Temporal knowledge graph inference based on attention mechanism and multi-relational events

Temporal knowledge graph inference is to introduce temporal information into the knowledge representation learning inference task, aiming to infer the evolution trend of events in the future. Most temporal knowledge graph inference methods have limited inference capability across temporal entities and relationship. To address the above mentioned problem, a temporal knowledge graph inference method based on multi-relational events and attention mechanisms is proposed. In order to use the correlation information between reasoning tasks and temporal events in the temporal knowledge graph, it is necessary to design a special encoder with high complexity. However, as one of the most commonly used sequence encoders, recurrent neural networks ignore the correlation between sequence nodes and tasks, and are not suitable for knowledge reasoning. In this paper, we use the self-attention mechanism as encoder model to fuse the historical information of the sequence, calculate the attention scalar of the reasoning task and the historical information of the sequence, so as to obtain more accurate historical event correlation information coding. On this basis, the attention mechanism is used to optimize the multi-relationship neighborhood aggregator, and the neighborhood representation of entities is calculated according to the degree of interest of events under different relationships, obtaining more accurate event codes and ultimately more accurate neighborhood vector representation of entities. Experiments on the WIKI and YAGO datasets show that the effect of Attn-Net is improved by 1.5% and 2%, respectively, effectively improving the inference capability of the temporal knowledge graph.