FBSN: A hybrid fine-grained neural network for biomedical event trigger identification

Abstract

Biomedical event extraction is one of the fundamental tasks in medical research and disease prevention. Event trigger usually signifies the occurrence of a biomedical event by adopting a word or a phrase. Meanwhile, the task of biomedical event trigger identification is a critical and prerequisite step for biomedical event extraction. The existing methods generally rely on the complex and unobtainable features engineering. To alleviate this problem, we propose a hybrid structure FBSN which consists of Fine-grained Bidirectional Long Short Term Memory (FBi-LSTM) and Support Vector Machine (SVM) to deal with the event trigger identification. The hybrid architecture makes the most of their advantages: FBi-LSTM is to mainly extract the higher level features by the fine-grained representations, and SVM is largely appropriate for small dataset for classifying the results of biomedical event trigger. After that, the popular dataset Multi Level Event Extraction (MLEE) is employed to verify our hybrid structure. Experimental results show that our method is able to achieve the state-of-the-art baseline approaches. Meanwhile, we also discuss the detailed experiments in trigger identification task.