Current knowledge base construction and semantic mining are often separated and it results in poor enough semantic understanding depth and weak interpretability in cross-language scenarios. This paper builds a knowledge base based on heterogeneous information network representation learning: it aggregates heterogeneous information of entities and relations extracted from multi-source texts via graph attention networks and introduces relation path encoding for improving the accuracy of entity alignment and relation fusion. Furthermore, a multi-task collaborative graph neural network semantic mining algorithm is designed, which combines a relation aware message passing mechanism and external semantic feature injection, while optimizing the relation classification and link prediction tasks at the same time to improve the semantic reasoning ability. In semantic relation mining, the classification accuracy of relation results is 89.4% by the spouse, and the score of link prediction is 0.523, indicating the proposed method is effective in enhancing the depth and interpretability of semantic understanding.
Nan Jiang (Thu,) studied this question.