INTENT-PRESERVING FEEDBACK-GROUNDED LARGE LANGUAGE MODEL QUERY EXPANSION TECHNIQUES

Abstract

Large language models can broaden query semantics but may introduce unsupported concepts and semantic drift when expanding short or ambiguous queries. To control this problem, this paper develops a feedback-grounded and intent-preserving query expansion method. The query encoder represents the original query, while top-ranked retrieval fragments provide external evidence for candidate generation. Query similarity, cross-document frequency, retrieval-rank gain, and context completeness are jointly scored, and a query-fidelity factor, relevance threshold, redundancy penalty, and constrained reformulation mechanism regulate semantic injection. Experiments use 18,000 queries from news, medical, e-commerce, and academic retrieval domains with a 7∶2∶1 split. The proposed method achieves P@10, MRR@10, and F1@20 of 0.792, 0.761, and 0.778, exceeding the previously evaluated expansion baselines, including RM3, Doc2Query, direct LLM, feedback-conditioned LLM, and Query2doc. Stage-level ablation shows that feedback grounding, candidate filtering, fidelity constraints, and reformulation contribute cumulative gains, while removing semantic similarity or redundancy control causes the largest degradation. At the optimal fidelity factor of 0.75, semantic consistency reaches 0.893 and semantic drift is reduced by 32.3% relative to direct LLM expansion. Leave-one-domain-out evaluation yields held-out P@10 values from 0.765 to 0.785, supporting transfer to unseen domains. Paired bootstrap tests confirm gains over the prompt-matched feedback control across primary retrieval metrics at p