Query Expansion and Verification with Large Language Model for Information Retrieval

Query expansion is a crucial task in information retrieval. Recent research has demonstrated that generative relevance feedback (GRF) based on large language models (LLMs) can significantly enhance the effectiveness of query expansion. However, user queries often suffer from issues such as unclear or incomplete semantics, and LLMs may also introduce hallucinations when assisting with query expansion, thereby reducing retrieval performance. In this paper, we semantically enrich queries by prompting a LLM to generate multiple query-related documents. Additionally, we employ a consistency verification method to remove meaningless information from the expanded results of LLM, mitigating the hallucination issues associated with large model generation. Given that dense retrievers have lower performance in encoding long texts expanded by LLMs, we adopt the BM25 sparse retrieval method in the downstream retrieval module. Our experiments show that the method proposed in this paper exhibits significant improvements compared to both the state-of-the-art sparse retrieval and dense retrieval methods.