Temporal Causality Discovery With Graph Attention Network

Extracting causal relation between entities is the basis for acquiring causal knowledge and a fundamental step in natural language understanding(NLU)tasks.Causality is often accompanied by temporal cues of entities: the causality of an entity is always a change in state or a change in behavior that is accompanied by a change in time.As far as we know,temporal causality extraction has not received enough attention in line with its importance,with few studies on temporal causality,and even fewer data sets with both temporal and causal relationships.In this paper,aiming at the existing mainstream methods,including sequence tagging,relation extraction and text classification,and according to the research task requirements based on temporal causal relation extraction,the method of sequence tagging combined with temporal relation extraction of causal entities in text is determined.No additional causal knowledge background and feature engineering is required.The main contributions of this dissertation are as follows:1.In terms of essential concepts and study methods,based on the existing causal relation extraction research methods,list the current research status at home and abroad and make a detailed analysis and summary,and combine the time relationship research applied in the direction of causal extraction,Domain research that educates readers about causal directions;2.On the algorithm model,the concept of graph attention network based on temporal relationship(T-GAT)is used,that is,the graph attention mechanism is applied to temporal relationship;the graph attention network based on causal knowledge graph(C-GAT)is used.The concept of,that is,using the causal knowledge graph to extract the adjacency matrix used by GAT;an Equilibrium mechanism is proposed to balance the outputs of the two modules and balance the negative impact of the temporal relationship.Combined with the knowledge of deep learning,a TC-GAT(BERT+T-GAT+CGAT+Equilibrium)causal relation extraction model is proposed to predict the causal label corresponding to each word in the sentence.3.In terms of labeling methods,the type of time relationship is first introduced,and the type of time labeling is determined according to actual needs.After multi-angle analysis of the existing causal sequence labeling methods,it is determined to use the method of ”core causal words” to label time and causal relationship,to reduce the difficulty of labeling;4.On the experimental data,a comprehensive analysis and summary of the existing causal relationship data sets are carried out,and their labeling methods and labeling criteria are listed;according to the needs of the experiment,the Sem Eval2010-Task8 and Altlex data sets are finally selected to correct and expand them,and label the causal relationship and time relationship,and create new time causal data sets TC-Sem Eval2010-Task8 and TC-Altlex;5.On the experimental results,by conducting experiments on the temporal causality data sets TC-Sem Eval2010-Task8 and TC-Altlex and contrasted with the traditional baseline models,the experimental results show that the performance of our model on the causal relation extraction task has been significantly improved,which proves that the temporal relationship has a significant improvement in the extraction of causal relationships.Finally,the model proposed in this paper further comprehensively explores its expansibility with temporal causality.