Biomedical Named Entity Recognition Based on Multi-task Learning

Under the background of big data era, the literature in the field of biomedicine has increased explosively. Named Entity Recognition (NER) is able to extract key information from large amounts of text quickly and accurately. But the problem of unclear boundary recognition and underutilization of hierarchical information has always existed in the task of entity recognition in the biomedical domain. Based on this, the paper proposes a novel Biomedical Named Entity Recognition (BioNER) model based on multi-task learning that incorporates syntactic dependency information. Syntactic dependency information is extracted through Graph Convolutional Network (GCN) and incorporated into the input sentences. Using a co-attentive mechanism, the input information and the label information encoded by BERT are fused to obtain the interaction matrix. Then, entity recognition is performed through the boundary detection and span classification tasks. The model was experimented on two English datasets, JNLPBA and BC5CDR, as well as a Chinese dataset, CCKS2017. The experimental results reflected the effectiveness of the entity recognition model proposed in this paper.