Medical Named Entity Recognition Based on Bidirectional Gated Pyramid Network

The rapid advancement of Internet-based healthcare technologies drives the daily generation of massive medical datasets, which hold substantial value for enhancing clinical decision support systems and facilitating evidence-based real-world medical research. Medical named entity recognition (NER) is important in the aforementioned research topics. In this paper, we propose a novel Bidirectional Gated Pyramid Network (BGPN), which consists of a convolution layer for extracting character-level features, a bidirectional LSTM (BiLSTM) layer for extracting local inter-sentence information, a Transformer layer for extracting long-distance textual information, and a gated fusion layer for dynamically updating the fusion weights of different levels. In addition, we incorporate a conditional random field (CRF), which enables the network to output the optimal prediction sequence of the BIO label. We validate our proposed method on the BC5CDR dataset, and the results show that our model achieves F1 scores of 0.79 and 0.70 for the two classes of named entities, chemical, and disease, with accuracies of 0.91 and 0.71, respectively.