A multi-granularity script event prediction method based on attention and graph networks
[Objective]When given an existing event context,script event prediction(SEP)aims to predict the subsequent counterpart.However,some disaster events,such as typhoons and floods,due to the complex event evolution process,their event contexts endure different degrees of information density,prompting the event reasoning very challenging.In this paper,we propose a multi-granularity SEP method based on attention mechanism and graph network to solve two problems in complex event reasoning,namely,how to merge different levels of event information to obtain deeper semantic information,and how to fully utilize dense event connections to improve the reasoning ability of the model.[Methods]First,we utilize the dependent syntax and semantic role labeling capabilities of the Language Technology Platform to extract event triples from News texts and construct event chains.Next,we employ the self-attention mechanism to obtain event fragments and merge them with the representations of individual events as inputs to the model.In addition,event chains are extracted from large-scale News texts to construct event evolution graphs and obtain transfer probability matrices between events.Based on these event representations and probability matrices,we construct a three-layer graph attention network(GAT)for inference.[Results]The performance of our proposed method is compared with other eight baseline methods on the benchmark dataset New York Times(NYT).To verify the effectiveness of the proposed method on the disaster events,we also construct a real dataset,which contains 20 typhoons and more than 2 800 related News reports.Overall experimental results on the typhoon dataset and benchmark dataset show that the accuracy of the proposed method reaches as high as 88.2％on the typhoon dataset and 54.23％on the NYT dataset,more satisfactory than other statistical and neural network-based methods.Specifically,on typhoon dataset EventComp,PairLSTM,and LSTM models,in which only strong temporal relationships in the event chain are considered,accuracies of 79.32％,78.45％,and 73.46％are obtained,respectively.On NYT,accuracy of 49.57％,50.83％,and 45.53％are obtained,respectively.The graph-based model SGNN performs more satisfactorily on both the typhoon dataset(87.27％)and the NYT dataset(52.45％)than those event-pair-based models do.Moreover,to investigate the intrinsic mechanisms of how the model improves event prediction performance,we conduct a series of ablation experiments on the typhoon dataset.Experimental results show that removing the event fragment representation leads to a significant performance degradation of 2.6 percentage points,thus highlighting the importance of event fragment representation in the model.Similarly,removing the attention mechanism in the event inference layer triggers a 0.4 percentage point performance degradation,likewise demonstrating the critical role of the attention mechanism in extracting valid information.[Conclusions]The proposed multi-granularity event inference method based on the attention mechanism and graph network attains significant improvements compared with those existing methods.Extracting and combining event information at different granularities,we can obtain richer event semantic information.Also,combining graph attention networks and event transition probability matrices,we can fully utilize dense event connections to enhance inference performance.The experimental analysis shows that the event fragment representation and graph-based learning method constitute the two major factors for enhancing prediction performance.The former shows that the event chain contains key correlation information among events,whereas the latter shows the event evolution graph can extract more effective interaction information among events and the event inference task can perform more satisfactorily.
万方数据
维普
