Multi-Path Reasoning for Multi-Hop Question Answering over Knowledge Graph

Multi-hop question answering over knowledge graph (KGQA) aims to find the answer entities that are multiple hops away from the entities in the question called seed entities in the knowledge graph. The main methods include rule and template based methods and deep learning based methods. At present, deep learning based methods is in the mainstream, with the advantages of good portability and high utilization of knowledge graph information. A significant challenge is the lack of information on intermediate entities along the reasoning path. However, most deep learning models are unable to learn the correct reasoning path. To address this challenge, we propose a multi-path reasoning model, which selects the correct reasoning path by constraining the consistency of multiple paths from the seed entity to the answer entity. Then, a teacher-student network is adopted for model compression, where the teacher model relies on the proposed multi-path reasoning model. To demonstrate our model's effectiveness on the KGQA task, we compared our model with four baselines on two benchmark datasets. The experimental results revealed that the Hits@1 values of the model reached 77.8% and 60.2% on WebQuestionsSP and Complex WebQuestions 1.1 datasets, respectively.