Scholarlink

<p>In interpretable multi-hop question answering, in addition to extracting the answer, there is a need to identify the reasoning chain that provides how the answer is extracted based on the supporting evidence. The reasoning chain is provided in only some existing approaches, and they are mainly based on supervised methods that need annotated data. This paper presents an approach for extracting the reasoning chain in multi-hop questions without using supporting evidence supervision. In this paper, texts are converted to triples using a rule-based approach. Based on the triples, a combination of documents with the most triples similar to question triples is selected as supporting documents. Then, an iterative clustering process including three main steps (clustering, filtering, and masking) is presented for supporting sentence identification. The K-Means and Bisecting K-Means algorithms have been used for clustering. The evaluation results on HotpotQA dataset show that the proposed approach has achieved better results on EM of the supporting fact identification than previous methods (even supervised methods). Also, we have better results on Average F1 and EM of supporting fact identification and on EM of supporting documents than other unsupervised approaches. Experimental results on 2WikiMultiHopQA demonstrate that our unsupervised method attains performance competitive with supervised baselines, especially when measured by F1 score.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-hop clustering for reasoning chain extraction in multi-hop question answering

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Maryam Jamehshourani,
  • Afsaneh Fatemi,
  • Mohammad Ali Nematbakhsh

摘要

In interpretable multi-hop question answering, in addition to extracting the answer, there is a need to identify the reasoning chain that provides how the answer is extracted based on the supporting evidence. The reasoning chain is provided in only some existing approaches, and they are mainly based on supervised methods that need annotated data. This paper presents an approach for extracting the reasoning chain in multi-hop questions without using supporting evidence supervision. In this paper, texts are converted to triples using a rule-based approach. Based on the triples, a combination of documents with the most triples similar to question triples is selected as supporting documents. Then, an iterative clustering process including three main steps (clustering, filtering, and masking) is presented for supporting sentence identification. The K-Means and Bisecting K-Means algorithms have been used for clustering. The evaluation results on HotpotQA dataset show that the proposed approach has achieved better results on EM of the supporting fact identification than previous methods (even supervised methods). Also, we have better results on Average F1 and EM of supporting fact identification and on EM of supporting documents than other unsupervised approaches. Experimental results on 2WikiMultiHopQA demonstrate that our unsupervised method attains performance competitive with supervised baselines, especially when measured by F1 score.