A Generative Theory of Relevance (eBook)

Victor Lavrenko is a lecturer at the School of Informatics at the University of Edinburgh, Scotland, UK. He received his Ph.D. in Computer Science from the University of Massachusetts Amherst in 2004. His dissertation focused on a generative framework for modeling relevance in Information Retrieval. In 2005 he joined the Center for Intelligent Information Retrieval at UMass as a post-doctoral research associate, working on statistical models for searching large semi-structured databases. From 2006 Victor worked as a language technology consultant for the Credit Suisse Group. Since 2000, he has served as a reviewer for SIGIR, CIKM, NAACL/HLT, IJCAI and NIPS conferences. Victor's current research interests include formal models for searching text in multiple languages, annotating and retrieving images, and detecting and tracking novel events in the news.

Introduction 17
Contributions 18
A new model of relevance 18
A new generative model 18
Minor contributions 19
Overview 19
Relevance 22
The many faces of relevance 22
A simple definition of relevance 22
User-oriented views of relevance 23
Logical views of relevance 24
The binary nature of relevance 25
Dependent and independent relevance 25
Attempts to Construct a Unified Definition of Relevance 27
Relevance in this book 30
Existing Models of Relevance 30
The Probability Ranking Principle 30
The Classical Probabilistic Model 32
The Language Modeling Framework 41
Contrasting the Classical Model and Language Models 47
A Generative View of Relevance 51
An Informal Introduction to the Model 51
Representation of documents and requests 52
Advantages of a common representation 53
Information retrieval under the generative hypothesis 56
Formal Specification of the Model 58
Representation of Documents and Queries 59
Document and query generators 59
Relevant documents 59
Relevance in the information space 60
Relevant queries 60
Summary of representations 61
Probability Measures 61
Distribution over the representation space 62
Distribution over documents and queries 63
Significance of our derivations 66
Summary of probability measures 66
Relevance Models 67
Frequentist interpretation: a sampling game 68
Bayesian interpretation: uncertainty about relevance 69
Multi-modal domains 69
Summary of relevance models 71
Ranked Retrieval 71
Probability ranking principle 71
Retrieval as hypothesis testing 73
Probability ratio or KL-divergence? 79
Summary of ranking methods 81
Discussion of the Model 82
Generative Density Allocation 85
Problem Statement 85
Objective 86
Existing Generative Models 86
The Unigram model 87
The Mixture model 88
The Dirichlet model 90
Probabilistic Latent Semantic Indexing (pLSI) 92
Latent Dirichlet Allocation 93
A brief summary 94
Motivation for a new model 95
A Common Framework for Generative Models 96
Unigram 97
Dirichlet 98
Mixture 99
pLSI 100
LDA 102
A note on graphical models 104
Kernel-based Allocation of Generative Density 105
Delta kernel 106
Dirichlet kernel 108
Advantages of kernel-based allocation 110
Predictive Effectiveness of Kernel-based Allocation 113
Summary 115
Retrieval Scenarios 117
Ad-hoc Retrieval 118
Representation 118
Examples of Relevance Models 121
Experiments 122
Relevance Feedback 130
Representation 130
Experiments 132
Cross-Language Retrieval 133
Representation 133
Example of a cross-lingual relevance model 138
Experiments 139
Significance of the cross-language scenario 145
Handwriting Retrieval 145
Definition 146
Representation 147
Experiments 150
Image Retrieval 152
Representation 154
Experiments 157
Video Retrieval 161
Representation 162
Experiments 163
Structured search with missing data 165
Representation of documents and queries 167
Probability distribution over documents and queries 167
Structured Relevance Model 168
Retrieving Relevant Records 168
Experiments 169
Topic Detection and Tracking 173
Definition 173
Representation 176
Link detection algorithm 177
Experiments 179
Conclusion 189
Limitations of our Work 193
Closed-universe approach 193
Exchangeable data 193
Computational complexity 194
Directions for Future Research 195
Relevance-based indexing 195
Hyper-linked and relational data 195
Order-dependent data 197
Dirichlet kernels 197
References 198
Index 208