Multiple-level biomedical event trigger recognition with transfer learning.

Area of Science:

• Biomedical informatics
• Computational biology
• Natural language processing

Background:

• Automatic biomedical event extraction aids in understanding biological systems by updating discoveries.
• Identifying event trigger words is crucial for subsequent event extraction steps.
• Current annotated resources primarily focus on molecular-level events, with limited data for multiple biological levels.

Purpose of the Study:

• To apply transfer learning for multi-level event trigger recognition.
• To leverage a source dataset with molecular-level annotations to improve performance on a target domain with sparse annotations and diverse trigger types.

Main Methods:

• Developed a generalized cross-domain neural network transfer learning architecture.
• Utilized the MLEE corpus as the target dataset for multi-level trigger recognition.
• Employed BioNLP'09 and BioNLP'11 Shared Task corpora as source datasets with varying label overlaps.

Main Results:

• The proposed transfer learning approach demonstrated improved recognition of multi-level event triggers.
• Performance gains were observed regardless of the degree of label overlap between source and target domains.
• The method surpassed previously reported results on the MLEE corpus.

Conclusions:

• Transfer learning significantly enhances multi-level event trigger recognition performance compared to traditional methods, especially with overlapping labels.
• The approach's success stems from a novel parameter sharing strategy (vertical sharing over horizontal sharing).
• Improved parameter sharing leads to better model generalization and performance on the target domain.