Finding good feedback documents

Abstract

Pseudo-relevance feedback finds useful expansion terms from a set of top-ranked documents. It is often crucial to identify those good feedback documents from which useful expansion terms can be added to the query. In this paper, we propose to detect good feedback documents by classifying all feedback documents using a variety of features such as the distribution of query terms in the feedback document, the similarity between a single feedback document and all top-ranked documents, or the proximity between the expansion terms and the original query terms in the feedback document. By doing this, query expansion is only performed using a selected set of feedback documents, which are predicted to be good among all top-ranked documents. Experimental results on standard TREC test data show that query expansion on the selected feedback documents achieves statistically significant improvements over a strong pseudo-relevance feedback mechanism, which expands the query using all top-ranked documents.