CE-QE: Cross-Encoder Query Expansion for Lexical Retrieval
A novel technique known as Cross-Encoder Query Expansion (CE-QE) tackles the issue of semantic vocabulary discrepancies in lexical retrieval, specifically where BM25 struggles to locate pertinent documents that phrase answers differently than the query. CE-QE employs a cross-encoder on the leading semantic search outcomes to pinpoint crucial terms that are subsequently added to the BM25 query. This approach diverges from traditional pseudo-relevance feedback, which depends on potentially inaccurate top results from BM25, by using results from the semantic retriever to initiate expansion, thereby preventing query drift. Additionally, it contrasts with generative query expansion techniques like HyDE and Query2doc, which rely on large language models to generate text. The research paper can be found on arXiv with the identifier 2608.00452.
Key facts
- CE-QE reads per-token relevance attributions from a cross-encoder applied to top semantic search results.
- CE-QE selects terms the cross-encoder treats as decisive and appends them to the BM25 query.
- CE-QE avoids self-reinforcing query drift by seeding expansion from the semantic retriever's results.
- CE-QE contrasts with classical pseudo-relevance feedback and generative query expansion (HyDE, Query2doc).
- The paper is on arXiv with identifier 2608.00452.
- BM25 is blind to semantic vocabulary gaps, missing documents that phrase answers differently.
- No amount of reranking or fusion can recover documents not in the candidate set.
- The method is called Cross-Encoder Query Expansion (CE-QE).
Entities
Institutions
- arXiv