Research On Joint Extraction Algorithm Of Chinese Entity Relationship In Financial Field Based On Deep Learning

In today’s information age,various types of financial transactions and activities in the financial field produce a large amount of information resources every day.How to quickly and accurately obtain useful knowledge from a large amount of data has become a difficult point in the research of information intelligent processing in the financial field.The entity relationship extraction method can automatically and efficiently extract triple from these unstructured financial information to provide data support for the construction of financial domain knowledge map.Entity relationship extraction methods can be divided into pipeline method and joint entity relationship extraction method.Pipeline method divides entity relationship extraction into two separate tasks,resulting in issues such as lack of correlation between tasks,error propagation,and redundancy of information.However,the joint entity relationship extraction method can handle both entities and relationships in a unified model,overcome the drawbacks of the pipeline method,and become the mainstream learning method.This paper conducts in-depth research on the joint extraction technology of entity relationships in the financial field,and the main research content is as follows:(1)Aiming at the problem of overlapping triples,this paper proposes BERT-FGM,a joint extraction model of entity relationships that integrates FGM(Fast Gradient Method)and pointer annotation.Based on the BERT(Bidirectional Encoder Representations from Transformers)model,this model effectively solves the overlapping triple problem by using a new pointer annotation strategy to model the relationships between pairs of entities as functions.In addition,the generalization performance of the model is improved by incorporating the FGM confrontation training algorithm in the process of training the word vector of the BERT model.In the comparative experiments with other entity-relationship joint extraction models on the public datasets Web NLG and NYT,the F1 value of the BERT-FGM model proposed in this paper reaches 90.7% and 88.3%,which effectively solves the problem of overlapping triple.(2)Aiming at the disadvantage of Chinese BERT model not optimizing financial field and word-granularity segmentation,this paper optimizes the BERT-FGM model based on the financial field pre-training model Fin BERT(BERT for Financial Text Mining).This model uses the whole-word MASK technology to pre-train the financial corpus and vocabulary,learn prior knowledge of the financial field,and overcome the disadvantage of BERT model in Chinese financial text recognition.Compared with the BERT-FGM model,the Fin BERT-FGM model has improved the accuracy and recall rates by 2.8% and 3.4% respectively.Using Python to design and develop a financial field relational triple extraction software based on the Fin BERT-FGM model,realize the visualization of the recognition results and the function of saving the recognition results,and provide data support for the construction of knowledge graphs in the financial field.