The Method Of Credit Risk Evaluation Of SMEs Based On Text Feature Extraction And Fusion

SMEs are difficult to obtain credit support from financial institutions due to the high information asymmetry between SMEs and financial institutions.Content-based information,including legal judgment content and the content posted on SMEs’ official website,contains much valuable information related to the SMEs’ credit risk,which can mitigates the information asymmetry between SMEs and financial institutions,and combines with financial information and enterprise-specific information to depict the credit level of SMEs from multiple views,like judicial punishment,publicity,performance,and operation.However,the above content-based information is characterized by strong timeliness,complex semantic relationship,and low value density.In addition,the completeness of multi-view information is poor.These bring great challenges to extract effective features from multi-view information and integrate features.Thus,this dissertation studies the method of credit risk evaluation of SMEs integrating multi-view text features.The detailed research contents and contributions are as follows.(1)Proposing a method for extracting features from legal judgments for credit risk evaluation of SMEs.This method can mine effective credit features from legal judgments with strong timeliness,great difference in categories,and different influence degrees,and improve the discriminate performance of credit risk evaluation models.Firstly,a framework is designed to extract features based on the legal judgments,which includes structuring judgments,identifying effective judgements,and mining effective features.Secondly,a feature construction approach is proposed,which considers the differences between the tolerances of SMEs for judgement decisions,and constructs the feature based on the relative value of the annual main business income of enterprises and the amount of judgement when constructing credit features.The dissertation finds that the features extracted from legal judgements significantly effect on the credit risk evaluation of SMEs and improve the accuracy of evaluation results.(2)Proposing a method for extracting features from content-based information posted on the official websites for credit risk evaluation of SMEs.This method can mine effective credit features from official websites content-based information with diverse forms of expression,complex semantic relations and strong dynamics,which can migrate the information asymmetry between SMEs and financial institutions.This method mines twelve credit features from three aspects: information breadth,information depth and information dynamics.Specifically,this method utilizes a word embedding model and a clustering model to cluster the semantic information of text corpus,construct two features of the information breadth by measuring the information richness of content,and construct eight features of the information depth by measuring the degree of detail or the concentration of content in a specific topic.In terms of information dynamics,this method employs a regression model to fit the trend of updating of the dynamic information posted on the official website and construct two features.The dissertation validates that the content-based information posted on the official websites of SMEs has effect on the credit risk evaluation,and the features of information depth and trend of updating can significantly improve the discriminate performance of evaluation models.(3)Proposing a method for constructing SMEs’ credit risk evaluation model integrating multi-view features.This method can construct a SMEs’ credit risk evaluation model with incomplete multi-view credit features and depict the credit level of SMEs from multi-dimensions.For the features in different views,this method utilizes different classification algorithms to build credit risk evaluation sub-models,and effectively fits the complex relationship between multi-view features and SMEs’ credit risk.In addition,the method solves the problem of building a credit risk evaluation model for SMEs with incomplete multi-view features by transferring knowledge between sub models.The dissertation finds that compared with a single view,the credit risk evaluation model for SMEs with multi-view features has better performance,and the model still has good performance when the incompleteness of view is high.Based on above research findings,this dissertation enriches the theory of credit risk evaluation of SMEs,expands the application of non-financial information in SME credit risk evaluation,improves the system of credit risk evaluation index,builds an effective credit risk evaluation model for multi-view features,and improves the accuracy of SME credit risk evaluation.In practice,this dissertation can provide decision-making support for financial institutions,help the SMEs with high credit level to obtain credit support,and promote the SMEs development further.