基于依存句法与多头注意力的生物医学事件抽取

Biomedical Event Extraction Based on Dependency Syntax and Multi-Head Attention

采用流水线方式抽取生物医学事件存在错误传播问题,且子任务之间的联系被忽略了。为了解决这些问题,采用端到端方法,对触发词识别、要素检测和事件评估三个独立的子任务同时建模,以联合的方式进行训练,再引入依存句法信息捕获同一句子中多个事件的关联性,将依存句法树转化为图,然后用基于多头注意力机制的图神经网络来建模图信息。实验结果表明,该方法在三大生物数据集上的事件抽取任务上效果较好。

The pipelined approach to biomedical event extraction has the problem of error propagation, and the associations among these subtasks is ignored. In order to solve these problems, this paper adopts an end-to-end approach to model three independent sub-tasks, namely trigger recognition, argument detection and event evaluation, simultaneously, and conducts training in a joint way. We introduce the dependency information of the sentence to capture the relevance of multiple events in the same sentence.Therefore, the proposed method converts the dependency parse tree into a graph and uses the graph neural network based on multiple heads mechanism to model the graph information. Experimental results show that this method is effective in event extraction tasks on the three major biological data sets.