Exploring sentence level query expansion in language modeling based information retrieval

Abstract

We introduce two novel methods for query expansion in information retrieval (IR). The basis of these methods is to add the most similar sentences extracted from pseudo-relevant documents to the original query. The first method adds a fixed number of sentences to the original query, the second a progressively decreasing number of sentences. We evaluate these methods on the English and Bengali test collections from the FIRE workshops. The major findings of this study are that: i) performance is similar for both English and Bengali; ii) employing a smaller context (similar sentences) yields a considerably higher mean average precision (MAP) compared to extracting terms from full documents (up to 5.9% improvemnent in MAP for English and 10.7% for Bengali compared to standard Blind Relevance Feedback (BRF); iii) using a variable number of sentences for query expansion performs better and shows less variance in the best MAP for different parameter settings; iv) query expansion based on sentences can improve performance even for topics with low initial retrieval precision where standard BRF fails.