Abstract: The introduction of deep neural network technology greatly improves the extraction accuracy of text semantic features of relation classification. The common sense knowledge graph is used to construct the contextual semantics other than the text′s own semantics, and the pre-trained model is used to obtain the contextual semantic features. Aiming at the semantic features of text, context and marked entity, a multiple semantic fusion mechanism is established to realize the relation classification model, which is named MSF-RC. The model is tested on two different datasets, SemEval-2010 task and TARCED. The experimental results show that the introduction of contextual information helps to strengthen the semantic understanding of labeled entities, and the hierarchical fusion of multiple semantics can further improve the performance of relation classification model.

Key words: relation classification, bidirectional encoder representation from Transformer(BERT), knowledge graph, feature fusion, semantic fusion