A Event Extraction Method of Document-Level Based on the Self-attention Mechanism.

Event extraction is an important task in the field of natural language processing. However, most of the existing event extraction techniques focus on sentence-level extraction, which inevitably ignores the contextual features of sentences and the occurrence of multiple event trigger words in the same sentence. Therefore, this paper mainly uses the multi-head self-attention mechanism to integrate text features from multiple dimensions and levels to achieve the task of event detection at the level of text. First, convolutional neural network combined with dynamic multi-pool strategy is used to extract sentence level features. Secondly, the discourse feature representation of full-text information fusion is obtained by multi-head self-attention mechanism model. Finally, using the classifier function to classify, and then detect the trigger word and category of the event. Experimental results show that the proposed method achieves good results in document-level event extraction.