Query expansion for language modeling using sentence similarities

• Authors: Debasis Ganguly;Johannes Leveling;Gareth J. F. Jones
• Venue: IRFC'11 Proceedings of the Second international conference on Multidisciplinary information retrieval facility
• Year: 2011

We propose a novel method of query expansion for Language Modeling (LM) in Information Retrieval (IR) based on the similarity of the query with sentences in the top ranked documents from an initial retrieval run. In justification of our approach, we argue that the terms in the expanded query obtained by the proposed method roughly follow a Dirichlet distribution which, being the conjugate prior of the multinomial distribution used in the LM retrieval model, helps the feedback step. IR experiments on the TREC ad-hoc retrieval test collections using the sentence based query expansion (SBQE) show a significant increase in Mean Average Precision (MAP) compared to baselines obtained using standard term-based query expansion using LM selection score and the Relevance Model (RLM). The proposed approach to query expansion for LM increases the likelihood of generation of the pseudo-relevant documents by adding sentences with maximum term overlap with the query sentences for each top ranked pseudorelevant document thus making the query look more like these documents. A per topic analysis shows that the new method hurts less queries compared to the base-line feedback methods, and improves average precision (AP) over a broad range of queries ranging from easy to difficult in terms of the initial retrieval AP. We also show that the new method is able to add a higher number of good feedback terms (the golden standard of good terms being the set of terms added by True Relevance Feedback). Additional experiments on the challenging search topics of the TREC-2004 Robust track show that the new method is able to improve MAP by 5.7% without the use of external resources and query hardness prediction typically used for these topics.