined slots in the target structure [2, 6, 7].

2.3.1 Preprocessing of Input Texts

Text corpus often consists of unstructured, “raw” natural language texts. A big part of the
relevant information can be distinguished by some regularity found in the linguistic properties of
texts. In this phase the language property i.e. its structure, the position where most relevant
information in the text are located, how the co-reference between sentences is presented in the
text and other language and domain specific information will be studied and implemented into
different linguistic components as part of the extraction system. The following linguistic
components are proved to be useful for information extraction as they are described in [2].

Tokenization: is the process of splitting the text into sentences and tokens. It Start with a
sequence of characters to identify the elementary parts of natural language such as words,
punctuation marks and separators. As a sentence is one of the most important components in the
natural language text for representation of interrelated information and for expressing a complete
thought or event. The resulting sequence of meaningful tokens is a base for further linguistic and
any text processing task [8]. Tokenization is similar to segmentation presented in [38].

15
POS: Certain facts are typically expressed by certain parts of speech like names and determining
parts of speech of tokens, known as POS tagging. Statistical systems can use POS tags as
classification features, knowledge based systems as elements of extraction rules [2, 15].

(Chunk) Parsing: While full sentence parsing is preferred by knowledge based systems, some
statistical approaches rely on chunk parsing, shallow syntactic analysis of the sentence fragments
performed on phrasal level. It is justified by the fact that the extracted information is often
completely included in a noun, verb or prepositional phrase that builds the most relevant context
for its recognition [2].

Co reference resolution and named entity recognition are also included as a preprocessing
component for IE.

2.3.2 Learning and Application of the Extraction Model

The learning and application phase gives much emphasis on the development of an IE system
which can be adopted to the different domain and language easily which is not yet successful as
IE is much bound to the language and the domain it is developed. Due to this reason the
adaptation effort to the other domain is high. Modern IE systems use a learning component to
reduce the dependence on specific domains and to decrease the amount of resources provided by
human. An extraction model is defined according to the pursued approach and its parameters are
optimized by a learning procedure. The three categories of approaches for IE use the learning
method. Statistical approaches learn relevant classification features, probabilities, and state
sequences, rule-based approaches learn a set of extraction rules and knowledge-based approaches
acquire structures to augment and interpret their knowledge for extraction [16, 17, 18].

Most of the IE systems use the supervised learning approach to train the extraction model used
about the domain specific information. The statistical approach uses annotated training corpus
which is divided in to two parts i.e. the training corpus and test corpus. The training corpus is
used to training the model about the different annotation in the text and the test corpus is used to
test the extraction model how much efficient it becomes after training [16].

16
2.3.3 Post Processing of Output
The main motivation for IE is the structured representation of information that enables formal
queries and automatic processing. One of the possibilities to structure the extracted data is to
model the target structure as a database relation. After the relevant information has been found
by applying the extraction model on the given text the identified text fragments are assigned to
the corresponding attributes of the target structure. They can be normalized according to the
expected format (e.g. representation of dates and numbers). Some identified facts may appear in
text more than once and there might be violation of primary key and other properties of the
database and all these things are handled at the post processing phase of IE [16].

2.4 Approaches for Information Extraction

The approaches for information extraction can be categorized under two primary categories
• Knowledge engineering approach and
• Machine learning approach
These two approaches for IE can be applied on the free text or semi structured or structured text
which is used as an input for IE system. Earlier when IE was introduced the main focus was to
extract the information from a short free text but latter its application is expanded to include the
semi structured and structured texts [19].

Free text: is unstructured collection of text. It can’t be easily managed as it doesn’t have the
structure or any predefined format in order to manage it by using computers. The natural
language components are applied in order to manage extraction from the free text [19, 20].

Semi Structured Text: is a data which is not in the form of tuples like structured text and is
different from free texts which rather exist in between the two. The information in the form of
HTML tags is semi structured text [19, 20].

Structured Text: is textual information which exists in a database or file following a predefined
and strict format. Such information can easily be extracted by using the format description as it
has a known format [19, 20].

17
2.4.1 Knowledge Engineering Approach
Knowledge engineering approach relies on knowledge engineers i.e. human experts, who are
skilled not only in the particular information extraction system, but also in the domain on which
the IE system is developed for. In this approach the person has to read a set of domain relevant
documents, find different rules by his observation and construct those rules by hand. In addition
to requiring skill and detailed knowledge on a particular IE system, the knowledge engineering
approach usually requires a lot of labor, a lengthy test-and-debug cycle, and it is dependent on
having linguistic resources at hand, such as appropriate lexicons, as well as someone with the
time, inclination, and ability to write rules. If any of these factors are missing, then the
knowledge engineering approach becomes problematic. Building a high performance system is
usually an iterative process whereby a set of rules is written, the system is run over a training
corpus of texts, and the output is examined to see where the rules under and over generate. The
knowledge engineer then makes appropriate modifications to the rules, and iterates the process.
The performance of the system is dependent on the competence of the knowledge engineer [19,
20, 21].

The advantage of knowledge engineering approach is good performing system is not hard to
build as long as the knowledge expert is there and best performing system can be achieved by
handcrafted rules. Its disadvantage is the development process is very laborious, some changes in
specification can be hard to accommodate and required expertise that may not be available [22].
Therefore developing an IE system will be very difficult as there are non experts that know IE
and the Amharic language and the NLP resources available in Amharic language are almost
none.

2.4.2 Machine Learning Approach

The Machine learning approach is quite different from knowledge engineering approach. In this
approach, it is not necessary to have someone on hand with detailed knowledge of how the IE
system works, or how to write rules. It is only necessary to have someone who knows enough
about the domain and the task to take a corpus of texts, and annotate the texts appropriately for
the IE system [11]. The machine learning approach for IE is then focused on the automatic
acquisition of the extraction patterns that will be used to extract the information relevant to the
particular task from each single document. When applying machine learning to IE, a learning
18
algorithm usually learns a model from a set of examples, grouped in documents, which have
been manually annotated by the user. Then the model can be used to extract information from
new documents. The accuracy of the learned model usually increases with the number of training
examples made available to the system.

The strengths and weaknesses of the machine learning approach are complementary to those of
the knowledge engineering approach. Rather than focusing on producing rules, the automatic
training approach focuses on the training data. Corpus probabilities or rules are then derived
automatically from the training data, and used to process novel or unseen data. As long as
someone familiar with the domain is available to annotate texts, systems can be customized to a
specific domain without intervention from any developers. The disadvantages of the automatic
training approach also revolve around the fact that it is based on training data. Training data may
be in short supply, or difficult and expensive to obtain. Sometimes one may wish to develop an
extraction system for a topic for which there are few relevant examples in a training corpus. Such
situations place a premium on the human intuition of a good rule designer. If the relations that
are sought are complex or technical, if it is hard to find annotators, and if it is difficult to produce
enough annotated data for training corpus then using machine learning approach will be
disadvantageous [21].

Current IE approaches supported on supervised machine learning technique are divided in to the
following there categories [23]
I Rule learning
II Linear separators
III Statistical learning
I Rule Learning
This approach is based on a symbolic inductive learning process. The extraction patterns
represent the training examples in terms of attributes and relations between textual elements.
Some IE systems use propositional learning (i.e. zero order logic), for instance, AutoSlog-TS and
CRYSTAL, while others perform a relational learning (i.e. first order logic), for instance
WHISK and SRV. This approach has been used to learn from structured, semi-structured and
free-text documents. Our method is related to the SRV system in that it models the IE task as a
19
classification problem. However, it applies Inductive Logic Programming and uses information
about negative examples [23].

II Linear Separators
In this approach the classifiers are learned as sparse networks of linear functions (i.e. linear
separators of positive and negative examples). It has been commonly used to extract information
from semi-structured documents. It has been applied in problems such as extraction of data from
job ads, and detection of an e-mail address change [23].

In general, the IE systems based on this approach present an architecture supported on the
hypothesis that looking at the words combinations around the interesting information is enough
to learn the required extraction patterns. Their main advantage is that a deep linguistic analysis is
not necessary instead classification techniques are used to find the desired information [23].

The method used in this research work is similar to all these systems. It is based on the same
hypothesis. Looking around the words combination around the fact need to be extracted from the
Amharic text is enough to know what type of information it is.

III Statistical Learning

This approach is focused on learning Hidden Markov Models (HMMs) as useful knowledge to
extract relevant fragments from documents [23].

These IE systems also differ from each other in the features that they use. Some use only basic
features such as token string, capitalization, and token type (word, number, etc.). In addition,
others use linguistic features such as part-of-speech, semantic information from gazetteer lists,
and the outputs of other IE systems (most frequently general purpose named entity recognizers).
A few systems also exploit genre-specific information such as document structure. In general,
the more features the system used, the better performance it could achieve.

One of the most successful machine learning methods for IE is Support Vector Machine (SVM),
which is a general supervised machine learning algorithm. It has achieved state-of-the-art
20
performance on many classification tasks, including named entity recognition. For instance, [19]
compares three commonly used methods for named entity recognition SVM with quadratic
kernel, maximal entropy, and a rule based learning system, and shows that the SVM-based
system outperforms the other two. More features usually result in better performance and
therefore, it is important to use the different features on the while using SVM algorithm to get a
good performance.

2.5 Information Extraction as a Text categorization Task

IE is a token classification task rather than a Text categorization task. With IE we are working
with texts but the basic unit that we are seeking to classify is tokens in the text rather than the
entire text. With Text categorization we are seeking to identify whether an entire text is a
member of particular category. With IE the categories are start and end, and the objects we seek
to assign to these categories are the individual tokens. With IE we are representing individual
tokens. We much encode additional information about the token to enable our learning algorithm
to generalize. For IE we encode several features of the token as well as relational information
about the surrounding tokens. The features include the specific token, as well as part-of-speech
(POS), chunking, orthographic and gazetteer information. In addition, we add features to
represent a fixed-width window of tokens on either side of the instance’s token. The learning
algorithm uses these features to create a model that can distinguish between tokens that are starts
of fields, ends of fields or neither [24].

Let us first consider a two class classification problem. Let {(x1, y1)… (xn, yn)} be a training
data set, in which xi denotes an instance (a feature vector) and yi ∈{−1,+1} denotes a
classification label. A classification model usually consists of two stages: learning and
prediction. In learning, one attempt to find a model from the labeled data that can separate the
training data, while in prediction the learned model is used to identify whether an unlabeled
instance should be classified as +1 or -1. (In some cases, the prediction results may be numeric
values, e.g. ranging from 0 to 1. Then an instance can be classified using some rules, e.g.
classified as +1 when the prediction value is larger than 0.5.) [25].

21
2.6 Evaluation Metrics
Since the classifier algorithms are employed as an algorithm applied for text categorization and
information extraction the standard evaluation techniques used for evaluating classifier algorithm
are presented. The accuracy of a classifier is calculated by the classifiers’ performance on the test
data set. The prediction of the category by the classifier can be evaluated using a confusion
matrix, which represents how well a classifier recognizes instances of different classes. For
example, if we have Politics news and economy news as two classes for prediction and if there
are 100 news instances for each category of the following confusion matrix table, The first row
of the confusion matrix table indicates that among the 100 instances of Politics news texts the
classifier correctly predict the category of 90 news instances correctly while 10 of the remaining
news instances are incorrectly categorized as economy news. The same works for the second
row which the classifier predicts the category of 96 economy news correctly while 4 of them are
categorized incorrectly as politics news.

Table 2.1: Two Class Confusion Matrix Predicted Class

Politics news Economy news classes

90 10 Politics news
4 96 Economy news

In addition to accuracy the experimental dataset may also require other performance measures to
analyze the experimental result in detail like sensitivity, specificity, recall, precision, F-measure,
and ROC Area. Sensitivity is also referred to as the true positive rate, while specificity is the true
negative rate (that is, the proportion of negative instances that are correctly identified).

Sensitivity or true positive rate is the proportion of positive instances that are correctly identified
by the classifier and it is calculated as
_

 = (2.1)

Where t_pos is true positives and pos is the number of positive documents

22
Specificity is the proportion of false positive instances that are correctly wrongly classified as

another class by the classifier

_

 = (2.2)


Where f_pos is false positive and pos is the number of positive documents

The recall, precision and ROC area are then calculated from the result of sensitivity or True

positive rate and specificity or false positive rate.

Precision is the proportion of instances that are correctly classified which a true positive instance


 = (2.3)
 + 

Where t-pos is the number of true positives and f-pos is the number of false positives

Recall is the proportion of instances which are classified correctly over the total number of

instances in the test dataset [44]


 = (2.4)
  !  

Where t-pos is number of true positive instances

F-Measure is an optimization criterion which often is used for tuning the threshold in binary

decision, which is defined in terms of precision and recall [44].

(β0 + 1)( X )

#β (precision, recall ) = … … … … … … 2.5
β0 (precision) + recall

Where β is the parameter allowing differential weighting of precision and recall

If recall and precision are given equal weight β=1, Equation (2.5) will be simplified to

23
 2( X )
#−  = … … … … … … … … … … … … … … 2.6
( +  )

ROC (Receiver Operating Characteristics) curves show the trade-off between the true positive

rate (sensitivity) and the false positive rate for a given model. The area under the ROC curve is a

measure of the accuracy of the model. If there a counter false positive for every true positive, the

ROC curve will be diagonal line with a ROC value 0.5. As the number of true positive instances

of the class increases and the false positive instances decrease the ROC value increase from 0.5

to 1. Is the ROC value is 1 the classifier is perfect with the all the instances are true positive [43].

24
 CHAPTER THREE

3. RELATED WORKS

The extraction of information from a free text for constructing databases is a recent technology.
Scholars have conducted IE researches using different approaches on different domains and
languages. Among these research works, some of the works that are more relevant to our
research, which are done using a machine learning approach are reviewed and presented in this
chapter.

3.1 Information Extraction from English Text

Most of the research works in the area of information extraction are conducted in different
domain on the English language. Among these research works the research work in [26] presents
a text categorization approach to extract information from business cards and automatically
process change of addresses from email messages. They scan 505 business cards to create a text
file for training and testing the system which is then annotated manually. After scanning some
cleaning operations are done to make sure that the text file is ready for being processed by the
text classifier to extract information. They use the Naïve Bayes algorithm to categorize the
different line of the text into Address, Name, Title, Company name, company Logo, Phone
number (either voice or fax), Web address, Email address, Telex number, cable number, and
other miscellaneous text from the scanned text. They claim that Naïve Bayes algorithm is
reasonably accurate but it ignores important structural constraints like how the business card is
arranged what attribute come after the other and so on. Such constraints are handled by HMM in
their work.

Another work presented in [27] focuses on using double classification approach to build a text
classifier that can extract information from the software job corpus. They review the different
proposed novel techniques that were used before for IE. They argued that IE is domain
dependent because it is based on an extensive syntactic and semantic analysis of the language
and the documents under consideration. However, this way the resulting IE system are less
applicable in domain where no semantic information is available or for corpora in other

25
languages than English, for which fewer syntactic processing tools exist. By considering these
issues in the IE research they propose a solution that can solve the problem which is an intuitive
approach consisting of two classification based machine learning loops. In the first loop they use
a text classifier to select the sentence that might contain relevant information from the document.
In the second loop they perform a deeper analysis of those relevant sentences by performing
word level classification. They use a Naïve Bayes classification for the first loop and rule based
classification for the second loop. They use POS tags for syntactic information to build rules for
the second loop. No semantic information is used. Their approaches is tested on software job
corpus and obtained good results on most of the template slot, although for some slots (language
and area) the word classification approach is not yet capable of finding good rules for the others
they got 77% recall and 65% precision.

Another work presented in [28] uses the SVM and different feature sets to build a text classifier
for information extraction. They review all the approaches and different algorithms that are used
so far in different IE researches. They use the approach that treats the identification of fragment
text to extract start and end position as distinct token classification task. The instances are all the
tokens in the document. All the tokens that begin a labeled field are positive instance for the start
classifier, while all the other tokens become negative instances for this classifier. Similarly, the
positive examples for the end classifier are the last tokens for each labeled field and the other
instances are negative examples. Each instance has a set of features that describe the given token.
The features used in the paper include the specific token as well as POS, chunking orthographic
and gazetteer information. They present the ELIE algorithm which has two distinct phases. In the
first phase ELIE simply learns to detect the start and end of the fragment to be extracted. Their
experiment shows that the first phase generally has high precision but low recall. The second
phase is designed to increase the recall. They find that very often false negatives are “almost”
extracted (the start but not the end is correctly identified, or the end but not the start). In the
second phase, which they call convergent boundary classification; ELIE is trained to detect either
the end of a fragment given its beginning, or the beginning of a fragment given its end. In this
way, ELIE iteratively converges to the correct boundary classification that is missed in the first
phase. They evaluate their approach on the seminar announcement dataset and the job posting
dataset.
26
Another work presented in [29] uses a hybrid approach to information extraction. The authors
present a hybrid knowledge based and statistical machine learning approach to extract entities
and relations at the sentence level. They present a hybrid entities and relation extraction system,
which combines the power of knowledge based and statistical machine learning approach. In this
approach, the rules for extraction are written manually, while the probabilities of the extracted
texts being part of the database slot are trained from an annotated corpus.

This approach allows the knowledge engineer to write very simple and naïve rules, while
retaining their power thus greatly reducing the required labor. In addition the size of the training
data is considerably smaller than the size of the training data needed for pure machine learning
system (for achieving comparable accuracy results). The Authors use DIAL, which is based on a
general purpose rule language for developing knowledge for extraction and statistical HMM for
machine learning approach which is used to train the system.

Effectiveness of their approach is tested by using three different corpora MUC-7, ACE -2 and an
industry corpus. On the MUC-7 corpus their hybrid approach called TEG outperforms pure
HMM model and DIAL rule based system for named entity recognition. On the ACE -2 corpus
they tested the relationship and in this case as well the hybrid approach they present performs
better than HMM and Markovian SCFG. They conclude that a small hand crafted rules when
combined with machine learning method will increase the performance of machine learning.

3.2 Information Extraction from Portuguese Text

A research work presented in [30] uses rule learning approach which applies rule learning
without ignoring stop words for learning text extraction rules from Portuguese text. The main
aim of the research work presented was to consider a collection of Portuguese text documents in
certain domain, annotate the text by identifying the element of interest from the text, and use the
annotated document to generate training set which in turn is used to generate extraction rules.
These rules can then be applied to new unseen documents.

27
The system presented in the paper use three phases for extraction of information
Phase 1: Data preparation – which is concerned with selecting the set of documents that are of
interest for a particular IE task and identify relevant elements from those relevant documents.
Phase 2: Generating text extraction rules – used to generate extraction rules which the authors
use the enhanced naïve Bayes method.
Phase 3: processing new documents- this phase is concerned with extracting relevant information
from new documents by using the extraction rules obtained during training of the system.

The stop words like conjunction (e.g. “and”, “or” etc), prepositions (with, without, from etc) or
other similar categories are considered quite irrelevant have no effect on the final results.
Therefore these stop words are removed during preprocessing in earlier works. Authors of this
paper proved that stop words will have a contribution to improve the result and they proved this
assertion by their experiment. Their approach has an F-measure of 80% after it is tested on
Portuguese language on domain concerning House/Flat sales.

3.3 Information Extraction from Thai Text

The research work presented in [31] uses the knowledge engineering and the machine learning
approach to extract information from Thai text. This work is different from the IE systems
developed for English language as it has its own unique features because of the nature of the
language structure in Thai text. Thai language doesn’t have capital letters, words themselves are
a combination of compound words, and there is no sentence demarcation and other language
specific ambiguities in the structure of the text. This language ambiguity which is most common
for Asian languages as the authors claim is difficult to apply the pattern or rule learning approach
directly.

To solve the ambiguity in the language structure of Thai texts they use the knowledge
engineering approach to solve the ambiguities and the machine learning approach to train the
system for new things that is not in the knowledge. The system they design has the following
different components

28
Word Tokenization: As the words used in Thai text are a combination of other words, the first
step is to segment these words in to individual words. In this phase the Head Words Dictionary
and rules for specific lexicon are used by the authors to improve the tokenization process.

Part of Speech Tagging: the next step is to tag the different part of the segmented text to its
appropriate part of speech tag. For this phase also they use lexicon dictionary and training corpus
to effectively tag the words in the text.

Surface Structure Analyses: Since there is no clear indication of sentence boundaries, it is not
possible for the system to process the document using a sentence-based approach. In Thai text it
is usually possible to identify paragraph breaks, taking a paragraph as one or more sentences
separated by empty line. In Thai, a space may be used as a separation between contents, but it
can also be used for other purposes, making the identification of contents unreliable and
complicating the syntactic analysis. In this step the regions of information are identified using
the spaces that separate between contexts. However, as there are fraudulent spaces, for instance
the spaces between numbers, labels and abbreviations, to prevent fault identification they remove
these spaces. To assist this phase they use Thai context free structure Grammar.

Extraction Engine: use a predefined concept definition and semantic class dictionary and the
output of the above three phases as an input and extract the specific information from the text.

The authors tested their approach on the Thai import/export domain derived from ministry of
commerce. The result of the experiment is a 42% precision and over 70% recall. The reason for
the precision percentage to be low comes from the ambiguity of the sentence structure in which
the parser is unable to determine the boundary of the sentences. They concluded that with
efficient grammar parser for Thai and a word sense disambiguation module the precision of their
system can be improved. However, the major problem with the system in its current form is that
it requires substantial knowledge acquisition.

29
3.4 Information Extraction from Chinese Text
A work in [32] presents an IE from Chinese free text which has similar nature with Thai text as
most of the Asian languages have similarities. The authors present an approach which combines
Automatic learning algorithm of pattern rule and employment of heuristic information for
Chinese free text. The authors present the different tasks they use to extract information from
Chinese free text.

Input Document Preprocessing: this phase contains different subtasks to make the Chinese free
text ready for the next phase. At first the input document is broken down into sentences. Then
the sentences are segmented into words by looking up the dictionary because Chinese sentence is
composed of characters without any natural delimiters such as space between words. After the
sentence segmentation the named entity recognition is done to identify the place, person, and
organization names.

Syntactic Analysis: uncategorized words during preprocessing are assigned in to the likely
category based on the HMM model probability approach.

VP and NP Recognition: verb and noun signify the important meaning of sentence at most part.
This process module uses syntactic patterns to identify small syntactic units through rule
primarily, such as basic noun groups (NG e.g.. position group, organization group…), which are
nouns with their left modifiers, and verb chains or verb groups (VG), which consist of a head
verb preceded by modals or adverbials.

Pronominal Anaphora Resolutions: It is very difficult to understand the extracted contents in

some instances that pronouns are contained without their antecedents. So, it is important to
resolve pronominal anaphora. Pronominal anaphora resolution is to track references to a frame
topic across sentences. For each finding, the probability is estimated that it co-refers to each
previously mentioned finding based on semantic features and dictionary cues. Their pronominal
anaphora resolution work uses rule and statistical methods.

30
Use of Heuristic Information: Also extraction pattern can play an important role in IE system,
some instances test are not accorded with existing pattern .In order to solve this problem, more
complicated pattern is needed to face to complex situations. This will bring out more difficult to
system and make conflict in future. They employ heuristic information to reduce this complexity.
The heuristic information is used to exploit the context structure of the source. It will be simplify
the structure of the text so that the pattern rules can be used to extract the information.

Then the extraction rules are used to extract the text fragment and fill the template slot after the
above different processing is done on the Chinese free text.

Their approach is tested on 50 articles they get from 1 China import official alteration to extract
information that will fill the following four slots person name, organization, old position and new
position. They test the system by applying two methods; method 1 is based on pattern matching
without heuristic information and method 2 is based on pattern matching with heuristic
information. And the result of the experiment was the following

Slot & result Method 1 Method 2

Person name Recall 64.1% 78.8%
precision 89.2% 87.2%
Organization Recall 62.3% 76.5%
precision 92.1% 89.3%
Old position Recall 64.5% 77.4%
Precision 86.3% 84.6%
New position Recall 68.3% 83.3%
precision 84.5% 81.3%

The authors conclude that the use of heuristic information in addition to pattern learning will
increase the efficiency of IE specially the Recall.

1
ttp://news.cina.com.cn/special/gov/index.shtml
31
3.5 Information Extraction from Spanish Text
Another research work presented in [23] uses machine learning approach for extracting
information from natural disaster news on Spanish news papers. The authors present a machine
learning approach in order to extract the information about characteristics and effects of natural
disasters that will be used as a relevant resource for populating determined database. They select
the machine learning approach as it is more appropriate to use when compared with knowledge
engineering as they require a human expert in generating extraction patterns while machine
learning is flexible by itself as it learn extraction patterns and as it can be easily trained to add
modification.

The authors present a linear separator approach for learning patterns in the sentence of the text.
This approach is based on the hypothesis that looking at the word combination around the
relevant data is enough for learning the required extraction patterns. This creates an advantage as
it doesn’t require deep linguistic analysis of text and can be used for a more generalized purpose.

The proposed system presented in the paper is used to extract the following information from the
disaster news. Information related to disaster itself such as Date, Place, magnitude of the
phenomenon Information related to the people for instance, the number of dead, wounded,
missing, damaged and affected person. Information related to the infrastructures, the number of
destroyed and affected houses Information related to economic impact, such as the number of
affected hectares, monetary lost among others

The proposed system architecture uses two main modules: text filtering and fact extraction. The
first module concentrates on selecting news report about natural disasters and the second on
extraction of information from the texts. Both the modules apply a supervised classification
approach, which is a machine learning technique that generates functions from training data that
allows mapping input object to a set of predefined categories.

Their approach was tested on Spanish news report about natural disasters. The corpus were
collected from several online Mexican news papers from 1996 to 2004 and get an F measure of

32
98 % in the detection of documents about natural disasters and 76% in extraction of relevant data
from those documents.

Most of the related work we review and the theoretical literatures in IE show that the use of
knowledge engineering approach by mixing it with the automatic training approach is more
efficient than the pure automatic training approach. The simple addition of basic rules to the
automatic training approach improves the performance of the IE system highly when compared
with pure automatic training approach.

3.6 Information Extraction and Related Works on Amharic Text

Different graduate students of Addis Ababa University and College of Telecommunication and
Information Technology (CTIT) have conducted research on NLP in different areas. Among
these research works that are more related to our work are reviewed and presented as follows.

3.6.1 Text Categorization from Amharic Text

Among the different categorization researches that are done on Amharic text, the research work
presented in [33] uses a machine learning approach to categorize the Amharic news text in to 15
different categories. The Ethiopian News Agency (ENA) news data and the manually given
category to the news are used as a corpus. The Weka package that has different machine learning
algorithm for classifier learning is employed which are SVM and decision making and has a
good result in classifying the news text. The different preprocessing tools are employed on the
Amharic text before it is given to the Weka classifier learning algorithms to learn the classifier
model. The tools employed are the following

Word Identification: The Amharic word separator which are single space,” netela serez”, “hulet
neteb”, “dereb serez”, “arat neteb” also written in Amharic as ፣, ፡ , ።, and ፤ respectively carriage
return, line feed, tab etc are used to identify words from the Amharic text. The hyphen between
words was also deleted to merge them in to a single word like ክፍለ-ከተማ into ክፍለከተማ. The
numbers weren’t considered in their work as it doesn’t differentiate one document from the
other.

33
Stop Word Removal: From the identified words they remove stop words (ነ ው, እ ና , and ነ በ ር )
as they doesn’t differentiate one document from the other as stop words exist in almost all
documents.

Stemming: varieties of words created by language requirement affixes could result in feature
words redundancy which in turn reduces the efficiency and accuracy of the text classifier. To
solve this problem stemmer is used to removes common suffix and prefix to change the word
variant into one common form, which will increase the efficiency of text classifier.

Controlling Spelling Variation: Observation of the Amharic news words that have been spelt in
different ways seems to indicate that the cause of the spelling variations is the difference in the
Amharic pronunciation of the words. Most of these words are foreign words adapted to Amharic.
For example, the word exhibition adapted from the English language has been spelt in the
following four different ways in ENA’s news items: ኤግዚቢሽን , ኤግዚብሽን
ኤግ ዚብሽን , ኢግዚቢሽን ,
ኢግዚብሽን . The spelling variations of words were controlled to have one form of representing
the same word since it is represented using different form due to its pronunciation.

Identifying Compound Words: The compound words in Amharic sometimes written as a single
word and sometimes as two separate words. This non-uniform way of writing could result in
semantic loss during document representation - thereby reducing the accuracy of document
representation. They identify the compound words and represent them as a single word instead of
two separate words.

Feature Word Selection with Their Frequency: the feature word, its frequency in the
documents, the number of documents it exist, weight given to the word is all stored in the
database as weka the machine learning package they use for implementation takes ARFF format
only. The stored data is then converted in to ARFF format to train the classifier and it is later
used to test the classifier model.

Their system is tested on Amharic news text from ENA using the Decision tree and SVM
algorithms that exist in Weka package. The experiment is conducted by using decision tree and
SVM algorithms among the different algorithms that Weka support. The SVM performs better
than the decision tree. The following is the summarized experimental result while using SVM

34
The average result of the classifier is shown as follows

Average accuracy for Average Accuracy for Average Accuracy for

Classifier
5-category data 10-category data 15-category data

LibSVM 95.21% 91.36% 81.15%

LMT 93.45% 89.98% 79.72%

This shows that the SVM, among the different classifier algorithms that Weka support is the best
algorithms in correctly predicting the category for the different news texts used as dataset for
training and testing the system.

3.6.2 Information Extraction from Amharic Text

Regarding the research on information extraction there is one research work presented [34],
which uses the hidden Markov model to extract information from the Amharic text. The IE tool
is developed for extracting information from a single sentence. It uses the slots subject, object,
action and reporter and tries to extract information from the news text which contain the above
listed four slots on a single sentence. If one of these components doesn’t exist in the sentence
the system doesn’t extract the information. The extracted information doesn’t have that much
usage as the facts with describe the different thing about entities are not part of the research
work. In this research work not much emphasis is given in studying the language as it is the
important building block of the IE tool.

Summary
The different related works that we reviewed so far uses machine learning or a hybrid approach
which is machine learning and knowledge based. It is known in the area of natural language
processing researches that Knowledge based approach is good to gain a good performance for
different NLP based systems. However, it is very difficult to build knowledge based system as it
requires much more time and knowledge expert in the domain of development. It is also not
easily modifiable to use it for other languages and domains and to adding new modifications is
difficult as it requires knowledge expert. To alleviate these problems different machine learning
algorithms are used and better results are acquired that can be compared to knowledge based

35
system. Text classifier approach to IE also shows a good result in most of the IE researches and
it uses off shelf machine learning algorithms which make it easy to develop IE system for
language like Amharic which has limited NLP resources. Therefore, the text categorization
approach with induced features is used in this work as there are no well developed standard NLP
resources for Amharic.

36
 CHAPTER FOUR

4. THE AMHARIC LANGUAGE

This chapter discusses the different issues about the Amharic language that is needed in the
development of an IE system and a sub text categorization system for Amharic news text. It
begins by introducing the Amharic writing system. The numerals and punctuation marks in
Amharic language are also described.

4.1 The Amharic Language

Amharic is a Semitic language spoken in many parts of Ethiopia, a country of 73.92 million
people by the as reported in the 2007 census from Ethiopia central statistics agency [45]. It is the
official working language of the Federal Democratic Republic of Ethiopia and thus has official
status nationwide. It is also the official or working language of several of the states/regions
within the federal system, including Amhara and the multi-ethnic Southern Nations, Nationalities
and Peoples region. Outside Ethiopia, Amharic is the language of millions of emigrants (notably
in Egypt, US, Israel, and Sweden), and is spoken in Eritrea. It is written using a writing system
called fidel, adapted from the one used by Ge'ez language [35].

4.2 Grammatical Structure of Amharic

Word Categorization in Amharic

The words in Amharic are categorized under five basic categories by Baye Yimam [36] that uses
the morphology and position of the word in Amharic sentence as criteria. These five categories
are ስ ም (noun), ግ ስ (verb), ቅ ፅ ል (adjective), ተውሳ ከ ግ ስ (Adverb) and መስ ተዋድድ (preposition) [36].

Noun: a word will be categorized as a noun, if it can be pluralized by adding the suffix ኦ ች / ዎች
(“owch”) and used as nominating something like person and animal. It is used as a subject in a
sentence. Pronouns, which were considered as independent category in the previous works by the
linguistics professionals is categorized under nouns after considering the unique nature of the
language as the earlier linguists just adopt the English language structure for Amharic lanaguage.

37
The following are some of the pronouns in Amharic ይህ , ያ , እ ሱ, እ ስ ዋ , እ ኔ , አ ን ተ , አ ን ች …;
quantitative specifiers, which includes አ ን ድ, አ ን ዳ ን ድ, ብዙ, ጥቂት , በ ጣም…; and possession specifiers
such as የ እ ኔ , የ አ ን ተ , የ እ ሱ.

Verb: any word which can be placed at the end of a sentence and which can accept suffixes as
/ህ /,/ሁ/,/ሽ /, etc. which is used to indicate masculine, feminine, and plurality is classified as a verb.
For example in “አ በ በ አ ን በ ሳ ገ ደ ለ ” “ገ ደ ለ ” is a verb since it appears at the end of the sentence.

Adjective: is a word that comes before a noun and add some kind of qualification to the noun.
But every word that comes before a noun is not an adjective. For it to be an adjective it should
also satisfy the condition when the word “በ ጣም” is added to it, it should be meaningful. For
example “ትልቅ በ ግ ” in this example “ትልቅ ” is an adjective to check it really is an adjective adding
the word “በ ጣም” before the adjective if it is meaningful it is an adjective if not is isn’t an
adjective. In this case it is meaningful and “ትልቅ ” is an adjective.

Adverb: a word that qualifies the verb by adding extra idea from time, place and situations point
of view. The following are adverbs in Amharic ትና ን ት , ገ ና , ዛ ሬ , ቶሎ, ምን ኛ , ክ ፉኛ , እ ን ደ ገ ና , ጅልኛ and
ግ ምኛ .

Preposition: a word that doesn’t take any kind of suffix and prefix, that can’t be used to create
other words and which doesn’t have meaning by itself but can represent different adverbial roles
when used with nouns. The different propositions include ከ ፣ ለ ፣ ወደ ፣ ስ ለ ፣ እ ን ደ …፣ ወዘተ

4.3 Amharic Punctuation Marks and Numerals

In Amharic, there are different punctuation marks used for different purposes [1]. In the old
scripture, a colon (two dots ፡ ) has been used to separate two words. These days the two dots are
replaced with whitespace. An end of a statement is marked with four dots (አ ራት ነ ጥብ ። ) while
ነ ጠላ ሰ ረ ዝ (፣ or ፥ ) is used to separate lists or ideas just like the comma in English. The
punctuation marks which are used for sentence demarcation and those which are used for
separating similar items are important to this research.

38
In Amharic, numbers can be represented using either the symbols of Arabic number system or
the symbols of the Ethiopic number system or using words and symbols of the Arabic number
system. Table 4.1 shows the Arabic, Amharic and alphanumeric representation of numbers.

Table 4.1: Number Representations in Amharic

Arabic Ethiopic Alphanumeric Arabic Ethiopic Alphanumeric

1 ፩ አንድ 20 ፳ ሃያ

2 ፪ ሁለ ት 30 ፴ ሰላሳ

3 ፫ ሦስ ት 40 ፵ አርባ

4 ፬ አ ራት 50 ፶ አ ምሳ /ሀ ምሳ

5 ፭ አ ምስ ት 60 ፷ ስ ልሳ /ስ ድሳ

6 ፮ ስ ድስ ት 70 ፸ ሰባ

7 ፯ ሰባት 80 ፹ ሰ ማኒ ያ

8 ፰ ስ ምን ት 90 ፺ ዘ ጠና

9 ፱ ዘ ጠኝ 100 ፻ መቶ

10 ፲ አስር 1000 ፼ ሺ/ሽ ህ

In Amharic, fractions and ordinals have their own way of representation [1]. Table 4.2 shows
fraction and ordinal representations in Amharic. As numbers are one of the information that is
extracted in this research work its representation in letters in Amharic text is important and it is
presented in the following table.

39
 Table 4.2: Amharic Fraction and Ordinal Representation
Fraction Amharic representation Ordinals representation

½ ግ ማሽ 1st አ ን ደ ኛ /ቀ ዳ ማዊ

1/3 ሲሶ 2nd ሁለ ተኛ /ዳ ግ ማዊ

¼ ሩብ/እ ር ቦ 3rd ሦስ ተኛ /ሳ ልስ

2/3 ሁለ ት ሲሶ / ሁለ ት ሦስ ተኛ 4th አ ራተኛ /ራብዕ

¾ ሶ ስ ት -አ ራተኛ . .

1/10 አ ስ ራት . .

2X እ ጥፍ 9th ዘ ጠን ኛ /ዘ ጠነ ኛ

2.X ሁለ ት ነ ጥብ 10th አስረኛ

Dates in Amharic can be written in different ways. It can be written using symbols in Arabic
number system like 12/01/2001 or using Ethiopic numeral representation and alphanumeric
representations like ጥቅምት 10/2003. Knowing the Amharic fraction and its representation in
Amharic language is important as they are mostly used in the Amharic text in the form of words
instead of digits.

4.4 Characteristics of Amharic writing

The characteristics of the Amharic writing system considered in this section are limited to those
that are common to news texts and IE systems.

Character Redundancy: Amharic took the whole Geez alphabet and uses it in the Amharic
writing system. It then added some more symbols for some other sounds that it has and that
could not be represented by the symbols of the Geez alphabet. This unsystematic borrowing from
Geez has resulted in redundant characters in the Amharic FIDEL [33]. Table 4.3 shows an
example of the character redundancy where more than one symbol is used for same sound.

40
 Table 4.3 Amharic Characters with Same Sound

Consonants Other symbols with the same sound

ሀ (hä) ሃ ሐሓኃ ኅ ኻ

ሰ (sä) ሠ

አ (ä) ኣ አ and ዓ

ጸ (tsä) ፀ

Spelling variations of a word would unnecessarily increase the number of words representing a
document which could reduce the efficiency and accuracy of the classifiers for sub text
categorization as well as for the classifier used for the IE system. Amharic document processing
for feature selection should therefore normalize word variants (spelling differences) caused by
inconsistent usage of redundant characters. During the pre-processing stage of Amharic
documents in this work, the different forms of a character that have the same sound are changed
to one common form. Table 4.4 shows examples of the different word spellings caused by the
redundant characters.

Table 4.4 Word Spelling Variations

The Word in English The word in Amharic Spelling Variants of the Word

Work ሥራ ስራ

Sun ፀ ሐይ ጸ ሐይ፣ ፀ ሀ ይ፣ ጸ ሀ ይ

World አ ለም አ ለ ም፣ ዓ ለ ም፣ ኣ ለ ም

Power ሀ ይል ሃ ይል፣ ኃ ይል

Compound Words in Amharic Language

In the Amharic writing system, inconsistency is often observed regarding the representation of
compound words. Some compound words are used as a single word in some instances (either by
fusing the two words or by inserting a hyphen between them) and as two separate words at other
instances. Inconsistent usage of compound words could result in redundant word features by
41
creating more words. when a compound word (example አዲስ አበባ) is treated as two separate
words አዲስ and አበባ.

Variations due to Pronunciations: usage of foreign language words in Amharic is also found to
be another source of word spelling variations. Observations of the ENA’s news documents
shows that in most cases in the writings of words adapted from foreign languages different
writers use different spellings. The cause of the difference in the Amharic spellings of these
foreign language words seems to be the difference in the pronunciations of these words. For
example, the word ሚቲዎሮሎጂ (Meteorology) have different Amharic spellings. Table 4.5 shows
examples of spelling variation in the writing of foreign words in Amharic.

Table 4.5 Word Variations due to Pronunciation

Foreign Word Equivalent Words in Amharic usage

Meteorology ሚትሪ ዮሎጂ፣ ሚትዎሮሎጂ፣ ሚትዮሮሎጂ፣ ሚቲዎሮሎጂ፣ ሚቲዎሮሎጅ፣ ሚቲዎሮሊጂ፣ ሜትዎሮሎጂ፣

ሜትሮዎሎጂ፣ ሜትሮዎሎጅ፣ ሜትሪ ዎሎጂ፣ ሜትሮሎጂ፣ ሜትር ዎሎጂ፣ ሜቲዎሮሎጂ፣ ሜቴዎሮሎጂ
Million ሚሊዮን ፣ ሚልዮን ፣ ሚሊየ ን
Television ቴሌቪዢን ፣ ቴሌቪዥን ፣ ቴሌቪዝን ፣ ቴሌቪዝን

Moreover there are word spelling variations that could be attributed to variations in
pronunciations at different parts of the country, like for example using the two words ጠባ ይ and
ፀ ባ ይ to mean temperament or using the three words ጢን ዚዛ , ጢን ዝዛ and ጥን ዚዛ to mean beetle [33].

Other Cases of Word Variations: Difference in word affixing has also been observed to cause
word spelling variations. For example difference in suffixing would result in the two writings
አእምሮአዊ and አእምሯዊ to refer to human intellect while difference in prefixing would give the two
writings ለአንድ and ላንድ to mean ‘for one’.

4.5 Sentences in Amharic

A sentence, in every language, is a group(s) of word(s) that comply with the grammatical
arrangement of the language and capable of conveying meaningful message to the audience. A
sentence in Amharic can be a statement which is used to declare, explain, or discuss an issue.
The combination of phrases to create another phrase that can express a full idea on something is
42
a sentence. When Amharic sentence is viewed from grammatical structure point of view it is a
combination of noun phrase and verb phrase. The noun phrase comes first and then the verb
phrase. Based on the number of phrases they contain sentences in Amharic are categorized under
two basic categories simple sentence and complex sentence. Simple sentence only contains a
single verb while complex sentence is constructed by combining more than one noun phrases and
verb phrases.

Declarative Sentence: is a sentence that is used to express the physical, psychological,

imaginary or real events. It is main objective is description of some issue. Most of the sentences
are declarative sentences. The news articles use the declarative sentence for expressing different
information on different issues. There are also Interrogative Sentence which is used to ask a
question, Exclamatory Sentence which is used for emphasis and emotion, and others.

Summary
The Amharic news text which we consider for training and testing of our system uses declarative
sentences to describe different issues about infrastructure news. Most of the facts that we try to
extract in this research work are names which are name of a person, name of an organization,
name of a place, numbers which describe different facts and date and times in the text. These
different facts in the Amharic text are either a noun or number. The ግ ስ (verb), ቅ ፅ ል (adjective),
ተውሳ ከ ግ ስ (Adverb) and መስ ተዋድድ (preposition) are not extracted in IE as they are descriptive
words to the noun and what the noun does or what is done on the noun.

43
 CHAPTER FIVE

5. THE AMHARIC TEXT INFORMATION EXTRACTION MODEL

In this chapter we will present our proposed Amharic Text Information Extraction (ATIE) Model
(Figure 5.1). The main components of the ATIE model along with their subcomponents and the
interaction between the main components will be presented.

5.1 Components of ATIE

Every IE system has three basic components which are the linguistic preprocessing, learning and
extraction and post processing regardless of the approach, language and domain on which the IE
system is developed for [2]. In addition to these three components other subcomponents are also
included in each of the main components. The Amharic Text IE model (ATIE) designed in this
work has one additional component called Text Categorization in addition to the above listed
three components. These four main components of ATIE also contain different subcomponents
which are language specific and general subcomponents that are required in IE. The main
components interact with each other to make information extraction from Amharic news text
possible.

The document preprocessing component is responsible for normalizing the different language
specific features which are caused due to its writing system. The language specific issues like
normalization of characters, normalization of numbers and tokenization are done at the document
preprocessing component. The text categorization component on the other hand handles the
categorization of preprocessed news text in to predefined categories. The categories for the news
which will be predicted using the text categorization component are investment, infrastructure
and others, which are among the subcategories under economy news main category. As IE is a
domain specific task text categorization component is used to select the specific news text that
can be used as an input to the IE system. The learning and extraction component extracts
candidate texts from the news text, which will be used as attributes to fill the database slot and
learns a classification model that will be used to predict the category of the candidate text. The
post processing component is responsible for the formatting of the extracted data according to
the structure of the predefined database slots.

44
 Document Preprocessing Text Categorization
Component Component

Punctuation Mark Remover

Tokenization

Input Text Stop Word Removal

Character Normalization

Number Prefix Separator Number Remover

Name prefix Separator

Number Normalization

Name Remover

Preprocessed Stemmer
Text
Component Word Selection

News text Category Selector TC Learning and Classification

Learning and
Extraction Component

Name Prefix Separator Categorized Post Processing

Text Index Component
POS
Data Formatting
Candidate Text Selection
Gazetteers

Feature Extractor (List of names)

Information Extraction Formatted Extracted

Data

Figure 5.1 The proposed Amharic Text IE Model

45
In the following subsections we give a more detailed description of the four main components
along with their subcomponents of our proposed ATIE model.

5.2 Document Preprocessing

The documents used as a corpus, training data and testing data for this work are Amharic news
texts obtained from ENA. The Amharic news texts need different type of preprocessing before
they are made ready for ATIE system as there are different language specific writing that should
be normalized. The document preprocessing component consists of tokenization, character
normalization, Number Prefix Separator, and number normalization as subcomponents to handle
language specific issues. These subcomponents and their function are described as follows.

Tokenization: is the task of chopping up the given text into pieces, called tokens [1], which are
then used by text categorization and IE. Since the text categorization components uses a
collection of words called features (attributes) to represent the news texts and IE uses token and
their POS as features for training and testing of the classifiers, tokenization is crucial for both
components. The tokenization is then used to chop up the news text into meaningful tokens that
can be used as features for both text categorization and IE. The Amharic punctuation marks, the
space between words are used for the tokenization process.

Character Normalization: The Amharic language has different characters with the same
meaning and pronunciation but with different symbols. The letters such as ሠ and ሰ ; ሀ, ኀ, ሃ, ኻ, ሓ,
ኃ and ሐ and ጸ and ፀ are examples of characters with the same meaning and pronunciation but
different symbol. Using them as they appear in Amharic literatures might have a meaning from
the language point of view but they need to be normalized when developing NLP tool for
Amharic language. These characters should be normalized to a single characters like ሠ and ሰ to
ሰ and ሀ, ኀ, and ሐ to ሀ and ጸ and ፀ to ፀ as well as their orders (ሠ, ሡ, ሢ, etc. to ሰ, ሱ, ሲ, etc.)
accordingly. In addition to this normalization, we further investigated and found that some other
orders of the letters should also be normalized. For example ሀ, ኀ, ሐ, ሃ, ኃ, and ሓ should be
normalized to ሀ. Using these characters as they appear in the news texts will create different
words that will be considered as different from one another in the training and testing of the
classifier for text categorization and IE. The role of character normalization is to normalize all
these characters of the same sound and meaning but different shapes to a single character.
46
Number Prefix Separator: Numbers in the Amharic news text appear using the number and
word representation. For example, the number might appear as ከ2ሚሊዮን, ከ2 ሚሊዮን, የ2ሚሊዮን, የ2
ሚሊዮን, በ2ሚሊዮን, and በ2 ሚሊዮን. This kind of number representation will cause a problem during
number normalization as they cause confusion in identifying them as numbers. The role of
Number Prefix Separator is then to separate the characters ከ, በ, and የ, which are the mostly
attached prefixes from the number, which will make the news text ready for number
normalization.

Number Normalization: the numbers in Amharic text appear by using the most widely used
number system in the world which is the Arabic number system. The problem is the numbers in
the Amharic text are mostly represented using words and digits. For example in most of the ENA
news the number 2,500,000 is written as “2 ነጥብ 5 ሚሊዮን” which is equivalent with 2.5 million.
The following are among the possible number representations in Amharic news texts 1ሚሊዮን,
1ሺ, 1ሺ ህ , 1ቢ ሊዮ ን , 1ሺ ህ 150, 1ሺ ህ 40, 1ሺ ህ 4, 1 ሺ ህ 1 መቶ ሃ ምሳ , 1 ሚሊዮ ን 1 መቶ ሺ ህ , 1
ቢ ሊዮ ን 5 ሚሊዮ ን 400 ሺ ህ , አ ን ድ ቢ ሊዮ ን አ ምስ ት ሚሊዮ ን አ ራ ት መቶ ሺ ህ , 1 ነ ጥ ብ 5 ሚሊዮ ን ,
and 1 ነ ጥ ብ 6 ቢ ሊዮ ን . These and other possible number representation in the news text causes
problem during text categorization and IE as it is difficult to uniquely identify the numbers in the
news text. The number normalizer then changes all above listed of number representation in to
their equivalent number representation. For example “2 ነጥብ 5 ሚሊዮን” will be normalized in to
2500000.

5.3 The Text Categorization Component

The text categorization component manages to categorize the news text as one of the predefined
categories by using a trained classifier model. It categorizes news texts so that the infrastructure
category news text will only be used as an input for IE. There are different subcomponents under
the text categorization which are used to preprocess the news text for the selection of words that
will be used as features (attributes) for training and testing of the classifier. The different
subcomponents of text categorization are punctuation mark remover, stop word removal, number
remover, Name Prefix Separator, name remover, stemmer, component word selection and text
categorization.

47
Punctuation mark remover: punctuation marks which exist in the news texts are not important
for text categorization. Therefore, the different punctuation marks that exist as independent token
or attached to another token are identified and removed.

Stop Words Removal: the words like አስታውቀዋል, አስታውሰዋል, እስካሁን, አሳሰበ, አሳስበዋል, አስፈላጊ,
አስገነዘቡ, አስገንዝበዋል, አብራርተዋል, አስረድተዋል and others listed in Appendix A, which are called stop
words are not important for text categorization. Since they exist in every category of Amharic
news texts they are not important to uniquely identify the one category of the news from the
other. Using them for text categorization will degrade the performance of the classifier.
Therefore the stop word removal removes all the above listed and other list of words in appendix
A, which are considered as stop words in the Amharic language from the news texts.

Stemming: reduces words to their root word so that the root word will be used for text
categorization. Ignoring removal of suffixes and prefixes from the word will increase the number
of words used as a feature for categorization which will consecutively degrade the performance
of the classifier. The purpose of stemming is to reduce all words to their root words so that the
classifier will be efficient since it uses the root words as features.

Number Remover: the different numbers which represent different aspects in the news texts or
any other text can’t uniquely identify a text as they can appear in different news and can be used
in different ways. The purpose of number remover is to remove all the numbers that exist in the
given text so that the number of features used to represent the news text will be minimized and
unique. Any token which contain numbers and any token which is a number will be removed by
using the number remover.

Name Remover: The names of a person, place, organization or any other type of names exist in
different news represent a different role than they have on other news. For example, the person
name in investment news might be used as the politician name in politics news or a place listed
in infrastructure news might also exist in sport news for conducting a national football match.
Therefore the names might be used in different news having different roles which makes them
not a feature to uniquely identify one news text from the other. The name remover then removes
all the names of a person, place, organization and titles attached to person name. The place and
organization names which exist in the news text mostly have prefixes attached to them which

48
make it difficult to remove names for that we developed another preprocessing tool called Name
Prefix Separator.

Name Prefix Separator

Most of the place and organization names which appear in the different ENA news we have used
as a corpus for the prototype development of our system has prefixes like በ, የ, ከ, and ለ attached
to them. These prefixes are identified after the analysis of the news texts considered as a corpus.
The function of the Name Prefix Separator is then to stem the prefix of all the names if they have
one of the above listed prefixes.

Component word selection: after the Amharic news text is preprocessed, the stop words, names
and numbers are removed and the remaining words are then used as features for text
categorization.

Text Categorization Model Learning and Classification: this subcomponent deals with
training a classifier model and uses it for prediction of the category of the unseen news texts. The
component words are used as an input for this subcomponent. The Weka preprocessing tools
and different classifiers which exist in Weka are used to train the classifier model. The selected
component words are then converted to ARFF format which is the dataset format that Weka
supports.

5.4 Learning and Extraction Component

This component is the main part of the ATIE model which is responsible for extracting candidate
texts from the infrastructure news texts and for learning a classifier model to predict the category
of the candidate texts, which will be used as an attribute for the database slots. This component
uses both the output of text categorization component and document preprocessing component.
The news texts which are classified as infrastructure news only passed to the learning and
extraction component and the other news are discarded as they are not considered for extraction
due to the domain specific nature of IE. The extraction and learning component also comprises
of different subcomponents which will be used to make the data ready for extraction.

49
Part of Speech Tagger: assigns a POS tag to each token. The set of tags includes the
conventional part of speech in Amharic such as noun, verb, adjective, adverb, prepositions and
pronoun. It is important for extraction as it is one of the features considered.

Candidate Text Selection: identifies possible candidate texts that are going to be extracted from
the news texts. The names of a person, place, organization, infrastructure and the different
numbers which exist in the news text are considered as candidate texts for extraction in this
research work. The candidate text selection then selects all the above listed text segments from
the news text as a candidate attribute values for extraction and tag them according to their
category that is as a name or a number.

Feature Extractor: extracts different features to uniquely identify the candidate text as one of
the predefined attribute. The features used for IE are POS tags, immediate attributes on both side
of the candidate token, the prefix and immediate next token of candidate token, the token
category of the candidate token which is the name of a person, place, organization, and
infrastructure or a number are extracted by the feature extractor. The extracted features by
feature extractor will be stored in the database for training and testing a classifier model.

Information Extraction: Any NLP tool developed for any language if it uses the machine
learning approach will have training and testing phase. The information extraction subcomponent
also has the training and testing phase. The following figure shows the detailed presentation of
Information Extraction subcomponent

50
 Training Phase Testing Phase

Predicting the category for

Preprocessing the candidate texts
Extracted
Using Weka
Features

Training a Token Category

classifier model with candidate texts

Classification
Model

Figure 5.2 The Information Extraction Subcomponent

During training phase the classifier model is generated based on the extracted features from
tagged training data. The extracted features are then preprocessed using Weka to make the data
ready for training a classifier model. After preprocessing, the classifier model will be generated
which will be used for predicting the category of the candidate texts during the testing phase.
The output of the information extraction subcomponent is the candidate texts with their
respective predicted category, which is used as an input to the post processing component.

5.5 Post Processing

It is the last component of ATIE model which formats the different attributes that are extracted
from the news text and store them in the database according to the predefined format of the data
base slots. The main function of the post processing component is to arrange the format of the
extracted data so that it will be flexible for data mining or any other application which want to
use the data.

Summary
In this chapter the ATIE model is presented and the main tasks of the different components are
described. The ATIE model comprised of document preprocessing, learning and extraction and
post processing and an additional text categorization component, which categorizes the Amharic
51
news text as one of the predefined categories. The document preprocessing handles the
processing of language related issues, the learning and extraction component learn a classifier
model and extract the candidate text segments from the news text, and the last component of the
system handles the formatting of the data to store it in the database. The ATIE model that is
proposed in this research work is a generic model which can be used for any other domain in the
Amharic language as long as there is a sufficient annotated training data.

52
 CHAPTER 6

6. IMPLEMENTATION
In this chapter, an Amharic text IE prototype system that is developed based on the proposed
model will be presented. The resources and algorithms used in the four main components of the
system and how the candidate texts for extraction are identified from unseen news texts are also
discussed.

6.1 Data Source

The data that we use to train and test ATIE system is obtained from Ethiopian News Agency
(ENA). ENA has a database containing different Amharic news from year 2005 to present. These
Amharic news texts are categorized under 16 main categories manually which are law and
justice, health, events directory, international relations, social affairs, culture, politics,
agriculture, defense and security, science and technology, sport, education, economy, accident,
weather and other classes. The main reason for selecting ENA as a data source is the availability
of large collection of manually categorized news texts in to one of the sixteen categories which
make it easy to select the specific news from the available category for training and testing
purpose.

Among the different categories from ENA news the economy category is used as a data source
for the training and testing of our ATIE prototype system. The reason for selecting economy
news category, among the other categories is the availability of factual information which can be
extracted and stored in the database. During selection of the news category as a data source for
our system we consider all the sixteen categories and analyzed which one of these categories
contain facts that can be extracted and stored in the database which can be used by different
application and individuals after extraction. Most of the other news categories are more
subjective which have less factual data that need to be extracted. The other reason is the
importance of the data after extraction. Economy news contain many vital information that
investors, financial organization or any individuals can use and extracting these information and
using them for later stage or by other application will be vital. Due to these reasons and the

53
nature of IE which is very domain specific the Economy main category and infrastructure
subcategory news texts are used as training and testing dataset in this research work.

6.2 Document Preprocessing

The document preprocessing component handles the different language specific issues that are
imposed by the nature of the language to make the data ready for remaining phases. The data
collected from ENA requires different normalization before it is used by the ATIE system. There
are different reasons for including document preprocessing in our system. One of the reasons is
the nature of the Amharic language characters which sound the same and used for the same
purpose but with different symbols. Using these characters as they appeared in a text without
normalization will make redundancy of words which in turn degrade the performance of text
categorization and IE as words are the features used for categorization purpose in both
components. The other reason is that number representation in the Amharic news text is
complicated. In some of the news it is represented using digits and in other news it is represented
using digits and Amharic characters and in some other news it is represented using characters
only. If we use them as they are they will create confusion during text categorization and IE as it
is difficult to consider them as numbers or words. The different document preprocessing
components are described as follows.

6.2.1 Tokenization
The purpose of tokenization in this research work is to break down a stream of text into words,
phrases, and symbols which will be used as meaningful elements called tokens for text
categorization and IE. As tokenization is a preprocessing subcomponent which will be used by
both text categorization and IE component, it is developed in a way it satisfy the requirements of
both components. The tokenization uses the ። ( አራት ነጥብ) the Amharic full stop, and ፣ (ነጠላ
ሰረዝ) the Amharic comma as the most commonly used punctuation mark in the news texts. These
punctuation marks play a vital role for extraction. The features for the tagged text are extracted at
the sentence level. The ። is used for identifying the sentence demarcation and ፣ is used to
separate different text segments which mostly are related. The tokenization then tokenizes all the
text segments which have space between each other as independent token. ። is used as

54
independent token and ፣ separated words are considered as a single token as they represent a list
of things under the same category.

6.2.2 Character Normalization

In Amharic languages there are characters which have the same sound and can be used
interchangeably but have a different symbols. Using these characters as they are, might have a
meaning. But when we come to implementing the Amharic text in natural language processing,
using them as they appear will reduce the efficiency of the system. For example, if we use them
as they are for text categorization we will have the same word with different representation as a
feature word which will degrade the performance of the system. If these characters are used as
they appear without normalization the same sound words will be written using different symbol
characters for example ጸሀይ፣ ፀሀይ፣ ጸሃይ፣ ፀሃይ which represent the same thing “sun” but appears in
different format will be used as independent features. Normalization of these characters in to a
single character is task of character normalization. The following are a list of characters with the
same sound but different shapes.

Table 6.1 Amharic Characters which have different symbols but similar sound

Characters with the same sound Normalized to

ሀ, ኀ, ሐ, ኃ, ኻ, ሓ, ሃ ሀ
ሰ, ሠ ሰ
ጸ, ፀ ጸ
ቸ, ቼ ቸ
ች, ቺ ች
ሸ, ሼ ሸ
የ, ዬ የ
ዪይ ይ
አ, ኣ, ዐ, ዓ አ
ኝ, ኚ ኝ
ሽ, ሺ ሽ

55
6.2.3 Number Prefix Separator
Numbers in the infrastructure news text are one of the facts that are considered for extraction in
this research work. The value for the attribute number of users and amount of money spent for
the development of the infrastructure are numbers. According to our analysis of the different
infrastructure news numbers have the two representation either they are attached to prefixes or
exist as independent tokens. The purpose of Number Prefix Separator is to stem the prefix from
the number and to consider the prefix and the number as independent tokens.

6.2.4 Number Normalization

Numbers are one of the text segments which are extracted by the ATIE system as facts to be
stored in the database. As the economy news texts are used as a data source for this research
work the different numbers representations in these news texts are considered. Numbers in these
news texts are represented using digits and words. For example the number 1.5 billion will exist
in the news text as “1 ነጥብ 5 ቢሊዮን” is needed to be extracted during IE and removed during text
categorization as it is not needed. Normalization of these number representations will make the
task of extracting the numbers for IE and removing for text categorization easy. The function of
number normalization is then to normalize all the numbers in the Amharic news texts which are
represented using digit or Amharic language characters or a combination of both in to their
equivalent number representation. The number “1 ነጥብ 5 ቢሊዮን” will be represented as
1500000000 after normalization. The following is the algorithm for number normalization.

For each news document in the corpus

Check for the presence of number or a word representation of a number or combination of
words and numbers existing as a single token
If present and the word is attached to a number
 Split the word or character from the number
 Initialnumber = the number separated from the character or word
 Candidatetext = the number representation of the separated word or character
or if it is not a number the character itself
 Token1= the immediate next token
 Token2= token next to token1

56
  Token3= token next to token2
 Token4= token next to token3
 Token5=token next to token4
Else if the word and the number are independent tokens
 Initialnumber = the number presented
 Token1= the immediate next token
 Token2= token next to token1
 Token3= token next to token2
 Token4= token next to token3
 Token5=token next to token4
End if
If Candidatetext is a word representation of a number
Candidate text will be the number equivalent representation of the word segment
Else if Token1 is a word equivalent representation of a number
Token1 will be the number representation of the word
Else if Token2 is a word equivalent representation of a number
Token2 will be the number representation of the word
Else if Token3 is a word equivalent representation of a numbr
Token3 will be the number representation of the word
Else if Token4 is a word equivalent representation of a number
Token4 will be the number representation of the word
Else if Token5 is a word equivalent representation of a number
Token5 will be the number representation of the word
End if
If Candidatetext is “NETB” or “.” and Token1 is a number and Token2 is a number and
Token3 is not a number
The normalized number will be the merger of Initialnumber, Token1 and Token2 by
removing one zero from Token2

57
Else If Token1 is not a number and Newnumber is not null
The normalized number will be the Merge InitialNumber and Newnumber
Else if Token1 is not a number and Candidatetext is null
The normalized number will be initial number
Else if Token1 is a number and Toekn2 is not a number
If the number of digits of Token1 is two
The normalized number will be Initialnumber+0+Token1
Else if the number of digits is one
The normalized number will be Initialnumber+00+Token1
Else if the number of digits are three
The normalized number will be Initialnumber+Token1
End if
Else if Token1 is a number and Toekn2 is a number and Token3 is not a number

The normalized number will be the Initialnumber+Token1+Token2

Else if Token1 is a number and Token2 is a number and Token3 is a number and token4 is not
a number

The normalized number will be the merger of Initialnumber, Token1 and Token3

Else if Token1 is a number and Token2 is a number and Token3 is a number and Token4 is a
number and Token5 is a number

If number of digits of Token3 are three

The normalized weight will be the merger of Initialnumber, Token1, Token3,

and Token4

Else if number of digits of Token3 are two

The normalized weight will be the merger of Initialnumber, Token1, the digit 0,
Token3 and Token4

58
 Else if the number of digits of Token3 is one

The normalized weight will be the merger of Initialnumber, Token1, the digit 00,
Token3 and Token4

End if

Else if

End if

End for

Figure 6.1 Algorithm for Number Normalization

6.3 Text Categorization

The text categorization component of ATIE prototype system is used to categorize the news text
as investment, infrastructure or others categories. The need of categorizing news texts is because
of the domain specific natures of IE. The different processes that are involved in the text
categorization component are organized under the following three activities

Training data preparation

Training a classifier model

Using the trained classifier model for text categorization

6.3.1 Training Data Preparation

The training data that is used for text categorization is economy news category which is obtained
from ENA. The different news texts which are under the investment, infrastructure and other sub
categories are used for training and testing the text categorization. The training data is prepared
by manually selecting all the news texts and storing them in the text file format. Then, all the text
file format news texts are organized in three different folders which have the same name to the
category of the news texts which are investment, infrastructure, and others. To obtain features
which will be used for training and testing the classifier the following preprocessing tools are

59
used for each of the news texts. For the text categorization subcomponent the work of Yohannes
Afework [33] which is done on text categorization using Weka is adopted.

Feature Word Selection

For the purpose of text categorization the news texts are considered as a bag of key words which
are called features or attributes of the document [3]. Not all words in a document are considered
as features. The purpose of feature selection is then to reduce the dimension of the data by
selecting features from the original words of the news text. Different preprocessing tools can be
used on the news text for feature selection. Redundant attributes are identified and removed as a
result of feature word selection process. The following processes are used for facilitating the
feature selection from the news text.

Stop Word Removal

Words that can’t represent the documents since they commonly exist in different news texts and
sentences are considered as stop words. If a word exists across documents it is considered as a
stop word in text categorization. Words like አስታውቀዋል, አስታውሰዋል, እስካሁን, አሳሰበ, አሳስበዋል,
አስፈላጊ, አስገነዘቡ, አስገንዝበዋል, አብራርተዋል, አበራርተው, አስረድተዋል are called stop words. This and
other stop words which are listed in the research work of Melese [40] are considered as stop
words and removed as they are not important for text categorization.

Stemmer

Stemming is the process for reducing inflected (or sometimes derived) words to their stem, base
or root word form [1]. Stemming is an important part of text categorization for language like
Amharic which has many morphemes for the same root word. Words might be expanded to
represent number, feminine, ownership, and other issues but for text categorization they are all
the same. Using all the words as they exist will degrade the performance of text categorization as
the feature words used for classification increases. For example, the words ልማትና, ልማትንና,
ልማትንናም are changed to their base term ልማት ‘lemat' (development). Similarly the variants of
selam such as ሰላምና, ሰላምንና, ሰላምንም are changed to ሰላም 'selam' (Peace). The stemming
algorithm developed in [41] and later used by Seid [1] is adopted for our work as it is developed
in the way that suits our work.

60
Number Remover

Numbers which exist in different Amharic news representing different amount associated with
them are not important for text categorization. The numbers exist in economy news might also
exist in sport, social or any other news category. Numbers are then not used for text
categorization. Number remover removes all the numbers and tokens which contain numbers
from the given news text.

Name Remover

In Amharic news text names of a person, place, organization, titles of a person exists having a
different meaning from one document to the other. The person names which exist in the
economy news as an investor might exist in other news having different role. Using names as a
feature for text categorization is not important as it doesn’t uniquely identify the news document.
The name remover then removes all the names by using the Gazetteer lists. The problem in
removing the names from the Amharic news text is that these names might have prefixes. Most
of the place and organization names in the news text that we use have one of the following four
prefixes ከ , በ, ለ, and የ attached to them. To solve this problem the Name Prefix Separator is
applied before the number remover is used, which removes all the names in the news text.

Name Prefix Separator

The Name Prefix Separators stem all the prefixes from names. If the first letter of the token is ከ,
በ, ለ, or የ the remaining part of the token is checked if it is a name or not. If it is a name it will be
Separated and both the prefix and the name will be removed.

Preprocessing using Weka

After the news texts are preprocessed using the language specific developed tools the next step is
to preprocess it using Weka to make the data ready for training the classifier and using the model
for prediction at later stage. Since the feature words that are used as dataset for training and
evaluating attributes are Strings the TexttoARFFLoader tool which Weka support is used to
convert all the text files news texts in to ARFF format. A string attribute can have in principle
infinite number of values and therefore it cannot be handled by any classifier in Weka. That is
why we have to convert string values into a set of attributes that represent the frequency of each
61
word in the strings. The StringToWordVector filtering tools also performs TF/IDF
transformation. TF/IDF weight (Term Frequency–Inverse Document Frequency) is a weight
often used in information retrieval and text mining. This weight is a statistical measure used to
evaluate how important a word is to a document in a collection or corpus. The importance
increases proportionally to the number of times a word appears in the document but is offset by
the frequency of the word in the corpus. The StringToWordVector filter is then is used to select
the feature words that can be used as attributes for training the text classifier [42].

6.3.2 Training a Classifier Model

Once the news text is preprocessed and the features are selected the next step is to train the
classifier model which later will be used to predict the category of unseen news texts. For
training a classifier Weka open source machine learning algorithm is used. Weka is an open
source machine learning algorithm for data mining task. It has open source machine learning
algorithm which contains tools for data pre-processing, classification, regression, clustering,
generate association rules, and visualization. All of the packages are developed in java at the
University of WAIKATO in New Zealand [42].

To use the Weka package one of the three interfaces that Weka provides can be used

A command line interface, which accept a command through the simple CLI command
line interface. There are different commands that can be used with simple CLI.

An Explorer GUI interface, which has the menus that the user can select for
preprocessing, classification and regression and apply them on the appropriate dataset.

An experimenter GUI interface: which allow running different algorithms in batch and
comparing the results

Weka provides different machine algorithms which can be applied on textual data and numeric
data. Among these different algorithms Naïve Bayes, SMO and Decision tree are used as they
are the most commonly known and used algorithms. The preprocessed dataset is then used to
train and evaluate the three classifiers that are used in this research work.

62
6.3.3 Using the Trained Classifier Model
After the classifier model is generated it can be used for predicting the category of the unseen
news text. To use the model to predict the category of unseen news text, the news text should be
preprocessed first in a same way the training data is prepared and converted in to ARFF format.
After that it will be loaded to the Weka system and the saved classifier model is then used to
predict the news category for the unseen news text.

6.4 Learning and Extraction

The learning and extraction component main task is to extract candidate texts and train and use
the classifier model for predicting the category of the extracted candidate texts. The main
hypothesis of this research work is that we human beings identify facts from a document by
using the informative words that come before and after the candidate text which we consider are
facts. For example, if we see the number in the text to know what it represents, we don’t need to
read all the document just looking at the words that are on both side of the number will be
enough. The main task in the learning and extraction component is then to extract candidate
tokens in the infrastructure news text and use the tokens that are on both sides of the candidate
token, their respective POS tags, and a token category as features to know the category of the
token. The classifier model can use the features to learn how to classify the candidate text as one
of the predefined attributes for extraction.

To train and use the classifier model for the learning and extraction component use the training
and testing phase. The training phase uses the manually annotated news data to train a classifier
model which will be used for extraction while the testing phase evaluates the performance of the
trained model on the unseen news texts.

6.4.1 Training Data Preparation

One of the time taking and tedious task of developing an IE system is preparation of training
data. For the sake of developing the prototype and experimenting the ATIE system the
infrastructure news subcategory which is under the economy news main category is used. But the
architecture that we develop for the ATIE system is generic one which can work on any domain
of the Amharic news text or any other text as long as there is plenty of a manually annotated data
is available for training and testing the system.
63
The predefined attributes that we set to extract from infrastructure news are the Infrastructure
name, Place where the infrastructure is built, the amount of money used to build the
infrastructure, the source of money for the infrastructure development, the number of users
which will be benefited from the infrastructure, the person who give the information to the news
agency. These six attributes are selected as an attribute after analysis of the different
infrastructure news and the common facts which exist in most of the infrastructure news and
which we thought are relevant facts that should be extracted.

During training data preparation all the text segments which are numbers and names are tagged
accordingly. If the name is the fact to be extracted it will be tagged according to the tag set for
the predefined attribute otherwise it will be tagged by other tags. The main purpose of tagging all
the text which are not facts considered for extraction to train the classifier to identify texts that
are part of the predefined database slot and those that are not.

B-INF for infrastructure name

B-PLACE for the place

O-PLACE for place which exist in the text but are not the facts considered for
extraction

B-FUND for the amount of money spent for the development of the infrastructure

O-NUM for numbers which are not tagged as fund or number of users

B-FSOURCE- for the financial source for the infrastructure development

O-ORG organizations which are not the financial source

B-NUSERS- for the number of users after the infrastructure development is
finished

B-REPORTER- for the name of the person who give the information to the news
Agency

O-REPO- person names other than the one who give the information

6.4.2 Processing of the News Texts for Feature Extraction

To make the news texts ready for IE different preprocessing are required. The features that are
used for IE are candidate token, prefix of the token, previous token, immediate next token, next
token, POS of the above listed tokens and the token category of the candidate token. The value

64
of the token category is a number, place name, person name, or organization name as the
candidate texts considered for extraction are these. To extract these features from the news texts
different processing tools are developed.

POS

The part of speech tagger is one of the features considered for IE. It is developed by using the
standard Amharic corpus from Walta. The Hidden Markov model is used to train the POS from
the tagged Amharic news text. The performance of the classifier is also evaluated using Ling
pipe experimentation option and it assign for the correct POS 79% tokens used for testing.

Name Prefix Separator

The Name Prefix Separator is used to stem the prefix from names. As place names, person
names, and organization names are the candidate tokens for extraction the prefix attached to
these names is stemmed and stored independently. Person names doesn’t have any prefix
attached to them which need stemming while organization and place names have a prefix
attached to them. Based on our analysis from different corpus the prefixes that are attached to
place and organization names are the letters ከ, በ, ለ or የ. The role of Name Prefix Separator is
then to stem the prefix from the name and write them to the text file as independent tokens. The
following is an algorithm for stemming prefix from names

65
 For each token in the corpus
Check if the token is a name by using the gazetteer list
If the token is a name
Write the token to a file
Else if the token is not a name
Check for the first character of the token
If the first character is ከ, በ, ለ or የ

Stem ከ, በ, ለ or የ from the token

Check if the remaining part of the token is a name

If true
Write the prefix as independent token
Write the remaining part of the token as independent token
Else
Write the token to the text file as it is
End if
End if
Else
Continue to the next token
End if
End for

Fig 6.2 Algorithm for Name Prefix Separator

6.4.3 Feature Extractor

The feature extractor component is developed to extract the different features and a category
from the tagged training data. The features that are used for ATIE system for prediction of the
category of the candidate texts are:

POS of the candidate token

66

POS of the prefix

POS of the previous token

POS of the immediate next token

POS of the next token

Candidate token category

Prefix token

Previous token

Immediate next token

Next token

The feature extractor extracts all these features from the tagged training data and stores them in
the database for latter processing by the Weka preprocessing component and classifier learning.
These features are selected after analyzing the nature of the language and by considering it will
provide a good prediction for the category of candidate tokens. The features are extracted on a
sentence level basis. The following is an algorithm developed for feature extraction in this
research work.

67
For each token in the news document
Check for the presence of a token which is tagged
If present
If the token is the first token
Previous token will be null and its POS will be null
Prefix will be null and its POS will be null
Immediate token will be the first next token to the tagged token and its POS will be the POS
attached to it
Next token will be the token next to immediate token and its POS will be the POS attached to it
Else if the token is the second token
Previous token will be null and its POS will be null
Prefix will be the first token and its POS will be the POS attached to it
Immediate token will be the first next token to the tagged token and its POS will be the POS
attached to it
Next token will be the token next to immediate token and its POS will be the POS attached to it
Else if the next token is :: (the Amharic full stop)
Previous token will be null and its POS will be the POS attached to it
Prefix will be the first token and its POS will be the POS attached to it
The immediate next token will be :: and its POS will be the POS of ::
The next token will be null
Else
Previous token will be null and its POS will be the POS attached to it
Prefix will be the first token and its POS will be the POS attached to it
Immediate token will be the first next token to the tagged token and its POS will be the POS
attached to it
Next token will be the token next to immediate token and its POS will be the POS attached to it
End if
End if
End for
Figure 6.3 Algorithm for Feature Extractor

68
6.4.4 Preprocessing the Data Using Weka
After the features are extracted the next step is to preprocess the data using Weka filtering tools.
The StringToWordVector conversion filtering tool is used. It has the same purpose to that of its
use for text categorization. The purpose of the filtering tool is to convert string attributes into a
set of attributes representing word occurrence (depending on the tokenizer) information from the
text contained in the strings. The set of words (attributes) is determined by the first batch filtered
(typically training data).

6.4.5 Training a Classifier Model

After the data is preprocessed it is passed to the classifier learners in Weka. The explorer or the
command line can be used. Weka classifiers can also be used from Java code which we use for
the implementation of the prototype system. The decision tree, naïve Bayes and SMO classifier
learning algorithms are selected among the different available machine learning algorithms that
Weka supports. These algorithms are selected after considering the possible machine learning
algorithms that can work on text data and numeric data. We select the three algorithms to see the
performance of the classifier and evaluate their efficiency in predicting the category for the
candidate token.

6.4.6 Using Trained Classifier for Information Extraction

The purpose of information extraction is to extract information from unseen news text. Once the
classifier model is generated using the training data the next step is to use the trained classifier
model to work on the unseen news text. The problem is the trained classifier model doesn’t
directly apply on the plain news texts. It rather works on the extracted features from the news
texts. To use the classifier model for extraction first the candidate texts should be identified and
the features of the candidate texts must be extracted using the feature extractor. In order to
achieve this there is a candidate text selection component which automatically tag the candidate
tokens. After the data is preprocessed and the POS of the tokens is identified and a name
stemmer is applied then the candidate text selection will be applied.

69
Candidate Text Selection

The candidate texts for extraction for this research work are names and numbers. The name of a
person is a candidate text for Reporter attribute, place name is for place the infrastructure is built
on, the number in the news text is considered as a candidate text for number of users and the for
the amount of money spent for the infrastructure development, organization names are used as
candidate texts for attribute financial source for infrastructure development and names used for
infrastructure are the candidate texts. The task of candidate text selection is to tag all these
candidate text accordingly to their attributes. The Gazetteers is used for candidate text selection
purpose. The gazetteer which consists of names for different places in Ethiopia, different names
that can be used for identification of persons, the different infrastructure names, and the different
governmental and nongovernmental organization list is used. The candidate text selection is
done using the Gazetteer list which comprises of the different names under consideration.

Feature Extraction

Once the candidate texts are selected and tagged, the feature extractors then extracts all the
features from the tagged candidate texts and store it in the database for later processing by Weka.

Extraction of data

After the features are extracted and preprocessed using weka the trained classifier model is used
to predict the category of the candidate text. Among the candidate texts those with the token
category of the predefined attributes will be stored in the database and others which are not under
the category of the predefined attributes will be discarded.

6.5 Post Processing of the Data

The post processing handles formatting of the data after the extraction. The attributes are
formatted according to the predefined order of the table formation. In this research work the six
attributes that are extracted are stored in the table as follows.

70
6.6 The Prototype User Interface

Based on the designed model the prototype system is developed to evaluate the performance of
the proposed model. The following figure shows the ATIE system after extracting the data from
the infrastructure news text. The extracted data from the news texts are then stored in the
database for further use by users or any other application.

Figure 6.4 Sample Prototype System after the Data Is Extracted From the News Text

For the development of the prototype ATIE system Java programming language, Ling pipe and
Weka open source machine learning algorithms are employed. The different subcomponents of
the text categorization and information extraction components, which are mostly language
specific issues are developed using java based on the designed algorithms. The main reason that
the java programming language is used is that machine learning algorithms of Weka are
developed using java which can easily be imported and used in the java environment. Among the
different machine learning algorithms that Weka supports the one that can be applied for textual
and numeric data are selected and used.

71
To use the prototype the news texts can be opened from the location where it is stored as a file
and by using “ክፈት“ button. After the file is opened will be first be preprocessed using the button
“የጽሁፍ ዜናን ቅድሚያ ማዘጋጀት “, which will tokenize the news text, normalize the different
characters, stem the prefix from the number, and normalize the number representation in the
news text. After the news text is preprocessed the output can be saved in a selected location by
using the “አስቀምጥ“ button. To make the preprocessed news text ready for text categorization
another button which is the “የጽሁፍ ዜናን ለፍረጃ ቅድሚያ ማዘጋጀት “ will be used. When the button is
pressed the different subcomponent under the text categorization component will be called and
applied on the preprocessed text to select features, which will be used for prediction purpose
later. After the news text is preprocessed for categorization the “ፈርጅ“ button is used to categorize
the news text as one of the three predefined categories. The “ከጽሁፍ ዜና ጠቃሚ መረጃ ለማዉጣት
ቅድሚያ ማዘጋጀት “button is used to preprocess the news text for information extraction. The
different subcomponents under learning and extraction component are applied on the
preprocessed infrastructure news text, which will select the candidate texts and select those
candidate texts which will be used as attributes for the predefined database slot. After it is
preprocessed, the “ጠቃሚ መረጃ አዉጣ “ button can be used to extract the information that can be
used as an attribute of the six predefined attributes. The candidate texts which are categorized as
one of the six predefined attributes will be stored in the database according to their attribute
value as the main purpose of the research is to extract and store the data in the database for
further use and management of the data by the users or applications.

Summary
The ATIE prototype system which is developed in this research work has four main components
which are interconnected with each other to extract information from infrastructure news text.
The document preprocessing component handles language specific issues, text categorization
component classify the news text as one of the predefined categories, the learning and extraction
component train a classifier and use the classifier for extract information and post processing
format the extract data and store them in the database. The different algorithms and
subcomponents under the four main components are also presented.

72
 CHAPTER 7

7. EXPERIMENT
In this chapter the test data set, the evaluation metrics and the evaluation result of the proposed
ATIE system will be presented. The standard methods that are used to evaluate a classifier model
are employed to evaluate the performance of our system.

7.1 Experimental Procedures

IE system is a domain specific activity which needs to be applied on a specific domain which
share similar patterns. Using IE system for a different domain for which they are not developed
for will not function correctly. Therefore, the following tasks are carried out in order to evaluate
the ATIE system we proposed.

7.1.1 Data Collection

Among the different news texts which exist in ENA’s database the economy news texts are used
for training and testing purpose of our system. As we are evaluating two different components of
ATIE i.e. text categorization and IE we prepare two different data sets, which will be used for
training and testing of both components. For text categorization 1200 news texts which
comprises of 400 news of infrastructure subcategory, 400 news of investment subcategory, and
400 news of subcategory others is used as a dataset for training and testing of the text
categorization component of the ATIE system. For the information extraction component 300
news texts from the infrastructure subcategory is used as a dataset for training and testing
purpose.

Test Data Set Preparation

Weka support four different evaluation techniques which are using training set, using supplied
test set, using fold cross validation and using percentage split. The training set option uses the
training set for training the classifier model and evaluating the classifier by using the same data
while the supplied test set uses independent test set for performance evaluation. In cross
validation option the dataset is randomly reordered and then split into n folds of equal size. In
each iterations, one fold is used for testing and the other n-1 folds are used for training the
73
classifier. The test results of each iterations are then collected and averaged over all folds which
give the cross-validation estimate of the accuracy. If the fold is 10 the data set will be divided
into 10 equal datasets and at a time one of the dataset is used as test dataset and the other 9 are
used as a training dataset it will continue 10 times by making each of the 10 dataset as testing
dataset. The output is then added together and the average will be the performance result. The
percentage split on the other hand allows splitting the dataset into training and test set randomly
by using the specified percentage by the user. [42].

Among the above different evaluation options in Weka we use the 10 fold stratified cross
validation technique for the text categorization and IE as it is good to get a better result out of the
classifier. Therefore the test data set are not prepared manually as they are automatically created
by the Weka itself.

7.2 Performance Evaluation

Since the Weka machine learning algorithms are used for training and testing the classifier model
we use the standard evaluation techniques used for evaluating the performance of classifier
algorithms. The accuracy of a classifier is measured by using classifier’s performance on a test
data. It is the percentage of test instances that are correctly classified by the classifier and those
that are classified incorrectly. The category assignment of a classifier is evaluated using the
confusion matrix which analyzes how well a classifier recognizes instances of different classes
and also clearly indicate where the confusion lies while predicting the category. The Weka itself
automatically generates the different experimental results.

7.2.1 Evaluation of Text Categorization Component

The text categorization component is responsible for categorizing any economy news text as
Infrastructure, Investment or others category. Among the different machine learning classifier
algorithms SMO, Decision tree and Naïve Bayes classifier are used for evaluating their
performance. The dataset contains 1200 instances which comprises 400 for each subcategory
under economy news.

74
7.2.1.1 Evaluation of Text Categorization Component Using Decision Tree Classifier
Decision tree contain different induction algorithms. The induction algorithms are ADTree,
BFTree, DecisionStamp, Id3, J48, LMT, MSP, NBTree, RandomForest, RandomTree, REPTree,
and Simplecast. All these tree induction algorithms are not used for training and testing the
classifier for text categorization component. ADTree cannot handle non-binary class (the
experimental data has nominal classes), Id3 cannot handle numeric attributes. MSP cannot
handle nominal class. Most of the decision tree classifiers require a large amount of time and
memory for training and evaluating the classifier. J48 is the only classifier which requires less
memory and time and we test the dataset using J48 and the following is the output result.

Experimental Result

Correctly Classified Instances 1007 83.9167 %

Incorrectly Classified Instances 193 16.0833 %

Confusion matrix

Table 7.1 Confusion Matrix for Text Categorization Component Using Decision Tree

Infrastructure Investment Others Classified as

331 14 55 Infrastructure

6 377 17 Investment

77 24 299 Others

The similarity between the news sources increases the confusion for the classifier in predicting
the category of the news document. For example in the first row of the confusion matrix table
among the 400 Infrastructure news 331 of them are correctly classified as infrastructure news
while the other 77 news confuses the classifier and predict 14 of them as Investments news and
55 of them as others news. The same is true for the second and third row which also show the
confusion of the classifier in predicting the Investment and others news category. The confusion
is minimum while predicting the Investment news by correctly classifying 377 of the investment

75
and confused only in predicting 33 of them as infrastructure and others news category. This is
due to the reason that the nature of the news text has a very similar attributes that confuse the
prediction of the classifier. The confusion matrix is later used to calculate the detail accuracy of
the classifier.

Detailed Accuracy by Class

Table 7. 2 Detailed Accuracy by Class for Text Categorization Component Using Decision
Tree

TP Rate FP Rate Precision Recall F-Measure Roc Area Class

0.828 0.104 0.8 0.828 0.813 0.868 Infrastructure

0.943 0.048 0.908 0.943 0.925 0.954 Investment

0.748 0.09 0.806 0.748 0.776 0.836 Others

0.839 0.08 0.838 0.839 0.838 0.886 Weighted Avg.

The detail accuracy by class result of the text categorization component using decision tree
describes the TP (True Positive), FP(False Positive), precision, recall, F-measure and Roc Area
for each subcategory of the news text the classifier predicts. TP Rate indicates the rate at which
the classifier correctly predicts the category of the news text. For example, in the first row the
table TP rate of Infrastructure news text is 0.828, which means the rate of the classifier in
correctly identifying the news text is 0.828 while is the FP rate is 0.104. As it is described in
chapter 2 the precision and recall are also presented. The ROC Area indicates the curves in
identifying the true positive over the false positive instances. Among the three categories the
ROC Area of Investment news is higher as the number of true positives is higher than the other
subcategories and the numbers of false positive instances are lower than other subcategories. The
J48 classifier in the decision tree is the better algorithm that can perform on limited memory and
requires less processing time while others require a tremendous amount of time and memory
which make it difficult to evaluate them. As it is claimed in the work of Yohannes Afework [33]
LMT decision tree classifier is better than all the other classifiers in decision tree however it is
difficult to evaluate it as it require large amount of memory and time.

76
7.2.1.2 Evaluation of Text Categorization Component Using Naïve Bayes
The other classifier algorithm that Weka support and which can be applied on textual data is
Naïve Bayes. Like decision tree Bayes also comprises different classifier algorithms and among
these algorithms the one which perform better is DMNBtext classifier. It is evaluated on the
same data set using the same test option which has the following experimental result.

Experimental result

Correctly Classified Instances 1114 92.8333 %

Incorrectly Classified Instances 86 7.1667 %

The accuracy of the DMNBtext classifier is better than the decision tree classifier by correctly
predicting the category of 1114 news texts. The confusion matrix also show that the Naïve
Bayes algorithm perform better than that of the decision tree which reduces the confusion of the
classifier in predicting the category for the news texts.

The confusion matrix

Table 7. 3 Confusion Matrix for Text Categorization Component Using Naïve Bayes

Infrastructure Investment Others Classified as

372 7 21 Infrastructure

5 383 12 Investment

29 12 359 Others

The confusion in categorizing the news text is better than that of the decision tree. For example,
in the first row of the confusion matrix table among the 400 Infrastructure news texts, 372 of
them are classified correctly while the classifier predicts 7 of them as Investment and 21 of them
as others. The same works for the remaining two line of the table. The increase in the correctly
classified news texts will minimize the confusion. Since the confusion is less and the accuracy of
the classifier in correctly predicting the category of the news text is higher than Decision tree ,
the precision, recall, TP rate values are increased while using Naïve Bayes algorithm.

77
 Detailed Accuracy by Class

Table 7. 4 Detailed Accuracy by Class for Text Categorization Component Using Naïve
Bayes

TP Rate FP Rate Precision Recall F-Measure Roc Area Class

0.93 0.043 0.916 0.93 0.923 0.976 Infrastructure

0.958 0.024 0.953 0.958 0.955 0.986 Investment

0.898 0.041 0.916 0.898 0.907 0.983 Others

0.928 0.928 0.036 0.928 0.928 0.928 Weighted Avg

7.2.1.3 Evaluation of Text Categorization Component Using SMO

The other classifier algorithm that Weka support and which can be applied on textual data is
SMO. The experimental results while using SMO as a classifier are presented as follows

Experimental result

Correctly Classified Instances 1099 91.5833 %

Incorrectly Classified Instances 101 8.4167 %

The Confusion Matrix

Table 7.5 Confusion Matrix for Text Categorization Component Using SMO

Infrastructure Investment Others Classified as

355 8 37 Infrastructure

6 386 8 Investment

34 8 358 Others

78
 Detailed Accuracy by Class

Table 7.6 Detail Accuracy by Class for Text Categorization Component Using SMO

TP Rate FP Rate Precision Recall F-Measure Roc Area Class

0.888 0.05 0.899 0.888 0.893 0.927 Infrastructure

0.965 0.02 0.96 0.965 0.963 0.979 Investment

0.895 0.056 0.888 0.895 0.892 0.931 Others

0.916 0.042 0.916 0.916 0.916 0.946 Weighted Avg.

7.2.1.4 Comparison of the Performance Classifiers Used for Text Categorization

Among the three classifiers that we used to evaluate the text categorization component of our
ATIE system the Naïve Bayes algorithm perform better than others. The experimental result of
SMO is also very similar to that of Naïve Bayes.. The Naïve Bayes algorithm doesn’t only
perform better in correctly classifying the news text but also require less time and memory to
execute the training and prediction. From the experimental result, we can conclude that when
there is an attribute similarity among the classes that are considered for text categorization the
Naïve Bayes will perform better.

7.2.2 Evaluation of IE component

The IE component is the crucial component of this research work which uses different features
for predicting the categories of the candidate texts. The main focus of the evaluation of IE
component is to see the role of the different features that are considered and their effect in
efficiently categorizing the candidate tokens as one of the predefined attributes. Four different
scenarios are considered for the evaluation to see the role of the different features in the
prediction of the category process. The different features that are considered in this research
work are POS of the candidate token, POS of the prefix, POS of the previous token, POS of the
immediate next token, POS of the next token, candidate token category, prefix token, previous
token, immediate next token, and next token.

Scenario 1: Using all the features

79
Scenario 2: Using all the features except the POS tags

Scenario 3: Using all the features except the token category

Scenario 4: Using all the features except token category and POS tags

For the above all the scenarios the Decision Tree, SMO and Naïve Bayes algorithms are used for
the experiment. The experimental result of the algorithm which performs better than others in
each of the considered scenarios is presented. The same testing option which is used for text
categorization component, which is 10 fold stratified cross validation is used.

Scenario 1:

Among the three algorithms selected for experiment the SMO classifier performs better than the
other classifiers by correctly classifying 94.6% of the instances correctly while the naïve bayes
perform by correctly classifying 93.6 and J48 decision tree algorithm by performing 90.6%. All
the instances that are used for training and testing the IE component are 1422 which includes the
different instances for the six predefined attributes for extraction and the other candidate texts
which are names and numbers but not considered as a fact for extraction. The detailed
experimental result of SMO algorithm is presented as follows.

Experimental result

Correctly Classified Instances 1345 94.5851 %

Incorrectly Classified Instances 77 5.4149 %

80
Confusion matrix

Table 7. 7 Confusion Matrix for IE Component for Scenario 1 Using SMO

Infrastructure Place Money Reporter Number Financial Category

Name Spent of Users Source

268 0 0 0 0 0 Infrastructure name

0 253 0 2 4 0 Place

0 0 249 0 0 1 Money spent

0 12 0 121 0 0 Reporter

0 0 0 0 172 0 Number of users

0 0 0 0 0 126 Financial source

From the confusion matrix table we can see that the confusion in predicting the different
categories considered in IE is very minimal.

Detailed accuracy by class

Table 7. 8 Detailed Accuracy by Class for IE Component for Scenario 1 Using SMO

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

0.996 0 1 0.996 0.998 1 Infrastructure Name

0.996 0.003 0.985 0.996 0.991 0.998 Place

0.977 0.003 0.984 0.977 0.981 0.997 Money Spent

0.956 0.007 0.95 0.956 0.953 0.994 Reporter

0.91 0.012 0.883 0.91 0.896 0.985 Number of users

0.955 0.014 0.875 0.955 0.913 0.991 Financial Source

0.946 0.006 0.942 0.946 0.943 0.992 Weighted Avg.

81
Scenario 2

Among the three algorithms for scenario 2 the Naïve Bayes algorithm performs a little bit better
than SMO. Naïve Bayes correctly classified 94.86 while SMO perform 94.58 % and J48 decision
tree algorithm correctly classifies 85.72%.

Correctly Classified Instances 1349 94.8664 %

Incorrectly Classified Instances 73 5.1336 %

Confusion Matrix

Table 7. 9 Confusion Matrix for IE Component for Scenario 2 Using Naïve Bayes

Infrastructure Place Money Reporter Number of Financial Category

Name Spent Users Source

250 0 0 0 0 0 Infrastructure name

0 269 0 0 0 0 Place

0 0 257 0 2 0 Money spent

0 0 5 180 0 0 Reporter

0 0 0 0 128 0 Number of users

0 0 0 0 0 132 Financial source

82
Detailed accuracy by class

Table 7. 10 Detailed Accuracy by Class for IE Component for Scenario 2 Using Naïve
Bayes

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

1 0 1 1 1 1 Infrastructure Name

1 0.003 0.985 1 0.993 0.999 Place

0.992 0.015 0.935 0.992 0.963 0.999 Money Spent

1 0.006 0.957 1 0.978 0.997 Reporter

0.962 0.021 0.826 0.962 0.889 0.995 Number of users

1 0.012 0.898 1 0.946 0.995 Financial Source

0.949 0.007 0.95 0.949 0.944 0.994 Weighted Avg.

Scenario 3

Among the three algorithms for scenario 3 the SMO algorithm once again performs better than
others. The SMO classified the 90.6% of the dataset correctly while the Naïve Bayes perform
90.6% and J48 decision tree algorithm performs 85.8%.

Correctly Classified Instances 1289 90.647 %

Incorrectly Classified Instances 133 9.353 %

83
Confusion matrix

Table 7. 11 Confusion Matrix for IE Component for Scenario 3 Using SMO

Infrastructure Place Money Reporter Number of Financial Category

Name Spent Users Source

0 0 0 0 0 0 Infrastructure name

8 252 0 0 0 9 Place

1 0 253 0 5 0 Money spent

0 0 0 173 0 0 Reporter

0 0 12 0 121 0 Number of users

0 4 6 0 0 116 Financial source

Detailed accuracy by class

Table 7. 12 Detail Accuracy by Class for IE Component for Scenario 3 Using SMO

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

0.92 0.022 0.898 0.92 0.909 0.973 Infrastructure Name

0.937 0.013 0.944 0.937 0.94 0.983 Place

0.977 0.003 0.984 0.977 0.981 0.996 Money Spent

0.961 0.006 0.956 0.961 0.958 0.996 Reporter

0.91 0.013 0.877 0.91 0.893 0.978 Number of users

0.879 0.02 0.817 0.879 0.847 0.947 Financial Source

0.906 0.013 0.901 0.906 0.903 0.975 Weighted Avg.

84
Scenario 4

Among the three algorithms for scenario 4 the SMO algorithm performs better than others. SMO
classifies 90.64% of the dataset correctly while Naïve Bayes and J48 decision tree algorithm
classifies 89.5218 % and 85.7243 % respectively. The detailed experimental result of the SMO
algorithm is presented as follows

Experimental result

Correctly Classified Instances 1289 90.647 %

Incorrectly Classified Instances 133 9.353 %

Confusion matrix

Table 7. 13 Confusion Matrix for IE Component for Scenario 4 Using SMO

Infrastructure Place Money Reporter Number of Financial Category

Name Spent Users Source

232 3 0 0 0 0 Infrastructure name

2 252 0 0 0 0 Place

0 0 253 0 6 0 Money spent

0 0 0 173 0 0 Reporter

0 0 0 0 123 0 Number of users

2 1 0 0 0 117 Financial source

85
Detailed Accuracy by Class

Table 7. 14 Confusion Matrix for IE Component for Scenario 4 Using SMO

TP Rate FP Rate Precision Recall F-Measure ROC Area Class

0.928 0.01 0.951 0.928 0.939 0.98 Infrastructure Name

0.937 0.008 0.966 0.937 0.951 0.982 Place

0.977 0.003 0.984 0.977 0.981 0.996 Money Spent

0.961 0.008 0.945 0.961 0.953 0.996 Reporter

0.925 0.012 0.891 0.925 0.908 0.984 Number of users

0.886 0.02 0.818 0.886 0.851 0.953 Financial Source

0.906 0.012 0.904 0.906 0.904 0.973 Weighted Avg.

7.2.2.1 Discussion on the Experimental Results of IE

The performances of the classifier that are used for IE component have more or less similar
output. Among all the three classifiers used on the four different scenarios, the SMO perform
better than all others except in one scenario which Naïve Bayes very slightly exceeds the
performance of SMO. As it can be seen easily from the confusion matrix and the higher value of
precision, recall, ROC rate, TP rate in each of the scenarios, the performance of the classifiers on
the IE component is higher. Among the different features considered the token category feature
plays a crucial role for prediction by minimizing the performance of the classifier by 4% when it
is excluded as a feature. The POS has almost no effect in Naïve Bayes and SMO which has the
same performance when it is used as a feature as well while it has a significant impact on the J48
decision tree algorithm which exceeds its performance by 5%. The confusion matrix for all the
experiment output shows that the confusion in predicting the candidate text in to its exact
attribute category is very minimum which sequentially increases the detailed experimental result
of the classifier on the data set which contains 1422 instances which are the collection of the six
attributes which are the predefined attributes for extraction and instances which are not important
for extraction but have a numeric and name value.

86
The experimental result shows that using the features that are considered in this research work
and open source machine learning classifier algorithms for IE can have a good performance
compared to the other state of the art machine learning algorithms.

Summary
The experiment is conducted on economy news category obtained from ENA. Three
subcategories which are Infrastructure, Investment, and other categories are considered for text
categorization and six predefined attributes are used for IE. The experimental result shows that
classifier algorithms can be used for IE and perform as that of the other algorithms for
information extraction. Among the three different classifier algorithms that are used for
experimentation the SMO algorithm performs better than other on both text categorization and
IE. The token category feature plays a crucial role in increasing the performance of the classifier
when compared to that of POS. Since our POS performance is 80% only using another POS
which perform better in predicting the POS tag might increase the performance of the prediction.
The confusion matrix in each of the four scenarios considered for IE is small.

87
 CHAPTER EIGHT

CONCLUSION, CONTRIBUTION AND RECOMMENDATION

8.1 Conclusion
The wide spread use of computers and the tremendous growth of information and
communication technology in all parts of the world increased the amount of available textual
information from time to time. Today there are numerous electronic documents on different
issues which can help the day to day life of individuals from teenage kids to older ones.
However, the availability of huge amount of information makes it difficult to manually search
and acquire the required information from the ocean of unstructured data. The text data in local
languages is also increasing from time to time. This is also true for Amharic as there is a growth
in development and use of different online news papers and contents. To alleviate this problem
different research works have been conducted to extract relevant information automatically.

IE is one of the research areas which mainly focus on extracting relevant facts from the
abundantly available text data and convert it to the database for easy use and management of the
data. There are different researches which are conducted so far on the area on different languages
and domains. The purpose of this research work is to contribute on the development of IE system
for Amharic news text. Though it is the first research work which is conducted on the extraction
of data from Amharic news text we can confidently say that it is promising to develop an IE
system using machine learning approach.

In this research work we proposed a generic model for Amharic Text information extraction
based on which we designed an IE system which we called it as ATIE (Amharic Text
Information Extraction) system. It has four main components and is developed using java and the
open source machine learning algorithms in Weka. The components of ATIE model are
document preprocessing, text categorization, learning and extraction and post processing. The
document preprocessing component which consists of tokenization, character normalization, and
number normalization handles the language specific issues. The text categorization component
predicts the category for the unseen news texts. It is used to identify the news text which the

88
learning and extraction component use it as an input. The learning and extraction component
uses different features that are extracted by using feature extractor on the candidate text for
predicting the category of the candidate text. The feature which are used for extraction are
previous token, prefix, immediate next token, next token, candidate token, and the respective
POS for the above listed tokens. The post processing component formats the extracted data
according to the format of the predefined database slots.

For experimental evaluation we used a 10 fold stratified cross validation technique which is one
of the testing option that Weka support. By using 1200 instances of dataset for training and
evaluating the text categorization component and 1422 instances for the IE component we got
the experimental result which is very high.

8.2 Contribution of the Work

A generic model is designed for Amharic text information extraction system that uses a
machine learning approach is proposed.

The use of text categorization with IE for identification of relevant documents and
extraction of data from those relevant documents is proposed

Algorithms are developed for language specific issues which can handle number
normalization, Name Prefix Separator, Number Prefix Separator

A feature extractor algorithms which extract features from tagged dataset is designed

An algorithm for automatic token tagger which tags candidate texts as a name or a
number is designed

A prototype system for Amharic Text Information Extraction from infrastructure news
text is developed

The experiment result shows the use of open source classification algorithm for IE from
Amharic news text is promising

89
8.3 Recommendation
We showed that machine learning classifier algorithms can be used for extracting information
from Amharic news text. However, using them for large scale dataset with higher numbers of
attributes might not perform well since it increases the complexity of the classifier as IE is a very
time taking task which requires different NLP processing tools for extraction. The IE system
proposed in this research work is not complete it requires different improvements to be used at
the large scale. The following are our recommendations for future works

Co-reference is not considered in this research work as the news data that are used for
training and testing discuss only about a single issue but all the Amharic text are not like
that therefore the inclusion of Co-reference in the further study will be vital to develop a
large scale IE system.

For recognizing names the gazetteer, which contains a different names which we try to
collect from the news and other source is used. Using an automatic named entity
recognition in later stages might minimize the burden of selecting the named entities

Incorporating a POS tagger which performs better than the one used in this research work
might increase the performance of the system

Incorporating Amharic spell checkers to minimize the spelling problems which mostly
happen in the news text might also have an impact as we manually modify the spelling
errors as they have impact for named entity recognition.

90
Reference
[1] Seid Yimam, Amharic Question Answering System for Factoid Questions, A Thesis
Submitted to the School of Graduate Studies of the Addis Ababa University in partial
fulfillment for the Degree of Master of Science in Computer Science, 2009
[2] Fabio Ciravegna,Claudio Giuliano,Nicholas Kushmerick,Alberto Lavelli and Ion Muslea,
Adaptive Text Extraction and Mining (ATEM 2006),11th Conference of the European
Chapter of the Association for Computational Linguistics Proceedings of the Workshop on
April 4, 2006, Trento, Italy
[3] Jim Cowie and Yorick Wilks, Information extraction, Lecture note on Information extraction
[4] Cunningham H , Information Extraction Automatic, Encyclopedia of Language & Linguistics
journal , Second Edition, volume 5, pp.665-677, Oxford: Elsevier ,2006
[5] Shubin Zhao, Information Extraction from Multiple Syntactic Sources, A dissertation
submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy,
Department of Computer Science New York University, May 2004
[6] Christian Siefkes and Peter Siniakov, An Overview and Classification of Adaptive
Approaches to Information Extraction, LNCS Journal on Data Semantics IV, 2005.
[7] Ralph Grishman, Silja Huttunen, Roman Yangarber, Information extraction for enhanced
access to disease outbreak reports, Journal of Biomedical Informatics 35(4): 236-246
(2002)
[8] Angel janevski, Information extraction from university web pages, unpublished Ms thesis,
Graduate School of the University of Kentucky.
[9] Marin Hassel, Evaluation of Automatic Text Summarization - A practical implementation,
Licentiate thesis, Department of Numerical Analysis and Computer Science, Royal
Institute of Technology, Stockholm, Sweden, 2004
[10] Nancy Chinchor, MUC-7 Information Extraction Task Definition, http://www-
24.nist.gov/related_projects/muc/proceedings/ie_task.html, last visited March 16, 2010
[11] Jun Ying, Analysis and Comparison of Existent Information Extraction Methods, TU
Darmstadt, Knowledge Engineering Group, 2006
[12] Un Yong Nahm and Raymond J. Mooney, A Mutually Beneficial Integration of Data
Mining and Information Extraction, In the Proceedings of the Seventeenth National
Conference on Artificial Intelligence(AAAI-2000), pp.627-632, Austin, TX, 2001
91
[13] Information Extraction, http://gate.ac.uk/ie/ , last visited December 29, 2009
[14] Dipanjan Das Andre F.T. Martins, A survey on automatic text summarization, unpublished,
Language Technologies Institute Carnegie Mellon University, Nov 21, 2007
[15] Jochen L. Leidner and Chris Callison-Burch, Evaluating Question Answering Systems
Using FAQ Answer Injection, In Proceedings of the 6th Annual CLUK Research
Colloquium, 2003
[16] Christian Siefkes and Peter Siniakov, An Overview and Classification of Adaptive
Approaches to Information Extraction, LNCS Journal on Data Semantics IV, 2005.
[17] Cunningham H , Information Extraction Automatic, Encyclopedia of Language &
Linguistics journal , Second Edition, volume 5, pp.665-677, Oxford: Elsevier , 2006
[18] Ying Yu, Xiao-Long Wang, Yi Guan, Information Extraction for Chinese Free Text Based
On Pattern Match Combine With Heuristic Information, Proceedings of the First
International Conference on Machine Learning and Cybernetics, Beijing, 4-5 November
2002
[19] Line Eikvil, Information extraction from World Wide Web, a survey July 1999
[20] Hyun Chul Lee, Jian Chang Mao, Information extraction by embedding HMM to the
induced linguistic feature space, International Conference on Intelligent Computing (ICIC),
China 2005
[21] Douglas E. Appelt, David J. Israel, Introduction to Information Extraction Technology, a
Tutorial Prepared for IJCAI-99
[22] Dwi H. Widyantoro1 , Yudi Wibisono, Information Extraction for E-Job Marketplace, The
4th International Conference TSSA 2007
[23] Alberto Téllez-Valero, Manuel Montes-y-Gómez, Luis Villaseñor-Pineda, Using Machine
Learning for Extracting Information from Natural Disaster News Reports, CICLing 2005:
539-547
[24] Aidan Finn, Nicholas Kushmerick, Information extraction by convergent boundary
classification, In AAAI-2004 Workshop on Adaptive Text Extraction and Mining. San Jose,
USA, 2004
[25] Jie Tang, Mingcai Hong, Duo Zhang, Bangyong Liang, and Juanzi Li, Information
Extraction: Methodologies and Applications, In the book of Emerging Technologies of Text

92
 Mining: Techniques and Applications, Hercules A. Prado and Edilson Ferneda (Ed.), Idea
Group Inc., Hershey, USA, 2007
[26] Nicholas Kushmerick , Edward Johnston , Stephen McGuinness, Information extraction
by text categorization, In The IJCAI-2001 Workshop on Adaptive Text Extraction and
Mining
[27] An De Sitter, Walter Daelemans, In Proceedings Of The ECML/PKDD 2003 Workshop On
Adaptive Text Extraction And Mining (ATEM 2003) , Cavtat-Dubronik
[28] Yaoyong Li, Kalina Bontcheva, Hamish Cunningham (2005). SVM Based Learning System
for Information Extraction. J. Winkler, M. Niranjan and N. Lawerence (Eds.): Deterministic
and Statistical Methods in Machine Learning, LNAI 3635, Springer Verlag, 319-339
[29] Benjamin Rosendfeld, Ronen Feldman, Moshe Fresko, TEG—a hybrid approach to
information extraction, Knowledge Information Systems (2005) 00: 1–18
[30] João Cordeiro, Pavel Brazdil, Learning Text Extraction Rules, Without Ignoring Stop
Words, PRIS 2004: 128-138
[31] Rattasit Sukhahuta and Dan Smith, Information Extraction for Thai Documents,
Proceedings of the fifth international workshop on Information retrieval with Asian
languages, 103 – 110, Hong Kong, China ,2000
[32] Ying Yu, Xiao-Long Wang, Yi Guan, Information Extraction for Chinese Free Text Based
On Pattern Match Combine with Heuristic Information, In Proceedings of the First
International Conference on Machine Learning and Cybernetics, Beijing, 4-5 November
2002

[33] Yohannes Afework, Automatic Amharic text categorization, A Thesis Submitted to the
School of Graduate Studies of the Addis Ababa University in partial fulfillment for the
Degree of Master of Science in Computer Science, 2007
[34] Ibrahim Yassin Hamid , Automatic Information Extraction For Amharic Language Text
Using Hidden Markov Model , A Thesis Submitted to Graduate School of
Telecommunication and Information Technology in partial fulfillment for the Degree of
Master of Science in Information Technology,2009
[35]Amharic, http://www.lonweb.org/link-amharic.htm, Accessed on June 20,2010
[36] ባ የ ይ ማም፤ 1987 የ አ ማር ኛ ሰ ዋ ሰ ው፤ ት .መ.ማ .ማ .ድ .
93
[37] Hamish Cunningham, Information extraction automatic, Elsevier Science 18th November
2004
[38] Goncalo Simoes, Helena Galhards, Luisa Coher, Information Extraction tasks: a survey
[39] Yannick Versley, A Modular Toolkit for Coreference Resolution, a hand book on
coreference resolution
[40] Melese Tamiru,2009, Automatic Amharic Text Summarization Using Latent Semantic
Analysis, A Thesis Submitted to the School of Graduate Studies of the Addis Ababa
University in partial fulfillment for the Degree of Master of Science in Computer Science.
[41] Tessema Mindaye, 2007, Design and Implementation of Amharic Search Engine, A Thesis
Submitted to the School of Graduate Studies of the Addis Ababa University in partial
fulfillment for the Degree of Master of Science in Computer Science.
[42] The Weka Documentation, http://www.cs.waikato.ac.nz/ml/weka/, Accessed on October 22,
2010
[43] J. Han and M. Kamber. Data Mining: Concepts and Techniques (2nd ed.). Morgan
Kaufmann Publishers, 2006.
[44] Y. Yang. An Evaluation of Statistical Approaches to Text Categorization. The Netherlands:
Kluwer Academic Publishers, 1999.
[45] Summary and statistical report of the 2007 population and housing census,
www.csa.gov.et/docs/cen2007_firstdraft.pdf, accessed on November 2, 2010

94
Appendices

Appendix A: List of Stop Words

ሁሉ ብዛት ናት እንደአስረዱት የታች

ሁሉም ብዙ ናቸው እንደገና የውስጥ
ኋላ ቦታ አሁን ወቅት የጋራ
ሁኔታ በርካታ አለ እንዲሁም ያ
ሆነ በሰሞኑ አስታወቀ እንጂ ይታወሳል
ሆኑ በታች አስታውቀዋል እዚህ ይህ
ሆኖም በኋላ አስታውሰዋል እዚያ ደግሞ
ሁል በኩል እስካሁን እያንዳንዱ ድረስ
ሁሉንም በውስጥ አሳሰበ እያንዳንዳችው ጋራ
ላይ በጣም አሳስበዋል እያንዳንዷ ግን
ሌላ ብቻ አስፈላጊ ከ ገልጿል
ሌሎች በተለይ አስገነዘቡ ከኋላ ገልጸዋል
ልዩ በተመለከተ አስገንዝበዋል ከላይ ግዜ
መሆኑ በተመሳሳይ አብራርተዋል ከመካከል ጥቂት
ማለት የተለያየ አበራርተው ከሰሞኑ ፊት
ማለቱ የተለያዩ አስረድተዋል ከታች ደግሞ
መካከል ተባለ እስከ ከውስጥ ዛሬ
የሚገኙ ተገለጸ እባክህ ከጋራ ጋር
የሚገኝ ተገልጿል እባክሸ ከፊት ተናግረዋል
ማድረግ ተጨማሪ እባክዎ ወዘተ የገለጹት
ማን ተከናውኗል አንድ ወይም ይገልጻል
ማንም ችግር አንጻር ወደ ሲሉ
ሰሞኑን ታች እስኪደርስ ዋና ብለዋል
ሲሆን ትናንት እንኳ ወደፊት ስለሆነ
ሲል ነበረች እስከ ውስጥ አቶ
ሲሉ ነበሩ እዚሁ ውጪ ሆኖም
ስለ ነበረ እና ያለ መግለጹን
ቢቢሲ ነው እንደ ያሉ አመልክተዋል
ቢሆን ነይ እንደገለጹት ይገባል ይናገራሉ
ብለዋል ነገር እንደተገለጸው የኋላ
ብቻ ነገሮች እንደተናገሩት የሰሞኑ

Appendix B: List of Titles

አቶ ክቡር ሻለቃ

ወ/ሮ ክብርት ጀኔራል

ወ/ሪት ሻምበል ጀነራል

ዶ/ር ኮሎኔል ፕሮፌሰር

ሸህ አስር አለቃ ወታደር

ቄስ አምሳ አለቃ ኢንጅነር

95
ድያቆን ሀጂ አምበል
ባላምበራስ አርቲስት ኡስታዝ
ብላቴን ጌታ አፈ-ጉባኤ ኢንስትራክተር
ፊታውራሪ የተከበሩ ወ/ሮ
ብላታ አምባሳደር ኢንጂነር
አባ ኮማንደር ሰአሊ
ደጃዝማች ብርጋድየር ጀኔራል ፒያኒስት
ኰሎኔል ሌተናል ኮሌኔል ሚ/ር
ሜጀር ሹም ጠ/ሚኒስትር
ጀነራል ፕ/ር ጠ/ ሚኒስትር
በጅሮንድ አፄ ፕሬዝዳንት
መምህር መቶ አለቃ ፕረዝዳንት
ግራዝማች ሚስተር ፕሬዚዳንት
ብላቴን ጌታ ጠ/ሚ ፕሬዚደንት
ባላምባራስ ሚኒስትር ድኤታ ፕረዝደንት
ሊቀ ጠበብት ብፁእ ካፒቴን
ዶክተር ዶክተር ፓትሪያርክ
ሻንበል ተመራማሪ ፕ/ት
ነጋድራስ ከንቲባ አፈ ጉባኤ
ኰሎኔል ሊቀመንበር ማእድንና ኢነርጂ ሚኒስትር
ልኡል ምክትል ወይዘሮ
ራስ ሳጅን ጠቅላይ ሚኒስትር
አቡነ አ/አለቃ ሸክ
መምህር ከንቲባ ዋና ዳይሬክተር
አለቃ ክቡር ዳይሬክተር
ብላታ ሎሬት ኢንስፔክተር
ሀኪም ሀምሳ አለቃ
ነጋድራስ አሰልጣኝ

96
Appendix C: List of Infrastructure Names

መንገድ ጤና ጣቢያ ማሰልጠኛ ተቋም

መብራት ሆስፒታል አስፋልት መንገድ
ስልክ ትምህርት ቤት ጠጠር መንገድ
መጠጥ ዉሃ አንደኛ ደረጃ ትምህርት ቤት ኮብል ስቶን መንገድ
ንጹህ መጠጥ ዉሃ ሁለተኛ ደረጃ ትምህርት ቤ ት ሞባይል ኔትወርክ
ጤና ተቋማት ዩኒቨርሲቲ አዉቶማቲክ ስልክ
ክሊኒክ ኮሌጅ ዋይርለስ ስልክ

97
Declaration

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

Declared by:

Name: ______________________________________

Signature: ______________________________________

Date: ______________________________________

Confirmed by advisor:

Name: ______________________________________

Signature: ______________________________________

Date: ______________________________________

98