Retrieve, Summarize, Plan: Advancing Multi-hop Question Answering with an Iterative Approach
This addresses context overload and planning inefficiencies in multi-hop QA, offering an incremental improvement for industrial applications.
The paper tackled multi-hop question answering by proposing ReSP, an iterative Retrieval-Augmented Generation method with a dual-function summarizer to compress retrieved information, which outperformed state-of-the-art methods on datasets like HotpotQA and 2WikiMultihopQA with improved robustness to context length.
Multi-hop question answering is a challenging task with distinct industrial relevance, and Retrieval-Augmented Generation (RAG) methods based on large language models (LLMs) have become a popular approach to tackle this task. Owing to the potential inability to retrieve all necessary information in a single iteration, a series of iterative RAG methods has been recently developed, showing significant performance improvements. However, existing methods still face two critical challenges: context overload resulting from multiple rounds of retrieval, and over-planning and repetitive planning due to the lack of a recorded retrieval trajectory. In this paper, we propose a novel iterative RAG method called ReSP, equipped with a dual-function summarizer. This summarizer compresses information from retrieved documents, targeting both the overarching question and the current sub-question concurrently. Experimental results on the multi-hop question-answering datasets HotpotQA and 2WikiMultihopQA demonstrate that our method significantly outperforms the state-of-the-art, and exhibits excellent robustness concerning context length.