Design And Implementation Of Event Extraction System Based On Deep Learning

With the development of Internet technology,the main media of information transmission has changed from paper to electronic website,thus a great quantity of natural language text data has been generated on the Internet.With the speedy increase of these data,the information becomes redundant and complex.An automated information extraction technology is required to extract and analyze unstructured texts so that people can quickly obtain relevant information about interesting events.Event extraction,as an important subtask of the information extraction,refers to identifying and classifying events triggers and related parameters of the events from the unstructured texts.The thesis mainly focuses on the research about the triggers extraction in the event extraction task,and realizes the event extraction prototype system,which can automatically recognize and classify the event triggers from the news texts.The main research content of the paper includes the following three aspects:(1)Firstly,the paper proposes an event extraction algorithm based on the weighted-enhanced graph attention neural network.In the existing graph neural network models such as Graph Convolutional Network(GCN)and Graph Attention Network(GAT),etc.,the attention weights that the center node assigned to the neighbor nodes are considered insufficient in the process of using neighbor nodes' features to update the central node's features.For this reason,the paper proposes a novel method for calculating attention weights,which not only considers the correlation between the central node and neighbor nodes,but also takes the weight of the central node in the sentence into account.Besides,to solve the problem of information loss in the Graph Attention Network(GAT),we propose to use the cascaded multi-head attention organization in our model.The experimental results show that the performance of the weighted-enhanced graph attention neural network model is better than that of the other types of graph neural network models used to solve the event extraction task.(2)Then,the paper implements a data augmentation algorithm based on Generative Adversarial Network(GAN)to improve the performance of event extraction tasks.To solve the problem of the complex predefined rules of traditional data augmentation algorithms,we use a simple predefined rule to discover a large amount of potential data,and use a pre-trained classifier to label the potential data to generate unreliable data.Then we use the generative adversarial network to filter these unreliable data sets,and get some reliable data as augmented data,which can solve the problem of a large amount of noisy data in the augmented data.Finally,we use these data to expand the training set and retrain the classifier.The experimental results show that compared with the classifier trained by the original training set,the classifier trained by the expanded training data has stronger generalization ability and get higher performance on event extraction task.(3)Finally,the event extraction system is implemented based on JQuery,Django and other frameworks.The system mainly includes article management module,user management module,authorization management module and event analysis module,which realizes the function of identifying and classifying event triggers from the article,and the function of user information and authority management.The test results show that the event extraction system meets the basic functional requirements and achieves the expected goals.