LIS 601
Week6:Relevance Feedback
General Comments
- Relevance Feedback (RF) found to be one of the most powerful methods for improving IR performance:
- Improvements of 40 to 60 percent (Precision) noted (Salton 1989, 322).
- RF is an iterative process, best modelled as a continuous loop
- RF is a user-centered approach.
- Based on the two-fold idea that:
- Relevant documents should be more strongly similar to each other than they are to non-relevant documents; and,
- Users are the best judges of relevance.
- RF is closely related to ranking.
- Vector-space model and RF work very well together.
- The Ide dec-hi query modification method found to be best
- Boolean approaches are not as amenable to RF.
- Choice of similarity measure arbitrary (Cosine Coefficient favoured by many)
- Most useful for exhaustive searches of large collections (Think about Blair and Maron (1985)).
Important Relevance Feedback Styles
- Query Expansion with Term Reweighting
- Query Expansion without Term Reweighting
- Query Reweighting without Expansion
- Also one can decide whether to include negative feedback information (and then one must decide how much importance to give negative feedback information).
Some Useful URLs
Page creator: J. Stephen Downie
Page created: 15 October 1997
Page updated: 16 October 1997