Llm-based evidence selection improves query expansion effectiveness
EviQE: Evidence Selection for LLM-Based Query Expansion
Information Retrieval
Summary
When computers search for information, they sometimes change the original question to get better answers, but picking which documents to read for these changes is tricky. The authors suggest a way to gather documents found by different methods, choose the best few as evidence, and then expand the query based on that. This method helps the computer find more relevant information than relying on just one source or guessing. Their tests showed that picking evidence smartly matters more than doing many rounds of searching and reformulating.
What this means in practice
- •For search engine developers: Improve search results by selecting diverse relevant documents from multiple retrieval methods for query reformulation.
- •For enterprise knowledge managers: Enhance internal document search by aggregating multiple retrieval perspectives to find better query expansions.
Authors
Hai Son Le, Amin Bigdeli, Shirin Seyedsalehi, Morteza Zihayat, Ebrahim Bagheri
Abstract
LLM-based query expansion increasingly conditions reformulation on documents retrieved from the target corpus, yet most work focuses on how to generate expansions rather than which documents the model should read. We propose EviQE, which aggregates documents retrieved by multiple reformulators, selects a compact evidence set, and uses it for one grounded expansion step. This separates evidence selection from generation and treats reformulators as complementary retrieval perspectives. Across three TREC DL and five BEIR benchmarks, reformulators frequently retrieve distinct relevant documents, so pooled candidates provide higher relevant-document coverage than any individual source. The strongest gains come from relevance-based evidence selection: LLM-Score consistently outperforms direct reformulation, cold-start expansion, and single-source seeded expansion. Additional retrieval-generation rounds provide little benefit once strong conditioning evidence has been selected and can reduce effectiveness.