A Document-level Event Causality Identification Method Based On Relational Graph Convolutional Network

Event causality identification is an important research topic in the field of natural language processing.As a kind of important semantic relation,event causalities organize events into an event graph according to causal logic,which can assist humans to make decisions through causal reasoning among events.It can be applied to event prediction,public opinion monitoring and other scenarios.Existing methods for identifying event causalities can be divided into three categories: pattern matching-based methods,statistical machine learning-based methods,and deep learning-based methods.The former two methods are limited by the complexity of feature engineering and the lack of model expressiveness,so the identification performance of implicit causalities is not good.Deep learning-based methods benefit from the powerful expression ability of neural networks,and can often capture the expression patterns of implicit causalities.However,most of these methods focus on the identification of intra-sentence causality with one-cause-one-effect or short-distance cross-sentence causality between adjacent sentences,which are difficult to apply to more complex scenarios,such as multiple causality and long-distance cross-sentence causality identification.In view of the shortcomings of the existing methods,this thesis proposes a documentlevel event causality identification method based on relational graph convolutional network DocEC(A Document-level Event Causality Identification Method based on Relational Graph Convolutional Network).By extending the sentence-level task to the document-level task,it can deal with the short-distance and long-distance crosssentence causalities while taking into account the intra-sentence causalities,so as to solve the multiple causalities.Specifically,DocEC takes a document as input,by constructing two different heterogeneous document graphs,namely text structure graph and mention relation graph,using relational graph convolutional network to model document graphs and capture global information.The text structure graph mainly captures the structural information including hierarchy,sequence,syntax and so on,while the mention relation graph mainly captures the potential correlational information among mentions.In the experiments,this thesis conducts comparative experiments of causality identification on two different datasets,and the effectiveness of DocEC can be verified by comparing the results of DocEC and baselines.At the same time,to verify the sensitivity of causality identification methods to the direction of causality,this thesis conducts a comparative experiment of causality direction identification.Through this experiment,the effectiveness of DocEC is verified again,which also indicates from the side that the task of causality direction identification is more challenging.In addition to the experimental comparisons with baselines,this thesis also conducts ablation study on each module designed in DocEC and different types of edges in the document graphs.The rationality of DocEC is proved by the results of ablation study,and each part of the model plays a role in different degrees.In the application,this thesis applies DocEC to the real scene,designs and implements a prototype system of news event analysis,which supports several interactive functions to facilitate users to view and analyze news.