Multi-layer Sequence Labeling-based Joint Biomedical Event Extraction

TL;DR

This paper tackles biomedical event extraction (BEE) by proposing MLSL, a multi-layer sequence labeling framework that jointly learns trigger word recognition and argument roles without relying on external prior knowledge. The model first identifies trigger words with a trigger layer, then merges trigger information via a merging layer before predicting theme and cause arguments, enabling end-to-end event assembly. MLSL achieves superior $F1$ scores on GE11 and GE13 compared with state-of-the-art baselines, largely due to explicit interaction with candidate triggers and a self-attention-based merging mechanism. The work contributes a simple, data-driven approach that reduces error propagation from pipelines and complex joint structures, with future directions including data augmentation and stronger generalization to handle imbalanced and rare event types.

Abstract

In recent years, biomedical event extraction has been dominated by complicated pipeline and joint methods, which need to be simplified. In addition, existing work has not effectively utilized trigger word information explicitly. Hence, we propose MLSL, a method based on multi-layer sequence labeling for joint biomedical event extraction. MLSL does not introduce prior knowledge and complex structures. Moreover, it explicitly incorporates the information of candidate trigger words into the sequence labeling to learn the interaction relationships between trigger words and argument roles. Based on this, MLSL can learn well with just a simple workflow. Extensive experimentation demonstrates the superiority of MLSL in terms of extraction performance compared to other state-of-the-art methods.

Figures (3)

• Figure 1: Overall workflow of biomedical event extraction.
• Figure 2: Multi-layer sequence labeling-based joint biomedical event extraction model
• Figure 3: An example of labeling schema