CeQe: Grounding Lexical Retrieval in Semantic Evidence
Adam Kahirov, Umesh Deshpande, Swaminathan Sundararaman
cs.IR, cs.AI
2026-08-01
CE-QE appends a cross-encoder's decisive tokens to BM25 queries to recover documents that phrase answers differently; NQ Recall@100 rises from 0.32 to 0.47, with the index untouched.
Lexical retrieval like BM25 matches exact keywords, so it misses answers phrased differently. This is the vocabulary gap: when a relevant document uses different terms from the query, BM25 never puts it in the candidate set. Reranking and fusion can only reorder documents already retrieved, so they cannot recover what was never pulled in. The recall ceiling is locked at the first stage.
This is a real problem. Hybrid retrieval (lexical plus semantic) is standard, but once the lexical arm misses a document, no amount of downstream weighting brings it back.
CE-QE (Cross-Encoder Query Expansion) borrows semantic-retrieval results to expand the lexical query, but with justification. The pipeline:
The design choices have explicit reasons. Expansion seeds come from semantic results, not BM25's own top results, because the latter is exactly what classical pseudo-relevance feedback (RM3) does, and it reinforces BM25's possibly-wrong results, drifting the query. Terms are chosen by cross-encoder attribution, not frequency, because frequency cannot tell generic terms from decisive ones; attribution shows which tokens the model actually treated as evidence for relevance. Every expansion term is copied verbatim from a retrieved passage, unlike HyDE or Query2doc, which have a large model generate text from parametric knowledge and can introduce vocabulary the corpus does not contain. The only added cost is attribution extraction on a cross-encoder the hybrid pipeline already runs for reranking.
SESF (Semantically Enriched Score Fusion) is the fusion variant: it fuses CE-QE-enriched BM25 results with semantic scores, then reranks.
On seven BEIR datasets, CE-QE lifts lexical recall, and the biggest gains land exactly where query and answer vocabulary diverge.
| Dataset | Recall@100 (BM25, then with CE-QE) |
| NQ | 0.32 to 0.47 |
| TREC-COVID | 0.56 to 0.67 |
| Climate-FEVER | 0.19 to 0.26 |
| FEVER | 0.71 to 0.70 (roughly flat) |
At the fusion level, SESF beats cross-encoder score fusion by 2.5% on Recall@100, and beats SPLADEv2 and ColBERTv2 by 5.3% and 4.6% on nDCG@10. The latency cost: RRF about 154 ms, cross-encoder score fusion about 400 ms, full SESF about 557 ms.
The authors also flag an overlooked cost structure: reranking is the real expense. Unbounded reranking pushes per-query time from tens of milliseconds to 25 to 86 seconds, a 485 to 2051x slowdown, worst on TREC-COVID. CE-QE's value is raising the recall ceiling at the source, which reranking cannot fix.
For retrieval engineers, this is a cheap, index-agnostic way to raise the recall ceiling of hybrid retrieval. The BM25 index stays untouched; expansion is just an ordinary query input that any lexical index accepts. Recall is the most valuable stage, because no reranker, however strong, can rank a document it never retrieved.
The flat result on FEVER also shows where to use it: when query and answer already share vocabulary, expansion is wasted money. A cheap signal like lexical-semantic agreement should gate whether to expand.
The authors concede several. It is tested only on flat BM25 up to 8.8M documents; billion-chunk scale and composition with hierarchical indexes are not measured. Only two rerankers are studied, and attribution quality bounds expansion quality, so a poorly calibrated cross-encoder picks less discriminative terms. A gating mechanism to decide when expansion pays off is suggested but not evaluated.
One gap: the NQ jump from 0.32 to 0.47 is striking, but the paper gives no number for how much recall gain actually converts to end-to-end QA accuracy. The link from recall to downstream task is left open here.