Improving Retrieval Performance for Verbose ƒeries via Axiomatic Analysis of Term Discrimination Heuristic

摘要

Number of terms in a query is a query-speci€c constant that isrntypically ignored in retrieval functions. However, previous studiesrnhave shown that the performance of retrieval models varies forrndi‚erent query lengths, and it usually degrades when query lengthrnincreases. A possible reason for this issue can be the extraneousrnterms in longer queries that makes it a challenge for the retrievalrnmodels to distinguish between the key and complementary conceptsrnof the query. As a signal to understand the importance of arnterm, inverse document frequency (IDF) can be used to discriminaternquery terms. In this paper, we propose a constraint to model therninteraction between query length and IDF. Our theoretical analysisrnshows that current state-of-the-art retrieval models, such as BM25,rndo not satisfy the proposed constraint. We further analyze thernBM25 model and suggest a modi€cation to adapt BM25 so that itrnadheres to the new constraint. Our experiments on three TREC collectionsrndemonstrate that the proposed modi€cation outperformsrnthe baselines, especially for verbose queries.