Utility-Based Information Distillation - Text Mining: Classification, Clustering, and Applications - page 216

Chapter 9

Utility-Based Information
Distillation

Yiming Yang and Abhimanyu Lad

9.1 Introduction ............................................................. 213

9.2 A Sample Task ........................................................... 216

9.3 Technical Cores .......................................................... 218

9.4 Evaluation Methodology ................................................. 221

.....................................................................

9.5 Data 225

................................................

9.6 Experiments and Results 226

9.7 Concluding Remarks ..................................................... 229

9.8 Acknowledgments ........................................................ 229

9.1 Introduction
Utility-based information distillation is a new challenge in information
retrieval, focusing on effective ways to combine technologies from adaptive
filtering (AF), novelty detection, anti-redundant passage ranking and flexible
user feedback. The ultimate goal is to improve the true utility of the system,
as well as to support effective and ecient user feedback. To see why utility-
based distillation is a practically important problem and an open challenge for
research, let us briefly outline the related work in Adaptive filtering and Topic
Detection and Tracking, and visit some limitations of the current solutions.

9.1.1 Related Work in Adaptive Filtering (AF)
Adaptive filtering is the task of online prediction of the relevance of each new
document in a temporally ordered sequence, with respect to a pre-specified
topic or query. Here we use the terms “topic” and “query” interchangeably
as synonyms. Based on the initial query words or topic description, and a few
positive examples of on-topic documents, the system maintains a profile for
each topic. The profile is incrementally updated whenever relevance feedback
is received from the user. The learning nature of AF systems makes them
more powerful than standard search engines without adaptation. A variety