ProofsNavigator：基于引导的可解释知识推理方法

ProofsNavigator: A Bootstrap Base Method for Explainable Knowledge Reasoning

摘要: 让大规模语言模型生成推理步骤有助于构建可解释的知识推理系统。现有的知识推理方法可能生成不可靠并且与目标无关的推理步骤。为解决这一问题，该文提出一种基于引导的逐步推理方法ProofsNavigator。首先，使用通过Beam搜索生成多个候选推理步骤；然后，通过分别对候选推理步骤的有效性以及跟假设的相关性进行验证，挑选高质量的推理步骤；最后，将所选择的推理结论加入知识集中以进行下一轮循环。实验结果显示，该方法在三个难度依次递增的任务上的准确率分别为40.0%、35.6%和7.1%，比先前最优对比方法分别高1.1%、2.3%和0.2%。此外，该方法在标注数据较少的情况下仍能保持较好的性能。

Abstract: Generating reasoning steps with large-scale language models aids in constructing explainable knowledge reasoning systems. Existing methods for knowledge reasoning might produce unreliable and irrelevant reasoning steps. To address this issue, this article introduces a guided, step-by-step reasoning approach named ProofsNavigator. Initially, it generates multiple candidate reasoning steps through Beam search. Then, it selects high-quality reasoning steps by validating the validity of each candidate step and its relevance to the hypothesis. Finally, the selected reasoning conclusions are added to the knowledge set for the next iteration cycle. Experimental results show that this method achieves accuracies of 40.0%, 35.6%, and 7.1% on three tasks of increasing difficulty, respectively, outperforming the previous best methods by 1.1%, 2.3%, and 0.2%. Moreover, this method maintains good performance even with less annotated data.