| Event extraction can help people quickly extract important information from massive data,and has a very important position in the field of natural language processing.To solve the problem that traditional event extraction models cannot span multiple sentences,researchers have proposed document-level event extraction.Existing document-level event extraction models usually start from whole articles with sequence continuity and single input,which leads to the model easily ignoring the event information contained within a single sentence and reduces the accuracy of event meta-extraction.To address the above issues,we proposes to fuse two levels of knowledge extracted from the input corpus,sentence-level and document-level,to enhance the hierarchical context representation capability of the model.And using a key-value memory network combined with a multi-tagged entity-oriented attention mechanism,our model can fully utilize the global contextual information in documents to ensure the completeness and accuracy of argument extraction.The main research of this paper is as follows:(1)A document-level event theory meta-extraction method based on hierarchical knowledge and key-value memory networks is proposed.The model is divided into two levels: sentence-level and document-level.At the sentence-level level,an end-to-end neural sequence model is used to obtain the contextual semantic representation of sentences to ensure the model’s attention to event information in individual sentences;at the document-level,a key-value memory network is introduced to prevent the model from forgetting prior knowledge by reading the key-value memory cache.Finally,the contextual information of the two levels is dynamically fused and the event theory elements in the documents are extracted.(2)A document-level event theorem element extraction method based on a multi-tagged entity-oriented attention mechanism is proposed.First,based on the document-level event meta-extraction model based on hierarchical knowledge,a Multi-Token Entities classifier is introduced to label the entities in a sentence into two types: single-token and multi-token.Secondly,sentences with Multi-Token Entities are fed into the document-level encoder,and the importance of different entities in the document is distinguished using a Multi-Token Entity-oriented attention mechanism.Finally,the knowledge of the two levels is fused using a gating mechanism and the event theory meta-extraction is completed by a sequence annotation method.To verify the effectiveness of the model in this paper,we use the MUC-4 dataset as the experimental data,compare it with the existing baseline model of document-level event-theoretic meta-extraction,and explore the effects of key-value memory network and multi-tagged entity-oriented attention mechanism on the model effect through ablation experiments.The experimental results all meet the expectations and prove the rationality of the method in this paper. |
