We are delighted to announce that one paper from the Intelligent Information Processing Lab (IIPL, Prof. YoungBin Kim) has been accepted to the 2025 ACM International Conference on Information and Knowledge Management (CIKM).

Title: 

Query, Decompose, Compress: Structured Query Expansion for Efficient Multi-Hop Retrieval

Authors:

Jungmin Yun, Youngbin Kim

Abstract:

This paper presents a novel information retrieval framework that leverages Large Language Models for effective query expansion in complex multi-hop scenarios. Existing query expansion methods have shown limitations when processing multi-hop queries, often degrading retrieval performance by introducing irrelevant or noisy information. To overcome this challenge, we propose a method that restructures and refines existing information instead of generating new content. Specifically, the framework consists of two key components: Query Decomposition, which breaks down a complex query into simpler sub-questions, and Query-aware Document Compression, which aggregates and distills relevant information from retrieved candidate documents using a Large Language Model. The overall pipeline of proposed method ensure a balanced contribution from each component, thereby increasing the stability and robustness of the final query embedding. Experimental results demonstrate that, in the context of multi-hop query-based information retrieval, our proposed method—despite utilizing a relatively small Large Language Model—surpasses strong baselines that rely on larger models. This suggests that effective information retrieval can be achieved not only by relying on the massive generative power of Large Language Models but also by sophisticatedly leveraging the reasoning and summarization capabilities of smaller models.