DTranNER

DTranNER integrates a deep learning-based label-label transition model within a conditional random fields (CRF)-based framework to improve biomedical named entity recognition (BioNER) by modeling dynamic contextual transitions between labels.

Key Features:

• CRF-based framework: Integrates a conditional random fields (CRF)-based framework to produce structured label outputs.
• Deep label-label transition model: Uses a deep learning-based label-label transition model to represent dynamic transitions between adjacent labels.
• Unary-Network: Unary-Network models input data to determine individual label assignments.
• Pairwise-Network: Pairwise-Network models contextual information to describe dynamic transitions between labels.
• Adaptive contextual relations: Adaptively explores fine-grained contextual relations between adjacent labels to enhance segmentation of biomedical entities.
• Benchmark performance: Evaluated on five BioNER corpora with F1 scores of 84.56% on BC2GM, 91.99% on BC4CHEMD, 94.16% on the chemical NER dataset, 87.22% on BC5CDR disease NER, and 88.62% on NCBI-Disease.
• Improved precision and recall: Dynamic transition modeling provides contextual clues that increase precision and recall in biomedical named entity recognition.

Scientific Applications:

• Biomedical named entity recognition (BioNER): Recognizing and segmenting entities such as gene mentions, chemicals/drugs, and diseases in biomedical text.
• Biomedical literature mining: Facilitating extraction and analysis of complex biomedical information from literature.
• Benchmarking and evaluation: Providing performance metrics for gene, chemical, and disease NER on standard corpora (BC2GM, BC4CHEMD, chemical NER dataset, BC5CDR, NCBI-Disease).

Methodology:

Integrates a deep learning-based label-label transition model within a CRF-based framework using two networks—a Unary-Network for individual label assignment modeling and a Pairwise-Network for dynamic pairwise transition modeling—and is evaluated on five benchmark BioNER corpora.