HUNER: improving biomedical NER with pretraining.

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• Deep neural networks have advanced biomedical named entity recognition (NER).
• Performance gains are limited by small gold standard corpora in the biomedical domain.
• Data sparsity poses a significant challenge for training robust NER models.

Purpose of the Study:

• To evaluate methods for alleviating data sparsity in biomedical NER.
• To improve the performance and robustness of deep neural networks for NER.
• To develop a practical, state-of-the-art NER tool for biomedical applications.

Main Methods:

• Pretraining a deep neural network (LSTM-CRF) followed by fine-tuning on specific corpora.
• Experimentation with supervised and semi-supervised pretraining strategies.
• Development and evaluation of the HUNER standalone NER tool.

Main Results:

• Pretraining yielded an average F1-score increase of ~2 percentage points across 34 corpora.
• Supervised and semi-supervised pretraining provided insights into the precision/recall trade-off.
• HUNER outperformed existing tools (GNormPlus, tmChem) by 5-13 pp on the CRAFT corpus for specific entity types.

Conclusions:

• Pretraining effectively mitigates data sparsity issues in biomedical NER.
• The HUNER tool provides a robust and high-performing solution for biomedical entity recognition.
• Freely available resources facilitate reproducible research and future comparisons in biomedical NER.