| In recent years,with the rapid development of the Internet and the continuous advancement of information construction,the amount of text data such as articles and news on the Internet has exploded.Information extraction,public opinion monitoring,intelligent dialogue and other tasks often need to extract event information from a large number of news reports,user dialogues and other text data.Among them,the amount of text data to be processed is usually very large,which may contain millions,tens of millions or even more texts,and these data are usually highly complex,heterogeneous and uncertain.Therefore,processing such natural language text data is an extremely challenging task.This thesis studies trigger extraction and argument extraction in event extraction,and implements an event extraction system based on article text.In order to break through the limitations of traditional event extraction methods,such as inefficiency of data processing,errors in signal transmission,and inability to make full use of external signals,this thesis proposes a prompt-based generative event extraction algorithm model.The model is guided by the predicted event type,and allows the model to automatically use label semantics in the input and output,which can provide a priori knowledge of the event type.The model is guided to focus on specific types of events,so as to guide the learning process of the model and effectively compensate for the problem caused by the sparsity of events.This model decouples trigger word extraction and argument role extraction,which reduces the interference and error transmission between modules.The experimental results show that this improvement significantly improves the precision rate,recall rate and F1 value of event detection.In order to further improve the accuracy of event extraction,this thesis designs and implements a constrained decoding algorithm for event extraction tasks,which adds constraints to the decoding process.This triebased constraint decoding method dynamically selects and deludes a candidate vocabulary according to the current generated state,which can effectively guide the decoding.Finally,the proposed algorithm is added to the prompt-based generative event extraction algorithm model.The experimental results show that the proposed method can more effectively generate event structures that conform to grammar rules and have reasonable semantics.Based on the above work,this thesis implements an event extraction prototype system based on deep learning technology,which includes event extraction,personal information management and other modules.Users import articles or texts into the system,and the system can automatically extract key information from the text content,aiming to identify and classify event triggers and event arguments in the text. |
