Phd Thesis - [PDF Document] (2024)

Phd Thesis - [PDF Document] (1)






Phd Thesis - [PDF Document] (2)

Phd Thesis - [PDF Document] (3)











Phd Thesis - [PDF Document] (4)

Phd Thesis - [PDF Document] (5)




I am indebted to my advisor, Shaul Markovitch, for the guidancehe afforded me during my studies. Shaul gave me plenty of freedomto explore the directions I was most interested in. At the same time,however, he constantly reminded me of the importance of focusingon the essential things and foregoing lower-hanging fruit that carrylittle benefit in the larger perspective. Shaul’s comments on paperdrafts were always to the point, and invariably led to much morecoherent and simply better papers. From our countless discussions,I adopted a useful habit of always thinking about evaluation beforelaunching a new study. Shaul also provided me with support at thetimes when my spirit was low, always insisting that there is light atthe end of the tunnel.

I am thankful to Susan Dumais and Eric Horvitz for hostingme as a summer intern at Microsoft Research. I also thank ArkadyBorkovsky and Hadar Shemtov for arranging my summer internshipat Yahoo.

I am grateful to Alex Gontmakher, my longtime friend and officeroommate. Alex offered his help on things too many to mention,from program optimization to idiosyncrasies of Linux. Alex wasalso solely responsible for persuading me to study Perl, which thenquickly became my scripting language of choice, and for spendinghours in teaching me new Perl idioms.

I am thankful to Erez Halahmi, my former manager at ComverseTechnology, for teaching me how to manage large software projects.Without his couching, I would hardly be able to write tens of thou-sands of lines of code during my Ph.D. studies, and would likely stillbe engrossed in debugging.

Phd Thesis - [PDF Document] (6)

Throughout my doctoral studies, I benefited from numerous dis-cussions with my friends and colleagues, who shared their wisdom,offered views from the perspective of their fields of expertise, andjust helped me when I needed some advice. I am grateful to allof them for their time and assistance: Gregory Begelman, BeataBeigman-Klebanov, Ron Bekkerman, Ido Dagan, Dmitry Davidov,Gideon Dror, Ofer Egozi, Saher Esmeir, Ran El-Yaniv, Anna Feld-man, Ariel Felner, Lev Finkelstein, Maayan Geffet, Oren Glickman,Yaniv Hamo, Alon Itai, Shyam Kapur, Richard Kaspersky, RonnyLempel, Dmitry Leshchiner, Irit Opher, Dan Roth, Eytan Rup-pin, Dmitry Rusakov, Fabrizio Sebastiani, Vitaly Skachek, FrankSmadja, Gennady Sterlin, Vitaly Surazhsky, Stan Szpakowicz, Pe-ter Turney, and Shuly Wintner.

I am especially thankful to my wife Lena for her unconditionallove, for putting up with the countless hours I spent on my thesiswork, and for being there when I needed it.

Finally, I am deeply beholden to my parents, Margarita Sterlinaand Solomon Gabrilovich, for their immeasurable love and support.I am also thankful to my parents for teaching me the value of knowl-edge and education. No words in any natural language would besufficient to thank my parents for all they have done for me. Thisthesis is dedicated to them.


Phd Thesis - [PDF Document] (7)

To my parents,Margarita Efimovna Sterlina and Solomon Isaakovich Gabrilovich

Phd Thesis - [PDF Document] (8)

Phd Thesis - [PDF Document] (9)

An investment in knowledge always pays the best interest.

– Benjamin Franklin

Phd Thesis - [PDF Document] (10)

Phd Thesis - [PDF Document] (11)


Abstract 1

1 Introduction 51.1 Proposed Solution . . . . . . . . . . . . . . . . . . . . . . . . . . 71.2 Contributions of This Thesis . . . . . . . . . . . . . . . . . . . . 101.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Background 132.1 Text Categorization . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.1.1 Document Features . . . . . . . . . . . . . . . . . . . . . 142.1.2 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . 162.1.3 Feature Valuation . . . . . . . . . . . . . . . . . . . . . . 192.1.4 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.1.5 Notes on Evaluation . . . . . . . . . . . . . . . . . . . . . 22

2.2 Problems With the Bag of Words Approach . . . . . . . . . . . . 222.3 Feature Generation . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3 Feature Generation Methodology 293.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293.2 Requirements on Suitable Knowledge Repositories . . . . . . . . 313.3 Building a Feature Generator . . . . . . . . . . . . . . . . . . . . 33

3.3.1 Attribute Selection . . . . . . . . . . . . . . . . . . . . . . 343.3.2 Feature Generation per se . . . . . . . . . . . . . . . . . . 34

3.4 Contextual Feature Generation . . . . . . . . . . . . . . . . . . . 353.4.1 Analyzing Local Contexts . . . . . . . . . . . . . . . . . . 363.4.2 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . 373.4.3 Feature Valuation . . . . . . . . . . . . . . . . . . . . . . 383.4.4 Revisiting the Running Example . . . . . . . . . . . . . . 38

3.5 Using Hierarchically-Structured Knowledge Repositories . . . . . 393.6 Using Knowledge Repositories that Define Arbitrary Relations Be-

tween Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

Phd Thesis - [PDF Document] (12)

4 Instantiation of the Feature Generation Methodology for theODP and Wikipedia 454.1 Using the Open Directory for Feature Generation . . . . . . . . . 45

4.1.1 Multiplying Knowledge Through Web Crawling . . . . . . 464.1.2 Noise Reduction and Attribute Selection . . . . . . . . . . 464.1.3 Implementation Details . . . . . . . . . . . . . . . . . . . 50

4.2 Using Wikipedia for Feature Generation . . . . . . . . . . . . . . 514.2.1 Wikipedia as a Knowledge Repository . . . . . . . . . . . 534.2.2 Feature Generator Design . . . . . . . . . . . . . . . . . . 544.2.3 Using the Link Structure . . . . . . . . . . . . . . . . . . 554.2.4 Implementation Details . . . . . . . . . . . . . . . . . . . 57

5 Empirical Evaluation of Feature Generation for Text Categoriza-tion 615.1 Test Collections . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.1.1 Reuters-21578 . . . . . . . . . . . . . . . . . . . . . . . . 615.1.2 20 Newsgroups (20NG) . . . . . . . . . . . . . . . . . . . 625.1.3 Movie Reviews . . . . . . . . . . . . . . . . . . . . . . . . 625.1.4 Reuters Corpus Version 1 (RCV1) . . . . . . . . . . . . . 635.1.5 OHSUMED . . . . . . . . . . . . . . . . . . . . . . . . . . 645.1.6 Short Documents . . . . . . . . . . . . . . . . . . . . . . . 655.1.7 Automatic Acquisition of Data Sets . . . . . . . . . . . . 67

5.2 Experimentation Procedure . . . . . . . . . . . . . . . . . . . . . 685.2.1 Text Categorization Infrastructure . . . . . . . . . . . . . 695.2.2 Baseline Performance of Hogwarts . . . . . . . . . . . . 705.2.3 Using the Feature Generator . . . . . . . . . . . . . . . . 70

5.3 ODP-based Feature Generation . . . . . . . . . . . . . . . . . . . 705.3.1 Qualitative Analysis of Feature Generation . . . . . . . . 705.3.2 The Effect of Feature Generation . . . . . . . . . . . . . . 765.3.3 The Effect of Contextual Analysis . . . . . . . . . . . . . 775.3.4 The Effect of Knowledge Breadth . . . . . . . . . . . . . . 775.3.5 The Utility of Feature Selection . . . . . . . . . . . . . . . 795.3.6 The Effect of Category Size . . . . . . . . . . . . . . . . . 815.3.7 The Effect of Feature Generation for Classifying Short Doc-

uments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825.3.8 Processing Time . . . . . . . . . . . . . . . . . . . . . . . 83

5.4 Wikipedia-based Feature Generation . . . . . . . . . . . . . . . . 855.4.1 Qualitative Analysis of Feature Generation . . . . . . . . 855.4.2 The Effect of Feature Generation . . . . . . . . . . . . . . 975.4.3 The Effect of Knowledge Breadth . . . . . . . . . . . . . . 995.4.4 Classifying Short Documents . . . . . . . . . . . . . . . . 99

Phd Thesis - [PDF Document] (13)

5.4.5 Using Inter-article links as Concept Relations . . . . . . . 100

6 Using Feature Generation for Computing Semantic Relatednessof Texts 103

6.1 Explicit Semantic Analysis . . . . . . . . . . . . . . . . . . . . . 103

6.2 Empirical Evaluation of Explicit Semantic Analysis . . . . . . . . 105

6.2.1 Test Collections . . . . . . . . . . . . . . . . . . . . . . . 106

6.2.2 The Effect of External Knowledge . . . . . . . . . . . . . 107

7 Related work 109

7.1 Beyond the Bag of Words . . . . . . . . . . . . . . . . . . . . . . 109

7.2 Feature Generation for Text Categorization . . . . . . . . . . . . 109

7.2.1 Feature Generation Using Electronic Dictionaries . . . . . 111

7.2.2 Comparing Knowledge Sources for Feature Generation: ODPversus WordNet . . . . . . . . . . . . . . . . . . . . . . . . 113

7.2.3 Using Unlabeled Examples . . . . . . . . . . . . . . . . . 116

7.2.4 Other Related Studies . . . . . . . . . . . . . . . . . . . . 117

7.3 Semantic Similarity and Semantic Relatedness . . . . . . . . . . . 118

8 Conclusions 123

A Text Categorization with Many Redundant Features: Using Ag-gressive Feature Selection to Make SVMs Competitive with C4.5127

A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

A.2 Experimental Methodology . . . . . . . . . . . . . . . . . . . . . 129

A.2.1 Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

A.2.2 Predicting the Utility of Feature Selection . . . . . . . . . 131

A.2.3 Extended Feature Set Based on WordNet . . . . . . . . . 132

A.2.4 Feature Selection Algorithms . . . . . . . . . . . . . . . . 133

A.2.5 Classification Algorithms and Measures . . . . . . . . . . 133

A.3 Empirical Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 134

A.3.1 Validation of Hogwarts Performance . . . . . . . . . . . 134

A.3.2 Predicting the Utility of Feature Selection with OutlierCount . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

A.3.3 Comparison of Classifiers . . . . . . . . . . . . . . . . . . 136

A.3.4 The Effect of Using Different Feature Sets . . . . . . . . . 138

A.3.5 The Effect of Using Different FS Algorithms . . . . . . . . 138

A.3.6 Testing the Relevancy of Features . . . . . . . . . . . . . 138

A.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

Phd Thesis - [PDF Document] (14)

B Parameterized Generation of Labeled Datasets for Text Catego-rization Based on a Hierarchical Directory 143B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143B.2 Parameterization of Dataset Generation . . . . . . . . . . . . . . 146

B.2.1 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147B.2.2 Properties of Individual Categories . . . . . . . . . . . . . 150

B.3 Methodology for Automatic Dataset Generation . . . . . . . . . . 151B.3.1 Acquisition of the Raw Data . . . . . . . . . . . . . . . . 151B.3.2 Filtering the Raw Data to Cope with Noise . . . . . . . . 153

B.4 Empirical Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 154B.4.1 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . 154B.4.2 Text Categorization Infrastructure . . . . . . . . . . . . . 154B.4.3 Correlation Between Distance Metrics and Text Catego-

rization Accuracy . . . . . . . . . . . . . . . . . . . . . . 155B.4.4 Correlation Between Distance Metrics and MAA . . . . . 156B.4.5 Versatility of Dataset Generation . . . . . . . . . . . . . . 157

B.5 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . 159

References 163

Hebrew Abstract i

Phd Thesis - [PDF Document] (15)

List of Figures

3.1 Standard approach to text categorization . . . . . . . . . . . . . . 303.2 Induction of text classifiers using the proposed framework for fea-

ture generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303.3 Building a feature generator . . . . . . . . . . . . . . . . . . . . . 353.4 Feature generation . . . . . . . . . . . . . . . . . . . . . . . . . . 353.5 Performing feature generation using the multi-resolution approach 363.6 Feature generation example . . . . . . . . . . . . . . . . . . . . . 393.7 Building a feature generator using hierarchically-structured knowl-

edge repositories . . . . . . . . . . . . . . . . . . . . . . . . . . . 413.8 Feature generation using a hierarchical ontology . . . . . . . . . . 423.9 Feature generation with arbitrary relations . . . . . . . . . . . . . 43

4.1 Feature generation with Wikipedia links as relations . . . . . . . . 574.2 Feature generation with Wikipedia links as relations, where only

more general features are constructed . . . . . . . . . . . . . . . . 58

5.1 Varying context length (Movies) . . . . . . . . . . . . . . . . . . . 795.2 Feature selection (Movies) . . . . . . . . . . . . . . . . . . . . . . 815.3 Feature selection (RCV1/Topic-16) . . . . . . . . . . . . . . . . . 815.4 RCV1 (Industry): Average improvement versus category size . . . 825.5 RCV1 (Topic): Average improvement versus category size . . . . 83

6.1 Knowledge-based semantic interpreter . . . . . . . . . . . . . . . . 106

A.1 Distribution of features by IG in several datasets . . . . . . . . . 132A.2 Hogwarts performance on existing datasets . . . . . . . . . . . 135A.3 Improvement in SVM accuracy at different FS levels vs. using 100%

features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135A.4 Comparison of performance of SVM, C4.5 and KNN with 100%

features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137A.5 Classification using a bag of words . . . . . . . . . . . . . . . . . 139A.6 Classification using an extended feature set . . . . . . . . . . . . 139A.7 SVM accuracy vs. FS level . . . . . . . . . . . . . . . . . . . . . . 140

Phd Thesis - [PDF Document] (16)

A.8 C4.5 accuracy vs. FS level . . . . . . . . . . . . . . . . . . . . . . 140A.9 Removing the best features by IG . . . . . . . . . . . . . . . . . . 141

B.1 Locating categories at requested text distance . . . . . . . . . . . 152B.2 SVM accuracy vs. graph distance . . . . . . . . . . . . . . . . . . 156B.3 SVM accuracy vs. text distance . . . . . . . . . . . . . . . . . . . 156B.4 Text distance vs. graph distance . . . . . . . . . . . . . . . . . . . 157B.5 MAA vs. graph distance . . . . . . . . . . . . . . . . . . . . . . . 158B.6 MAA vs. text distance . . . . . . . . . . . . . . . . . . . . . . . . 158B.7 Distribution of graph distances in ODP . . . . . . . . . . . . . . . 159B.8 Distribution of text distances in ODP (sample) . . . . . . . . . . 159

Phd Thesis - [PDF Document] (17)

List of Tables

4.1 Examples of attribute selection using information gain . . . . . . 49

5.1 Definition of RCV1 category sets used in the experiments . . . . . 655.2 Definition of OHSUMED category sets used in the experiments . . 665.3 Baseline performance of Hogwarts text categorization platform 715.4 Top ten ODP concepts generated for a sample sentence . . . . . . 765.5 Top ten ODP concepts generated for a sample sentence . . . . . . 775.6 Text categorization with and without feature generation . . . . . 785.7 Text categorization with and without feature generation, when

only a subset of the ODP is used . . . . . . . . . . . . . . . . . . 805.8 Text categorization of short documents . . . . . . . . . . . . . . . 845.9 Test collection sizes . . . . . . . . . . . . . . . . . . . . . . . . . . 855.10 The effect of feature generation . . . . . . . . . . . . . . . . . . . 985.11 Comparison of two Wikipedia snapshots . . . . . . . . . . . . . . 995.12 The effect of feature generation using a newer Wikipedia snapshot 1005.13 Feature generation for short documents . . . . . . . . . . . . . . . 1015.14 Feature generation for short documents using inter-article links . . 102

6.1 Correlation of word relatedness scores with human judgements . . 1086.2 Correlation of text relatedness scores with human judgements . . 108

A.1 Classifier accuracy at different FS levels . . . . . . . . . . . . . . . 137A.2 Statistical significance of differences in classifier accuracy . . . . . 137

Phd Thesis - [PDF Document] (18)

Phd Thesis - [PDF Document] (19)


Imagine an automatic news filtering system that tracks company news. Giventhe news item “FDA approves ciprofloxacin for victims of anthrax inhalation”,how can the system know that the drug mentioned is an antibiotic producedby Bayer? Or consider an information professional searching for data on RFIDtechnology—how can a computer understand that the item “Wal-Mart supplychain goes real time” is relevant for the search? Algorithms we present can dojust that.

When humans approach text processing tasks, such as text categorization,they interpret documents in the context of their background knowledge and ex-perience. On the other hand, conventional information retrieval systems repre-sent documents as bags of words, and are restricted to learning from individualword occurrences in the (necessarily limited) training set. We propose to enrichdocument representation through automatic use of vast repositories of humanknowledge. To this end, we use Wikipedia and the Open Directory Project, thelargest encyclopedia and Web directory, respectively. Wikipedia articles and ODPcategories represent knowledge concepts. In the preprocessing phase, a featuregenerator analyzes the input documents and maps them onto relevant concepts.The latter give rise to a set of generated features that either augment or replacethe standard bag of words. Feature generation is accomplished through contextualanalysis of document text, thus implicitly performing word sense disambiguation.Coupled with the ability to generalize from words to concepts, this approach ad-dresses the two main problems of natural language processing—synonymy andpolysemy.

Categorizing documents with the aid of knowledge-based features leverages in-formation that cannot be deduced from the training documents alone. Empiricalresults confirm that this knowledge-intensive representation brings text catego-rization to a qualitatively new level of performance across a diverse collection ofdatasets.

We also adapt our feature generation methodology for another task in nat-ural language processing, namely, automatic assessment of semantic relatednessof words and texts. Previous state of the art results are based on Latent Se-mantic Analysis, which represents documents in the space of “latent concepts”


Phd Thesis - [PDF Document] (20)

computed using Singular Value Decomposition. We propose Explicit SemanticAnalysis, which uses the feature generator methodology to represent the mean-ing of text fragments in a high-dimensional space of features based on naturalconcepts identified and described by humans. Computing semantic relatedness inthis space yields substantial improvements, as judged by the very high correlationof computed scores with human judgments.


Phd Thesis - [PDF Document] (21)

List of Abbreviations

20NG 20 NewsgroupsBOW Bag of wordsDF Document FrequencyESA Explicit Semantic AnalysisFG Feature GenerationFS Feature SelectionFV Feature ValuationHTML Hyper Text Markup LanguageIDF Inverse Document FrequencyIG Information GainIR Information RetrievalKNN K Nearest NeighborsLSA Latent Semantic AnalysisLSI Latent Semantic IndexingMAA Maximum Achievable AccuracyML Machine LearningODP Open Directory ProjectRCV1 Reuters Corpus Volume 1RCV1-v2 Reuters Corpus Volume 1 (version 2)SGML Standard Generalized Markup LanguageSVM Support Vector MachineTC Text CategorizationTF Term FrequencyTFIDF Term Frequency/Inverse Document FrequencyTREC Text REtrieval ConferenceURL Uniform Resource LocatorWWW World Wide WebWN WordNetXML eXtensible Markup Language


Phd Thesis - [PDF Document] (22)


Phd Thesis - [PDF Document] (23)

Chapter 1


Recent proliferation of the World Wide Web, and common availability of inex-pensive storage media to accumulate over time enormous amounts of digital data,have contributed to the importance of intelligent access to this data. It is thesheer amount of data available that emphasizes the intelligent aspect of access—no one is willing to or capable of browsing through but a very small subset of thedata collection, carefully selected to satisfy one’s precise information need.

The branch of Computer Science that deals with facilitating access to largecollections of data is called Information Retrieval (IR). The field of InformationRetrieval1 spans a number of sub-areas, including information retrieval per se, asperformed by users of Internet search engines or digital libraries; text categoriza-tion, which labels text documents with one or more predefined categories (possiblyorganized in a hierarchy); information filtering (or routing), which matches inputdocuments with users’ interest profiles, and question answering, which aims toextract specific (and preferably short) answers rather then provide full documentscontaining them.

Text categorization (TC) deals with assigning category labels to natural lan-guage documents. Categories come from a fixed set of labels (possibly organizedin a hierarchy) and each document may be assigned one or more categories. Textcategorization systems are useful in a wide variety of tasks, such as routing newsand e-mail to appropriate corporate desks, identifying junk email, or correctlyhandling intelligence reports.

The majority of existing text classification systems represent text as a bag ofwords, and use a variant of the vector space model with various weighting schemes(Salton and McGill, 1983). State-of-the-art systems for text categorization use a

1While the term “Information Retrieval” does not by itself imply that the information beingretrieved is hom*ogeneous (and in fact multimedia IR systems dealing with collections of soundand image files are becoming more popular), in what follows we only discuss IR applications totextual data.


Phd Thesis - [PDF Document] (24)

variety of induction techniques, such as support vector machines, k-nearest neigh-bor algorithm, and neural networks. The bag of words (BOW) method is veryeffective in easy to medium difficulty categorization tasks where the category of adocument can be identified by several easily distinguishable keywords. However,its performance becomes quite limited for more demanding tasks, such as thosedealing with small categories or short documents.

Early text categorization systems were predominantly manually crafted, andsince the advent of machine learning techniques to the field in early 1990s, sig-nificant improvements have been obtained. However, after a decade of steadyimprovements, the performance of the best document categorization systems ap-pears to have reached a plateau. No system is considerably superior to others,and improvements are becoming evolutionary (Sebastiani, 2002). In his landmarksurvey, Sebastiani (2002) even hypothesized that “[the effectiveness of automatedtext categorization] is unlikely to be improved substantially by the progress ofresearch.”

There have been various attempts to extend the basic BOW approach. Sev-eral studies augmented the bag of words with n-grams (Caropreso, Matwin, andSebastiani, 2001; Peng and Shuurmans, 2003; Mladenic, 1998b; Raskutti, Ferra,and Kowalczyk, 2001) or statistical language models (Peng, Schuurmans, andWang, 2004). Others used linguistically motivated features based on syntacticinformation, such as that available from part-of-speech tagging or shallow pars-ing (Sable, McKeown, and Church, 2002; Basili, Moschitti, and Pazienza, 2000).Additional studies researched the use of word clustering (Baker and McCallum,1998; Bekkerman, 2003; Dhillon, Mallela, and Kumar, 2003), as well as dimen-sionality reduction techniques such as LSA (Deerwester et al., 1990; Hull, 1994;Zelikovitz and Hirsh, 2001; Cai and Hofmann, 2003). However, these attemptshad mostly limited success.

We believe that the bag of words approach is inherently limited, as it can onlyuse those pieces of information that are explicitly mentioned in the documents,and only if the same vocabulary is consistently used throughout. The BOWapproach cannot generalize over words, and consequently words in the testingdocument that never appeared in the training set are necessarily ignored. Norcan synonymous words that appear infrequently in training documents be used toinfer a more general principle that covers all the cases. Furthermore, consideringthe words as an unordered bag makes it difficult to correctly resolve the sense ofpolysemous words, as they are no longer processed in their native context. Mostof these shortcomings stem from the fact that the bag of words method has noaccess to the wealth of world knowledge possessed by humans, and is thereforeeasily puzzled by facts and terms that cannot be easily deduced from the trainingset.

To illustrate the limitations of the BOW approach, consider document #15264


Phd Thesis - [PDF Document] (25)

in Reuters-21578, which is one of the most frequently used datasets in text catego-rization research. This document discusses a joint mining venture by a consortiumof companies, and belongs to the category “copper.” However, this fairly longdocument mentions only briefly that the aim of this venture is mining copper;rather, its main focus is on the mutual share holdings of the companies involved(Teck Corporation, Cominco, and Lornex Mining), as well as other mining ac-tivities of the consortium. Consequently, the three very different text classifiersthat we used (SVM, KNN and C4.5) failed to classify the document correctly.This comes as no surprise—“copper” is a fairly small category, and none of thesecompanies, nor the location of the venture (Highland Valley in British Columbia)is ever mentioned in the training set for this category. The failure of the bag ofwords approach is therefore unavoidable, as it cannot reason about the importantcomponents of the story.

We analyze typical problems and limitations of the BOW method in moredetail in Section 2.2.

1.1 Proposed Solution

In order to break through the existing performance barrier, a fundamentally newapproach is apparently necessary. One possible solution is to depart completelyfrom the paradigm of induction algorithms in an attempt to perform deep un-derstanding of the document text. Yet, considering the current state of naturallanguage processing systems, this does not seem to be a viable option (at leastfor the time being). Lacking full natural language understanding, we believethat in many cases common-sense knowledge and domain-specific knowledge maybe used to improve the effectiveness of text categorization by generating moreinformative features than the mere bag of words.

Over a decade ago, Lenat and Feigenbaum (1990) formulated the knowledgeprinciple, which postulated that “If a program is to perform a complex task well,it must know a great deal about the world it operates in.” The recognition of theimportance of world knowledge led to the launching of the CYC project (Lenatand Guha, 1990; Lenat, 1995).

We therefore propose an alternative solution that capitalizes on the powerof existing induction techniques while enriching the language of representation,namely, exploring new feature spaces. Prior to text categorization, we employa feature generator that uses common-sense and domain-specific knowledge toenrich the bag of words with new, more informative and discriminating features.Feature generation is performed automatically, using machine-readable reposito-ries of knowledge. Many sources of world knowledge have become available inrecent years, thanks to rapid advances in information processing, and Internet


Phd Thesis - [PDF Document] (26)

proliferation in particular. Examples of general purpose knowledge bases includethe Open Directory Project (ODP), Yahoo! Web Directory, and the Wikipediaencyclopedia.

Feature generation (also known as constructive induction) studies methodsthat endow the learner with the ability to modify or enhance the representationlanguage. Feature generation techniques search for new features that describethe target concepts better than the attributes supplied with the training in-stances. These techniques were found useful in a variety of machine learningtasks (Matheus, 1991; Fawcett, 1993; Markovitch and Rosenstein, 2002). A num-ber of feature generation algorithms were proposed (Pagallo and Haussler, 1990;Matheus and Rendell, 1989; Hu and Kibler, 1996; Murphy and Pazzani, 1991),which led to significant improvements in performance over a range of classificationtasks.

Feature generation methods were also attempted in the field of text cate-gorization (Kudenko and Hirsh, 1998; Mikheev, 1999; Scott, 1998). However,their application did not yield any substantial improvement over the standardapproaches. By design, these methods were mostly limited to the informationpresent in the texts to be classified. Specifically, they made little use of linguisticor semantic information obtained from external sources.

Our aim is to empower machine learning techniques for text categorizationwith a substantially wider body of knowledge than that available to a humanworking on the same task. This abundance of knowledge will to some extentcounterbalance the superior inference capabilities of humans.

In this thesis we use two repositories of world knowledge, which are the largestof their kind—the Open Directory and the Wikipedia encyclopedia. The Open Di-rectory catalogs millions of Web sites in a rich hierarchy of 600,000 categories, andrepresents the collective knowledge of over 70,000 volunteer editors. Thus, in theabove Reuters example, the feature generator “knows” that the companies men-tioned are in the mining business, and that Highland Valley happens to host a cop-per mine. This information is available in Web pages that discuss the companiesand their operations, and are cataloged in corresponding ODP categories such asMining and Drilling and Metals. Similarly, Web pages about Highland Valleyare cataloged under Regional/North America/Canada/British Columbia.Wikipedia is by far the largest encyclopedia in existence with over 1 million arti-cles contributed by hundreds of thousands of volunteers. Even though Wikipediaeditors are not required to be established researchers or practitioners, the openediting approach yields remarkable quality. A recent study (Giles, 2005) foundWikipedia accuracy to rival that of Encyclopaedia Britannica. We discuss theODP and Wikipedia in more detail in Sections 4.1 and 4.2, respectively.

To tap into this kind of knowledge, we build an auxiliary text classifier thatis capable of matching documents with the most relevant concepts of the Open


Phd Thesis - [PDF Document] (27)

Directory and Wikipedia. We then augment the conventional bag of words withnew features that correspond to these concepts. Representing documents for textcategorization in this knowledge-rich space of words and constructed featuresleads to substantially greater categorization accuracy. It is essential to mentionthat this entire scheme works fully automatically. Given a knowledge repository,the feature generator examines documents and enriches their representation in acompletely mechanical way.

We chose text categorization (TC) as the first exploration area of informationretrieval. Lewis (1992a) suggested that text categorization is more suitable forstudying feature effectiveness than text retrieval. This is because in the case ofretrieval, user requests are usually short and ambiguous, limiting the possibilitiesto experiment with different indexing terms. On the other hand, in the caseof TC, documents are long and manually classified, allowing statistical analysisof features without any additional user intervention. This allows to study textrepresentation separately from query interpretation.

In order to intuitively explain the necessity for feature construction, let usdraw a parallel with a (remotely related) field of speech processing. Speech sig-nals are usually sampled at a fixed rate of several dozen times a second, yielding afeature vector for each signal frame of 20–50 milliseconds. Obviously, understand-ing the contents of speech using these vectors alone would be a Sisyphean task.On the other hand, analyzing these feature vectors at the macro-level and com-bining them into much longer sequences allows one to achieve very good results.A similar situation occurs in image processing, where values of individual pixelsare combined into higher-level features. Of course, text words carry significantlymore meaning than speech frames or image pixels. Nevertheless, as we show inthis thesis, feature construction based on background knowledge leads to moresophisticated features that greatly contribute to the performance of automatictext processing.

It is interesting to observe that traditional machine learning data sets, suchas those available from the UCI data repository (Blake and Merz, 1998), are onlyavailable as feature vectors, while their feature set is essentially fixed. On theother hand, textual data is almost always available in raw text format. Thus, inprinciple, possibilities for feature generation are more plentiful and flexible.

Our approach is not limited to text categorization and can be applied to othertasks in natural language processing. In order to demonstrate the generality ofour approach, we also apply our feature generation methodology for assessingsemantic relatedness of natural language texts.

Prior work on semantic relatedness of words and texts was based on purelystatistical techniques that did not make use of background knowledge, such as theVector Space Model (Baeza-Yates and Ribeiro-Neto, 1999) or LSA (Deerwesteret al., 1990), as well as on using the WordNet electronic dictionary (Fellbaum,


Phd Thesis - [PDF Document] (28)

1998). Here we propose a novel method, called Explicit Semantic Analysis (ESA),for fine-grained semantic interpretation of unrestricted natural language texts.ESA uses our feature generation techniques to represent meaning of input textsin a high-dimensional space of concepts derived from the ODP and Wikipedia.The feature generator maps a text fragment into a long feature vector in thisspace. Comparing vectors in this space using any familiar distance metric (e.g.,the cosine metric (Zobel and Moffat, 1998)) allows to automatically compute thedegree of semantic relatedness between input fragments of natural language text.Empirical evaluation confirms that the use of ESA improves the existing state ofthe art in the field by 34% for computing relatedness of individual words, and by20% for longer texts.

1.2 Contributions of This Thesis

This thesis embodies several contributions.

1. We proposed a framework and a collection of algorithms that perform fea-ture generation using very large-scale repositories of human knowledge. Per-forming feature generation using external information effectively capitalizeson human knowledge encoded in these repositories, leveraging informationthat cannot be deduced solely from the texts being classified.

2. We proposed a novel kind of contextual analysis performed during featuregeneration, which views the document text as a sequence of local contexts,and implicitly performs word sense disambiguation.

3. Instantiating our feature generation methodology for two specific knowl-edge repositories, the Open Directory and Wikipedia, led to major im-provements in text categorization performance over a broad range of testcollections, breaking the existing performance plateau. Particularly notableimprovements have been observed in categorizing short documents, as wellas categories with few training examples.

4. We formulated a new approach to automatic semantic interpretation ofnatural language texts using repositories of knowledge concepts. To thisend, we used our feature generation methodology that transforms an in-put fragment of text into a high dimensional concept space. The ExplicitSemantic Analysis we proposed based on this methodology led to majorimprovements in assessing semantic relatedness of texts.

5. We also describe a way to further enhance the knowledge embedded in theOpen Directory by several orders of magnitude through crawling the WorldWide Web.


Phd Thesis - [PDF Document] (29)

1.3 Thesis Outline

The rest of the paper is organized as follows. In Section 2 we provide backgroundon text categorization and feature generation. Section 3 describes our featuregeneration methodology that uses repositories of human knowledge to overcomelimitations of the conventional bag of words approach. Section 4 instantiatesthis methodology with two particular knowledge resources, the Open DirectoryProject and the Wikipedia encyclopedia. In Section 5 we outline the implemen-tation details of our system, and report the results of evaluating the proposedmethodology empirically on a variety of test collections. Section 6 presents anapplication of our feature generation methodology to the task of automatically as-sessing the degree of semantic relatedness of natural language texts. In Section 7we discuss our methodology in the context of prior work and related literature.Section 8 concludes the thesis and outlines directions for future research.


Phd Thesis - [PDF Document] (30)


Phd Thesis - [PDF Document] (31)

Chapter 2


In this section we provide some background in a number of related areas. Sec-tion 2.1 reviews the existing approaches to text categorization, while Section 2.3presents an account of feature generation.

2.1 Text Categorization

Text categorization (TC, also known as text classification) deals with assigningcategory labels to natural language documents. Categories normally come from afixed set of labels, and may optionally be organized in a hierarchy. Under variousdefinitions, the documents may be labeled with one or many categories. If eachdocument is labeled with precisely one category, the problem is called single-labeled. If documents may be assigned either no categories or several categoriesat once, then we are dealing with multi-labeled categorization.

Research in automatic text categorization started in the 1960s, while most ar-ticles cite Maron’s work on probabilistic indexing (Maron, 1961) as the first majorwork in the field. At the beginning, many text categorization systems were builtaround manually-defined sets of rules, as exemplified by the Construe system(Hayes et al., 1990) developed for Reuters. Obviously, it is very time-consumingto acquire rules by manual labor, and moreover, such rules cannot be easily reusedacross domains (or even across data sets from the same domain that have differ-ent category focus and hence different word usage patterns). Consequently, themachine learning approach prevails, where the classifier is built automatically bylearning from a training set of documents.

In order to improve categorization accuracy, researchers occasionally augmentthe automatic induction process through some manual intervention. Most fre-quently this is done by defining additional features; for example, for a junk emailfiltering problem, Sahami et al. (1998) used a set of non-textual features such asthe time of the day the message was sent, or whether it had any files attached.


Phd Thesis - [PDF Document] (32)

In the operational text categorization setting (that is, in commercial systems),incorporation of fine-tuned manually defined rules is also occasionally used. Tocomplete this description of human involvement, we shall mention the issue ofcomprehensibility of the learned model. As in other application areas of machinelearning, humans who operate the computer tend to feel more confident aboutthe classification result if they can “understand” the way it was produced. Somesystems, notably, those using decision trees or explicitly manipulating decisionrules (e.g., Ripper (Cohen, 1995)), construct models that can be readily inter-preted by humans. In other systems, such as those built around neural networks,this issue might constitute a considerable challenge.1

Sebastiani (2002) and Yang (1999) present two very elaborate surveys in thearea of text categorization.

2.1.1 Document Features

The absolute majority of works in the field use plain language words as features.In the dichotomy defined by Sebastiani (2002), these works only use endogenousknowledge (i.e., extracted from the documents proper, as opposed to externallysupplied, or exogenous, knowledge). Whenever plain words are used as features,they may be optionally stemmed, by collapsing morphological variants to thesame indexing term. Note, however, that results on the usefulness of stemmingremain inconclusive (Sebastiani, 2002, Section 5.1). Our experimental systemhas a stemming component whose invocation is subject to run-time configura-tion. This component is based on our own, enhanced implementation of thePorter (Porter, 1980) algorithm. A number of works investigated the usefulnessof phrases (either syntactically or statistically motivated), but the results weremostly discouraging, even though intuitively one would expect that phrases docarry information important for classification.

Fuernkranz, Mitchell, and Riloff (2000) used linguistic phrases based on in-formation extraction patterns produced by the AutoSlog-TS system. Themotivation behind these features is that they are supposed to capture some ofthe syntactic structure of natural language text. For example, given a sentence“I am a student of CS at CMU”, the following features are extracted: “I am ”,“ is student”, “student of ”, and “student at ”. The main conclusion ofthis study was that linguistic features can improve the precision of TC at the lowrecall end; they do not improve precision at the high recall end, since they havevery narrow focus.

1Models built by support vector machines, which operate feature vectors in multidimensionalhyperspaces, are usually considered to be non-trivial for human interpretation. Interestingly, awork by Dumais et al. (1998), sketches a possible way of such interpretation by examining themagnitude of values in the model vectors.


Phd Thesis - [PDF Document] (33)

Dumais et al. (1998) used several kinds of NLP-derived phrases, namely,factoids (e.g., “Salomon Brothers International”, “April 8”), multi-word dictio-nary entries (e.g., “New York”, “interest rate”), and noun phrases (e.g., “first -quarter”, “modest growth”). Again, these features had no impact on classifica-tion accuracy with Naive Bayes, and even hurt slightly with SVMs.

A series of works by Mladenic and Grobelnik (Mladenic and Grobelnik, 1998b;Mladenic and Grobelnik, 1998a; Mladenic, 1998b) used Naive Bayes to classifyWeb documents collected with the aid of Yahoo! search engine. In addition toplain words, they considered n-grams (up to 5 -grams) that were built iteratively,constructing i -grams on the ith pass, and deleting infrequent features after eachiteration.

Caropreso, Matwin, and Sebastiani (2001) used a more sophisticated notionof n-grams, where each n-gram comprised an alphabetically ordered sequence ofn word stems. Using an ordered sequence of stems (with stop words removed)allowed to approximate concept indexing ; for example, expressions such as “infor-mation retrieval” and “the retrieving of information” were effectively collapsed tothe same feature. N -grams and unigrams (regular words) competed against eachother to be selected by the feature selection algorithm. The authors evaluatedthe n-grams in a so-called “learner-independent” way, by scoring the candidatefeatures with different feature selection functions rather then directly analyzingtext categorization performance. This research concluded that albeit bigrams canfrequently be better predictors of class membership (as judged by their featureselection scores such as information gain), their addition does not necessarilyimproves classification results, and sometimes may even adversely affect the per-formance.

Lewis (1992a) conjectured that phrase indexing is less effective than wordindexing and requires more features. He argued that although phrases have bettersemantic qualities (expressing more complex concepts) than plain words, they areused far less frequently, therefore, their poor statistical properties outweigh anysemantic advantages they may have.

Sahami et al. (1998) designed a system for junk email filtering, which useddomain-specific features such as the (Internet) domain from which the messagewas sent, the time of the day, the percentage of punctuation characters, or thepresence of attachments. The authors defined approximately 20 non-phrasal man-ually crafted features that “required very little person-effort to create”. Thesedomain-specific features were combined with automatically collected terms (dueto the regular bag-of-words representation), and then feature selection was per-formed using the mutual information criterion. This work further suggested theuse of domain-specific features for the TC task in general, and proposed examplesof appropriate features, such as document authors, author affiliations, and pub-lishers. It should be noted that these are actually extra-linguistic features, since


Phd Thesis - [PDF Document] (34)

they do not carry genuine linguistic knowledge, or knowledge about the domainof the texts to be classified.

Ghani et al. (2000) described a data mining system that used several types offeatures to discover new facts about public companies. The authors did not usefeature generation per se, but nevertheless their sources of features are interest-ing, and relevant to our present discussion. The database of companies compriseda collection of HTML pages from Web sites describing these companies’ activi-ties. Three kinds of features were used. Extracted features (e.g., performs-activity,officers) were obtained from the HTML pages using information extraction tech-niques, while additional extracted features (e.g., sector) were determined by NaiveBayes classification of Web pages. Wrapped features (e.g., competitor, subsidiary)resulted from a collection of wrappers developed for the Hoovers Internet databaseof companies2. The wrappers assumed a uniform format of Hoovers pages con-taining information about the companies, and extracted the values of predefinedfields. Abstracted features (e.g., same-state, reciprocally-competes) were specifiedmanually to provide the data mining algorithms with background knowledge;the values for these features were obtained from cross-referencing other features.Based on these features, the authors employed C5.0 (an improved version of C4.5(Quinlan, 1993)) to mine previously unobserved dependencies in the data. Forexample, the system could detect that many companies that offer computer soft-ware and services retain Pricewaterhouse Coopers or Ernst & Young as theirauditors.

2.1.2 Feature Selection

Term (or feature) selection is necessary to reduce noise, as well as to preventoverfitting. Some machine learning techniques exhibit inferior performance whenpresented with too many attributes3, so it is essential to select only the best ones.

Lewis (1992a) and Sebastiani (2002) note that in order to avoid overfitting,the number of training examples should be commensurate with the number offeatures; a common rule of thumb is that the number of training examples perclass should be at least ten times the number of features (Jain, Duin, and Mao,2000). Sebastiani (1999) brings a simple yet instructive example for the case ofpossible overfitting. If a classifier for category Cars for sale were trained on onlythree examples, in two of which the car sold was yellow, then the classifier mightmistakenly consider “yellowness” an essential property of this category. In thelight of this example, it would be interesting to apply knowledge-based feature

2Hoovers Online, notable example is the k-nearest neighbor (KNN) algorithm, which usually does not

weigh features differently according to their discriminative ability, and thus spurious featuressimply increase the amount of noise.


Phd Thesis - [PDF Document] (35)

selection techniques, to ascertain which features represent intrinsic properties ofthe concept to be learned.

Two main approaches for feature selection are filtering and wrapper model(John, Kohavi, and Pfleger, 1994). The filtering approach receives a set of fea-tures, and filters it independently from the induction algorithm. The wrappermodel searches for good feature subsets, and evaluates them using n-fold cross-validation on the training data. This scheme may be used in conjunction withany induction algorithm, which is used for evaluating feature subsets on the val-idation set. The search for feature subsets can be performed using simple greedyalgorithms such as backward elimination or forward selection, or more complexones that can both add and delete features at each step.

Since the wrapper model requires much more computation, filtering is themore common type of feature selection. This is especially true in the domainof textual information retrieval, where using the bag-of-words model results in ahuge number of features. A number of feature selection techniques were describedin the TC literature, while Yang and Pedersen (1997) found document frequency(DF), information gain (IG) and χ2 (CHI) to be the most effective (reducing thefeature set by 90-98% with no performance penalty, or even a small performanceincrease due to removal of noise). Yang and Pedersen (1997) also observed thatcontrary to a popular belief in information retrieval that common terms are lessinformative, document frequency, which prefers frequent terms (except for stopwords), was found to be quite effective for text categorization.

Feature selection may be either local, resulting in category-specific features,or global, yielding collection-wide features. Document frequency is immediatelysuitable for global feature selection, while in order to adapt information gain andχ2 to global operation, the sum of scores, weighted average or the maximum score(over categories) can be used.

When category-specific features are used, a problem may arise for categorieswith little training data. When only a few documents are available for a category,the number of candidate features is simply too small, so even if all of them areselected, document vectors in this category may be extremely sparse (and somemay even be empty). To overcome this problem, one can unconditionally selectsome minimum number of features regardless of their actual scores, or “back off”to global features whenever an insufficient number of local features are available.

Recently, Joachims (1998) argued that support vector machines are very ro-bust even in the presence of numerous features. He further claimed that themultitude of text features are indeed useful for text categorization. To substan-tiate this claim, Joachims used the Naive Bayes classifier with feature sets ofincreasing size, where features were first ordered by their discriminative capac-ity (as predicted by the information gain criterion), and then the most powerful


Phd Thesis - [PDF Document] (36)

features were removed. The classifier trained on these “low-utility” features per-formed markedly better than random assignment of categories to documents, thusimplying that all features are relevant and should be used. Based on these find-ings, many later works using SVMs did not perform any feature selection at all(Leopold and Kindermann, 2002; Lewis et al., 2004). At the same time, othersachieved very decent results while using some form of feature selection (notably,the result reported by Dumais et al. (1998) is considered the best one for theReuters-21578 collection), so the evidence in this respect remains inconclusive.4

Observe that filtering techniques ignore mutual dependence between features,even though features are usually not completely independent. In the domain oftext categorization, where features are plain words, there is naturally a consid-erable dependence among them. One simple approach to address this issue hasbeen proposed by Soucy and Mineau (2001). They first select a small numberof features according to the information gain criterion, and then further selectonly those features that both have high information gain (above some prede-fined threshold) and do not co-occur too often with the features already selected.Another feature selection technique recently proposed for text categorization ad-dresses a situation that arises in processing huge data sets, such as Reuters CorpusVolume 1 (Lewis et al., 2004). In such cases, there exists a trade-off between thesize of the feature space and the amount of training documents that can be usedfor learning. To circumvent this problem, Brank et al. (2002) proposed to firsttrain a linear SVM classifier in the full feature space using only a fraction of thetraining data, then use the trained model to rank the features and retain onlythe best ones. This way, feature selection involves examining the normal vectorto the hyperplane separating the classes, and removing features that correspondto the vector components with low absolute values (since they have less impacton the classification outcome than those with high values). Finally, a new modelis trained (using either the same or a different classifier), which only makes useof the features selected in the previous step, but now taking advantage of all thetraining data.

In addition to the “principled” feature selection schemes described above, twoadditional steps are frequently performed, namely, removal of stop words5 and

4In our own experiments, using 10% of features instead of the entire feature set has littleeffect for Reuters-21578 (-0.1% . . . +0.3% depending on the category set), and a small positiveeffect of 1.3% for the Movie Reviews data set (Pang, Lee, and Vaithyanathan, 2002). For the20 Newsgroups collection (Lang, 1995), using all the features improves SVM results by as muchas 4.7%, but this is due to the particular nature of newsgroup postings, which exhibit a verylarge and diversified vocabulary.

5In an application of text classification techniques to information extraction, where the taskwas to estimate the relevancy of extracted patterns to various categories, Riloff (1995) foundthat stemming and removal of function words (e.g., prepositions) may harm the performance ofTC considerably. This happens because specific words forms may be more characteristic than


Phd Thesis - [PDF Document] (37)

words occurring in the collection less than some predefined number of times (orless than in some predefined number of documents). For the former task, manyresearchers adopted the stop word lists developed by Lewis (1992b) and Salton(1971).

2.1.3 Feature Valuation

After the features have been selected, they need to be assigned values for eachdocument vector. This step is commonly known as feature valuation or termweighting. Numerous term weighting schemes are available, while most can bedescribed as particular cases of the tf.idf family introduced by Salton and Buckley(1988) in the SMART project.

Each scheme can be represented as a triple of parameters XY Z, where Xstands for the term frequency factor, Y for the document frequency, and Z forthe normalization method. A list of the most frequently used schemes is givenbelow6, and further details are available in (Hersh et al., 1994; Salton and Buckley,1988; Singhal, 1998; Manning and Schuetze, 2000, pp.541–544).

The following definitions describe the weighting of term tk in document dj.We use N to denote the total number of documents in the collection, count(tk, dj)— the number of times tk occurs in dj (term frequency), and dfk — the numberof documents in the collection that contain tk (document frequency).

• X (term frequency)

– l (logarithmic) = 1 + log(count(tk, dj))

– L (log-average) = 1+log(count(tk,dj))


where tfavg(dj) — average term frequency in the document

– a (augmented) = 0.5 + 0.5∗count(tk,dj)


where tfmax(dj) = maxi count(ti, dj) — maximum term frequency inthe document

– n (natural) = count(tk, dj)

– b (binary) =

{1, if tk ∈ dj

0, otherwise

• Y (document frequency)

– t (inverse document frequency) = log( Ndfk


others of particular categories. Manning and Schuetze (2000) also note that “little” words oftenprove useful for the task of author identification (which can be easily cast as a classificationproblem).

6All these schemes are implemented in our Hogwarts text categorization system.


Phd Thesis - [PDF Document] (38)

– n (natural) = 1.0

• Z (normalization)

– c (cosine) = 1√∑i(weight of termi)2

– u (pivoted unique normalization) =1

(1−slope)+slope∗ number of unique words in the documentaverage number of unique words per document

– n (no normalization)

Of these, the ltc scheme has been found the most effective (Yang, 1999; Sebas-tiani, 2002). Putting everything together, ltc stands for logarithmic weighting ofoccurrence counts (l), inverse document frequency (t), and cosine normalization(c):

tfidf(tk, dj) = tf(tk, dj) · logN




tf(tk, dj) =

{1 + log count(tk, dj), if count(tk, dj) > 00, otherwise


Finally, cosine normalization is applied to tf.idf weights to disregard differencesin document length, by weighting all components of the feature vector as follows:


tfidf(tk, dj)√∑rs=1 tfidf(ts, dj)2


where r is the number of selected features.

2.1.4 Metrics

Following (Yang, 1999), we define text categorization performance measures usingthe following two-way contingency table:

Classification = Yes Classification = No

Correct = Yes a bCorrect = No c d

Then the following metrics can be introduced:

• precision: p = a/(a + c) (undefined when a + c = 0);

• recall: r = a/(a + b) (undefined when a + b = 0);


Phd Thesis - [PDF Document] (39)

• fallout: f = c/(c + d) (undefined when c + d = 0);

• accuracy: Acc = (a + d)/n, where n = a + b + c + d > 0;

• error: Err = (b + c)/n, where n = a + b + c + d > 0.

Historically, accuracy and error are the standard metrics used in machinelearning experiments. However, these metrics are hardly suitable for most textcategorization applications. In real-life text collections there are many more neg-ative examples than positive ones (usually, by orders of magnitude). Therefore,since accuracy and error have the total number of examples in the denominator,they are very insensitive to changes in true classification performance, and com-monly produce very large (in case of accuracy) and very small (in case of error)values (fallout suffers from a similar problem).

The basic metrics commonly used to evaluate text categorization performanceare precision and recall, taken from the mainstream information retrieval research.When it is desirable to visualize the performance of a TC system, the value ofprecision may be plotted for a number of values of recall; borrowing from electricalengineering terminology, the resulting graph is usually called a receiver operatingcurve (ROC).

Sometimes, it is convenient to have a single measure instead of two, in whichcase the F -measure (van Rijsbergen, 1979, Chapter 7) may be used:

Fβ(p, r) =(β2 + 1)pr

β2p + r,

where p and r denote precision and recall, respectively.The β parameter allows fine-tuning the relative importance of precision over

recall. When both metrics are equally important, the F1 measure is used: F1 =2pr/(p + r).

The precision-recall Break-Even Point (BEP) is occasionally used as an al-ternative to F1. It is obtained by either tuning the classifier so that precisionis equal to recall, or sampling several (precision, recall) points that bracket theexpected BEP value and then interpolating (or extrapolating, in the event thatall the sampled points lie on the same side).

In the presence of two or more categories, it is handy to have a single value thatreflects the overall performance. In such a case, either micro-averaging or macro-averaging of category-specific performance measures is used. The former accu-mulates a, b, c, d values over the documents in all categories, and then computesprecision and recall (as well as F1 and BEP) values. The latter simply averagesprecision and recall computed for each category individually.7 Macro-averaging

7In the case of macro-averaging, there is an issue of evaluating the ratio 0/0 that can resultin some underpopulated categories. Various approaches set it to either 0, 1, or some very small


Phd Thesis - [PDF Document] (40)

ignores the relative sizes of categories, which may or may not be appropriate ineach case at hand. Therefore, although most researchers report micro-averagedmetrics, others perform both types of averaging.

2.1.5 Notes on Evaluation

It should be remembered that there is an upper limit on the performance oftext categorization systems, as even humans occasionally disagree on assignmentof categories to documents. This is a manifestation of inter-indexer inconsis-tency (Sebastiani, 2002) common in information retrieval. Rose, Stevenson, andWhitehead (2002) studied this phenomenon in depth, analyzing the consistencyof classification of Reuters news items by a group of Reuters editors. They foundthe inter-editor consistency to be quite high (about 95%), but still not 100%.Observe also that this high correlation was probably due to the fact that Reutersemploys seasoned information professionals, whose judgement is further boundby strict in-house policies on labeling news stories; therefore, in less constrainedcirc*mstances the agreement among humans would probably be lower.

2.2 Problems With the Bag of Words Approach

1. Words that appear in testing documents but not in training documents arecompletely ignored by the BOW approach. Since the classification model isbuilt with a subset of words that appear in the training documents, wordsthat do not appear there are excluded by definition. Lacking the abilityto analyze such words, the system may overlook important parts of thedocument being classified.

Example: Document #15264 from Reuters-21578 described in the Intro-duction presents a perfect example of this limitation. This document de-scribes a copper-mining venture formed by a group of companies, whosenames are not mentioned even once in the training set, and are thus ig-nored by the classification model.

2. Words that appear infrequently in the training set, or appear just once, aremostly ignored even if they are essential for proper classification. It oftenhappens that human annotators assign a document to a certain categorybased on some notion briefly mentioned in the document. If the wordsthat describe this notion do not appear with sufficient frequency elsewherein the training set, then the system will overlook the real reason for this

positive value. Observe that the different decisions can adversely affect the overall averagedperformance score.


Phd Thesis - [PDF Document] (41)

document’s annotation. Consequently, it will either come up with somespurious association between the actual category and unrelated words orignore this document as a training example altogether.

Example: Suppose we have a collection of pharmaceutical documents andare trying to learn the concept of antibiotics. If a particular training docu-ment describes the results of a clinical trial for a new antibiotic drug, andmentions it only by a brand name that does not appear elsewhere in thetraining set, the system will likely miss this important piece of evidence.

3. The problem described in the previous item can manifest itself in a moreextreme way. Suppose we have a group of related words, where each wordappears only a few times in the collection, and few documents containmore than one word of the group. As a result, the connection betweenthese words remains implicit and cannot be learned without resorting toexternal knowledge. External knowledge, however, allows us to determinethat certain words are related. Furthermore, we can use the generalizationability of hierarchical knowledge organization to establish that the wordscorrespond to specific instances of the same general notion.

Example: Consider a collection of clinical narrative reports on administer-ing various antibiotic drugs. Since such reports are circulated among med-ical professionals, they are likely to refer to specific drugs by name, whileomitting the knowledge already shared by the target audience. Hence, thereports will likely not explain that each drug is actually an antibiotic. Inthe absence of this vital piece of knowledge, the BOW approach can easilyfail to learn the notion shared by the reports.

4. A critical limitation of the BOW approach lies in its ignorance of the con-nections between the words. Thus, even more difficult than the problemdescribed in the previous item, is the one where we have several relatedphrases or longer contexts, while the connection between them is not statedin any single document.

Example: Consider again a collection of clinical reports, which are in-herently rich in diverse medical terminology. Often, each report describesthe case of a single patient. Thus, without extensive medical knowledge itwould be nearly impossible to learn that Lown-Ganong-Levine Syndromeand Wolff-Parkinson-White Syndrome are different kinds of arrhythmia,while Crigler-Najjar Syndrome and Gilbert Syndrome are two kinds of liverdiseases.

5. Because contextual adjacency of words is not taken into account by theBOW approach, word sense disambiguation can only be performed at the


Phd Thesis - [PDF Document] (42)

level of entire documents, rather than at much more linguistically plausiblelevels of a single sentence or paragraph.

Example: As an extreme example of this limitation, consider a documentabout the Jaguar company establishing a conservation trust to protect itsnamesake8 animal. This fairly long document is devoted mainly to thepreservation of wildlife, while briefly covering the history of the car manu-facturer in its last paragraph. Taken as a single bag of words, the documentwill likely be classified as strongly related to jaguar the animal, while thecursory mention of Jaguar the company will likely be ignored.

Some of these limitations are due to data sparsity—after all, if we had infi-nite amounts of text on every imaginable topic, the bag of words would performmuch better. Many studies in machine learning and natural language processingaddressed the sparsity problem. Simple approaches like smoothing (Chen andGoodman, 1996) allocate some probability mass for unseen events and thus elim-inate zero probabilities. Although these approaches facilitate methods that aresensitive to zero probabilities (e.g., Naive Bayes), they essentially do not intro-duce any new information. More elaborate techniques such as transfer learning(Bennett, Dumais, and Horvitz, 2003; Do and Ng, 2005; Raina, Ng, and Koller,2006) and semi-supervised learning (Goldberg and Zhu, 2006; Ando and Zhang,2005a; Ando and Zhang, 2005b), leverage cooccurrence information from similarlearning tasks or from unlabeled data. Other studies that addressed the sparsityproblem include using the EM algorithm with unlabeled data (Nigam, McCal-lum, and Mitchell, 2006; Nigam et al., 2000), latent semantic kernels (Cristianini,Shawe-Taylor, and Lodhi, 2002), transductive inference (Joachims, 1999b), andgeneralized vector space model (Wong, Ziarko, and Wong, 1985).

Humans avoid these limitations due to their extensive world knowledge, aswell as their ability to understand the words in context rather than just viewthem as an unordered bag. In this thesis we argue that the limitations of the bagof words can be overcome by endowing computers with access to the wealth ofhuman knowledge. Recall the sample Reuters document we considered above—when a Reuters editor originally handled this news item, she most likely knewquite a lot about the business of the companies mentioned, and based on her deepdomain knowledge she easily assigned the document to the category “copper.” Itis this kind of knowledge that we would like machine learning algorithms to haveaccess to.

8 events/archive/Jaguar Conserva-tion trust longcopy.htm


Phd Thesis - [PDF Document] (43)

2.3 Feature Generation

Feature generation (FG), also known as feature construction, constructive induc-tion or bias shift, is a process of building new features based on those present inthe examples supplied to the system, possibly using the domain theory (i.e., in-formation about goals, constraints and operators of the domain) (Fawcett, 1993).Feature construction techniques can be useful when the attributes supplied withthe data are insufficient for concise concept learning.

Matheus (1991) proposed to use constructive induction to address the prob-lems of disjunctive regions in the instance space (i.e., discontinuous concepts). Heposed the following issues as the main questions of constructive induction, andsuggested approaches to answer them from the instance-based, hypothesis-based,and knowledge-based points of view.

1. When should new features be constructed?

• Instance-based detection: estimate irregularity of the membershipfunction from the distribution of observed instances.

• Hypothesis-based detection: an initial hypothesis fails to meet someperformance criterion.

• Knowledge-based detection: use domain knowledge.

2. What constructive operators should be used and which of the existing fea-tures should they be applied to?

• Instance-based selection: search for patterns among training instances.

• Hypothesis-based selection: for example, analyze the structure ofbranches in decision trees.

• Knowledge-based selection: use domain theory.

3. Which (if any) features should be discarded?

• Instance-based evaluation: use probabilistic or information-theoreticmeasures.

• Hypothesis-based evaluation: for example, rank features according tohow well they are used within the hypothesis.

• Knowledge-based evaluation: Domain knowledge is generally used forthe generation rather than evaluation of features. Nevertheless, fea-tures can be evaluated according to how well they conform to thedomain knowledge.


Phd Thesis - [PDF Document] (44)

Feature generation has been studied in a number of works over the recentyears, and several notable algorithms have been proposed. Most prominent ex-amples include the FRINGE (Pagallo and Haussler, 1990), CITRE (Matheusand Rendell, 1989) and GALA (Hu and Kibler, 1996) algorithms that manip-ulate boolean combinations of features, the ID2-of-3 (Murphy and Pazzani,1991) algorithm that uses M-of-N concepts, various genetic algorithms that ap-ply crossover and mutation operations to feature bit strings, and the FICUSalgorithm (Markovitch and Rosenstein, 2002) that generalizes over previous ap-proaches by using constructor functions and searches the space of generated fea-tures.

Callan (1993) suggested a feature generation technique suitable for searchdomains (e.g., the n-queens problem). He observed that “search goal descriptionsare usually not monolithic, but rather consist of subexpressions, each describinga goal or constraint”. Taken together they characterize the goal state, but theycan also be used independently to measure progress in reaching the goal. Thiswork proposed a set of heuristics for decomposing goal specifications into theirconstituent parts, in order to use them as features.

Classical approaches to feature generation belong to two main classes:

• Data-driven, in which new features are created by combining existing fea-tures in various ways. Feedback from the learned concept is typically usedto suggest plausible feature combinations.Limitations: The amount of the improvement is limited, since the algo-rithm starts with the example features, and combines features one step ata time. If useful feature are complex combinations of example features, thesystem will have to generate and test prohibitively many features until auseful one is derived.

• Analytical, using domain theory to deduce appropriate new features.Using information about the domain helps create complex features in onestep.Limitations: such systems can only create features that follow deductivelyfrom the domain theory. Many real-world domains require useful featuresthat are not deducible from the domain theory, and these analytical systemsare incapable of deriving them.

Fawcett (1991) proposed a hybrid theory of feature generation, so that usefulfeatures can be derived from abstractions and combinations of abstractions ofthe domain theory. Abstractions are created by relaxing conditions specified ina domain theory, using a hybrid of data-driven (bottom-up) and theory-driven(top-down) approaches.

It is important to distinguish between feature generation and feature selection.While the former attempts to construct new features not present in the original


Phd Thesis - [PDF Document] (45)

description of the data, the latter starts with a set of features and attempts todecimate it. Feature selection is frequently used in tandem with feature genera-tion. Some approaches to feature generation employ a generate and test strategy,where a set of new features is created, which is then filtered according to somefitness criterion, then another FG iteration is performed using the original andselected constructed features, and so on. Using all the generated features with-out pruning the set heavily after each iteration may result in a combinatorialexplosion of the number of features.


Phd Thesis - [PDF Document] (46)


Phd Thesis - [PDF Document] (47)

Chapter 3

Feature Generation Methodology

In Section 2.2 we discussed a number of problems with the BOW approach. Wenow proceed to developing a feature generation methodology that will addressand alleviate these problems using repositories of human knowledge.

3.1 Overview

The proposed methodology allows principled and uniform integration of one ormore sources of external knowledge to construct new features. These knowledgesources define a collection of concepts that are assigned to documents to qualifytheir text. In the preprocessing step, we build a feature generator capable ofrepresenting documents in the space of these concepts. The feature generator isthen invoked prior to text categorization to assign a number of relevant conceptsto each document. Subsequently, these concepts give rise to a set of constructedfeatures that provide background knowledge about the document’s contents. Theconstructed features can then be used either in conjunction with or in place of theoriginal bag of words. The resulting set optionally undergoes feature selection,and the most discriminative features are retained for document representation.

We use traditional text categorization techniques to learn a text categorizer inthe new feature space. Figure 3.1 depicts the standard approach to text catego-rization. Figure 3.2 outlines the proposed feature generation framework; observethat the “Feature generation” box replaces the “Feature selection” box framedin bold in Figure 3.1.


Phd Thesis - [PDF Document] (48)

Labeled documents

Feature valuation

Induction algorithm


labeled feature vectors

Feature selection

Basic features



Testing documents

Feature valuation

Classifier Classified documents

Selected features

Figure 3.1: Standard approach to text categorization

Labeled documents

Feature valuation

Induction algorithm


labeled feature vectors

Feature generation

Feature construction

Feature selection

Generated features



General purpose


Figure 3.2: Induction of text classifiers using the proposed framework for featuregeneration


Phd Thesis - [PDF Document] (49)

3.2 Requirements on Suitable Knowledge


We impose the following requirements on knowledge repositories for feature gen-eration:

1. The repository contains a collection of concepts, which are defined byhumans and correspond to notions used by humans in commonsense ordomain-specific reasoning. Formally, let KR be a knowledge repositorythat contains concepts C = {c0, . . . , cn}.

2. There is a collection of texts associated with each concept. The featuregenerator uses these texts to learn the definition and scope of the concept,in order to be able to assign it to relevant documents. We refer to thesetexts as textual objects, and denote the set of such objects associated withconcept ci as Ti = {t0,1, . . . , ti,mi


3. Optionally, there is a collection of relations between concepts, R ={r1, . . . , rl}, where each relation is a set of pairs of concepts, rk = {〈ci, cj〉}.For example, one such relation could be a generalization (“is-a”) relation,which organizes the concepts into a hierarchical structure. In what follows,Section 3.5 discusses the extension of our methodology to hierarchically-structured knowledge bases, and Section 3.6 discusses the use of arbitraryrelations.

Let W be a set of words that appear in documents to be classified. Our goalis to build a mapping function f : W ∗ → 2C . We propose building the mappingfunction using text categorization techniques. This is a very natural thing todo, as text categorization is all about assigning documents or parts thereof to apredefined set of categories (concepts in our case). One way to do so is to usea binary learning algorithm L(Pos,Neg) to build a set of n binary classifiers,f0, . . . , fn, such that fi : W ∗ → {0, 1}. This way, individual classifiers are builtusing the chosen learning algorithm: fi = L(Ti,

⋃0≤j≤n,j 6=i Tj). Another way to

build such a mapping function is to devise a hierarchical text classifier that takesadvantage of the hierarchical organization of categories. In this paper, we use asimpler approach of building a single classifier that simultaneously considers allthe concepts for each input sequence of words.

We believe that the above requirements are not overly restrictive. In fact,there are quite a few sources of common-sense and domain-specific knowledgethat satisfy these requirements. We list below several notable examples.


Phd Thesis - [PDF Document] (50)

• Internet directories such as the Yahoo Web Directory1, the Open Direc-tory Project2 and the LookSmart directory3 catalog huge numbers of URLsorganized in an elaborate hierarchy. The Web sites pointed at by theseURLs can be crawled to gather a wealth of information about each direc-tory node. Here each directory node defines a concept, and crawling theWeb sites cataloged under the node provides a collection of textual objectsfor that node.

• The Medical Subject Headings (MeSH) taxonomy (MeSH, 2003), which de-fines over 18,000 categories and is cross-linked with the MEDLINE databaseof medical articles, is a notable example of a domain-specific knowledgebase. The MEDLINE links allow to easily associate MeSH nodes with nu-merous scientific articles that are highly relevant to the node, yielding a setof textual objects for that node.

• Other domain-specific knowledge repositories are also available, notably inthe terminology-rich law domain, which includes the KeySearch taxonomyby WestLaw4 and the Web-based FindLaw hierarchy5 (both of them cross-linked with material relevant for each node).

• The US Patent Classification6 and the International Patent Classification7

are exceptionally elaborate taxonomies, where each node is linked to rele-vant patents.

• The online Wikipedia encyclopedia8 has a fairly shallow hierarchy but itsnodes contain very high-quality articles, which are mostly noise-free (exceptfor occasional spamming).

• In the brick-and-mortar world, library classification systems such as theUniversal Decimal Classification (UDC) (Mcilwaine, 2000), the Dewey Dec-imal Classification (Dewey et al., 2003) or the Library of Congress Classifi-cation (Chan, 1999) provide structuring of human knowledge for classifyingbooks. By the very virtue of their definition, each classification node can beassociated with the text of books cataloged under the node. Interestingly,



Phd Thesis - [PDF Document] (51)

modern book-scanning efforts such as those underway at Google and Ama-zon can eventually make it possible to build feature generators powered bythe knowledge available in printed books.

In this work we use the ODP and Wikipedia as our knowledge repositories,due to the easy accessibility of their data on the Web. In the next section, we shalldiscuss the instantiation of our methodology for these two knowledge repositories.However, our methodology is general enough to facilitate other knowledge sourcessuch as those listed above, and in our future work we intend to explore their utilityas well, focusing in particular on the MeSH hierarchy for domain-specific featuregeneration.

A note on terminology is in order here. The most commonly used term fornodes of directories of knowledge is “category.” In text categorization, however,this term normally refers to topical labels assigned to documents. To preventpossible confusion, we use the word “concept” to refer to the former notion. Werepresent such concepts as vectors in a high-dimensional space of “attributes.”Again, we avoid using the term “features,” which is reserved for denoting indi-vidual entries of document vectors in text categorization per se.

3.3 Building a Feature Generator

The first step in our methodology is preprocessing, performed once for all futuretext categorization tasks. In the preprocessing step we induce a text classifierthat maps pieces of text onto relevant knowledge concepts, which later serve asgenerated features. The resulting classifier is called a feature generator accordingto its true purpose in our scheme. The feature generator represents concepts asvectors of their most characteristic words, which we call attributes (reserving theterm features to denote the properties of documents in text categorization).

The feature generator operates similarly to a regular text classifier—it firstlearns a classification model in the space of concept attributes, and then identi-fies a set of concepts that are most appropriate to describe the contents of theinput text fragment. Observe that the number of concepts to which the featuregenerator can classify document text is huge, as suitable knowledge repositoriesmay contain tens and even hundreds of thousands of concepts. Few machinelearning algorithms can efficiently handle so many different classes and about anorder of magnitude more of training examples. Suitable candidates include thenearest neighbor and the Naive Bayes classifier (Duda and Hart, 1973), as well asprototype formation methods such as Rocchio (Rocchio, 1971) or centroid-based(Han and Karypis, 2000) classifiers.


Phd Thesis - [PDF Document] (52)

3.3.1 Attribute Selection

Prior to learning a text classifier that will act as feature generator, we representeach concept as an attribute vector. To this end, we pool together all the textualobjects for the concept, and represent the accumulated description with a vectorof words. Using all encountered words as attributes is impractical because it yieldsa classification model that is too big, and because this would inevitably increasethe level of noise. The former consideration is essential to allow fitting the inducedmodel into computer memory. The latter consideration is particularly importantfor Web-based knowledge repositories, which are inherently plagued with noiseranging from intentional directory spamming to merely irrelevant information.To remedy the situation, we perform attribute selection for each concept prior tolearning the feature generator.

To this end, we use standard attribute selection techniques (Sebastiani, 2002)such as information gain, and identify words that are most characteristic of a con-cept versus all other concepts. This approach to attribute selection is reminiscentof the approaches described by Chakrabarti et al. (1997) and by Koller and Sa-hami (1997). Let us denote by Di the collection of textual objects associatedwith concept ci, Di =


ti,k, and by Di the collection of textual objects forall other concepts, Di =

⋃l=0...n,l 6=i


tl,k. Then, we can assess the discrim-inative capacity of each word w ∈ Di with respect to Di. It is essential to notethat conventional attribute selection techniques select attributes for ci from theentire lexicon, Di ∪Di. In our case, however, we aim at selecting words that aremost characteristic for the concept, and therefore we limit the selection only towords that actually appear in the textual objects for that concept, that is, Di.

Figure 3.3 shows the algorithm for building a feature generator. The algorithmuses a global structure Text(ci) that accumulates textual objects for concept ci

(attributes for the concept are then selected from the words occurring in thispool). We manipulate Text(ci) as an unordered bag of words. Attribute vectorsfor each category are stored in V ector(ci).

3.3.2 Feature Generation per se

Given a fragment of text for which we desire to generate features, we representit as an attribute vector, and then compare it to the vectors of all knowledgeconcepts. The comparison can use any distance metric for comparing vectorsin a high-dimensional space; in this work, we use the cosine metric. The de-sired number of highest-scoring concepts are then returned as generated features.Figure 3.4 outlines this process.


Phd Thesis - [PDF Document] (53)

Algorithm BuildFeatureGenerator# Compute attribute vectors for all conceptsBuildVectors()

# Use an induction algorithm to train a feature generator FG# using the attribute vectors V ector(ci)FG ← InduceClassifier({V ector(ci)})

———————————————————————————————-Algorithm BuildVectors()

For each ci ∈ C = {c0, . . . , cn} doText(ci) ← ⋃


# Build the attribute vector by performing attribute selection# among the words of Text(ci)V ector(ci) ← AttributeSelection(Text(ci))# Assign values to the selected attributesV ector(ci) ← tfidf(V ector(ci))

Figure 3.3: Building a feature generator

Algorithm FG(text, distanceMetric, numConcepts)TextV ector ← tfidf(text)For each ci ∈ C = {c0, . . . , cn} do

Score(ci) ← distanceMetric(TextV ector, V ector(ci))Let GeneratedConcepts be a set of numConcepts concepts

with the highest Score(ci)Return GeneratedConcepts

Figure 3.4: Feature generation

3.4 Contextual Feature Generation

Feature generation precedes text categorization, that is, before the inductionalgorithm is invoked to build the text categorizer, the documents are fed to thefeature generator.

Traditionally, feature generation uses the basic features supplied with thetraining instances to construct more sophisticated features. In the case of textprocessing, however, important information about word ordering will be lost ifthe traditional approach is applied to the bag of words. Therefore, we argue that


Phd Thesis - [PDF Document] (54)

Algorithm ContextualFeatureGeneration(D)Let CT be a series of contexts for DCT ← words(D) ∪ sentences(D) ∪ paragraphs(D) ∪ {D}Let F be a set of features generated for DF ← ∅For each context ct ∈ CT perform feature generation:

F ← F ∪ FG(ct)Represent D as BagOfWords(D) ∪ F

Figure 3.5: Performing feature generation for document D using the multi-resolutionapproach

feature generation becomes much more powerful when it operates on the rawdocument text. But should the generator always analyze the whole document asa single unit, as do regular text classifiers?

3.4.1 Analyzing Local Contexts

We believe that considering the document as a single unit can often be misleading:its text might be too diverse to be readily mapped to the right set of concepts,while notions mentioned only briefly may be overlooked. Instead, we propose topartition the document into a series of non-overlapping segments (called contexts),and then generate features at this finer level. Each context is classified into anumber of concepts in the knowledge base, and pooling these concepts togetherto describe the entire document results in multi-faceted classification. This way,the resulting set of concepts represents the various aspects or sub-topics coveredby the document.

Potential candidates for such contexts are simple sequences of words, or morelinguistically motivated chunks such as sentences or paragraphs. The optimalresolution for document segmentation can be determined automatically using avalidation set. We propose a more principled multi-resolution approach that si-multaneously partitions the document at several levels of linguistic abstraction(windows of words, sentences, paragraphs, up to taking the entire document asone big chunk), and performs feature generation at each of these levels. Werely on the subsequent feature selection step (Section 3.4.2) to eliminate extra-neous features, preserving only those that genuinely characterize the document.Figure 3.5 presents the feature generation algorithm.

In fact, the proposed approach tackles the two most important problems innatural language processing, namely, synonymy (the ability of natural languagesto express many notions in more than one way), and polysemy (the property


Phd Thesis - [PDF Document] (55)

of natural language words to convey more than a single sense, while certainwords may have as many as dozens of different, sometimes unrelated senses).When individual contexts are classified, word sense disambiguation is implicitlyperformed, thus resolving word polysemy to some degree. A context that containsone or more polysemous words is mapped to the concepts that correspond to thesense shared by the context words. Thus, the correct sense of each word isdetermined with the help of its neighbors. At the same time, enriching documentrepresentation with high-level concepts and their generalizations addresses theproblem of synonymy, as the enhanced representation can easily recognize thattwo (or more) documents actually talk about related issues, albeit using differentvocabularies.

For each context, the feature generator yields a list of concepts ordered by theirscore, which quantifies their appropriateness to the context. A number of top-scoring concepts are used to actually generate features. For each of these conceptswe generate one feature that represents the concept itself. If the knowledgerepository also defines relations between concepts, these relations can be used forgenerating additional features (see Sections 3.5 and 3.6).

3.4.2 Feature Selection

Using support vector machines in conjunction with bag of words, Joachims (1998)found that SVMs are very robust even in the presence of numerous features. Hefurther observed that the multitude of features are indeed useful for text cate-gorization. These findings were corroborated in more recent studies (Rogati andYang, 2002; Brank et al., 2002; Bekkerman, 2003) that observed either no im-provement or even small degradation of SVM performance after feature selection.Consequently, many later works using SVMs did not apply feature selection atall (Leopold and Kindermann, 2002; Lewis et al., 2004).

This situation changes drastically as we augment the bag of words with gen-erated features. First, nearly any technique for automatic feature generation caneasily generate huge numbers of features, which will likely aggravate the “curseof dimensionality.” Furthermore, it is feature selection that allows the featuregenerator to be less than a perfect classifier. When some of the concepts assignedto the document are correct, feature selection can identify them and seamlesslyeliminate the spurious ones. We further analyze the utility of feature selection inSection 5.3.5.

Note also that the categories to which the documents are categorized mostlikely correspond to a mix of knowledge repository concepts rather than a singleone. Therefore, as the feature generator maps documents to a large set of relatedconcepts, it is up to feature selection to retain only those that are relevant to theparticular categorization task in hand.


Phd Thesis - [PDF Document] (56)

In a related study (Gabrilovich and Markovitch, 2004) we described a classof problems where feature selection can improve SVM performance even for abag of words. In this work, we formulated the notion of feature redundancy,and proposed a criterion for quantifying this phenomenon in order to predict theusefulness of feature selection. Further details can be found in Appendix A.

3.4.3 Feature Valuation

In regular text categorization, each word occurrence in document text is initiallycounted as a unit, and then feature valuation is performed, usually by subjectingthese counts to TFIDF weighting (Salton and Buckley, 1988; Debole and Sebas-tiani, 2003). To augment the bag of words with generated features and to usea single unified feature set, we need to assign weights to generated features in acompatible manner.

Each generated feature is assigned the basic weight of 1, as in the single occur-rence of a word in the bag of words. However, this weight is further multiplied bythe classification score produced for each classified concept by the feature genera-tor (Score(ci) in Figure 3.4). This score quantifies the degree of affinity betweenthe concept and the context it was assigned to.

3.4.4 Revisiting the Running Example

Let us revisit the example from Section 1, where we considered a document thatbelongs to the “copper” category of Reuters-21578. Figure 3.6 illustrates the pro-cess of feature generation for this example. While building the feature generator inthe preprocessing stage, our system learns the scope of mining-related ODP cate-gories such as Business/Mining and Drilling, Science/Technology/Mining

and Business/Industrial Goods and Services/Materials/Metals. Thesecategories contain related URLs, such as and, which belong to the (now merged) Teck Com-inco company. The company’s prominence and frequent mention causes the words“Teck” and “Cominco” to be included in the set of attributes selected to representthe above categories.

During feature generation, the document is segmented into a sequence of con-texts The feature generator analyzes these contexts and uses their words (e.g.,“Teck” and “Cominco”) to map the document to a number of mining-relatedconcepts in the ODP (e.g., Business/Mining and Drilling). These concepts, aswell as their ancestors in the hierarchy, give rise to a set of generated features thataugment the bag of words. Observe that the training documents for the category“copper” underwent similar processing when a text classifier was induced. Con-sequently, features based on these concepts were selected during feature selection


Phd Thesis - [PDF Document] (57)


Attributes selected for this concept: …, Teck, Cominco, …

“ … Cominco and Teck's 22 pct-owned Lornex agreed in January 1986 to form the joint venture, merging their Highland Valley operations …” Bag of Words

Generated features: …, Metallurgy, Metallic_Deposits Mining_and_Drilling, …

Feature vector

Text classifier

Web sites catalogued under Business/Mining_and_Drilling

Figure 3.6: Feature generation example

and retained in document vectors, thanks to their high predictive capacity. Itis due to these features that the document is now categorized correctly, whilewithout feature generation it consistently caused BOW classifiers to err.

3.5 Using Hierarchically-Structured Knowledge


We now elaborate on Requirement 3 (Section 3.2) that allows knowledge repos-itories to optionally define relations between concepts. The simplest and mostcommon organization of a set of concepts is using a hierarchical structure, whichestablishes an “is-a” relation between concepts. This way, each concept is moregeneral that all of its children.

The Open Directory Project that we use in this study is an ex-ample of such knowledge repository. Consider, for instance, the pathTop/Computers/Artificial Intelligence/Machine Learning/Datasets,which leads to a leaf concept in the ODP tree. In this example, the parentconcept Top/Computers/Artificial Intelligence/Machine Learning

is more general than the leaf concept Datasets, and concept


Phd Thesis - [PDF Document] (58)

Top/Computers/Artificial Intelligence is more general than Ma-

chine Learning. The root concept Top is more general than any otherODP concept.

Let us now formalize this extended setting. Let c0 be the root concept, which ismore general than any other concept. Let Parent(ci) be a function that uniquelyassociates a node with its parent in the hierarchy, whereas Parent(c0) is un-defined. Let Children(ci) be a function that associates a node with a set ofits children, where for leaf nodes Children(ci) = ∅. When concept ci is moregeneral than another concept cj, we denote this by ci v cj; this happens whencj ∈ Children∗(ci), where Children∗ denotes the recursive application of thefunction (obviously, ∀j > 0 : c0 v cj). Similarly, let Parent∗(ci) denote the set ofancestors of ci, obtained through the recursive application of the Parent function(of course, ∀j > 0 : c0 ∈ Parents∗(cj)).

Two aspects of our methodology can benefit from hierarchical organizationof concepts. First, while building the feature generator, we make use of the textobjects associated with each concept to learn its scope, in order to be able toassign this concept to documents in text categorization. Hierarchical organiza-tion allows us to greatly extend the amount of text associated with each concept,by taking the texts associated with all of the descendants of this concept. Thisis possible because the descendants represent more specific concepts, and thusit makes perfect sense to use their sample texts to enrich the text pool for theancestor concept. Thus, for example, if a certain concept ci is only associatedwith a few textual objects, we can learn its scope much more reliably by aggre-gating the textual objects associated with all of its descendants, Children∗(ci).Figure 3.7 provides pseudocode of the algorithm for building a feature generatorusing hierarchically-structured knowledge repositories.

Accumulation of textual objects from the descendants of a con-cept has implications for attribute selection. Let us denote by Dhier


the collection of textual objects of ci and its descendants, Dhieri =⋃

k=0...miti,k ∪ ⋃

j=0...n s.t. cj∈Children∗(ci)


tj,k, and by Dhieri the collection

of textual objects for all other concepts. Attribute selection now has to identifyattributes from Dhier

i that are most characteristic of ci.

However, there is an additional benefit in using a hierarchical ontology, asit allows us to perform powerful generalizations during feature construction. Asexplained in Section 3.3.2, the default feature construction strategy is to use thefeature generator to map the document text into one or more pertinent conceptsthat are classified based on the document text. For the sake of this discussion,let ct denote a particular document context that undergoes feature construction,and let it be classified into concepts c1, . . . , cp. In the presence of hierarchicalorganization of concepts, we can now map this context to additional concepts,namely,

⋃j=1...p Parent∗(cj). That is, the context is also mapped to concepts


Phd Thesis - [PDF Document] (59)

Algorithm BuildFeatureGenerator# Compute attribute vectors for all conceptsBuildVectors(c0)

# Use an induction algorithm to train a feature generator FG# using the attribute vectors V ector(ci)FG ← InduceClassifier({V ector(ci)})

———————————————————————————————-Algorithm BuildVectors(ci)

Text(ci) = ∅

# Traverse the hierarchy bottom-up, collecting the textual objects# of the descendants of each conceptFor each child cj ∈ Children(ci) doBuildVectors(cj)Text(ci) ← Text(ci) ∪ ⋃


# Now add the textual objects for the concept itselfText(ci) ← Text(ci) ∪ ⋃


# Build the attribute vector by performing attribute selection# among the words of Text(ci)V ector(ci) ← AttributeSelection(Text(ci))# Assign values to the selected attributesV ector(ci) ← tfidf(V ector(ci))

Figure 3.7: Building a feature generator using hierarchically-structured knowledgerepositories

that are more general than the originally classified ones. Figure 3.8 presents theextended feature generation algorithm.

When knowledge concepts are organized hierarchically, the feature generationalgorithm can take advantage of such organization. This way, instead of consider-ing all existing concepts simultaneously, it can work top-down into the hierarchy,identifying several most suitable concepts at each level, as in the hierarchical textclassifiers described in the literature (Koller and Sahami, 1997; Dumais and Chen,2000; Ruiz and Srinivasan, 2002). One possible drawback of such approach, how-ever, is that erroneous decisions made early in the process cannot be correctedlater.


Phd Thesis - [PDF Document] (60)

Algorithm FG(text, distanceMetric, numConcepts)TextV ector ← tfidf(text)For each ci ∈ C = {c0, . . . , cn} do

Score(ci) ← distanceMetric(TextV ector, V ector(ci))Let GeneratedConcepts be a set of numConcepts concepts

with highest Score(ci)

Ancestors ← ∅For each cj ∈ GeneratedConcepts do

Ancestors ← Ancestors ∪ Parent∗(cj)

Return GeneratedConcepts ∪ Ancestors

Figure 3.8: Feature generation using a hierarchical ontology

3.6 Using Knowledge Repositories that Define

Arbitrary Relations Between Concepts

Knowledge concepts can be subject to many other relations in addition togeneralization. Examples of such relations include meronymy (“part of”) andholonymy, synonymy, as well as more specific relations such as “capital of”,“birthplace/birthdate of” etc. A notable example of a knowledge repositorythat features such relations is the Wikipedia encyclopedia, where relations arerepresented by hypertext links between Wikipedia articles.

As opposed to strict hierarchical organization built on the “is-a” relation, itmakes little sense to use arbitrary relations to enrich the text pool associatedwith each concept, as explained in the previous section. However, we can stilluse these relations for feature construction. This way, whenever a text fragmentis classified to a certain concept ci, we consider generating features based onthe concepts that stand in some relation to ci. Since different relations mightreflect different strength of connection between concepts, it might be necessaryto quantify this strength in some way, in order to construct features based onconcepts that are truly relevant to the input text.

Figure 3.9 presents the pseudocode of the feature generation algorithm ex-tended for the case of arbitrary relations.


Phd Thesis - [PDF Document] (61)

Algorithm FG(text, distanceMetric, numConcepts)TextV ector ← tfidf(text)For each ci ∈ C = {c0, . . . , cn} do

Score(ci) ← distanceMetric(TextV ector, V ector(ci))Let GeneratedConcepts be a set of numConcepts concepts

with highest Score(ci)

Related ← ∅For each cj ∈ GeneratedConcepts do

For each ck such that ∃rl : 〈cj, ck〉 ∈ rl doIf Strength(〈cj, ck〉) > threshold then

Related ← Related ∪ {ck}

Return GeneratedConcepts ∪Related

Figure 3.9: Feature generation with arbitrary relations


Phd Thesis - [PDF Document] (62)


Phd Thesis - [PDF Document] (63)

Chapter 4

Instantiation of the FeatureGeneration Methodology for theODP and Wikipedia

In this Chapter we instantiate our feature generation methodology for two spe-cific knowledge repositories—the Open Directory Project (ODP, 2006) and theWikipedia encyclopedia (Wikipedia, 2006).

4.1 Using the Open Directory for Feature Gen-


We now instantiate the general methodology presented in Section 3 to use theOpen Directory project as a knowledge repository (Gabrilovich and Markovitch,2005).

The Open Directory comprises a hierarchy of approximately 600,000 nodesthat catalog over 4,000,000 Web sites, each represented by a URL, a title, anda brief summary of its contents. The directory is organized as a tree whereeach node has a title (defined by its location within the directory, e.g., Comput-

ers/Artificial Intelligence), and about one-third of all nodes have a shorttextual description. Every ODP node is associated with a collection of URLs toWeb sites cataloged under that node, while each URL has a title and a concisesummary of the corresponding Web site. The project constitutes an ongoing ef-fort promoted by over 65,000 volunteer editors around the globe, and is arguablythe largest publicly available Web directory.1 Being the result of pro bono work,

1Although the actual size of Yahoo! has not been publicly released in the re-cent years, it is estimated to be about half the size of the Open Directory. Thisestimate is based on brute-force exhaustive crawling of the Yahoo! hierarchy. See


Phd Thesis - [PDF Document] (64)

the Open Directory has its share of drawbacks, such as non-uniform coverage,duplicate subtrees in different branches of the hierarchy, and sometimes biasedcoverage due to peculiar views of the editors in charge. At the same time, how-ever, ODP embeds a colossal amount of human knowledge in a wide variety ofareas, covering even very specific scientific and technical concepts. Armed withthis knowledge, the ODP-based feature generator constructs new features thatdenote ODP categories, and adds them to the bag of words. The augmented fea-ture space provides text classifiers with a cornucopia of additional information.

4.1.1 Multiplying Knowledge Through Web Crawling

We can use the titles and summaries of the URLs as training examples for learningthe feature generator. Although these descriptions alone constitute a sizeableamount of information, we devised a way to increase the volume of training databy several orders of magnitude. We do so by crawling the Web sites pointed atby all cataloged URLs, and obtain a small representative sample of each site.Following the scheme introduced by Yang, Slattery, and Ghani (2002), each linkcataloged in the ODP is used to obtain a small representative sample of the targetWeb site. To this end, we crawl each cataloged site in the BFS order, startingfrom the URL listed in the directory. A predefined number of Web pages aredownloaded, and then concatenated into a synthetic meta-document. This meta-document, along with the URL title and summary, constitutes the textual objectfor that site. Pooling together the meta-documents for all sites associated withan ODP node gives us a wealth of additional information about it.

4.1.2 Noise Reduction and Attribute Selection

Using so much knowledge requires a host of filtering mechanisms that control thequality and utility of the generated features. We now describe these mechanismsin detail. In what follows, we distinguish between structural noise, which isinherent to the ODP structure, and content noise, which is found in the texts weobtain through crawling the cataloged URLs.

Structural noise

However elaborate the Open Directory is, it necessarily contains concepts that aredetrimental to feature generation. These include concepts too specific or situatedtoo deep in the hierarchy, or having too few textual objects to build a represen-tative attribute vector. It is important to observe, however, that whenever we and formore details.


Phd Thesis - [PDF Document] (65)

prune small concepts, we assign all of their textual content to their parents. Hereagain we benefit from the hierarchical organization of the directory, which allowsus to aggregate small fragments of specific knowledge at a higher conceptual level,where its accumulated mass becomes sufficient to define a more general concept.

We identified the following potential sources of noise in the Open Directory:

1. The branch Top/World concentrates material in languages other thanEnglish. This entire branch is therefore pruned.

2. Some top-level branches contain concepts that are hardly useful for subse-quent text categorization.

(a) Top/News is a very elaborate subtree devoted to listing numerousCNN stories on various topics organized by date. The nodes of thissubtree represent past dates, and do not correspond to useful knowl-edge concepts.

(b) Top/Adult lists adult-oriented Web sites, and we believe that theconcepts of this subtree are of little use for general purpose text cate-gorization.

(c) Top/Kids and Teens roughly duplicates the structure of the ODPbut only lists resources suitable for children.

All these branches are pruned as well.

3. Overly small categories (usually situated very deep in the hierarchy) thatonly contain a handful of URLs, and therefore their scope cannot be learnedreliably. We therefore eliminate categories with fewer than 10 URLs orthose situated below depth level 7 (the textual content of pruned categoriesis assigned to their parents).

4. The Top/Regional branch contains approximately one third of the entiremass of the ODP data, and is devoted to listing English language sitesabout various geographical regions of the world. This branch is furtherdivided into continents, countries and smaller localities, up to the levelof cities, towns and landmarks. However, the hierarchy does not stop atthis level, and for most localities it provides much more elaborate classifi-cation, similar to that of the higher ODP levels. For example, under thepath Top/Regional/North America/United States/New York/Lo-

calities/N/New York City one finds further subdivisions such asArts and Entertainment, Business and Economy, Health, Shopping

and Society and Culture. A similar set of categories duplicatinghigher-level notions (Top/Arts, Top/Business etc.) can be also


Phd Thesis - [PDF Document] (66)

found at the state level (i.e., at Top/Regional/North America/

United States/New York).

ODP classification principles2 prescribe that businesses that operate in aparticular locality (in this example, local to the State of New York or toNew York City) should normally be catalogued under the most specificapplicable categories, while businesses with global reach should be cata-logued somewhere under Top/Business; the rationale for choosing othercategories (e.g., Top/Society/... vs. Top/Regional/North America/

United States/New York/Society and Culture) is similar. However,we believe that when the ODP is used as a knowledge repository to supporttext categorization, such fine-grained distinctions (e.g., architect offices inManhattan) are of little use. These categories only pollute the hierarchywith numerous small nodes, each of which only has a small chance of beingassigned to any given context.

Therefore, we eliminate overly specific categories under Top/Regional bypruning all paths at the level of geographical names. When the featuregenerator operates on a context describing a particular New York business,it will map the latter to the New York City node, as well as to one or moreappropriate nodes under Top/Business.

5. Web spam, which comes in the form of URLs that are hardly authoritativeor representative of their host category, but are nonetheless included inthe directory by a minority of unscrupulous editors. We do not explicitlyaddress the problem of spam here, as it lies beyond the scope of our currentstudy.

Content noise

Texts harvested from the WWW are quite different from clean passages in for-mal written English, and without adequate noise reduction crawled data may domore harm than good. To reduce content noise we perform attribute selection asexplained in Section 3.3.1. For example, Table 4.1 shows the top 10 attributesselected for sample ODP concepts using information gain as the attribute selec-tion criterion. As we can see, the attributes selected for all the sample conceptsare very intuitive and plausible.

2See and for generalODP editorial guidelines, and for Regional-specific is-sues.

3Many crawled Web pages under Top/Regional/Europe/Switzerland contain non-English material, therefore we observe words like “Schweiz” (German for Switzerland) and “der”(German masculine definite article), which survived stop words removal that is only performedfor English.


Phd Thesis - [PDF Document] (67)

ODP concept Top 10 selected attributesTop/Business/Financial Services finance, loan, mortgage, equity,

insurance, lender, bank, investment,transaction, payment

Top/Computers/Artificial Intelligence neural, artificial, algorithm, intelligence,AAAI, Bayesian, probability, IEEE,cognitive, inference

Top/Health/Nutrition nutrition, diet, nutrient, vitamin, dietary,cholesterol, carbohydrate, intake,protein, fat

Top/Home/Cooking recipe, sauce, ingredient, soup, salad,casserole, stew, bake, butter, cook

Top/Recreation/Travel travel, itinerary, trip, destination, cruise,hotel, tour, adventure, travelogue,departure

Top/Regional/Europe/Switzerland3 Switzerland, Swiss, Schweiz, und, Suiss,sie, CHF, der, Zurich, Geneva

Top/Science science, research, scientific, biology,laboratory, analysis, university, theory,study, scientist

Top/Shopping/Gifts gift, birthday, occasion, basket, card,shipping, baby, keepsake, order, wedding

Top/Society/History war, history, military, army, civil,historian, soldier, troop, politics, century

Top/Sports/Golf golf, golfer, tee, hole, fairway,tournament, championship, clubhouse,PGA, par

Table 4.1: Examples of attribute selection using information gain

Learning the Feature Generator

In our current implementation, the feature generator works as a centroid-basedclassifier (Han and Karypis, 2000), which represents each category as a centroidvector of the pool of textual objects associated with it.4 Given a fragment oftext supplied as input for feature generation, the classifier represents it as anattribute vector in the same space. It then compares this vector to those of allthe concepts, and returns the desired number of best-matching ones. Attributevectors are compared using the cosine metric (Zobel and Moffat, 1998); the value

4The centroid classifier offers a simple and efficient way for managing the multitude ofconcepts in the Open Directory; additional machine learning techniques suitable for learningthe feature generator have been mentioned in Section 3.3.


Phd Thesis - [PDF Document] (68)

of the metric is treated as the classification score. A number of top-scoringconcepts are retained for each input text as generated features. The featuregenerator also performs generalization of these concepts, and constructs featuresfrom the classified concepts per se as well as their ancestors in the hierarchy.

4.1.3 Implementation Details

To evaluate the utility of knowledge-based feature generation, we implemented theproposed methodology using the Open Directory as a source of world knowledge.Throughout the experiments we used an ODP snapshot as of April 2004. Crawlingof URLs cataloged in the Open Directory was performed over the period of April–August 2004. In what follows, we describe the implementation details and designchoices of our system.

Constructing the Feature Generator

All ODP data is publicly available in machine-readable RDF format at We used the file structure.rdf.u8, which defines thehierarchical structure of the directory, as well as provides category names anddescriptions, and the file content.rdf.u8, which associates each category witha list of URLs, each having a title and a concise summary of the correspondingWeb site. After pruning the Top/World branch, which contains non-Englishmaterial, and Top/Adult branch, which lists adult-oriented Web sites, we ob-tained a collection of over 400,000 concepts and 2,800,000 URLs, organized ina very elaborate hierarchy with maximum depth of 13 levels and median depthof 7. Further pruning of too small and deep categories, as well as pruning ofthe Top/Regional subtree at the level of geographical names as explained inSection 4.1.2, reduced the number of concepts to 63,000 (the number of URLswas not reduced, since the entire URL population from pruned nodes is movedto their parents).

Titles and summaries of the URLs amounted to 436 Mb of text. In order toincrease the amount of information available for training the feature generator, wefurther populated the ODP hierarchy by crawling all of its URLs, and taking thefirst 10 pages (in the BFS order) encountered at each site to create a representativemeta-document of that site. As an additional noise removal step, we discardedmeta-documents containing fewer than 5 distinct terms. This operation yielded425 Gb worth of HTML files. After eliminating all the markup and truncatingoverly long files at 50 Kb, we ended up with 70 Gb of additional textual data.Compared to the original 436 Mb of text supplied with the hierarchy, we obtainedover a 150-fold increase in the amount of data.

Applying our methodology to a knowledge repository of this scale required an


Phd Thesis - [PDF Document] (69)

enormous engineering effort. After tokenization and removal of stop words, num-bers and mixed alphanumeric strings (e.g., “Win2k” or “4Sale”), we obtained20,800,000 distinct terms. Further elimination of rare words (occurring in lessthan 5 documents) and applying the Porter stemming algorithm (Porter, 1980)resulted in a more manageable number of 2,900,000 distinct terms that were usedto represent ODP nodes as attribute vectors. Up to 1000 most informative at-tributes were selected for each ODP node using the Document Frequency criterion(other commonly used feature selection techniques, such as Information Gain, χ2

and Odds Ratio (Yang and Pedersen, 1997; Rogati and Yang, 2002; Mladenic,1998a), yielded slightly inferior results in text categorization).

In order to speed up consequent classification of document contexts, we alsobuilt an inverted index that, given a word, provides a list of concepts that haveit in their attribute vector (i.e., the word has been selected for this concept).

When assigning weights to individual entries in attribute vectors, we took intoconsideration the location of original word occurrences. For example, words thatoccurred in URL titles were assigned higher weight than those in the summaries.Words originating from the summaries or meta-documents corresponding to linksprioritized5 by the ODP editors were also assigned additional weight. We com-pletely ignored node descriptions since these are only available for about 40% ofthe nodes, and even then the descriptions are rarely used to actually describethe corresponding concept; in many cases they just contain instructions to theeditors or explain what kinds of sites should not be classified under the node.

Finally, the set of attribute vectors undergoes tf.idf weighting, and serves forbuilding a centroid-based feature generator.

4.2 Using Wikipedia for Feature Generation

From time immemorial, the human race strived to organize its collective knowl-edge in a single literary work. From “Naturalis Historiae” by Pliny the Elderto the contemporary mammoth “Encyclopaedia Britannica”, encyclopedias havebeen major undertakings to systematically assemble all the knowledge availableto the mankind.

Back in the early years of AI research, Buchanan and Feigenbaum (1982) for-mulated the knowledge as power hypothesis, which postulated that “The power ofan intelligent program to perform its task well depends primarily on the quantityand quality of knowledge it has about that task.” Lenat et al. (1990) arguedthat without world knowledge computer programs are very brittle, and can only

5ODP editors can highlight especially prominent and important Web sites; sites marked assuch appear at the top of category listings and are emphasized with an asterisk (in RDF datafiles, the corresponding links are marked up with a <priority> tag).


Phd Thesis - [PDF Document] (70)

carry out tasks that have been fully foreseen by their designers.

When computer programs face tasks that require human-level intelligence, itis only natural to use an encyclopedia to endow the machine with the breadth ofknowledge available to humans. There are, however, several obstacles on the wayto using encyclopedic knowledge. First, such knowledge is available in textualform, and using it requires natural language understanding, a major problemin its own right. Furthermore, language understanding may not be enough, astexts written for humans normally assume the reader possesses a large amountof common-sense knowledge, which is omitted even from most detailed encyclo-pedia articles (Lenat, 1997). To address this situation, Lenat and his colleagueslaunched the CYC project, which aims to explicitly catalog the common senseknowledge of the humankind.

In this thesis we propose and evaluate a way to render text categorizationsystems with true encyclopedic knowledge, based on the largest encyclopedia inexistence—Wikipedia.

Let us illustrate the importance of encyclopedic knowledge with a coupleof examples. Given a very brief news title “Bernanke takes charge”, a casualobserver can infer little information from it. However, using the algorithm wedeveloped for consulting Wikipedia, we find out the following relevant concepts:Ben Bernanke, Federal Reserve, Chairman of the Federal Reserve,Alan Greenspan (Bernanke’s predecessor), Monetarism (an economic theoryof money supply and central banking), inflation and deflation. As anotherexample, consider the title “Apple patents a Tablet Mac”. Unless the reader iswell-versed in the hi-tech industry and gadgets, she will likely find it hard topredict the contents of the news item. Using Wikipedia, we identify the follow-ing related concepts: Mac OS (the Macintosh operating system) Laptop (thegeneral name for portable computers, of which Tablet Mac is a specific example),Aqua (the GUI of Mac OS X), iPod (another prominent product by Apple),and Apple Newton (the name of Apple’s early personal digital assistant).

Observe that documents manipulated by a text categorization system aregiven in the same form as the encyclopedic knowledge we intend to use—plaintext. Therefore, we can use text similarity algorithms to automatically identifyencyclopedia articles relevant to each document, and then leverage the knowledgegained from these articles in subsequent processing. It is this key observation thatallows us to circumvent the obstacles we enumerated above, and use encyclope-dia directly, without the need for deep language understanding or pre-catalogedcommon-sense knowledge. Also, it is essential to note that we do not use encyclo-pedia to simply increase the amount of the training data for text categorization;neither do we use it as a text corpus to collect word cooccurrence statistics.Rather, we use the knowledge distilled from the encyclopedia to enrich the rep-resentation of documents, so that a text categorizer is induced in the augmented,


Phd Thesis - [PDF Document] (71)

knowledge-rich feature space.

4.2.1 Wikipedia as a Knowledge Repository

What kind of knowledge repository should be used for feature generation? In theprevious section, we assumed the external knowledge is available in the form ofa generalization hierarchy, and used the Open Directory Project as an example.This method, however, had a number of drawbacks, which can be corrected byusing Wikipedia.

First, requiring the knowledge repository to define an “is-a” hierarchy limitsthe choice of appropriate repositories. Moreover, hierarchical organization em-bodies only one particular relation between the nodes (generalization), while nu-merous other relations, such as relatedness or meronymy/holonymy, are ignored.Second, large-scale hierarchies tend to be extremely unbalanced, so that the rela-tive size of some branches is disproportionately large or small due to peculiar viewsof the editors. Such phenomena are indeed common in the ODP. For example,the Top/Society branch is heavily dominated by one of its children—Religion

and Spirituality; the Top/Science branch is dominated by its Biology child;a considerable fraction of the mass of Top/Recreation is concentrated in Pets.Finally, to learn the scope of every ODP concept, short URL summaries associ-ated with the concepts were augmented by crawling the URLs themselves. Thisprocedure allowed us to accumulate many gigabytes worth of textual data, butat a price, as texts obtained from the Web are often quite far from formal writingand plagued with noise. Crawling a typical Web site often brings auxiliary ma-terial that has little to do with the site theme, such as legal disclaimers, privacystatements, and help pages.

In this section we propose to perform feature generation using Wikipedia,which is currently the largest knowledge repository on the Web (Gabrilovichand Markovitch, 2006b). Wikipedia is available in dozens of languages, whileits English version is the largest of all, containing 300+ million words innearly one million articles contributed by over 160,000 volunteer editors. Forthe sake of comparison, the other well-known encyclopedia, Britannica, isabout an order of magnitude smaller, with 44 million words in 65,000 articles(, visited on February 10, 2006).

Compared to the ODP, Wikipedia possesses several advantageous properties.First, its articles are much cleaner than typical Web pages, and mostly qualify asstandard written English. Although Wikipedia offers several orthogonal browsinginterfaces, their structure is fairly shallow, and we propose to treat Wikipedia ashaving essentially no hierarchy. This way, mapping documents onto relevantWikipedia concepts yields truly multi-faceted classification of the document text,and avoids the problem of unbalanced hierarchy branches. Moreover, by not


Phd Thesis - [PDF Document] (72)

requiring the knowledge repository to be hierarchically organized, our approachis suitable for new domains, for which no ontology is available. Finally, Wikipediaarticles are heavily cross-linked, in a way reminiscent of linking on the Web. Webelieve that these links encode many interesting relations between the concepts,and constitute an important source of information in addition to the article texts.We explore using inter-article links in Section 4.2.3.

4.2.2 Feature Generator Design

Although Wikipedia has almost a million articles, not all of them are equallyuseful for feature generation. Some articles correspond to overly specific concepts(e.g., Metnal, the ninth level of the Mayan underworld), or are otherwise un-likely to be useful for subsequent text categorization (e.g., specific dates or a listof events in a particular year). Other articles are just too short, so we cannot reli-ably classify texts onto the corresponding concepts. We developed a set of simpleheuristics for pruning the set of concepts, by discarding articles that have fewerthan 100 non stop words or fewer than 5 incoming and outgoing links. We alsodiscard articles that describe specific dates, as well as Wikipedia disambiguationpages.

The feature generator performs classification of texts onto Wikipedia concepts.Observe that input texts are given in the same form as Wikipedia articles, thatis, in the form of plain text. Therefore, we can use conventional text classificationalgorithms (Sebastiani, 2002) to rank the concepts represented by these articlesaccording to their relevance to the given text fragment. It is this key observationthat allows us to use encyclopedia directly, without the need for deep languageunderstanding or pre-cataloged common-sense knowledge.

However, this is a very peculiar classification problem with hundreds of thou-sands of classes, each having a single positive example—the article text. Conven-tional induction techniques can hardly be applied in these settings, so we optedto use a simple and efficient centroid classifier (Han and Karypis, 2000), whichrepresents each concept with an attribute vector of the article text.

When using a centroid classifier, it is essential to perform attribute selectionto reduce noise. However, since we only have a single article for each concept,standard attribute selection techniques cannot be applied, so we postpone noisecontrol to the next step. Each concept is represented as an attribute vector,whose entries are assigned weights using a tf.idf scheme (Debole and Sebastiani,2003). Then, we build an inverted index that maps each attribute into a listof concepts in which it appears. The primary purpose of inverted index is tospeed up vector matching. In addition to that we use it to discard insignificantassociations between attributes and concepts. This is done by removing thoseconcepts whose weights for a given attribute are too low. This scheme allows


Phd Thesis - [PDF Document] (73)

us to circumvent the scarceness of text objects for each concept—we cast theproblem of attribute selection per concept as concept selection per attribute.

4.2.3 Using the Link Structure

It is only natural for an electronic encyclopedia to provide cross-references in theform of hyperlinks. As a result, a typical Wikipedia article has many more linksto other entries than articles in conventional printed encyclopedias.

This link structure can be used in several ways. Observe that each link isassociated with an anchor text (clickable highlighted phrase). The anchor textis not always identical to the canonical name of the target article, and differ-ent anchor texts are used to refer to the same article in different contexts. Forexample, anchor texts pointing at Federal Reserve include “Fed”, “U.S. Fed-eral Reserve Board”, “U.S. Federal Reserve System”, “Board of Governors of theFederal Reserve”, “Federal Reserve Bank”, “foreign reserves” and “Free BankingEra”. Thus, anchor texts provide alternative names, variant spellings, and re-lated phrases for the target concept, which we use to enrich the article text forthe target concept.

Similarly to the WWW, incoming links contribute to the significance of anarticle. Indeed, the highest number of incoming links—over 100,000—point at thearticle United States. We use the number of incoming links to express a slightpreference for more significant concepts in feature generation, by multiplying theFG score of each concept by log(log(number of incoming links)).

Finally, inter-article links often reflect important relations between conceptsthat correspond to the linked articles. We evaluate the use of such relations forfeature generation in the next section.

Inter-article Links as Concept Relations

As a rule, the presence of a link implies some relation between the concepts itconnects. For example, the article on the United States links to Washing-ton, D.C. (country capital) and North America (the continent where thecountry is situated). It also links to a multitude of other concepts, which aredefinitely related to the source concept, albeit it is more difficult to define thoserelations; examples include United States Declaration of Independence,President of the United States, and Elvis Presley.

Let us briefly recap the way we would like to use inter-concept relations forfeature generation. Let ct be a text fragment, and let it be mapped by thefeature generator to a sequence of concepts Cct = c1, . . . , cp. We would like togenerate additional features for ct, based on concepts that stand in some rela-tion to Cct. When using Wikipedia, it is therefore logical to consider generating


Phd Thesis - [PDF Document] (74)

features based on concepts that are linked from the articles corresponding to theinitially classified concepts, namely, Cct. This way, we will generate features usingthe knowledge encoded in the links connecting the concepts.

However, our observations reveal that the existence of a link does not alwaysimply the two articles are strongly related.6 In fact, many words and phrasesin a typical Wikipedia article link to other articles just because there are entriesfor the corresponding concepts. For example, the Education subsection in thearticle on the United States has gratuitous links to concepts High school,College, and Literacy rate.

Therefore, in order to use Wikipedia links for feature generation, it is essentialto filter the linked concepts according to their relevance to the context. To thisend, we examine the related concepts linked to those in Cct, and retain those withhighest scores for the original context ct. If a newly considered concept is linkedto more than one concept in Cct, its FG score is multiplied accordingly. Finally,the desired number of highest-scoring related concepts is retained to produceadditional features. Figure 4.1 illustrates the proposed algorithm.

Concept generality filter

Recall that when using the Open Directory, we generated additional features thatwere by definition more general than the originally classified ones. Wikipediaprovides numerous relations in addition to the simple “is-a”, but are featuresconstructed from them equally useful for text categorization?

Relevance of the newly constructed features is certainly important, but is notthe only criterion. Suppose that we are given an input text “Google search”.Which additional feature is likely to be more useful: Nigritude ultramarine (aspecially crafted meaningless phrase used in a search engine optimization contest)or Website? Now suppose the input is “artificial intelligence”—which feature islikely to contribute more to the representation of this input, John McCarthy

(computer scientist) or Logic? We believe that in both examples, the secondfeature would be more useful because it is not overly specific.

Consequently, we conjecture that in text categorization we should generateadditional link-based features sparingly, taking only those features that are “moregeneral” than those that triggered them. But how can we judge the generalityof concepts? While this may be tricky to achieve in the general case (no punintended), we propose the following task-oriented criterion. Given two conceptsca and cb, we compare the numbers of links pointing at them. Then, we say thatca is “more general” than cb if its number of incoming links is at least an orderof magnitude larger, that is, if log10(#inlinks(ca))− log10(#inlinks(cb)) > 1.

6The opposite is also true—the absence of a link may simply be due to an oversight. Adafreand de Rijke (2005) studied the problem of discovering missing links in Wikipedia.


Phd Thesis - [PDF Document] (75)

Algorithm FG(ct, distanceMetric, numConcepts, numSearched,numExamined, numRelated)

TextV ector ← tfidf(ct)For each ci ∈ C = {c0, . . . , cn} do

Score(ci) ← distanceMetric(TextV ector, V ector(ci))Let Generated be a set of numConcepts concepts with highest Score(ci)Let Searched be a set of numSearched concepts with highest Score(ci)Let Examined be a set of numExamined concepts with highest Score(ci)

Let Links = {〈ca, cb〉} be a set of links between Wikipedia conceptsFor each ck ∈ Searched do

RelWeight(ck) ← 0For each cj ∈ Examined do

For each ck such that 〈cj, ck〉 ∈ Links doIf ck ∈ Searched then

RelWeight(ck) ← RelWeight(ck) + Score(ck)

Let Related be a set of numRelated concepts with highest RelWeight(ck)Return Generated ∪Related

Figure 4.1: Feature generation with Wikipedia links as relations

Figure 4.2 illustrates the algorithm that only generates more general features.We use boldface font to highlight the difference from the previous version.

We show examples of additional features generated using inter-article links inSection 5.4.1. In Section 5.4.5 we report the results of using inter-article linksfor feature generation. In that section we also specifically examine the effect ofconstructing features from concepts that are more general than the concepts thattriggered them.

4.2.4 Implementation Details

We used Wikipedia snapshot as of November 11, 2005. After parsing theWikipedia XML dump, we obtained 1.8 Gb of text in 910,989 articles. Upon re-moving small and overly specific concepts that have fewer than 100 words, fewerthan 5 incoming or outgoing links, category pages, disambiguation pages and thelike, 171,332 articles were left that defined concepts used for feature generation.We processed the text of these articles by first tokenizing it, removing stop wordsand rare words (occurring in fewer than 3 articles), and stemmed the remaining


Phd Thesis - [PDF Document] (76)

Algorithm FG(ct, distanceMetric, numConcepts, numSearched,numExamined, numRelated)

TextV ector ← tfidf(ct)For each ci ∈ C = {c0, . . . , cn} do

Score(ci) ← distanceMetric(TextV ector, V ector(ci))Let Generated be a set of numConcepts concepts with highest Score(ci)Let Searched be a set of numSearched concepts with highest Score(ci)Let Examined be a set of numExamined concepts with highest Score(ci)

Let Links = {〈ca, cb〉} be a set of links between Wikipedia conceptsFor each ck ∈ Searched do

RelWeight(ck) ← 0For each cj ∈ Examined do

For each ck such that 〈cj, ck〉 ∈ Links doIf ck ∈ Searched thenIf log10(#inlinks(ck))− log10(#inlinks(cj)) > 1 then

RelWeight(ck) ← RelWeight(ck) + Score(ck)

Let Related be a set of numRelated concepts with highest RelWeight(ck)Return Generated ∪Related

Figure 4.2: Feature generation with Wikipedia links as relations, where only moregeneral features are constructed

words; this yielded 296,157 distinct terms, which were used to represent conceptsas attribute vectors.

Preprocessing of Wikipedia XML dump

Wikipedia data is publicly available online at the data is distributed in XML format, and several packaged versions areavailable: article texts, edit history, list of page titles, interlanguage links etc. Inthis project, we only use the article texts, but ignore the information on articleauthors and page modification history. Before building the feature generator, weperform a number of operations on the distributed XML dump:

• We simplify the original XML by removing all those fields that are not usedin feature generation, such as author ids and last modification times.

• Wikipedia syntax defines a proprietary format for inter-article links,whereas the name of the article referred to is enclosed in brackets (e.g.,


Phd Thesis - [PDF Document] (77)

“[United States]”). We map all articles to numeric ids, and for each ar-ticle build a list of ids of the articles it refers to. We also count the numberof incoming and outgoing links for each article.

• Wikipedia defines a redirection mechanism, which maps frequently usedvariant names of entities into canonical names. For examples, UnitedStates of America is mapped to United States. We resolve all suchredirections during initial preprocessing.

• Another frequently used mechanism is templates, which allows articles to in-clude frequently reused fragments of text without duplication, by includingpre-defined and optionally parameterized templates on the fly. To speed upsubsequent processing, we resolve all template inclusions at the beginning.

• We also collect all anchor texts that point at each article.

This preprocessing stage yields a new XML file, which is then used for buildingthe feature generator.

Inverted Index Pruning

The algorithm for pruning the inverted index operates as follows. We first sort allthe concepts for a given word according to their tf.idf weights in the decreasingorder. We then scan the resulting sequence of concepts with a sliding windowof length 100, and truncate the sequence when the difference in scores betweenthe first and last concepts in the window drops below 5% of the highest-scoringconcept for this word (which is positioned first in the sequence).


Phd Thesis - [PDF Document] (78)


Phd Thesis - [PDF Document] (79)

Chapter 5

Empirical Evaluation of FeatureGeneration for TextCategorization

In this chapter we evaluate the benefits of using external knowledge for textcategorization.

5.1 Test Collections

We used the following test collections to evaluate our methodology.

5.1.1 Reuters-21578

This data set contains one year worth of English-language stories distributedover the Reuters newswire in 1986–1987, and is arguably the most often usedtest collection in text categorization research. Reuters-21578 is a cleaned versionof the earlier release named Reuters-22173, which contained errors and duplicatedocuments.

The collection contains 21578 documents (hence the name) in SGML for-mat. Of those, 12902 documents are categorized, i.e., assigned a category label ormarked as not belonging to any category. Other documents do not have an ex-plicit classification; that is, they can reasonably belong to some categories (judgedby their content), but are not marked so. Several train/test splits of the collectionhas been defined, of which ModApte (Modified Apte) is the most commonly usedone. The ModApte split divides the collection chronologically, and allocates thefirst 9603 documents for training, and the rest 3299 documents for testing.

The documents are labeled with 118 categories; there are 0–16 labels per doc-ument, with the average of 1.04. The category distribution is extremely skewed:


Phd Thesis - [PDF Document] (80)

the largest category (“earn”) has 3964 positive examples, while 16 categories haveonly one positive example. Several category sets were defined for this collection:

• 10 largest categories (“earn”, “acq”, “money-fx”, “grain”, “crude”, “trade”,“interest”, “ship”, “wheat”, “corn”).

• 90 categories with at least one document in the training set and one in thetesting set (Yang, 2001).

• Galavotti, Sebastiani, and Simi (2000) used a set of 115 categories with atleast one training example (three categories, “cottonseed”, “f-cattle” and“sfr” have no training examples under the ModApte split).

• The full set of 118 categories with at least one positive example either inthe training or in the testing set.

Following common practice, we used the ModApte split and two categorysets, 10 largest categories and 90 categories with at least one training and testingexample.

5.1.2 20 Newsgroups (20NG)

The 20 Newsgroups collection (Lang, 1995) is comprised of 19997 postings to 20Usenet newsgroups. Most documents have a single label, defined as the nameof the newsgroup it was sent to; about 4% of documents have been cross-posted,and hence have several labels. Each newsgroup contains exactly 1000 positiveexamples, with the exception of “soc.religion.christian” which contains 997 doc-uments.

Some categories are quite close in scope, for example, “” and “comp.sys.mac.hardware”, or “talk.religion.misc” and “soc.reli-gion.christian”. A document posted to a single newsgroup may be reasonablyconsidered appropriate for other groups too (the author may have simply notknown of other similar groups, and thus not cross-posted the message); this nat-urally poses additional difficulty for classification.

It should be noted that Internet news postings are very informal, and thereforethe documents frequently contain non-standard and abbreviated words, foreignwords, and proper names, as well as a large amount of markup characters (usedfor attribution of authorship or for message separation).

5.1.3 Movie Reviews

The Movie Reviews collection (Pang, Lee, and Vaithyanathan, 2002) represents aslightly different classification task than standard text categorization, referred to


Phd Thesis - [PDF Document] (81)

as sentiment classification. The collection contains 1400 reviews of movies, halfof which express positive sentiment (opinion) about the movie, and half nega-tive. The reviews were collected from the “” newsgroup,archived at the Internet Movie Database (IMDB, Theclassification problem in this case is to determine the semantic orientation ofthe document, rather than to relate its content to one of the predefined topics.This problem is arguably more difficult than topical text categorization, sincethe notion of semantic orientation is quite general. We saw this collection as anopportunity to apply feature generation techniques to this new task.

Recent works on semantic orientation include (Turney and Littman, 2002;Turney, 2002; Pang, Lee, and Vaithyanathan, 2002).1 The two former studiesused unsupervised learning techniques based on latent semantic indexing, esti-mating semantic distance between a given document and two reference wordsthat represent polar opinions, namely, “excellent” and “poor”. The latter workused classical TC techniques.

5.1.4 Reuters Corpus Version 1 (RCV1)

RCV1 is the newest corpus released by Reuters (Lewis et al., 2004; Rose, Steven-son, and Whitehead, 2002). It is considerably larger than its predecessor, andcontains over 800,000 news items, dated between August 20, 1996 and August19, 1997. The stories are labeled with 3 category sets, Topics, Industries andRegions.

• Topics are most close in nature to the category set of the old Reuterscollection (Reuters-21578). There are 103 topic codes, with 3.24 categoriesper document on the average. The topics are organized in a hierarchy, andthe Hierarchy Policy required that if a category is assigned to a document,all its ancestors in the hierarchy should be assigned as well. As a result,as many as 36% of all Topic assignments are due to the four most generalcategories, CCAT, ECAT, GCAT, and MCAT. Consequently, the micro-averaged performance scores are dominated by these categories (Lewis et

1The field of genre classification, which attempts to establish the genre of document, is some-what related to sentiment classification. Examples of possible genres are radio news transcriptsand classified advertisem*nts. The work by Dewdney, VanEss-Dykema, and MacMillan (2001)cast this problem as text categorization, using presentation features in addition to words. Theirpresentation features included part of speech tags and verb tenses, as well as mean and vari-ance statistics of sentence and word length, punctuation usage, and the amount of whitespacecharacters. Using support vector machines for actual classification, the authors found that theperformance due to the presentation features alone was at least as good as that achieved withplain words, and that the combined feature set usually resulted in an improvement of severalpercentage points.


Phd Thesis - [PDF Document] (82)

al., 2004), and macro-averaging becomes of interest.2 The Minimum CodePolicy required that each document was assigned at least one Topic andone Region code.

• Industries are more fine-grained than Topics, and are therefore harder forclassification. These categories are also organized in a hierarchy, althoughthe Hierarchy Policy was only partially enforced for them.

• Region codes correspond to geographical places, and are further subdividedinto countries, regional groupings and economic groupings. Lewis et al.(2004) argue that Region codes might be more suitable for named entityrecognition than for text categorization.

As noted by Lewis et al. (2004), the original RCV1 distribution contains anumber of errors; in particular, there are documents that do not conform to eitherMinimum Code or Hierarchy Policy, or labeled with erratic codes. Lewis et al.(2004) proposed a procedure to correct these errors, and defined a new version ofthe collection, named RCV1-v2 (as opposed to the original distribution, referredto as RCV1-v1 ). All our experiments are based on RCV1-v2.

In our experiments we used Topic and Industry categories. Due to the sheersize of the collection, processing all the categories in each set would take unrea-sonably long, allowing us to conduct only few experiments. Following the schemeintroduced by Brank et al. (2002), we used 16 Topic and 16 Industry categories,which constitute a representative sample of the full groups of 103 and 354 cate-gories, respectively. We also randomly sampled the Topic and Industry categoriesinto 5 sets of 10 categories each. Table 5.1 gives the full definition of the categorysets we used. To further speed up experimentation, we used a subset of the cor-pus with 17,808 training documents (dated August 20–27, 1996) and 5341 testingdocuments (dated August 28–31, 1996).


OHSUMED (Hersh et al., 1994) is a subset of the MEDLINE database, whichcontains 348,566 references to documents published in medical journals over theperiod of 1987–1991. Each reference contains the publication title, and abouttwo-thirds (233,445) also contain an abstract. Each document is labeled withseveral MeSH categories (MeSH, 2003). There are over 14,000 distinct categoriesin the collection, with an average of 13 categories per document. OHSUMED isfrequently used in information retrieval and text categorization research.

2This is why micro-averaged scores for Topic codes are so much higher than macro-averagedones, see Section 5.2.2.


Phd Thesis - [PDF Document] (83)

Set name Categories comprising the setTopic-16 e142, gobit, e132, c313, e121, godd, ghea, e13, c183, m143,

gspo, c13, e21, gpol, m14, c15Topic-10A e31, c41, c151, c313, c31, m13, ecat, c14, c331, c33Topic-10B m132, c173, g157, gwea, grel, c152, e311, c21, e211, c16Topic-10C c34, c13, gtour, c311, g155, gdef, e21, genv, e131, c17Topic-10D c23, c411, e13, gdis, c12, c181, gpro, c15, g15, c22Topic-10E c172, e513, e12, ghea, c183, gdip, m143, gcrim, e11, gvioIndustry-16 i81402, i79020, i75000, i25700, i83100, i16100, i1300003, i14000,

i3302021, i8150206, i0100132, i65600, i3302003, i8150103,i3640010, i9741102

Industry-10A i47500, i5010022, i3302021, i46000, i42400, i45100, i32000, i81401,i24200, i77002

Industry-10B i25670, i61000, i81403, i34350, i1610109, i65600, i3302020, i25700,i47510, i9741110

Industry-10C i25800, i41100, i42800, i16000, i24800, i02000, i34430, i36101,i24300, i83100

Industry-10D i1610107, i97400, i64800, i0100223, i48300, i81502, i34400, i82000,i42700, i81402

Industry-10E i33020, i82003, i34100, i66500, i1300014, i34531, i16100, i22450,i22100, i42900

Table 5.1: Definition of RCV1 category sets used in the experiments

Following Joachims (1998), we used a subset of documents from 1991 that haveabstracts, taking the first 10,000 documents for training and the next 10,000 fortesting. To limit the number of categories for the experiments, we randomlygenerated 5 sets of 10 categories each. Table 5.2 gives the full definition of thecategory sets we used.

5.1.6 Short Documents

We conjectured that knowledge-based feature generation should be particularlybeneficial for categorization of short documents. To verify this conjecture, wederived several datasets of short documents from the test collections describedabove. Recall that about one-third of OHSUMED documents have titles but noabstract, and can therefore be considered short documents “as-is.” We used thesame range of documents as defined in Section 5.1.5, but considered only thosewithout abstracts; this yielded 4,714 training and 5,404 testing documents. Forall other datasets, we created a short document from each original document bytaking only the title of the latter (with the exception of Movie Reviews, where


Phd Thesis - [PDF Document] (84)

Set name Categories comprising the set(parentheses contain MeSH identifiers)

OHSUMED-10A B-Lymphocytes (D001402);Metabolism, Inborn Errors (D008661);Creatinine (D003404); Hypersensitivity (D006967);Bone Diseases, Metabolic (D001851); Fungi (D005658);New England (D009511); Biliary Tract (D001659);Forecasting (D005544); Radiation (D011827)

OHSUMED-10B Thymus Gland (D013950); Insurance (D007341);Historical Geographic Locations (D017516);Leukocytes (D007962); Hemodynamics (D006439);Depression (D003863); Clinical Competence (D002983);Anti-Inflammatory Agents, Non-Steroidal (D000894);Cytophotometry (D003592); Hydroxy Acids (D006880)

OHSUMED-10C Endothelium, Vascular (D004730);Contraceptives, Oral, Hormonal (D003278);Acquired Immunodeficiency Syndrome (D000163);Gram-Positive Bacteria (D006094); Diarrhea (D003967);Embolism and Thrombosis (D016769);Health Behavior (D015438); Molecular Probes (D015335);Bone Diseases, Developmental (D001848);Referral and Consultation (D012017)

OHSUMED-10D Antineoplastic and Immunosuppressive Agents (D000973);Receptors, Antigen, T-Cell (D011948);Government (D006076); Arthritis, Rheumatoid (D001172);Animal Structures (D000825); Bandages (D001458);Italy (D007558); Investigative Techniques (D008919);Physical Sciences (D010811); Anthropology (D000883)

OHSUMED-10E HTLV-BLV Infections (D006800);Hemoglobinopathies (D006453); Vulvar Diseases (D014845);Polycyclic Hydrocarbons, Aromatic (D011084);Age Factors (D000367); Philosophy, Medical (D010686);Antigens, CD4 (D015704);Computing Methodologies (D003205);Islets of Langerhans (D007515); Regeneration (D012038)

Table 5.2: Definition of OHSUMED category sets used in the experiments


Phd Thesis - [PDF Document] (85)

documents have no titles).

It should be noted, however, that substituting a title for the full documentis a poor man’s way to obtain a collection of classified short documents. Whendocuments were originally labeled with categories, the human labeller saw eachdocument in its entirety. In particular, a category might have been assigned to adocument on the basis of facts mentioned in its body, even though the informa-tion may well be missing from the (short) title. Thus, taking all the categoriesof the original documents to be “genuine” categories of the title is often mislead-ing. However, because we know of no publicly available test collections of shortdocuments, we decided to construct datasets as explained above. Importantly,OHSUMED documents without abstracts have been classified as such by humans;working with the OHSUMED-derived dataset can thus be considered a “pure”experiment.

5.1.7 Automatic Acquisition of Data Sets

Although numerous works studied text categorization in the past, good test col-lections are by far less abundant. In part, this scarcity can be attributed tothe huge manual effort required to collect a sufficiently large body of text, cat-egorize it, and ultimately produce in a machine-readable format (usually XMLor SGML). Most works use the Reuters-21578 collection (Reuters, 1997) as theprimary benchmark. Others use 20 Newsgroups (Lang, 1995) and OHSUMED(Hersh et al., 1994), while TREC3 filtering experiments often use the data fromthe TIPSTER corpus.

Although the Reuters corpus became a standard reference in the field, it hasa number of significant shortcomings. According to Dumais and Chen (2000),“the Reuters collection is small and very well organized compared with manyrealistic applications”. Scott (1998) also notes that the Reuters corpus has avery restricted vocabulary, since Reuters in-house style prescribes using uniformunambiguous terminology to facilitate quick comprehension. As a consequence,good classifiers (e.g., SVM or KNN) yield very reasonable performance even usinga simple bag-of-words approach, without the need for more elaborate features.

Mainly due to these achievements in Reuters classification, Sebastiani (2002)notes that “[automated TC] has reached effectiveness levels comparable to thoseof trained professionals . . . and, more importantly, it is unlikely to be improvedsubstantially by the progress of research”. While this argument might be appro-priate for the Reuters-21578 corpus, we believe it does not apply to the generalcase. For example, the state of the art performance on the OHSUMED collection

3Text REtrieval Conferences administered by the U.S. National Institute of Science andTechnology (NIST).


Phd Thesis - [PDF Document] (86)

is only around 50–60% (Yang, 2001; Yang, 1999). We believe that the perfor-mance of TC on more representative real-life corpora still has way to go. Therecently introduced new Reuters corpus (RCV1), which features very large sizeand three orthogonal label sets definitely constitutes a new challenge. At thesame time, acquisition of additional corpora suitable for TC research remains amajor challenge.

To this end, as a part of this research we developed a methodology for auto-matic acquisition of labeled datasets for text categorization. This methodologyallows one to define a set of parameters in order to generate datasets with desiredproperties, based on the Open Directory. We present and evaluate this method-ology in Appendix B. It is essential to note that since these datasets have beenderived from the Open Directory, we cannot use them to test the effect of usingthe ODP for feature generation. Indeed, we did not use these datasets to evaluateour feature generation methodology. In Appendix A we used these datasets in astudy of feature selection.

5.2 Experimentation Procedure

We used support vector machines4 as our learning algorithm to build text cate-gorizers, since prior studies found SVMs to have the best performance for textcategorization (Sebastiani, 2002; Dumais et al., 1998; Yang and Liu, 1999). Fol-lowing established practice, we use the precision-recall break-even point (BEP)to measure text categorization performance. For the two Reuters datasets andOHSUMED we report both micro- and macro-averaged BEP, since their cate-gories differ in size significantly. Micro-averaged BEP operates at the documentlevel and is primarily affected by categorization performance on larger categories.On the other hand, macro-averaged BEP averages results for individual cate-gories, and thus small categories with few training examples have large impacton the overall performance.

For both Reuters datasets (Reuters-21578 and RCV1) and OHSUMED weused a fixed train/test split as defined in Section 5.1, and consequently usedmacro sign test (S-test) (Yang and Liu, 1999) to assess the statistical significanceof differences in classifier performance. For 20NG and Movies we performed 4-foldcross-validation, and used paired t-test to assess the significance. We also usedthe Wilcoxon signed-ranks test (Demsar, 2006) to compare the baseline and theFG-based classifiers over multiple data sets.

4We used the SVM light implementation (Joachims, 1999a).


Phd Thesis - [PDF Document] (87)

5.2.1 Text Categorization Infrastructure

We conducted the experiments using a text categorization platform of ourown design and development named Hogwarts 5 (Davidov, Gabrilovich, andMarkovitch, 2004). We opted to build a comprehensive new infrastructure fortext categorization, as surprisingly few software tools are publicly available forresearchers, while those that are available allow only limited control over theiroperation. Hogwarts facilitates full-cycle text categorization including text pre-processing, feature extraction, construction, selection and valuation, followed byactual classification. The system currently provides XML parsing, part-of-speechtagging (Brill, 1995), sentence boundary detection, stemming (Porter, 1980),WordNet (Fellbaum, 1998) lookup, a variety of feature selection algorithms, andtf.idf feature weighting schemes. Hogwarts has over 250 configurable parame-ters that control its modus operandi in minute detail. Hogwarts interfaces withSVM, KNN and C4.5 text categorization algorithms, and computes all standardmeasures of categorization performance. Hogwarts was designed with a par-ticular emphasis on processing efficiency, and portably implemented in the ANSIC++ programming language and C++ Standard Template Library. The systemhas built-in loaders for Reuters-21578 (Reuters, 1997), RCV1 (Lewis et al., 2004),20 Newsgroups (Lang, 1995), Movie Reviews (Pang, Lee, and Vaithyanathan,2002), and OHSUMED (Hersh et al., 1994), while additional datasets can beeasily integrated in a modular way.

Each document undergoes the following processing steps. Document text isfirst tokenized, and title words are replicated twice to emphasize their importance.Then, stop words, numbers and mixed alphanumeric strings are removed, and theremaining words are stemmed. The bag of words is next merged with the set offeatures generated for the document by analyzing its contexts as explained inSection 3.4, and rare features occurring in fewer than 3 documents are removed.

Since earlier studies found that most BOW features are indeed useful forSVM text categorization (Section 3.4.2), we take the bag of words in its entirety(with the exception of rare features removed in the previous step). The gen-erated features, however, undergo feature selection using the information gaincriterion. Finally, feature valuation is performed using the “ltc” tf.idf function(logarithmic term frequency and inverse document frequency, followed by cosinenormalization) (Salton and Buckley, 1988; Debole and Sebastiani, 2003).

5Hogwarts School of Witchcraft and Wizardry is the educational institution attended byHarry Potter (Rowling, 1997).


Phd Thesis - [PDF Document] (88)

5.2.2 Baseline Performance of Hogwarts

We now demonstrate that the performance of basic text categorization in ourimplementation (column “Baseline” in Table 5.3) is consistent with the stateof the art as reflected in other published studies (all using SVM). On Reuters-21578, Dumais et al. (1998) achieved micro-BEP of 0.920 for 10 categories and0.870 for all categories. On 20NG, Bekkerman (2003) obtained BEP of 0.856.6

Pang, Lee, and Vaithyanathan (2002) obtained accuracy of 0.829 on Movies. Theminor variations in performance are due to differences in data preprocessing in thedifferent systems; for example, for the Movies dataset we worked with raw HTMLfiles rather than with the official tokenized version, in order to recover sentenceand paragraph structure for contextual analysis. For RCV1 and OHSUMED,direct comparison with published results is more difficult because we limited thecategory sets and the date span of documents to speed up experimentation.

5.2.3 Using the Feature Generator

We used the multi-resolution approach to feature generation, classifying docu-ment contexts at the level of individual words, complete sentences, paragraphs,and finally the entire document.7 For each context, features were generated fromthe 10 best-matching concepts produced by the feature generator, as well as forall of their ancestors (in the case of the ODP-based FG).

5.3 ODP-based Feature Generation

5.3.1 Qualitative Analysis of Feature Generation

We now study the process of feature generation on a number of actual examples.

Feature Generation per se

In this section we demonstrate ODP-based feature generation for a number ofsample sentences taken from CNN and other Web sites. For each example, wediscuss a number of highly relevant features found among the top ten generatedones.

6Using distributional clustering of words, Bekkerman et al. (2003) obtained BEP of 0.886on this dataset in the multi-labeled setting.

7The 20NG dataset is an exception, owing to its high level of intrinsic noise that rendersidentification of sentence boundaries extremely unreliable, and causes word-level feature gen-eration to produce too many spurious classifications. Consequently, for this dataset we restrictthe multi-resolution approach to individual paragraphs and the entire document only.


Phd Thesis - [PDF Document] (89)

Dataset Baseline

micro macroBEP BEP

Reuters-2157810 categories 0.925 0.87490 categories 0.877 0.602RCV1Industry-16 0.642 0.595Industry-10A 0.421 0.335Industry-10B 0.489 0.528Industry-10C 0.443 0.414Industry-10D 0.587 0.466Industry-10E 0.648 0.605Topic-16 0.836 0.591Topic-10A 0.796 0.587Topic-10B 0.716 0.618Topic-10C 0.687 0.604Topic-10D 0.829 0.673Topic-10E 0.758 0.742OHSUMEDOHSUMED-10A 0.518 0.417OHSUMED-10B 0.656 0.500OHSUMED-10C 0.539 0.505OHSUMED-10D 0.683 0.515OHSUMED-10E 0.442 0.54220NG 0.854Movies 0.813

Table 5.3: Baseline performance of Hogwarts text categorization platform

• Text: “Rumsfeld appeared with Gen. Richard Myers, chairman of the JointChiefs of Staff.”

Sample generated features:

– Society/Issues/Government Operations, Society/Politics —both Donald Rumsfeld and Richard Myers are senior governmentofficers, hence the connection to government operations and pol-itics. Their names have been selected for these ODP concepts,since they appear in many Web sites cataloged under them, such asthe National Security Archive at the George Washington University


Phd Thesis - [PDF Document] (90)

(̃ nsarchiv) and the John F. Kennedy School ofGovernment at Harvard University (

– Society/Issues/Warfare and Conflict/Specific Conflicts/

Iraq, Science/Technology/Military Science, Society/Issues/

Warfare and Conflict/Weapons—again, both persons mentionedwere prominent during the Iraq campaign.

– Society/History/By Region/North America/United States/

Presidents/Bush, George Walker — Donald Rumsfeld serves asSecretary of Defense under President George W. Bush.

– Society/Politics/Conservatism — Rumsfeld is often seen as holdingconservative views on a variety of political issues.

• Text: “The new film follows Anakin’s descent into evil and lust for power.”

Sample generated features:

– Arts/Movies/Titles/Star Wars Movies is the root of the ODPsubtree devoted to the “Star Wars” movie series. The word“Anakin” has been selected as an attribute for this concept dueto its numerous occurrences in the cataloged Web sites such as and

– Arts/Performing Arts/Acting/Actors and Actresses/Chris-

tensen, Hayden is the actor who played Anakin Skywalker; thisparticular piece of information cannot be inferred from the shortinput sentence without elaborate background knowledge.

• Text: “On a night when Dirk Nowitzki (34 points), Jerry Stackhouse (29),Josh Howard (19) and Jason Terry (17) all came up big, he couldn’t matchtheir offensive contributions.”

Sample generated features:

– Sports/Basketball/Professional/NBA/Dallas Mavericks—even though the sentence mentions neither the particular sport northe name of the team, the power of context is at its best, immediatelyyielding the correct classification as the best-scoring generated feature.The names of the players mentioned in the context occur often in theWeb sites cataloged under this concept, including such resources as,,and

• Text: “Herceptin is a so-called targeted therapy because of its ability toattack diseased cells and leave healthy ones alone.”


Phd Thesis - [PDF Document] (91)

Sample generated features:

– Health/Conditions and Diseases/Cancer/Breast, Society/

Issues/Health/Conditions and Diseases/Cancer/Alternative -

Treatments, Health/Support Groups/Conditions and Disea-

ses/Cancer provide relevant additional information for Herceptin,a medication for breast cancer. The name of this medicine has beenselected for these concepts due to its occurrences in cataloged Websites such as,

breastcenter and

• Finally, we give an example of how the power of context can be used forword sense disambiguation. The following pair of sentences use the word“tie” in two different meanings—once as a necktie and once as a kind ofconnection. Even though these sentences contain no distinguishing propernames, the context of the polysemous words allows the feature generator toproduce correct suggestions in both cases

Text: “Kinship with others is based either on blood ties or on maritalties.”

Sample generated features:

– Society/Genealogy

– Home/Family

– Society/Relationships

– Science/Social Sciences/Sociology

Text: “Our tie shop includes plain solid colour ties, novelty ties, patternedsilk ties, and men’s bow ties.”

Sample generated features:

– Shopping/Clothing/Men’s/Neckties

– Shopping/Clothing/Accessories/Men’s

– Business/Consumer Goods and Services/Clothing/Accessories/

Ties and Scarves

Evidently, many of the generated features could not have been accessed byconventional text classification methods, since heavy use of world knowledge isrequired to deduce them.


Phd Thesis - [PDF Document] (92)

Actual Text Categorization Examples Under a Magnifying Glass

Thanks to feature generation, our system correctly classifies the running exampledocument #15264. Let us consider additional testing examples from Reuters-21578 that are incorrectly categorized by the BOW classifier. Document #16143belongs to the category “money-fx” (money/foreign exchange) and discussesthe devaluation of the Kenyan shilling. Even though “money-fx” is one of the10 largest categories, the word “shilling” does not occur in its training documentseven once. However, the feature generator easily recognizes it as a kind of cur-rency, and produces features such as Recreation/Collecting/Paper Money

and Recreation/Collecting/Coins/World Coins. While analyzing docu-ment contexts it also uses other words such as “Central Bank of Kenya”and “devaluation” to correctly map the document to ODP concepts Soci-

ety/Government/Finance, Science/Social Sciences/Economics and Busi-

ness/Financial Services/Banking Services. Even though the behavior of theKenyan shilling was never mentioned in the training set, these high-level fea-tures were also constructed for many training examples, and consequently thedocument is now classified correctly.

Similarly, document #18748 discusses Italy’s balance of payments and be-longs to the category “trade” (interpreted as an economic indicator), while theword “trade” itself does not occur in this short document. However, whenthe feature generator considers document contexts discussing Italian deficitas reported by the Bank of Italy, it correctly maps them to concepts suchas Society/Government/Finance, Society/Issues/Economic/Internatio-

nal/Trade, Business/International Business and Trade. These features,which were also generated for training documents in this category (notably, docu-ment #271 on Japanese trade surplus, document #312 on South Korea’s accountsurplus, document #354 on tariff cuts in Taiwan and document #718 on U.S.-Canada trade pact), allow the document to be categorized correctly.

Let us also consider a few documents from the Movie Reviews dataset thatconfuse the BOW classifier (here we consider a training/testing split inducedby one particular cross-validation fold). Recall that this dataset represents asentiment classification task, where documents are classified according to thesentiment of the review (positive or negative) rather than its topic. Docu-ment #19488 contains a negative review of Star Wars Episode 1, but at theword level it is difficult to judge its true sentiment since positive and negativewords are interspersed. For instance, the sentence “Anakin is annoying and un-likeable, instead of cute and huggable as Lucas no doubt intended” contains twowords with positive connotation (“cute and huggable”) that counterbalance thetwo words with negative ones (“annoying and unlikeable”). However, given con-texts like “The two leads are hideously boring, static characters given little to


Phd Thesis - [PDF Document] (93)

do and too much time to do it,” the feature generator produces features such asArts/Movies/Reviews/Top Lists/Bad Films. This ODP node catalogs Websites devoted to reviews of bad movies, and the wording of this sample contextlooks similar to that used in known negative reviews (as cataloged in the ODP).In fact, this particular feature is one of the most informative ones generated forthis dataset, and it is also produced for contexts like “Next up we have the dia-logue, which is amusingly bad at its best, painful at its worst” and “What ensuesis a badly scripted and horribly directed 114 minutes of cinema hell,” both foundin negative reviews.

As another example, consider document #15111, which contains a posi-tive review of the movie “Soldier.” This review, which constantly switchesbetween criticizing and praising the film, easily perplexes the BOW classi-fier. Interestingly, given the sentence “It is written by David Webb Peoples,who penned the screenplay to the classic Blade Runner and the critically-acclaimed 12 Monkeys,” the feature generator constructs the highly informativefeature Arts/Movies/Reviews/Top Lists/Good Films. This is made possibleby the references to known good films (“Blade Runner” and “12 Monkeys”) thatare listed in Web sites devoted to good films ( and 250 films, for example). The same feature was alsogenerated for a number of training documents, and thus helps the classifier tocategorize the document correctly.

The Importance of Feature Selection

To understand the utility of feature selection, consider a sample sentence fromour running example, Reuters document #15264: “Cominco’s share of produc-tion was 43,000 short tons of copper, 340,000 ounces of silver and 800 ounces ofgold.” Table 5.4 gives the top ten ODP concepts generated as features for thiscontext. Most of the assigned concepts deal with mining and drilling, and willeventually be useful features for document classification. However, the conceptsBusiness/Investing/Commodities, Futures/Precious Metals, Shopping andBusiness/Investing/Commodities, Futures/Precious Metals/Gold havebeen triggered by the words “gold” and ”silver,” which are mentioned inciden-tally and do not describe the gist of the document. Feature selection is thereforeneeded to eliminate features based on these extraneous concepts.

As another example, consider the following sentence taken from the samedocument: “‘Cominco, 29.5 percent owned by a consortium led by Teck, is op-timistic that the talks will soon be concluded,’ spokesman Don Townson toldReuters,” along with its top ten classifications given in Table 5.5. Here, the con-cept Society/Issues is triggered by the word “Reuters.” In turn, the conceptBusiness/Marketing and Advertising/Consulting/Sales is triggered by the


Phd Thesis - [PDF Document] (94)

# ODP concept

1 Business/Mining and Drilling/Mineral Exploration and Extraction

2 Business/Mining and Drilling

3 Business/Mining and Drilling/Mineral Exploration and Extraction/

Base Metals

4 Science/Technology/Mining

5 Business/Mining and Drilling/Consulting

6 Business/Investing/Commodities, Futures/Precious Metals

7 Shopping

8 Business/Mining and Drilling/Mining Equipment

9 Business/Investing/Commodities, Futures/Precious Metals/Gold

10 Science/Technology/Mining/Investments

Table 5.4: The top ten ODP concepts generated for the sentence “Cominco’s shareof production was 43,000 short tons of copper, 340,000 ounces of silver and 800ounces of gold.”

name of the company spokesman, Don Townson. As it happens, a sales consultingcompany named “Townson & Alexander Consulting Services” is catalogued underthis concept. Based on the crawled content of this site, the word “Townson” andother sales-related words in the context (e.g., “percent,” “owned,” “optimistic,”and “consortium”) taken together yield this concept in the results. Again, thissales-related concept is hardly useful for categorizing copper-related documents,and features based on it would therefore not be selected.

5.3.2 The Effect of Feature Generation

Table 5.6 shows the results of using feature generation for text categorization,with significant improvements (p < 0.05) shown in bold. We consistently observedlarger improvements in macro-averaged BEP, which is dominated by categoriza-tion effectiveness on small categories. This goes in line with our expectationsthat the contribution of external knowledge should be especially prominent forcategories with few training examples. As can be readily seen, categorizationperformance was improved for all datasets, with notably high improvements forReuters RCV1, OHSUMED and Movies. Given the performance plateau cur-rently reached by the best text categorizers, these results clearly demonstrate theadvantage of knowledge-based feature generation.


Phd Thesis - [PDF Document] (95)

# ODP concept

1 Business/Mining and Drilling/Mineral Exploration and Extraction/

Base Metals2 Business/Mining and Drilling/Mineral Exploration and Extraction

3 Business/Mining and Drilling

4 Business/Mining and Drilling/Consulting

5 Society/Issues

6 Regional/North America/Canada/British Columbia/Localities/

Kimberley7 Science/Technology/Mining

8 Business/Marketing and Advertising/Consulting/Sales

9 Regional/North America/Canada/Quebec/Regions/Northern Quebec

10 Science/Environment/Mining

Table 5.5: The top ten ODP concepts generated for the sentence “‘Cominco, 29.5percent owned by a consortium led by Teck, is optimistic that the talks will soon beconcluded,’ spokesman Don Townson told Reuters.”

5.3.3 The Effect of Contextual Analysis

We now explore the various possibilities for defining document contexts for fea-ture generation, i.e., chunks of document text that are classified onto the ODP toconstruct features. Figure 5.1 shows how text categorization performance on theMovies dataset changes for various contexts. The x-axis measures context lengthin words, and the FG/words curve corresponds to applying the feature generatorto the context of that size. With these word-level contexts, maximum perfor-mance is achieved when using pairs of words (x=2). The Baseline line representstext categorization without feature generation. The FG/doc line shows what hap-pens when the entire document is used as a single context. In this case, the resultsare somewhat better than without feature generation (Baseline), but are still infe-rior to the more fine-grained word-level contexts (FG/words). However, the bestperformance by far is achieved with the multi-resolution approach (FG/multi),in which we use a series of linguistically motivated chunks of text, starting withindividual words, and then generating features from sentences, paragraphs, andfinally the entire document.

5.3.4 The Effect of Knowledge Breadth

In the experiments reported in Section 5.3.2 we performed feature generation us-ing the entire ODP. It is interesting to observe, however, that four out of the five


Phd Thesis - [PDF Document] (96)

Dataset Baseline Feature Improvementgeneration vs. baseline

micro macro micro macro micro macroBEP BEP BEP BEP BEP BEP

Reuters-2157810 categories 0.925 0.874 0.930 0.884 +0.5% +1.1%90 categories 0.877 0.602 0.880 0.614 +0.3% +2.0%

RCV1Industry-16 0.642 0.595 0.648 0.613 +0.9% +3.0%Industry-10A 0.421 0.335 0.457 0.420 +8.6% +25.4%Industry-10B 0.489 0.528 0.530 0.560 +8.4% +6.1%Industry-10C 0.443 0.414 0.468 0.463 +5.6% +11.8%Industry-10D 0.587 0.466 0.588 0.496 +0.2% +6.4%Industry-10E 0.648 0.605 0.657 0.639 +1.4% +5.6%Topic-16 0.836 0.591 0.840 0.660 +0.5% +11.7%Topic-10A 0.796 0.587 0.803 0.692 +0.9% +17.9%Topic-10B 0.716 0.618 0.727 0.655 +1.5% +6.0%Topic-10C 0.687 0.604 0.694 0.618 +1.0% +2.3%Topic-10D 0.829 0.673 0.836 0.687 +0.8% +2.1%Topic-10E 0.758 0.742 0.762 0.756 +0.5% +1.9%

OHSUMEDOHSUMED-10A 0.518 0.417 0.537 0.479 +3.7% +14.9%OHSUMED-10B 0.656 0.500 0.659 0.548 +0.5% +9.6%OHSUMED-10C 0.539 0.505 0.547 0.540 +1.5% +6.9%OHSUMED-10D 0.683 0.515 0.688 0.549 +0.7% +6.6%OHSUMED-10E 0.442 0.542 0.452 0.573 +2.3% +5.7%

20NG 0.854 0.858 +0.5%Movies 0.813 0.842 +3.6%

Table 5.6: Text categorization with and without feature generation

datasets we used have a fairly narrow scope.8 Specifically, both Reuters datasets(Reuters-21578 and RCV1) contain predominantly economic news and thereforematch the scope of the Top/Business branch of the ODP. Similarly, Movie Re-views contains opinions about movies, and therefore fits the scope of Top/Arts.OHSUMED contains medical documents, which can be modelled within the scopeof Top/Health and Top/Science. In the light of this, it could be expected thatrestricting the feature generator to a particular ODP branch that corresponds

8The 20 Newsgroups dataset consists of 20 diverse categories, each of which corresponds toone or more ODP branches.


Phd Thesis - [PDF Document] (97)














l BE


Context window length (words)



Figure 5.1: Varying context length (Movies)

to the scope of the test collection would result in much better categorizationaccuracy due to the elimination of noise in “unused” ODP branches.

Experimental results (Table 5.7) disprove this hypothesis. As can be seen, inthe absolute majority of cases the improvement over the baseline is much largerwhen the entire ODP is used (cf. Table 5.6). These findings show the superiorityof wide general-purpose knowledge over its domain-specific subsets.

5.3.5 The Utility of Feature Selection

Under the experimental settings defined in Section 5.2.3, feature generation con-structed approximately 4–5 times as many features as are in the bag of words(after rare features that occurred in less than 3 documents were removed). Weconducted two experiments to understand the effect of feature selection in con-junction with feature generation.

Since earlier studies found that feature selection from the bag of words im-pairs SVM performance (Section 3.4.2), in our first experiment we apply featureselection only to the generated features and use the selected ones to augment the(entire) bag of words. In Figures 5.2 and 5.3, the BOW line depicts the baselineperformance without generated features, while the BOW+GEN curve shows theperformance of the bag of words augmented with progressively larger fractions ofgenerated features (sorted by information gain). For both datasets, the perfor-mance peaks when only a small fraction of the generated features are used, whileretaining more generated features has a noticeable detrimental effect.

Our second experiment examined the performance of the generated featuresalone, without the bag of words (GEN curve in Figures 5.2 and 5.3). For Movies,


Phd Thesis - [PDF Document] (98)

Dataset Domain-specific ODP subset Full ODPSubset micro macro micro macrodescription BEP BEP BEP BEP

Reuters-21578 Top/Business10 categories +0.4% +0.6% +0.5% +1.1%90 categories +0.1% +1.2% +0.3% +2.0%

RCV1 Top/BusinessIndustry-16 +1.9% +2.2% +0.9% +3.0%Topic-16 +0.5% +1.4% +0.5% +11.7%

OHSUMED Top/HealthOHSUMED-10A +2.1% +1.7% +3.7% +14.9%OHSUMED-10B +0.2% +1.2% +0.5% +9.6%OHSUMED-10C +1.7% +2.8% +1.5% +6.9%OHSUMED-10D +0.3% +1.9% +0.7% +6.6%OHSUMED-10E +2.7% +1.8% +2.3% +5.7%

OHSUMED Top/Health +Top/Science

OHSUMED-10A +5.4% +3.6% +3.7% +14.9%OHSUMED-10B +0.3% +3.4% +0.5% +9.6%OHSUMED-10C +0.6% +3.8% +1.5% +6.9%OHSUMED-10D +0.9% +5.8% +0.7% +6.6%OHSUMED-10E +1.6% +1.8% +2.3% +5.7%

Movies Top/Arts +2.6% +3.6%

Table 5.7: Text categorization with and without feature generation, when only asubset of the ODP is used

discarding the BOW features leads to somewhat worse performance, but thedecrease is far less significant than what could be expected—using only the gen-erated features we lose less than 3% in BEP compared with the BOW baseline.For 20NG, a similar experiment sacrifices about 10% of the BOW performance, asthis dataset is known to have a very diversified vocabulary, for which many stud-ies found feature selection to be particularly harmful. Similarly, for OHSUMED,using only the generated features sacrifices up to 15% in performance, reinforcingthe value of precise medical terminology that is discarded in this experiment.However, the situation is reversed for both Reuters datasets. For Reuters-21578,the generated features alone yield a 0.3% improvement in micro- and macro-BEPfor 10 categories, while for 90 categories they only lose 0.3% in micro-BEP and3.5% in macro-BEP compared with the bag of words. For RCV1/Industry-16, dis-posing of the bag of words reduces BEP performance by 1–3%. Surprisingly, for


Phd Thesis - [PDF Document] (99)














l BE


Fraction of generated features selected



Figure 5.2: Feature selection (Movies)














l BE


Fraction of generated features selected



Figure 5.3: Feature selection (RCV1/Topic-16)

RCV1/Topic-16 (Figure 5.3) the generated features per se command a 10.8% im-provement in macro-BEP, rivalling the performance of BOW+GEN, which gainsonly another 1% (Table 5.6). We interpret these findings as further reinforcementthat the generated features improve the quality of the representation.

5.3.6 The Effect of Category Size

We saw in Section 5.3.2 that feature generation greatly improves text categoriza-tion for smaller categories, as can be evidenced in the greater improvements inmacro-BEP. To explore this phenomenon further, we depict in Figures 5.4 and 5.5


Phd Thesis - [PDF Document] (100)






0 100 200 300 400 500 600 700





ent (

in %


Category size (binned)






(0) (1) (0) (0) (0) (0) (0) (0) (1)

Figure 5.4: RCV1 (Industry): Average improvement versus category size

the relation between the category size and the improvement due to feature gen-eration for RCV1 (the number of categories in each bin appears in parenthesesabove the bars). To this end, we pooled together the categories that comprisedthe individual sets (10A–10E) in the Industry and Topic groups, respectively.

As we can readily see, smaller categories tend to benefit more from knowledge-based feature generation. These graphs also explain the more substantial improve-ments observed for Industry categories compared to Topic categories—as can beseen from the graphs, Topic categories are larger than Industry categories, andthe average size of Topic categories (among those we used in this study) is almost6 times larger than that of Industry categories.

5.3.7 The Effect of Feature Generation for ClassifyingShort Documents

We conjectured that knowledge-based feature generation might be particularlyuseful for classifying short documents. To evaluate this hypothesis, we used thedatasets defined in Section 5.1.6.

Table 5.8 presents the results of this experiment. As we can see, in the ma-jority of cases (except for RCV1 Topic category sets), feature generation leads togreater improvement on short documents than on regular documents. Notably,the improvements are particularly high for OHSUMED, where “pure” experimen-tation on short documents is possible (see Section 5.1.6).


Phd Thesis - [PDF Document] (101)










0 500 1000 1500 2000 2500 3000 3500 4000





ent (

in %


Category size (binned)






Figure 5.5: RCV1 (Topic): Average improvement versus category size

5.3.8 Processing Time

Using the ODP as a source of background knowledge requires additional compu-tation. This extra computation includes the (one-time) preprocessing step wherethe feature generator is built, as well as the actual feature generation performedon documents prior to text categorization. The processing times reported be-low were measured on a workstation with dual Xeon 2.2 GHz CPU and 2 GbRAM running the Microsoft Windows XP Professional operating system (ServicePack 1).

Parsing the ODP structure (file structure.rdf.u8) took 3 minutes. Parsingthe list of ODP URLs (file content.rdf.u8) required 3 hours, and parsing thecrawled ODP data (meta-documents collected from all cataloged URLs) required2.6 days. Attribute selection for ODP concepts took 1.5 hours. The cumulativeone-time expenditure for building the feature generator was therefore just under3 days (not counting the actual Web crawling that was performed beforehand).

We benchmarked feature generation in two scenarios—individual words and10-word windows. In the former case, the feature generator classified approxi-mately 310 words per second, while in the latter case it classified approximately45 10-word windows per second (i.e., 450 words per second).9 These times con-stitute the additional overhead required by feature generation compared withregular text categorization. Table 5.9 lists the sizes of the test collections we ex-perimented with (see Section 5.1). To speed up experimentation, we used subsetsof the entire RCV1 and OHSUMED collections; these subsets are comparable in

9Classifying word windows is more efficient due to the sharing of data structures whenprocessing the words in a single context.


Phd Thesis - [PDF Document] (102)

Dataset Short documents Full documentsBaseline Feature Improvement Improvement

generation vs. baseline vs. baselinemicro macro micro macro micro macro micro macroBEP BEP BEP BEP BEP BEP BEP BEP

Reuters-2157810 categories 0.868 0.774 0.868 0.777 +0.0% +0.4% +0.5% +1.1%90 categories 0.793 0.479 0.794 0.498 +0.1% +4.0% +0.3% +2.0%

RCV1Industry-16 0.454 0.400 0.466 0.415 +2.6% +3.7% +0.9% +3.0%Industry-10A 0.249 0.199 0.278 0.256 +11.6% +28.6% +8.6% +25.4%Industry-10B 0.273 0.292 0.348 0.331 +27.5% +13.4% +8.4% +6.1%Industry-10C 0.209 0.199 0.295 0.308 +41.1% +54.8% +5.6% +11.8%Industry-10D 0.408 0.361 0.430 0.431 +5.4% +19.4% +0.2% +6.4%Industry-10E 0.450 0.410 0.490 0.459 +8.9% +12.2% +1.4% +5.6%Topic-16 0.763 0.529 0.763 0.534 +0.0% +0.9% +0.5% +11.7%Topic-10A 0.718 0.507 0.720 0.510 +0.3% +0.6% +0.9% +17.9%Topic-10B 0.647 0.560 0.644 0.560 -0.5% +0.0% +1.5% +6.0%Topic-10C 0.551 0.471 0.561 0.475 +1.8% +0.8% +1.0% +2.3%Topic-10D 0.729 0.535 0.730 0.553 +0.1% +3.4% +0.8% +2.1%Topic-10E 0.643 0.636 0.656 0.646 +2.0% +1.6% +0.5% +1.9%

OHSUMEDOHSUMED-10A 0.302 0.221 0.357 0.253 +18.2% +14.5% +3.7% +14.9%OHSUMED-10B 0.306 0.187 0.348 0.243 +13.7% +29.9% +0.5% +9.6%OHSUMED-10C 0.441 0.296 0.494 0.362 +12.0% +22.3% +1.5% +6.9%OHSUMED-10D 0.441 0.356 0.448 0.419 +1.6% +17.7% +0.7% +6.6%OHSUMED-10E 0.164 0.206 0.211 0.269 +28.7% +30.6% +2.3% +5.7%

20NG 0.699 0.740 +5.9% +0.5%

Table 5.8: Text categorization of short documents with and without feature genera-tion (the improvement percentage in the two rightmost columns is computed relativeto the baseline shown in Table 5.6)

size with 20 Newsgroups and Reuters-21578.

In the light of the improvements in categorization accuracy due to featuregeneration, we believe that the extra processing time is well compensated for. Inoperational text categorization systems, documents rarely arrive in huge batchesof hundreds of thousands at a time. For example, the RCV1 dataset contains allEnglish-language news items published by Reuters over the period of one year.Therefore, in practical settings, once the classification model has been trained, thenumber of documents it needs to classify per time unit is much more reasonable,


Phd Thesis - [PDF Document] (103)

Dataset Number of documents Number of words10

20NG 19,997 5.5 millionMovies 1,400 0.95 millionReuters-21578 21,902 2.8 millionRCV1- full 804,414 196 million- used in this study 23,149 5.5 millionOHSUMED- full 348,566 57 million- used in this study 20,000 3.7 million

Table 5.9: Test collection sizes

and can be easily facilitated by our system.

5.4 Wikipedia-based Feature Generation

In this section we evaluate the feature generator based on Wikipedia.

5.4.1 Qualitative Analysis of Feature Generation

We start with demonstrating the results of feature generation on a number ofactual examples.

Feature Generation per se

To illustrate our approach, we show features generated for several text fragments.Whenever applicable, we provide short explanations of the generated concepts; inmost cases, the explanations are taken from Wikipedia itself (Wikipedia, 2006).

• Text: “Wal-Mart supply chain goes real time”

Sample generated features:

– Wal-Mart

– Sam Walton — Wal-Mart founder

– Sears Holdings Corporation, Target Corporation, Albertsons

— prominent competitors of Wal-Mart

10Measured using the ‘wc’ utility available on UNIX systems.


Phd Thesis - [PDF Document] (104)

– RFID — Radio Frequency Identification, a technology that Wal-Martuses very extensively to manage its stock

– Hypermarket — superstore (a general concept, of which Wal-Mart isa specific example)

– United Food and Commercial Workers — a labor union that hasbeen trying to organize Wal-Mart’s workers

• Text: “Scientific methods in biology”

Sample generated features:

– Biology

– Scientific classification

– Science

– Chemical biology

– Binomial nomenclature — the formal method of naming species inbiology

– Nature (journal)

– Social sciences

– Philosophy of biology

– Scientist

– History of biology

• Text: “With quavering voices, parents and grandparents of those killed atthe World Trade Center read the names of the victims in a solemn recitationtoday, marking the third anniversary of the terror attacks. The ceremonyis one of many planned in the United States and around the world to honorthe memory of the nearly 3,000 victims of 9/11.”

Sample generated features:

– September 11, 2001 attack memorials and services

– United Airlines Flight 93 — one of the four flights hijacked onSeptember 11, 2001

– Aftermath of the September 11, 2001 attacks

– World Trade Center

– September 11, 2001 attacks

– Oklahoma City bombing — a terrorist attack in Oklahoma City in1995


Phd Thesis - [PDF Document] (105)

– World Trade Center bombing

– Arlington National Cemetery — an American military cemetery

– World Trade Center site

– Jewish bereavement

• Text: “A group of European-led astronomers has made a photograph ofwhat appears to be a planet orbiting another star. If so, it would be the firstconfirmed picture of a world beyond our solar system.”

Sample generated features:

– Planet

– Solar system

– Astronomy

– Planetary orbit

– Extrasolar planet

– Pluto

– Jupiter

– Neptune

– Minor planet

– Mars

• Text: “Nearly 70 percent of Americans say they are careful about whatthey eat, and even more say diet is essential to good health, according to anew nationwide health poll in which obesity ranked second among the biggesthealth concerns.”

Sample generated features:

– Veganism — a philosophy of avoiding animal-derived food

– Vegetarianism

– Obesity

– Atkins Nutritional Approach

– Binge eating disorder

– Dick Gregory — an American nutritionist

– Nutrition

– Super Size Me — a documentary film about an individual who ateonly McDonald’s fast food for one full month


Phd Thesis - [PDF Document] (106)

– Health insurance

– Eating disorder

• Text: “U.S. intelligence cannot say conclusively that Saddam Hussein hasweapons of mass destruction, an information gap that is complicating WhiteHouse efforts to build support for an attack on Saddam’s Iraqi regime. TheCIA has advised top administration officials to assume that Iraq has someweapons of mass destruction. But the agency has not given President Bush a“smoking gun,” according to U.S. intelligence and administration officials.”

Sample generated features:

– Iraq disarmament crisis

– Yellowcake forgery — falsified intelligence documents about Iraq’salleged attempt to purchase yellowcake uranium

– Senate Report of Pre-War Intelligence on Iraq

– Iraq and weapons of mass destruction

– Iraq Survey Group

– September Dossier — a paper on Iraq’s weapons of mass destructionpublished by the UK government in 2002

– Iraq war

– Scott Ritter — UN weapons inspector in Iraq

– Iraq War Rationale

– Operation Desert Fox — US and UK joint military campaign inIraq in 1998

• Text: ‘The development of T-cell leukaemia following the otherwise suc-cessful treatment of three patients with X-linked severe combined immunedeficiency (X-SCID) in gene-therapy trials using haematopoietic stem cellshas led to a re-evaluation of this approach. Using a mouse model for genetherapy of X-SCID, we find that the corrective therapeutic gene IL2RG itselfcan act as a contributor to the genesis of T-cell lymphomas, with one-thirdof animals being affected. Gene-therapy trials for X-SCID, which have beenbased on the assumption that IL2RG is minimally oncogenic, may thereforepose some risk to patients.”

Sample generated features:

– Leukemia

– Severe combined immunodeficiency


Phd Thesis - [PDF Document] (107)

– Cancer

– Non-Hodgkin lymphoma — a particular cancer type


– ICD-10 Chapter II: Neoplasms; Chapter III: Diseases of the

blood and blood-forming organs, and certain disorders involv-

ing the immune mechanism — a disease code of the ICD (InternationalStatistical Classification of Diseases and Related Health Problems)

– Bone marrow transplant

– Immunosuppressive drug

– Acute lymphoblastic leukemia

– Multiple sclerosis

• Finally, it is particularly interesting to juxtapose the features generated forfragments that contain ambiguous words. To this end, we show featuresgenerated for two phrases that contain the word “bank” in two differentsenses, “Bank of America” (financial institution) and “Bank of Amazon”(river bank). As can be readily seen, our feature generation methodology iscapable of performing word sense disambiguation by considering ambiguouswords in the context of their neighbors.

Text: “Bank of America”

Sample generated features:

– Bank

– Bank of America

– Bank of America Plaza (Atlanta)

– Bank of America Plaza (Dallas)

– MBNA — a bank holding company acquired by Bank of America

– VISA (credit card)

– Bank of America Tower, New York City


– MasterCard

– Bank of America Corporate Center

Text: “Bank of Amazon”

Sample generated features:


Phd Thesis - [PDF Document] (108)

– Amazon River

– Amazon Basin

– Amazon Rainforest


– Rainforest

– Atlantic Ocean

– Brazil

– Loreto Region - a region in Peru, located in the Amazon Rainforest

– River

– Economy of Brazil

• As another example, consider a pair of contexts that contain the word“jaguar”, where the first context contains this ambiguous word in the senseof a car model, and the second one—in the sense of an animal.

Text: “Jaguar car models”

Sample generated features:

– Jaguar (car)

– Jaguar (S-Type) — a particular Jaguar car model

– Jaguar X-type — a particular Jaguar car model

– Jaguar E-Type — a particular Jaguar car model

– Jaguar XJ — a particular Jaguar car model

– Daimler Motor Company — a car manufacturing company that be-came a part of Jaguar in 1960

– British Leyland Motor Corporation - another vehicle manufactur-ing company that merged with Jaguar

– Luxury vehicles

– V8 engine — an internal combustion engine used in some Jaguar carmodels

– Jaguar Racing — a Formula One team used by Jaguar to promoteits brand name

Text: “Jaguar (Panthera onca)”

Sample generated features:

– Jaguar


Phd Thesis - [PDF Document] (109)

– Felidae — a family that include lions, tigers, jaguars, and other re-lated feline species

– Black panther

– Leopard

– Puma

– Tiger

– Panthera hybrid

– Cave lion

– American lion

– Kinkajou — another carnivore mammal

Using Inter-article Links for Generating Additional Features

In Section 4.2.3, we presented an algorithm that generates additional featuresusing inter-article links as relations between concepts. In what follows, we showa series of text fragments, where for each fragment we show (a) features generatedwith the regular FG algorithm, (b) features generated using Wikipedia links, and(c) more general features generated using links. As we can see from the examples,the features constructed from the links are in most cases highly relevant to theinput text.

• Text: “Google search”

Regular feature generation:

– Search engine

– Google Video

– Google

– Google (search)

– Google Maps

– Google Desktop

– Google (verb)

– Google News

– Search engine optimization

– Spamdexing — search engine spamming

Features generated using links:


Phd Thesis - [PDF Document] (110)

– PageRank

– AdWords

– AdSense

– Gmail

– Google Platform

– Website

– Sergey Brin

– Google bomb

– MSN Search

– Nigritude ultramarine — a meaningless phrase used in a searchengine optimization contest in 2004

More general features only:

– Website

– Mozilla Firefox

– Portable Document Format

– Algorithm

– World Wide Web

• Text: “artificial intelligence”

Regular feature generation:

– Artificial intelligence

– A.I. (film)

– MIT Computer Science and Artificial Intelligence Laboratory

– Artificial life

– Strong AI

– Swarm intelligence

– Computer Science

– Frame problem

– Cognitive science

– Carl Hewitt

Features generated using links:


Phd Thesis - [PDF Document] (111)

– Robot

– John McCarthy (computer scientist)

– Artificial consciousness

– Marvin Minsky

– Planner programming language

– Actor model — a model of concurrent computation formulated byCarl Hewitt and his colleagues

– Logic

– Scientific Community Metaphor

– Natural language processing

– Lisp programming language

More general features only:

– Robot

– Massachusetts Institute of Technology

– Psychology

– Consciousness

– Lisp programming language

• Text: “Israel terror”

Regular feature generation:

– Israel

– Palestinian political violence

– Terrorism

– Labour (Israel)

– Terrorism against Israel

– Israel Defense Forces

– Shabak

– Steve Israel — an American politician who has worked extensivelyon military and terrorism-related issues; his interests include nationalsecurity and the State of Israel

– Israeli peace camp


Phd Thesis - [PDF Document] (112)

– Agudat Israel — a religious political party in Israel, which has re-cently become more conscious of issues related to Israel’s security

Features generated using links:

– Oslo Accords

– Al-Aqsa Intifada

– Israeli-Palestinian conflict

– 1982 Lebanon War

– British Mandate of Palestine

– Israel Border Police

– Israel’s unilateral disengagement plan

– History of Israel

– Israeli Security Forces

– Israel-Jordan Treaty of Peace

More general features only:

– Jew

– Gaza Strip

– West Bank

– British Mandate of Palestine

– Six-Day War

• Text: “programming tools”

Regular feature generation:

– Tool

– Programming tool

– Computer software

– Integrated development environment

– Computer-aided software engineering

– Macromedia Flash

– Borland

– Game programmer

– C programming language


Phd Thesis - [PDF Document] (113)

– Performance analysis

Features generated using links:

– Compiler

– Debugger

– Source code

– Software engineering

– Microsoft

– Revision control

– Scripting language


– Make

– Linux

More general features only:

– Microsoft

– Software engineering

– Linux

– Compiler


• Text: “A group of European-led astronomers has made a photograph ofwhat appears to be a planet orbiting another star. If so, it would be the firstconfirmed picture of a world beyond our solar system.”

Regular feature generation:

– Planet

– Solar system

– Astronomy

– Planetary orbit

– Extrasolar planet

– Pluto

– Jupiter

– Neptune


Phd Thesis - [PDF Document] (114)

– Minor planet

– Mars

Features generated using links:

– Asteroid

– Earth

– Oort cloud — a postulated cloud of comets

– Comet

– Sun

– Saturn

– Moon

– Mercury (planet)

– Asteroid belt

– Orbital period

More general features only:

– Earth

– Moon

– Asteroid

– Sun

– National Aeronautics and Space Administration

• Text: “Nearly 70 percent of Americans say they are careful about whatthey eat, and even more say diet is essential to good health, according to anew nationwide health poll in which obesity ranked second among the biggesthealth concerns.”

Regular feature generation:

– Veganism — a philosophy of avoiding animal-derived food

– Vegetarianism

– Obesity

– Atkins Nutritional Approach

– Binge eating disorder

– Dick Gregory — an American nutritionist


Phd Thesis - [PDF Document] (115)

– Nutrition

– Super Size Me — a documentary film about an individual who eatsonly McDonald’s fast food for one full month

– Health insurance

– Eating disorder

Features generated using links:

– Raw food diet

– Diabetes mellitus

– Healthy eating

– Body mass index

– Omega-3 fatty acid — an important nutritional component

– Dieting

– Milk

– United States — this classification is quite interesting, as the issuediscussed in the input text fragment is very characteristic of the Amer-ican life style

– Hypertension

– Egg (food)

More general features only:

– United States

– Diabetes mellitus

– Cancer

– Food

– McDonald’s

5.4.2 The Effect of Feature Generation

Table 5.10 shows the results of using Wikipedia-based feature generation, withsignificant improvements (p < 0.05) shown in bold. We consistently observedlarger improvements in macro-averaged BEP, which is dominated by categoriza-tion effectiveness on small categories. This goes in line with our expectationsthat the contribution of encyclopedic knowledge should be especially prominent


Phd Thesis - [PDF Document] (116)

Dataset Baseline Wikipedia Improvementmicro macro micro macro micro macro


Reuters-21578 (10 cat.) 0.925 0.874 0.932 0.887 +0.8% +1.5%Reuters-21578 (90 cat.) 0.877 0.602 0.883 0.603 +0.7% +0.2%

RCV1 Industry-16 0.642 0.595 0.645 0.617 +0.5% +3.7%RCV1 Industry-10A 0.421 0.335 0.448 0.437 +6.4%+30.4%RCV1 Industry-10B 0.489 0.528 0.523 0.566 +7.0% +7.2%RCV1 Industry-10C 0.443 0.414 0.468 0.431 +5.6% +4.1%RCV1 Industry-10D 0.587 0.466 0.595 0.459 +1.4% -1.5%RCV1 Industry-10E 0.648 0.605 0.641 0.612 -1.1% +1.2%RCV1 Topic-16 0.836 0.591 0.843 0.661 +0.8% +11.8%RCV1 Topic-10A 0.796 0.587 0.798 0.682 +0.3% +16.2%RCV1 Topic-10B 0.716 0.618 0.723 0.656 +1.0% +6.1%RCV1 Topic-10C 0.687 0.604 0.699 0.618 +1.7% +2.3%RCV1 Topic-10D 0.829 0.673 0.839 0.688 +1.2% +2.2%RCV1 Topic-10E 0.758 0.742 0.765 0.755 +0.9% +1.8%

OHSUMED-10A 0.518 0.417 0.538 0.492 +3.9%+18.0%OHSUMED-10B 0.656 0.500 0.667 0.534 +1.7% +6.8%OHSUMED-10C 0.539 0.505 0.545 0.522 +1.1% +3.4%OHSUMED-10D 0.683 0.515 0.692 0.546 +1.3% +6.0%OHSUMED-10E 0.442 0.542 0.462 0.575 +4.5% +6.1%

20NG 0.854 0.862 +1.0%

Movies 0.813 0.842 +3.6%

Table 5.10: The effect of feature generation

for categories with few training examples. Categorization performance was im-proved for virtually all datasets, with notable improvements of up to 30.4% forRCV1 and 18% for OHSUMED. Using the Wilcoxon test, we found that theWikipedia-based classifier is significantly superior to the baseline with p < 10−5

in both micro- and macro-averaged cases. Given the performance plateau cur-rently reached by the best text categorizers, these results clearly demonstrate theadvantage of knowledge-based feature generation.


Phd Thesis - [PDF Document] (117)

Wikipedia snapshot Wikipedia snapshotas of November 11, 2005 as of March 23, 2006

Combined article text 1.8 Gb 2.9 GbNumber of articles 910,989 1,187,839Concepts used 171,332 241,393Distinct terms 296,157 389,202

Table 5.11: Comparison of two Wikipedia snapshots

5.4.3 The Effect of Knowledge Breadth

Wikipedia is being constantly expanded with new material as volunteer editorscontribute new articles and extend the existing ones. Consequently, we conjec-tured that such addition of information should be beneficial for feature generation,as it would rely on a larger knowledge base.

To test this assumption, we acquired a new Wikipedia snapshot as ofMarch 26, 2006. Table 5.11 presents a comparison in the amount of informationbetween two Wikipedia snapshots we used. Table 5.12 shows the effect of featuregeneration using the newer snapshot. As we can see, using the larger amount ofknowledge leads on average to greater improvements in text categorization per-formance. Although the difference between the performance of the two versionsis admittedly small, it is consistent across datasets (a similar situation happenswhen assessing the role of external knowledge for computing semantic relatedness,see Section 6.2.2).

5.4.4 Classifying Short Documents

We conjectured that Wikipedia-based feature generation should be particularlyuseful for classifying short documents, similarly to using ODP (cf. Section 5.3.7).

Table 5.13 presents the results of this evaluation on the datasets defined inSection 5.1.6. In the majority of cases, feature generation yielded greater im-provement on short documents than on regular documents. Notably, the im-provements are particularly high for OHSUMED, where “pure” experimentationon short documents is possible (see Section 5.1.6). According to the Wilcoxontest, the Wikipedia-based classifier is significantly superior to the baseline withp < 2 · 10−6. These findings confirm our hypothesis that encyclopedic knowl-edge should be particularly useful when categorizing short documents, which areinadequately represented by the standard bag of words.


Phd Thesis - [PDF Document] (118)

Dataset Baseline Wikipedia Improvement Improvement(26/03/06) (26/03/06) (05/11/05)

micro macro micro macro micro macro micro macro


Reuters-21578 (10 cat.) 0.925 0.874 0.935 0.891 +1.1% +1.9% +0.8% +1.5%Reuters-21578 (90 cat.) 0.877 0.602 0.883 0.600 +0.7% -0.3% +0.7% +0.2%

RCV1 Industry-16 0.642 0.595 0.648 0.616 +0.9% +3.5% +0.5% +3.7%RCV1 Industry-10A 0.421 0.335 0.457 0.450 +8.6% +34.3% +6.4% +30.4%RCV1 Industry-10B 0.489 0.528 0.527 0.559 +7.8% +5.9% +7.0% +7.2%RCV1 Industry-10C 0.443 0.414 0.458 0.424 +3.4% +2.4% +5.6% +4.1%RCV1 Industry-10D 0.587 0.466 0.607 0.448 +3.4% -3.9% +1.4% -1.5%RCV1 Industry-10E 0.648 0.605 0.649 0.607 +0.2% +0.3% -1.1% +1.2%RCV1 Topic-16 0.836 0.591 0.842 0.659 +0.7% +11.5% +0.8% +11.8%RCV1 Topic-10A 0.796 0.587 0.802 0.689 +0.8% +17.4% +0.3% +16.2%RCV1 Topic-10B 0.716 0.618 0.725 0.660 +1.3% +6.8% +1.0% +6.1%RCV1 Topic-10C 0.687 0.604 0.697 0.627 +1.5% +3.8% +1.7% +2.3%RCV1 Topic-10D 0.829 0.673 0.838 0.687 +1.1% +2.1% +1.2% +2.2%RCV1 Topic-10E 0.758 0.742 0.762 0.752 +0.5% +1.3% +0.9% +1.8%

OHSUMED-10A 0.518 0.417 0.545 0.490 +5.2% +17.5% +3.9% +18.0%OHSUMED-10B 0.656 0.500 0.667 0.529 +1.7% +5.8% +1.7% +6.8%OHSUMED-10C 0.539 0.505 0.553 0.527 +2.6% +4.4% +1.1% +3.4%OHSUMED-10D 0.683 0.515 0.694 0.550 +1.6% +6.8% +1.3% +6.0%OHSUMED-10E 0.442 0.542 0.461 0.588 +4.3% +8.5% +4.5% +6.1%

20NG 0.854 0.859 +0.6% +1.0%

Movies 0.813 0.850 +4.5% +3.6%

Average +2.50% +6.84% +2.11% +6.71%

Table 5.12: The effect of feature generation using a newer Wikipedia snapshot(dated March 26, 2006)

5.4.5 Using Inter-article links as Concept Relations

Using inter-article links for generating additional features, we observed furtherimprovements in text categorization performance on short documents. As we cansee in Table 5.14, in the absolute majority of cases using links to generate moregeneral features only is a superior strategy.


Phd Thesis - [PDF Document] (119)

Dataset Baseline Wikipedia Improvementmicro macro micro macro micro macro


Reuters-21578 (10 cat.) 0.868 0.774 0.877 0.793 +1.0% +2.5%Reuters-21578 (90 cat.) 0.793 0.479 0.803 0.506 +1.3% +5.6%

RCV1 Industry-16 0.454 0.400 0.481 0.437 +5.9% +9.2%RCV1 Industry-10A 0.249 0.199 0.293 0.256 +17.7% +28.6%RCV1 Industry-10B 0.273 0.292 0.337 0.363 +23.4% +24.3%RCV1 Industry-10C 0.209 0.199 0.294 0.327 +40.7% +64.3%RCV1 Industry-10D 0.408 0.361 0.452 0.379 +10.8% +5.0%RCV1 Industry-10E 0.450 0.410 0.474 0.434 +5.3% +5.9%RCV1 Topic-16 0.763 0.529 0.769 0.542 +0.8% +2.5%RCV1 Topic-10A 0.718 0.507 0.725 0.544 +1.0% +7.3%RCV1 Topic-10B 0.647 0.560 0.643 0.564 -0.6% +0.7%RCV1 Topic-10C 0.551 0.471 0.573 0.507 +4.0% +7.6%RCV1 Topic-10D 0.729 0.535 0.735 0.563 +0.8% +5.2%RCV1 Topic-10E 0.643 0.636 0.670 0.653 +4.2% +2.7%

OHSUMED-10A 0.302 0.221 0.405 0.299 +34.1% +35.3%OHSUMED-10B 0.306 0.187 0.383 0.256 +25.2% +36.9%OHSUMED-10C 0.441 0.296 0.528 0.413 +19.7% +39.5%OHSUMED-10D 0.441 0.356 0.460 0.402 +4.3% +12.9%OHSUMED-10E 0.164 0.206 0.219 0.280 +33.5% +35.9%

20NG 0.699 0.749 +7.1%

Table 5.13: Feature generation for short documents


Phd Thesis - [PDF Document] (120)

Dataset Baseline Wikipedia Wikipedia Wikipedia+ links + links

(more generalfeatures only)

micro macro micro macro micro macro micro macroBEP BEP BEP BEP BEP BEP BEP BEP

Reuters-21578 (10 cat.) 0.868 0.774 0.877 0.793 0.878 0.796 0.880 0.801Reuters-21578 (90 cat.) 0.793 0.479 0.803 0.506 0.804 0.506 0.809 0.507RCV1 Industry-16 0.454 0.400 0.481 0.437 0.486 0.445 0.488 0.444RCV1 Topic-16 0.763 0.529 0.769 0.542 0.769 0.539 0.775 0.54520NG 0.699 0.749 0.753 0.756Dataset Improvement Improvement Improvement

over baseline over baseline over baselineReuters-21578 (10 cat.) – – +1.0% +2.5% +1.2% +2.8% +1.4% +3.5%Reuters-21578 (90 cat.) – – +1.3% +5.6% +1.4% +5.6% +2.0% +5.8%RCV1 Industry-16 – – +5.9% +9.2% +7.1% +11.3% +7.5% +11.0%RCV1 Topic-16 – – +0.8% +2.5% +0.8% +1.9% +1.6% +3.0%20NG – +7.1% +7.7% +8.1%

Table 5.14: Feature generation for short documents using inter-article links


Phd Thesis - [PDF Document] (121)

Chapter 6

Using Feature Generation forComputing Semantic Relatednessof Texts

How related are “cat” and “mouse”? And what about “preparing a manuscript”and “writing an article”? The ability to quantify semantic relatedness of textsunderlies many fundamental tasks in computational linguistics, including wordsense disambiguation, information retrieval, word and text clustering, and errorcorrection (Budanitsky and Hirst, 2006). Reasoning about semantic relatednessof natural language utterances is routinely performed by humans but remains anunsurmountable obstacle for computers. Humans do not judge text relatednessmerely at the level of text words. Words trigger reasoning at a much deeper levelthat manipulates concepts—the basic units of meaning that serve humans toorganize and share their knowledge. Thus, humans interpret the specific wordingof a document in the much larger context of their background knowledge andexperience. Lacking such elaborate resources, computers need alternative waysto represent texts and reason about them.

In this Chapter, we discuss the application of our feature generation method-ology to automatic assessment of semantic relatedness of words and texts(Gabrilovich and Markovitch, 2007; Gabrilovich and Markovitch, 2006a).

6.1 Explicit Semantic Analysis

In supervised text categorization, one is usually given a collection of labeled textdocuments, from which one can induce a text categorizer. Consequently, wordsthat occur in the training examples can serve as valuable features—this is howthe bag of words approach was born. There are, however, other tasks in natural


Phd Thesis - [PDF Document] (122)

language processing, in which labeled training examples can hardly be produced,because the decisions are essentially “one-off”. A notable example of such a taskis automatic assessment of semantic relatedness of words and texts. Here, givena pair of text fragments, we need to quantify their relatedness on some scale, say,between 0 and 1. In such cases, the very words of the text fragments are likelyto be of marginal usefulness, and when the two fragments are one word long, thewords are probably useless at all. This happens because all the data available tous is limited to the two input fragments, which in most cases share few words, ifat all.

Besides the inappropriateness of the bag of words, another obvious conjectureis that external knowledge is likely to be of substantial benefit for assessing se-mantic relatedness, as this is exactly the kind of knowledge that humans applyto this task. Therefore, we propose to apply our feature generation methodologyto the task of computing semantic relatedness. Given a pair of text fragmentswhose semantic relatedness needs to be established, we use the feature generatorto construct knowledge-based features for each fragment that will replace its bagof words. However, instead of using only a few top-scoring concepts, we nowconsider all the available concepts. The role of the feature generator is, therefore,to quantify the affinity of the input text fragment to each of the knowledge con-cepts. The output of feature generation in this case is a vector of weights, oneper concept, which quantifies the relevance of the concept to the text fragment.In other words, text fragments are represented in the space of all knowledge con-cepts. Since our task is inherently not supervised, we cannot perform featureselection in the conventional sense. Instead, we rely on the weights assigned tothe concepts in order to only select those concepts that are strongly relevant tothe input. Concepts that are marginally relevant to the input have their weightsdropped to zero, thus eliminating spurious associations.

We represent text fragments as a weighted mixture of a predetermined setof natural concepts, which are defined by humans themselves and can be eas-ily explained. An important advantage of our approach is thus the use of vastamounts of highly organized human knowledge. Compared to Latent SemanticAnalysis, our methodology explicitly uses the knowledge collected and organizedby humans. Compared to lexical resources such as WordNet, our methodologyleverages knowledge bases that are orders of magnitude larger and more compre-hensive. Importantly, the Web-based knowledge repositories we use in this workundergo constant development so their breadth and depth steadily increase overtime.

Viewed more generally, our methodology can be seen as building a semanticinterpreter, which maps fragments of natural language text into a weighted se-quence of concepts ordered by their relevance to the input. This way, weightedvectors of concepts that represent input texts can be viewed as their interpretation


Phd Thesis - [PDF Document] (123)

vectors. The meaning of a text fragment is thus interpreted in terms of its affin-ity with a host of knowledge concepts. Computing semantic relatedness of textsthen amounts to comparing their vectors in the space defined by the concepts,for example, using the cosine metric (Zobel and Moffat, 1998). Our semanticanalysis is explicit in the sense that we manipulate manifest concepts groundedin human cognition, rather than “latent concepts” used by LSA. Therefore, wecall our approach Explicit Semantic Analysis (ESA).

To speed up semantic interpretation, we build an inverted index, which mapseach word into a list of concepts in which it appears. The inverted index isalso used to discard insignificant associations between words and concepts byremoving those concepts whose weights for a given word are too low.

Given a text fragment, we first represent it as an attribute vector using tf.idfscheme. The semantic interpreter iterates over the text words, retrieves corre-sponding entries from the inverted index, and merges them into a weighted vectorof concepts that represents the given text. Let T = {wi} be the input text, andlet 〈vi〉 be its attribute vector, where vi is the weight of word wi. Let 〈kj〉 be aninverted index entry for word wi, where kj quantifies the strength of associationof word wi with knowledge concept cj ∈ {c0, . . . , cn} (where n is the total numberof concepts). Then, the semantic interpretation vector V for text T is a vectorof length n, in which the weight of each concept cj is defined as

∑wi∈T vi · kj.

Entries of this vector reflect the affinity of the corresponding concepts to text T .Figure 6.1 illustrates the processes of building and using the semantic interpreter.

6.2 Empirical Evaluation of Explicit Semantic


Humans have an innate ability to judge semantic relatedness of texts. Humanjudgements on a reference set of text pairs can thus be considered correct bydefinition, a kind of “gold standard” against which computer algorithms are eval-uated. Several studies measured inter-judge correlations and found them to beconsistently high (Budanitsky and Hirst, 2006; Jarmasz, 2003; Finkelstein et al.,2002a), r = 0.88 − 0.95. These findings are to be expected—after all, it is thisconsensus that allows people to understand each other. Consequently, our evalua-tion amounts to computing the correlation of ESA relatedness scores with humanjudgments.


Phd Thesis - [PDF Document] (124)

Relatedness estimation

Building Semantic Interpreter

Building weighted inverted index

Weighted list of knowledge concepts

Weighted inverted index

Using Semantic Interpreter

Semantic Interpreter



Vector comparison

Text vector: <c1, c2, …, ck>





cweightwweightc )()(

word1 wordi wordn

Weighted vector of knowledge concepts

Knowledge repository

Figure 6.1: Knowledge-based semantic interpreter

6.2.1 Test Collections

In this work, we use two datasets that to the best of our knowledge are the largestpublicly available collections of their kind.1 To assess word relatedness, we use

1Recently, Zesch and Gurevych (2006) discussed automatic creation of datasets for assessingsemantic similarity. However, the focus of their work was on automatic generation of a set of


Phd Thesis - [PDF Document] (125)

the WordSimilarity-353 collection (Finkelstein et al., 2002b; Finkelstein et al.,2002a), which contains 353 word pairs. Each pair has 13–16 human judgementsmade by individuals with university degrees having either mother-tongue-levelor otherwise very fluent command of the English language. Word pairs wereassigned relatedness scores on the scale from 0 (totally unrelated words) to 10(very much related or identical words). Judgements collected for each word pairwere then averaged to produce a single relatedness score.

For document similarity, we used a collection of 50 documents from the Aus-tralian Broadcasting Corporation’s news mail service (Lee, Pincombe, and Welsh,2005; Pincombe, 2004). The documents were between 51 and 126 words long, andcovered a variety of topics. The judges were 83 students from the University ofAdelaide, Australia, who were paid a small fee for their work. These documentswere paired in all possible ways, and each of the 1,225 pairs has 8–12 humanjudgements (averaged for each pair). To neutralize the effects of ordering, docu-ment pairs were presented in random order, and the order of documents withineach pair was randomized as well.

Importantly, instructions for human judges in both test collections specificallydirected the participants to assess the degree of relatedness of words and textsinvolved. For example, in the case of antonyms, judges were instructed to considerthem as “similar” rather than “dissimilar”.

For both test collections, we use the correlation of computer-assigned scoreswith human scores to assess the algorithm performance.

6.2.2 The Effect of External Knowledge

Table 6.1 shows the results of applying our methodology to estimating related-ness of individual words. As we can see, both ESA techniques yield substantialimprovements over previous state of the art results. Notably, ESA also achievesmuch better results than another recently introduce method based on Wikipedia(Strube and Ponzetto, 2006). We provide a detailed comparison of our approachwith this latter work in Section 7.3. Table 6.2 shows the results for computingrelatedness of entire documents.

In Section 5.4.3 we examined the effect of knowledge breadth by comparingfeature generators based on two Wikipedia versions. Here we also evaluate thebenefits of using a larger knowledge base for ESA. As we can see in both exper-iments, using a newer Wikipedia snapshot leads to better results (although thedifference between the performance of two versions is admittedly small).

sufficiently diverse word pairs, thus relieving the humans of the need to construct word listsmanually. Obviously, establishing the “gold standard” semantic relatedness for each word pairis still performed manually by human judges.


Phd Thesis - [PDF Document] (126)

Algorithm Correlation withhuman judgements

WordNet-based techniques (Jarmasz, 2003) 0.33–0.35Roget’s Thesaurus-based technique (Jarmasz, 2003) 0.55LSA (Finkelstein et al., 2002a) 0.56WikiRelate! (Strube and Ponzetto, 2006) 0.19–0.48ESA-Wikipedia (March 26, 2006 version) 0.75ESA-Wikipedia (November 11, 2005 version) 0.74ESA-ODP 0.65

Table 6.1: Correlation of word relatedness scores with human judgements on theWordSimilarity-353 collection

Algorithm Correlation withhuman judgements

Bag of words (Lee, Pincombe, and Welsh, 2005) 0.1–0.5LSA (Lee, Pincombe, and Welsh, 2005) 0.60ESA-Wikipedia (March 26, 2006 version) 0.72ESA-Wikipedia (November 11, 2005 version) 0.71ESA-ODP 0.69

Table 6.2: Correlation of text relatedness scores with human judgements on Lee etal.’s document collection

On both test collections, Wikipedia-based semantic interpretation is supe-rior to the ODP-based one. We believe that two factors contribute to this phe-nomenon. First, axes of a multi-dimensional interpretation space should ideallybe as independent as possible. The hierarchical organization of the Open Direc-tory reflects the generalization relation between concepts and obviously violatesthis independence requirement. Second, to increase the amount of training datafor building the ODP-based semantic interpreter, we crawled all the URLs listedin the ODP. This allowed us to increase the amount of textual data by severalorders of magnitude, but also brought about a non-negligible amount of noise,which is common in Web pages. On the other hand, Wikipedia articles are vir-tually noise-free, and mostly qualify as Standard Written English. Thus, thetextual descriptions of Wikipedia concepts are arguably more focused than thoseof the ODP concepts.


Phd Thesis - [PDF Document] (127)

Chapter 7

Related work

This section puts our methodology in the context of related prior work.

7.1 Beyond the Bag of Words

To date, quite a few attempts have been made to deviate from the orthodox bagof words paradigm, usually with limited success. In particular, representationsbased on phrases (Lewis, 1992a; Dumais et al., 1998; Fuernkranz, Mitchell, andRiloff, 2000), named entities (Kumaran and Allan, 2004), and term clustering(Lewis and Croft, 1990; Bekkerman, 2003) have been explored. However, noneof these techniques could possibly overcome the problem underlying the variousexamples we reviewed in this paper—lack of world knowledge.

In mainstream Information Retrieval, query expansion techniques are used toaugment queries with additional terms. However, this approach does not enhancequeries with high-level concepts beyond words or phrases. It occasionally usesWordNet (Fellbaum, 1998) as a source of external knowledge, but queries aremore often enriched with individual words, which are chosen through relevancefeedback (Mitra, Singhal, and Buckley, 1998; Xu and Croft, 2000), by consult-ing dictionaries and thesauri (Voorhees, 1994; Voorhees, 1998), or by analyzingthe context around the query term (Finkelstein et al., 2002a). Ballesteros andCroft (1997) studied query expansion with phrases in the context of cross-lingualinformation retrieval.

7.2 Feature Generation for Text Categorization

Feature generation techniques were found useful in a variety of machine learningtasks (Markovitch and Rosenstein, 2002; Fawcett, 1993; Matheus, 1991). Thesetechniques search for new features that describe the target concept better than


Phd Thesis - [PDF Document] (128)

the ones supplied with the training instances. A number of proposed featuregeneration algorithms (Pagallo and Haussler, 1990; Matheus and Rendell, 1989;Hu and Kibler, 1996; Murphy and Pazzani, 1991) led to significant improvementsin performance over a range of classification tasks. However, even though featuregeneration is an established research area in machine learning, only a few workshave applied it to text processing (Kudenko and Hirsh, 1998; Mikheev, 1999;Cohen, 2000; Scott, 1998). In contrast to our approach, these techniques did notuse any exogenous knowledge.

Kudenko and Hirsh (1998) proposed a domain-independent feature generationalgorithm that uses Boolean features to test whether certain sub-sequences appeara minimum number of times. They applied the algorithm to three toy problemsin topic spotting and book passage categorization.

Mikheev (1999) used a feature collocation lattice as a feature generation enginewithin maximum entropy framework, and applied it to document categorization,sentence boundary detection and part-of-speech tagging. This work utilized in-formation about individual words, bigrams and trigrams to pre-build the featurespace, and then selected a set of feature cliques with the highest log-likelihoodestimate.

Cohen (2000) conducted research on the following problem: given a set oflabeled instances not accompanied by a feature set, is it possible to automaticallydiscover features useful for classification according to the given labels? Problemsof this kind occur, for example, when classifying names of musical artists by musicgenres, or names of computer games by categories such as quest or action. Thepaper proposed to collect relevant Web pages, and then define features based onwords from HTML headers that co-occur with the names to be classified. Thefact that a word appears in an HTML header usually signifies its importance, andhence potential usefulness for classification. The author also identified anothersource of features based on positions inside HTML documents, where position isdefined as a sequence of tags in the HTML parsing tree, between the root of thetree and the name of interest. For example, if a name frequently appears insidetables, this characteristic may be defined as a feature.

Fuhr (1985) introduced the Darmstadt Indexing Approach (DIA), which de-fines features as properties of terms, documents or categories, rather than mereterms or phrases. Thus, meta information such as positions of words withindocuments, document lengths or the cardinality of category training sets mayall be considered as features. Sebastiani (2002) notes that the DIA allows uni-form usage of these new features along with conventional term- or phrase-basedrepresentations.

Bekkerman et al. (2001) represented documents by word clusters rather thanby individual words, within the framework of the information bottleneck approach(Pereira, Tishby, and Lee, 1993; Tishby, Pereira, and Bialek, 1999). The resulting


Phd Thesis - [PDF Document] (129)

clusters were then used as new features that replaced the original words.

7.2.1 Feature Generation Using Electronic Dictionaries

Several studies performed feature construction using the WordNet electronic dic-tionary (Fellbaum, 1998) and other domain-specific dictionaries (Scott, 1998;Urena-Lopez, Buenaga, and Gomez, 2001; Wang et al., 2003; Bloehdorn andHotho, 2004).

Scott (1998) attempted to augment the conventional bag-of-words represen-tation with additional features, using the symbolic classification system Ripper(Cohen, 1995). This study evaluated features based on syntactically1 and statis-tically motivated phrases, as well as on WordNet synsets2. In the latter case, thesystem performed generalizations using the hypernym hierarchy of WordNet, andcompletely replaced a bag of words with a bag of synsets. While using hyper-nyms allowed Ripper to produce more general and more comprehensible rulesand achieved some performance gains on small classification tasks, no perfor-mance benefits could be obtained for larger tasks, which even suffered from somedegradation in classification accuracy. Consistent with other published findings(Lewis, 1992a; Dumais et al., 1998; Fuernkranz, Mitchell, and Riloff, 2000), thephrase-based representation also did not yield any significant performance bene-fits over the bag-of-words approach.3

Urena-Lopez, Buenaga, and Gomez (2001) used WordNet in conjunction withRocchio (Rocchio, 1971) and Widrow-Hoff (Lewis et al., 1996; Widrow andStearns, 1985, Ch. 6) linear classifiers to fine-tune the category vectors. Wang etal. (2003) used Medical Subject Headings (MeSH) (MeSH, 2003) to replace thebag of words with canonical medical terms; Bloehdorn and Hotho (2004) useda similar approach to augment Reuters-21578 documents with WordNet synsetsand OHSUMED medical documents with MeSH terms.

It should be noted, however, that WordNet was not originally designed to be apowerful knowledge base, but rather a lexical database more suitable for peculiarlexicographers’ needs. Specifically, WordNet has the following drawbacks whenused as a knowledge base for text categorization:

• WordNet has a fairly small coverage—for the test collections we used inthis paper, up to 50% of their unique words are missing from WordNet. In

1Identification of syntactic phrases was performed using a noun phrase extractor built ontop of a part of speech tagger (Brill, 1995).

2A synset is WordNet notion for a sense shared by a group of synonymous words.3Sebastiani (2002) casts the use of bag of words versus phrases as utilizing lexical semantics

rather than compositional semantics. Interestingly, some bag-of-words approaches (notably,KNN) may be considered context-sensitive as they do not assume independence between eitherfeatures (terms) or categories (Yang and Pedersen, 1997).


Phd Thesis - [PDF Document] (130)

particular, many proper names, slang and domain-specific technical termsare not included in WordNet, which was designed as a general-purposedictionary.

• Additional information about synsets (beyond their identity) is very limited.This is because WordNet implements a differential rather than constructivelexical semantics theory, so that glosses that accompany the synsets aremainly designed to distinguish the synsets rather than provide a definitionof the sense or concept. Usage examples that occasionally constitute partof the gloss serve the same purpose. Without such auxiliary information,reliable word sense disambiguation is almost impossible.

• WordNet was designed by professional linguists who are trained to recognizeminute differences in word senses. As a result, common words have far toomany distinct senses to be useful in information retrieval (Mihalcea, 2003);for example, the word “make” has as many as 48 senses as a verb alone. Suchfine-grained distinctions between synsets present an additional difficulty forword sense disambiguation.

Both our approach and the techniques that use WordNet manipulate a collec-tion of concepts. However, there are a number of crucial differences. All previousstudies only performed feature generation for individual words only. Our ap-proach can handle arbitrarily long or short text fragments alike. Consideringwords in context allows our approach to perform word sense disambiguation.Approaches using WordNet cannot achieve disambiguation because informationabout synsets is limited to merely a few words, while in both the ODP andWikipedia concepts are associated with huge amounts of text. Even for individ-ual words, our approach provides much more sophisticated mapping of words toconcepts, through the analysis of the large bodies of texts associated with con-cepts. This allows us to represent the meaning of words (or texts) as a weightedcombination of concepts, while mapping a word in WordNet amounts to simplelookup, without any weights. Furthermore, in WordNet the senses of each wordare mutually exclusive. In our approach, concepts reflect different aspects of theinput, thus yielding weighted multi-faceted representation of the text.

In the next section we illustrate the limitations of WordNet on two specificexamples.


Phd Thesis - [PDF Document] (131)

7.2.2 Comparing Knowledge Sources for Feature Genera-tion: ODP versus WordNet

To demonstrate the shortcomings of WordNet as a source for knowledge-basedfeature generation, we juxtapose WordNet-based and ODP-based feature gener-ation for two sample sentences examined in Section 5.3.1 (we repeat the ODPcontext classifications for readers’ convenience). We opted to compare WordNetwith the Open Directory, since both knowledge repositories are hierarchicallyorganized.

We used WordNet version 1.6. In what follows, WordNet synsets are denotedwith curly braces, and noun and verb synsets are followed by their immediatehypernym (more general synset), if applicable.

• Text: “Rumsfeld appeared with Gen. Richard Myers, chairman of the JointChiefs of Staff.”

ODP classifications:

– Society/Issues/Government Operations

– Society/Politics

– Society/Issues/Warfare and Conflict/Specific Conflicts/


– Science/Technology/Military Science

– Society/Issues/Warfare and Conflict/Weapons

– Society/History/By Region/North America/United States/

Presidents/Bush, George Walker

– Society/Politics/Conservatism

WordNet classifications:

– {Rumsfeld} → { }; (word not present in WordNet)

– {look, appear, seem} → {be}; {appear}; {appear, come out} →{happen, materialize}; {appear, seem} → {be}; {appear, comealong}; {appear} → {perform, execute, do}

– {Gen} → {information, info}– {Richard} → { }; (word not present in WordNet)

– {Myers} → { }; (word not present in WordNet)

– {president, chairman, chairwoman, chair, chairperson} → {presidingofficer}; {chair, chairman} → {head, lead}


Phd Thesis - [PDF Document] (132)

– {joint, articulation, articulatio} → {body part}; {joint} → {spot};{articulation, join, joint, juncture, junction} → {connection, connex-ion, link}; {roast, joint} → {cut, cut of meat}; {joint} → {junction,conjunction}; {joint, marijuana cigarette, reefer, stick} → {cigarette,cigaret, coffin nail, butt, fa*g}

– {joint} → {fit, go}; {joint, articulate} → {supply, provide, render,furnish}; {joint} → {fasten, fix, secure}

– {joint (vs. separate)}; {joint}– {head, chief, top dog} → {leader}; {foreman, chief, gaffer, honcho,

boss} → {supervisor}– {staff} → {force, personnel}; {staff} → {stick}; {staff, faculty} →{body}; {staff} → {symbol}; {staff, stave} → {musical notation}

– {staff} → {provide, supply, ply, cater}

• Text: “Herceptin is a so-called targeted therapy because of its ability toattack diseased cells and leave healthy ones alone.”

ODP classifications:

– Health/Conditions and Diseases/Cancer/Breast

– Society/Issues/Health/Conditions and Diseases/Cancer/Al-

ternative Treatments

– Health/Support Groups/Conditions and Diseases/Cancer

WordNet classifications:

– {Herceptin} → { }; (word not present in WordNet)

– {alleged (prenominal), so-called, supposed} → {questionable (vs. un-questionable)}

– {target, aim, place, direct, point} → {aim, take, train, take aim,direct}

– {therapy} → {medical care, medical aid}– {ability} → {quality}– {ability, power} → {cognition, knowledge}


Phd Thesis - [PDF Document] (133)

– {attack, onslaught, onset, onrush} → {operation}; {attack} →{turn, play}; {fire, attack, flak, blast} → {criticism, unfavor-able judgment}; {approach, attack, plan of attack} → { concep-tualization, conceptualisation, formulation, formularizing, formula-rising}; {attack, attempt} → {battery, assault, assault and bat-tery}; {attack, tone-beginning} → {beginning, start, commence-ment}; {attack} → {affliction}; {attack, assault} → {attention, at-tending};

– {attack, assail} → {fight, struggle}; {attack, round, assail, lash out,snipe, assault} → {criticize, criticise, pick apart}; {attack, aggress}→ {act, move}; {assail, assault, set on, attack}; {attack} → {begin,get, start out, start, set about, set out, commence}; {attack} →{affect}

– {assault (prenominal), attack (prenominal)} → {offensive (vs. defen-sive)};

– {diseased, morbid, pathologic, pathological} → {unhealthy (vs.healthy)};

– {cell} → {compartment}; {cell} → {entity, something}; {cell, elec-tric cell} → {electrical device}; {cell, cadre} → {political unit}; {cell,cubicle} → {room}; {cell, jail cell, prison cell} → {room}

– {leave, leave of absence} → {time off}; {leave} → {permission};{farewell, leave, leave-taking, parting} → {departure, going, goingaway, leaving};

– {leave, go forth, go away}; (16 more verb senses omitted for brevity)

– {healthy (vs. unhealthy)}; {healthy} → {sound (vs. unsound)};{healthy, salubrious, good for you (predicate)} → {wholesome (vs.unwholesome)}; {fit (vs. unfit), healthy} → {able, able-bodied};{healthy, intelligent, levelheaded, sound} → {reasonable (vs. unrea-sonable), sensible};

– {one, 1, I, ace, single, unity} → {digit}; {one} → {unit}– {alone (predicate)} → {unsocial (vs. social)}; {alone (predicate),

lone (prenominal), lonely (prenominal), solitary} → {unaccompanied(vs. accompanied)}; {alone (predicate), only} → {exclusive (vs. in-clusive)}; {alone (predicate), unique, unequaled, unequalled, unpar-alleled} → {incomparable (vs. comparable), uncomparable}

– {entirely, exclusively, solely, alone, only}; {alone, unaccompanied}Evidently, WordNet classifications are overly general and diverse because con-

text words cannot be properly disambiguated. Furthermore, owing to lack of


Phd Thesis - [PDF Document] (134)

proper names, WordNet cannot possibly provide the wealth of information en-coded in the Open Directory, which easily overcomes the drawbacks of WordNet.Crawling all the Web sites cataloged in the Open Directory results in exception-ally wide word coverage. Furthermore, the crawled texts provide a plethora ofinformation about each ODP concept.

7.2.3 Using Unlabeled Examples

To the best of our knowledge, with the exception of the above studies that usedWordNet, there have been no attempts to date to automatically use large-scalerepositories of structured background knowledge for feature generation. An in-teresting approach to using non-structured background knowledge was proposedby Zelikovitz and Hirsh (2000). This work uses a collection of unlabeled examplesas intermediaries in comparing testing examples with the training ones. Specif-ically, when an unknown test instance does not appear to resemble any labeledtraining instances, unlabeled examples that are similar to both may be used as“bridges.” Using this approach, it is possible to handle the situation where thetraining and the test document have few or no words in common. The unlabeleddocuments are utilized to define a cosine similarity metric, which is then usedby the KNN algorithm for actual text categorization. This approach, however,suffers from efficiency problems, as looking for intermediaries to compare everytwo documents makes it necessary to explore a combinatorial search space.

In a subsequent paper, Zelikovitz and Hirsh (2001) proposed an alternativeway to use unlabeled documents as background knowledge. In this work, unla-beled texts are pooled together with the training documents to compute a LatentSemantic Analysis (Deerwester et al., 1990) model. LSA analyzes a large cor-pus of unlabeled text, and automatically identifies so-called “latent concepts”using Singular Value Decomposition. The resulting LSA metric then facilitatescomparison of test documents to training documents. The addition of unlabeleddocuments significantly increases the amount of data on which word cooccurrencestatistics is estimated, thus providing a solution to text categorization problemswhere training data is particularly scarce. However, subsequent studies foundthat LSA can rarely improve the strong baseline established by SVM, and ofteneven results in performance degradation (Wu and Gunopulos, 2002; Liu et al.,2004). In contrast to LSA, which manipulates virtual concepts, our methodologyrelies on using concepts identified and described by humans.

In Section 6.2 we reported the results of applying our methodology to theproblem of computing semantic relatedness of words and texts, for which previousstate of the art results have been based on LSA. To this end, we formulatedExplicit Semantic Analysis (ESA), which represents fragments of text in the spaceof knowledge concepts defined in the Open Directory or in Wikipedia. Compared


Phd Thesis - [PDF Document] (135)

with the existing state of the art, using ESA resulted in substantial improvementsin correlation of computed relatedness scores with human judgments. Thesefindings prove that the benefits of using distilled human knowledge are muchgreater than merely using cooccurrence statistics gathered from a collection ofauxiliary unlabeled texts.

7.2.4 Other Related Studies

There have been numerous previous attempts to add knowledge to machine learn-ing techniques. Transfer learning approaches (Bennett, Dumais, and Horvitz,2003; Do and Ng, 2005; Raina, Ng, and Koller, 2006) leverage information fromdifferent but related learning tasks. Pseudo-relevance feedback (Ruthven and Lal-mas, 2003) uses information from the top-ranked documents, which are assumedto be relevant to the query; for example, characteristic terms from such documentsmay be used for query expansion (Xu and Croft, 1996). Recent studies on semi-supervised methods (Goldberg and Zhu, 2006; Ando and Zhang, 2005a; Ando andZhang, 2005b) infer information from unlabeled data, which is often available inmuch larger amounts than labeled data. However, all these approaches amountto using shallow cooccurrence-style knowledge. On the other hand, the methodswe propose in this thesis use much deeper knowledge cataloged by humans, whichcomes in the form of concepts that correspond to the nodes of the Open Directoryor to the articles of Wikipedia.

While our approach relies on existing repositories of classified knowledge, thereis a large body of research on extracting facts through Web mining (Cafarella etal., 2005; Etzioni et al., 2004), so it would be interesting to consider using such ex-tracted facts to drastically increase the amount of available knowledge, especiallywhen measures are taken to ascertain correctness of the extracted information(Downey, Etzioni, and Soderland, 2005).

Our use of local contexts to facilitate fine-grained feature generation is reminis-cent of the intra-document dynamics analysis proposed by Gabrilovich, Dumais,and Horvitz (2004) for characterization of news article types. The latter workmanipulated sliding contextual windows of the same size to make their scores di-rectly comparable. As we showed in Section 5.3.3, the multi-resolution approach,which operates at several levels of linguistic abstraction, is superior to fixed-sizewindows for the case of text categorization. Incidentally, the term “Local ContextAnalysis” is also used in an entirely different branch of Information Retrieval. Xuand Croft (2000) used this term to refer to a particular kind of query expansion,where a query is expanded in the context of top-ranked retrieved documents.

In our methodology, the feature generator is implemented as a text classi-fier that maps local document contexts onto knowledge concepts, which thenserve as additional document features. This approach is similar to the use of


Phd Thesis - [PDF Document] (136)

classifiers as features. Bennett, Dumais, and Horvitz (2005) used the reliability-indicator methodology (Toyama and Horvitz, 2000) to combine several regulartext classifiers (decision tree, SVM, Naive Bayes and unigram) with the aid of ameta-classifier.

Our approach that uses structured background knowledge is somewhat remi-niscent of explanation-based learning (Mitchell, Keller, and Kedar-Cabelli, 1986;Dejong and Mooney, 1986), where generalizations of previously seen examples arereused in future problem solving tasks, thus mimicking humans’ ability to learnfrom a single example.

7.3 Semantic Similarity and Semantic Related-


In this thesis we dealt with “semantic relatedness” rather than “semantic sim-ilarity” or “semantic distance”, which are also often used in the literature. Intheir extensive survey of relatedness measures, Budanitsky and Hirst (2006) ar-gued that the notion of relatedness is more general than that of similarity, as theformer subsumes many different kind of specific relations, including meronymy,antonymy, functional association, and others. They further maintained that com-putational linguistics applications often require measures of relatedness ratherthan the more narrowly defined measures of similarity. For example, word sensedisambiguation can use any related words from the context, and not merely sim-ilar words. Budanitsky and Hirst (2006) also argued that the notion of semanticdistance might be confusing due to the different ways it has been used in theliterature.

Prior work on computing semantic relatedness pursued three main directions:comparing text fragments as bags of words in vector space (Baeza-Yates andRibeiro-Neto, 1999; Rorvig, 1999), using lexical resources, and using Latent Se-mantic Analysis (Deerwester et al., 1990). The former technique is the simplest,but performs sub-optimally when the texts to be compared share few words, forinstance, when the texts use synonyms to convey similar messages. This tech-nique is also trivially inappropriate for comparing individual words. The lattertwo techniques attempt to circumvent this problem.

Lexical databases such as WordNet (Fellbaum, 1998) or Roget’s Thesaurus(Roget, 1852) encode relations between words such as synonymy, hypernymy,and meronymy. Quite a few metrics have been defined that compute related-ness using various properties of the underlying graph structure of these resources(Budanitsky and Hirst, 2006; Jarmasz, 2003; Resnik, 1999; Lin, 1998; Jiang andConrath, 1997). The obvious drawback of this approach is that creation of lexi-cal resources requires lexicographic expertise as well as a lot of time and effort,


Phd Thesis - [PDF Document] (137)

and consequently such resources cover only a small fragment of the language lexi-con. Specifically, such resources contain few proper names, neologisms, slang, anddomain-specific technical terms. Furthermore, these resources have strong lexi-cal orientation in that they predominantly contain information about individualwords, but little world knowledge in general.

On the other hand, LSA (Deerwester et al., 1990) is a purely statistical tech-nique, which leverages word cooccurrence information from a large unlabeledcorpus of text. LSA does not rely on any human-organized knowledge; rather,it “learns” its representation by applying Singular Value Decomposition to thewords-by-documents cooccurrence matrix. LSA is essentially a dimensionality re-duction technique that identifies a number of most prominent dimensions in thedata, which are assumed to correspond to “latent concepts”. Meanings of wordsand documents are then compared in the space defined by these concepts. Latentsemantic models are notoriously difficult to interpret, since the computed con-cepts cannot be readily mapped into natural concepts manipulated by humans.The Explicit Semantic Analysis method we proposed circumvents this problem,as it represents meanings of text fragments using natural concepts defined byhumans.

Our approach to estimating semantic relatedness of words is somewhat rem-iniscent of distributional (or co-occurrence) similarity (Lee, 1999; Dagan, Lee,and Pereira, 1999). Indeed, we compare the meanings of words by comparingthe occurrence patterns across a large collection of natural language documents.However, the compilation of these documents is not arbitrary, rather, the docu-ments are aligned with encyclopedia articles, while each of them is focused on asingle topic. Furthermore, distributional similarity methods are inherently suit-able for comparing individual words, while our method can compute similarity ofarbitrarily long texts.

Prior work in the field mostly focused on semantic similarity of words, us-ing R&G (Rubenstein and Goodenough, 1965) list of 65 word pairs and M&C(Miller and Charles, 1991) list of 30 word pairs. When only the similarity relationis considered, using lexical resources was often successful enough, reaching thecorrelation of 0.70–0.85 with human judgements (Budanitsky and Hirst, 2006;Jarmasz, 2003). In this case, lexical techniques even have a slight edge over ESA-Wikipedia, whose correlation with human scores is 0.723 on M&C and 0.816 onR&G.4 However, when the entire language wealth is considered in an attempt tocapture more general semantic relatedness, lexical techniques yield substantiallyinferior results (see Table 6.1). WordNet-based techniques, which only considerthe generalization (“is-a”) relation between words, achieve correlation of only0.33–0.35 with human judgements (Budanitsky and Hirst, 2006; Jarmasz, 2003).

4WikiRelate! (Strube and Ponzetto, 2006) achieved relatively low scores of 0.31–0.54 onthese domains.


Phd Thesis - [PDF Document] (138)

Jarmasz & Szpakowicz’s ELKB system (Jarmasz, 2003) based on Roget’s The-saurus (Roget, 1852) achieves a higher correlation of 0.55 due to its use of a richerset of relations.

Studying semantic similarity and relatedness of words is related to assessingthe similarity of relations. An example of this task is to establish that word pairscarpenter:wood and mason:stone are relationally similar, as the words in bothpairs stand in the same relation (profession:material). State of the art resultson relational similarity are based on Latent Relational Analysis (Turney, 2006;Turney, 2005).

Sahami and Heilman (2006) proposed to use the Web as a source of additionalknowledge for measuring similarity of short text snippets. To this end, theydefined a kernel function that sends two snippets as queries to a search engine,and compares the bags of words for the two sets of returned documents. A majorlimitation of this technique is that it is only applicable to short texts, becausesending a long text as a query to a search engine is likely to return few or evenno results at all. On the other hand, our approach is applicable to text fragmentsof arbitrary length.

The above-mentioned WordNet-based techniques are inherently limited to in-dividual words, and their adaptation for comparing longer texts requires an extralevel of sophistication (Mihalcea, Corley, and Strapparava, 2006). In contrast,our method treats both words and texts in essentially the same way.

A recent paper by Strube and Ponzetto (2006) also used Wikipedia for com-puting semantic relatedness. However, their method, called WikiRelate!, is radi-cally different from ours. Given a pair of words w1 and w2, WikiRelate! searchesfor Wikipedia articles, p1 and p2, that respectively contain w1 and w2 in theirtitles. Semantic relatedness is then computed based on various distance mea-sures between p1 and p2. These measures either rely on the texts of the pages,or path distances within the category hierarchy of Wikipedia. Our approach, onthe other hand, represents each word as a weighted vector of Wikipedia concepts.Semantic relatedness is then computed by comparing the two concept vectors.

Thus, the differences between the two approaches are:

1. WikiRelate! can only process words that actually occur in titles ofWikipedia articles. ESA only requires that the word appears within thetext of Wikipedia articles.

2. WikiRelate! is limited to single words while ESA can compare texts of anylength.

3. WikiRelate! represents the semantics of a word by either the text of thearticle associated with it, or by the node in the category hierarchy. ESA


Phd Thesis - [PDF Document] (139)

has a much more structured semantic representation consisting of a vectorof Wikipedia concepts.

Indeed, as we have shown in Section 6.2, the richer representation of ESAyields much better results.


Phd Thesis - [PDF Document] (140)


Phd Thesis - [PDF Document] (141)

Chapter 8


In this thesis we proposed a feature generation methodology for textual informa-tion retrieval. In order to render machine learning algorithms with common-senseand domain-specific knowledge of humans, we use very large-scale knowledgerepositories to build a feature generator. These knowledge repositories, whichhave been manually crafted by human editors, provide a fully automatic way totap into the collective knowledge of tens and hundreds of thousands of people.The feature generator analyzes document text and augments the conventionalbag of words representation with relevant concepts from the knowledge reposi-tory. The enriched representation contains information that cannot be deducedfrom the document text alone.

In Section 2.2 we listed several limitations of the bag of words approach, andin the subsequent sections we showed how they are resolved by our methodology.In particular, external knowledge allows us to reason about words that appearin the testing set but not in the training set. We use hierarchically organizedknowledge to make powerful generalizations, making it possible to know thatcertain infrequent words belong to more general classes of words. Externallysupplied knowledge can also help in those cases when some information vital forclassification is omitted from training texts because it is assumed to be sharedby the target readership.

In this work we instantiated our feature generation methodology with two spe-cific knowledge repositories, the Open Directory Project and the Wikipedia en-cyclopedia. We succeeded to make use of an encyclopedia without deep languageunderstanding, specially crafted inference rules or relying on additional common-sense knowledge bases. This was made possible by applying standard text clas-sification techniques to match document texts with relevant Wikipedia articles.The Wikipedia-based results are superior to the ODP-based ones on a numberof datasets, and are comparable to it on others. Moreover, using Wikipedia im-poses fewer restrictions on suitable knowledge repositories, and does not assume


Phd Thesis - [PDF Document] (142)

the availability of an ontology. In our future work, we intend to study possi-ble ways for combining two or more knowledge repositories for improving textcategorization performance even further.

We also described multi-resolution analysis, which examines the documenttext at several levels of linguistic abstraction and performs feature generation ateach level. When polysemous words are considered in their native context, wordsense disambiguation is implicitly performed. Considering local contexts allowsthe feature generator to cope with word synonymy and polysemy. Furthermore,when the document text is processed at several levels of granularity, even brieflymentioned aspects can be identified and used. These might easily have beenoverlooked if the document were processed as one large chunk of text.

Empirical evaluation definitively confirmed the value of knowledge-based fea-ture generation for text categorization across a range of datasets. Recently,the performance of the best text categorization systems became similar, as ifa plateau has been reached, and previous work mostly achieved small improve-ments. Using the ODP and Wikipedia allowed us to reap much greater benefitsand to bring text categorization to a qualitatively new level of performance, withdouble-digit improvements observed on a number of datasets. Given the domain-specific nature of some test collections, we also compared the utility of narrowdomain-specific knowledge with that of larger amounts of information coveringall branches of knowledge (Section 5.3.4). Perhaps surprisingly, we found thateven for narrow-scope test collections, a wide coverage knowledge base yieldedsubstantially greater improvements than its domain-specific subsets. This obser-vation reinforces the breadth hypothesis, formulated by Lenat and Feigenbaum(1990), that “to behave intelligently in unexpected situations, an agent must becapable of falling back on increasingly general knowledge.”

We also applied our feature generation methodology to the problem of au-tomatically assessing semantic relatedness of words and texts. To this end, wepresented a novel technique, called Explicit Semantic Analysis, for represent-ing semantics of natural language texts using natural concepts. In contrast toexisting methods, ESA offers a uniform way for computing relatedness of bothindividual words and arbitrarily long text fragments. Moreover, using naturalconcepts makes the ESA model easy to interpret, as can be seen in the exampleswe provided. Compared with the previous state of the art, using ESA results insubstantial improvements in correlation of computed relatedness scores with hu-man judgements: from r = 0.56 to 0.75 for individual words and from r = 0.60 to0.72 for texts. Consequently, we anticipate ESA to give rise to the next generationof natural language processing tools.

We believe that this research constitutes a step towards the long-standingaspiration of Artificial Intelligence to endow natural language processing withhumans’ knowledge about the world. However, our study only scratches the


Phd Thesis - [PDF Document] (143)

surface of what can be achieved with knowledge-rich features. In our futurework, we plan to investigate new algorithms for mapping document contextsonto knowledge concepts, as well as new techniques for selecting attributes thatare most characteristic of every concept. We intend to apply focused crawling tocollect only relevant Web pages when cataloged URLs are crawled; we also planto apply page segmentation techniques to eliminate noise from crawled pages (Yuet al., 2003). In this work we capitalized on inter-article links of Wikipedia inseveral ways, including the use of anchor text and the number of incoming linksfor each article, as well as creating additional features from linked concepts. Inour future work we intend to investigate more elaborate techniques for leveragingthe high degree of cross-linking between Wikipedia articles.

The Wiki technology underlying the Wikipedia project is often used nowadaysin a variety of open-editing initiatives. These include corporate intranets thatuse Wiki as a primary documentation tool, as well as numerous domain-specificencyclopedias on topics ranging from mathematics to Orthodox Christianity.1

Therefore, we believe our methodology may be used for augmenting documentrepresentation in domains for which no ontologies exist. It is also essential tonote that Wikipedia is available in numerous languages, while different languageversions are cross-linked at the level of concepts. We believe this informationcan be leveraged to use Wikipedia-based semantic interpretation for improvingmachine translation.

In addition to the ODP and Wikipedia, we also plan to make use of additionalknowledge repositories. Among domain-specific knowledge bases, it would be par-ticularly interesting to use the Medical Subject Headings (MeSH) hierarchy toimprove classification of biomedical documents. Recently, several projects havebeen launched that intend to digitize large numbers of books. The largest and ar-guably best known among these projects are Google Print2 and Amazon’s SearchInside the Book3. If the content of numerous books is made available for researchpurposes, it would be extremely interesting to use their text in conjunction withone of the library classification schemes (e.g., UDC) to build a book-based featuregenerator.

Over the recent years, collaborative tagging projects (also known as folk-sonomies) became widespread on the Internet (Marlow et al., 2006). We believeit would be very interesting to use the data accumulated through these taggingefforts as knowledge sources for feature generation. This way, we would use tagsas concepts, and the tagged textual objects (such as blog postings and Web pages)as material for learning the scope of these concepts.

1See encyclopedias for alonger list of examples.



Phd Thesis - [PDF Document] (144)

We conjecture that knowledge-based feature generation will also be usefulfor other information retrieval tasks beyond text categorization, and we intendto investigate this in our future work. Specifically, we intend to apply featuregeneration to information search and word sense disambiguation. In the searchscenario, we are studying ways to augment both queries and documents with gen-erated features. This way, documents will be indexed in the augmented space ofwords and concepts. Current approaches to word sense disambiguation representcontexts that contain ambiguous words using the bag of words augmented withpart-of-speech information. We believe representation of such contexts can begreatly improved if we use feature generation to map these contexts into relevantknowledge concepts. Anecdotal evidence (such as the examples presented in Sec-tions 5.3.1 and 5.4.1) implies our method has much promise for improving thestate of the art in word sense disambiguation.

In its present form, our method can inherently be applied only for improvingrepresentation of textual documents. Indeed, to date we applied our feature gen-eration methodology for improving the performance of text categorization and forcomputing semantic relatedness of texts. However, we believe our approach canalso be applied beyond mere text, as long as the objects to be manipulated are ac-companied with some textual description. As an example, consider a collection ofmedical records containing test results paired with narrative reports. Performingfeature generation from narrative reports is likely to produce pertinent conceptsthat can be used for augmenting the original record. Indeed, prior studies (Hripc-sak et al., 1995) showed that natural language processing techniques can be usedto extract vital information from narrative reports in automated decision-supportsystems.


Phd Thesis - [PDF Document] (145)

Appendix A

Text Categorization with ManyRedundant Features: UsingAggressive Feature Selection toMake SVMs Competitive withC4.5

Text categorization algorithms usually represent documents as bags of words andconsequently have to deal with huge numbers of features. Most previous studiesfound that the majority of these features are relevant for classification, and thatthe performance of text categorization with support vector machines peaks whenno feature selection is performed. We describe a class of text categorizationproblems that are characterized with many redundant features. Even though mostof these features are relevant, the underlying concepts can be concisely capturedusing only a few features, while keeping all of them has substantially detrimentaleffect on categorization accuracy. We develop a novel measure that capturesfeature redundancy, and use it to analyze a large collection of datasets. We showthat for problems plagued with numerous redundant features the performance ofC4.5 is significantly superior to that of SVM, while aggressive feature selectionallows SVM to beat C4.5 by a narrow margin.

A.1 Introduction

Text categorization deals with assigning category labels to natural language doc-uments. Categories come from a fixed set of labels, and each document maybe assigned one or more categories. The absolute majority of works in the field


Phd Thesis - [PDF Document] (146)

employ the so-called “bag of words” approach and use plain language words asfeatures (Sebastiani, 2002). Using a bag of words usually leads to an explosionin the number of features, so that even moderately-sized test collections oftenhave thousands or even tens of thousands of features. In such high-dimensionalspaces, feature selection (FS) is often necessary to reduce noise and avoid overfit-ting. Prior studies found support vector machines (SVM) and K-Nearest Neigh-bor (KNN) to be the best performing algorithms for text categorization (Dumaiset al., 1998; Yang and Liu, 1999).

Joachims (1998) found that support vector machines are very robust even inthe presence of numerous features, and further observed that the multitude oftext features are indeed useful for text categorization. To substantiate this claim,Joachims used a Naive Bayes classifier with feature sets of increasing size, wherefeatures were first ordered by their discriminative capacity (using the informationgain criterion), and then the most informative features were removed. The clas-sifier trained on the remaining “low-utility” features performed markedly betterthan random labeling of documents with categories, thus implying that all fea-tures are relevant and should be used. These findings were later corroborated inmore recent studies (Brank et al., 2002; Rogati and Yang, 2002) that observedeither no improvement or even small degradation of SVM performance after fea-ture selection. On the 20 Newsgroups collection (Lang, 1995), which is one of thestandard text categorization datasets, feature selection significantly degrades theaccuracy of SVM classification (Bekkerman, 2003) due to a very large and diver-sified vocabulary of newsgroup postings. Consequently, many later works usingSVMs did not perform any feature selection at all (Leopold and Kindermann,2002; Lewis et al., 2004).

In this paper we describe a class of text categorization problems that arecharacterized by many redundant features. The corresponding datasets were col-lected in the course of our prior work (Davidov, Gabrilovich, and Markovitch,2004), where we proposed a methodology for parameterized generation of labeleddatasets for text categorization based on the Open Directory Project (ODP).Further details on parameterized generation of labeled datasets can be found inAppendix B. In our present work we use a subset of 100 datasets whose catego-rization difficulty (as measured by baseline SVM accuracy) is evenly distributedfrom very easy to very hard. We observed that even though the datasets differsignificantly in their difficulty, many of them are comprised of categories thatcan be told apart using a small number of words. For example, consider dis-tinguishing the documents about Boulder, Colorado, from those about Dallas,Texas. A few proper names of local landmarks and a handful of words describinglocal industries and other peculiarities often suffice to distinguish texts aboutthe two cities. Given these discriminators, other words add little differentiationpower, and are therefore redundant. As we show in Section A.3, support vector


Phd Thesis - [PDF Document] (147)

machines—which are usually quite robust in the presence of many features—donot fare well when a few good discriminators are vastly outnumbered by featureswith little additional differentiation power.

We further demonstrate that on such datasets C4.5 significantly outperformsSVM and KNN, although the latter are usually considered substantially superiortext classifiers. When no feature selection is performed, C4.5 constructs smalldecision trees that capture the concept much better then either SVM or KNN.Surprisingly, even when feature selection is optimized for each classifier, C4.5formulates a powerful classification model, significantly superior to that of KNNand only marginally less capable than that of SVM. We also show the crucialimportance of aggressive feature selection for this class of problems on a differentdocument representation. In this experiment we extend the conventional bagof words with features constructed using the WordNet electronic dictionary bygeneralizing original words; again, SVM performance steadily increases as fewerfeatures are selected.

To account for this phenomenon, we developed a novel measure that predictsfeature redundancy in datasets. This measure analyzes the distribution of featuresby their information gain, and reliably predicts whether feature selection will bebeneficial or harmful for a given dataset. Notably, computation of this measuredoes not require to actually build a classifier, nor to invoke it on a validation setto determine an optimal feature selection level.

The main contributions of this paper are threefold. First, we describe a class oftext categorization problems that have many redundant features, and for whichaggressive feature selection is essential to achieve decent level of SVM perfor-mance. The existence of such class of problems is in contrast to most of priorresearch in text categorization, which found the majority of features (except therarest ones) to be relevant, and specifically beneficial for SVM classification. Sec-ond, we use two different feature sets to show that without an aggressive featureselection, SVM classification is substantially inferior to that of C4.5, which waspreviously shown to be a less capable text classifier. Finally, we develop a mea-sure that, given a dataset, predicts whether feature selection would be beneficialfor it. This measure performs outlier detection in the distribution of features byinformation gain, without actually classifying the documents.

A.2 Experimental Methodology

We conducted a series of experiments to explore the utility of feature selectionfor datasets plagued with redundant features. In what follows, we first describethe construction of the datasets used in the experiments, and then proceed todeveloping a measure that predicts the utility of feature selection for a given


Phd Thesis - [PDF Document] (148)


A.2.1 Datasets

Acquiring datasets for text categorization based on Web directories has been oftenperformed in prior studies, which used Yahoo! (Mladenic and Grobelnik, 1998b),ODP (Chakrabarti et al., 2002; Cohen et al., 2002) and the Hoover’s Onlinecompany database (Yang, Slattery, and Ghani, 2002). This approach allows toeliminate the huge manual effort required to actually label the documents, byfirst selecting a number of categories (= directory nodes) to define the labels,and then collecting the documents from the subtrees rooted at these categoriesto populate the dataset.

In our prior work (Davidov, Gabrilovich, and Markovitch, 2004) we de-veloped a methodology for automatically acquiring labeled datasets for textcategorization from hierarchical directories of documents, and implemented asystem that performed such acquisition based on the Open Directory Project( In the present paper we use a subset of 100 datasets ac-quired using this methodology. Each dataset consists of a pair of ODP categorieswith an average of 150 documents, and corresponds to a binary classification taskof telling these two categories apart (documents are single-labeled, that is, ev-ery document belongs to exactly one category). When generating datasets fromWeb directories, where each category contains links to actual Internet sites, weconstruct text documents representative of those sites. Following the scheme in-troduced by Yang, Slattery, and Ghani (2002), each link cataloged in the ODP isused to obtain a small representative sample of the target Web site. To this end,we crawl the target site in BFS order, starting from the URL listed in the direc-tory. A predefined number of Web pages (5 in this work) are downloaded, andconcatenated into a synthetic document, which is then filtered to remove HTMLmarkup; the average document size after filtering is 11.2 Kilobytes.

The datasets vary significantly by their difficulty for text categorization, andbaseline SVM accuracy obtained on them is nearly uniformly distributed between0.6 and 0.92. To list a few examples, datasets in our collection range from easyones containing such pairs of ODP categories as Games/Video Games/Shooter

and Recreation/Autos/Makes and Models/Volkswagen, to medium difficultyones with Arts/Music/Bands and Artists vs. Arts/Celebrities, to hardones such as Regional/North America/United States/Virginia/Richmond/

Business and Economy vs. Regional/North America/United States/Florida/

Fort Myers/Business and Economy. The full collection of 100 datasets, along withadditional statistics and all the raw data used in our experiments is available at .


Phd Thesis - [PDF Document] (149)

A.2.2 Predicting the Utility of Feature Selection

In Section A.3 we show that the majority of datasets we used in this study benefitgreatly from aggressive feature selection. We conjectured that these datasets havea small number of features that together allow to learn the underlying conceptconcisely, while the rest of the features do more harm than good. To understandthis phenomenon, we examined the distribution of features in each dataset bytheir information gain.

Figure A.1 shows this distribution for several sample datasets.1 Empirically,we observed that datasets with feature distribution similar to Dataset 46 benefitfrom feature selection immensely (for this particular dataset, aggressive featureselection improved SVM accuracy from 0.60 to 0.93). Such datasets have severalfeatures with high information gain, while the rest of their features have markedlylower IG scores. In contrast to these, datasets similar to Dataset 1 are charac-terized with smooth spectrum of IG values—in such cases feature selection willoften eliminate features that carry essential information; indeed, for this datasetfeature selection caused SVM accuracy to drop from 0.86 to 0.74. For comparison,we show a similarly looking graph for the 20 Newsgroups (20NG) dataset, whichis often used for text categorization experiments and for which feature selectionwas found particularly harmful (Bekkerman, 2003).

Interestingly, high IG values of best-scoring features do not necessarily im-ply that feature selection will substantially improve the accuracy. For instance,Dataset 31 has several features with very high information gain, but its IG graphdeclines gracefully over subsequent features, and does not fall as sharp as forDataset 46. Consequently, feature selection only improves SVM accuracy from0.92 to 0.95—a much more modest gain than for Dataset 46. On the other hand,Dataset 61 has somewhat lower initial IG values, but its IG graph declines verysharply. Feature selection was shown to be of high utility for this dataset as well,boosting the accuracy from 0.64 to 0.84.

The above results imply that the absolute values of information gain are ofless importance than the speed of decline of IG values across features. To quan-tify this phenomenon, we need to assess the number of outliers—features whoseinformation gain is highly above that of all other features. Under this definitionthe desired measure becomes easy to formulate. We first compute the informationgain for all features, and then count the number of features whose informationgain is higher than 3 standard deviations above the average. Although the un-derlying distribution cannot be assumed to be normal, this familiar statisticalcriterion works very reliably in practice. Formally, let D be a dataset and let F

1Dataset ids refer to the full listing table at .


Phd Thesis - [PDF Document] (150)












0 10 20 30 40 50





Features sorted by information gain(only first 50 features are shown)

Dataset 46Dataset 1Dataset 31Dataset 61


Figure A.1: Distribution of features by IG in several datasets

be a set of its features. We define the Outlier Count (OC) as

OC(D,F) = |{f ∈ F : IG(f) > µIG + 3 · σIG }| ,

where µIG and σIG are the average and standard deviation of information gainof the features in F . In Section A.3 we show that Outlier Count reliably predictsthe utility of feature selection for a variety of datasets.

A.2.3 Extended Feature Set Based on WordNet

Several studies in text categorization performed feature construction using theWordNet electronic dictionary (Fellbaum, 1998). In this work we show thataggressive feature selection can significantly improve categorization accuracy fordocument representation extended with constructed features.

Scott and Matwin (1999), and later Wermter and Hung (2002), used WordNetto change document representation from a bag of words to a bag of synsets(WordNet notion of concepts), by using the hypernymy relation to generalizeword senses. Since many words are not found in WordNet (e.g., neologisms,narrow technical terms, and proper names), we opted for extending a bag ofwords with WordNet-based features rather than completely changing documentrepresentation to a bag of synsets. To this end, we first perform feature generationby generalizing document words using WordNet, and then decimate the generatedfeatures through feature selection. In Section A.3.4 we demonstrate that featureselection is as important for generated features as it is for regular features (plainlanguage words).


Phd Thesis - [PDF Document] (151)

A.2.4 Feature Selection Algorithms

A variety of feature selection techniques have been tested for text categorization,while Information Gain, χ2, Document Frequency (Yang and Pedersen, 1997;Rogati and Yang, 2002), Bi-Normal Separation (Forman, 2003) and Odds Ratio(Mladenic, 1998a) were reported to be the most effective. Adopting the prob-abilistic notation from Sebastiani (2002), we use P (tk, ci) to denote the jointprobability that a random document contains term tk and belongs to category ci,and N to denote the number of training documents. The above feature selectiontechniques are then defined as follows:

1. Information Gain (IG):∑c∈{ci,ci}

∑t∈{tk,tk} P (t, c) · log P (t,c)

P (t)P (c)

2. χ2 (CHI): N · P (tk,ci)P (tk,ci)−P (tk,ci)P (tk,ci)

P (tk)P (tk)P (ci)P (ci)

3. Document Frequency (DF): N · P (tk)

4. Bi-Normal Separation (BNS):|F−1(P (tk|ci))−F−1(P (tk|ci))|, where F is the cumulative probability func-tion of the standard Normal distribution

5. Odds Ratio (OR): P (tk|ci)·(1−P (tk|ci))(1−P (tk|ci))·P (tk|ci)

6. Random (RND)

Actual feature selection is performed by selecting the top scoring features, us-ing either a predefined threshold on the feature score or a fixed percentage ofall the features available. In addition to these “principled” selection schemes,we unconditionally remove stop words and words occurring in less than threedocuments.

A.2.5 Classification Algorithms and Measures

We used the datasets described in Section A.2.1 to compare the performance ofSupport Vector Machines (Vapnik, 1995), C4.5 (Quinlan, 1993), and K-NearestNeighbor (Duda and Hart, 1973). In this work we used the SV M light implemen-tation (Joachims, 1999a) with a linear2 kernel.

We used classification accuracy as a measure of text categorization perfor-mance. Studies in text categorization usually work with multi-labeled datasets

2Joachims (1998) observed that most text categorization problems are linearly separable, andconsequently most studies in the field used a linear SVM kernel (Bekkerman, 2003; Forman,2003; Brank et al., 2002).


Phd Thesis - [PDF Document] (152)

in which each category has much fewer positive examples than negative ones.In order to adequately reflect categorization performance in such cases, othermeasures of performance are conventionally used (Sebastiani, 2002), includingprecision, recall, F1, and precision-recall break-even point (BEP). However, forsingle-labeled datasets all these measures can be proved to be equal to accuracy,which is the measure of choice in the machine learning community.

A.3 Empirical Evaluation

In this section we evaluate the role of feature selection for several classificationalgorithms operating on datasets with many redundant features. We conductedthe experiments using a text categorization platform of our own design and de-velopment called Hogwarts. All accuracy values reported below were obtainedusing 4-fold cross-validation scheme.

When working with support vector machines, it is essential to perform ade-quate parameter tuning. In the case of a linear kernel (and under the assumptionof equal cost of errors on positive and negative examples), the only relevantparameter is C, namely, the trade-off between training error and margin. Tooptimize this parameter, we set aside one fold of the training data as a validationset, and for each feature selection level selected the best C value from among{10−4, 10−3, 10−2, 10−1, 1, 101, 102, 103, 104}.

A.3.1 Validation of Hogwarts Performance

In this section we demonstrate that the results of classifying existing datasetswith Hogwarts are consistent with those in other published studies. Figure A.2shows the results of using SVM in conjunction with IG feature selection to classifythree datasets frequently used in text categorization studies: 10 largest categoriesof Reuters-21578 (Reuters, 1997), 20 Newsgroups (Lang, 1995), and Movie Re-views (Pang, Lee, and Vaithyanathan, 2002).3 Using all features, Hogwartsachieved BEP of 0.922 on Reuters, 0.854 on 20 Newsgroups and 0.818 on MovieReviews. These results are very similar to the performance obtained by otherresearchers (all using SVM). Dumais et al. (1998) achieved BEP of 0.92 for the10 largest Reuters categories. Bekkerman (2003) obtained BEP of 0.856 on the20 Newsgroups dataset. Pang, Lee, and Vaithyanathan (2002) obtained accuracyof 0.829 on the Movie Reviews dataset.

As can be seen in Figure A.2, any level of feature selection harms the per-formance on all of these datasets. The graphs for χ2 and BNS feature selection

3Since the former two of these datasets are multi-labeled, we use precision-recall break-evenpoint (BEP) as a measure of classification performance rather than accuracy (see Section A.2.5).


Phd Thesis - [PDF Document] (153)








0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1





l BE


Feature selection level

Reuters-2157820 NG

Movie Reviews

Figure A.2: Hogwarts performance on existing datasets (feature selection withIG)










ent i

n ac



Datasets ordered by Outlier Count


Figure A.3: Improvement in SVM accuracy at different FS levels vs. using 100%features

algorithms exhibit behavior very similar to IG, so we do not show them hereowing to lack of space. Note that all the experiments reported in the rest of thepaper use the 100 datasets we acquired as explained in Section A.2.1.


Phd Thesis - [PDF Document] (154)

A.3.2 Predicting the Utility of Feature Selection withOutlier Count

We now show that the Outlier Count measure defined in Section A.2.2 reliablypredicts the utility of feature selection. Figure A.3 shows the improvement inSVM accuracy at several feature selection levels versus the baseline accuracyobtained using 100% of features. As we can see, Outlier Count strongly corre-lates with the magnitude of improvement that can be obtained through featureselection. We observe that at lower values of Outlier Count aggressive feature se-lection is highly beneficial. Conversely, at higher OC values much more moderate(if any) feature selection should be performed, while aggressive selection causesdegradation in accuracy. The next section examines the correlation of OutlierCount with the differences in performance between individual classifiers.

The Outlier Count for the datasets we used is nearly uniformly distributed be-tween 6 and 62, with a single outlier value (no pun intended!) of 112 for Dataset 1(Figure A.1), for which feature selection caused SVM accuracy to drop from 0.86to 0.74. For other datasets frequently used for text categorization, Outlier Countfor Reuters-21578 is 78, Movie Reviews—154, and 20 Newsgroups—391, whichexplains why feature selection does for them more harm than good.

Based on these findings, we conclude that using Outlier Count for orderingdatasets reflects the degree to which a dataset can be concisely described by onlya few features, while the rest of the features are predominantly redundant andhave detrimental effect on classification results.

A.3.3 Comparison of Classifiers

Figure A.4 compares the performance of SVM, KNN and C4.5 on the 100 datasetsordered by Outlier Count. When no feature selection is employed, the perfor-mance of C4.5 mostly dominates that of SVM and KNN, and only declines in therightmost part of the graph, which contains datasets where a few features are notsufficient for learning the concept.

Table A.1 shows classifier accuracy without feature selection and with theoptimal feature selection level for each classifier. We used paired t-test to assessthe significance of differences in classifier accuracy over the 100 datasets (seeTable A.2). Without any feature selection, the differences between classifiers werefound to be very significant at p < 5 · 10−3 or lower. For individual classifiers,the improvement in accuracy due to feature selection was extremely significantat p < 10−13.


Phd Thesis - [PDF Document] (155)











Datasets ordered by Outlier Count

SVM (100%)C4.5 (100%)KNN (100%)

Figure A.4: Comparison of performance of SVM, C4.5 and KNN with 100% features

Table A.1: Classifier accuracy at different FS levels

Accuracy with Accuracy with theClassifier 100% features optimal FS level

SVM 0.769 0.853 (using 0.5% features)C4.5 0.800 0.843 (using 0.5% features)KNN 0.741 0.827 (using 2% features)

Table A.2: Statistical significance of differences in classifier accuracy (p values)

Classifier C4.5 KNN SVM C4.5 KNN(FS level) (100%) (100%) (0.5%) (0.5%) (2%)

SVM (100%) 5 · 10−3 4 · 10−9 4 · 10−15 2 · 10−10 6 · 10−11

C4.5 (100%) 2 · 10−5 6 · 10−14 2 · 10−15 3 · 10−4

KNN (100%) 2 · 10−16 6 · 10−13 6 · 10−14

SVM (0.5%) 9 · 10−3 4 · 10−8

C4.5 (0.5%) 5 · 10−3


Phd Thesis - [PDF Document] (156)

A.3.4 The Effect of Using Different Feature Sets

Figure A.5 compares the performance of classifiers at different feature selectionlevels (using Information Gain). As we can see, C4.5 performs better than SVMexcept for the most aggressive FS levels, where their accuracy becomes nearlyequal. Interestingly, C4.5 stays high above KNN at most FS levels.

Figure A.6 presents a similar graph for the extended feature set based onWordNet. Here we use all features of the conventional bag of words, and onlyapply feature selection to the constructed features. C4.5 clearly manages themultitude of redundant features much better than both SVM and KNN. It isalso noteworthy that the accuracy of SVM and KNN increases steadily as featureselection becomes more aggressive, while the improvement in their performancewith 0.5% features vs. 100% features is strongly significant at p < 10−18.

When using the optimal FS level (0.5% for both regular words and WordNetconcepts), the addition of WordNet features is only responsible for a minor im-provement in SVM accuracy from 0.853 to 0.854.

A.3.5 The Effect of Using Different FS Algorithms

Figures A.7 and A.8 show the effect of using different feature selection algorithms(see Section A.2.4) with SVM and C4.5. Consistently with prior studies (Forman,2003; Rogati and Yang, 2002), we observe that IG, CHI and BNS are the bestperformers, while the difference between them is not statistically significant.4 Incontrast with prior studies, we observe that on the family of datasets we described,the best performance of SVM is obtained when only using a tiny fraction offeatures (0.5% for the three best FS techniques).

A.3.6 Testing the Relevancy of Features

In previous sections we showed that text categorization can greatly benefit fromaggressive feature selection. We now address the question whether the featuresdiscarded by selection are at all relevant for classification. Following Joachims(1998), we sorted all features by their information gain, and then removed pro-gressively larger fractions (0.1%, 0.5%, 1%, . . . , 10%, 20%, . . . , 100%) of themost informative features. As can be seen in Figure A.9, the performance ofan SVM classifier trained on the remaining features is noticeably better thanrandom up to very high levels of such harmful “selection”. These results corrob-orate earlier findings by Joachims (1998), and support our hypothesis that the

4The graph for KNN looks substantially similar and also confirms the superiority of IG, CHIand BNS (with negligible differences), so we omit it owing to lack of space.


Phd Thesis - [PDF Document] (157)






0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1




Feature selection level


Figure A.5: Classification using a bag of words






0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1




Feature selection level


Figure A.6: Classification using an extended feature set

features removed through selection are redundant, even though most of them maybe considered relevant.

A.4 Discussion

Studies in text categorization usually represent documents as a bag of words, andconsequently have to manage feature spaces of very high dimensionality. Mostprevious works in the field found that these numerous features are relevant forclassification, and that in particular the performance of SVM text categorization


Phd Thesis - [PDF Document] (158)












0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1




Feature selection level




Figure A.7: SVM accuracy vs. FS level









0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1




Feature selection level




Figure A.8: C4.5 accuracy vs. FS level

peaks when no feature selection is performed.

We described a class of datasets plagued with redundant features, such thattheir elimination significantly boosts categorization accuracy of a host of classi-fiers. Specifically, we showed that when no feature selection is employed on suchdatasets, SVMs are significantly outperformed by C4.5. To explain this phe-nomenon, we analyzed the distribution of features by their information gain, andobserved that this effect occurs when a small number of features are sufficient forconcisely learning the underlying concept. We defined a measure named OutlierCount that, for a given dataset and fixed representation scheme, estimates the


Phd Thesis - [PDF Document] (159)








0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1





Fraction of worst features retained

Figure A.9: Removing the best features by IG

amount of feature redundancy through outlier analysis.

In a series of experiments, we demonstrated that Outlier Count reliably pre-dicts the amount of improvement that can be gained through feature selection.These findings are backed by empirical evidence both for the conventional bagof words, and for a representation extended through feature generation basedon WordNet. We further performed a controlled ablation study to verify thatthe redundant features are in fact relevant for classification. To this end, weremoved progressively larger fractions of most informative features, and foundthe remaining ones to suffice for better than random performance. Finally, weanalyzed several established benchmarks for text categorization with respect toOutlier Count, and explained why they do not benefit from feature selection.

Following the established practice in text categorization, throughout this pa-per we used an SVM classifier with a linear kernel. In an ancillary experimentwe sought to determine whether a non-linear SVM kernel may fare better thana linear one when dealing with numerous redundant features. Without featureselection, switching from a linear kernel to an RBF one reduced the accuracyfrom 0.769 to 0.687. Even at the optimal feature selection level, the accuracyachieved with an RBF kernel was slightly below that of a linear one (0.849 vs.0.853), contradicting our anticipation of better performance by a more sophis-ticated kernel. However, this experiment should be considered preliminary, andin our future work we plan to conduct a thorough investigation of the ability ofnon-linear SVM kernels to withstand high rates of redundant features.

In a recent study, Forman (2003) proposed a novel feature selection algorithm


Phd Thesis - [PDF Document] (160)

named Bi-Normal Separation, which improved the performance of SVM text cat-egorization on a range of datasets. Peak performance was obtained when using500–1000 features (approximately 10% of all available features on the average).More aggressive feature selection led to sharp degradation of the results—usingless than 100 features caused macro-F1 to decrease by 5%–10% depending on theselection algorithm used.

Our work corroborates the findings that feature selection can help text catego-rization with SVMs, and describes a class of problems where the improvement dueto feature selection is particularly large. We showed that for this class of problemsthe improvement in accuracy can be twice as high as found by Forman (2003)(namely, 8.4% vs. 4.2%), while optimal performance is achieved when using muchfewer features (between 5 and 40, depending on the dataset). We also evaluatedseveral feature selection algorithms on text categorization problems characterizedwith many redundant features. Our results support earlier findings that Informa-tion Gain, Bi-Normal Separation and χ2 are the most powerful feature selectionalgorithms, while the differences between them are not significant.

It should be noted that for all the datasets we used, the utility of featureselection could be established by setting aside part of the training data to serveas a validation set. Indeed, the high redundancy level was so pronounced, thatthe optimal selection level for the testing data could almost always be correctlydetermined on the validation fold. However, we believe that the introduction ofOutlier Count and the use of ablation experiments that systematically eliminatemost informative features, allow deeper understanding of the issues of featureredundancy and relevancy.


Phd Thesis - [PDF Document] (161)

Appendix B

Parameterized Generation ofLabeled Datasets for TextCategorization Based on aHierarchical Directory

Although text categorization is a burgeoning area of IR research, readily availabletest collections in this field are surprisingly scarce. We describe a methodologyand system (named Accio) for automatically acquiring labeled datasets for textcategorization from the World Wide Web, by capitalizing on the body of knowl-edge encoded in the structure of existing hierarchical directories such as the OpenDirectory. We define parameters of categories that make it possible to acquirenumerous datasets with desired properties, which in turn allow better control overcategorization experiments. In particular, we develop metrics that estimate thedifficulty of a dataset by examining the host directory structure. These metricsare shown to be good predictors of categorization accuracy that can be achievedon a dataset, and serve as efficient heuristics for generating datasets subject touser’s requirements. A large collection of automatically generated datasets aremade available for other researchers to use.

B.1 Introduction

While numerous works studied text categorization (TC) in the past, good testcollections are by far less abundant. This scarcity is mainly due to the hugemanual effort required to collect a sufficiently large body of text, categorize it,and ultimately produce it in machine-readable format. Most studies use theReuters-21578 collection (Reuters, 1997) as the primary benchmark. Others use


Phd Thesis - [PDF Document] (162)

20 Newsgroups (Lang, 1995) and OHSUMED (Hersh et al., 1994), while TRECfiltering experiments often use the data from the TIPSTER corpus (Harman,1992).

Even though the Reuters-21578 dataset became a standard reference in thefield, it has a number of significant shortcomings. According to Dumais and Chen(2000), “the Reuters collection is small and very well organized compared withmany realistic applications”. Scott (1998) also noted that the Reuters corpus hasa very restricted vocabulary, since Reuters in-house style prescribes using uni-form unambiguous terminology to facilitate quick comprehension. As observedby Joachims (1998), large Reuters categories can be reliably classified by virtuallyany reasonable classifier. We believe that TC performance on more representativereal-life corpora still has way to go. The recently introduced new Reuters corpus(Lewis et al., 2004), which features a large number of documents and three or-thogonal category sets, definitely constitutes a substantial challenge. At the sametime, acquisition of additional corpora for TC research remains a major issue.

In the past, developing a new dataset for text categorization required extensivemanual effort to actually label the documents. However, given today proliferationof the Web, it seems reasonable to acquire large-scale real-life datasets from theInternet, subject to a set of constraints. Web directories that catalog Internet sitesrepresent readily available results of enormous labeling projects. We thereforepropose to capitalize on this body of information in order to derive new datasetsin a fully automatic manner. This way, the directory serves as a source of URLs,while its hierarchical organization is used to label the documents collected fromthese URLs with corresponding directory categories. Since many Web directoriescontinue to grow through ongoing development, we can expect the raw materialfor dataset generation to become even more abundant as the time passes.

In what follows, we propose a methodology for automatic acquisition of up-to-date datasets with desired properties. The automatic aspect of acquisitionfacilitates creation of numerous test collections, effectively eliminating a consid-erable amount of human labor normally associated with preparing a dataset. Atthe same time, datasets that possess predefined characteristics allow researchersto exercise better control over TC experiments and to collect data geared towardstheir specific experimentation needs. Choosing these properties in different waysallows one to create focused datasets for improving TC performance in certainareas or under certain constraints, as well as to collect comprehensive datasetsfor exhaustive evaluation of TC systems.

After the data has been collected, the hierarchical structure of the directorymay be used by classification algorithms as background world knowledge—theassociation between the data and the corresponding portion of the hierarchy isdefined by virtue of dataset construction. The resulting datasets can be used forregular text categorization, hypertext categorization, as well as hierarchical text


Phd Thesis - [PDF Document] (163)

classification. Moreover, many Web directories cross-link related categories withso-called “symbolic links”, which allow one to construct datasets for multi-labeledTC experiments.

We developed a software system named Accio 1 that lets the user specifydesired dataset parameters, and then efficiently locates suitable categories andcollects documents associated with them. It should be observed that Web docu-ments are far less fluent and clean compared to articles published in the “brickand mortar” world. To ensure the coherence of the data, Accio represents eachWeb site with several pages gathered from it through crawling, and filters thepages gathered both during and after the crawling. The final processing stepcomputes a number of performance metrics for the generated dataset.

In this paper we describe generation of datasets based on the Open Direc-tory Project (ODP,, although the techniques we propose arereadily applicable to other Web directories, as well as to non-Web hierarchies ofdocuments (see Section B.2). A number of previous studies in hypertext and hi-erarchical text classification used document sets collected from Yahoo! (Mladenicand Grobelnik, 1998b; Labrou and Finin, 1999), ODP (Chakrabarti et al., 2002;Cohen et al., 2002; Meng et al., 2002) and the Hoover’s Online company database(Ghani et al., 2000; Yang, Slattery, and Ghani, 2002). To the best of our knowl-edge, all these studies performed standard acquisition of Web documents pointedat from the explicitly specified directory nodes; specifically, no properties of cat-egories were considered or defined, and no attempt to predict the classificationperformance was made. Interestingly, a recent study in word sense disambigua-tion (Santamaria, Gonzalo, and Verdejo, 2003) used ODP to automatically ac-quire labeled datasets for disambiguation tasks. In this work, a collection of ODPcategories were first automatically mapped to WordNet (Fellbaum, 1998) senses,and then the descriptions of links classified under these categories were collectedto serve as sentences with sense-labeled words. In contrast to our approach, thismapping only considered category paths, while we also analyze the full text ofcategory and link descriptions (see Section B.2).

The main contributions of this paper are threefold. First, we present amethodology for automatically acquiring labeled data sets for text categoriza-tion experiments, which allows parameterized generation of datasets with desiredproperties. Second, we establish a connection between similarity metrics for doc-ument sets and the classification accuracy achieved on these sets. The similaritymetrics we developed are shown to be good predictors of classification accuracy,and can therefore be used as efficient heuristics for locating datasets of desireddegree of hardness. We also propose to use classification accuracy as a new simi-larity metric that reflects how separable two document sets are. Finally, we make

1Accio (Latin - to call to, summon)—incantation for the Summoning Charm, which causesan object called for to fly to the caster (Rowling, 2001).


Phd Thesis - [PDF Document] (164)

publicly available a large collection of text categorization datasets that we col-lected and evaluated in the course of this work, along with a variety of metricscomputed for them. Using the same datasets allows different research groupsto conduct repeatable experiments and to compare their results directly. Thisrepository, which is similar in purpose to the UCI Repository of machine learningdatabases (Blake and Merz, 1998), is available for research use and is publiclyaccessible at All datasets are available inplain text form and in the form of preprocessed feature vectors; the latter distri-bution can be used by researchers in machine learning who are less interested inthe specifics of text processing. Furthermore, for each dataset we provide baselineperformance numbers using SVM, KNN, and C4.5. We also plan to release thesoftware system for automatic generation of datasets. Other researchers will beable to use Accio to acquire new datasets subject to their specific requirements.

B.2 Parameterization of Dataset Generation

Throughout this paper we discuss generation of datasets that contain two cat-egories and are single-labeled, that is, every document belongs to exactly onecategory. In Section B.5 we consider possible relaxations to this rule.

We assume the availability of a hierarchical directory of documents that sat-isfies the following requirements:

1. The directory is organized as a tree where each node is labeled with acategory.

2. There is a collection of documents associated with each category (directorynode).

3. Categories are provided with text descriptions. Documents associated withthe categories may optionally be accompanied by short annotations.

Suitable directories come in a variety of forms. Some are major Web directo-ries that catalog actual Web sites, such as Yahoo! or the Open Directory. TheMedical Subject Headings (MeSH) hierarchy (MeSH, 2003) maintained by theU.S. National Library of Medicine is cross-linked with the MEDLINE database,and therefore can be used for automatic generation of labeled datasets of medi-cal texts. Library classification schemes such as UDC and Dewey are hierarchicalcatalogs of books that can also be used for automatical acquisition of text catego-rization datasets; samples of books can be used if shorter documents are required.The open content Wikipedia encyclopedia2 collaboratively developed by Internet

2 .


Phd Thesis - [PDF Document] (165)

users offers tantalizing opportunities for harnessing high quality datasets. As ofthis writing, Wikipedia contains over 170,000 articles in English and 150,000 inother languages, thus allowing acquisition of datasets on similar topics in a vari-ety of languages. Yet another option is to use the new Reuters collection (Lewiset al., 2004) that contains over 800,000 documents labeled with categories com-ing from three distinct hierarchies. In this project we generate datasets based onthe Open Directory Project, which is arguably the largest publicly available Webdirectory.3

We employ two kinds of parameters that define the nature of generateddatasets: those characterizing the dataset as a whole (i.e., describe pairs of cate-gories), and those characterizing individual categories that comprise the datasets.Varying these parameters allows one to create classification tasks with differentproperties.

B.2.1 Metrics

Metrics quantify conceptual distance between a pair of categories. Intuitively,the larger the distance, the easier it is to induce a classifier for separating thecategories. From the machine learning perspective, the difficulty of a datasetfor existing categorization algorithms is an important parameter. The ability tocreate datasets with varying degree of difficulty would be instrumental in the questfor better learning algorithms. In other words, we would like to retain controlover the degree of separability of the two categories comprising the dataset. Inthis section we first define an exact but computationally expensive measure ofdataset hardness, and then propose two metrics that are highly correlated withit but are much more efficient to compute.

Achievable Categorization Accuracy as a Measure of Dataset Hardness

A straightforward way to assess how difficult a given dataset is for currently avail-able learning algorithms is simply to run these algorithms on it. It is apparentlyappealing to use the accuracy of a single best classification algorithm as an ulti-mate measure, especially in the light of the fact that a number of studies showedsupport vector machines to be the best performing text classifier (Joachims, 1998;Dumais et al., 1998). However, as we show in Section B.4.4, SVM does not nec-essarily produce the best results for every dataset. Several researchers observed

3Although the actual size of Yahoo! has not been publicly released in the re-cent years, it is estimated to be about half the size of the Open Directory (see and formore details).


Phd Thesis - [PDF Document] (166)

similar phenomena, and used various learning approaches to decide which clas-sifier should be used for a given category (Lam and Lai, 2001) or for a givendocument (Bennett, Dumais, and Horvitz, 2002).

We believe that such sophisticated classifier combination schemes might bean overkill for establishing a measure of category separability. We suggest usingsome function of the accuracy values achieved by a number of classifiers as the“gold standard” of hardness. While there are many ways to define a suitablecombination scheme, we propose to use the maximum accuracy among a setof classifiers, as we believe it reflects how difficult the dataset is for the bestavailable algorithm (obviously, without an oracle predicting which classifier touse, this value cannot always be attained in practice). Formally, we define

distclass max(c1, c2) = maxalg∈C

Accuracyalg(c1, c2) ,

where c1, c2 are a pair of categories comprising a dataset and C is a set of classifi-cation algorithms. In the sequel we refer to this metric as Maximum AchievableAccuracy (MAA). In the experiments reported in Section B.4 we compute MAAusing classifiers based on support vector machines, decision trees and the K-Nearest Neighbor algorithm.

Nothing seems simpler than defining the hardness of a dataset by actual clas-sification accuracy. The only problem with this approach is that it is grosslyinefficient. When we search for datasets in a certain difficulty range, using MAAas part of “generate-and-test” strategy is too computationally intensive to bepractical. Computing MAA requires to actually crawl the Web to download thedocuments, clean the data and organize it as a dataset, and finally subject it toa number of classifiers. If MAA turns out to be too low or too high comparedwith the requirements, we have to test another pair of categories, then anotherone, and so on.

We developed two metrics that estimate the difficulty of a dataset by onlyexamining the hierarchical structure of the host directory, without analyzing thetext of actual documents. In Section B.4 we show that these metrics are stronglycorrelated with MAA and the accuracies of individual classifiers, and this canserve good predictors of how difficult it is to build a classifier that tells twocategories apart.

Historically, the idea of partitioning categories by similarity of meaning (aswell as by importance or frequency) was first mentioned by Lewis (1991), whenhe suggested to group categorization results over different kinds of categories.

In order to develop metrics for computing similarity of categories drawn froma hierarchical directory, let us review a similar setting of assessing similarity ofwords using a hierarchical dictionary or taxonomy. The metrics we define assignlower values to more similar categories, therefore, in what follows we use the termdistance metric (rather than similarity metric) to emphasize this fact.


Phd Thesis - [PDF Document] (167)

Edge-counting Graph Metric

The edge-counting metric (called graph metric below) measures the distance be-tween a pair of categories by the length of the shortest4 path connecting them inthe hierarchy. We conjecture that the closer two categories are in the underlyinggraph, the closer they are in meaning, and hence the smaller the distance betweenthem is. Formally, we define

distgraph(c1, c2) = #edges in the tree path from c1 to c2 .

Rada and Bicknell (1989) also used hierarchy path length as a measure ofconceptual distance. However, this study focused on estimating the similarity ofindividual terms rather than entire sets of documents.

WordNet-based Textual Metric

The above metric only uses the graph structure underlying the hierarchy as asole source of information. We now propose a more elaborate metric (called textmetric in the sequel) that compares textual descriptions of the categories thatare assumed to be provided with the hierarchy.

Our text metric builds upon the similarity metric for individual words sug-gested by Resnik (1999), which uses the WordNet electronic dictionary (Fellbaum,1998) as a source of additional background knowledge. Given two words w1 andw2 whose similarity needs to be established, let us denote by S1 the set of allWordNet nodes (called synsets) that contain w1 and by S2—the set of all synsetsthat contain w2. Resnik defined the similarity between two words as

simResnik(w1, w2) = maxsj[− log p(sj)] , (B.1)

where {sj} is a set of synsets that subsume at least one synset from S1

and one synset from S2 (i.e., the set of all concepts that subsume bothgiven words), p(sj) is the probability of synset sj computed as a func-tion of the frequencies of words that belong to it measured on a ref-erence corpus, and − log p(sj) is the information content of this synset.No word sense disambiguation is performed, and all senses of a polysemous wordare considered equally probable.

We generalize this metric to make it applicable to entire category descrip-tions rather than individual words. In the preprocessing phase we represent eachcategory by pooling together (i) the title and description of the category itselfand all of its descendants (sub-categories), and (ii) the titles and descriptions

4Using the shortest path is important when the hierarchy is actually a graph rather than atree (for example, when symbolic links of the Open Directory are considered).


Phd Thesis - [PDF Document] (168)

(annotations) of the links to actual documents classified under this category orone of its sub-categories. We denote the union of all these textual descriptionsfor category ci as Di. Each pooled description Di is represented as an unorderedbag of words.

The (asymmetric) distance between a pair of such descriptions is canonicallydefined as an average distance from the words of the first description to those ofthe second one:

dist(D1, D2) =1



dist(w,D2) ,

where the distance between a word and a bag of words is defined as the shortestdistance between this word and the bag (i.e., the distance to the nearest word inthe bag):

dist(w, D) = minw′∈Ddist(w,w′) . (B.2)

The distance between two words is defined using Resnik’s similarity met-ric, except the score it returns is subtracted from the maximum possible score(simMAX) to transform the similarity metric into a measure of distance:

dist(w, w′) = simMAX − simResnik(w,w′) .

To estimate the word frequencies needed for the computation of p(sj) in (B.1),we used a training corpus composed of the descriptions of all ODP categories;this step effectively tunes the metric to a specific text collection at hand.

Finally, the metric that operates on entire textual descriptions of categoriesis symmetrically defined as

disttext(c1, c2) = dist(D1, D2) + dist(D2, D1) .

Computing disttext requires some preprocessing computation to build categorydescriptions Di, and then use the frequency of words found in these descriptionsto train a language model that underlies the computation of − log p(sj). Observethat even without the preprocessing phase performed offline, computing the textmetric is a computationally intensive process, as it considers every pair of words inthe two category descriptions, and for each such pair maximizes the informationcontent of the subsuming synsets.

Budanitsky and Hirst (2001) provide a good survey of other word similaritymetrics based on WordNet.

B.2.2 Properties of Individual Categories

The following parameters can be configured for individual categories:


Phd Thesis - [PDF Document] (169)

1. The cardinality of a category specifies the desired number of documents itshould contain. In general, the more examples (documents) are available,the easier the learning task is due to a better representation of the category.

2. Recall that the documents we collect actually represent Web sites theywere downloaded from. Exploring Web sites to different depths affects thequality of this representation. However, taking too many documents fromeach site is not necessarily good, as moving further away from the site’sroot page may lead to barely related pages. The parameter that controlsthis fine balance is called coherence, and is expressed as a number of pagesdownloaded from each Web site and concatenated into a single document.

3. Limiting the selection of categories to a certain part of the hierarchy effec-tively allows to restrict the contents of the documents to a particular topic.For example, generating datasets from the Open Directory Top/Health

subtree may be useful for testing operational TC systems for the medicaldomain. The language of documents may be restricted in a similar way.

B.3 Methodology for Automatic Dataset Gen-


In this section we outline the methodology for automatic generation of datasets.

B.3.1 Acquisition of the Raw Data

Generating a new dataset starts with locating a pair of categories subject to user’sspecification, which consists of a set of desired parameters (or characteristics) ofthe dataset to build (see Section B.2). Finding a pair of categories at specifiedgraph distance is easy, as it involves pursuing a corresponding number of edgesin the graph underlying the hierarchy. On the other hand, identifying pairs ofcategories at a specified text distance is far from trivial. Although the experimentspresented in Section B.4.3 do show high correlation between the two metrics, ingeneral counting the number of edges can only give a rough estimation of the textdistance between two categories.

Since the text metric is much more computationally intensive than the graphone, we cache its values for all pairs of categories considered so far. Given thedesired text distance, we first consult the cache to see if a suitable pair of cate-gories was already found. If this simple test fails, we randomly sample the cacheand identify a pair in the sample whose distance is closest to the required one.We then perform a hill-climbing search in the hierarchy graph starting from thatpair. This search is limited in the number of steps, and if no appropriate pair


Phd Thesis - [PDF Document] (170)

Algorithm LocateCategoryPair TextDist(d)if (∃(p, q) ∈ Cache s.t. disttext(p, q) = d)

then return (p, q)found ← falsewhile (¬found)

Draw a random sample S ⊂ CacheLet (p, q) ∈ S s.t. ∀(p′, q′) ∈ S :|d− disttext(p, q)| ≤ |d− disttext(p

′, q′)|Starting from (p, q), perform n-step hill climbing

until a pair (pd, qd) is found s.t. disttext(p, q) = d

Figure B.1: Locating categories at requested text distance

is found after the limit is exhausted, we randomly sample the cache again, andrepeat the entire process until a suitable pair of categories is found. Figure B.1outlines the pseudocode of the search algorithm.

It is essential to emphasize that the above algorithm only analyzes the hierar-chy structure and category descriptions, but never examines the contents of actualdocuments. It is this feature of our methodology that makes finding datasets ofconfigurable difficulty much more computationally tractable than if MAA was tobe used (Section B.2.1). In our future work we plan to develop more sophisti-cated algorithms for efficiently locating pairs of categories at specified conceptualdistance (see Section B.5).

After locating an appropriate pair of categories, we collect the documents as-sociated with them. Importantly, if a certain category c has several sub-categoriesunder it in the given hierarchy (c1 . . . cn), we collect the documents from the unionof all these categories. The hierarchy structure allows us to view c1 . . . cn as par-ticular cases of c, and thus we can find many more relevant documents than iflooking into category c alone.

When generating datasets from Web directories such as the ODP, where eachcategory contains links to actual Internet sites, we need to construct text docu-ments representative of those sites. Following the scheme introduced in (Yang,Slattery, and Ghani, 2002), each link cataloged in the ODP is used to obtain asmall representative sample of the target Web site. To this end, we crawl thetarget site in the BFS order, starting from the URL listed in the directory. Apredefined number of Web pages are downloaded, and then concatenated into asynthetic document. We refer to these individual pages as sub-documents, sincetheir concatenation yields one document for the categorization task. We usu-ally refer to synthetic documents created by pooling sub-documents simply asdocuments to be consistent with TC terminology; alternatively, we call them


Phd Thesis - [PDF Document] (171)

meta-documents to avoid ambiguity when necessary.Finally, HTML documents are converted into plain text and organized as a

dataset, which we render in a simple XML-like format. It should be noted thatconverting HTML to text is not always perfect, since some small auxiliary textsnippets (as found in menus and the like) may survive this procedure; we viewsuch remnants as a (low) residual noise inherent in automated data acquisition.

B.3.2 Filtering the Raw Data to Cope with Noise

Data collected from the Web can be quite noisy. Common examples of this noiseare textual advertisem*nts, numerous unrelated images, and text rendered inbackground color aimed at duping search engines. To reduce the amount of noisein generated datasets we employ filtering mechanisms before, during, and afterdownloading the data.

Pre-processing filtering eliminates certain categories from consideration. Forexample, we unconditionally disregard the entire Top/World subtree of the OpenDirectory that catalogs Web sites in languages other than English. Similarly, theTop/Adult subtree may be pruned to eliminate inappropriate adult content.

Recall that for every directory link we download a number of pages whoseconcatenation represents the corresponding Web site. Online filtering performedduring the download restricts the crawler to the site linked from the directory,and does not allow it to pursue external links to other sites.

Post-processing filtering analyzes all the downloaded documents as a group,and selects the ones to be concatenated into the final meta-document. In practice,we download more sub-documents than requested by the user, and then decimatethem. We developed two post-processing filters:

1. Weak filtering discards Web pages that contain HTTP error messages, oronly have a few words.

2. Strong filtering attempts to eliminate unrelated pages that do not ad-equately represent the site they were collected from (e.g., legal no-tices or discussion forum rules). To eliminate such pages, we try toidentify obvious outliers. We use the root page of a Web site (i.e.,the page linked from the directory) as a “model” deemed to be rep-resentative of the site as a whole. Whenever the root page con-tains enough text for comparison, we use the text metric developed inSection B.2.1 to compute the distance between it and every other pagedownloaded from the site. We then discard all pages located “further”from the root than one standard deviation above the average.

Comparing weak and strong filtering, we found the latter to improve TCaccuracy by about 0.5%–1.5%.


Phd Thesis - [PDF Document] (172)

B.4 Empirical Evaluation

In this section we show that the datasets generated using the proposed method-ology are sufficiently versatile and allow adequate degree of control over TC ex-periments.

B.4.1 Data Acquisition

We used the methodology outlined in Section B.3 to automatically generate acollection of datasets based on the Open Directory Project ( Open Directory is a public directory that catalogs selected Internet sites.At the time of this writing, ODP covers over 4 million sites organized in morethan 540,000 categories. The project constitutes an ongoing effort promoted bynon-professional users around the globe; currently, ODP advertises a staff of over60,500 editors. Being the result of pro bono work, the Open Directory has itsshare of drawbacks, such as non-uniform coverage, duplicate subtrees in differentbranches of the hierarchy, and sometimes biased coverage due to peculiar viewsof the editors in charge. At the same time, however, ODP embeds a considerableamount of human knowledge.

Based on the Open Directory, we generated 300 datasets of varying difficulty,by using the metrics defined in Section B.2.1 to find categories located at differentgraph or text distances. Each dataset consists of a pair of categories with 100–200 documents per category, while each document was created by concatenating5 sub-documents.

B.4.2 Text Categorization Infrastructure

The following learning algorithms were used to induce actual text classifiers: sup-port vector machines (Vapnik, 1995) (using SV M light implementation (Joachims,1999a)), decision trees (C4.5 (Quinlan, 1993)), and K-Nearest Neighbor (Dudaand Hart, 1973). The motivation behind this choice of algorithms is that theybelong to very different families, and thus allow comprehensive evaluation of thedatasets generated.

We used classification accuracy as a measure of text categorization perfor-mance. Studies in text categorization usually work with multi-labeled datasetsin which each category has much fewer positive examples than negative ones. Inorder to adequately reflect categorization performance in such cases, other mea-sures of performance are conventionally used, including precision, recall, F1, andprecision-recall break-even point (Sebastiani, 2002). However, for single-labeleddatasets all these measures can be proved to be equal to accuracy, which is the


Phd Thesis - [PDF Document] (173)

measure of choice in the machine learning community. All accuracy values re-ported in this paper were obtained under the 10-fold cross-validation scheme.

We conducted the experiments using a text categorization platform of ourown design and development called Hogwarts5. We opted to build a compre-hensive new infrastructure for text categorization, as surprisingly few softwaretools are publicly available for researchers, while those available only allow lim-ited control over their operation. Hogwarts performs text preprocessing, fea-ture extraction, construction, selection and valuation, followed by cross-validatedclassification. Hogwarts interfaces with SVM, KNN and C4.5, and computesall standard measures of categorization performance. At a later stage we plan tomake Hogwarts publicly available for research use.

B.4.3 Correlation Between Distance Metrics and TextCategorization Accuracy

Recall that our primary aim is to generate datasets with predefined properties.Specifically, one of the most important properties we introduced in Section B.2is the ability to exercise control over the difficulty of separation of two categoriescomprising a dataset. The experiments reported below were designed to verifywhether the metrics we developed in Section B.2.1 can serve as reliable predictorsof category separability. We first juxtapose metric predictions with the accuracyof an SVM classifier, and then compare them with the Maximum AchievableAccuracy (MAA).

Figure B.2 shows the correlation between the graph metric and SVM cate-gorization accuracy, while Figure B.3 shows a similar plot for the text metric.Both figures demonstrate that the metrics have strong prediction power for SVMaccuracy. The value of Pearson’s linear correlation coefficient (Press et al., 1997)that we computed to quantify this dependence is 0.533 for the graph metric and0.834 for the text one. Interestingly, the two metrics are fairly strongly correlatedbetween themselves, as implied by their correlation of 0.614 (see Figure B.4).

As follows from the experimental results, there is a trade-off between thecomputational efficiency and the prediction power of the two metrics. The graphmetric is much faster to compute, but only offers a rough estimation of the degreeof separability of a pair of categories. The text metric is much less efficient tocompute, but offers by far more reliable distance assessment.

5Hogwarts school of witchcraft and wizardry is the educational institution attended by HarryPotter (Rowling, 2001).


Phd Thesis - [PDF Document] (174)










0 2 4 6 8 10 12 14 16 18 20




n ac


cy (



Graph distance

Figure B.2: SVM accuracy vs. graph distance










0 3 6 9 12 15 18 21 24 27 30




n ac


cy (



Text distance

Figure B.3: SVM accuracy vs. text distance

B.4.4 Correlation Between Distance Metrics and MAA

In Section B.2.1 we defined the difficulty of a dataset as a function of performanceof a number of classifiers. Instead of using the accuracy produced by any singleclassifier, we proposed to use the maximum value among several classifiers thatwere shown to be good performers in previous studies.

Let us first provide empirical support for the choice of MAA as a rea-sonable measure of conceptual distance between a pair of categories. The


Phd Thesis - [PDF Document] (175)






0 2 4 6 8 10 12 14 16 18 20


t dis



Graph distance

Figure B.4: Text distance vs. graph distance

average accuracy achieved by SVM on the datasets tested is 0.896, KNN—0.874, and C4.5—0.878. These results are consistent with previously pub-lished studies (Sebastiani, 2002), and show that the generated datasets ex-hibit similar performance properties to the manually collected ones used inprior research. However, a closer look at classifier performance on individualdatasets reveals that SVM—although a superior technique in the majority ofcases—does not always yield the best accuracy compared to other classifiers.Specifically, SVM was outperformed by KNN on 58 datasets (19%) and by C4.5on 80 datasets (27%). Furthermore, C4.5 outperformed KNN on 119 datasets(40%), even though decision trees are usually deemed an inferior approach totext categorization compared to SVM and KNN. Therefore, the performance ofthe best currently available algorithm for a particular dataset constitutes a morereliable measure of its true difficulty.

The experiments we conducted prove that the correlation of the graph andtext metrics to MAA is consistently high. Specifically, the correlation betweendistgraph and MAA is 0.550, and between disttext and MAA—0.790. Figures B.5and B.6 depict these correlations with standard error bars. Based on thesefindings, we conclude that the metrics we developed are good predictors of datasetdifficulty.

B.4.5 Versatility of Dataset Generation

We now show that the proposed methodology can be used to automatically gen-erate a continuum of non-trivial categorization tasks of varying difficulty. Having


Phd Thesis - [PDF Document] (176)










0 2 4 6 8 10 12 14 16 18 20M

ax a



Graph distance

Figure B.5: MAA vs. graph distance










0 3 6 9 12 15 18 21 24 27 30





Text distance

Figure B.6: MAA vs. text distance

established in the previous section that the distance metrics are good predic-tors of categorization accuracy, we demonstrate that it is possible to find enoughcategory pairs of adequate size at different conceptual distances.

To prove this, we examine two graphs with pertinent ODP statistics. Fig-ure B.7 depicts the number of category pairs that reside at various distances asmeasured by the graph metric. Since the text metric is much more computa-tionally expensive, showing in full the similar distribution of text distances isnot feasible. For machine learning tasks, we are usually interested in categorieswith a sufficient number of examples to make (statistical) learning meaningfuland allow adequate generalization. Figure B.8 shows a sampled distribution oftext distances among mid-size category pairs having 100–3000 links. ODP hasapproximately 13,000 categories in this size range (and therefore 13,0002/2 pairs);Figure B.8 was built by randomly sampling 3,500 pairs of such categories.

These graphs suggest that the Open Directory is large and versatile enoughto produce numerous datasets with desired properties.


Phd Thesis - [PDF Document] (177)








0 5 10 15 20



of c




Graph distance

Figure B.7: Distribution of graph distances in ODP








0 5 10 15 20 25 30 35



of c




Text distance

Figure B.8: Distribution of text distances in ODP (sample)

B.5 Conclusions and Future Work

Text categorization is an active area of research in information retrieval, yetgood test collections are scarce. We presented a methodology and system namedAccio for automatically acquiring labeled datasets for text categorization fromhierarchical directories of documents. We applied this methodology to gener-ate 300 datasets from the largest Web directory to date—the Open DirectoryProject—as an example. The datasets thus generated can be used in a varietyof learning tasks, including regular text categorization, hypertext categorization,and hierarchical text classification.


Phd Thesis - [PDF Document] (178)

To allow acquisition of new datasets with predefined characteristics, we de-fined a set of properties that characterize datasets as a whole, as well as individualcategories that comprise them. We first introduced Maximum Achievable Accu-racy (MAA) as an intrinsic measure of dataset difficulty, and then developed twokinds of distance metrics that predict the categorization difficulty of a datasetwithout actually examining the full text of the documents. These metrics analyzethe location of categories in the hierarchy tree, as well as textual descriptions ofcategories and annotations of documents. We empirically showed that the text-based metric possesses high predictive power for estimating the separability ofa pair of categories. The edge-counting graph metric is somewhat less reliable,but is much more efficient computationally. We also observed that MAA can beused as a measure of similarity between sets of documents, quantifying the easeof separating them with a text classifier. Since texts acquired from the WWWare often plagued with noise and are generally quite different in nature from for-mal written English found in printed publications, we reported specific steps weundertook to filter the data and monitor its quality during acquisition.

Finally, we established a new repository of text categorization datasets, whichcurrently contains several hundred datasets at various levels of difficulty thatwe generated using the proposed methodology. This collection is available at, along with ancillary statistics and mea-sured classifier performance. The collection continues to grow, and its growthrate is only limited by bandwidth and storage resources. Having a wide varietyof datasets in a centralized repository will allow researchers to perform a widerange of repeatable experiments. The Accio system that performs parameter-ized dataset acquisition from the ODP will be released at a later stage. Usinga subset of these datasets, we developed a novel criterion that assesses featureredundancy and predicts the utility of feature selection for TC (Gabrilovich andMarkovitch, 2004).

This research can be extended in several directions. We plan to investigatemore sophisticated distance metrics that overcome the drawbacks of the basicmetrics we described herein. The graph metric does not account for the factthat two nodes whose common ancestor is close to the hierarchy root are muchless related, than two nodes at the same edge distance whose common ancestorresides deep in the tree. The graph metric may also produce unreliable valuesfor extremely long hierarchy paths, which contain too many intermediate gener-alizations. The WordNet-based text metric is obviously undefined for words notfound in WordNet (e.g., neologisms, narrow technical terms, and proper names);currently, if such a word is present in both documents, we take the value in equa-tion (B.2) to be zero, otherwise, we ignore this word. The text metric may also beinaccurate for documents with only a few words. Following standard IR practice,we also tested the conventional cosine metric to compare bag-of-word vectors of


Phd Thesis - [PDF Document] (179)

categories and documents, but empirically found it to be inadequate. Most ofthe values of the cosine measure clustered near its extremes (0 and 1), while themid-range was very sparsely populated; we attribute this phenomenon to the lackof any background knowledge about word semantics (as, for example, providedby WordNet in the text metric).

We intend to investigate additional parameters of categories that will allowto exercise better control over the properties of generated datasets. Of particularinterest and practical importance are filtering techniques for cleaning the datadownloaded from the Web, and we plan to study this issue in greater depth usingfocused crawling techniques. We also plan to develop more elaborate algorithmsthat locate pairs of categories subject to user’s requirements.

We further intend to construct larger datasets consisting of more than twocategories; to do so, category similarity metrics will need to be generalized ap-propriately to consider mutual distances in a group of categories. We also intendto generate datasets from additional document directories that contain high qual-ity noise-free articles.


Phd Thesis - [PDF Document] (180)


Phd Thesis - [PDF Document] (181)


Adafre, Sisay Fissaha and Maarten de Rijke. 2005. Discovering missing linksin wikipedia. In Proceedings of the Workshop on Link Discovery: Issues,Approaches and Applications (LinkKDD-2005).

Ando, Rie Kubota and Tong Zhang. 2005a. Framework for learning predictivestructures from multiple tasks and unlabeled data. JMLR, pages 1817–1853.

Ando, Rie Kubota and Tong Zhang. 2005b. A high-performance semi-supervisedlearning method for text chunking. In Proceedings of the 43rd Annual Meetingof the ACL, pages 1–9, Ann Arbor, MI, June.

Baeza-Yates, Ricardo and Berthier Ribeiro-Neto. 1999. Modern InformationRetrieval. Addison Wesley, New York, NY.

Baker, L. Douglas and Andrew K. McCallum. 1998. Distributional clustering ofwords for text classification. In W. Bruce Croft, Alistair Moffat, Cornelis J.Van Rijsbergen, Ross Wilkinson, and Justin Zobel, editors, Proceedings ofSIGIR-98, 21st ACM International Conference on Research and Developmentin Information Retrieval, pages 96–103, Melbourne, AU. ACM Press, NewYork, US.

Ballesteros, Lisa and W. Bruce Croft. 1997. Phrasal translation and queryexpansion techniques for cross-language information retrieval. In Proceedingsof the 20th ACM International Conference on Research and Development inInformation Retrieval, pages 84–91.

Basili, Roberto, Alessandro Moschitti, and Maria T. Pazienza. 2000. Language-sensitive text classification. In Proceedings of RIAO-00, 6th InternationalConference “Recherche d’Information Assistee par Ordinateur”, pages 331–343, Paris, France.

Bekkerman, Ron. 2003. Distributional clustering of words for text categorization.Master’s thesis, Technion.

Bekkerman, Ron, Ran El-Yaniv, Naftali Tishby, and Yoad Winter. 2001. Onfeature distributional clustering for text categorization. In Proceedings of the24th International Conference on Research and Development in InformationRetrieval, pages 146–153, September.

Bekkerman, Ron, Ran El-Yaniv, Naftali Tishby, and Yoad Winter. 2003. Distri-butional word clusters vs. words for text categorization. Journal of MachineLearning Research, 3:1183–1208, March.


Phd Thesis - [PDF Document] (182)

Bennett, Paul N., Susan T. Dumais, and Eric Horvitz. 2002. Probabilistic com-bination of text classifiers using reliability indicators: Models and results.In Proceedings of the 25rd ACM International Conference on Research andDevelopment in Information Retrieval, pages 207–215.

Bennett, Paul N., Susan T. Dumais, and Eric Horvitz. 2003. Inductive transferfor text classification using generalized reliability indicators. In Proceedingsof the ICML-2003 Workshop on The Continuum from Labeled to UnlabeledData.

Bennett, Paul N., Susan T. Dumais, and Eric Horvitz. 2005. The combination oftext classifiers using reliability indicators. Information Retrieval, 8(1):67–100.

Blake, C.L. and C.J. Merz. 1998. UCI Repository of machine learning databases.̃ mlearn/MLRepository.html.

Bloehdorn, Stephan and Andreas Hotho. 2004. Boosting for text classificationwith semantic features. In Proceedings of the MSW 2004 Workshop at the10th ACM SIGKDD Conference on Knowledge Discovery and Data Mining,pages 70–87.

Brank, Janez, Marko Grobelnik, Natasa Milic-Frayling, and Dunia Mladenic.2002. Interaction of feature selection methods and linear classification models.In Workshop on Text Learning held at ICML-2002.

Brill, Eric. 1995. Transformation-based error-driven learning and natural lan-guage processing: A case study in part of speech tagging. ComputationalLinguistics, 21(4):543–565.

Buchanan, B. G. and E. A. Feigenbaum. 1982. Forward. In R. Davis and D. B.Lenat, editors, Knowledge-Based Systems in Artificial Intelligence. McGraw-Hill.

Budanitsky, Alexander and Graeme Hirst. 2001. Semantic distance in WordNet:An experimental, application-oriented evaluation of five measures. In Proc.of the NAACL 2001 Workshop on WordNet and Other Lexical Resources.

Budanitsky, Alexander and Graeme Hirst. 2006. Evaluating wordnet-based mea-sures of lexical semantic relatedness. Computational Linguistics, 32(1):13–47.

Cafarella, Michael, Doug Downey, Stephen Soderland, and Oren Etzioni. 2005.Knowitnow: Fast, scalable information extraction from the web. In Proceed-ings of the Conference on Empirical Methods in Natural Language Processing,Vancouver, Canada, October.


Phd Thesis - [PDF Document] (183)

Cai, Lijuan and Thomas Hofmann. 2003. Text categorization by boosting auto-matically extracted concepts. In Proceedings of the 26th International Confer-ence on Research and Development in Information Retrieval, pages 182–189.

Callan, James. 1993. Knowledge-Based Feature Generation for Inductive Learn-ing. Ph.D. thesis, Department of Computer and Information Science, Univer-sity of Massachusetts. Also available as CMPSCI Technical Report 93-07.

Caropreso, Maria Fernanda, Stan Matwin, and Fabrizio Sebastiani. 2001. Alearner-independent evaluation of the usefulness of statistical phrases for au-tomated text categorization. In Amita G. Chin, editor, Text Databases andDocument Management: Theory and Practice. Idea Group Publishing, Her-shey, US, pages 78–102.

Chakrabarti, Soumen, Byron Dom, Rakesh Agrawal, and Prabhakar Raghavan.1997. Using taxonomy, discriminants, and signatures for navigating in textdatabases. In Proceedings of the 23rd VLDB Conference, pages 446–455.

Chakrabarti, Soumen, Mukul M. Joshi, Kunal Punera, and David M. Pennock.2002. The structure of broad topics on the web. In Proc. of the Int’l WorldWide Web Conference.

Chan, Lois Mai. 1999. A Guide to the Library of Congress Classification. Li-braries Unlimited, 5th edition.

Chen, Stanley and Joshua Goodman. 1996. A high-performance semi-supervisedlearning method for text chunking. In Proceedings of the 34th Annual Meetingof the ACL.

Cohen, Doron, Michael Herscovici, Yael Petruschka, Yoelle S. Maarek, Aya Soffer,and Dave Newbold. 2002. Personalized pocket directories for mobile devices.In Proc. of the Int’l World Wide Web Conference.

Cohen, William W. 1995. Fast effective rule induction. In Proceedings of the 12thInternational Conference on Machine Learning (ICML-95), pages 115–123.

Cohen, William W. 2000. Automatically extracting features for concept learningfrom the web. In Proceedings of the 17th International Conference on MachineLearning.

Cristianini, Nello, John Shawe-Taylor, and Huma Lodhi. 2002. Latent semantickernels. Journal of Intelligent Information Systems, 18(2/3):127–152.

Dagan, Ido, Lillian Lee, and Fernando C. N. Pereira. 1999. Similarity-basedmodels of word cooccurrence probabilities. Machine Learning, 34(1–3):43–69.


Phd Thesis - [PDF Document] (184)

Davidov, Dmitry, Evgeniy Gabrilovich, and Shaul Markovitch. 2004. Parame-terized generation of labeled datasets for text categorization based on a hier-archical directory. In Proceedings of the 27th ACM International Conferenceon Research and Development in Information Retrieval, pages 250–257.

Debole, Franca and Fabrizio Sebastiani. 2003. Supervised term weighting forautomated text categorization. In SAC’03, pages 784–788.

Deerwester, S., S. Dumais, G. Furnas, T. Landauer, and R. Harshman. 1990.Indexing by latent semantic analysis. JASIS, 41(6):391–407.

Dejong, Gerald and Raymond Mooney. 1986. Explanation-based learning: Analternative view. Machine Learning, 1(2):145–176.

Demsar, Janez. 2006. Statistical comparison of classifiers over multiple data sets.JMLR, 7:1–30.

Dewdney, Nigel, Carol VanEss-Dykema, and Richard MacMillan. 2001. The formis the substance: Classification of genres in text. In Workshop on HLT andKM held at ACL-2001.

Dewey, Melvil, Joan S. Mitchell, Julianne Beall, Giles Martin, Winton E.Matthews, and Gregory R. New, editors. 2003. Dewey Decimal Classifica-tion and Relative Index. Online Computer Library Center (OCLC), 22ndedition.

Dhillon, Inderjit, Subramanyam Mallela, and Rahul Kumar. 2003. A divisiveinformation-theoretic feature clustering algorithm for text classification. Jour-nal of Machine Learning Research, 3:1265–1287, March.

Do, Chuong and Andrew Ng. 2005. Transfer learning for text classification. InProceedings of Neural Information Processing Systems (NIPS).

Downey, Doug, Oren Etzioni, and Stephen Soderland. 2005. A probabilisticmodel of redundancy in information extraction. In Proceedings of the 19thInternational Joint Conference on Artificial Intelligence, Edinburgh, Scotand,August.

Duda, R.O. and P.E. Hart. 1973. Pattern Classification and Scene Analysis.John Wiley and Sons.

Dumais, Susan and Hao Chen. 2000. Hierarchical classification of web content.In Proceedings of the 23rd ACM International Conference on Research andDevelopment in Information Retrieval, pages 256–263.


Phd Thesis - [PDF Document] (185)

Dumais, Susan, John Platt, David Heckerman, and Mehran Sahami. 1998. In-ductive learning algorithms and representations for text categorization. InCIKM’98, pages 148–155.

Etzioni, Oren, Michael Cafarella, Doug Downey, Stanley Kok, Ana-MariaPopescu, Tal Shaked, Stephen Soderland, Daniel Weld, and AlexanderYates. 2004. Webscale information extraction in knowitall (preliminary re-sults). In Proceedings of the 13th International World Wide Web Conference(WWW’04), New York, USA, May. ACM Press.

Fawcett, Tom. 1991. Feature discovery for inductive concept learning. Tech-nical Report COINS Technical Report 91-8, Department of Computer andInformation Science, University of Massachusetts, Amherst, Massachusetts.

Fawcett, Tom. 1993. Feature Discovery for Problem Solving Systems. Ph.D.thesis, UMass, May.

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

Finkelstein, Lev, Evgeniy Gabrilovich, Yossi Matias, Ehud Rivlin, Zach Solan,Gadi Wolfman, and Eytan Ruppin. 2002a. Placing search in context: Theconcept revisited. ACM TOIS, 20(1):116–131, January.

Finkelstein, Lev, Evgeniy Gabrilovich, Yossi Matias, Ehud Rivlin, Zach Solan,Gadi Wolfman, and Eytan Ruppin. 2002b. WordSimilarity-353 testcollection.̃ gabr/resources/data/wordsim353/


Forman, George. 2003. An extensive empirical study of feature selection metricsfor text classification. Journal of Machine Learning Research, 3:1289–1305,March.

Fuernkranz, Johannes, Tom Mitchell, and Ellen Riloff. 2000. A case study inusing linguistic phrases for text categorization on the WWW. In MehranSahami, editor, Learning for Text Categorization: Proceedings of the 1998AAAI/ICML Workshop. AAAI Press, Madison, Wisconsin, pages 5–12.

Fuhr, Norbert. 1985. A probabilistic model of dictionary-based automatic in-dexing. In Proceedings of RIAO-85, 1st International Conference “Recherched’Information Assistee par Ordinateur”, pages 207–216, Grenoble, France.

Gabrilovich, Evgeniy, Susan Dumais, and Eric Horvitz. 2004. Newsjunkie:Providing personalized newsfeeds via analysis of information novelty. In


Phd Thesis - [PDF Document] (186)

Proceedings of the Thirteenth International World Wide Web Conference(WWW2004), pages 482–490, New York, NY, May. ACM Press.

Gabrilovich, Evgeniy and Shaul Markovitch. 2004. Text categorization withmany redundant features: Using aggressive feature selection to make SVMscompetitive with C4.5. In Proceedings of the 21st International Conferenceon Machine Learning, pages 321–328.

Gabrilovich, Evgeniy and Shaul Markovitch. 2005. Feature generation for textcategorization using world knowledge. In IJCAI’05, pages 1048–1053.

Gabrilovich, Evgeniy and Shaul Markovitch. 2006a. Computing semantic re-latedness of words and texts in wikipedia-derived semantic space. Technicalreport CIS-2006-04, Computer Science Department, Technion—Israel Insti-tute of Technology, Haifa, Israel, July.

Gabrilovich, Evgeniy and Shaul Markovitch. 2006b. Overcoming the brittlenessbottleneck using Wikipedia: Enhancing text categorization with encyclopedicknowledge. In AAAI’06, pages 1301–1306, July.

Gabrilovich, Evgeniy and Shaul Markovitch. 2007. Computing semantic related-ness using wikipedia-based explicit semantic analysis. In Proceedings of the20th International Joint Conference on Artificial Intelligence, January.

Galavotti, Luigi, Fabrizio Sebastiani, and Maria Simi. 2000. Experiments on theuse of feature selection and negative evidence in automated text categoriza-tion. In Jose Borbinha and Thomas Baker, editors, Proceedings of ECDL-00,4th European Conference on Research and Advanced Technology for DigitalLibraries, pages 59–68, Lisbon, Portugal. Published in the “Lecture Notes forComputer Science” series, number 1923, Springer Verlag.

Ghani, Rayid, Rosie Jones, Dunja Mladenic, Kamal Nigam, and Sean Slattery.2000. Data mining on symbolic knowledge extracted from the web. In Work-shop on Text Mining at the 6th ACM International Conference on KnowledgeDiscovery and Data Mining (SIGKDD).

Giles, Jim. 2005. Internet encyclopaedias go head to head. Nature, 438:900–901.

Goldberg, Andrew and Xiaojin Zhu. 2006. Seeing stars when there aren’t manystars: Graph-based semi-supervised learning for sentiment categorization. InWorkshop on Textgraphs: Graph-based Algorithms for Natural Language Pro-cessing, HLT-NAACL 2006.

Han, Eui-Hong (Sam) and George Karypis. 2000. Centroid-based documentclassification: Analysis and experimental results. In PKDD’00, September.


Phd Thesis - [PDF Document] (187)

Harman, Donna. 1992. The DARPA TIPSTER project. In SIGIR Forum,volume 26(2). ACM, pages 26–28.

Hayes, Philip J., Peggy M. Andersen, Irene B. Nirenburg, and Linda M.Schmandt. 1990. TCS: a shell for content-based text categorization. InProceedings of CAIA-90, 6th IEEE Conference on Artificial Intelligence Ap-plications, pages 320–326, Santa Barbara, US. IEEE Computer Society Press,Los Alamitos, US.

Hersh, William, Chris Buckley, T.J. Leone, and David Hickam. 1994.OHSUMED: An interactive retrieval evaluation and new large test collectionfor research. In SIGIR’94, pages 192–201.

Hripcsak, George, Carol Friedman, Philip O. Alderson, William DuMouchel,Stephen B. Johnson, and Paul D. Clayton. 1995. Unlocking clinical datafrom narrative reports: a study of natural language processing. Annals ofInternal Medicine, 122(9):681–688.

Hu, Yuh-Jyh and Dennis Kibler. 1996. A wrapper approach for constructiveinduction. In AAAI-96.

Hull, David A. 1994. Improving text retrieval for the routing problem usinglatent semantic indexing. In W. Bruce Croft and Cornelis J. Van Rijsbergen,editors, Proceedings of SIGIR-94, 17th ACM International Conference on Re-search and Development in Information Retrieval, pages 282–289, Dublin, IE.Springer Verlag, Heidelberg, DE.

Jain, Anil K., Robert P.W. Duin, and Jianchang Mao. 2000. Statistical patternrecognition: A review. IEEE Transactions on Pattern Analysis and MachineIntelligence, 22(1), January.

Jarmasz, Mario. 2003. Roget’s thesaurus as a lexical resource for natural lan-guage processing. Master’s thesis, University of Ottawa.

Jiang, Jay J. and David W. Conrath. 1997. Semantic similarity based on corpusstatistics and lexical taxonomy. In ROCLING’97.

Joachims, Thorsten. 1998. Text categorization with support vector machines:Learning with many relevant features. In ECML’98, pages 137–142.

Joachims, Thorsten. 1999a. Making large-scale SVM learning practical. InB. Schoelkopf, C. Burges, and A. Smola, editors, Advances in Kernel Methods– Support Vector Learning. The MIT Press.


Phd Thesis - [PDF Document] (188)

Joachims, Thorsten. 1999b. Transductive inference for text classification usingsupport vector machines. In Proceedings of the 13th International Conferenceon Machine Learning.

John, George H., Ron Kohavi, and Karl Pfleger. 1994. Irrelevant features and thesubset selection problem. In Proceedings of the 11th International Conferenceon Machine Learning, pages 121–129.

Koller, Daphne and Mehran Sahami. 1997. Hierarchically classifying documentsusing very few words. In Proceedings of the 14th International Conference onMachine Learning, pages 170–178.

Kudenko, Daniel and Haym Hirsh. 1998. Feature generation for sequence catego-rization. In Proceedings of the 15th Conference of the American Associationfor Artificial Intelligence, pages 733–738.

Kumaran, Giridhar and James Allan. 2004. Text classification and named enti-ties for new event detection. In Proceedings of the 27th ACM InternationalConference on Research and Development in Information Retrieval, pages297–304.

Labrou, Yannis and Tim Finin. 1999. Yahoo! as an ontology—using Yahoo!categories to describe documents. In Proc. of the 8th Int’l Conference onInformation and Knowledge Management, pages 180–187.

Lam, Wai and Kwok-Yin Lai. 2001. A meta-learning approach for text catego-rization. In Proceedings of the 24th International Conference on Research andDevelopment in Information Retrieval, pages 303–309.

Lang, Ken. 1995. Newsweeder: Learning to filter netnews. In ICML’95, pages331–339.

Lee, Lillian. 1999. Measures of distributional similarity. In Proceedings of the37th Annual Meeting of the ACL.

Lee, Michael D., Brandon Pincombe, and Matthew Welsh. 2005. An empiricalevaluation of models of text document similarity. In CogSci2005, pages 1254–1259.

Lenat, D. and R. Guha. 1990. Building Large Knowledge Based Systems. AddisonWesley.

Lenat, D. B. 1995. CYC: A large-scale investment in knowledge infrastructure.CACM, 38(11), November.


Phd Thesis - [PDF Document] (189)

Lenat, D.B. and E.A. Feigenbaum. 1990. On the thresholds of knowledge. Arti-ficial Intelligence, 47:185–250.

Lenat, Douglas B. 1997. From 2001 to 2001: Common sense and the mind ofHAL. In HAL’s Legacy. The MIT Press, pages 194–209.

Lenat, Douglas B., Ramanathan V. Guha, Karen Pittman, Dexter Pratt, andMary Shepherd. 1990. CYC: Towards programs with common sense. CACM,33(8), August.

Leopold, Edda and Joerg Kindermann. 2002. Text categorization with supportvector machines: How to represent texts in input space. Machine Learning,46:423–444.

Lewis, David D. 1991. Evaluating text categorization. In Proc. of the Speech andNatural Language Workshop, pages 312–318. Morgan Kaufmann, February.

Lewis, David D. 1992a. An evaluation of phrasal and clustered representationson a text categorization task. In Proceedings of the 15th ACM InternationalConference on Research and Development in Information Retrieval, pages 37–50.

Lewis, David D. 1992b. Representation and learning in information rerieval.Ph.D. thesis, Department of Computer Science, University of Massachusetts,February.

Lewis, David D. and W. Bruce Croft. 1990. Term clustering of syntactic phrases.In Proceedings of the 13th ACM International Conference on Research andDevelopment in Information Retrieval, pages 385–404.

Lewis, David D., Robert E. Schapire, James P. Callan, and Ron Papka. 1996.Training algorithms for linear text classifiers. In Proceedings of the 19th ACMInternational Conference on Research and Development in Information Re-trieval, pages 298–306.

Lewis, David D., Yiming Yang, Tony Rose, and Fan Li. 2004. RCV1: A newbenchmark collection for text categorization research. JMLR, 5:361–397.

Lin, Dekang. 1998. An information-theoretic definition of word similarity. InICML’98.

Liu, Tao, Zheng Chen, Benyu Zhang, Wei-ying Ma, and Gongyi Wu. 2004. Im-proving text classification using local latent semantic indexing. In ICDM’04,pages 162–169.


Phd Thesis - [PDF Document] (190)

Manning, Christopher D. and Hinrich Schuetze. 2000. Foundations of StatisticalNatural Language Processing. The MIT Press.

Markovitch, Shaul and Danny Rosenstein. 2002. Feature generation using generalconstructor functions. Machine Learning, 49(1).

Marlow, Cameron, Mor Naaman, danah boyd, and Marc Davis. 2006. Positionpaper, tagging, taxonomy, flickr, article, toread. In Collaborative Web TaggingWorkshop, The 15th International World Wide Web Conference, Edinburgh,Scotland, May.

Maron, M. 1961. Automatic indexing: an experimental inquiry. Journal of theACM, 8(3):404–417.

Matheus, Christopher J. 1991. The need for constructive induction. In L.A.Birnbaum and G.C. Collins, editors, 8th Int’l Workshop on Machine Learning,pages 173–177.

Matheus, Christopher J. and Larry A. Rendell. 1989. Constructive inductionon decision trees. In Proceedings of the 11th International Conference onArtificial Intelligence, pages 645–650.

Mcilwaine, I.C. 2000. The Universal Decimal Classification: Guide to its Use.UDC Consortium.

Meng, Weiyi, Wenxian Wang, Hongyu Sun, and Clement Yu. 2002. Concepthierarchy-based text database categorization. Knowledge and InformationSystems, 4:132–150.

MeSH. 2003. Medical subject headings (MeSH). National Library of Medicine.

Mihalcea, Rada. 2003. Turning wordnet into an information retrieval resource:Systematic polysemy and conversion to hierarchical codes. International Jour-nal of Pattern Recognition and Artificial Intelligence (IJPRAI), 17(1):689–704.

Mihalcea, Rada, Courtney Corley, and Carlo Strapparava. 2006. Corpus-basedand knowledge-based measures of text semantic similarity. In AAAI’06, July.

Mikheev, Andrei. 1999. Feature lattices and maximum entropy models. Infor-mation Retrieval.

Miller, George A. and Walter G. Charles. 1991. Contextual correlates of semanticsimilarity. Language and Cognitive Processes, 6(1):1–28.


Phd Thesis - [PDF Document] (191)

Mitchell, Tom, Richard Keller, and Smadar Kedar-Cabelli. 1986. Explanation-based generalization: A unifying view. Machine Learning, 1(1):47–80.

Mitra, Mandar, Amit Singhal, and Chris Buckley. 1998. Improving automaticquery expansion. In Proceedings of the 21st ACM International Conferenceon Research and Development in Information Retrieval, pages 206–214.

Mladenic, Dunja. 1998a. Feature subset selection in text learning. In Proceedingsof ECML-98, 10th European Conference on Machine Learning, pages 95–100.

Mladenic, Dunja. 1998b. Turning Yahoo into an automatic web-page classifier.In Proceedings of ECAI’98, 13th European Conference on Artificial Intelli-gence, pages 473–474.

Mladenic, Dunja and Marko Grobelnik. 1998a. Feature selection for clasificationbased on text hierarchy. In Working notes of Learning from Text and theWeb, Conference on Automated Learning and Discovery (CONALD-98).

Mladenic, Dunja and Marko Grobelnik. 1998b. Word sequences as features intext-learning. In Proc. of ERK-98, 7th Electrotechnical and Computer ScienceConference, pages 145–148.

Murphy, Patrick M. and Michael J. Pazzani. 1991. ID2-of-3: Constructive induc-tion of M-of-N concepts for discriminators in decision trees. In Proceedings ofthe 8th International Conference on Machine Learning, pages 183–188. Mor-gan Kaufmann.

Nigam, Kamal, Andrew McCallum, and Tom Mitchell. 2006. Semi-supervisedtext classification using EM. In Olivier Chapelle, Bernhard Schoelkopf, andAlexander Zien, editors, Semi-Supervised Learning. MIT Press, Boston, MA.

Nigam, Kamal, Andrew McCallum, Sebastian Thrun, and Tom Mitchell. 2000.Text classification from labeled and unlabeled documents using EM. MachineLearning, 39(2-3):103–134.

2006. Open Directory Project.

Pagallo, Giulia and David Haussler. 1990. Boolean feature discovery in empiricallearning. Machine Learning, 5(1):71–99.

Pang, Bo, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? Sen-timent classification using machine learning techniques. In EMNLP’02, pages79–86.


Phd Thesis - [PDF Document] (192)

Peng, Fuchun, Dale Schuurmans, and Shaojun Wang. 2004. Augmenting naiveBayes classifiers with statistical language models. Information Retrieval, 7(3-4):317–345.

Peng, Fuchun and Dale Shuurmans. 2003. Combining naive Bayes and n-gramlanguage models for text classification. In Proceedings of the 25th EuropeanConference on Information Retrieval Research (ECIR-03).

Pereira, Fernando, Naftali Tishby, and Lillian Lee. 1993. Distributional clus-tering of english words. In Proceedings of the 30th Annual Meeting of theAssociation for Computational Linguistics, pages 183–190.

Pincombe, Brandon. 2004. Comparison of human and latent semantic analysis(LSA) judgements of pairwise document similarities for a news corpus. Tech-nical Report DSTO-RR-0278, Information Sciences Laboratory, Defence Sci-ence and Technology Organization, Department of Defense, Australian Gov-ernment, September.

Porter, M.F. 1980. An algorithm for suffix stripping. Program, 14(3):130–137.

Press, William H., Saul A. Teukolsky, William T. Vetterling, and Brian P. Flan-nery. 1997. Numerical Recipes in C: The Art of Scientific Computing. Cam-bridge University Press.

Quinlan, J. Ross. 1993. C4.5: Programs for Machine Learning. Morgan Kauf-mann.

Rada, Roy and Ellen Bicknell. 1989. Ranking documents with a thesaurus.JASIS, 40(5):304–310, September.

Raina, Rajat, Andrew Ng, and Daphne Koller. 2006. Constructing informa-tive priors using transfer learning. In Proceedings of the 23rd InternationalConference on Machine Learning (ICML), Pittsburgh, PA.

Raskutti, Bhavani, Herman Ferra, and Adam Kowalczyk. 2001. Second or-der features for maximizing text classification performance. In L. De Raedtand P. Flach, editors, Proceedings of the European Conference on MachineLearning (ECML), Lecture notes in Artificial Intelligence (LNAI) 2167, pages419–430. Springer-Verlag.

Resnik, Philip. 1999. Semantic similarity in a taxonomy: An information-basedmeasure and its application to problems of ambiguity in natural language.JAIR, 11:95–130.


Phd Thesis - [PDF Document] (193)

Reuters, 1997. Reuters-21578 text categorization test collection, Distribution 1.0.Reuters.

Riloff, Ellen. 1995. Little words can make a big difference for text classification.In Edward A. Fox, Peter Ingwersen, and Raya Fidel, editors, Proceedings ofSIGIR-95, 18th ACM International Conference on Research and Developmentin Information Retrieval, pages 130–136, Seattle, US. ACM Press.

Rocchio, J. 1971. Relevance feedback in information retrieval. In The SMARTRetrieval System: Experiments in Automatic Document Processing. PrenticeHall, pages 313–323.

Rogati, Monica and Yiming Yang. 2002. High-performing feature selection fortext classification. In Proceedings of the International Conference on Infor-mation and Knowledge Management (CIKM’02), pages 659–661.

Roget, Peter. 1852. Roget’s Thesaurus of English Words and Phrases. LongmanGroup Ltd.

Rorvig, M.E. 1999. Images of similarity: A visual exploration of optimal simi-larity metrics and scaling properties of TREC topic-document sets. Journalof the American Society of Information Science, 50(8):639–651.

Rose, Tony, Mark Stevenson, and Miles Whitehead. 2002. The Reuters CorpusVolume 1—from yesterday’s news to tomorrow’s language resources. In Pro-ceedings of the Third International Conference on Language Resources andEvaluation.

Rowling, J.K. 1997. Harry Potter and the Philosopher’s Stone. Bloomsbury.

Rowling, J.K. 2001. Harry Potter and the Goblet of Fire. Bloomsbury.

Rubenstein, Herbert and John B. Goodenough. 1965. Contextual correlates ofsynonymy. Communications of the ACM, 8(10):627–633.

Ruiz, Miguel E. and Padmini Srinivasan. 2002. Hierarchical text categorizationusing neural networks. Information Retrieval, 5:87–118.

Ruthven, Ian and Mounia Lalmas. 2003. A survey on the use of relevancefeedback for information access systems. Knowledge Engineering Review,18(2):95–145.

Sable, Carl, Kathleen McKeown, and Kenneth W. Church. 2002. NLP foundhelpful (at least for one text categorization task). In Conference on EmpiricalMethods in Natural Language Processing.


Phd Thesis - [PDF Document] (194)

Sahami, Mehran, Susan Dumais, David Heckerman, and Eric Horvitz. 1998. Abayesian approach to filtering junk e-mail. In AAAI 98 Workshop on TextCategorization, July.

Sahami, Mehran and Timothy Heilman. 2006. A web-based kernel function formeasuring the similarity of short text snippets. In WWW’06. ACM Press,May.

Salton, G. and M.J. McGill. 1983. An Introduction to Modern InformationRetrieval. McGraw-Hill.

Salton, Gerard, editor. 1971. The SMART Retrieval System: Experiments inAutomatic Document Processing. Prentice Hall.

Salton, Gerard and Chris Buckley. 1988. Term weighting approaches in automatictext retrieval. Information Processing and Management, 24(5):513–523.

Santamaria, Celina, Julio Gonzalo, and Felisa Verdejo. 2003. Automatic associ-ation of web directories to word senses. Computational Linguistics, 29(3).

Scott, Sam. 1998. Feature engineering for a symbolic approach to text classifica-tion. Master’s thesis, U. Ottawa.

Scott, Sam and Stan Matwin. 1999. Feature engineering for text classification. InProceedings of the 16th International Conference on Machine Learning, pages379–388.

Sebastiani, Fabrizio. 1999. A tutorial on automated text categorisation. InAnalia Amandi and Ricardo Zunino, editors, Proceedings of ASAI-99, 1stArgentinian Symposium on Artificial Intelligence, pages 7–35, Buenos Aires,Argentina.

Sebastiani, Fabrizio. 2002. Machine learning in automated text categorization.ACM Comp. Surv., 34(1):1–47.

Singhal, Amit. 1998. AT&T at TREC-6. In Proceedings of the 6th Text REtrievalConference (TREC-6), pages 215–226.

Soucy, Pascal and Guy W. Mineau. 2001. A simple feature selection method fortext classification. In IJCAI, pages 897–902.

Strube, Michael and Simon Paolo Ponzetto. 2006. WikiRelate! Computingsemantic relatedness using Wikipedia. In AAAI’06, Boston, MA.


Phd Thesis - [PDF Document] (195)

Tishby, Naftali, Fernando Pereira, and William Bialek. 1999. The informationbottleneck method. In Invited paper to the 37th Annual Allerton Conferenceon Communication, Control,and Computing.

Toyama, K. and E. Horvitz. 2000. Bayesian modality fusion: Probabilistic inte-gration of multiple vision algorithms for head tracking. In Proceedings of the4th Asian Conference on Computer Vision.

Turney, Peter. 2002. Thumbs up or thumbs down? Semantic orientation appliedto unsupervised classification of reviews. In Proceedings of the ACL.

Turney, Peter. 2005. Measuring semantic similarity by latent relational analysis.In Proceedings of the Nineteenth International Joint Conference on ArtificialIntelligence (IJCAI-05), pages 1136–1141, Edinburgh, Scotland.

Turney, Peter. 2006. Similarity of semantic relations. Computational Linguistics,32(3):379–416.

Turney, Peter and Michael L. Littman. 2002. Unsupervised learning of semanticorientation from a hundred-billion-word corpus. Technical Report ERB-1094,National Research Council Canada, May.

Urena-Lopez, L. Alfonso, Manuel Buenaga, and Jose M. Gomez. 2001. Integrat-ing linguistic resources in TC through WSD. Computers and the Humanities,35:215–230.

van Rijsbergen, C. J. 1979. Information Retrieval. Butterworths, London, 2edition.

Vapnik, V.N. 1995. The nature of statistical learning theory. Springer-Verlag.

Voorhees, Ellen M. 1994. Query expansion using lexical-semantic relations. InProceedings of the 17th International Conference on Research and Develop-ment in Information Retrieval, pages 61–69.

Voorhees, Ellen M. 1998. Using wordnet for text retrieval. In Christiane Fell-baum, editor, WordNet, an Electronic Lexical Database. The MIT Press.

Wang, Bill B., R.I. McKay, Hussein A. Abbass, and Michael Barlow. 2003. Acomparative study for domain ontology guided feature extraction. In Pro-ceedings of the 26th Australian Computer Science Conference (ASCS-2003),pages 69–78.

Wermter, Stefan and Chihli Hung. 2002. Selforganizing classification on theReuters news corpus. In Proceedings of the International Conference on Com-putational Linguistics (COLING).


Phd Thesis - [PDF Document] (196)

Widrow, B. and S.D. Stearns. 1985. Adaptive Signal Processing. Prentice Hall.

Wikipedia. 2006. Wikipedia, the free encyclopedia.

Wong, S.K.M., Wojciech Ziarko, and Patrick C.N. Wong. 1985. Generalizedvector spaces model in information retrieval. In Proceedings of the 8th ACMInternational Conference on Research and Development in Information Re-trieval, pages 18–25.

Wu, Huiwen and Dimitrios Gunopulos. 2002. Evaluating the utility of statisticalphrases and latent semantic indexing for text classification. In ICDM’02,pages 713–716.

Xu, Jinxi and W. Bruce Croft. 1996. Query expansion using local and globaldocument analysis. In Proceedings of the 19th International Conference onResearch and Development in Information Retrieval, pages 4–11.

Xu, Jinxi and W. Bruce Croft. 2000. Improving the effectiveness of informa-tion retrieval with local context analysis. ACM Transactions on InformationSystems, 18(1):79–112.

Yang, Yiming. 1999. An evaluation of statistical approaches to text categoriza-tion. Information Retrieval, 1(1-2):69–90.

Yang, Yiming. 2001. A study on thresholding strategies for text categoriza-tion. In Proceedings of the 24th International Conference on Research andDevelopment in Information Retrieval, pages 137–145.

Yang, Yiming and Xin Liu. 1999. A re-examination of text categorization meth-ods. In SIGIR’99, pages 42–49.

Yang, Yiming and Jan Pedersen. 1997. A comparative study on feature selectionin text categorization. In Proceedings of the 14th International Conference onMachine Learning, pages 412–420.

Yang, Yiming, Sean Slattery, and Rayid Ghani. 2002. A study of approachesto hypertext categorization. Journal of Intelligent Information Systems,18(2/3):219–241.

Yu, Shipeng, Deng Cai, Ji-Rong Wen, and Wei-Ying Ma. 2003. Improvingpseudo-relevance feedback in web information retrieval using web page seg-mentation. In Proceedings of the 12th International World Wide Web Con-ference (WWW’03), Budapest, Hungary, May. ACM Press.


Phd Thesis - [PDF Document] (197)

Zelikovitz, Sarah and Haym Hirsh. 2000. Improving short-text classification usingunlabeled background knowledge to assess document similarity. In Proceedingsof the 17th International Conference on Machine Learning.

Zelikovitz, Sarah and Haym Hirsh. 2001. Using LSI for text classification in thepresence of background text. In Proceedings of the Conference on Informationand Knowledge Management.

Zesch, Torsten and Iryna Gurevych. 2006. Automatically creating datasets formeasures of semantic relatedness. In Proceedings of the ACL Workshop onLinguistic Distances, pages 16–24, Sydney, Australia, July.

Zobel, Justin and Alistair Moffat. 1998. Exploring the similarity space. ACMSIGIR Forum, 32(1):18–34.


Phd Thesis - [PDF Document] (198)


Phd Thesis - [PDF Document] (199)

il`ehqwh rcin xefgi`l zeipekz zipa

mlerd rci zxfra

uiaelixab ipbai

Phd Thesis - [PDF Document] (200)

Phd Thesis - [PDF Document] (201)

il`ehqwh rcin xefgi`l zeipekz zipa

mlerd rci zxfra

xwgn lr xeaig

x`ezd zlawl zeyixcd ly iwlg ielin myl

diteqelitl xehwec

uiaelixab ipbai

l`xyil ibelepkh oekn — oeipkhd hpql ybed

2006 xanvc dtig f"qyz elqk

Phd Thesis - [PDF Document] (202)

Phd Thesis - [PDF Document] (203)

uiaewxn le`y 'xc ziigpda dyrp xwgndaygnd ircnl dhlewta

dcez zxkd

izenlzyda daicpd zitqkd dkinzd lr oeipkhl dcen ip`

Phd Thesis - [PDF Document] (204)

Phd Thesis - [PDF Document] (205)

.xzeia ddeabd d`eyzd z` daipn cinz rcia drwyd

oilwpxt oin'bpa --

Phd Thesis - [PDF Document] (206)

Phd Thesis - [PDF Document] (207)

mipipr okez

1 xivwz

5 `ean 17 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . rvend oexztd 1.110 . . . . . . . . . . . . . . . . . . . . . . . . . xwgnd ly zeircn zenexz 1.211 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xeaigd dpan 1.3

13 rwx 213 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . mihqwh beeiq 2.114 . . . . . . . . . . . . . . . . . . . . . . . . miknqn ly zeipekz 2.1.116 . . . . . . . . . . . . . . . . . . . . . . . . . . . zeipekz zxiga 2.1.219 . . . . . . . . . . . . . . . . . . . . . . . . zeipekz ikxr zriaw 2.1.320 . . . . . . . . . . . . . . . . . . . . beeiqd weic ocne`l zehiy 2.1.422 . . . . . . . . . . . . . . . . . . . . . miieqip zkxrd lr zexrd 2.1.522 . . . . . . . . . . . . . . . . . milin sqe` lr zqqeand dyiba zeira 2.225 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zeipekz zipa 2.3

29 zeipekz zipal dibelecezn 329 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . `ean 3.131 . . . . . . . . . . . . . . . . . . . . . . mini`zn rci ixb`n lr miveli` 3.233 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zeipekzd dpea zxivi 3.334 . . . . . . . . . . . . . . . . . . . . . . . . . miheaixh` zxiga 3.3.134 . . . . . . . . . . . . . . . . . . . . . . . . . . . . zeipekz zipa 3.3.235 . . . . . . . . . . . . . . . . . . . . . . . xywd zqqean zeipekz zipa 3.436 . . . . . . . . . . . . . . . . . . . . . . miinewn mixywd gezip 3.4.137 . . . . . . . . . . . . . . . . . . . . . . . . . . . zeipekz zxiga 3.4.238 . . . . . . . . . . . . . . . . . . . . . . . . zeipekz ikxr zriaw 3.4.338 . . . . . . . . . . . . . . . . . . `nbecd ly xzei winrn gezip 3.4.439 . . . . . . . . . . . . . . . . . . . . . . . . miikxxid rci ixb`na yeniy 3.542 . . . . . . . . . . . . mibyen oia miqgi mixicbny rci ixb`na yeniy 3.6

Phd Thesis - [PDF Document] (208)

45 dictiwiee geztd jixcnd zxfra zeipekz zipal dibeleceznd yenin 445 . . . . . . . . . . . . . . . . . . . geztd jixcnd zxfra zeipekz zipa 4.146 . . . . . . . . . . . . . . hpxhpi`d zervn`a sqep rci seqi` 4.1.146 . . . . . . . . . . . . . . . . . . miheaixh` zxigae yrx oepiq 4.1.250 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . yenin ihxt 4.1.351 . . . . . . . . . . . . . . . . . . . . . . . dictiwie zxfra zeipekz zipa 4.253 . . . . . . . . . . . . . . . . . . . . . . . . rci xb`nk dictiwie 4.2.154 . . . . . . . . . . . . . . . . . . . . . . . . zeipekzd dpea oekiz 4.2.255 . . . . . . . . . . . . . . . . . . mixn`n oia mixeyiwa yeniy 4.2.357 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . yenin ihxt 4.2.4

61 mihqwh beeiq xear zeipekz zipa ly dkxrd 561 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . mipezp itqe` 5.161 . . . . . . . . . . . . . . . . . . . . . . . . . . . Reuters-21578 5.1.162 . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 Newsgroups 5.1.262 . . . . . . . . . . . . . . . . . . . . . . . . . . . Movie Reviews 5.1.363 . . . . . . . . . . . . . . . . . . . . Reuters Corpus Version 1 5.1.464 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . OHSUMED 5.1.565 . . . . . . . . . . . . . . . . . . . . . . . . . . . mixvw miknqn 5.1.667 . . . . . . . . . . . . . . mipezp itqe` ly zihnehe` dykxd 5.1.768 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . miieqipd zhiy 5.269 . . . . . . . . . . . . . . . . . . . . . . mihqwh beeiql zizyz 5.2.170 . . . . . . . . . . . . . . . . . . . . . ziqiqad zkxrnd irevia 5.2.270 . . . . . . . . . . . . . . . . . . . . . . zeipekzd dpeaa yeniy 5.2.370 . . . . . . . . . . . . . geztd jixcnd qqean zeipekzd dpeaa yeniy 5.370 . . . . . . . . . . . . . . . . . zeipekzd zipa ly izeki` gezip 5.3.176 . . . . . . . . . . . . . . . . . . . zeipekzd zipa ly drtydd 5.3.277 . . . . . . . . . xywdd zqqean zeipekzd zipa ly drtydd 5.3.377 . . . . . . . . . . . . . . . . . . . . . . rcid agex ly drtydd 5.3.479 . . . . . . . . . . . . . . . . . . zeipekzd zxiga ly drtydd 5.3.581 . . . . . . . . . . . . . . . . . . . dixebhwd lceb ly drtydd 5.3.682 . . . . mixvw miknqn beeiq lr zeipekzd zipa ly drtydd 5.3.783 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ceair onf 5.3.885 . . . . . . . . . . . . . . . . . dictiwie qqean zeipekzd dpeaa yeniy 5.485 . . . . . . . . . . . . . . . . . zeipekzd zipa ly izeki` gezip 5.4.197 . . . . . . . . . . . . . . . . . . . zeipekzd zipa ly drtydd 5.4.299 . . . . . . . . . . . . . . . . . . . . . . rcid agex ly drtydd 5.4.399 . . . . . . . . . . . . . . . . . . . . . . . mixvw miknqn beeiq 5.4.4100 . . . . . . . . . . . . . . . . . . mixn`n oia mixeyiwa yeniy 5.4.5

Phd Thesis - [PDF Document] (209)

103 mihqwh oia zihpnq daxw ocne`l zeipekz zipaa yeniy 6103 . . . . . . . . . . . . . . . . . . . . . . . . . . zyxetn zihpnq dfilp` 6.1105 . . . . . . . . . . . . . . . . . . . . zyxetn zihpnq dfilp` ly dkxrd 6.2106 . . . . . . . . . . . . . . . . . . . . . . . . . . . . mipezp itqe` 6.2.1107 . . . . . . . . . . . . . . . . . . . . ipevigd rcid ly drtydd 6.2.2

109 zextq zxiwq 7109 . . . . . . . . . . . . . . . . . zeipekzd zveaw zagxdl zencew zeyib 7.1109 . . . . . . . . . . . . . . . . . . . . . mihqwh beeiq xear zeipekz zipa 7.2111 . . . . . . . . . . . . miaygenn mipelin zxfra zeipekz zipa 7.2.1113 . . . WordNet znerl geztd jixcnd :rci zexewn z`eeyd 7.2.2116 . . . . . . . . . . . . . . . . . . zebiiezn `l ze`nbeca yeniy 7.2.3117 . . . . . . . . . . . . . . . . . . . . . zetqep zeihpeelx zecear 7.2.4118 . . . . . . . . . . . . . . . . . . . . . . . . zihpnq daxwe ihpnq oeinc 7.3

123 mekiq 8

127 zeax zexizi zeipekz zegkepa mihqwh beeiq `127 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . `ean `.1129 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . miieqip zibelecezn `.2130 . . . . . . . . . . . . . . . . . . . . . . . . . . . . mipezp itqe` `.2.1131 . . . . . . . . . . . . . . . . zeipekz zxiga ly drtydd iefig `.2.2132 . . . . . . . . . . . . . . WordNet zeqqean zeipekza yeniy `.2.3133 . . . . . . . . . . . . . . . . . . . zeipekz zxigal minzixebl` `.2.4133 . . . . . . . . . . . . . . . . . . . mihqwh beeiql minzixebl` `.2.5134 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zixitn` dkxrd `.3134 . . . . . . . . . . . . . . . . . . . . . ziqiqad zkxrnd irevia `.3.1136 . . . . . . . . . . . . . . . . zeipekz zxiga ly drtydd iefig `.3.2136 . . . . . . . . . . . . . . . . . . . . . . . . . . mibeeqn z`eeyd `.3.3138 . . . . . . . . . . . . . zepeyd zeipekzd zeveaw ly drtydd `.3.4138 . . . . . zeipekz zxigal mipeyd minzixebl`d ly drtydd `.3.5138 . . . . . . . . . . . . . . . . . zeipekzd ly zeihpeelxd zpiga `.3.6139 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ze`veza oeic `.4

143 mihqwh beeiql mibiiezn mipezp itqe` zxivi a143 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . `ean a.1146 . . . . . . . . . . . . . . . . . . . . . . . . . mipezp itqe` ly mixhnxt a.2147 . . . . . . . . . . . . . . . . . . . . . . wgxn ocne`l zeivwpet a.2.1150 . . . . . . . . . . . . . . . . . . . . . . . . zeixebhw ly zepekz a.2.2151 . . . . . . . . . . . . mipezp itqe` ly zihnehe` dxivil dibelecezn a.3151 . . . . . . . . . . . . . . . . . . . . . . . . . . mipezpd zykxd a.3.1

Phd Thesis - [PDF Document] (210)

153 . . . . . . . . . . . . . . . . . . . yrxd zxwal mipezpd oepiq a.3.2154 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zixitn` dkxrd a.4154 . . . . . . . . . . . . . . . . . . . . . . . . . . mipezpd zykxd a.4.1154 . . . . . . . . . . . . . . . . . . . . . . mihqwh beeiql zizyz a.4.2155 . . . . . . . beeiqd weic oial wgxnd zewixhn oia divlxewd a.4.3156 . . . . . . . . . . MAA oial wgxnd zewixhn oia divlxewd a.4.4157 . . . . . . . . . . . . . . . mipezpd itqe` zxivi ly zeipeb-ax a.4.5159 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . mekiq a.5

163 zexewn zniyx

i xivwz

Phd Thesis - [PDF Document] (211)

mixei` zniyx

30 . . . . . . . . . . . . . . . . . . . . . . . mihqwh beeiql zniiwd dyibd 3.130 . . . . . . . . . . . . . . . zrvend dhiyd zxfra mihqwh beeqn zipa 3.235 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zeipekzd dpea zxivi 3.335 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zeipekz zipa 3.436 . . . . . . . . . . . . . . . . . . . daexn divelefx zhiya zeipekz zipa 3.539 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . zeipekz zipal `vbec 3.641 . . . . . . . . . . . . . . . ikxxid rci xb`n zxfra zeipekz dpea zxivi 3.742 . . . . . . . . . . . . . . . . . zikxxid dibelehpe` zxfra zeipekz zipa 3.843 . . . . . . . . . . . . . . . . . . mibyen oia miqgi zxfra zeipekz zipa 3.9

57 . . . . . . . . . . . . . . . . . . . . mixeyiwa yeniy jez zeipekz zipa 4.158 . . . . xzei miillk mibyenl mixeyiwa wx yeniy jez zeipekz zipa 4.2

79 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xywdd jxe` iepiy 5.181 . . . . . . . . . . . . . . . . . . . . . . . . . . . (Movies) zeipekz zxiga 5.281 . . . . . . . . . . . . . . . . . . . . . (RCV1/Topic-16) zeipekz zxiga 5.382 . . . . . . . . (RCV1/Industry) dixebhwd lceba zelzk rvenn xetiy 5.483 . . . . . . . . . . (RCV1/Topic) dixebhwd lceba zelzk rvenn xetiy 5.5

106 . . . . . . . . . . . . . . . . . . . . . . . . . . . rci qqean ihpnq gzpn 6.1

132 . . . . . . . . . . . . . . . . . mipey mipezp itqe`a zeipekzd zebltzd `.1135 . . . . . . . . . . . . . . . . . . . . . . . . . . ziqiqad zkxrnd irevia `.2135 . . . . . . . zeipekz zxiga ly zepeyd zenxa SVM ly weica xetiy `.3137 . . . . . . . . . . . . . . . zeipekzd lka yeniy jez mibeeqnd z`eeyd `.4139 . . . . . . . . . . . . . . . . . . . . . . . . . milin sqe` zervn`a beeiq `.5139 . . . . . . . . . . . . . . . . zagxend zeipekzd zveaw zervn`a beeiq `.6140 . . . . . . . . . . . . zeipekzd zxiga oial SVM ly weicd oia zelzd `.7140 . . . . . . . . . . . . zeipekzd zxiga oial C4.5 ly weicd oia zelzd `.8141 . . . . . . . . . . . . . . . . . . . . . . . xzeia zeaehd zeipekzd lehia `.9

152 . . . . . . . . . . . . . . . . . ievxd wgxna ze`vnpd zeixebhw xezi` a.1156 . . . . . . . . . . . . . . . . . . . . . . itxbd wgxna zelzk SVM weic a.2

Phd Thesis - [PDF Document] (212)

156 . . . . . . . . . . . . . . . . . . il`ehqwhd wgxna zelzk SVM weic a.3157 . . . . . . . . . . . . . . . . itxbd wgxna zelzk il`ehqwhd wgxnd a.4158 . . . . . . . . . . . . . . . . . . . . . . . . . itxbd wgxna zelzk MAA a.5158 . . . . . . . . . . . . . . . . . . . . . il`ehqwhd wgxna zelzk MAA a.6159 . . . . . . . . . . . . . . . geztd jixcna miitxbd miwgxnd zebltzd a.7159 . . . . . . . . . . . geztd jixcna miil`ehqwhd miwgxnd zebltzd a.8

Phd Thesis - [PDF Document] (213)

ze`lah zniyx

49 . . . . . . . . . . . . . . . . . . . . . . . . miheaixh` zxigal ze`nbec 4.1

65 . . . . . . . . . . . . RCV1 mipezpd sqe`a zeixebhw zeveaw zxcbd 5.166 . . . . . . . . . OHSUMED mipezpd sqe`a zeixebhw zeveaw zxcbd 5.271 . . . . . . . . . . . . . . . . . . . . . . . . . . ziqiqad zkxrnd irevia 5.376 . . . . . . . . . . . . . . `nbecl htyn xear zepey`xd zeipekzd xyr 5.477 . . . . . . . . . . . . . . `nbecl htyn xear zepey`xd zeipekzd xyr 5.578 . . . . . . . . . . . . . . . . . . . zeipekz zipa ilae mr mihqwh beeiq 5.680 . . . . . . . . . . . geztd jixcnd ly miwlg zervn`a zeipekz zipa 5.784 . . . . . . . . . . . . . . . zeipekz zipa ilae mr mixvw miknqn beeiq 5.885 . . . . . . . . . . . . . . . . . . . . . . . . . . mipezpd itqe` ly milcb 5.998 . . . . . . . . . . . . . . . . . . . . . . . . . zeipekzd zipa ly drtydd 5.1099 . . . . . . . . . . . . . . dictiwie ly ze`qxb izy ly milcb z`eeyd 5.11100 . . . . . . . . . . . . . dictiwie ly xzei dycgd dqxbd ly drtydd 5.12101 . . . . . . . . . . . . . . . . . . . . mixvw miknqn xear zeipekz zipa 5.13102 . mixvw miknqn xear zeipekz zipal mixn`n oia mixeyiwa yeniy 5.14

108 . . . . . . . . . . . . . . . . . . . . . . . . zeccea milin xear ze`vezd 6.1108 . . . . . . . . . . . . . . . . . . . . . . minly mihqwh xear ze`vezd 6.2

137 . . . . . . . . . . . . zeipekz zxiga ly zepeyd zenxa mibeeqnd weic `.1137 . . . . . . . . . . mibeeqnd oia weicd ilcad ly zihqihhq zewdaen `.2

Phd Thesis - [PDF Document] (214)

Phd Thesis - [PDF Document] (215)


mpyi ef dniynl .zeixebhwa mihqwh ly ihnehe` beiza wqer mihqwh beeiq

migeeic xezi` ,`yep itl miknqn itqe` oebx` ,laf x`ec iedif oebk ,miax miyeniy

z` zbviin mihqwh beeiql zlaewnd dyibd .cere ,oiricend megza miihpeelx

zexiaq ze`vez ozep df bevii lr qqaznd beeiq .xecq izla milin sqe`k hqwhd

mr cgi .beeqnd oeni`l ze`nbec daxd yiy i`pzae mikex` mihqwha xaecn xy`k

zveawa e` mixvw mihqwha xaecn xy`k zxkip dxeva micxei dyibd irevia ,z`f

.dphw oeni`

ynzydl dleki `id oky ,dceqin zlaben dpid milin sqe` lr zqqeand dyib

-ynzyn mipeyd miknqndy i`pzae ,yxetn ote`a miknqna xkfed xy` rcina wx

lrn zellkd zeyrl zlbeqn dpi` dyibd ,hxta .ziawr dxeva oewiqwl eze`a mi

.oeni`d iknqna zeriten opi` xy` dwicad iknqnay milinn mlrzz okle ,milin

-xicza oeni`d zveawa zeriten opi` xy` zetcxp milina mb lthl dywzn dyibd

,milin ly zernyn-ax zxzd rval zlbeqn dpi` ef dyib ,ok enk .witqn ddeab ze

.odly ixewnd xywda zecaern opi` zeirnyn-ax milin oky

oebk ,zepey zewipkh zervn`a hqwhd bevii z` xtyl miax zepeiqip eyrp xara

.ixiagz gzpna yeniy oke ,xaic iwlga milin beiz ,(n-grams) milin zexcqa yeniy

dirad icnin mevnvl zeyiba oke (clustering) milin ixava eynzyd zexg` zehiy

jxca lw miyp`l ,z`f mr cgi .zlaben dglvd aexl dzid el` zepeiqipl .LSA enk


Phd Thesis - [PDF Document] (216)

.miyp`l yiy axd rcid xve` lya z`fe ,hqwhd `yep z` zedfl llk

myl mlerd rci ixb`n lr zqqazn xy` dycg dyib mirivn ep` ef dceara

ynzyi xy` zeipekz dpeaa ynzyp ,hqwhd beeiq iptl .hqwhd bevii zxyrd

xyek zelra zeycg zeipekza heytd milind sqe` z` xiyrdl zpn lr ipevig rcia

,zihnehe` dxeva zrvazn zeipekzd ziipa .mbeeql yiy mi`yepd oia aeh dcxtd

,zepexg`d mipya hpxhpi`d zyx zehytzd zece` .iyep` rci ixb`na yeniy jez

ly hpxhpi`d jixcn ,geztd jixcnd oebk ,lkl miyibpd miax rci ixb`n meid mpyi

.dictiwie oke ,Yahoo

z` bxcyl dzxhny ,zecnel zekxrn megza dreci dhiy dpid zeipekz ziipa

.xteyn dcxtd xyek zelra ,zeycg zeipekz zxfra ze`nbecd ly ixewnd beviid

minrt dk cr dhiyd dzqep mihqwh ceair megza j` ,miax miyeniy ef dhiyl

.ziwlg dglvdae ,zehren

yiy rcid ixb`nl dyib zecnel zekxrnl zepwdl dpid zigkepd dceard zxhn

miynnne ,ipevig rcia yeniyl zillk dibelecezn mixicbn ep` ef dceara .mc` ipal

ynzydl zpn lr .dictiwiee geztd jixcnd - minieqn rci ixb`n ipy xear dze`

mibyenl hqwh irhw zetnl lbeqn xy` ,xfr beeqn mipea ep` ,dl`d rcid zexve`a

dl` mibyena miynzyn ep` ,okn xg`l .dictiwieae geztd jixcna miihpeelxd

agxend zeipekzd agxna mihqwh bevii .milind sqe` z` zexiyrny zeipekz xeza

.beeiqd weic xetiyl zeax mxez

ixt mdipy .mbeqn xzeia milecbd rcid ixb`n mpid dictiwiee geztd jixcnd

ziyteg dcxedl mipzip mixb`nd ipy .lazd iagxa miacpzn itl` ze`n ly dcear

dxiyre dwenr dikxxida obxe`n rcid geztd jixcna .aygna ceairl dlwy dxeva

inegze zepne` ,d`etx oebk miitivtq minegza rci ode illk rci od dlikn xy` ,c`n

`ide ,mlera xzeia dlecbd zaygennd dictelwivp`d dpid dictiwie .mipeyd rcnd

ynzydl zexyt`d lr zeax xaec xara .mixn`n ly lecb sqe` xeza zpbxe`n


Phd Thesis - [PDF Document] (217)

dtyd zpad ly ax iyewa lwzp df oeirx j` ,miaygnl rci ziipwdl dictelwivp`a

ynzydl aygnl zxyt`ny dyib dpey`xl mirivn ep` ef dceara .aygn ici lr

rci ly xfr xb`n `lle dtyd zpadl miicerii milecen `ll ,oixyina dictelwivp`a

.mlerd lr illk

.zeirah zety ceair megza miax miyeniy zrvend dhiyl ik mipin`n ep`

,epiidc ,sqep megza mb dhiy z` milirtn ep` mihqwh beeiql sqepa ,z`fd dceara

mipeyd minegza c`n dvetp ef dniyn .hqwh irhw oia zihpnq daxw ocne`

mihqwhe milin ueaiw ,milin ly zernyn-ax zxzd oebk ,ziaeyig zepyla ly

mipelina exfrp ef diral zepexzt ,xara .'eke ze`iby iedif ,zernyna daxw itl

mirivn ep` o`k .(Latent Semantic Analysis) dieag zihpnq dfilp`ae miaygenn

xy` ,(Explicit Semantic Analysis) zyxetn zihpnq dfilp` z`xwpd dycg dyib

ly icnin-axd agxna mihqwh ly zernyn zbviine zeipekzd dpeaa zynzyn

.dictiwiene geztd jixcndn mibyen

-lk dibelecezn mirivn ep` ,ziy`x .zeircn zenexz xtqn zigkepd dcearl

ziipa .mlera milecbdn rci ixb`n zxfra zeipekz ziipal minzixebl` xtqne zil

rcin lvpl zlbeqne ,ax iyep` rcia zynzyn el` mixb`na yeniy jez zeipekz

dycg dyib mirivn ep` ,ok enk .cala beeqnd hqwhdn dwqdl ozip epi` xy`

rvan jk ici lre zepey divelefx zenxa mihqwh gzpn xy` ,xywd qqean gezipl

ixb`n ipy xear zrvend dibeleceznd z` epynin .milin ly zernyn-ax zxzd

xzeia milecbd rcid ixb`n mpid xy` ,dictiwiee geztd jixcnd - minieqn rci

zeaygenn zekxrn ly mirevia zxkip dxeva xtiy zrvend dhiya yeniy .mbeqn

oia zihpnq daxw ocne`e mihqwh beeiq ,epiidc ,ziaeyig zepyla ly minegz ipya

mipyay zexnl beeiqd weic z` xytl epglvd mihqwh beeiq megza .hqwh irhw

ik ,oiivl aeyg ok enk .zeniiwd zekxrnd ireviaa xetiy did `le hrnk zencewd

milin oia od daxwd aeyigl cig` oexzt zpzep zihpnq daxw ocne`l eply dyibd


Phd Thesis - [PDF Document] (218)

.c`n mikex` hqwh irhw oia ode ,zeccea


Phd Thesis - [PDF Document] (2024)


Where can I download PhD thesis for free? › aims to be the best possible resource for finding open access graduate theses and dissertations published around the world. Metadata (information about the theses) comes from over 1100 colleges, universities, and research institutions. OATD currently indexes 6,912,508 theses and dissertations.

Is 100 pages enough for a PhD thesis? ›

How long should a PhD thesis be? A PhD thesis (or dissertation) is typically 60,000 to 120,000 words (100 to 300 pages in length) organised into chapters, divisions and subdivisions (with roughly 10,000 words per chapter) – from introduction (with clear aims and objectives) to conclusion.

Should you submit your dissertation as a PDF? ›

This must be a single file. If your manuscript is in Word or RTF format, you must convert it into a PDF. Optional Supplementary files (images, data, etc.) that are an integral part of the dissertation/thesis, but not part of the full text.

How many months to write PhD thesis? ›

On average, it can take several months to a few years to complete a PhD thesis. The research, data collection, and analysis stages can span several years, with the PhD thesis writing phase itself often lasting several months.

What does a PhD thesis look like? ›

The typical PhD thesis structure will contain four chapters of original work sandwiched between a literature review chapter and a concluding chapter. There is no universal rule for the length of a thesis, but general guidelines set the word count between 70,000 to 100,000 words.

Can I write my PhD thesis in one month? ›

The answer to this question is yes! You absolutely can write a thesis in 30 days. And you can write your thesis, from start to finish, without the emotional distress that often comes with such a monumental task.

How much does it cost to print a PhD thesis? ›

A4 Hardback Binding (wording on front and spine included) on a 48 hour service
48 hour service
1 Copy£27.50
2 Copies£52.50
3 Copies£75
4 Copies£95

Can I pay someone to write my PhD thesis? ›

Hiring a professional writer for your dissertation can have its perks. Firstly, it relieves you of the stress and pressure that comes with writing such a significant academic piece. Professional writers bring expertise, experience, and a fresh perspective to your work.

What is the shortest PhD thesis? ›

Ever seen the film A Beautiful Mind? The mathematician that film was based on, John Nash, has one of the shortest PhD dissertations ever published: 'Non-Cooperative Games'. It has a grand total of 26 pages, and only cites two references. That thesis went on to found the basis for his paper on the devel…

Is 30k words enough for a PhD thesis? ›

A PhD thesis must not exceed 80,000 words, and will normally be over 60,000 words. This word limit includes footnotes and endnotes, but excludes appendices and reference list / bibliography.

How many publications are considered good for a PhD? ›

While there is no hard-and-fast, field-independent rule for how many publications you should publish in grad school, some departments waive the requirement to write a dissertation if a candidate has published three peer reviewed articles. I therefore dutifully submitted three first-author manuscripts during my PhD.

What is the minimum word count for a PhD thesis? ›

Word count

Arts and Humanities: 40,000 words (MPhil); 75,000 words (PhD) Health: 40,000 words (MPhil); 75,000 words (PhD, MD) Science: 40,000 words (MPhil); 80,000 words (PhD) Social Sciences: 40,000 words (MPhil); 75,000-100,000 words (PhD)

Do employers read your dissertation? ›

You're generally only going to be asked to go into depth about your research if it's relevant to your role (eg interviewing for a PhD).

Should I use Grammarly for my dissertation? ›

Yes, it is okay to use Grammarly on a thesis or dissertation, and I will explain why. Grammarly is an online writing tool that assists you in making sure your writing is clear and error-free. It can check for grammar, spelling, and punctuation errors and make suggestions to help you improve your writing.

What is the format of a PhD thesis? ›

The thesis should be written in English. It must be typewritten on A4 size paper (21 cm x 29.7 cm) in a clear and legible font (e.g., Times New Roman 12). As far as possible, use the same font for the entire thesis, but, if necessary, different fonts may be used within Tables, Figures, and Appendices.

How do I start writing my PhD thesis? ›

A Step-by-Step Guide To Writing An Excellent Ph. D. Thesis
  1. Selecting a Topic. The first step in Ph. ...
  2. Developing The Proposal. Upon the committee's approval, you can already begin writing a proposal that will give an overview of your dissertation. ...
  3. Conducting Research. ...
  4. Writing The Thesis. ...
  5. Editing and Proofreading.

What are the steps of a PhD thesis? ›

The Dissertation Process Explained in 6 Simple Steps
  • Step 1: Brainstorm Topics. ...
  • Step 2: Find a Faculty Mentor and Committee Assignment. ...
  • Step 3: Develop and Submit a Proposal. ...
  • Step 4: Conduct Research and Data Analysis. ...
  • Step 5: Draft Your Manuscript. ...
  • Step 6: Defend Your Thesis.

How should a PhD thesis be structured? ›

Theses will usually contain most or all of the following sections:
  1. Title page.
  2. Abstract.
  3. Acknowledgements.
  4. Contents page(s)
  5. Introduction.
  6. Literature review (sometimes within the introduction)
  7. Materials/sources and methods (can be part of every chapter if these are different per chapter)
  8. Themed topic chapters.

Top Articles
Latest Posts
Article information

Author: Rubie Ullrich

Last Updated:

Views: 6500

Rating: 4.1 / 5 (52 voted)

Reviews: 83% of readers found this page helpful

Author information

Name: Rubie Ullrich

Birthday: 1998-02-02

Address: 743 Stoltenberg Center, Genovevaville, NJ 59925-3119

Phone: +2202978377583

Job: Administration Engineer

Hobby: Surfing, Sailing, Listening to music, Web surfing, Kitesurfing, Geocaching, Backpacking

Introduction: My name is Rubie Ullrich, I am a enthusiastic, perfect, tender, vivacious, talented, famous, delightful person who loves writing and wants to share my knowledge and understanding with you.