Research on Multi-hop Knowledge Graph Reasoning

Knowledge graph reasoning is a key task in the field of natural language processing, aiming to infer missing entities or relations through existing information in the graph. Multi-hop knowledge graph reasoning has been one of the hot topics in knowledge graph research in recent years. However, due to the inherent incompleteness of knowledge graphs, effectively completing knowledge graphs has become a challenge. Multi-hop knowledge graph reasoning provides an interpretable solution by predicting missing links through exploring paths between entities. Most previous works have used reinforcement learning (RL) based methods to learn paths towards target entities. However, these methods may lack the ability to capture long-distance dependencies between entities, and they may fail to infer a path when there are missing edges along the path. To address these issues, we propose a multi-hop reasoning model based on sequence-to-sequence Transformer model named MHKGR. We introduce two modeling improvements to address the above problems: (1) we introduce a position-aware attention mechanism, which allows the model to flexibly focus on entities on different paths during the reasoning process, thereby improving the accuracy of multi-hop reasoning paths; (2) we optimize the output distribution of the model when generating reasoning paths by introducing a global response normalization module, which reduces noise and interference from irrelevant paths, enhancing the model’s ability to identify correct reasoning paths. We conducted extensive experiments on multiple standard and sparse knowledge graph datasets, and the results show that our method outperforms previous methods significantly.