📄 Q2EI: Query-to-Entity Inference for Semantic Condensation in Domain-Specific Retrieval

Summary

Retrieval in specialized fields such as medicine and law often fails because users describe symptoms or situations in everyday language, while the relevant documents use expert terminology. Common LLM-based fixes like HyDE and Query2Doc lengthen the query with generated text. The authors show that this extra material dilutes the useful signal and shifts query embeddings away from the target. They instead propose Generative Query Condensation (GQC), which compresses a query rather than enlarging it. Their implementation, Q2EI, prompts an LLM (zero-shot or few-shot, without fine-tuning or a hand-built knowledge base) to infer the domain entity behind a lay query, such as "Botulism", and then builds a compact, entity-focused query around it. On MedQuAD and COLIEE, Q2EI outperforms strong baselines with both sparse and dense retrievers while using far fewer rewriting tokens. The work matters because it offers a cheap, interpretable pre-retrieval step that makes RAG more dependable in professional domains and degrades gracefully when the entity guess is wrong.

Key points

• Headline gains: On COLIEE, mE5 gains 3.35 nDCG@10 over the best baseline. On MedQuAD, BM25 gains 13.97 Recall@10 over the best baseline. BGE-M3 was also evaluated.
• Motivation experiment (mE5, MedQuAD): Appending generated descriptions to queries hurts retrieval even when the text is relevant. Recall@10 falls from 86.88 to 77.94 with 100 added words and to 67.10 with 300 words.
• Correct entity inference: When the inferred entity matches the target, Q2EI few-shot on mE5 reaches 52.11 nDCG@10, compared with 18.34 for HyDE and 22.53 for Q2D.
• Wrong entity inference: When the inferred entity is incorrect, Q2EI still scores slightly better (6.61 vs. 4.82 for Q2D), suggesting that compact rewrites limit error amplification.
• Token cost per query on MedQuAD: HyDE uses about 13,409 tokens and Q2D about 2,768. Q2EI uses about 905 tokens zero-shot (6.8% of HyDE) and about 1,718 few-shot (12.8% of HyDE). Code and data are released on GitHub (xu2023-ICT/Q2EI).