Research On Key Technologies Of Financial Donain-oriented Eventic Graph Construction

Traditional knowledge graphs mostly center on attribute and relation knowledge of entities,but ignores the knowledge of evolutionary patterns between events.In order to make up for this deficiency,researchers gradually put forward the concept of eventic graph.The nodes in eventic graph are highly generalized events,and the edges are the evolutionary relationships between events,such as causal relationship and sequential relationship.With the development of natural language processing technology,the accuracy of information extraction has been improved,making it possible to automatically mine eventic knowledge and build the eventic graph from the text.Starting from the financial domain,this paper explores the key technologies of financial domain-oriented eventic graph construction,including the extraction of event causality based on sequence labeling,the commonsense-enhanced event representation learning and the data-driven method for causal strength calculation.The acquisition of event knowledge is the basis of eventic graph construction.This paper explores the acquisition of causal relationship between events.In this paper,event causality extraction is modeled as a sequence labeling task,and a method of causality extraction based on pre-trained model is proposed.To alleviate the problem of insufficient labeled data,this paper further proposes a semi-supervised learning method based on noise model,which uses a large number of unlabeled data to improve the effect of causal extraction.The experimental results on two causal extraction datasets in Chinese and English prove the effectiveness of this method.Event is the core concept of eventic graph.In order to model event semantics better,this paper proposes a commonsense-enhanced event representation learning method,which integrates commonsense information such as intent,sentiment and entity relationship into the learned event representation,to help the construction of event graph and its application in other tasks.The experimental results on event similarity,script event prediction,stock market prediction and other tasks show that our method can model event semantics more precisely and improve the results on downstream tasks.Towards better modeling the causal strength between events,this paper explores the calculation of causal strength based on statistical method and pre-trained models,which automatically learn causal strength information from a large number of causal event pairs.The experimental results on COPA causal inference dataset show that pre-trained models can effectively learn causal knowledge from a large number of causal event pairs,and accurately model the strength of causal relationship.Finally,based on the above researches,this paper designs and implements an eventic graph construction system in financial domain,and constructs a eventic graph that contains millions of events and causal relationships from large-scale financial corpus.The experiment verified the feasibility of eventic graph construction methods proposed in this paper.