A hybrid deep-learning approach for complex biochemical named entity recognition

Named entity recognition (NER) of chemicals and drugs is a critical domain information extraction in biochemical research. NER provides support for text mining reactions, including relation extraction, attribute metabolic response relationship extraction. However, the existence complex naming characteristics biomedical field, such as polysemy special characters, make task very challenging. Here, we propose hybrid deep learning approach to improve accuracy NER. Specifically, our applies Bidirectional Encoder Representations from Transformers (BERT) model extract underlying features text, learns representation context through Bi-directional Long Short-Term Memory (BILSTM), incorporates multi-head attention (MHATT) mechanism chapter-level features. In this approach, MHATT aims abbreviations efficiently deal with problem inconsistency full-text labels. Moreover, conditional random field (CRF) used label sequence tags because probabilistic method does not need strict independence assumptions can accommodate arbitrary information. The experimental evaluation on publicly-available dataset shows that proposed achieves best performance; particular, it substantially improves performance recognizing abbreviations, polysemes, low-frequency entities, compared state-of-the-art approaches. For instance, accuracies entities produced by BILSTM-CRF algorithm, those two datasets (MULTIPLE IDENTIFIER) have been increased 80% 21.69%, respectively.