Selective relevance feedback using term characteristics

This paper presents a new relevance feedback technique; selectively combining evidence based on the usage of terms within documents. By considering how terms are used within documents, we can better describe the features that might make a document relevant and thus improve retrieval effectiveness. In this paper we present an initial, experimental investigation of this technique, incorporating new and existing measures for describing the information content of a document. The results from these experiments positively support our hypothesis that extending relevance feedback to take into account how terms are used within documents can improve the performance of relevance feedback.