Research On Chinese Event Extraction Method Based On Attention Mechanism

In recent years,the amount of event text on the Chinese Internet has exploded,and these event texts are rich in information.Therefore,how to accurately and quickly extract valuable event information from natural language text on the Internet is an urgent problem to be solved.Event extraction task is divided into two sub-tasks of event detection and event theory element extraction,most of the current researchers use deep learning research methods to complete the event extraction task,based on deep learning research methods have the advantages of strong learning ability,strong portability and wide coverage.However,there are still two problems with this type of method:(1)In the event detection task,the ambiguity of the event trigger word has a negative impact on the experimental results.(2)In the meta-extraction task,the positive role of semantic dependence is ignored.Therefore,this paper presents a Chinese event extraction method based on attention mechanism,and the specific research work and contributions include the following two aspects:1)A Chinese event detection model based on argument-aware attention mechanism(Args-based ATT-Bi GRU)is proposed to fully mine the contextual features of the text and the type features of event arguments,and solve the ambiguity of event trigger words.Firstly,the text is semantically encoded by the Bert pre-training model,and the text word vector with rich semantic information is obtained,and the argument type vector is constructed by corpus annotation;then,the sequence encoding and argument-aware learning of the gated recurrent unit neural network are used.,obtain the context features and argument type features of the candidate trigger words;then,dynamically fuse the two features through the gating mechanism,and then use the self-attention mechanism to learn the internal dependencies between words,and obtain the final feature fusion vector to enhance the Semantic representation of candidate event trigger words;finally,in order to eliminate the noise features generated by training in multiple contexts,a two-channel GAN network is constructed for noise reduction and trigger word classification.The experimental results show that the method achieves better event detection results on the CEC2.0 Chinese corpus.2)A Chinese event argument extraction model(DP-based ATT-Bi GRU)based on semantic dependency attention mechanism is proposed to make full use of the semantic dependency features of texts to improve the extraction effect of event arguments.First,the text is semantically encoded by the Bert pre-training model to obtain text word vectors with rich semantic information,and then spliced with trigger word type vectors to become joint word vectors;then,events are obtained through recurrent neural networks and graph convolutional networks.The contextual features and semantic-dependent features of the arguments;then,by building a semantic-dependent attention network,the two features are fused to obtain the final feature fusion vector to enhance the semantic expression of the candidate event arguments;finally,in order to reduce the multiple The layer semantic dependent attention network training causes the loss of some contextual features.The fused feature vector and the contextual feature vector are spliced,and the event argument is classified by the softmax function.The experimental results show that the method achieves better event argument extraction results on the CEC2.0 Chinese corpus.Figure [24] Table [8] Reference [71]...