Improving Zero-Shot Information Retrieval with Mutual Validation of Generative and Pseudo-Relevance Feedback

Information retrieval systems often suffer from poor performance due to brief or ambiguous user queries. A direct and effective method is to expand these queries by incorporating additional information. Traditional Pseudo-Relevance Feedback (PRF) approaches enhance queries by extracting information from the top-k retrieved documents during the initial retrieval, with their effectiveness closely correlated to retrieval quality. Meanwhile, recent studies on Generative-Relevance Feedback (GRF) utilize Large Language Models (LLMs) to capture the latent search intent behind user queries, but may generate corpus-irrelevant contexts due to the inherent issues of hallucination in LLMs. To mitigate the limitations and maintain the advantages of both PRF and GRF, we propose a zero-shot sentence-level mutual validation framework. Specifically, we enrich the generative relevance feedback by synthesizing various prompting strategies. The mutual validation process incorporates a comprehensive scoring mechanism that considers both consistency and relevance dimensions, facilitating alignment between GRF and PRF while retaining query-relevant information. Lastly, our fine-grained dual-filtering approach ensures relevant and reliable sentences for robust query expansion. By conducting extensive analytical experiments on four low-resource datasets, we showcase the effectiveness of our proposed method compared to existing approaches, particularly in resource-constrained settings.