Comparing two blind relevance feedback techniques

Abstract

Query expansion based on Blind Relevance Feedback (BRF) has been demonstrated to be an effective technique for improving retrieval results. There are two types of BRF-based query expansion. BRF Type 1 (BRFT1) is the original version of BRF, where query expansion is performed on the BRF information extracted from top N documents selected from an initial search on the same collection that the target documents are in [1]. This collection is called “target collection” in this paper. BRF Type 2 (BRFT2) has been explored as an alternative to BRFT1. The query expansion is performed based on the BRF information of the top N documents selected from the initial search on a DIFFERENT collection. Such a collection is called “expansion collection” in this paper. The expanded query is then used to search on the target collection to find the relevant documents. The effectiveness of BRF depends on two key factors: 1) the documents selected from the initial search for BRF should contain reasonable number of topically relevant documents to the query; and 2) those selected documents should share the similar genre with the target relevant documents so that there is high chance that the important content terms used in these two sets of documents are the same[2]. Both BRFT1 and BRFT2 may encounter situations that at least one of the two conditions cannot be satisfied. For example, there are not enough truly relevant documents in the target collection for many topics in Robust track of TREC evaluation, which makes it difficult to utilize BRFT1 based query expansion techniques to improve the search results. However, with the amount of electronic resources available, it is often possible that both BRFT1 and BRFT2 can