Abduction, explanation and relevance feedback
Ian Ruthven · 1999
Relevance feedback (RF) is an interactive technique that is intended to automatically improve an information retrieval system's representation of a search based on documents that a user has assessed as being relevant. Our research centres around defining a formal model of RF that is based on detecting what aspects of a term's use in a document, “term characteristics”, are good indicators of relevant material. This model also incorporates behavioural aspects of searching in order to offer better support to how users search. Term characteristics give values to terms according to how they are used within documents or collections. In order to deal with the diversity of system and user variables that can affect how term characteristics should be selected we formalise a model of RF that is based on abductive inference or abduction. Abduction is primarily a model of explanation. Our RF approach is a process of active selection based on inference: inferring which documents, terms and characteristics of terms to use in RF. The inference mechanisms will select good components of explanations and weight them according to their value in explaining the relevance assessments. The resulting components can be assembled into different types of explanation according to the type of search. (2 pages)