Internal compliance management system based on financial compliance large model

Abstract

The invention relates to the field of management systems, and discloses an internal compliance management system based on a financial compliance large model, which comprises an information import module, a compliance examination module, a risk monitoring module, an information consultation module and a training education module, the information import module is used for importing business application information submitted by business personnel; the compliance review module is used for performing compliance review on the service application information, and entering the next step after the compliance review is passed; the risk monitoring module is used for carrying out risk analysis on the business application information passing the compliance examination; the information consultation module is used for providing consultation service for business personnel; and the training education module is used for performing compliance education training for business personnel. According to the invention, more comprehensive and intelligent financial industry internal compliance management can be realized.

Description

Technical Field

[0001] The present invention relates to the field of management systems, and in particular to an internal compliance management system based on a financial compliance big model. Background Art

[0002] In today's complex, ever-changing and highly digitalized financial field, the traditional compliance management model is facing unprecedented severe challenges, which has become the key technical background for the birth of an internal compliance management system based on the financial compliance big model.

[0003] On the one hand, the financial regulatory environment is becoming increasingly stringent. Globally, financial regulators frequently issue and update a large number of regulations and policies to maintain financial market stability and protect the rights and interests of investors. Hundreds of new and revised financial regulations are added and revised every year. Their content not only covers various business links in traditional financial fields such as banking, securities, and insurance, such as market access conditions, business operation specifications, risk management indicators, etc., but also continues to expand to emerging financial technology fields with financial innovation. These regulations and policies are complicated and highly professional, and there are differences in requirements between different regions and different regulatory entities, which makes it difficult for financial institutions to interpret and implement them. If they are not careful, they may fall into compliance difficulties and face serious consequences such as huge fines, business restrictions, and even reputation damage.

[0004] On the other hand, the complexity of financial institutions' own businesses is growing exponentially. With the deep opening and integration of financial markets, financial products and service innovations are emerging in an endless stream, and cross-market and cross-field integrated financial business models are becoming more and more common. For example, a structured financial product may involve transactions and risk characteristics in multiple fields such as credit, securities, and derivatives at the same time. The internal business processes are intricate and involve the collaborative operations of multiple departments and multiple systems. This has led to an explosive growth in business data, with massive and diverse data types, including structured transaction data, semi-structured customer information, and unstructured contract texts, market research reports, etc., scattered across various business systems. Traditional data processing and compliance review methods are difficult to quickly and accurately mine potential compliance risk information from them.

[0005] Furthermore, the limitations of traditional compliance management methods are becoming increasingly prominent. In the past, the methods that relied on manual interpretation of regulations, manual review of business documents, and risk judgment based on experience were extremely inefficient. Faced with massive business and frequently updated regulations, timeliness and accuracy could not be guaranteed. Manual review is easily affected by subjective factors, resulting in negligence and misjudgment, and it is difficult to cope with complex and changing new compliance risk scenarios. At the same time, the compliance processes scattered across various departments lack effective integration, and information flow is not smooth, resulting in duplication of work and low coordination efficiency, which further weakens the ability of financial institutions to cope with compliance challenges. Therefore, an internal compliance management system based on a large financial compliance model is proposed. Summary of the invention

[0006] The technical problem to be solved by the present invention is: how to solve the problem of poor management effect of the existing internal compliance management system, and provide an internal compliance management system based on a financial compliance big model.

[0007] The present invention solves the above technical problems through the following technical solutions, which include an information import module, a compliance review module, a risk monitoring module, an information consulting module and a training and education module; The information import module is used to import the business application information submitted by the business personnel; The compliance review module is used to conduct compliance review on the business application information, and proceed to the next step after the compliance review is passed; The risk monitoring module is used to perform risk analysis on business application information that has passed the compliance review; The information consultation module is used to provide consulting services to business personnel; The training and education module is used to provide compliance education and training for business personnel.

[0008] Furthermore, when the business application information submitted by the business personnel is imported through the information import module, the type of the submitted business application information is analyzed, and different importing methods are selected according to the type of the submitted business application information to import the business application information.

[0009] Furthermore, the specific process of selecting different importing methods according to the type of submitted business application information to import the business application information is as follows: The types of business application information submitted include electronic version, paper print version and paper handwritten version; For electronic business application information, directly import it into the compliance review module for compliance review; For business application information in paper printed versions and paper handwritten versions, the paper printed versions and paper handwritten versions will be scanned and converted into electronic versions for compliance review.

[0010] Furthermore, the specific process of conducting compliance review after scanning the paper printed version and the paper handwritten version into an electronic version is as follows: Put the printed or handwritten version of the business application information into the adaptive pressure smoothing device to smooth the printed or handwritten version of the business application information. The pressure sensors in the adaptive pressure smoothing device are evenly distributed on the pressure plate of the device. When a wrinkled paper document is placed, the sensor can sense the pressure difference at different positions on the surface of the document in real time. The shape memory alloy material automatically adjusts its shape according to the pressure information fed back by the sensor, so that the pressure plate can apply uniform and appropriate pressure to the document, thereby smoothing the wrinkles to the greatest extent.

[0011] Workflow: Put the paper document into the device and the system starts. The pressure sensor starts to collect pressure data and transmits it to the control unit. The control unit analyzes the data according to the preset algorithm and sends instructions to the shape memory alloy. The alloy deforms and adjusts the shape of the pressure plate, and applies pressure to the document for a period of time (such as 5-10 minutes) until the sensor detects that the pressure is evenly distributed and considers that the document has been basically smoothed.

[0012] After smoothing is completed, use a multi-spectral scanning device to scan the smoothed paper print version or paper handwritten version of the business application information Technical architecture: A scanning system that integrates three light sources: visible light, infrared light, and ultraviolet light. Each light source has independent transmitting and receiving modules and can work alone or in combination. During the scanning process, different light sources can capture document image information from different angles and characteristics.

[0013] Working mode: Single light source mode: For different types of documents, such as handwritten documents, which may show ink penetration more clearly under infrared light, and printed documents, which can obtain regular images under visible light, you can select the appropriate light source for scanning.

[0014] Multi-light source fusion mode: For documents with severe wrinkles and blurred text, three light sources are turned on at the same time for scanning to obtain multiple sets of image data, which are then processed using an image fusion algorithm.

[0015] The image preprocessing includes wrinkle repair through a deep learning-based wrinkle repair algorithm as well as image enhancement and denoising: Wrinkle repair algorithm based on deep learning Model architecture: The Generative Adversarial Network (GAN) architecture is used, including a generator and a discriminator. The generator takes a wrinkled image as input and tries to generate a wrinkle-free restoration image; the discriminator is responsible for determining whether the input image is a real wrinkle-free image or an image generated by the generator.

[0016] Training data: A large number of paper document images of different types and different degrees of wrinkles are collected, and artificially restored or simulated wrinkle-free images are used as real samples to form a training data set.

[0017] Training process: During the training process, the generator and the discriminator are constantly engaged in adversarial training. The generator continuously adjusts parameters to make the generated repaired image closer and closer to the real wrinkle-free image in order to deceive the discriminator; the discriminator continuously improves its judgment ability. After multiple iterations of training, the generator can learn effective wrinkle repair rules.

[0018] Application stage: The scanned wrinkled image is input into the trained generator, and the generator outputs the repaired image.

[0019] Enhancement algorithm: A method combining histogram equalization and adaptive gamma correction is used. Histogram equalization can enhance the overall contrast of the image, making the text and background in the image clearer and more discernible; adaptive gamma correction adjusts the gamma value according to the local features of the image to further enhance the image's detail performance.

[0020] Denoising algorithm: Use the non-local mean filtering algorithm. This algorithm searches for similar pixel blocks in the image and uses the information of these similar blocks to estimate the true value of the current pixel, thereby effectively removing noise from the image while retaining the image's detailed information; After wrinkle repair, image enhancement and denoising are completed, text recognition is performed, two different ORC engines are selected, and the pre-processed images are input into the two OCR engines at the same time to obtain recognition results respectively; The two results are merged through voting mechanism and similarity analysis, and the same recognition content is directly retained; For different contents, calculate the semantic similarity between them and the surrounding texts, and select the result with higher similarity as the final recognition result; Then, context-aware correction is performed using pre-trained financial language models, such as the financial BERT model based on the Transformer architecture. This model has been pre-trained on a large amount of text data such as financial regulations and business documents, and can understand professional vocabulary and semantics in the financial field.

[0021] The OCR recognition results are input into the financial domain language model, and the model performs grammatical and semantic analysis on the recognition results based on the context information; For existing recognition errors, such as typos, missing words, and incoherent semantics, the model makes corrections based on the language habits and logic in the financial field. Once the proofreading is completed, the business application information is converted into an electronic paper print version or a paper handwritten version.

[0022] Furthermore, the specific process of the compliance review module for conducting compliance review on the business application information is as follows: First, perform data preprocessing to remove noise data in business application information, handle missing values, and standardize data; For numerical data, normalization is performed to scale the data to the [0,1] interval using the minimum-maximum normalization formula: , where x is the original data, x min and x max are the minimum and maximum values ​​of the data column respectively; For text data, perform lexical analysis and part-of-speech tagging to convert the text into a vector representation that can be processed by a computer, such as using the Bag-of-Words model or word embedding technology; After preprocessing, rule matching is performed. Compliance rules are stored in a structured manner in the database. Each rule contains a rule number, a rule description, a rule condition, and a rule result. For simple text rules, a method combining pattern matching and logical judgment is used (for example, to determine whether the loan amount A in the business application exceeds the maximum loan amount M). For complex text rules, natural language processing technology is used for semantic matching (for example, to determine whether the product description in the business application contains keywords that are prohibited from sale, a method combining keyword matching and word vector similarity calculation can be used). Identify factors that may lead to compliance risks from business application information, including customer credit rating, transaction amount and transaction frequency; The weighted summation method is used to quantitatively assess the risk. Each risk factor is assigned a weight Wi, and the risk score Si of each risk factor is determined based on historical data and expert experience. The total risk score R is calculated as follows: ; Where n is the number of risk factors. For example, assuming that the three risk factors of customer credit rating, transaction amount and transaction frequency are considered, the weights are w1=0.4, w2=0.3, w3=0.3, and the corresponding risk scores are s1=3, s2=2, s3=1, then the total risk score R=0.4×3+0.3×2+0.3×1=1.2+0.6+0.3=2.1; Risk levels are divided according to risk scores; for example: R<1: low risk, the review result is passed.

[0023] 1≤R<2: Medium risk, further review required.

[0024] R≥2: high risk, failed the review; Finally, an audit report is generated based on the rule matching results and risk quantification results. If the audit report shows that there are no abnormalities, the next step is carried out.

[0025] Furthermore, the specific process of the risk monitoring module performing risk analysis on the business application information that has passed the compliance review is as follows: Data access and integration are performed to obtain real-time and historical data related to business applications from multiple data sources, including the core business systems within financial institutions (such as transaction systems, customer relationship management systems) and external data providers (such as credit reporting agencies, market data platforms), etc. For example, for a loan application, the accessed data may include the customer's credit report, historical repayment records, current market interest rate fluctuations, etc.

[0026] Clean, convert and integrate data from different sources and formats, unify data formats and standards, and build a comprehensive business application information data set; Risk indicator definition and calculation: Construction of advanced risk indicator system: define risk indicators according to the characteristics and risk types of financial business; Risk indicators include credit risk indicators, market risk indicators and operational risk indicators; Indicator calculation method: different calculation methods and models are used to calculate risk indicators; Credit risk indicator calculation: Use machine learning algorithms such as logistic regression models and decision tree models to predict the probability of default based on the customer's historical data and characteristic variables. For example, the formula for the logistic regression model is: ; Where P(Y=1 / X) is the probability of default, X1, X2, …, Xn are the customer’s characteristic variables, and β0, β1, …, βn are the parameters of the model; Calculation of market risk indicators: Calculate VaR value using historical simulation method, Monte Carlo simulation method, etc.; The steps are as follows: collect market data over a historical period; Calculate the portfolio return for each historical period; Sort the yields from small to large and select the quantile at the preset level as the VaR value.

[0027] Risk model construction and training: Select risk model based on risk type and data characteristics; Model training: Use historical data to train and optimize the risk model, divide the data set into a training set and a test set, adjust the model parameters on the training set so that the similarity between the model's prediction results and the actual situation is greater than the preset value, and use cross-validation and grid search methods to select the optimal model parameters.

[0028] Real-time monitoring and early warning: Real-time data update: Continuously access real-time data, update business application information and risk indicators in real time, and use message queues (such as Kafka) and streaming computing frameworks (such as Flink) to achieve real-time data processing and update.

[0029] Risk threshold setting: Set a reasonable risk threshold for each risk indicator. When the risk indicator exceeds the threshold, the early warning mechanism is triggered; Warning rule definition: define warning rules according to different risk types and business scenarios. Warning rules include single indicator warning and multi-indicator combination warning; Warning notification and processing: When a warning is triggered, the system automatically sends a notification to relevant personnel; Risk analysis and assessment: Analyze the risk events that trigger early warnings to find out the causes and influencing factors that lead to the risks; Generate risk assessment reports regularly to evaluate and summarize the overall risk status of business applications. The report content includes the changing trend of risk indicators, the distribution of risk events, the effectiveness of risk response measures, etc.

[0030] Furthermore, the specific process of the information consulting module for providing consulting services to business personnel is as follows: When business personnel consult information, the content of the business personnel's consultation is identified and imported into the preset AI database for identification and reply. When the preset AI database cannot reply, it is transferred to manual reply; Mark the message sent by the business personnel as Hi, and the message replied by the system as Yu, where i is the number of messages sent by the business personnel, and u is the number of messages replied by the system; After the business personnel completes the consultation, the reply time difference between each business personnel sending Hi and the system reply message Yu is extracted to obtain i reply time differences; Process the i reply time differences to obtain reply evaluation parameters. When the reply evaluation parameters are abnormal, a prompt message is generated. The process of obtaining the response evaluation parameters and determining the abnormality is as follows: The average of the i response times is calculated, that is, the response evaluation parameter is obtained. When the response evaluation parameter is greater than the preset value, it indicates that there is an abnormality.

[0031] Furthermore, the specific process of the training and education module for conducting compliance education and training for business personnel is as follows: Randomly select a preset number of business personnel to send test questions for testing and obtain the test results; Business personnel whose test result scores are less than the preset value are selected as personnel to be trained; After multiple tests, when the number of people to be trained is greater than the preset value, compliance education training information will be sent to all people to be trained to prompt them to receive training.

[0032] Compared with the prior art, the present invention has the following advantages: the internal compliance management system based on the financial compliance big model can improve the accuracy and efficiency of compliance review, and the compliance review module uses advanced data preprocessing, rule matching and risk quantification technology to conduct a comprehensive analysis of business application information. Whether it is a complex text rule or a numerical business indicator, it can accurately determine whether it is compliant. For example, when reviewing a loan business application, it can accurately compare key factors such as customer credit rating and transaction amount with preset rules to avoid subjective misjudgment of manual review, greatly improve the accuracy of the review, and effectively reduce the risk of illegal business passing.

[0033] Efficient processing: Through automated processes and intelligent algorithms, the import, analysis and review of business application information can be completed quickly. Compared with traditional manual review, the process that may have taken hours or even days can now be completed in minutes, greatly shortening the business waiting time, accelerating the business operation of financial institutions, and enhancing market competitiveness.

[0034] The risk monitoring module integrates multi-source data, builds a complete risk indicator system, and uses machine learning, time series and other models to monitor risks from multiple dimensions such as credit, market, and operation. For example, the logistic regression model for predicting the probability of default can provide early insight into the potential default risk of customers; the historical simulation method for calculating the VaR value can accurately grasp the market volatility risk and provide financial institutions with a comprehensive and in-depth risk perspective.

[0035] Real-time warning response: With the help of message queues and streaming computing frameworks, business application information and risk indicators are updated in real time. Once the risk indicator reaches the preset threshold, a warning notification is immediately sent to relevant personnel through SMS, email, system pop-up windows, etc., to ensure that risk events can be discovered and handled in a timely manner at the embryonic stage, minimizing potential losses.

[0036] The information consultation module uses the financial compliance big model to deeply understand the business personnel's consultation content, quickly search the knowledge map, regulations and policy database and case library, and generate accurate answers. Most common questions can be answered instantly, saving business personnel time to search for information and ask experts, and improving work efficiency.

[0037] By marking the interaction messages between business personnel and the system, calculating the response time difference and evaluating the response situation, abnormalities in consulting services can be discovered in a timely manner. If the average response time is too long or the response fluctuations are too large, the problem can be quickly identified, the service process can be optimized, and the satisfaction of consulting services can be continuously improved.

[0038] The training and education module accurately screens out personnel and key knowledge areas that need training based on compliance knowledge testing, behavior assessment, and business development needs analysis of business personnel. This avoids indiscriminate training for all personnel, concentrates limited resources on improving weak links, and improves the input-output ratio of training.

[0039] For paper-printed and handwritten business application documents, a series of innovative technologies such as adaptive pressure smoothing, multi-spectral scanning, deep learning image restoration, hybrid OCR engine recognition, etc., paper documents are successfully converted into high-quality electronic versions, ensuring the smooth progress of subsequent compliance reviews, breaking the bottleneck of paper document processing, and realizing full-process electronic management, making the system more worthy of promotion and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0041] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0042] like Figure 1 As shown, this embodiment provides a technical solution: an internal compliance management system based on a financial compliance big model, including an information import module, a compliance review module, a risk monitoring module, an information consulting module and a training and education module; The information import module is used to import the business application information submitted by the business personnel; The compliance review module is used to conduct compliance review on the business application information, and proceed to the next step after the compliance review is passed; The risk monitoring module is used to perform risk analysis on business application information that has passed the compliance review; The information consultation module is used to provide consulting services to business personnel; The training and education module is used to provide compliance education and training for business personnel.

[0043] When the business application information submitted by the business personnel is imported through the information import module, the type of the submitted business application information is analyzed, and different import methods are selected according to the type of the submitted business application information to import the business application information.

[0044] The specific process of selecting different importing methods to import the business application information according to the type of the submitted business application information is as follows: The types of business application information submitted include electronic version, paper print version and paper handwritten version; For electronic business application information, directly import it into the compliance review module for compliance review; For business application information in paper printed versions and paper handwritten versions, the paper printed versions and paper handwritten versions will be scanned and converted into electronic versions for compliance review.

[0045] The specific process of conducting compliance review after scanning the paper printed version and the paper handwritten version into an electronic version is as follows: Put the printed or handwritten version of the business application information into the adaptive pressure smoothing device to smooth the printed or handwritten version of the business application information. The pressure sensors in the adaptive pressure smoothing device are evenly distributed on the pressure plate of the device. When a wrinkled paper document is placed, the sensor can sense the pressure difference at different positions on the surface of the document in real time. The shape memory alloy material automatically adjusts its shape according to the pressure information fed back by the sensor, so that the pressure plate can apply uniform and appropriate pressure to the document, thereby smoothing the wrinkles to the greatest extent.

[0046] Workflow: Put the paper document into the device and the system starts. The pressure sensor starts to collect pressure data and transmits it to the control unit. The control unit analyzes the data according to the preset algorithm and sends instructions to the shape memory alloy. The alloy deforms and adjusts the shape of the pressure plate, and applies pressure to the document for a period of time (such as 5-10 minutes) until the sensor detects that the pressure is evenly distributed and considers that the document has been basically smoothed.

[0047] After smoothing is completed, use a multi-spectral scanning device to scan the smoothed paper print version or paper handwritten version of the business application information Technical architecture: A scanning system that integrates three light sources: visible light, infrared light, and ultraviolet light. Each light source has independent transmitting and receiving modules and can work alone or in combination. During the scanning process, different light sources can capture document image information from different angles and characteristics.

[0048] Working mode: Single light source mode: For different types of documents, such as handwritten documents, which may show ink penetration more clearly under infrared light, and printed documents, which can obtain regular images under visible light, you can select the appropriate light source for scanning.

[0049] Multi-light source fusion mode: For documents with severe wrinkles and blurred text, three light sources are turned on at the same time for scanning to obtain multiple sets of image data, which are then processed using an image fusion algorithm.

[0050] The image preprocessing includes wrinkle repair through a deep learning-based wrinkle repair algorithm as well as image enhancement and denoising: Wrinkle repair algorithm based on deep learning Model architecture: The Generative Adversarial Network (GAN) architecture is used, including a generator and a discriminator. The generator takes a wrinkled image as input and tries to generate a wrinkle-free restoration image; the discriminator is responsible for determining whether the input image is a real wrinkle-free image or an image generated by the generator.

[0051] Training data: A large number of paper document images of different types and different degrees of wrinkles are collected, and artificially restored or simulated wrinkle-free images are used as real samples to form a training data set.

[0052] Training process: During the training process, the generator and the discriminator are constantly engaged in adversarial training. The generator continuously adjusts parameters to make the generated repaired image closer and closer to the real wrinkle-free image in order to deceive the discriminator; the discriminator continuously improves its judgment ability. After multiple iterations of training, the generator can learn effective wrinkle repair rules.

[0053] Application stage: The scanned wrinkled image is input into the trained generator, and the generator outputs the repaired image.

[0054] Enhancement algorithm: A method combining histogram equalization and adaptive gamma correction is used. Histogram equalization can enhance the overall contrast of the image, making the text and background in the image clearer and more discernible; adaptive gamma correction adjusts the gamma value according to the local features of the image to further enhance the image's detail performance.

[0055] Denoising algorithm: Use the non-local mean filtering algorithm. This algorithm searches for similar pixel blocks in the image and uses the information of these similar blocks to estimate the true value of the current pixel, thereby effectively removing noise from the image while retaining the image's detailed information; After wrinkle repair, image enhancement and denoising are completed, text recognition is performed, two different ORC engines are selected, and the pre-processed images are input into the two OCR engines at the same time to obtain recognition results respectively; The two results are merged through voting mechanism and similarity analysis, and the same recognition content is directly retained; For different contents, calculate the semantic similarity between them and the surrounding texts, and select the result with higher similarity as the final recognition result; Then, context-aware correction is performed using pre-trained financial language models, such as the financial BERT model based on the Transformer architecture. This model has been pre-trained on a large amount of text data such as financial regulations and business documents, and can understand professional vocabulary and semantics in the financial field.

[0056] The OCR recognition results are input into the financial domain language model, and the model performs grammatical and semantic analysis on the recognition results based on the context information; For existing recognition errors, such as typos, missing words, and incoherent semantics, the model makes corrections based on the language habits and logic in the financial field. Once the proofreading is completed, the business application information is converted into an electronic paper print version or a paper handwritten version.

[0057] The specific process of the compliance review module for conducting compliance review of business application information is as follows: First, perform data preprocessing to remove noise data in business application information, handle missing values, and standardize data; For numerical data, normalization is performed to scale the data to the [0,1] interval using the minimum-maximum normalization formula: , where x is the original data, x min and x max are the minimum and maximum values ​​of the data column respectively; For text data, perform lexical analysis and part-of-speech tagging to convert the text into a vector representation that can be processed by a computer, such as using the Bag-of-Words model or word embedding technology; After preprocessing, rule matching is performed. Compliance rules are stored in a structured manner in the database. Each rule contains a rule number, a rule description, a rule condition, and a rule result. For simple text rules, a method combining pattern matching and logical judgment is used (for example, to determine whether the loan amount A in the business application exceeds the maximum loan amount M). For complex text rules, natural language processing technology is used for semantic matching (for example, to determine whether the product description in the business application contains keywords that are prohibited from sale, a method combining keyword matching and word vector similarity calculation can be used). Identify factors that may lead to compliance risks from business application information, including customer credit rating, transaction amount and transaction frequency; The weighted summation method is used to quantitatively assess the risk. Each risk factor is assigned a weight Wi, and the risk score Si of each risk factor is determined based on historical data and expert experience. The total risk score R is calculated as follows: ; Where n is the number of risk factors. For example, assuming that the three risk factors of customer credit rating, transaction amount and transaction frequency are considered, the weights are w1=0.4, w2=0.3, w3=0.3, and the corresponding risk scores are s1=3, s2=2, s3=1, then the total risk score R=0.4×3+0.3×2+0.3×1=1.2+0.6+0.3=2.1; Risk levels are divided according to risk scores; for example: R<1: low risk, the review result is passed.

[0058] 1≤R<2: Medium risk, further review required.

[0059] R≥2: high risk, failed the review; Finally, an audit report is generated based on the rule matching results and risk quantification results. If the audit report shows that there are no abnormalities, the next step is carried out.

[0060] Furthermore, the specific process of the risk monitoring module performing risk analysis on the business application information that has passed the compliance review is as follows: Data access and integration are performed to obtain real-time and historical data related to business applications from multiple data sources, including the core business systems within financial institutions (such as transaction systems, customer relationship management systems) and external data providers (such as credit reporting agencies, market data platforms), etc. For example, for a loan application, the accessed data may include the customer's credit report, historical repayment records, current market interest rate fluctuations, etc.

[0061] Clean, convert and integrate data from different sources and formats, unify data formats and standards, and build a comprehensive business application information data set; Risk indicator definition and calculation: Construction of advanced risk indicator system: define risk indicators according to the characteristics and risk types of financial business; Risk indicators include credit risk indicators (such as default probability, default loss rate), market risk indicators (such as volatility, VaR value) and operational risk indicators (such as business process error rate). For example, for credit risk, the default probability can be calculated based on factors such as the customer's credit score and financial status.

[0062] Indicator calculation method: different calculation methods and models are used to calculate risk indicators; Credit risk indicator calculation: Use machine learning algorithms such as logistic regression models and decision tree models to predict the probability of default based on the customer's historical data and characteristic variables. For example, the formula for the logistic regression model is: ; Where P(Y=1 / X) is the probability of default, X1, X2, …, Xn are the customer’s characteristic variables (such as age, income, debt, etc.), β0, β1, …, βn are the parameters of the model; Calculation of market risk indicators: Calculate VaR value using historical simulation method, Monte Carlo simulation method, etc.; The steps are as follows: collect market data (such as stock prices, interest rates, etc.) over a historical period; Calculate the portfolio return for each historical period; Sort the yields from small to large and select the quantile at the preset level as the VaR value.

[0063] Risk model construction and training: Select a risk model based on the risk type and data characteristics. In addition to the machine learning models mentioned above, you can also use time series models (such as ARIMA, GARCH) to predict market risks, and use neural network models (such as multi-layer perceptron, recurrent neural network) for complex risk prediction.

[0064] Model training: Use historical data to train and optimize the risk model, divide the data set into a training set and a test set, adjust the model parameters on the training set so that the model's prediction results are more similar to the actual situation than the preset value, and use cross-validation and grid search methods to select the optimal model parameters. For example, for a neural network model, the stochastic gradient descent algorithm can be used to update the model's weights and biases to minimize the prediction error.

[0065] Real-time monitoring and early warning: Real-time data update: Continuously access real-time data, update business application information and risk indicators in real time, and use message queues (such as Kafka) and streaming computing frameworks (such as Flink) to achieve real-time data processing and update.

[0066] Risk threshold setting: Set a reasonable risk threshold for each risk indicator. When the risk indicator exceeds the threshold, the early warning mechanism is triggered. For example, the threshold of the probability of default is set to 5%. When the calculated probability of default of a customer exceeds 5%, an early warning signal is issued. Definition of warning rules: Define warning rules based on different risk types and business scenarios. Warning rules include single indicator warning and multi-indicator combination warning. For example, when the customer's default probability exceeds the threshold and the transaction amount exceeds a certain amount, a higher level warning is issued; Warning notification and processing: When a warning is triggered, the system automatically sends a notification to relevant personnel, such as risk management personnel and business personnel, through SMS, email, system pop-up windows, etc. At the same time, the system provides detailed warning information, including risk indicators, warning reasons, related business information, etc., to facilitate timely processing by relevant personnel.

[0067] Risk Analysis and Assessment: Analyze the risk events that trigger early warnings to find out the causes and influencing factors that lead to the occurrence of risks; use causal analysis methods (such as Granger causality test) and sensitivity analysis methods to determine the impact of various factors on risks. For example, analyze whether the reason for the increase in the probability of customer default is due to changes in the economic environment, the deterioration of the customer's own financial situation, or business process loopholes.

[0068] Generate risk assessment reports regularly to evaluate and summarize the overall risk status of business applications. The report content includes the changing trend of risk indicators, the distribution of risk events, the effectiveness of risk response measures, etc. Use visualization tools (such as Tableau, PowerBI) to display the report content in the form of charts and graphs to facilitate management decision-making.

[0069] The specific process of the information consulting module for providing consulting services to business personnel is as follows: When business personnel consult information, the content of the business personnel's consultation is identified and imported into the preset AI database for identification and reply. When the preset AI database cannot reply, it is transferred to manual reply; Mark the message sent by the business personnel as Hi, and the message replied by the system as Yu, where i is the number of messages sent by the business personnel, and u is the number of messages replied by the system; After the business personnel completes the consultation, the reply time difference between each business personnel sending Hi and the system reply message Yu is extracted to obtain i reply time differences; Process the i reply time differences to obtain reply evaluation parameters. When the reply evaluation parameters are abnormal, a prompt message is generated. The process of obtaining the response evaluation parameters and determining the abnormality is as follows: The average of the i response times is calculated, that is, the response evaluation parameter is obtained. When the response evaluation parameter is greater than the preset value, it indicates that there is an abnormality.

[0070] The specific process of compliance education and training for business personnel in the training and education module is as follows: Randomly select a preset number of business personnel to send test questions for testing and obtain the test results; Business personnel whose test result scores are less than the preset value are selected as personnel to be trained; After multiple tests, when the number of people to be trained is greater than the preset value, compliance education training information will be sent to all people to be trained to prompt them to receive training.

[0071] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0072] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0073] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. An internal compliance management system based on a financial compliance big model, characterized in that: It includes information import module, compliance review module, risk monitoring module, information consultation module and training and education module; The information import module is used to import the business application information submitted by the business personnel; The compliance review module is used to conduct compliance review on business application information; The risk monitoring module is used to perform risk analysis on business application information that has passed the compliance review; The information consultation module is used to provide consulting services to business personnel; The training and education module is used to provide compliance education and training for business personnel.

2. According to claim 1, an internal compliance management system based on a financial compliance big model is characterized by: When the business application information submitted by the business personnel is imported through the information import module, the type of the submitted business application information is analyzed, and different import methods are selected according to the type of the submitted business application information to import the business application information.

3. The internal compliance management system based on the financial compliance big model according to claim 2 is characterized by: The specific process of selecting different import methods to import business application information according to the type of business application information submitted is as follows: The types of business application information submitted include electronic version, paper print version and paper handwritten version; For electronic business application information, directly import it into the compliance review module for compliance review; For business application information in paper printed versions and paper handwritten versions, the paper printed versions and paper handwritten versions will be scanned and converted into electronic versions for compliance review.

4. The internal compliance management system based on the financial compliance big model according to claim 3 is characterized by: The specific process of conducting compliance review after scanning the paper printed version and the paper handwritten version into an electronic version is as follows: Putting the paper printed version or the paper handwritten version of the business application information into the adaptive pressure smoothing device to smooth the paper printed version or the paper handwritten version of the business application information; After smoothing is completed, use a multi-spectral scanning device to scan the smoothed paper print version or paper handwritten version of the business application information; The image preprocessing includes wrinkle repair through a deep learning-based wrinkle repair algorithm as well as image enhancement and denoising: After wrinkle repair, image enhancement and denoising are completed, text recognition is performed, two different ORC engines are selected, and the pre-processed images are input into the two OCR engines at the same time to obtain recognition results respectively; The two results are merged through voting mechanism and similarity analysis, and the same recognition content is directly retained; For different contents, calculate the semantic similarity between them and the surrounding texts, and select the result with higher similarity as the final recognition result; Then perform context-aware correction, using a pre-trained language model for the financial domain; The OCR recognition results are input into the financial domain language model, and the model performs grammatical and semantic analysis on the recognition results based on the context information; For existing recognition errors, the model makes corrections based on the language habits and logic of the financial field; Once the proofreading is completed, the business application information is converted into an electronic paper print version or a paper handwritten version.

5. The internal compliance management system based on the financial compliance big model according to claim 1 is characterized by: The specific process of the compliance review module for conducting compliance review of business application information is as follows: First, perform data preprocessing to remove noise data in business application information, handle missing values, and standardize data; For numerical data, normalization is performed to scale the data to the [0,1] interval using the minimum-maximum normalization formula: , where x is the original data, x min and x max are the minimum and maximum values ​​of the data column respectively; For text data, perform lexical analysis and part-of-speech tagging to convert the text into a vector representation that can be processed by a computer; After preprocessing, rule matching is performed. Compliance rules are stored in a structured manner in the database. Each rule contains the rule number, rule description, rule condition, and rule result. For simple text rules, a combination of pattern matching and logical judgment is used. For complex text rules, natural language processing technology is used for semantic matching. Identify factors that may lead to compliance risks from business application information, including customer credit rating, transaction amount and transaction frequency; The weighted summation method is used to quantitatively assess the risk, assign a weight Wi to each risk factor, and determine the risk score Si of each risk factor based on historical data and expert experience. The total risk score R is calculated as follows: ; in n is the number of risk factors; Classify risk levels based on risk scores; Finally, an audit report is generated based on the rule matching results and risk quantification results. If the audit report shows that there are no abnormalities, the next step is carried out.

6. The internal compliance management system based on the financial compliance big model according to claim 1 is characterized by: The specific process of the risk monitoring module performing risk analysis on the business application information that has passed the compliance review is as follows: Access and integrate data to obtain real-time and historical data related to business applications from multiple data sources, including the core business systems within financial institutions and external data providers; Clean, convert and integrate data from different sources and formats, unify data formats and standards, and build business application information data sets; Risk indicator definition and calculation: Construction of advanced risk indicator system: define risk indicators according to the characteristics and risk types of financial business; Risk indicators include credit risk indicators, market risk indicators and operational risk indicators; Indicator calculation method: different calculation methods and models are used to calculate risk indicators; Credit risk indicator calculation: Use a logistic regression model or decision tree model machine learning algorithm to predict the probability of default based on the customer's historical data and characteristic variables. The formula for the logistic regression model is: ； Where P(Y=1 / X) is the probability of default, X1, X2, …, Xn are the customer’s characteristic variables, and β0, β1, …, βn are the parameters of the model; Calculation of market risk indicators: Calculate VaR value using historical simulation method or Monte Carlo simulation method; The steps are as follows: collect market data over a historical period; Calculate the portfolio return for each historical period; Sort the yields from small to large, and select the quantile at the preset level as the VaR value; Risk model construction and training: Select risk model based on risk type and data characteristics; Model training: Use historical data to train and optimize the risk model, divide the data set into a training set and a test set, adjust the model parameters on the training set so that the similarity between the model's prediction results and the actual situation is greater than the preset value, and use cross-validation and grid search methods to select the optimal model parameters; Real-time monitoring and early warning: Real-time data update: Continuously access real-time data, update business application information and risk indicators in real time, and use message queues and streaming computing frameworks to achieve real-time data processing and update; Risk threshold setting: Set a reasonable risk threshold for each risk indicator. When the risk indicator exceeds the threshold, the early warning mechanism is triggered; Warning rule definition: define warning rules according to different risk types and business scenarios. Warning rules include single indicator warning and multi-indicator combination warning; Warning notification and processing: When a warning is triggered, the system automatically sends a notification to relevant personnel; Risk Analysis and Assessment: Analyze the risk events that trigger early warnings to find out the causes and influencing factors that lead to the risks; Generate risk assessment reports regularly to evaluate and summarize the overall risk status of business applications. The report content includes the changing trends of risk indicators, the distribution of risk events, and the effectiveness of risk response measures.

7. The internal compliance management system based on the financial compliance big model according to claim 1 is characterized by: The specific process of the information consulting module for providing consulting services to business personnel is as follows: When business personnel consult information, the content of the business personnel's consultation is identified and imported into the preset AI database for identification and reply. When the preset AI database cannot reply, it is transferred to manual reply; Mark the message sent by the business personnel as Hi, and the message replied by the system as Yu, where i is the number of messages sent by the business personnel, and u is the number of messages replied by the system; After the business personnel completes the consultation, the reply time difference between each business personnel sending Hi and the system reply message Yu is extracted to obtain i reply time differences; Process the i reply time differences to obtain reply evaluation parameters. When the reply evaluation parameters are abnormal, a prompt message is generated. The process of obtaining the response evaluation parameters and determining the abnormality is as follows: The average of the i response times is calculated, that is, the response evaluation parameter is obtained. When the response evaluation parameter is greater than the preset value, it indicates that there is an abnormality.

8. The internal compliance management system based on the financial compliance big model according to claim 1 is characterized by: The specific process of compliance education and training for business personnel in the training and education module is as follows: Randomly select a preset number of business personnel to send test questions for testing and obtain the test results; Business personnel whose test result scores are less than the preset value are selected as personnel to be trained; After multiple tests, when the number of people to be trained is greater than the preset value, compliance education training information will be sent to all people to be trained to prompt them to receive training.

Images