Research On Event-oriented Event Element Identification And Event Relation Classification

Event is the basic unit of knowledge for human beings to store memory and carry out logical thinking activities,and the related research is gaining more and more attention in philosophy,linguistics,artificial intelligence and other related academic fields.Event-oriented research can provide support and services for professional research and specific applications such as knowledge graph,event analysis,search engine,and recommendation system.In this dissertation,two directions of event element identification and event relation classification in event-oriented research are investigated.And solutions are proposed to address some of the problems existing in previous research and found in the research process,including.(1)Research on event element identification based on bi-dimension attention and dynamic target model.To address the common problems of difficulty in extracting deep semantic features and sample imbalance in sequence annotation tasks such as event element identification.This dissertation propose an event element identification method based on bi-dimension attention and dynamic target model by using bi-dimension attention network to extract features and combining with dynamic target algorithm.First,the pre-trained embedding neural network is used to map the pre-processed samples into a low-dimensional dense matrix.Then the bi-dimension attention network calculates the attention score from two dimensions of matrix rows and columns and extracts the deep semantic features in the text combined with the fully connected layer.Finally,the linear neural network is used to classify the output according to the effective features and perform sequence labeling according to the output results to complete the event element identification.In addition,in the model training stage,the dynamic target algorithm is used to assign weights to the target labels and calculate the loss values so as to update the model,which effectively alleviates the sample imbalance problem.Comparative experiments show that the method in this dissertation can better improve the effect of event element identification.(2)Research on event relation classification based on multi-feature and multi-attention model.To address the problem of traditional relation classification methods rely on manually designed features,which lack deep-level features.This dissertation proposes an event relation classification method based on multi-feature multi-attention model.First,when constructing the feature matrix,three additional features are added: location features,event element features and event context features,and multiple embedded neural networks are used to map and fuse each feature to construct the feature matrix.Then,deep semantic features,inter-event features and context-dependent features are sequentially extracted from the feature matrix by combining bi-dimensional attention networks,dot-product attention networks and bi-directional gated recurrent unit networks.Finally,the linear neural network is used to classify the output,and the relationship category with the highest probability value is used as the classification result to complete the event relationship classification.The experimental results show that by fusing multiple features,this method achieves a more satisfactory classification effect.Figure 14 Table 17 Reference 68...