CeQe boosts BM25 recall by 47% with cross-encoder query expansion

Lexical retrieval like BM25 is fast but blind to semantic paraphrasing: if a document phrases an answer differently from the query, it never enters the candidate set, and downstream reranking can't recover it. The new paper from Kahirov, Deshpande, and Sundararaman proposes Cross-Encoder Query Expansion (CE-QE), a method that attacks this vocabulary gap by reading the per-token relevance attributions of a cross-encoder already used in hybrid retrieval pipelines. CE-QE looks at the top semantic search results, selects the terms the cross-encoder deems decisive, and appends them verbatim to the BM25 query.

Unlike pseudo-relevance feedback (which reuses BM25's own possibly wrong top results) or generative query expansion like HyDE (which can hallucinate out-of-corpus vocabulary), CE-QE copies every expansion term directly from a retrieved passage. This ensures no vocabulary is introduced that doesn't exist in the corpus, and the only added cost is attribution extraction on a cross-encoder that's already running for reranking. On seven BEIR datasets, CE-QE delivers substantial lexical recall gains where query and answer vocabularies diverge—NQ Recall@100 jumps from 0.32 to 0.47. The score-fusion variant, SESF, beats standard cross-encoder score fusion by 2.5% on Recall@100 and outperforms SPLADEv2 and ColBERTv2 by 5.3% and 4.6% on nDCG@10, all while leaving the underlying BM25 index completely untouched.