Joint entity and relation extraction model based on directed-relation GAT oriented to Chinese patent texts

The joint extraction task aims to construct an entity-relation triple comprising two entities and the relation between them. Existing joint models make it difficult to process too many overlapping relations in Chinese patent texts (CPT). This article introduces a joint entity and relation extraction model based on directed-relation graph attention network (DGAT) oriented to CPT to locate this problem. First, word-character tokens are obtained from CPT using BERT as the DGAT model input. Global tokens are expanded using the BiLSTM network to enhance contextual connection from the model input. Second, the DGAT model encodes the global tokens as a fully connected graph whose nodes represent the global tokens and edges denote the relations between global tokens. The edges with directed relation in the fully connected graph are assigned weights by the DGAT model, and other edges are pruned, resulting in a directed-relation-connected graph. Finally, the entity-relation triples are decoded using conditional random fields (CRF) from the directed relation-connected graph. Experimental results show that the proposed model was highly accurate based on the CPT dataset.