Information Retrieval Advances using Relevance Feedback

Erik W. Selberg
Department of Computer Science and Engineering
University of Washington
Seattle, WA 98195
selberg@cs.washington.edu
http://www.cs.washington.edu/homes/selberg

August 12, 1997

Abstract:

Relevance Feedback (RF) is a powerful technique whereby a user can instruct an Information Retrieval (IR) system to find additional relevant documents by providing relevance information on certain documents or query terms. There are several major design decisions that can affect how a RF system can help users. In this paper, I present a brief literature survey covering two emphases of RF: term selection and term weighting in the two prevalent IR models: probabilistic and vector-based. I then summarize how both the techniques and the underlying models can be combined and to what ends. Finally, I hypothesize why Relevance Feedback is not prevalent on World Wide Web search services, and give suggestions on which methods may be appropriate and what, if anything, should be done to add feedback techniques to Web searching.

 
• Introduction & Motivation
• High Level Overview of IR
  • User-level View
  • Query Expansion and Relevance Feedback
  • Term Weighting and Term Selection
  • Common IR Techniques
  • Evaluation and Comparison of IR systems
• Formal Background
  • Definitions
  • Vector Model of IR
  • Probabilistic Model of IR
    • Robertson / Sparck Jones Formalization
    • Inference Networks
  • Boolean Model of IR
• Two Vector-Based Systems
  • Term Weighting in a Vector Model
  • Term Selection in a Vector Model
• Two Probabilistic Systems
  • Term Weighting in a Probabilistic Model
  • Term Selection in a Probabilistic Model
  • Term Selection using Traditional Probabilistic Methods
    • The Swets Model of IR System Performance
    • The wpq formula
• Compositions of Technologies
  • Combining Weighting and Selection under Similar Models
  • Combining Weighting and Selection under Different Models
  • Best Methods
  • Combining Redundant Techniques
  • Unstated Assumptions
    • Documents and abstracts are ``well-formed''
    • Users queries are ``well-formed''
    • Users make ``perfect'' choices
    • Closed World Assumption
• Speculation
  • Relevance Feedback is Out There
  • Feedback is Painful
  • Resource Requirements
  • User Access Patterns
  • Hyperlinks obviate Relevance Feedback
• Conclusions
• Acknowledgments
• References
• About this document ...