Research And Implementation Of Inference Method For Financial Knowledge Graph

In the long-term development,financial field accumulated structured data in line with database specification and semi-structured and unstructured data dominated by text,financial statements and relevant news reports,which together constitute financial big data.Building a financial knowledge graph with strong semantic understanding ability can make full use of the value of financial big data,mine the hidden information and provide help for financial business.Although the knowledge graph stores a lot of knowledge in the form of triple,there is also the problem of lack of facts.Therefore,using the existing entities,relationships or fact information in the knowledge graph to infer and predict the new relationship between the entities of the knowledge graph,that is,knowledge inference,is an important part of the knowledge graph research at present.This paper starts from the construction of financial knowledge graph.In order to improve the completeness of the graph,the knowledge inference method for financial knowledge graph is studied.Firstly,according to the professional knowledge,rich content and wide sources of financial data,this paper designs a top-down and bottom-up construction method of financial knowledge graph,which guides the construction of the data layer through the mode layer,and uses the content of the data layer to feed the mode layer to supplement the lack of ontology.Based on this method,this paper constructs a Chinese financial knowledge graph with equity knowledge and the same trade relationship knowledge of financial institutions to support the research on the equity penetration and the same trade relationship in the financial field.Secondly,in this paper,two challenges are noted in the research of knowledge inference algorithm:how to make the algorithm use the context information in large-scale text data while using the structure information of knowledge graph;how to alleviate the problems of low quality and false negative of negative sampling triple in algorithm model training.Under this background,a knowledge inference algorithm combining distributed representation learning and PLM is proposed in this paper,and on this basis,a knowledge inference algorithm introduced GAN is proposed to enhance triple negative sampling.Distributed representation learning can better learn the structural information of knowledge graph,and the PLM can better learn the semantic information in large-scale text data due to its strong feature capture ability.The joint training of the two model is conducive to meeting the first challenge;through the adversarial learning of generator and discriminator in the GAN,the second challenge can also be met.Finally,the adversarial learning model ScKGAN which integrates semantic and structural information was obtained in this paper.The experiment shows that the model has good knowledge inference effect.Based on the algorithm,this paper implements a prototype system which provides users with the functions of graph query and inference result quality evaluation.