Multi-hop Temporal Knowledge Graph Reasoning over Few-Shot Relations with Novel Method

Abstract

Multi-hop knowledge graph reasoning is a method to predict the target entity via reasoning paths. This method can not only get the effective target entity, but also obtain the interpretable reasoning path. Most of the previous reasoning methods require that every relation should be supported by enough triples, without considering the triples of few-shot relations, and ignored the gain from temporal information in the reasoning process. Based on the above analysis, we propose a model (FS-Path) which is suitable for temporal knowledge graph reasoning over few-shot relations. This model extends the triple representation for knowledge graph to the 4-tuple representation with temporal information, and it can increase the accuracy of reasoning path by the temporal information. In addition, the FS-Path utilizes the meta-learning to study meta-parameters from high-frequency relations, and then uses the meta-parameters to adapt to the tasks of few-shot relations to improve the generalization ability of the model over few-shot relations. Meanwhile, we construct a new reward function which use the LSTM to capture observation results from environment, and then output the relation vector and temporal vector to environment. Experiments show that the model we proposed can actualize the multi-hop reasoning in temporal knowledge graph, and has outstanding effect over few-shot relations.