Biomedical Named Entity Recognition Based on Self-supervised Deep Belief Network

Named entity recognition is a fundamental and crucial issue of biomedical data mining. For effectively solving this issue, we propose a novel approach based on Deep belief network (DBN). We select nine entity features, and construct feature vector mapping tables by the recognition contribution degree of different values of them. Using the mapping tables, we transform words in biomedical texts to feature vectors. The DBN will identify entities by reducing dimensions of vector data. The extensive experimental results reveal that the novel approach has achieved excellent recognition performance, with 69.96% maximum value ofF-measure on GENIA 3.02 testing corpus. We propose a self-supervised DBN, which can decide whether to add supervised fine-tuning or not according to the recognition performance of each layer, can overcome the errors propagation problem, while the complexity of model is limited. Test analysis shows that the new DBN improves recognition performance, theF-measure increases to 72.12%.