Adversarial Transfer Learning for Biomedical Named Entity Recognition.

Biomedical Named Entity Recognition (BioNER) is one of the basic tasks of biomedical text mining. In reality, the labeled biomedical data is relatively limited, there is a lack of large enough training data to train a strong model, and manual labeling is expensive. To solve this problem, this paper proposes a network model based on deep transfer learning to improve the performance of entity recognition by learning text knowledge in the general domain (source resource) and migrating to the biomedical domain (target resource). In addition, in order to solve the problem of model training bias caused by the imbalance of data volume between the two domains and the large difference between the data, we construct an adversarial neural network model to extract domain-independent features to effectively alleviate the problem of negative migration. Without adding any artificial features, the proposed model is able to learn transferable feature representations better than existing methods and achieve better results on two biomedical field datasets.