Document-level attention-based BiLSTM-CRF incorporating disease dictionary for disease named entity recognition.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Disease Named Entity Recognition (NER) is crucial for biomedical research.
• Challenges include identifying rare/complex diseases and handling tagging inconsistencies.

Purpose of the Study:

• To introduce a novel neural network method, Dic-Att-BiLSTM-CRF (DABLC), for improved disease NER.
• To leverage external disease dictionaries and deep learning for enhanced accuracy.

Main Methods:

• DABLC utilizes exact string matching with a dictionary derived from the Disease Ontology.
• It incorporates a dictionary attention layer combined with document-level attention.
• A bidirectional long short-term memory network and conditional random field (BiLSTM-CRF) architecture is employed.

Main Results:

• DABLC achieved superior performance on the NCBI disease and BioCreative V CDR corpora.
• Highest F1 scores were obtained: 0.886 on NCBI and 0.883 on BioCreative V CDR.
• The model demonstrated state-of-the-art results, outperforming existing methods.

Conclusions:

• DABLC effectively combines dictionary resources with deep attention networks for disease NER.
• The method improves the identification of rare/complex diseases and mitigates tagging inconsistency.
• Future work may focus on specialized NER and deep learning for long sentences.