Joint Event Extraction Based on CNN-BiGRU and Attention Mechanism

Biomedical event extraction is a very challenging task of information extraction, which plays a key role in medical research, disease analysis and other applications. At present, the task of biomedical event extraction mainly consists of two steps: trigger identification and argument classification. Most research methods use a pipelining approach to accomplish two sub-tasks in stages, which leads to cascading errors. Therefore, a joint event extraction method based on CNN-BiGRU and attention mechanism is proposed, which can extract deeper and more comprehensive features more effectively to complete the task. Firstly, the word vector representation obtained by pretraining language model is combined with part-of-speech vector and position vector. Then input them into Convolutional Neural Network (CNN) and Bi-directional Gated Recurrent Unit (BiGRU) respectively to obtain the local and global feature representations of sentences. Finally, the attention mechanism is used to integrate these two feature representations and jointly deal with these two subtasks. Experiments on MLEE data sets show that the proposed method is superior to the previously proposed biological event extraction method and can effectively extract biomedical events.

I. Introduction

In recent years, the number of biomedical literatures has gradually increased. It is the main way for medical researchers to obtain the latest research knowledge from medical literatures, which requires extracting the information they want from numerous literatures. Therefore, biomedical event extraction has become a hot research field of bioin-formation extraction, which aims to extract useful information from unstructured texts and express it in a structured form. Moreover, it can extract more detailed relationships and make up for the shortcomings of binary relationship extraction.