MTBC-BioNER: Multi-task Learning Using BioBERT and CharCNN for Biomedical Named Entity Recognition

In recent years, a large amount of biomedical text data has been generated with the rapid development of biomedical technology and computer technology. How to effectively process these data has become an important issue. Biomedical Named Entity Recognition (BioNER) is an important task in biomedical text information processing. It has an important impact on its downstream tasks, such as entity relation extraction, question answering systems, document classification, etc. However, existing works only relying on a single deep learning model are difficult to learn feature representations for biomedical text data. This paper proposes MTBC-BioNER model, which combines multi-task learning methods to train different related tasks at the same time, and combines the contextual word embedding vector generated by Biomedical Bidirectional Encoder Representation from Transformers (BioBERT) with the character embedding vector generated by Character-level Convolution Neural Network (CharCNN) to effectively represent the feature information of biomedical entities. This model regards each data set in 15 biomedical data sets as an independent task, and uses a specific module for different tasks. Through joint training of all data sets, the model can obtain common features among different tasks, so as to improve the universality of the model. Experiments on 15 biomedical datasets show the effectiveness of MTBC-BioNER.