Deep learning with word embeddings improves biomedical named entity recognition.

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Bioinformatics

Background:

• Text mining is crucial for biomedical research, with Named Entity Recognition (NER) as a fundamental task.
• Current NER methods are often entity-specific, requiring costly development and limiting generalizability.
• Existing approaches rely on hand-crafted features that are difficult to optimize and transfer across datasets.

Purpose of the Study:

• To develop and evaluate a generic, deep learning-based method for biomedical NER.
• To demonstrate the superiority of this new method over existing state-of-the-art entity-specific tools.
• To assess the performance across diverse biomedical entity types and datasets.

Main Methods:

• Implemented a long short-term memory network-conditional random field (LSTM-CRF) model.
• Utilized statistical word embeddings for feature representation.
• Compared LSTM-CRF performance against entity-specific NER tools and an entity-agnostic CRF on 33 datasets covering five entity classes.

Main Results:

• The LSTM-CRF model achieved an average F1-score 5% higher than baseline methods.
• Performance gains were primarily driven by a significant increase in recall.
• The generic LSTM-CRF approach outperformed specialized NER tools, often by a substantial margin.

Conclusions:

• A generic deep learning approach (LSTM-CRF) can effectively perform biomedical NER without entity-specific tuning.
• LSTM-CRF offers a more efficient and broadly applicable solution for extracting biomedical entities from text.
• This method has the potential to advance biomedical text mining by improving the accuracy and recall of entity recognition.