Recognizing Biomedical Named Entities Based on the Sentence Vector/Twin Word Embeddings Conditioned Bidirectional LSTM

摘要

As a fundamental step in biomedical information extraction tasks, biomedical named entity recognition remains challenging. In recent years, the neural network has been applied on the entity recognition to avoid the complex hand-designed features, which are derived from various linguistic analyses. However, performance of the conventional neural network systems is always limited to exploiting long range dependencies in sentences. In this paper, we mainly adopt the bidirectional recurrent neural network with LSTM unit to identify biomedical entities, in which the twin word embeddings and sentence vector are added to rich input information. Therefore, the complex feature extraction can be skipped. In the testing phase, Viterbi algorithm is also used to filter the illogical label sequences. The experimental results conducted on the BioCreative Ⅱ GM corpus show that our system can achieve an F-score of 88.61 %, which outperforms CRF models using the complex hand-designed features and is 6.74 % higher than RNNs.