Research On The Event Template Construction And Event Extraction Of Chinese Financial Text

Financial texts contain a large amount of event information,which plays an important role in corporate investment decisions and stock predictions,and the extracted events can also lay a solid foundation for downstream tasks such as corporate portraits and matter mapping construction.Therefore,the task of event extraction in finance and economics has become a research direction for many scholars.This thesis focuses on improving 2 shortcomings in the current state of research.First,there is a lack of basic resources for research in this field.Most of the current event extraction is based on generic templates,and there is a lack of specialized event templates in the financial field.Secondly,how to mine more semantic structure information implicit in the text and embed it into the deep learning model.This thesis addresses the current situation that existing event templates are inadequate in the application of finance and economics,and constructs a Chinese financial event template,mainly including six categories of equity changes,corporate dynamics,institutional behaviors,judicial behaviors,macroeconomics,and stock market.The rationality of the template is elaborated from a theoretical perspective,mainly including the research significance of each type of events and the ideas and reasons for setting up the theoretical roles of each type of events.In order to verify whether the argument role in the template exists and is reasonable,this thesis annotates and explains the roles involved in various events in the real financial text.In this thesis,based on the GIT model,we further obtain the rich semantic information in financial texts and embed them into the model,and propose the SMTAF＿GIT model.Firstly,the corpus construction work of the experimental data is introduced;secondly,the general framework of the model and the internal structure of each sub-module are elaborated and introduced;finally,the evaluation process and evaluation indexes are introduced,and the experimental results are presented and analyzed in multiple dimensions.The experimental results show that the SMTAF＿GIT model proposed in this thesis can better learn the interargument relationships,resulting in an overall F1 value that is 1.7 percentage points higher than the F1 value of the GIT model.