Investigating Cross-Domain Binary Relation Classification in Biomedical Natural Language Processing.

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Computational Biology
• Health Informatics

Background:

• Binary relation classification is crucial for extracting knowledge from biomedical texts.
• Existing methods struggle with diverse domains like gene-disease and social determinants of health (SDOH).
• Evaluating transformer-based models in low-data scenarios is essential.

Purpose of the Study:

• To assess the performance of fine-tuned Bidirectional Encoder Representations from Transformers (BERT) and generative Large Language Models (LLMs) for biomedical relation classification.
• To investigate model capabilities in zero-shot and few-shot learning settings.
• To introduce a novel annotated dataset for social and clinical entity relation extraction.

Main Methods:

• Fine-tuning domain-specific BERT models.
• Evaluating generative Large Language Models (LLMs).
• Performance assessment in zero-shot and few-shot scenarios using a new biomedical dataset.

Main Results:

• BERT models, particularly when fine-tuned on domain-specific data, demonstrated strong performance across various biomedical relation classification tasks.
• Generative LLMs showed comparable performance and generalization capabilities to BERT in certain domains.
• Both model types still fall short of human-level performance, highlighting task complexity.

Conclusions:

• Domain-specific fine-tuning significantly impacts the performance of transformer-based models in biomedical NLP.
• While LLMs offer promise, specialized BERT models remain competitive for relation classification.
• High-quality annotated data and domain expertise are critical for advancing biomedical NLP research.