Biomedical Event Trigger Identification via Multiple Self-attention Mechanisms

Biomedical events represent the interrelationship between various biomedical entities. Extracting biomedical events from biomedical works is a research hotspot. Among all event extraction methods, biomedical event trigger recognition is a crucial stage. However, there is insufficient capability to extract semantic information from complex biological texts in the current research models. It is necessary to design a new model to enhance the extraction ability. Herein, we propose a biomedical event trigger identification model based on a multi-layer self-attention mechanism. Our model applies the currently popular multilevel self-attention structure, which can imitate how humans browse text. By configuring weights for each word, it can better extract the critical information contained in biomedical text. To validate the effectiveness of the self-attention structure, a series of experiments are conducted on the Multi-Level Event Extraction (MLEE) dataset. The as-obtained results uncover that the proposed model has achieved comparable properties in the task of biomedical event trigger identification.

I. Introduction

With the rapidly growing information technology (IT), increasing text data are produced and openly accessible in the biomedical field. Biomedical event contains complex interaction between action and some corresponding entities or action, which exerts a dominant impact on facilitating biomedical investigation. And it is noted that there are numerous functions for biomedical event tasks, including domain-specific semantic search engines [1], pathway curtains [2], and the construction of medical knowledge graphs [3]. Many biomedical tasks have been organized to gain researchers' attraction and promote progress in this field, such as BioNLP2009, BioNLP2011, and BioNLP2013. Figure 1.

Example of two biomedical events.