An attention-based BiLSTM-CRF approach to document-level chemical named entity recognition

Area of Science:

• Biomedical research
• Bioinformatics
• Natural Language Processing

Background:

• Chemical named entity recognition (NER) is crucial for biomedical information extraction.
• Traditional machine learning methods for chemical NER rely heavily on feature engineering and suffer from tagging inconsistencies.
• Sentence-level analysis limits the contextual understanding in chemical NER.

Purpose of the Study:

• To develop a novel neural network approach for document-level chemical NER.
• To address the tagging inconsistency problem inherent in sentence-level methods.
• To improve the performance of chemical NER with reduced feature engineering.

Main Methods:

• Proposed an attention-based bidirectional Long Short-Term Memory with a conditional random field layer (Att-BiLSTM-CRF) model.
• Leveraged document-level global information via an attention mechanism for enhanced tagging consistency.
• Applied the model to established biomedical corpora for evaluation.

Main Results:

• Achieved superior performance compared to state-of-the-art methods on the BioCreative IV CHEMDNER and BioCreative V CDR datasets.
• Attained F-scores of 91.14% and 92.57% on the respective corpora.
• Demonstrated the effectiveness of the attention mechanism in enforcing consistent tagging across documents.

Conclusions:

• The proposed Att-BiLSTM-CRF model significantly advances document-level chemical NER.
• The approach offers a robust and efficient alternative to traditional feature-dependent methods.
• The model's ability to utilize document-level context enhances accuracy and consistency in identifying chemical entities.