Research On News Event Extraction Based On Joint Model

With the advent of the big data era,the increase in information has made information processing more difficult.Therefore,it has become important to automatically extract human-perceived information from large amounts of text.Event extraction technology has been widely used in fields such as news reporting,sentiment analysis,information recommendation,and financial analysis.Excellent event extraction algorithms can improve work efficiency,promote the development of practical application scenarios,and help improve the accuracy and efficiency of natural language processing.At present,there are three main problems with event extraction methods:first,the performance is poor when dealing with long texts,which usually requires the construction of large neural network models,resulting in increased training costs;second,in news texts,an entity usually plays different roles in different events,and the existing event extraction methods cannot accurately identify the classification of entities;third,the construction of event extraction models usually adopts a pipeline Third,a pipeline approach is often used to construct event extraction models,which brings about problems such as error propagation and thus reduces the model performance.To address the above problems,this paper proposes a joint model-based event extraction method,including a trigger word assistance module and an event parameter detection module,which constructs dependency relationships by establishing correlations between trigger words and event parameters,and uses the idea of joint learning to achieve multi-task learning and overall optimization.The main contents of this study include:First,a Fast-Transformer-based module for encoding semantic information of events is proposed.First,word-level and word-level encoding is performed using pre-trained language models and lexical annotation tools to retain more semantic information of the original text.To obtain the global semantic features of the text,the text is encoded with semantic information using the Transformer designed based on the self-attentive mechanism.Meanwhile,in order to solve the problem of high computational complexity of the self-attentive mechanism,this paper optimizes the self-attentive computational mechanism and proposes Fast-Transformer.by changing the computational process,the time and space complexity is reduced fromO(n~2)to O(n),compared with the traditional Transformer model,this module not only solves the long-distance dependency problem,but also significantly reduces the training cost.Second,a sequence labeling-based event extraction scheme is proposed.The event extraction task is transformed into a sequence labeling task by combining the event type labels and event parameter labels in text data into one entity label and using a sequence labeling model to recognize them as a whole.To address the problem that entities play different roles in different events,this paper proposes the use of a multi-layer label pointer network for sequence annotation.The pointer network can identify the starting position of an entity in a sentence,and superimpose multiple pointer networks to classify the role of the entity,and finally achieve the extraction of trigger words and event parameters.Finally,a joint model of event extraction is constructed.In this paper,we construct a dependency relationship between two subtasks,event type detection and event parameter detection,and train the model as a whole using the idea of joint learning.Compared with the traditional pipelined model,this method can effectively alleviate the error propagation problem and thus improve the performance of the model.The experimental results show that the model performs superiorly on public datasets,which fully demonstrates its effectiveness.