Deep learning for biomedical event extraction

Abstract

Biomedical Event Extraction (BEE) is a crucial task in biomedical natural language processing, aiming to identify molecular events involving genes, proteins, and other biological entities. This thesis presents two approaches to improve BEE: a classification-based model for event trigger detection and a generation-based model for event extraction. First, the BioLSL model enhances event trigger detection by leveraging label-based synergistic representation learning, capturing dependencies between event type labels and trigger words. Experimental results on three benchmark BioNLP datasets demonstrate its state-of-the-art performance, particularly in data-scarce scenarios. Second, the GenBEE model formulates BEE as a sequence generation problem, integrating structured prompts and prefix-based representations to incorporate event semantics and argument dependencies. Besides, the structured prompts and prefix-guided learning further improve model performance by effectively integrating event structure into the generative framework, leading to more accurate and comprehensive event extraction across multiple datasets.