English Drug Name Entity Recognition Method Based on Attention Mechanism BiLSTM-CRF

摘要

Drug Name Entity Recognition (DNER) in biomedical literature is the basis of drug information extraction. Existing drug name entity recognition methods depend heavily on hand-designed features and domain knowledge. To avoid tedious features engineering, an attention based BiLSTM-CRF English drug name entity recognition method is proposed. Firstly, in BiLSTM, Words are learned at character level vectors and Word level vectors are obtained, Which are joined together and input into the attention mechanism. Then, the vectors containing the semantic and syntactic roles of Words are learned by LSTM, truncated and input into the attention mechanism. Finally, the attention mechanism dynamically determines the Weights of the two inputs through matrix. The Weights obtained are put into the CRF layer for joint decoding to do sentence classification and prediction. Experiments on the corpus of 2013DDI Extraction show that the F1 value of the model is 80.06%.