Research On Script Event Prediction Based On Event Evolutionary Graph

Given a sequence of events that have occurred,the goal of script event prediction is to infer what happens next.It is essential for many natural language understanding applications,such as text understanding,intent recognition and question answering systems.Script event prediction is a downstream task of event extraction.There are still some problems in this field,such as lack of datasets,incomplete event representation,and insufficient information fusion.The existing approaches are not yet able to solve these problems well.This paper takes script event prediction as the theme,and the main research contents are as follows:Firstly,this paper constructs a Chinese dataset based on Sina News called SinaNews.Currently,script event prediction has only one English dataset called NYT based on the New York Times,which greatly hinders the research on this task.In accordance with the processing flow of the script event prediction task,we construct a Chinese dataset which covers entertainment,economics and military sectors based on the Sina News corpus.The SinaNews dataset is similar to the NYT in scale and has a wide range of applications.Secondly,this paper studies the effect of event representation methods on script event prediction,and proposes a new event representation method.In the script event prediction task,events are represented in the form of four-tuple,i.e.,(predicate,subject,object,preposition object).The existing methods concatenate the word vectors of the four elements as the event representation,which causes the event representation overly relies on the word vectors but ignores the semantic association between words.We propose a Transformer-based event representation method,which uses the Transformer to capture the semantic association between words within the event to obtain a more accurate event representation.Thirdly,this paper proposes a script event prediction method combining event chains and event evolutionary graphs(ECGNet).The orders in event chains and the event evolutionary patterns in constructed event evolutionary graphs are very important for script event prediction.But the existing approaches cannot consider the orders and event evolutionary patterns simultaneously.We propose a neural network that combines event chains and event evolutionary graphs(ECGNet).It fully learns the orders and the event evolutionary patterns,and uses the features to j ointly predict subsequent events,which improves the prediction results.Finally,this paper further proposes a script event prediction method(Gate-ECGNet)that dynamically integrates multiple information.The Gate-ECGNet improves ECGNet's event representation layer and feature fusion method.The event representation layer not only considers the semantic association between words within the event,but integrates the co-occurrence relationship between contextual event verbs.In addition,the GateECGNet model designs a gated attention mechanism to learn the confidence of orders and event evolutionary patterns on different samples,thereby the model can dynamically fuse features to predict subsequent events and further improve the prediction performance.This paper evaluates the ECGNet and Gate-ECGNet models through extensive experiments.The results show that the models proposed in this paper can better fuse multiple information to predict the subsequent event,so they outperform the existing methods on the Chinese and English datasets.The Gate-ECGNet model can dynamically fuse multiple information,thus its prediction ability is better.