Script Event Prediction Based On Deep Learning

Script event prediction is not only an important challenge in the field of knowledge reasoning,but also a basic problem in the field of event law analysis.It can play an important role in crisis event early warning,disease and health prediction,story text generation and so on.With the emergence of a large number of network texts and the development of word representation technology,script event prediction has achieved the data and technical basis.However,the existing methods have some problems,such as insufficient event interpretability analysis and poor event prediction effect.How to improve the interpretability and prediction performance of script event prediction is an important challenge.From the perspective of interpretability and accuracy in script event prediction,this paper deeply analyzes the characteristics of event data,and designs the corresponding methods based on the deep learning structure of shallow word embedding and pre training encoder combined with the law of event development.The main research contents and relevant innovations are as follows:(1)Interpretable Script Event Prediction Based on Argument-level Attention and Multi-Level Scores: To solve the problem that the existing event representation methods do not analyze the rationality of events,resulting in the unrealistic prediction results,this paper proposes an Argument-level attention mechanism.This method calculates the attention value between the predicate verb of the event and other event Argument(subject,predicate and object),adjusts the vector value of each element,and obtains the event vector representation with rationality information.On this basis,we construct the event representation method and multi-level scoring module into a model(called Arg-attention).The multi-level scoring module comprehensively considers the candidate events from the similarity of events and the overall characteristics of event sequences.The training stage is supplemented by confrontation training to enhance the training effect.Several experiments show that the designed Argument-level attention mechanism is helpful to the prediction of script events.The intermediate data of it can intuitively explain the rationality information of events.Compared with other methods based on shallow word embedding,the Arg-attention Model constructed in this paper performs better in script event prediction task,and the prediction accuracy is improved to a certain extent.(2)Script Event Prediction Based on Action Sequence and Tail Event Enhancement: To highlight the action information in the event and improve the ability to deal with the turning point of the event,this paper proposes the event text processing method of action sequence and the tail event enhancement module.The event text processing method hides the protagonist information in the event text and uniformly uses ”X” instead to highlight the information of the action sequence,so that the model pays more attention to the coherence of the action of the event rather than the name of the participant.In addition,this paper also designs a tail event enhancement module,which comprehensively considers the possible turning of tail events by adding the transition probability of tail events and candidate events to the score.The prediction model based on the two methods can greatly improve the prediction effect compared with the existing pre training encoder methods,which verifies the positive effect of event text processing and tail event enhancement on script event prediction.