MAQ-Retrieval: Multi-Aspect Queries Retrieval for Large Language Models

Retrieval-Augmented Generation (RAG) typically relies on single-query retrieval. Various query optimization techniques have been proposed to enhance retrieval by decomposing the single query into multiple queries focusing on different aspects. This query optimization is to supplement the original query to increase the retrieval confidence from diverse aspects. Existing practices individually retrieve relevant texts based on the generated multi-aspect queries, aggregate these texts, and apply reranking methods to select the Top-$K$most relevant texts. Such existing individual retrieval and reranking mechanisms often overlook the connections between queries. Consequently, these methods fail to adequately capture the nuances introduced by distinct aspects in multiple queries, and may amplify retrievalinduced hallucinations when semantically fragmented evidence is presented to large models. To bridge this gap, we introduce a novel Multi-Aspect Queries Retrieval (MAQ-Retrieval) method. It follows query optimization for multi-aspect queries, integrates a query-weighting mechanism, and leverages the concept of convex hull to perform joint retrieval of multi-aspect queries. In particular, the convex region we constructed can include the set of candidate texts that may be overlooked in traditional approaches, ensuring comprehensive coverage of relevant texts while suppressing inconsistent or noisy fragments that may contribute to model hallucinations. Extensive experiments on both multiplechoice and summarization tasks in either English or Chinese demonstrate that MAQ-Retrieval consistently outperforms four baselines, achieving up to an increased accuracy of 15.2% and the best F1 and ROUGE-Lsum scores across all evaluated datasets, confirming its advantages in RAG with multiple queries, highlighting its ability to enhance retrieval accuracy and mitigate hallucination propagation in large models.