Stepwise relation prediction with dynamic reasoning network for multi-hop knowledge graph question answering

Multi-hop knowledge graph question answering (KGQA) targets at pinpointing the answer entities to a natural language question by reasoning across multiple triples in knowledge graphs (KGs). When faced with multi-hop questions, existing methods take the whole relation paths into consideration, whereas the number of candidate paths grows exponentially with the increasement of path length, resulting in high search space complexity. Meanwhile, due to the complex semantic information, it is important to focus on different parts of the question at different steps. Moreover, previous studies only give the predicted answers but lack a relational path to indicate the reasoning process. To address these challenges, this paper proposes an interpretable neural model for multi-hop KGQA, namely D ynamic R easoning N etwork (DRN). Inspired by human’s hop-by-hop reasoning behavior, DRN employs an interpretable, stepwise reasoning process to predict a relation at each step, all the intermediate relations form a traceable reasoning path. With effectively stepwise path search over KGs, DRN reduces the search space significantly. Furthermore, to facilitate semantic parsing, DRN dynamically updates the representation of relations and questions for each step based on attention mechanism. Extensive experiments conducted over four benchmark datasets from football, movie and general domain well demonstrate the effectiveness of our method.