Effective Contrastive Weighting for Dense Query Expansion

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12688–12704 July 9-14, 2023 ©2023 Association for Computational Linguistics Effective Contrastive Weighting for Dense Query Expansion Xiao Wang‡, Sean MacAvaney†, Craig Macdonald†, Iadh Ounis†, School of Computing Science, University of Glasgow, UK ‡x.wang.8@research.gla.ac.uk †{sean.macavaney, craig.macdonald, iadh.ounis}@glasgow.ac.uk Abstract Verbatim queries submitted to search engines often do not sufficiently describe the user’s search intent. Pseudo-relevance feedback (PRF) techniques, which modify a query’s representation using the top-ranked documents, have been shown to overcome such inade- quacies and improve retrieval effectiveness for both lexical methods (e.g., BM25) and dense methods (e.g., ANCE, ColBERT). For instance, the recent ColBERT-PRF approach heuristically chooses new embeddings to add to the query representation using the inverse document frequency (IDF) of the underlying tokens. However, this heuristic potentially ignores the valuable context encoded by the embeddings. In this work, we present a contrastive solution that learns to select the most useful embeddings for expansion. More specifically, a deep language model-based contrastive weighting model, called CWPRF, is trained to learn to discriminate between relevant and non-relevant documents for semantic search. Our experimental results show that our contrastive weighting model can aid to select useful expansion embeddings and outperform various baselines. In particular, CWPRF can improve nDCG@10 by upto to 4.1% compared to an existing PRF approach for ColBERT while maintaining its efficiency. 1 Introduction When using search engines, users frequently enter queries that insufficiently express their desired intent. For instance, a user who issues the query georgia run off elections may indeed be looking for details about a specific electoral procedure in the US state of Georgia. For search algorithms that rely on lexical matching, such as BM25, this can result in a lexical gap, since relevant documents may just as easily use different terms (e.g., GA and 2nd-round). Pseudo-Relevance Feedback (PRF) techniques are often employed to overcome such a lexical gap. Indeed, classical PRF techniques, 0 500 1000 1500 2000 2500 3000 PRF Stage Mean Execution time (ms) 0.66 0.68 0.70 0.72 0.74 nDCG@10 CWPRF (ours) ANCE-PRF Vector-PRF ColBERT-PRF CWPRF-AAAT CWPRF-OAAT ColBERT-PRF KMeans (default) ColBERT-PRF KMeans-Closest ColBERT-PRF KMedoids ANCE-PRF Vector-PRF Figure 1: Effectiveness (nDCG@10) versus dense PRF stage mean execution time on the TREC 2019 query set. such as RM3 (Abdul-Jaleel et al., 2004), have been widely used to enrich the user’s query with terms selected from the initial retrieval top-ranked docu- ments, i.e. the pseudo-relevance feedback set (Am- ati and Van Rijsbergen, 2002; Roy et al., 2016; Cao et al., 2008). This expanded query is capable of overcoming the lexical gap if the pseudo-relevance feedback documents are relevant and additional related terms can be identified (for instance adding GA to the query). However, there is a risk that the added terms drift the intent of the query (for instance, adding terms such as Tbilisi that relate to the country of Georgia rather than the US state). An alternative approach for overcoming the lexi- cal gap is to perform semantic search over …

Content truncated to 500 words