Credit risk detection method and device, electronic equipment and program product

Abstract

The invention discloses a credit risk detection method and device, electronic equipment and a program product, and relates to the field of financial science and technology, and the method comprises the steps: responding to a credit business request, collecting structured financial data and unstructured news text data of a customer, screening the preprocessed financial data, outputting a financial index set, and storing the financial index set in a database; the method comprises the steps of preprocessing news text data, performing topic representation on the preprocessed news text data, outputting text topic probability distribution data of each news document, fusing a screened financial index set with the text topic probability distribution data after topic representation to generate a detection information set, and inputting the detection information set into a credit risk detection model. And outputting a risk assessment report and a credit business approval suggestion. According to the method and the device, the technical problems that the dependent data source is relatively single, the collected information is not comprehensive, the credit risk detection is influenced and the property loss of the financial institution is easily caused when the financial institution carries out risk detection on the customer credit business in the related technology are solved.

Description

Technical Field

[0001] The present invention relates to the field of financial technology or other related fields, and in particular, to a credit risk detection method and device, electronic equipment, and program product. Background Art

[0002] In today's financial market, credit business is one of the core services of financial institutions, and its risk control ability is directly related to the sound operation of financial institutions and the healthy development of financial markets. Accurate assessment of credit risk is a necessary step before financial institutions issue loans, which involves consideration of many factors such as the borrower's financial status, credit history, and market environment. Traditional credit risk assessment mainly relies on manual analysis and structured data of financial statements. Existing credit risk detection models usually rely on a single data source and use structured data such as corporate financial statements as input. They lack the use of unstructured data such as text involved in traditional credit risk assessment, resulting in incomplete and inaccurate information, which ultimately affects the effectiveness of risk detection.

[0003] In the relevant technologies, financial institutions still use unstructured data at a low level in credit risk assessment, mainly due to the following technical problems: 1. Insufficient information utilization: Traditional credit risk assessment models are mainly based on digital information in financial statements, ignoring the text descriptions and detailed explanations in the notes to the financial statements, which may contain important clues to the business conditions and financial health of the company. 2. Limited data processing capabilities: Unstructured data such as news texts are usually large and complex, and existing text analysis methods are often inefficient in processing such data, and are not effective in processing short texts and sparse information.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present invention provide a credit risk detection method and device, electronic equipment, and program product to at least solve the technical problem in the related art that when financial institutions conduct risk detection on customer credit business, they rely on a relatively single data source and collect incomplete information, which affects credit risk detection and easily causes property losses to the financial institutions.

[0006] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a credit risk detection method is provided, comprising: after receiving a credit business request, collecting structured financial data and unstructured news text data associated with a target customer in the credit business request, and preprocessing the structured financial data and the news text data; screening the preprocessed financial data, outputting a set of financial indicators for credit risk detection, and using a Dirichlet polynomial mixture model to perform topic characterization on the preprocessed news text data, and outputting text topic probability distribution data of each news document; fusing the screened financial indicator set with the text topic probability distribution data after topic characterization to generate a detection information set containing structured financial information and unstructured text information; inputting the detection information set into a credit risk detection model, and outputting a risk assessment report of the current credit business initiated by the target customer and a credit business approval recommendation.

[0007] Optionally, the step of collecting structured financial data and unstructured news text data associated with the target customer in the credit business request includes: collecting financial reports related to the target customer, and screening out balance sheets, income statements and cash flow statements from the financial reports to obtain financial data tables; extracting multiple financial indicators and financial data corresponding to each financial indicator from the financial data tables to obtain the structured financial data, wherein the multiple financial indicators include at least one of the following: current ratio, quick ratio, total asset growth rate, gross profit margin, net profit margin, period expense ratio, cash cost ratio and financial expense cash ratio; collecting news reports related to the target customer from financial news websites and financial engine databases, wherein the types of the news reports include at least one of the following: corporate profit announcements, goodwill warnings, violation penalties and industry dynamic news; extracting text data from the news reports to obtain the unstructured news text data.

[0008] Optionally, the step of preprocessing the structured financial data and the news text data includes: performing data cleaning and missing value processing on the structured financial data, wherein the data cleaning includes: deleting duplicate values ​​and deleting abnormal values, and the missing value processing includes: filling missing data; performing cleaning processing on the news text data, wherein the cleaning processing includes removing stop words, removing punctuation marks, removing numbers and special characters; performing word segmentation processing on the news text obtained after the cleaning processing to obtain a text word set and a text sentence set.

[0009] Optionally, the step of screening the pre-processed financial data and outputting a set of financial indicators for credit risk detection includes: calculating the IV value of each financial indicator in the structured financial data, screening out financial indicators with IV values ​​greater than or equal to a preset threshold, and obtaining financial indicators associated with credit risk detection; discretizing the continuous financial data in the structured financial data, and converting continuous variables into discrete variables using a decision tree binning strategy; and using a chi-square test strategy to perform feature screening on the discretized financial data, removing features whose correlation with credit risk detection is lower than a preset correlation threshold, and outputting the set of financial indicators for credit risk detection.

[0010] Optionally, the step of using the Dirichlet polynomial mixture model to perform topic characterization on the preprocessed news text data includes: using a pre-trained Word2Vec model to perform word vectorization on the preprocessed news text to construct a word vector matrix; using the Dirichlet polynomial mixture model to perform topic extraction on the word vector matrix to obtain the probability distribution of each news document on each topic.

[0011] Optionally, the Dirichlet polynomial mixture model is trained in the following manner: initializing model parameters, including the number of topics and hyperparameters of the Dirichlet distribution; obtaining preprocessed historical news text data, and randomly assigning each historical news document to a topic as the initial state of model training; a topic probability calculation step, iteratively calculating the probability that each historical news document belongs to each topic, wherein, when calculating the probability that each historical news document belongs to each topic, a Gibbs sampling algorithm is used for probability estimation based on the topic features of each document and model parameters; a probability redistribution step, reallocating each of the historical news documents to a target topic according to the calculated probability; repeating the iterative topic probability calculation step and the probability redistribution step until the change rate of the document topic allocation is lower than a preset threshold or the maximum number of iterations is reached, completing the convergence of the topic model, and obtaining the trained Dirichlet polynomial mixture model.

[0012] Optionally, the step of fusing the screened financial indicator set with the text topic probability distribution data after topic characterization to generate a detection information set containing structured financial information and unstructured text information includes: based on a data table connection strategy, merging the financial indicator set of the target customer and the text topic probability distribution data into the detection information set, wherein in the detection information set, an independent column is allocated to each financial indicator in the financial indicator set and each text body in the text topic probability distribution data.

[0013] Optionally, the credit risk detection model is trained in the following manner: obtaining a historical detection information set including historical structured financial information and historical unstructured text information, and dividing the historical detection information set into a training set and a test set; using the Adaboost ensemble learning algorithm to build an initial risk detection model; selecting a decision tree as a weak classifier, and assigning an equal weight to each training sample in the training set; classifier weight distribution adjustment step: according to the iterative process of the Adaboost algorithm, selecting a weak classifier in each round, and performing risk detection training based on the weight distribution of the training samples, wherein the weight of the weak classifier The method comprises the following steps: adjusting the weight distribution of the weak classifiers according to the classification accuracy; updating the weight of the training samples after each round of weak classifier training, wherein the weight of the training samples that are incorrectly classified is increased, and the weight of the training samples that are correctly classified is reduced; repeating the classifier weight distribution adjustment step and the sample weight update step until a preset number of weak classifiers is reached, and obtaining a strong classifier composed of a plurality of weak classifiers; using the test set to perform a performance evaluation on the credit risk detection model, and confirming that the training is completed when the performance evaluation result meets the evaluation accuracy, thereby obtaining the credit risk detection model.

[0014] According to another aspect of an embodiment of the present invention, a credit risk detection device is also provided, comprising: a data collection unit, which is used to collect structured financial data and unstructured news text data associated with a target customer in the credit business request after receiving a credit business request, and preprocess the structured financial data and the news text data; a data screening unit, which is used to screen the preprocessed financial data, output a set of financial indicators for credit risk detection, and use a Dirichlet polynomial mixed model to perform topic characterization on the preprocessed news text data, and output text topic probability distribution data of each news document; a data fusion unit, which is used to fuse the screened financial indicator set with the text topic probability distribution data after topic characterization, and generate a detection information set containing structured financial information and unstructured text information; a risk assessment unit, which is used to input the detection information set into a credit risk detection model, and output a risk assessment report on the current credit business initiated by the target customer and a credit business approval recommendation.

[0015] Optionally, the data collection unit includes: a financial report collection module, which is used to collect financial reports related to the target customer, and screen out the balance sheet, income statement and cash flow statement from the financial report to obtain a financial data table; a financial data extraction module, which is used to extract multiple financial indicators and financial data corresponding to each financial indicator from the financial data table to obtain the structured financial data, wherein the multiple financial indicators include at least one of the following: current ratio, quick ratio, total asset growth rate, gross profit margin, net profit margin, period expense ratio, cash cost ratio and financial expense cash ratio; a news report collection module, which is used to collect news reports related to the target customer from financial news websites and financial engine databases, wherein the types of the news reports include at least one of the following: corporate profit announcements, goodwill warnings, violation penalties and industry dynamic news; a news text extraction module, which is used to extract text data in the news reports to obtain unstructured news text data.

[0016] Optionally, the data collection unit includes: a data preprocessing module, used to perform data cleaning and missing value processing on the structured financial data, wherein the data cleaning includes: deleting duplicate values ​​and deleting abnormal values, and the missing value processing includes: filling missing data; performing cleaning processing on the news text data, wherein the cleaning processing includes removing stop words, removing punctuation marks, removing numbers and special characters; a word segmentation module, used to perform word segmentation processing on the news text obtained after the cleaning processing to obtain a text word set and a text sentence set.

[0017] Optionally, the data screening unit includes: an IV value calculation module, which is used to calculate the IV value of each financial indicator in the structured financial data, screen out financial indicators with IV values ​​greater than or equal to a preset threshold, and obtain financial indicators associated with credit risk detection; a discretization module, which is used to discretize the continuous financial data in the structured financial data, and convert continuous variables into discrete variables using a decision tree binning strategy; a feature screening module, which is used to use a chi-square test strategy to perform feature screening on the discretized financial data, remove features whose correlation with credit risk detection is lower than a preset correlation threshold, and output the set of financial indicators used for credit risk detection.

[0018] Optionally, the data screening unit also includes: a word vectorization processing module, which is used to use a pre-trained Word2Vec model to perform word vectorization processing on the preprocessed news text and construct a word vector matrix; a topic extraction module, which is used to use the Dirichlet polynomial mixture model to extract topics from the word vector matrix and obtain the probability distribution of each news document on each topic.

[0019] Optionally, the Dirichlet polynomial mixture model is trained in the following manner: initializing model parameters, including the number of topics and hyperparameters of the Dirichlet distribution; obtaining preprocessed historical news text data, and randomly assigning each historical news document to a topic as the initial state of model training; a topic probability calculation step, iteratively calculating the probability that each historical news document belongs to each topic, wherein, when calculating the probability that each historical news document belongs to each topic, a Gibbs sampling algorithm is used for probability estimation based on the topic features of each document and model parameters; a probability redistribution step, reallocating each of the historical news documents to a target topic according to the calculated probability; repeating the iterative topic probability calculation step and the probability redistribution step until the change rate of the document topic allocation is lower than a preset threshold or the maximum number of iterations is reached, completing the convergence of the topic model, and obtaining the trained Dirichlet polynomial mixture model.

[0020] Optionally, the data fusion unit includes: a data merging module, which is used to merge the financial indicator set of the target customer and the text topic probability distribution data into the detection information set based on the connection strategy of the data table, wherein in the detection information set, an independent column is allocated to each financial indicator in the financial indicator set and each text body in the text topic probability distribution data.

[0021] Optionally, the credit risk detection model is trained in the following manner: obtaining a historical detection information set including historical structured financial information and historical unstructured text information, and dividing the historical detection information set into a training set and a test set; using the Adaboost ensemble learning algorithm to build an initial risk detection model; selecting a decision tree as a weak classifier, and assigning an equal weight to each training sample in the training set; classifier weight distribution adjustment step: according to the iterative process of the Adaboost algorithm, selecting a weak classifier in each round, and performing risk detection training based on the weight distribution of the training samples, wherein the weight of the weak classifier The method comprises the following steps: adjusting the weight distribution of the weak classifiers according to the classification accuracy; updating the weight of the training samples after each round of weak classifier training, wherein the weight of the training samples that are incorrectly classified is increased, and the weight of the training samples that are correctly classified is reduced; repeating the classifier weight distribution adjustment step and the sample weight update step until a preset number of weak classifiers is reached, and obtaining a strong classifier composed of a plurality of weak classifiers; using the test set to perform a performance evaluation on the credit risk detection model, and confirming that the training is completed when the performance evaluation result meets the evaluation accuracy, thereby obtaining the credit risk detection model.

[0022] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned credit risk detection methods.

[0023] According to another aspect of an embodiment of the present invention, there is also provided an electronic device, comprising one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement any one of the above-mentioned credit risk detection methods.

[0024] According to another aspect of an embodiment of the present invention, a computer program product is provided, including a computer program, wherein when the computer program is executed by a processor, the steps of any one of the above-mentioned credit risk detection methods are implemented.

[0025] In the present disclosure, after receiving a credit business request, structured financial data and unstructured news text data associated with a target customer in the credit business request can be collected, and the structured financial data and news text data can be preprocessed, the preprocessed financial data can be screened, a set of financial indicators for credit risk detection can be output, and a Dirichlet polynomial mixed model can be used to perform topic characterization on the preprocessed news text data, and text topic probability distribution data of each news document can be output. The screened financial indicator set can be fused with the text topic probability distribution data after topic characterization to generate a detection information set containing structured financial information and unstructured text information, the detection information set can be input into a credit risk detection model, and a risk assessment report on the current credit business initiated by the target customer and a credit business approval recommendation can be output.

[0026] Based on the above-mentioned public content, on the basis of traditional structured financial data, unstructured news text data can be introduced to analyze customer credit risk from multiple information dimensions, which can enable more comprehensive and accurate corporate credit risk detection. At the same time, in terms of the representation method of unstructured text data, the advantages of pre-trained models and topic models are integrated, and the vectorized representation of text and topic probability distribution are considered at the same time, text information is enriched, the accuracy of credit risk detection is improved, the risk prevention and control capabilities and market competitiveness of financial institutions are improved, and the property losses of financial institutions are reduced. The technical problem that financial institutions rely on a relatively single data source and incomplete information collection when conducting risk detection on customer credit business in related technologies is solved, which affects credit risk detection and easily causes property losses to financial institutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0028] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a credit risk detection method is shown;

[0029] Figure 2 is a flow chart of an optional credit risk detection method according to an embodiment of the present invention;

[0030] Figure 3 is a flow chart of credit risk detection based on a credit risk detection model of multi-dimensional information according to an embodiment of the present invention;

[0031] Figure 4 is a schematic diagram of a probabilistic graphical model of a Dirichlet polynomial mixture model according to an embodiment of the present invention;

[0032] Figure 5 is a schematic diagram of an optional credit risk detection device according to an embodiment of the present invention;

[0033] Figure 6 It is a structural block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0035] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0036] To facilitate those skilled in the art to understand the present invention, some terms or nouns involved in the embodiments of the present invention are explained below:

[0037] The word2vec model is a deep learning model that converts words in a text into vector representations.

[0038] IV value, Information Value, is a statistic used to quantify the strength of the relationship between a feature variable and a target variable, and is mainly used for feature selection. In the field of credit risk, it is used to measure the predictive ability of a financial indicator for credit risk. The higher the V value, the stronger the correlation between the feature variable and the target variable, and the greater the contribution to the model.

[0039] Chi-squared test, Chi-squared test, is a statistical test method used to determine whether there is a significant association between two categorical variables. In feature selection, the chi-squared test can be used to test whether the feature variable is associated with the target variable (such as "default" or "non-default").

[0040] Dirichlet Multinomial Mixture Model, Dirichlet Multinomial Mixture Model, a non-parametric Bayesian model for text topic modeling. Unlike the traditional LDA model, the DMM model assumes that the terms in the document follow a multinomial distribution, while the prior distribution of the document topic follows a Dirichlet distribution. In the present invention, during the model training process, the model parameters are iteratively optimized by methods such as Gibbs sampling or variational inference to obtain the topic distribution of the document.

[0041] AdaBoost model, Adaptive Boosting, is an integrated learning algorithm for classification, which mainly forms a strong classifier by combining multiple weak classifiers. In each round of training, AdaBoost adjusts the weight of the samples according to the error rate of the previous round of classifiers. Those samples that are misclassified will be given higher weights in the next round of training, forcing subsequent classifiers to pay more attention to these difficult samples. In this way, a series of weak classifiers gradually adjust and focus on the most difficult samples to classify, and finally combine the outputs of these classifiers to form a more powerful model that can make classification predictions more accurately. In credit risk detection, AdaBoost can effectively improve the sensitivity and accuracy of the model in identifying high-risk customers.

[0042] It should be noted that the credit risk detection method and device in the present disclosure can be used in the field of financial technology to implement risk detection of customer credit business based on multi-dimensional customer-related data, and can also be used in any field other than the field of financial technology to implement risk detection of customer credit business based on multi-dimensional customer-related data. The present disclosure does not limit the application field of the credit risk detection method and device.

[0043] It should be noted that the information (including but not limited to customer identity information, customer financial account information, customer news reports, customer personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) collected by this disclosure are information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of relevant data are in compliance with the relevant laws, regulations and standards of the relevant regions, necessary confidentiality measures are taken, and public order and good customs are not violated, and corresponding operation portals are provided for users to choose to authorize or refuse. For example, an interface is set up between this system and relevant users or institutions. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or institution through the interface, and obtain relevant information after receiving the consent information fed back by the aforementioned user or institution.

[0044] It should be noted that in the present disclosure, when collecting and analyzing customer information, corresponding operation entrances are provided for users to choose to agree or reject the automated decision results; if the user chooses to reject, the expert decision-making process is entered.

[0045] The following embodiments of the present invention can be applied to various credit risk detection systems / applications / devices. Based on traditional structured financial data, the present invention focuses on the annotation data of the customer's financial statements and introduces unstructured news text data to analyze corporate credit risks from multiple information dimensions. Compared with the existing technology, it can perform corporate credit risk detection more comprehensively and accurately. In addition, in terms of the representation of unstructured text data, the advantages of pre-trained models and topic models are integrated, and the vectorized representation of text and topic probability distribution are considered at the same time to enrich text information.

[0046] The present invention is described in detail below in conjunction with various embodiments.

[0047] Embodiment 1

[0048] According to an embodiment of the present invention, an embodiment of a credit risk detection method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0049] The credit risk detection method embodiment provided in the first embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal (or mobile device) for implementing a credit risk detection method. Figure 1 As shown, the computer terminal 10 (or mobile device) may include one or more ( Figure 1 , 102n are used to illustrate) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU (Microcontroller Unit) or a programmable logic device FPGA (Field Programmable Gate Array)), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a Universal Serial Bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. Those skilled in the art can understand that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0050] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10 (or mobile device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0051] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the credit risk detection method in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, the above-mentioned credit risk detection method is realized. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0052] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0053] The display may be, for example, a touch screen liquid crystal display (LCD), which may enable a user to interact with a user interface of the computer terminal 10 (or mobile device).

[0054] It should be noted that the credit risk detection method provided by the embodiment of the present invention can solve the challenges encountered by financial institutions when evaluating the credit risk of customers (such as enterprises and factories) before issuing loans. It is applicable to various financial institutions in scenarios such as enterprise credit assessment, risk management, and investment decision-making. By integrating structured financial data and unstructured text data, the accuracy, timeliness, and comprehensiveness of credit risk assessment are improved, thereby effectively improving the risk prevention and control capabilities and market competitiveness of financial institutions. Specific application scenarios may include: 1. Enterprise loan approval: When an enterprise applies for a loan from a bank or other financial institution, the credit risk detection model of this application can quickly and accurately assess the financial status and potential risks of the applicant enterprise. By integrating structured data in the enterprise's financial statements with unstructured data such as news reports, the model can provide a more comprehensive risk assessment and help financial institutions make more informed loan decisions. 2. Credit portfolio management: For loans that have been issued, by continuously monitoring changes in the company's financial indicators and related news reports, the credit risk detection model of this application can regularly assess the overall risk level of the credit portfolio. This dynamic monitoring capability is essential for timely identification and handling of possible credit risks and adjustment of credit strategies and portfolio structures. 3. Risk warning system: Combined with real-time news streams and financial data updates, the credit risk detection model of this application can be used as part of a risk warning system to automatically identify borrowing companies whose financial conditions have deteriorated or there are negative news reports, provide early warning signals, and help financial institutions take measures in advance to avoid or reduce potential losses. 4. Credit rating update: Credit rating agencies or financial institutions can use the model of this application to update the credit rating of a company. Compared with traditional rating methods, this credit risk detection model can consider more dimensions of information, such as the business environment, industry trends, emergencies, etc., to ensure the accuracy and timeliness of the rating. 5. Risk preference management: Financial institutions can adjust the parameters in the model according to their risk preferences, such as the screening thresholds of financial indicators, the parameters of the theme model, etc., to adapt to different credit policies and market environments.

[0055] Under the above operating environment, this application provides Figure 2 The credit risk detection method shown. Figure 2 is a flow chart of an optional credit risk detection method according to an embodiment of the present invention, such as Figure 2 As shown, the method comprises the following steps:

[0056] Step S201, after receiving a credit business request, collecting structured financial data and unstructured news text data associated with a target customer in the credit business request, and preprocessing the structured financial data and the news text data.

[0057] After receiving the credit business request, the credit business request is parsed to obtain the customer identification and associated financial accounts of the target customer who initiated the request, and to determine the customer's specific credit business products and credit terms. Then the first step is to collect structured financial data and unstructured news text data associated with the target customer. Structured data mainly comes from the customer's financial reports, including but not limited to balance sheets, income statements, and cash flow statements. These data are obtained through the bank's internal system or a third-party credit assessment platform, reflecting the financial health of the company. Unstructured data mainly comes from financial news websites, industry reports, social media, and corporate financial report notes. This type of data can provide multi-dimensional information such as corporate operating conditions, market reputation, industry trends, and emergencies, which plays an important supplementary role in credit risk assessment.

[0058] Optionally, the step of collecting structured financial data and unstructured news text data associated with the target customer in the credit business request includes: collecting financial reports related to the target customer, and screening out the balance sheet, income statement and cash flow statement from the financial report to obtain a financial data table; extracting multiple financial indicators and financial data corresponding to each financial indicator from the financial data table to obtain structured financial data, wherein the multiple financial indicators include at least one of the following: current ratio, quick ratio, total asset growth rate, gross profit margin, net profit margin, period expense ratio, cash cost ratio and financial expense cash ratio; collecting news reports related to the target customer from financial news websites and financial engine databases, wherein the types of news reports include at least one of the following: corporate profit announcements, goodwill warnings, violation penalties and industry dynamic news; extracting text data from news reports to obtain unstructured news text data.

[0059] In this embodiment, when collecting data, it includes collecting structured financial data and unstructured news text data. Here, the collection and screening of structured financial data are described. First, financial institutions will collect all financial reports involving target customers through various legal channels, such as financial information submitted by enterprises, public financial reports or data from third-party credit assessment agencies. These reports are the basis for assessing the financial health and credit risk of enterprises. From the collected financial reports, the balance sheet, income statement and cash flow statement are selected. These three statements are the core reflection of the financial status of the enterprise, providing detailed information on assets, liabilities, income, profits and cash flows, and are indispensable data sources for credit risk assessment. Multiple key indicators are extracted from the financial statements, including but not limited to the current ratio, quick ratio, total asset growth rate, gross profit margin, net profit margin, period expense ratio, cash cost ratio and financial expense cash ratio. These indicators are directly related to the debt repayment ability, profitability, cost control and cash flow status of the enterprise, and have a direct impact on credit risk assessment. The extracted financial indicators and their corresponding values ​​are converted into structured data tables to facilitate subsequent model processing and analysis. This step ensures the standardization and format consistency of the data.

[0060] For the collection and preprocessing of unstructured news text data, here, news reports related to target customers can be collected through channels such as financial news websites, financial database engines or social media. These reports cover a wide range of information such as corporate profits, goodwill changes, legal proceedings, violations and penalties, and industry dynamics, and can provide important clues beyond financial statements. Information directly related to credit risk is screened out from news reports, including corporate profit announcements, goodwill warnings, violations and penalties, and industry dynamics news. These types of information can reveal the real-time dynamics of corporate operations and help to adjust risk assessments in real time. Text data is extracted from the screened news reports, and irrelevant formats, advertising information, stop words, and punctuation marks are removed, and only text content with potential value for risk assessment is retained. Text cleaning is a necessary step before text analysis, which can improve data quality and analysis efficiency. The extracted text data is preprocessed, including steps such as word segmentation, part-of-speech tagging, and topic modeling, to convert unstructured news text into quantifiable features, such as word vectors and topic probability distributions, to provide rich information input for subsequent credit risk assessment models.

[0061] After obtaining structured financial data and unstructured news text data, these two types of data need to be preprocessed. Optionally, the steps of preprocessing structured financial data and news text data include: data cleaning and missing value processing for structured financial data, wherein data cleaning includes: deleting duplicate values ​​and outliers, and missing value processing includes: filling missing data; cleaning processing for news text data, wherein the cleaning processing includes removing stop words, punctuation marks, numbers and special characters; and word segmentation processing for the news text obtained after the cleaning processing to obtain a text word set and a text sentence set.

[0062] In the preprocessing stage, structured financial data needs to be merged, cleaned, and transformed, including filling missing values, deleting outliers, and standardizing amount variables to ensure data consistency and reliability. For unstructured text data, first perform text cleaning to remove irrelevant symbols, numbers, and stop words, and then use word segmentation tools to segment the text into words or phrases to prepare for subsequent vectorization and topic extraction.

[0063] Specifically, the above-mentioned data cleaning may include: deleting duplicate values ​​and outliers. Deleting duplicate values ​​here means that in the collected financial data, the same data may appear repeatedly in different records. This may be caused by human errors in the data entry process or inconsistencies in the data source. Deleting duplicate values ​​can ensure that each piece of financial information is unique and avoid deviations in model training. Deleting outliers refers to the occasional values ​​in the financial data that are significantly deviated from the normal range. These values ​​may be caused by data entry errors, abnormal financial activities or special circumstances (such as changes in accounting policies). Identifying and deleting these outliers is crucial to maintaining the accuracy and reliability of the data set. Optionally, the outlier detection method used in this embodiment includes using box plots, Z-scores, and IQR (interquartile range) to determine the degree of abnormality of the values.

[0064] In addition, in structured financial data, some fields may be missing due to various reasons (such as delayed reporting, incomplete data). There are many ways to handle missing values, including deleting records with missing values, filling with mean or median, and using regression model prediction to fill. In this application, considering the continuity and importance of financial data, you may choose to use a prediction model or fill according to historical data trends to maintain the integrity and information content of the data.

[0065] Further, it is necessary to explain the preprocessing of news text data. Among them, the cleaning process includes: removing stop words. Stop words refer to words that appear frequently in the text but contribute little to the text theme, such as "de", "shi", "he", etc. Removing stop words can reduce the redundancy of text data and improve the efficiency and accuracy of subsequent analysis. Removing punctuation marks, numbers, and special characters: Punctuation marks, numbers, and special characters usually do not have a direct impact on the text theme or sentiment analysis, but instead increase the complexity of the data. Removing these elements helps simplify the text and make it more suitable for analysis by natural language processing algorithms.

[0066] After that, it is necessary to perform word segmentation on the news text to obtain a text word set and a text sentence set. Here, word segmentation is the process of splitting continuous text into independent words or phrases, which is a key step in text preprocessing. For news text data, word segmentation helps convert the text into a form that can be understood by machines. Through word segmentation, a text word set is obtained, that is, the set of all words in the text.

[0067] Cleaning and missing value processing of structured financial data can eliminate inconsistencies and incompleteness in the data and improve the accuracy of the data. For news text data, removing stop words, punctuation marks, numbers, and special characters, as well as performing word segmentation, can convert unstructured information into structured features for further analysis.

[0068] In step S202, the preprocessed financial data is screened to output a set of financial indicators for credit risk detection, and the Dirichlet polynomial mixture model is used to perform topic characterization on the preprocessed news text data to output the text topic probability distribution data of each news document.

[0069] The preprocessed financial data is screened through IV value calculation and chi-square test to select financial indicators with a high contribution to credit risk prediction and construct a set of financial indicators. The IV value measures the discrimination ability of variables for credit risk prediction, and the chi-square test helps screen variables significantly related to credit risk, thus ensuring the efficiency and accuracy of the model. Optionally, the step of screening the preprocessed financial data to output a set of financial indicators for credit risk detection includes: calculating the IV value of each financial indicator in the structured financial data, screening out financial indicators with an IV value greater than or equal to a preset threshold to obtain financial indicators related to credit risk detection; discretizing the continuous financial data in the structured financial data, and using the decision tree binning strategy to convert continuous variables into discrete variables; using the chi-square test strategy to perform feature screening on the discretized financial data, removing features with a correlation degree lower than the preset correlation degree threshold for credit risk detection, and outputting a set of financial indicators for credit risk detection.

[0070] First, the IV value of each financial indicator in the structured financial data is calculated to find out those indicators that contribute significantly to credit risk detection. The calculation formula of the IV value can be:

[0071] [IV=\sum(GoodPercent-BadPercent)*Weight];

[0072] Among them, (GoodPercent) and (BadPercent) represent the proportion of risk-free customers and customers with credit risk in the financial indicator segment, respectively, and (Weight) is the proportion of the number of customers in this segment to the total number of customers. By calculating the IV value, the ability of each financial indicator to distinguish credit risk can be measured, and then the financial indicators with IV values ​​greater than or equal to the preset threshold can be screened out as the input of the subsequent risk assessment model. Indicators with IV values ​​greater than 0.3 are considered strong predictive features, while those less than 0.02 are considered almost useless and should be excluded.

[0073] For example, this embodiment extracts a total of 176 financial indicators from the balance sheet, income statement, cash flow statement and financial statement notes, and incorporates them into the initial feature pool, and then merges, cleans and transforms the data. Considering the correlation between the initial features, the IV value of the variable is calculated to measure the contribution of the variable to the target prediction, and the variables with higher IV values ​​are screened out. For discrete variables, the IV value and the grading results are directly calculated as shown in Table 1:

[0074] Table 1

[0075] Grade IV value range Useless <0.02 Weak [0.02,0.1) Medium [0.1,0.3) Strong [0.3,0.5) Suspicious >＝0.5

[0076] After the calculation is completed, all fields with IV values ​​belonging to the Use less level are deleted. For continuous variables, they are first discretized through decision tree binning, and then their IV values ​​are calculated, and features with IV values ​​less than 0.02 are deleted. The chi-square test is used for feature screening, and the chi-square test strategy is used to evaluate the correlation between the discretized financial data and credit risk. In the specific operation of the chi-square test, a significance level of the chi-square test is set (such as a p-value less than 0.01), and the chi-square value of each discretized indicator is calculated. Features whose correlation with credit risk detection is lower than the preset threshold will be excluded to reduce the complexity of the model, avoid overfitting, and ensure the prediction accuracy and efficiency of the model.

[0077] Among the selected financial indicators, for those continuous financial data, such as current ratio and gross profit margin, the decision tree binning strategy is used for discretization. Discretization is the process of converting continuous values ​​into categorical values, which can simplify the complexity of the model and improve the prediction performance. The decision tree binning strategy automatically finds the best split point of the data by building a decision tree, converting continuous variables into discrete variables, which is helpful for subsequent statistical analysis and model training. The discretized financial data needs further feature screening to ensure that the features of the model input have a practical contribution to credit risk detection.

[0078] For unstructured news text data, the pre-trained word2vec model is used for word embedding processing to convert the text into a vector representation. At the same time, this embodiment innovatively uses the Dirichlet polynomial mixture model (DMM) to represent the topic of news text. The DMM model can extract news topics more accurately based on the characteristics of short text and sparse information. During the model training process, each news document will be assigned to a different topic to obtain text topic probability distribution data, which provides a rich source of unstructured information for subsequent risk detection.

[0079] Optionally, the step of using the Dirichlet polynomial mixture model to perform topic characterization on the preprocessed news text data includes: using a pre-trained Word2Vec model to perform word vectorization on the preprocessed news text to construct a word vector matrix; using the Dirichlet polynomial mixture model to extract topics from the word vector matrix to obtain the probability distribution of each news document on each topic.

[0080] In this embodiment, the pre-trained Word2Vec model is used to perform word vectorization on the pre-processed news text to construct a word vector matrix. Word2Vec is a word embedding technology based on a neural network, which converts words in the text into vectors of fixed size. These vectors can capture the semantic relationship between words and reduce the dimension of the processed text data. In this implementation step, the pre-processed news text (stop words, punctuation marks, etc. have been removed) is input into the Word2Vec model to generate a word vector matrix.

[0081] It should be noted that the Dirichlet polynomial mixture model (DPMM) mentioned in this embodiment is obtained by training in the following manner: initializing model parameters, including the number of topics and hyperparameters of the Dirichlet distribution; obtaining preprocessed historical news text data, and randomly assigning each historical news document to a topic as the initial state of model training; a topic probability calculation step, iteratively calculating the probability that each historical news document belongs to each topic, wherein, when calculating the probability that each historical news document belongs to each topic, the Gibbs sampling algorithm is used for probability estimation based on the topic features of each document and the model parameters; a probability redistribution step, reallocating each historical news document to the target topic according to the calculated probability; repeating the iterative topic probability calculation step and the probability redistribution step until the change rate of the document topic allocation is lower than the preset threshold or the maximum number of iterations is reached, completing the convergence of the topic model, and obtaining the trained Dirichlet polynomial mixture model.

[0082] During the model training process, at the initial stage of model training, it is necessary to set the number of topics (K) (although DPMM is adaptive to the number of topics, an initial value must be set at the beginning of training), the hyperparameters of the Dirichlet distribution (\a lpha) (control the concentration of topic distribution) and (\beta) (affect the distribution of words in topics). Then collect the preprocessed historical news text data, which will be used for model training to help the model learn different topics and their corresponding word distributions. Each historical news document is randomly assigned to a topic as the initial state of model training. During the iteration process, for each historical news document, based on the topic features of each document and the current model parameters, the Gibbs sampling algorithm is used for probability estimation to calculate the probability that the document belongs to each topic. Gibbs sampling is a sampling method used to process complex probabilistic models. It can gradually update the model parameters according to the existing information until the model reaches a stable state. According to the probability of each document belonging to each topic, each historical news document is redistributed to the most likely topic. This process is to optimize the document topic allocation, so that documents with the same topic are as concentrated as possible, and documents with different topics are as dispersed as possible. The iterative calculation of topic probability and probability redistribution steps are repeated until the rate of change of document topic allocation is lower than the preset threshold (indicating that the model has reached a stable state) or the maximum number of iterations is reached (to prevent infinite iterations). Model convergence marks the completion of the training process. At this time, the model parameters and topic allocation can better characterize the topic structure of the news text. Through the above process, DPMM can characterize the news text data and obtain the probability distribution of each news document on each topic. These probability distribution data will be used together with structured financial data to build a credit risk detection model to improve the prediction ability of the model.

[0083] Step S203, fusing the screened financial indicator set with the text topic probability distribution data after topic characterization to generate a detection information set including structured financial information and unstructured text information.

[0084] Optionally, step S203 includes: based on the connection strategy of the data table, merging the target customer's financial indicator set and the text topic probability distribution data into the detection information set, wherein in the detection information set, an independent column is allocated to each financial indicator in the financial indicator set and each text body in the text topic probability distribution data.

[0085] During the data fusion process, a table join strategy is used to merge the structured financial data with the topic probability distribution of the unstructured text data. This usually involves using the customer ID as the join key to ensure that each financial data record can be associated with the corresponding text data record. The target customer has a unique identifier in both the financial indicator set and the text topic probability distribution data, which will be used to accurately join the data.

[0086] Through the connection strategy, each financial indicator record of the target customer and the corresponding text topic probability distribution record are merged into a unified data table, which is the detection information set. In the detection information set, the probability distribution of each financial indicator and each text topic will be assigned to a separate column, forming a multidimensional data table, in which each row represents a customer and each column contains information of different dimensions.

[0087] It should be noted that during the merging process, special attention was paid to the independent column assignment for each financial indicator and text topic probability distribution to ensure that the data structure is clear and distinguishable. For financial indicators, each column will contain the value of a specific indicator (such as current ratio, quick ratio, etc.), which reflects the performance of the customer in terms of financial health. For text topics, each column will contain the probability value of a specific topic, which quantifies the degree of association between the news text and the topic. In order for the financial indicators and text topic probability distribution data to work effectively together in the detection information set, it may be necessary to standardize the data. This includes converting the financial numerical indicators into a form that can be compared with the probability distribution data, or adjusting the range of the text topic probability distribution to match the magnitude of the financial data. Standardization ensures that different types of features can be treated equally during the model training process, avoiding the difference in numerical amplitude affecting the model's decision. The final detection information set is a data set that integrates structured financial information and unstructured text information, providing a comprehensive analytical basis for the credit risk detection model. Through this multi-dimensional data integration, the model will be able to conduct credit risk assessment based on richer and more diverse information sources, thereby improving the accuracy and reliability of detection.

[0088] Step S204, input the detection information set into the credit risk detection model, and output the risk assessment report of the current credit business initiated by the target customer and the credit business approval suggestion.

[0089] The detection information set is input into the credit risk detection model constructed in this embodiment. The model is based on the Adaboost ensemble learning algorithm, and through the collaboration of multiple weak classifiers, a strong classifier that is highly sensitive to credit risk is formed. After the model training is completed, for the current credit business initiated by the target customer, the model outputs a risk assessment report, which contains information such as the customer's default probability, credit rating, and potential risk factors. Based on the risk assessment report, the model will also give approval suggestions for the credit business, such as approving the loan, requiring additional collateral, rejecting the loan, etc., to provide a scientific basis for credit decision-making.

[0090] Through the above steps, after receiving a credit business request, the structured financial data and unstructured news text data associated with the target customer in the credit business request can be collected, and the structured financial data and the news text data can be preprocessed, the preprocessed financial data can be screened, and a set of financial indicators for credit risk detection can be output. The preprocessed news text data can be topic characterized by using a Dirichlet polynomial mixture model, and the text topic probability distribution data of each news document can be output. The screened financial indicator set can be fused with the text topic probability distribution data after topic characterization to generate a detection information set containing structured financial information and unstructured text information, the detection information set can be input into a credit risk detection model, and a risk assessment report on the current credit business initiated by the target customer and a credit business approval recommendation can be output. In this embodiment, on the basis of traditional structured financial data, unstructured news text data can be introduced to analyze customer credit risk from multiple information dimensions, so that corporate credit risk detection can be performed more comprehensively and accurately. At the same time, in terms of the representation method of unstructured text data, the advantages of pre-trained models and topic models are integrated, and the vectorized representation of text and topic probability distribution are considered at the same time, text information is enriched, the accuracy of credit risk detection is improved, the risk prevention and control capabilities and market competitiveness of financial institutions are improved, and the property losses of financial institutions are reduced. The technical problem that financial institutions rely on a relatively single data source and collect incomplete information when conducting risk detection on customer credit business in related technologies is solved, which affects credit risk detection and easily causes property losses of financial institutions.

[0091] Optionally, the credit risk detection model is trained in the following manner: obtaining a historical detection information set including historical structured financial information and historical unstructured text information, and dividing the historical detection information set into a training set and a test set; using the Adaboost ensemble learning algorithm to construct an initial risk detection model; selecting a decision tree as a weak classifier, and assigning equal weights to each training sample in the training set; a classifier weight distribution adjustment step: according to the iterative process of the Adaboost algorithm, selecting a weak classifier in each round, and performing risk detection training based on the weight distribution of the training samples, wherein the weight of the weak classifier is adjusted according to the classification accuracy; a sample weight update step: after each round of weak classifier training is completed, updating the weight of the training sample, wherein when updating the weight of the training sample, the weight of the wrongly classified training sample is increased, and the weight of the correctly classified training sample is reduced; repeatedly executing the classifier weight distribution adjustment step and the sample weight update step until a preset number of weak classifiers is reached, and a strong classifier composed of multiple weak classifiers is obtained; using the test set to perform performance evaluation on the credit risk detection model, and when the performance evaluation result meets the evaluation accuracy, confirming that the training is completed, and obtaining the credit risk detection model.

[0092] In the process of training the credit risk detection model, first, a detection information set is obtained from historical data. This set contains structured financial information and unstructured text information. In order to evaluate the performance of the model and avoid overfitting, the historical detection information set is divided into a training set and a test set. The training set is used for model training, and the test set is used for model performance evaluation.

[0093] Adaboost is an iterative ensemble learning method that aims to build a strong classifier from multiple weak classifiers. The Adaboost algorithm is used as the basis to build the initial risk detection model. In the initialization phase of Adaboost, a decision tree is selected as a weak classifier, and each training sample in the training set is assigned an equal weight of 1 / N, where N is the total number of training samples. In the iterative training process of Adaboost, a weak classifier (decision tree) is selected in each round and trained for risk detection based on the current sample weight distribution. The weight of the weak classifier is adjusted according to its classification accuracy, that is, for weak classifiers with high classification accuracy, its weight increases; for weak classifiers with low classification accuracy, its weight decreases. The adjustment of classifier weights is achieved by calculating the error rate of each weak classifier and updating it according to the weight update rule of Adaboost. After each round of weak classifier training is completed, Adaboost updates the weight distribution of the training samples. Specifically, the weights of samples that are misclassified by the current weak classifier will increase, while the weights of correctly classified samples will decrease. This weight update mechanism ensures that the model can gradually focus on samples that are difficult to classify correctly during the training process, thereby improving the classification effect of the model as a whole. Through the iterative process of the Adaboost algorithm, a series of weak classifiers with different weights are generated. When the preset number of weak classifiers is reached, these weak classifiers will be combined into a strong classifier for the final credit risk detection.

[0094] After the strong classifier is built, the previously divided test set is used to evaluate the performance of the credit risk detection model. The main evaluation indicators include accuracy, recall, F1 score, etc. If the performance evaluation result of the model on the test set meets the preset accuracy threshold (for example, the accuracy is greater than 90%), the model training is considered complete and the final credit risk detection model is confirmed.

[0095] The model can effectively integrate structured financial information and unstructured text information, providing financial institutions with comprehensive credit risk assessment capabilities.

[0096] Another optional specific implementation is described in detail below.

[0097] Figure 3 is a flow chart of performing credit risk detection based on a credit risk detection model of multi-dimensional information according to an embodiment of the present invention. Figure 3 As shown, including:

[0098] Step 1: Data collection.

[0099] like Figure 3As shown, first collect corporate financial report data, which mainly includes data tables such as balance sheet, income statement and cash flow statement. Select financial indicators with more information in the financial report, including current ratio, quick ratio, total asset growth rate, etc. in the balance sheet, gross profit margin, net profit margin, period expense rate, etc. in the income statement, and cash cost ratio, financial expense cash ratio and other indicators in the cash flow statement.

[0100] Considering that corporate financial data is updated regularly, the published data has a certain lag and cannot keep up with the business conditions of the company in a timely manner. Therefore, in addition to structured financial data, this embodiment also considers information in unstructured data. There are often many clues hidden in corporate news reports on financial websites. At the same time, news reports are timely and objective. Introducing news text features into the risk detection model can detect corporate credit risks more timely.

[0101] Step 2: Data preprocessing and feature screening.

[0102] For structured financial data, this embodiment extracts a total of 176 financial indicators from the three financial statements of the balance sheet, income statement, and cash flow statement, as well as the notes to the financial statements, and incorporates them into the initial feature pool. The data is then merged, cleaned, and transformed, including missing value filling, singular value deletion, and outliers in the amount variable are deleted. Taking into account the correlation between the initial features, the IV value of the variable is calculated to measure the contribution of the variable to the target prediction, and the variables with higher IV values ​​are screened out. After the calculation is completed, all fields whose IV values ​​belong to the Useless level are deleted. For continuous variables, they are first discretized through decision tree binning, and then their IV values ​​are calculated, and features with IV values ​​less than 0.02 are deleted.

[0103] The chi-square test is used for feature screening, and the chi-square test strategy is used to evaluate the correlation between the discretized financial data and credit risk. In the specific operation of the chi-square test, a significance level of the chi-square test is set (such as a p-value less than 0.01), and the chi-square value of each discretized indicator is calculated. Features whose correlation with credit risk detection is lower than the preset threshold will be excluded to reduce the complexity of the model, avoid overfitting, and ensure the prediction accuracy and efficiency of the model. The smaller the p-value, the less likely the H0 hypothesis is to hold under the current observation, that is, the current field has a significant impact on the sample distribution, and the variables with a p-value less than 0.01 are deleted. For the fields that still exist, the Pearson coefficient between each pair of fields is calculated. If the coefficient is greater than 0.95, delete the one with the lower IV value of the two fields. For example, 89 financial indicators are screened out for credit risk detection.

[0104] For unstructured text data, we first use the word2vec pre-trained model to vectorize the text, and then embed the word vectors by setting the window size to achieve the purpose of dimensionality reduction. At the same time, the context window in the model takes into account that words with similar distribution of adjacent words often have similar meanings, so the word meaning is also included in the model.

[0105] Step 3: Topic model construction and training.

[0106] The keywords of news reports can summarize the main content of the news. For example, when a report contains keywords such as "year-on-year", "growth", "net profit", "100 million yuan" and so on for many times, the report is generally related to the company's profitability; when a report contains words such as "goodwill", "claim", "fine", "warning", "violation" and so on for many times, it is mostly negative news about the company. News reports are mostly about companies that disclose profitability with small credit, while news reports are mostly about companies with negative news and have greater credit risks. Therefore, on the basis of vectorized representation of text, the topic model is further used to extract the text keyword representation, and on the basis of the traditional topic model, a Dirichlet polynomial mixed model is proposed for short text topic representation, and a more effective Dirichlet polynomial mixed model solution method is explored in short text topic representation, in an effort to improve the accuracy and adaptability of short text topic representation.

[0107] This embodiment uses the Dirichlet polynomial mixture model for short text topic representation. The model is a non-parametric Bayesian model that can well balance the sparsity and high dimensionality problems in short text topic representation.

[0108] The traditional LDA topic model assumes that each document in the document collection is distributed on a topic. When generating words in a document, a topic is first generated according to the topic distribution of the document, and then a word is generated through the distribution of the topic on the word. The Dirichlet polynomial mixture model assumes that the document is generated by a mixture model, and there is a one-to-one correspondence between the mixture components and the clusters. When generating a document, first a topic is selected from K topics according to the weights p(z=k) of each topic, assuming it is the Kth topic, and then a document w is generated according to the distribution p(u|z=k) corresponding to the topic. The probability that document w is generated by the Dirichlet polynomial mixture model is:

[0109]

[0110] Where K is the number of all topics. The weight of each topic corresponds to the multinomial distribution p(z=k)=Mult(z=k|θ)=θ k Generated, the prior distribution of this multinomial distribution is the Dirichlet distribution in The distribution corresponding to each topic is the multinomial distribution p(x|z=k)=Mult(x|φ k )=φ k,x , the prior distribution of each topic is Dirichlet distribution in

[0111] like Figure 4 The figure shows the probabilistic graphical model of the Dirichlet polynomial mixture model, where the generation process of the Dirichlet polynomial mixture model is as follows:

[0112]

[0113] z u |θ~Mult(θ)

[0114]

[0115] Among them, p(u|φ z,u )=Π x∈u Mult(x|φ z,u )

[0117] Compared with the traditional LDA topic model, the Dirichlet polynomial mixture model does not need to generate a high-dimensional vector with the length of the dictionary, but only needs to generate words in the text, which can better describe the generation process of short texts. When the Dirichlet polynomial mixture model is used for clustering problems, the Gibbs sampling algorithm is usually used to solve it. Considering that in the representation of short texts, it is not necessary to obtain the random variables θ and φ in the Dirichlet polynomial mixture model, the variables θ and φ are marginalized in the solution process, and the conditional distribution is directly solved. The conditional probability that short text u is generated by topic z is:

[0118]

[0119] Where N is the total number of all texts, It is the number of texts in topic z without considering document u. When there are more texts in topic z, text u will tend to choose topics with more documents, so it has better completeness, that is, texts that should actually be generated by the same topic model can be assigned to the same topic.

[0120] is the total number of words in topic z without considering text u, It represents the frequency of word x in topic z without considering text u. When there are more texts in topic z that have similar words to text w, text w will have a greater probability of being generated by topic z, and therefore have better similarity, that is, texts assigned to the same topic actually belong to the same topic.

[0121] The learning process of the topic model is a parameter estimation process. First, K topics are initialized and the text is randomly assigned to a topic. During the iteration process, the probability of each short text belonging to each topic is calculated, and it is redistributed according to the probability distribution calculation result. Each iteration makes the topic model develop in a more accurate direction. Finally, the probability distribution of each news document on the topic can be obtained for the subsequent risk detection model.

[0122] Next, we will explain the effect of parameters α and β on the results. According to the conditional probability formula that short text u is generated by topic z, we can see that when α = 0, the probability of the text selecting an empty topic is zero. When α increases, the probability of the text selecting an empty topic will increase accordingly. A larger α is more likely to produce a topic with only one text. When β is small, the probability of a text selecting a topic will be affected. The value of β is more sensitive and pays more attention to the number of times word x appears in topic z, that is, the similarity of topics between texts. Therefore, relatively more topics and higher similarities can be formed. When β is larger, relatively fewer topics and higher similarities are formed.

[0123] In addition, in order to explore the impact of the initial number of topics K on the results, the number of iterations was set to 10, and the initial α = 0.2, β = 0.2. It can be found that the larger the initial number of topics K, the closer the number of topics is to the actual number of topics when the algorithm tends to stabilize.

[0124] Step 4: Risk detection model construction and training.

[0125] The processed structured financial data and unstructured text data are spliced ​​into a wide table (corresponding to the detection information set obtained by merging the customer's financial indicator set and the text topic probability distribution data in the above embodiment), where the representation of the unstructured text data includes word2vec vectorization representation and the probability distribution of the text on the topic. The Adaboost ensemble learning algorithm is used to build a risk detection model.

[0126] During the training process, the data is divided into training set X train and the test set X test , let the training sample set be T = {(x1,y1),(x2,y2),…,(x m ,y m )}, the output is {-1, +1}. In the AdaBoost model, decision trees are used as basic classifiers. Different weights are assigned to classifiers during training, and multiple weak classifiers are promoted to strong classifiers. The input sample weight of the kth weak learner in the training set is D(k) = (w k1 ,w k2 ,…,w km)}, the weight of the training samples of the next round of classifiers will be adjusted accordingly with the classification effect of the previous round of classifiers.

[0127] Through the above implementation, a credit risk detection model integrating multi-dimensional information is constructed. On the basis of traditional structured financial data, it focuses on the annotation data of financial statements and introduces unstructured news text data to analyze corporate credit risk from multiple information dimensions. Compared with the existing technology, it can detect corporate credit risk more comprehensively and accurately. In addition, in terms of the representation of unstructured text data, it integrates the advantages of pre-trained models and topic models, and considers the vectorized representation of text and the probability distribution of topics at the same time to enrich text information.

[0128] The following is a detailed description in conjunction with another embodiment.

[0129] Embodiment 2

[0130] A credit risk detection device provided in this embodiment includes multiple implementation units, each of which corresponds to each implementation step in the above-mentioned embodiment 1. Its specific implementation methods and beneficial effects can refer to the above-mentioned method embodiments and will not be repeated here.

[0131] Figure 5 is a schematic diagram of an optional credit risk detection device according to an embodiment of the present invention, such as Figure 3 As shown, the credit risk detection device may include: a data collection unit 51 , a data screening unit 52 , a data fusion unit 53 , and a risk assessment unit 54 .

[0132] The data collection unit 51 is used to collect structured financial data and unstructured news text data associated with the target customer in the credit business request after receiving the credit business request, and pre-process the structured financial data and the news text data.

[0133] The data screening unit 52 is used to screen the pre-processed financial data, output a set of financial indicators for credit risk detection, and use the Dirichlet polynomial mixture model to perform topic characterization on the pre-processed news text data, and output text topic probability distribution data of each news document.

[0134] The data fusion unit 53 is used to fuse the screened financial indicator set with the text topic probability distribution data after topic characterization to generate a detection information set containing structured financial information and unstructured text information.

[0135] The risk assessment unit 54 is used to input the detection information set into the credit risk detection model, and output a risk assessment report of the current credit business initiated by the target customer and a credit business approval suggestion.

[0136] The above-mentioned credit risk detection device can collect structured financial data and unstructured news text data associated with the target customer in the credit business request through the data collection unit 51 after receiving the credit business request, and pre-process the structured financial data and news text data, filter the pre-processed financial data through the data screening unit 52, output a set of financial indicators for credit risk detection, and use the Dirichlet polynomial mixed model to perform topic characterization on the pre-processed news text data, output text topic probability distribution data of each news document, fuse the screened financial indicator set with the text topic probability distribution data after topic characterization through the data fusion unit 53, generate a detection information set containing structured financial information and unstructured text information, input the detection information set into the credit risk detection model through the risk assessment unit 54, and output a risk assessment report on the current credit business initiated by the target customer and a credit business approval recommendation. In this embodiment, on the basis of traditional structured financial data, unstructured news text data can be introduced to analyze customer credit risk from multiple information dimensions, so that corporate credit risk detection can be performed more comprehensively and accurately. At the same time, in terms of the representation method of unstructured text data, the advantages of pre-trained models and topic models are integrated, and the vectorized representation of text and topic probability distribution are considered at the same time, text information is enriched, the accuracy of credit risk detection is improved, the risk prevention and control capabilities and market competitiveness of financial institutions are improved, and the property losses of financial institutions are reduced. The technical problem that financial institutions rely on a relatively single data source and collect incomplete information when conducting risk detection on customer credit business in related technologies is solved, which affects credit risk detection and easily causes property losses of financial institutions.

[0137] Optionally, the data collection unit includes: a financial report collection module, which is used to collect financial reports related to target customers, and screen out balance sheets, income statements and cash flow statements from the financial reports to obtain financial data tables; a financial data extraction module, which is used to extract multiple financial indicators and financial data corresponding to each financial indicator from the financial data table to obtain structured financial data, wherein the multiple financial indicators include at least one of the following: current ratio, quick ratio, total asset growth rate, gross profit margin, net profit margin, period expense ratio, cash cost ratio and financial expense cash ratio; a news report collection module, which is used to collect news reports related to target customers from financial news websites and financial engine databases, wherein the types of news reports include at least one of the following: corporate profit announcements, goodwill warnings, violation penalties and industry dynamic news; a news text extraction module, which is used to extract text data in news reports to obtain unstructured news text data.

[0138] Optionally, the data collection unit includes: a data preprocessing module, used to perform data cleaning and missing value processing on structured financial data, wherein data cleaning includes: deleting duplicate values ​​and deleting abnormal values, and missing value processing includes: filling missing data; performing cleaning processing on news text data, wherein the cleaning processing includes removing stop words, removing punctuation marks, removing numbers and special characters; a word segmentation module, used to perform word segmentation processing on the news text obtained after the cleaning processing to obtain a text word set and a text sentence set.

[0139] Optionally, the data screening unit includes: an IV value calculation module, which is used to calculate the IV value of each financial indicator in the structured financial data, screen out financial indicators with IV values ​​greater than or equal to a preset threshold, and obtain financial indicators associated with credit risk detection; a discretization module, which is used to discretize continuous financial data in the structured financial data, and convert continuous variables into discrete variables using a decision tree binning strategy; a feature screening module, which is used to use a chi-square test strategy to perform feature screening on the discretized financial data, remove features whose correlation with credit risk detection is lower than a preset correlation threshold, and output a set of financial indicators for credit risk detection.

[0140] Optionally, the data screening unit also includes: a word vectorization processing module, which is used to use a pre-trained Word2Vec model to perform word vectorization processing on the preprocessed news text and construct a word vector matrix; a topic extraction module, which is used to use a Dirichlet polynomial mixture model to extract topics from the word vector matrix and obtain the probability distribution of each news document on each topic.

[0141] Optionally, the Dirichlet polynomial mixture model is trained in the following manner: initializing model parameters, including the number of topics and hyperparameters of the Dirichlet distribution; obtaining preprocessed historical news text data, and randomly assigning each historical news document to a topic as the initial state of model training; a topic probability calculation step, iteratively calculating the probability that each historical news document belongs to each topic, wherein, when calculating the probability that each historical news document belongs to each topic, a Gibbs sampling algorithm is used for probability estimation based on the topic features of each document and model parameters; a probability redistribution step, reallocating each historical news document to a target topic according to the calculated probability; repeating the iterative topic probability calculation step and the probability redistribution step until the rate of change of the document topic allocation is lower than a preset threshold or the maximum number of iterations is reached, completing the convergence of the topic model, and obtaining a trained Dirichlet polynomial mixture model.

[0142] Optionally, the data fusion unit includes: a data merging module, which is used to merge the target customer's financial indicator set and the text topic probability distribution data into a detection information set based on the connection strategy of the data table, wherein in the detection information set, an independent column is allocated to each financial indicator in the financial indicator set and each text body in the text topic probability distribution data.

[0143] Optionally, the credit risk detection model is trained in the following manner: obtaining a historical detection information set including historical structured financial information and historical unstructured text information, and dividing the historical detection information set into a training set and a test set; using the Adaboost ensemble learning algorithm to construct an initial risk detection model; selecting a decision tree as a weak classifier, and assigning equal weights to each training sample in the training set; a classifier weight distribution adjustment step: according to the iterative process of the Adaboost algorithm, selecting a weak classifier in each round, and performing risk detection training based on the weight distribution of the training samples, wherein the weight of the weak classifier is adjusted according to the classification accuracy; a sample weight update step: after each round of weak classifier training is completed, updating the weight of the training sample, wherein when updating the weight of the training sample, the weight of the wrongly classified training sample is increased, and the weight of the correctly classified training sample is reduced; repeatedly executing the classifier weight distribution adjustment step and the sample weight update step until a preset number of weak classifiers is reached, and a strong classifier composed of multiple weak classifiers is obtained; using the test set to perform performance evaluation on the credit risk detection model, and when the performance evaluation result meets the evaluation accuracy, confirming that the training is completed, and obtaining the credit risk detection model.

[0144] The above-mentioned credit risk detection device may also include a processor and a memory. The above-mentioned data collection unit 51, data screening unit 52, data fusion unit 53, risk assessment unit 54, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize corresponding functions.

[0145] The processor includes a kernel, which retrieves the corresponding program unit from the memory. One or more kernels can be provided, and credit risk detection based on multi-dimensional information can be achieved by adjusting kernel parameters.

[0146] The above-mentioned memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (fl ash RAM), and the memory includes at least one storage chip.

[0147] Embodiment 3

[0148] An embodiment of the present application may provide an electronic device, Figure 6is a structural block diagram of an electronic device according to an embodiment of the present application. Figure 6 As shown, the electronic device may include: one or more ( Figure 6 (only one is shown) processor 602, memory 604, storage controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0149] Among them, the memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the credit risk detection method and device in the embodiment of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, realizing the above-mentioned credit risk detection method. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory may further include a memory remotely arranged relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0150] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: after receiving a credit business request, collect structured financial data and unstructured news text data associated with the target customer in the credit business request, and preprocess the structured financial data and news text data; screen the preprocessed financial data, output a set of financial indicators for credit risk detection, and use the Dirichlet polynomial mixed model to perform topic characterization on the preprocessed news text data, and output text topic probability distribution data of each news document; merge the screened financial indicator set with the text topic probability distribution data after topic characterization to generate a detection information set containing structured financial information and unstructured text information; input the detection information set into the credit risk detection model, and output a risk assessment report on the current credit business initiated by the target customer and a credit business approval recommendation.

[0151] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: collect financial reports related to the target customers, and screen out the balance sheet, income statement and cash flow statement from the financial reports to obtain a financial data table; extract multiple financial indicators and the financial data corresponding to each financial indicator from the financial data table to obtain structured financial data, wherein the multiple financial indicators include at least one of the following: current ratio, quick ratio, total asset growth rate, gross profit margin, net profit margin, period expense ratio, cash cost ratio and financial expense cash ratio; collect news reports related to the target customers from financial news websites and financial engine databases, wherein the types of news reports include at least one of the following: corporate profit announcements, goodwill warnings, violation penalties and industry dynamic news; extract text data from news reports to obtain unstructured news text data.

[0152] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: perform data cleaning and missing value processing on the structured financial data, wherein data cleaning includes: deleting duplicate values ​​and deleting abnormal values, and missing value processing includes: filling missing data; perform cleaning processing on the news text data, wherein the cleaning processing includes removing stop words, removing punctuation marks, removing numbers and special characters; perform word segmentation processing on the news text obtained after the cleaning processing to obtain a text word set and a text sentence set.

[0153] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: calculate the IV value of each financial indicator in the structured financial data, screen out the financial indicators with IV values ​​greater than or equal to the preset threshold, and obtain the financial indicators associated with credit risk detection; discretize the continuous financial data in the structured financial data, and use the decision tree binning strategy to convert the continuous variables into discrete variables; use the chi-square test strategy to perform feature screening on the discretized financial data, remove the features whose correlation with credit risk detection is lower than the preset correlation threshold, and output a set of financial indicators for credit risk detection.

[0154] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: use the pre-trained Word2Vec model to perform word vectorization on the pre-processed news text and construct a word vector matrix; use the Dirichlet polynomial mixture model to extract topics from the word vector matrix and obtain the probability distribution of each news document on each topic.

[0155] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: The Dirichlet polynomial mixture model is trained in the following manner: Initialize the model parameters, including the number of topics and the hyperparameters of the Dirichlet distribution; obtain the preprocessed historical news text data, and randomly assign each historical news document to a topic as the initial state of the model training; a topic probability calculation step, iteratively calculate the probability that each historical news document belongs to each topic, wherein, when calculating the probability that each historical news document belongs to each topic, the Gibbs sampling algorithm is used for probability estimation based on the topic features of each document and the model parameters; a probability redistribution step, reallocating each historical news document to the target topic according to the calculated probability; repeating the iterative topic probability calculation step and the probability redistribution step until the change rate of the document topic distribution is lower than the preset threshold or the maximum number of iterations is reached, the topic model converges, and the trained Dirichlet polynomial mixture model is obtained.

[0156] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: based on the connection strategy of the data table, merge the target customer's financial indicator set and the text topic probability distribution data into the detection information set, wherein in the detection information set, an independent column is allocated to each financial indicator in the financial indicator set and each text body in the text topic probability distribution data.

[0157] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: the credit risk detection model is trained in the following manner: a historical detection information set containing historical structured financial information and historical unstructured text information is obtained, and the historical detection information set is divided into a training set and a test set; an Adaboost ensemble learning algorithm is used to build an initial risk detection model; a decision tree is selected as a weak classifier, and an equal weight is assigned to each training sample in the training set; a classifier weight distribution adjustment step: according to the iterative process of the Adaboost algorithm, a weak classifier is selected in each round, based on the weight distribution of the training samples Conduct risk detection training, wherein the weight of the weak classifier is adjusted according to the classification accuracy; a sample weight updating step: after each round of weak classifier training is completed, the weight of the training sample is updated, wherein when updating the weight of the training sample, the weight of the wrongly classified training sample is increased, and the weight of the correctly classified training sample is reduced; the classifier weight distribution adjustment step and the sample weight updating step are repeatedly executed until a preset number of weak classifiers is reached, and a strong classifier composed of a plurality of weak classifiers is obtained; a test set is used to perform a performance evaluation on the credit risk detection model, and when the performance evaluation result meets the evaluation accuracy, the training is confirmed to be completed, and a credit risk detection model is obtained.

[0158] It can be understood by those skilled in the art that Figure 6 The structure shown is for illustration only, and the electronic device may also be a terminal device such as a smart phone, a tablet computer, a PDA, a mobile Internet device (Mobile Internet Devices, MID), a PAD, etc. Figure 6 The structure of the electronic device is not limited. Figure 6 More or fewer components (such as network interfaces, display devices, etc.) shown in, or having Figure 6 Different configurations shown.

[0159] A person skilled in the art can understand that all or part of the steps in the various credit risk detection methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a program, and the program can be stored in a computer-readable storage medium, and the storage medium can include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0160] Embodiment 4

[0161] The embodiment of the present application further provides a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the credit risk detection method provided in the first embodiment.

[0162] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium including a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute any one of the credit risk detection methods in the above-mentioned embodiment 1.

[0163] Optionally, in this embodiment, the above storage medium may be located in any computer terminal in a computer terminal group in a computer network, or in any mobile terminal in a mobile terminal group.

[0164] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the credit risk detection method described in each embodiment of the present application.

[0165] The present application also provides a computer program product, including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the credit risk detection method described in each embodiment of the present application are implemented.

[0166] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0167] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0168] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0169] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0170] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0171] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0172] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A credit risk detection method, characterized in that: include: After receiving a credit business request, collecting structured financial data and unstructured news text data associated with a target customer in the credit business request, and preprocessing the structured financial data and the news text data; Screening the pre-processed financial data, outputting a set of financial indicators for credit risk detection, and using a Dirichlet polynomial mixture model to perform topic characterization on the pre-processed news text data, and outputting text topic probability distribution data of each news document; The screened financial indicator set is integrated with the text topic probability distribution data after topic characterization to generate a detection information set including structured financial information and unstructured text information; The detection information set is input into a credit risk detection model, and a risk assessment report of the current credit business initiated by the target customer and a credit business approval recommendation are output.

2. The method according to claim 1, characterized in that: The step of screening the pre-processed financial data and outputting a set of financial indicators for credit risk detection includes: Calculating the IV value of each financial indicator in the structured financial data, screening out financial indicators with IV values ​​greater than or equal to a preset threshold, and obtaining financial indicators associated with credit risk detection; Discretizing the continuous financial data in the structured financial data, and converting the continuous variables into discrete variables using a decision tree binning strategy; The chi-square test strategy is used to perform feature screening on the discretized financial data, remove features whose correlation with credit risk detection is lower than a preset correlation threshold, and output the financial indicator set used for credit risk detection.

3. The method according to claim 1, characterized in that The step of using the Dirichlet polynomial mixture model to perform topic characterization on the pre-processed news text data includes: Use the pre-trained Word2Vec model to perform word vectorization on the pre-processed news text and construct a word vector matrix; The Dirichlet polynomial mixture model is used to extract topics from the word vector matrix to obtain the probability distribution of each news document on each topic.

4. The method according to claim 3, characterized in that: The Dirichlet polynomial mixture model is trained in the following way: Initialize model parameters, including the number of topics and hyperparameters of Dirichlet distribution; Obtain preprocessed historical news text data and randomly assign each historical news document to a topic as the initial state for model training; The topic probability calculation step is to iteratively calculate the probability that each historical news document belongs to each topic. When calculating the probability that each historical news document belongs to each topic, the Gibbs sampling algorithm is used to perform probability estimation based on the topic features of each document and model parameters. A probability redistribution step, redistributing each of the historical news documents to a target topic according to the calculated probability; Repeat the iterative topic probability calculation step and the probability redistribution step until the change rate of the document topic distribution is lower than a preset threshold or the maximum number of iterations is reached, the topic model converges, and the trained Dirichlet polynomial mixture model is obtained.

5. The method according to claim 1, characterized in that: The step of fusing the screened financial indicator set with the text topic probability distribution data after topic characterization to generate a detection information set containing structured financial information and unstructured text information includes: Based on the connection strategy of the data table, the financial indicator set of the target customer and the text topic probability distribution data are merged into the detection information set, wherein in the detection information set, an independent column is allocated to each financial indicator in the financial indicator set and each text body in the text topic probability distribution data.

6. The method according to claim 1, characterized in that The credit risk detection model is trained in the following way: Acquire a historical detection information set including historical structured financial information and historical unstructured text information, and divide the historical detection information set into a training set and a test set; The Adaboost ensemble learning algorithm is used to build the initial risk detection model; Selecting a decision tree as a weak classifier and assigning equal weight to each training sample in the training set; Classifier weight distribution adjustment step: According to the iterative process of the Adaboost algorithm, a weak classifier is selected in each round, and risk detection training is performed based on the weight distribution of the training samples, wherein the weight of the weak classifier is adjusted according to the classification accuracy; Sample weight updating step: after each round of weak classifier training is completed, the weight of the training sample is updated, wherein when the weight of the training sample is updated, the weight of the wrongly classified training sample is increased, and the weight of the correctly classified training sample is reduced; Repeating the classifier weight distribution adjustment step and the sample weight update step until a preset number of weak classifiers is reached, thereby obtaining a strong classifier composed of a plurality of weak classifiers; The test set is used to perform performance evaluation on the credit risk detection model. When the performance evaluation result meets the evaluation accuracy, the training is confirmed to be completed and the credit risk detection model is obtained.

7. The method according to claim 1, characterized in that The step of collecting structured financial data and unstructured news text data associated with the target customer in the credit business request includes: Collect financial reports related to the target customer, and select balance sheets, income statements, and cash flow statements from the financial reports to obtain financial data tables; Extracting multiple financial indicators and financial data corresponding to each financial indicator from the financial data table to obtain the structured financial data, wherein the multiple financial indicators include at least one of the following: current ratio, quick ratio, total asset growth rate, gross profit margin, net profit margin, period expense ratio, cash cost ratio and financial expense cash ratio; Collecting news reports related to the target customer from financial news websites and financial engine databases, wherein the types of the news reports include at least one of the following: corporate profit announcements, goodwill warnings, violation penalties, and industry dynamic news; The text data in the news report is extracted to obtain the unstructured news text data.

8. The method according to claim 1, characterized in that The step of preprocessing the structured financial data and the news text data includes: Performing data cleaning and missing value processing on the structured financial data, wherein the data cleaning includes: deleting duplicate values ​​and deleting abnormal values, and the missing value processing includes: filling missing data; Performing a cleaning process on the news text data, wherein the cleaning process includes removing stop words, removing punctuation marks, removing numbers and special characters; The news text obtained after the cleaning process is segmented to obtain a text word set and a text sentence set.

9. A credit risk detection device, characterized in that: include: A data collection unit, configured to, after receiving a credit business request, collect structured financial data and unstructured news text data associated with a target customer in the credit business request, and pre-process the structured financial data and the news text data; A data screening unit, used to screen the pre-processed financial data, output a set of financial indicators for credit risk detection, and use a Dirichlet polynomial mixture model to perform topic characterization on the pre-processed news text data, and output text topic probability distribution data of each news document; A data fusion unit, used to fuse the screened financial indicator set with the text topic probability distribution data after topic characterization to generate a detection information set containing structured financial information and unstructured text information; The risk assessment unit is used to input the detection information set into the credit risk detection model, and output a risk assessment report of the current credit business initiated by the target customer and a credit business approval suggestion.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the credit risk detection method according to any one of claims 1 to 8.

11. An electronic device, characterized in that: The invention comprises one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the credit risk detection method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the credit risk detection method according to any one of claims 1 to 8 are implemented.

Images