Weakly supervised relevance feedback based on an improved language model

Abstract

Relevance feedback, which traditionally uses the terms in the relevant documents to enrich the user's initial query, is an effective method for improving retrieval performance. This approach has another problem is that Relevance feedback assumes that most frequent terms in the feedback documents are useful for the retrieval. In fact, the reports of some experiments show that it does not hold in reality many expansion terms identified in traditional approaches are indeed unrelated to the query and harmful to the retrieval. In this paper, we propose to select better and more relevant documents with a clustering algorithm. And then we present an improved Language Model to help us identify the good terms from those relevant documents. Ours experiments on the 2008 TREC collection show that retrieval effectiveness can be much improved when the improved Language Model is used.