Unrestricted multi-hop reasoning network for interpretable question answering over knowledge graph

Knowledge graphs significantly boost the answer retrieval quality for natural language questions. The knowledge graph based question answering (KGQA) task returns accurate answer entities instead of keyword matches. For the more challenging task of multi-hop KGQA, existing methods either address fixed-length multi-hop reasoning, or perform a delayed detection of termination that requires an extra hop of reasoning. In addition, they suffer from two mapping problems between the question and relations: (1) one-to-many mapping when an individual question word corresponds to multiple hops of reasoning; (2) many-to-one mapping when a single hop of reasoning corresponds to multiple question words. Therefore, in this paper, we address these two issues of delayed determination and mapping problems by proposing an Unrestricted Multi-Hop Reasoning Network for Interpretable KGQA named UMRNet. Specifically, the proposed dynamic update strategy of question embeddings based on our attention redistribution mechanism is capable of handling the mapping problems. Furthermore, to avoid the need for an extra hop of reasoning, we propose a non-delayed termination detection mechanism that performs effective evaluation of the remaining reasoning information based on history attention. Extensive ablation studies and comparative experiments have been conducted on four KGQA benchmark datasets. The results demonstrate that the major modules of UMRNet are effective, and UMRNet outperforms the state-of-the-art methods regarding both accuracy and efficiency.