Research And Implementation Of Document-level Financial Event Extraction System Based On Deep Learning

In the financial field,event is the main carrier of information transmission,and also an important reference object for relevant personnel to carry out investment analysis and asset management.Event extraction is a necessary process to build a financial event graph and perform inference analysis,which aims to extract event types and event arguments related to event topic from unstructured text and finally present them in a structured form.However,the corpus texts in the financial field tend to be in document-level form,which leads to event arguments usually distributed randomly in multiple different sentences of the text,and a text may consist of multiple events,which greatly increases the difficulty of event extraction.The existing document-level event extraction methods based on deep learning have serious error propagation problems,and cannot fully perceive the dependencies between events when facing multi-event text.Therefore,this thesis proposes a reverse inference model for document-level event extraction named RIDEE which provides an effective solution to the above problems based on the two key technologies of reverse joint inference and event dependency perception.With the core technical support of the model trained on the dataset,a document-level financial event extraction system based on deep learning is designed and implemented.Specifically,the main works of this thesis are as follows:(1)This thesis proposes a key technique of reverse joint inference,which defines the basic processing framework of RIDEE model for reverse inference.Firstly,candidate event arguments in the text are extracted and their argument roles are identified in the way of traditional sequence tagging methods,then the semantic representation of candidate event arguments is enhanced by incorporating the context information of the text,and then event arguments of the same event type are combined based on the event argument roles.Finally,the authenticity of combined events is predicted as a whole.Thus,the overall prediction of event types and event arguments can be realized,and the problems of error propagation caused by subtask serialization of existing methods and sequential event argument extraction process based on preceding predicted event argument are solved.(2)This thesis proposes a key technology of event dependent perception.Based on the basic framework of reverse joint inference,an event dependency perception module is added to the model.Firstly,compress the real events obtained from the reverse joint inference module into the event dependency pool,and then concatenate the combined events to be inferenced with the historical events in the event dependency pool,and use the transformer encoder to fully perceive the dependencies between the events,thus improving the extraction performance of the reverse joint inference module for multi-event text.Thus,the problem that existing methods cannot fully perceive the dependencies between events when facing multi-event text is solved.(3)Based on the RIDEE model trained on the financial dataset as the algorithm support,and the framework of Django,Bootstrap,Jquery,and Pytorch as the development technical support,a document-level financial event extraction system based on deep learning is constructed to realize the automatic extraction of single text or batch text,and the extraction results are returned to users in the form of structured text,event graph,or file.