Improving dictionary-based named entity recognition with deep learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Natural Language Processing

Background:

• Dictionary-based named entity recognition (NER) is crucial for normalizing biomedical terms.
• Current methods for creating blocklists of problematic names are manual, inefficient, and do not scale well.
• Adapting NER to new entity types requires extensive manual curation of dictionaries and blocklists.

Purpose of the Study:

• To develop an automated approach for generating improved blocklists for biomedical NER.
• To enhance the precision of text mining by reducing false positives in entity recognition.
• To create context-aware blocklists for more accurate entity identification across different documents.

Main Methods:

• Generated a large dataset of over 12.5 million text spans by comparing three established biomedical NER systems.
• Developed positive and negative context examples for four entity types (genes, diseases, species, chemicals).
• Trained a Transformer-based model (BioBERT) for entity type classification to identify names requiring blocking.

Main Results:

• The trained BioBERT model achieved a high F1-score of 96.7% for entity type classification.
• The automated method generated a significantly larger blocklist, doubling the previous corpus-wide list.
• Text mining precision increased by approximately 5.5% on average (over 8.5% for chemicals, 7.5% for genes) with a minimal recall drop of 0.6%.

Conclusions:

• Automated generation of context-aware blocklists substantially improves biomedical NER performance.
• The developed method effectively reduces false positives, enhancing precision in biological databases like STRING.
• This approach offers a scalable and efficient solution for maintaining and improving biomedical NER systems.