Two Heads Are Better Than One: Improving Search Effectiveness Through LLM Generated Query Variants

Abstract

User query quality impacts retrieval effectiveness. This study categorizes thousands of user queries spanning one hundred topics into three groups – low, medium, and high quality – based on NDCG@10 scores. The study investigates the impact of fusing search results of acLLM-generated query variants with results retrieved from user queries drawn from the three groups, similar to a collaborative search approach where users with diverse queries collaborate in locating relevant information. The findings indicate that a traditional search system can be significantly improved by fusing results for low-quality queries, offering a promising solution for users who struggle to find relevant information, particularly in contexts where advanced search systems are impractical due to technical or resource constraints, or where access to query logs are unavailable.

Type
Publication
Proceedings of the 2025 ACM SIGIR Conference on Human Information Interaction and Retrieval