Chinese Biomedical Entity Relation Extraction Based On Directed Graph Convolutional Network

In the field of biomedical text mining, biomedical entity relation extraction is the core task to assist researchers in completing the text mining. However, the special characteristics of medical literature, such as long and complex syntax of medical text and a large number of overlapping relations, pose a challenge for biomedical entity relation extraction. In this paper, we propose and apply a Directed Graph Convolutional Network (D-GCN) to encode syntactic information based on the existing neural model as a backbone, thus enhancing the representational ability of the input sequence. Experiments on the evaluation data set showed that the framework could effectively enhance the baseline models based on LSTM, Transformer and pre-trained model BERT at three different levels. Compared with the baseline model PRGC, our model improved F1-score by 1.4% on average.