End-to-End Beam Retrieval for Multi-Hop Question Answering (2308.08973v2)

Abstract: Multi-hop question answering (QA) involves finding multiple relevant passages and step-by-step reasoning to answer complex questions, indicating a retrieve-and-read paradigm. However, previous retrievers were customized for two-hop questions, and most of them were trained separately across different hops, resulting in a lack of supervision over the entire multi-hop retrieval process and leading to poor performance in complicated scenarios beyond two hops. In this work, we introduce Beam Retrieval, an end-to-end beam retrieval framework for multi-hop QA. This approach models the multi-hop retrieval process in an end-to-end manner by jointly optimizing an encoder and two classification heads across all hops. Moreover, Beam Retrieval maintains multiple partial hypotheses of relevant passages at each step, expanding the search space and reducing the risk of missing relevant passages. To establish a complete QA system, we incorporate a supervised reader or a LLM. Experimental results demonstrate that Beam Retrieval achieves a nearly 50% improvement compared with baselines on challenging MuSiQue-Ans, and it also surpasses all previous retrievers on HotpotQA and achieves 99.9% precision on 2WikiMultiHopQA. Providing high-quality context, Beam Retrieval helps our supervised reader achieve new state-of-the-art performance and substantially improves the few-shot QA performance of LLMs.

Introducing Beam Retrieval: Enhancing Multi-Hop Question Answering with End-to-End Passage Retrieval

Overview

Multi-Hop Question Answering (QA) tasks necessitate the identification and reasoning across multiple relevant pieces of information from a corpus to accurately answer a query. This complex challenge has prompted the development of systems that can effectively navigate through passages to retrieve and subsequently utilize the necessary information. In this research, we present "Beam Retrieval", a novel, generalized framework aimed at significantly improving the performance of Multi-Hop QA systems through an innovative end-to-end retrieval approach.

Beam Retrieval Framework

Beam Retrieval differs fundamentally from existing retrievers by employing a beam search strategy, traditionally used in auto-regressive language generation, for the retrieval process. This method maintains multiple hypotheses of relevant passages at each step of the retrieval, thereby broadening the search scope and mitigating the risk of overlooking pertinent information. The framework leverages a joint optimization of an encoder and two classification heads across all hops, refining the selection of passages with regard to the question at hand.

• Beam Search Integration: By applying beam search, Beam Retrieval maintains several partial hypotheses of relevant passages, significantly expanding the search space compared to conventional methods.
• Joint Optimization: A key innovation of Beam Retrieval is its end-to-end training and inference mechanism, which optimizes the encoder and classification heads across all hops, ensuring a coherent and robust retrieval process.
• Enhanced Multi-Hop QA Performance: The application of Beam Retrieval has demonstrated remarkable improvements in Multi-Hop QA tasks, setting new state-of-the-art performances on benchmark datasets such as MuSiQue-Ans, HotpotQA, and 2WikiMultiHopQA.

Empirical Evaluation

Empirical results underline the efficacy of Beam Retrieval across multiple datasets. On the challenging MuSiQue-Ans benchmark, the system achieved an almost 50% improvement in retrieval accuracy compared with baseline methods. Moreover, it outperformed all previous retrievers on HotpotQA and 2WikiMultiHopQA, providing high-quality context that enabled a supervised reader to achieve new state-of-the-art performance, and notably enhanced the question-answering capabilities of a zero-shot GPT-3.5.

Implications and Future Directions

The introduction of Beam Retrieval has several significant implications for the development of Multi-Hop QA systems and potentially for other NLP tasks that involve complex information retrieval and reasoning.

• Generalizability: The framework's design enables its application to questions requiring varied numbers of hops, depicting its adaptability to different complexity levels.
• Reduction in Early-Stage Retrieval Errors: By keeping track of multiple hypotheses, Beam Retrieval lessens the impact of potential early-stage errors, ensuring more reliable information retrieval.
• Integration with LLMs: The remarkable improvements observed with GPT-3.5 suggest that Beam Retrieval can effectively complement the capabilities of LLMs, forwarding the frontier in generative AI and NLP.

Looking ahead, the potential integration of Beam Retrieval with more advanced LLMs and its adaptation for other complex NLP tasks present exciting avenues for future research. Additionally, further optimization of the beam search strategy and investigation into the retrieval of even more nuanced information could amplify the system's capabilities and applicability.

Conclusion

In summary, Beam Retrieval presents a significant advancement in the field of Multi-Hop QA, showcasing the power of integrating beam search into the retrieval process and optimizing the system in an end-to-end manner. Its superior performance across benchmark datasets underscores the effectiveness of this approach, offering promising prospects for future exploration and development in NLP and AI.