HunFlair

HunFlair performs biomedical named entity recognition to extract and categorize mentions of biomedical entities from scientific text.

Key Features:

• Integration with Flair: Integrates into the Flair NLP framework.
• Comprehensive entity recognition: Recognizes five distinct biomedical entity types and handles variations in text genre and style.
• High accuracy and robustness: Matches or exceeds state-of-the-art performance on multiple evaluation corpora and shows an average improvement of 7.26 percentage points over other off-the-shelf biomedical NER tools in cross-corpus evaluation.
• Cross-corpus training and pretraining: Trained using a character-level language model that was pretrained on approximately 24 million biomedical abstracts and three million full texts to mitigate corpus-specific biases.
• Harmonized corpora: Provides harmonized versions of 23 biomedical NER corpora for comparative evaluation and research.

Scientific Applications:

• Biomedical text mining: Extraction and categorization of biomedical entities from literature for text-mining workflows.
• Drug discovery: Supporting extraction of entity mentions from biomedical literature relevant to drug discovery research.
• Disease gene identification: Enabling identification of gene and disease mentions for studies of disease–gene relationships.
• Corpus development and benchmarking: Facilitating corpus harmonization and benchmarking using the provided harmonized corpora.

Methodology:

Training uses a character-level language model pretrained on ~24 million biomedical abstracts and ~3 million full texts and is applied within the Flair NLP framework with evaluation reported in cross-corpus settings.