Medical Named Entity Recognition Using Weakly Supervised Learning

Electronic medical record named entity recognition can extract important clinical information from unstructured text, which is helpful for clinical diagnosis and medical decision-making. However, due to the particularity of the medical field, it is difficult for researchers to obtain sufficient labeled electronic medical records. Models trained using traditional supervised learning methods with insufficient data are not promising. To solve this problem, this paper proposes two weakly supervised learning methods, sampling-based active learning and parameter-based transfer learning, to achieve better performance. In sampling-based active learning, two uncertainty sampling strategies, least confidence sampling and entropy sampling, are used to select data from unlabeled dataset for retraining. In parameter-based transfer learning, the parameters of word representation layer and encoding layer in the source domain are initialized to the corresponding layer of the target domain, and the objective is to learn generalized linguistic knowledge from the source domain. Finally, we use a voting mechanism to ensemble these individual models to get better prediction results. Experiment on the CCKS2017 official test set shows that our system for MER achieves 0.8972 F1 score and gets better performance than the supervised methods, which obtains 0.8921 F1 score and proves the effectiveness of our approaches. The experimental results show that the weakly supervised learning methods proposed in this paper achieve the satisfactory performance as the supervised methods under comparable conditions.