Integrating deep learning architectures for enhanced biomedical relation extraction: a pipeline approach.

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Bioinformatics
• Computational Biology

Background:

• Biomedical relation extraction (RE) is crucial for building knowledge bases and accelerating evidence synthesis from scientific literature.
• Existing methods often face challenges in comprehensive extraction and novelty detection.
• Previous work in the BioCreative VIII BioRED Track provided a foundation for this research.

Purpose of the Study:

• To develop an enhanced end-to-end pipeline for biomedical relation extraction (RE) and novelty detection (ND).
• To integrate state-of-the-art deep learning methods and leverage existing datasets for improved performance.
• To address limitations in document-level RE and ND.

Main Methods:

• An integrated pipeline comprising Named Entity Recognition (NER), Entity Linking (EL), RE, and ND.
• Exploration of BERT-based sequence labeling and span classification for NER.
• Development of a Convolutional Neural Network (CNN) for disease/chemical EL, supplemented by PubTator 3.0.
• Adaptation of the BERT-based PURE model for bidirectional and document-level RE and ND.
• Extensive hyperparameter tuning.

Main Results:

• The enhanced pipeline achieved substantial improvements over the previous submission: NER (93.53%, +3.09), EL (83.87%, +9.73), RE (46.18%, +15.67), and ND (38.86%, +14.9).
• BERT-based models for NER, RE, and ND, combined with a hybrid EL approach, yielded the best performance.
• NER and EL models demonstrated high performance, while RE and ND remain challenging at the document level.

Conclusions:

• The proposed enhanced pipeline offers significant improvements for biomedical relation extraction and novelty detection.
• Document-level RE and ND present ongoing challenges requiring further research.
• Future dataset enhancements could lead to more accurate and practical models for biomedical NLP applications.