Leveraging Perturbation Consistency to Improve Multi-hop Knowledge Base Question Answering

Abstract

Multi-hop knowledge base question answering aims to answer natural language questions through multi-hop relation reasoning in the knowledge base. An important challenge of the task is the lack of labels for reasoning paths, which leads to the possibility to produce correct answers through incorrect paths in the training, and cannot generalize well in testing. Recently research has attempted to handle the challenge by devising reward shaping or introducing additional information to generate supervision signals of intermediate paths. But they required extra expert experience and label information. To address this situation, we propose a novel method under the teacher-student framework, it leverages perturbation consistency to learn intermediate paths. In the teacher network, we construct close data points for intermediate path prediction by applying random perturbations. Inspired by the data smoothing assumption that labels of close data points should be the same, a consistency loss over predictions of constructed data points and original ones is evaluated. The student network is used to answer questions more precisely by leveraging the intermediate distribution learned from the teacher network. Extensive experiments on two benchmark datasets are conducted, and the results have demonstrated the effectiveness of the proposed method.