Intelligent customer risk assessment system and method based on large language model

Abstract

The invention provides an intelligent customer risk assessment system and method based on a large language model, and relates to the technical field of risk assessment, and the method comprises the steps: obtaining multi-modal data of a customer, carrying out the preprocessing, and extracting structured and unstructured features; constructing a hierarchical risk knowledge system, and realizing adaptive evolution of the knowledge system through a graph neural network and a generative model; constructing an initial negative sample library, and constructing a negative sample database in combination with a non-risk mode labeled by an expert and derivative layer analysis; optimizing the large language model by adopting a strong supervision, weak supervision and reinforcement learning cooperative training mechanism under each classification according to the customer type; mining risk features in a text by using the optimized large language model, processing multi-modal data through a multi-level attention network, and generating a positioning report including contradiction type coding, service influence dimension evaluation and risk level quantification; the accuracy, efficiency and flexibility of customer risk assessment are improved, and the risk management strategy is optimized.

Description

Technical Field

[0001] The present application relates to the field of risk assessment technology, and in particular to a customer risk intelligent assessment system and method based on a large language model. Background Art

[0002] In modern financial services, customer risk assessment is an essential component of the sound operation of financial institutions. It directly impacts the accuracy of credit approvals, the effectiveness of risk control, and the overall security and sustainability of the business. However, traditional customer risk assessment methods primarily rely on structured information such as financial data and credit scores. While these methods can provide a certain degree of insight into a customer's creditworthiness, they overlook the rich risk clues contained in unstructured data such as customer statements.

[0003] Customers' expressions in various interactions, such as credit applications, customer service communications, and online feedback, often reveal key risk information, including their business performance, willingness to repay, and adaptability to industry trends. For example, a customer might optimistically forecast profits in a credit application, only to reveal cost control challenges in subsequent customer service communications. Or, a customer might frequently discuss industry trends on social media, but their business transformation plans might not align with market trends or company resources. The risk information contained in this unstructured data is crucial for a comprehensive and accurate assessment of customer risk.

[0004] However, unstructured data is massive, multi-source, and heterogeneous, and includes various forms such as text and speech-to-text transcription. Conventional data processing processes often face difficulties in data integration and feature extraction when faced with this data. On the one hand, due to the wide range of data sources and different formats, how to effectively integrate this data becomes a major challenge; on the other hand, the risk information in unstructured data is often implicit in complex contexts and logical relationships. How to extract valuable feature information from it places extremely high demands on data processing technology; traditional risk assessment methods are often unable to cope with unstructured data, resulting in the loss of a large amount of key risk information; not only does it reduce the comprehensiveness and accuracy of risk assessments, it may also catch financial institutions off guard when facing potential risks, thereby affecting their stable operations and business development.

[0005] In terms of risk knowledge system construction, the traditional hierarchical risk knowledge system is relatively fixed and static, and it is difficult to adapt to the rapidly changing market environment and the endless emergence of new risks; when new risk patterns emerge, they cannot be incorporated into the knowledge system in a timely and accurate manner and reasonably classified and evaluated, which greatly reduces the timeliness and accuracy of risk assessment. Summary of the Invention

[0006] Based on the above problems, this application proposes a customer risk intelligent assessment system and method based on a large language model. Through multimodal data integration and a layered risk knowledge system, combined with an adaptive learning mechanism, it significantly improves the accuracy, efficiency and flexibility of customer risk assessment and optimizes risk management strategies.

[0007] The purpose of this application is achieved by the following technical solutions: In a first aspect, the present application provides a method for intelligently assessing customer risk based on a large language model, the method comprising: S1. Obtain and preprocess the customer's multimodal data to extract structured and unstructured features; the multimodal data includes structured business data, unstructured text description data, business scenario images, and multiple rounds of voice interaction data; S2. Construct a layered risk knowledge system and implement adaptive evolution of the knowledge system through graph neural networks and generative models. The layered risk knowledge system includes a core layer, a scenario layer, a derivative layer, and an archive layer. A risk pattern vector library is established based on multiple anomaly types, including logical fallacies, contradictory expression paradigms, and abnormal semantics. S3. Randomly sample multimodal data that meets the preset conditions to build an initial negative sample library. Combined with expert-annotated non-risk patterns and derivative layer analysis, a negative sample database is constructed. S4. Classify and label risk text data by customer type, and optimize the large language model using a collaborative training mechanism of strong supervision, weak supervision, and reinforcement learning for each category; S5. Use the optimized large language model to mine risk features in text, process multimodal data through a multi-level attention network, locate logical contradictions through intra-modal enhancement, cross-modal fusion, and contradiction detection, and generate a positioning report that includes contradiction type coding, business impact dimension assessment, and risk level quantification.

[0008] Preferably, the S2 includes: Perform cluster analysis on unlabeled text to obtain new risk patterns; Calculate the risk confidence of the new risk model based on the maximum similarity between the new risk model and each historical risk model in the core layer and the frequency of occurrence of the new risk model, and determine the storage level of the new risk model data; Integrate multi-source data, including expert feedback, model discovery, policy analysis, and public opinion monitoring, to automatically trigger updates through NLP and time series analysis. Integrate external knowledge bases through semantic fusion and transfer learning, and combine them with generative large language models to expand and optimize the knowledge base content.

[0009] Preferably, the step of calculating the risk confidence of the new risk model based on the maximum similarity between the new risk model and each historical risk model of the core layer and the frequency of occurrence of the new risk model, and determining the storage level of the new risk model data, includes: If the maximum similarity is less than the similarity threshold, then calculating the abnormality score of the new risk pattern; The abnormality score of the new risk pattern is calculated by using the new risk pattern vector, the mean and standard deviation of multiple historical pattern vectors in the core layer, and the frequency of occurrence of the new risk pattern. The abnormality score is obtained by the following formula:

[0010] in, is the characteristic vector of the new risk pattern; is the mean of multiple historical pattern vectors in the core layer; is the standard deviation of the historical pattern eigenvector; To quantify the difference between the new pattern feature vector and the mean of the historical pattern vector by the square of the Euclidean distance; is the risk pattern frequency; is the frequency reference value; If the abnormality score is greater than or equal to the first score threshold, it is pushed to the expert annotation queue; the risk pattern data annotated by the expert is stored in the core layer, and the non-risk pattern data is stored in the negative sample library; If the abnormality score is less than the first score threshold, the new risk pattern data is stored in the derivative layer; If the maximum similarity is greater than or equal to the similarity threshold, the corresponding risk knowledge system layer is selected and stored according to the risk confidence level; including: If the risk confidence of the new risk model is greater than or equal to the first confidence threshold, the data of the risk model is stored in the core layer; If the risk confidence of the new risk model is greater than or equal to the second confidence threshold and less than the first confidence threshold, the data of the risk model is stored in the scenario layer; If the risk confidence of the new risk model is lower than the second confidence threshold, the data of the new risk model is stored in the derivative layer; When the frequency of risk patterns in the derivative layer or scenario layer continues to increase for multiple statistical periods, its abnormality score and risk confidence level are recalculated and migrated to the core layer or scenario layer after meeting the standards.

[0011] Preferably, the method for obtaining the risk confidence level includes: Extract historical risk pattern feature vectors and new risk pattern feature vectors through pre-trained semantic encoders; calculate the maximum similarity between the new risk pattern and each historical risk pattern feature vector as the basic confidence level; Determine the final risk confidence level through basic confidence level, risk pattern frequency and policy correlation coefficient; Among them, the final risk confidence is:

[0012] in, is the final risk confidence level, is the basic confidence level, ∈[0,1], is the risk pattern frequency; is the frequency reference value; is the policy relevance coefficient, ∈[0.7, 1.6]; min() is the minimum value; is the risk confidence of the historical risk pattern that is most similar to the new risk pattern.

[0013] Preferably, the S4 includes: Perform main task training based on manually labeled samples in the core layer; The weakly supervised training includes positive samples and negative samples, including: The new risk pattern samples in the core layer are used as positive samples, and positive and negative sample pairs for comparative learning are constructed based on the positive samples; Use non-risk patterns annotated by experts as negative samples; Generate adversarial samples based on the derivative layer risk model and expand the training set of old customers with multiple conflicts; Convert the scenario-level business model into regularization constraints; The feature representation learned by the strong supervision branch is distilled to the weak supervision branch to construct a hybrid attention guidance mechanism.

[0014] Preferably, the intra-modal enhancement includes: A temporal self-attention mechanism is used on structured business data to capture the cyclical fluctuation patterns of historical business indicators; Implement BERT-based semantic enhancement on unstructured text data and reconstruct key risk description fragments through masked language models; An improved SwinTransformer is used to extract local-global features from business scene images, strengthening the focus on digital areas in financial voucher images. A voiceprint feature encoder is constructed for voice interaction data to separate emotional features from semantic features and generate an emotion-semantic joint embedding vector.

[0015] Preferably, cross-modal fusion is achieved by constructing a fourth-order modal interaction matrix, including: First-order fusion uses a parameter-sharing cross-modal Transformer to achieve semantic alignment between text and speech; Second-order fusion introduces graph attention network to transform the time series features of structured data into dynamic knowledge graph nodes; Third-order fusion builds a joint latent space of image and text through variational autoencoders; The fourth-order fusion design gated fusion unit adaptively adjusts the weight coefficient of each modality according to the business scenario type.

[0016] Preferably, performing contradiction detection by building a dual-channel contradiction detection engine includes: In the first channel, a lightweight detector based on knowledge distillation loads the core layer risk pattern vector library and matches the cross-modal fusion features with the risk pattern vector library through a graph neural network. Based on the matching results, it outputs the core layer conflict type code and preliminary location. Pattern matching verification is also performed on the negative sample database. The second channel maps the multimodal fusion features into a time-series risk map. The spatiotemporal graph convolutional network is used to calculate the abnormality of the map nodes. Based on the abnormality of the map nodes, the archive layer historical cases are triggered for verification, and the derivative layer contradiction type code and precise positioning are output. When the dual-channel detection results conflict, the reinforcement learning arbitration module is triggered to dynamically allocate confidence weights based on the risk level.

[0017] Preferably, the customer is automatically routed to the corresponding detection channel based on the customer's historical risk level and customer type: If the customer type is a new customer, an initial risk level is assigned by default and the first channel detection is enabled; a new risk level is determined based on the first channel detection results; If the customer type is an old customer with no risk and the historical risk level is less than the level threshold, the first channel is activated; If the customer type is a single-conflict old customer or a multiple-conflict old customer, or the historical risk level is greater than or equal to the level threshold, the first channel and the second channel are enabled at the same time; the risk knowledge system is dynamically updated based on the contradiction detection results.

[0018] In a second aspect, the present application provides a customer risk intelligent assessment system based on a large language model, which is used to implement any of the customer risk intelligent assessment methods based on a large language model described in the present application, and the system includes: The data acquisition module is used to obtain and pre-process the customer's multimodal data to extract structured and unstructured features; the multimodal data includes structured business data, unstructured text description data, business scenario images, and multi-round voice interaction data; A system construction module is used to construct a layered risk knowledge system and achieve adaptive evolution of the knowledge system through graph neural networks and generative models. The layered risk knowledge system includes a core layer, a scenario layer, a derivative layer, and an archive layer. A risk pattern vector library is established based on multiple anomaly types, including logical fallacies, contradictory expression paradigms, and abnormal semantics. The negative sample library construction module is used to randomly sample multimodal data that meets the preset conditions to build an initial negative sample library. It combines expert-annotated non-risk patterns and derivative layer analysis to build a negative sample database. The model training module is used to classify and label risk text data by customer type, optimizing the large language model using a collaborative training mechanism of strong supervision, weak supervision, and reinforcement learning for each classification. The customer types include new customers, existing customers without risks, existing customers with a single conflict, and existing customers with multiple conflicts. The output module is used to use the optimized large language model to mine risk features in text, process multimodal data through a multi-level attention network, locate logical contradictions through intra-modal enhancement, cross-modal fusion and contradiction detection, and generate a positioning report that includes contradiction type coding, business impact dimension assessment and risk level quantification.

[0019] In a third aspect, the present application also provides an electronic device comprising a memory, a processor, and a computer program stored on and executable on the memory, wherein the processor executes the program to implement the steps of any method described in the present application.

[0020] The beneficial effects of the present invention include: by processing multimodal data such as structured business data, unstructured textual representations, business scenario images, and multi-round voice interaction data, the method can capture richer customer risk information, reduce information omissions, and thus significantly improve the accuracy of risk assessment. The design of a layered risk knowledge system (core layer, scenario layer, derivative layer, and archive layer) enables more refined risk assessment and accurately identifies risk patterns at different levels. An adaptive evolution mechanism automatically updates the risk knowledge system through graph neural networks and generative models, reducing manual intervention and improving the efficiency of risk knowledge base maintenance. A collaborative training mechanism (strong supervision, weak supervision, and reinforcement learning) combines expert annotation and non-risk pattern analysis to construct an efficient negative sample database, accelerating the model training process. A multi-level attention network, through intra-modal enhancement and cross-modal fusion, enables the model to better adapt to different scenarios and customer types, enhancing the flexibility of risk assessment. A dynamic weight allocation mechanism (such as a fourth-order modal interaction matrix and a gated fusion unit) adaptively adjusts the weight coefficients of each modality based on the business scenario type, further improving the model's performance in different scenarios. The contradiction detection engine uses dual-channel detection (a lightweight detector based on knowledge distillation and a spatiotemporal graph convolutional network) to locate logical contradictions in real time, improving the real-time nature of risk assessment. When conflicting dual-channel detection results are encountered, the reinforcement learning arbitration module dynamically assigns confidence weights based on risk levels, ensuring the timeliness and accuracy of risk assessment results. The generated risk location report includes contradiction type coding, business impact dimension assessment, and risk level quantification, providing detailed decision-making support for risk management. By dynamically updating the risk knowledge system and optimizing detection strategies, the method can continuously adapt to changes in the risk environment, providing long-term support for risk management. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of a method for intelligently assessing customer risk based on a language model provided in an embodiment of the present application; Figure 2 This is a schematic diagram of a flow chart for the adaptive evolution of a hierarchical risk knowledge system provided in an embodiment of the present application; Figure 3 This is a schematic diagram of the detection channel routing selection flow chart provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] Below, the present application is further described in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of no conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0023] Refer to the attached Figure 1 , this application provides a customer risk intelligent assessment method based on a large language model, the method comprising: S1. Obtain and preprocess the customer's multimodal data to extract structured and unstructured features; the multimodal data includes structured business data, unstructured text description data, business scenario images, and multiple rounds of voice interaction data; S2. Construct a layered risk knowledge system and implement adaptive evolution of the knowledge system through graph neural networks and generative models. The layered risk knowledge system includes a core layer, a scenario layer, a derivative layer, and an archive layer. A risk pattern vector library is established based on multiple anomaly types, including logical fallacies, contradictory expression paradigms, and abnormal semantics. S3. Randomly sample multimodal data that meets the preset conditions to build an initial negative sample library. Combined with expert-annotated non-risk patterns and derivative layer analysis, a negative sample database is constructed. S4. Classify and label risk text data by customer type, and optimize the large language model using a collaborative training mechanism of strong supervision, weak supervision, and reinforcement learning for each category. The customer types include new customers, existing customers without risks, existing customers with a single conflict, and existing customers with multiple conflicts. S5. Use the optimized large language model to mine risk features in text, process multimodal data through a multi-level attention network, locate logical contradictions through intra-modal enhancement, cross-modal fusion, and contradiction detection, and generate a positioning report that includes contradiction type coding, business impact dimension assessment, and risk level quantification.

[0024] The working principle of the above technical solution is: Acquire and pre-process multi-dimensional customer data through multi-channel interfaces. This includes collecting structured operational data from bank credit management systems and corporate financial systems, including corporate balance sheets, cash flow statements, and operating performance indicators. Furthermore, it extracts structured information such as business plans and financial forecasts submitted by customers during the bank credit application process to form the foundation of structured operational data. Collect unstructured text data from public documents, application forms, and social media, including public documents such as annual reports, news reports, and regulatory announcements, as well as internal corporate meeting minutes and business reports. On the other hand, for customer credit application scenarios, collect in real time text information such as business plans and financial forecasts filled out by customers in application forms, as well as industry insights and corporate dynamics shared by customers on social media, to comprehensively collect customer-related unstructured text data. Collect call records in which customers explain details such as the purpose of funds and repayment sources during telephone conversations with bank customer service. At the same time, data from other voice interaction scenarios involving customer business communications and demand expressions is also collected to ensure the integrity of voice interaction data. When acquiring multiple rounds of voice interaction data, support is provided for customers to proactively initiate voice inquiries or for the system to proactively initiate conversations based on the risk assessment process. Through voice interaction, further explanations and supplementary information from customers on risk-related issues are obtained, and this information is incorporated into the risk assessment data source. For structured business data, missing values ​​are identified and processed, and numerical missing data are filled using methods such as mean filling and multiple filling. For categorical data, filling is performed according to business rules or majority classes; outliers are detected and corrected, and abnormal data are found using methods such as box plots and the 3σ principle, and then corrected or eliminated based on actual business conditions.

[0025] For unstructured text data, we remove duplicate text, invalid characters, and garbled text; we use regular expressions to filter out distracting content such as advertisements and irrelevant statements. For voice interaction data, we remove invalid audio clips that are too short, too noisy, or unrecognizable.

[0026] For structured business data, numerical data is normalized or standardized, such as using Min-Max normalization to map data to the [0, 1] range or using Z-Score normalization to conform to a standard normal distribution. Categorical data is encoded, converting it to numerical data using methods such as one-hot encoding and label encoding. The encoding mapping table aligns heterogeneous indicators in corporate financial reports and credit data. For unstructured text data, word segmentation is first performed, using natural language processing tools to break the text into words or phrases. Then, word embedding is performed, mapping words into low-dimensional dense vector representations using models such as Word2Vec and BERT. For voice interaction data, speech recognition technology is used to convert audio into text data, and subsequent conversion operations are performed according to the processing methods for unstructured text data.

[0027] A layered risk knowledge system is constructed, consisting of a core layer, a scenario layer, a derivative layer, and an archive layer. The core layer (with real-time memory updates covering logical fallacies, contradictory paradigms, and anomalous semantics) stores expert-verified and annotated key risk patterns, as well as high-value new risk patterns. The scenario layer corresponds to risk patterns in different business scenarios, as well as some new risk modules (such as the integrity of new customer information and historical contradictions with existing customers). The derivative layer stores new or unverified risk patterns, and the archive layer preserves historical risk cases. Furthermore, a risk pattern vector library is established based on various anomaly types, such as logical fallacies, contradictory expression paradigms, and anomalous semantics.

[0028] Graph neural networks and generative models are used to dynamically update the knowledge system. By integrating multi-source data such as expert feedback, model discovery, policy analysis, and public opinion monitoring, knowledge system updates are automatically triggered through natural language processing (NLP) and time series analysis. External knowledge bases are integrated through semantic fusion and transfer learning, and generative large language models are used to expand and optimize knowledge base content, enabling the risk knowledge system to continuously evolve as the market environment and risk characteristics change.

[0029] Random sampling is performed on multimodal data that meets the preset conditions to build an initial negative sample library to cover various potential non-risk situations; combined with expert-annotated non-risk patterns and analysis of the derivative layer, the negative sample database is further enriched to ensure the quality and representativeness of the negative sample data, providing sufficient non-risk pattern learning materials for model training. Specifically, text data with a risk score below the scoring threshold and a complaint rate of 0, and business transaction behavior data that conforms to normal patterns (such as transaction frequency and amount within the industry average fluctuation range) are screened from the business database, and a negative sample library is generated after random sampling. The preset threshold may be determined based on a preset percentile (e.g., the 20th percentile) of the historical data risk score; The risk score and complaint rate of business text data are recalculated based on a rolling time window (such as 30 days). Text data that newly meets the "risk_score < preset score threshold, complaint_rate = 0" and whose business transaction behavior data conforms to the normal pattern are added to the negative sample library at a preset sampling ratio (such as 5%); the preset score threshold is obtained through pre-training.

[0030] Customers are divided into new customers, old customers without risks, old customers with a single conflict, and old customers with multiple conflicts, and the risk text data of each type of customer is labeled.

[0031] Under each customer classification, a collaborative training mechanism of strong supervision, weak supervision and reinforcement learning is used to optimize the large language model; this enables the model to effectively learn and identify the risk characteristics of different types of customers; the classification optimization strategy enables the model to more accurately adapt to the risk characteristics of different customer types, thereby improving the targeted nature of risk assessment.

[0032] The processed customer data is fed into an optimized large language model. Leveraging the model's semantic understanding and generation capabilities, it deeply analyzes the text content. Using a multi-layered attention network to process multimodal data, it achieves intra-modal enhancement, cross-modal fusion, and contradiction detection. This allows for precise identification of risk characteristics and logical contradictions within the text, including but not limited to inconsistent operating data (e.g., a company's profit statements don't match its cost control statements), logical flaws in repayment plans (e.g., planned repayment dates don't match projected cash inflows), and industry deviation risks (e.g., a customer's business transformation statements don't align with market trends and company resources). Furthermore, through multi-round conversation tracking and contextual analysis, it identifies coherence risks within the customer's statements and generates a location report that includes contradiction type coding, business impact assessment, and quantified risk levels.

[0033] Based on the identified risk characteristics, and incorporating factors such as risk type, frequency, and severity, a detailed risk assessment report is generated for the client. This report uses intuitive visualizations (such as risk radar charts and trend line graphs) combined with clear text descriptions to present a multi-dimensional risk profile of the client, encompassing credit risk, fraud risk, and more. Based on the risk assessment results, financial institutions are provided with decision-making recommendations covering areas such as credit approval, credit limit setting, and risk monitoring strategy development, helping them accurately manage risk and achieve robust operations.

[0034] In a possible implementation, the method further includes: A visual management interface facilitates comprehensive system control by financial institution managers. It supports operations such as configuring risk assessment rules and setting risk thresholds, allowing for flexible adjustment of risk assessment criteria based on business needs and risk appetite. A log is maintained throughout the risk assessment process, covering key information such as client information, assessment time, risk characteristics, and assessment results, providing detailed data support for subsequent audits and model optimization. A feedback mechanism is established to collect feedback from financial institution risk control personnel on risk assessment results. This feedback is then used to conduct targeted optimization of the large language model and risk assessment rules, forming a closed-loop continuous improvement process to continuously enhance the accuracy and practicality of system assessments.

[0035] In one possible implementation, the core layer memory real-time update adopts a distributed memory database architecture, specifically including: RedisCluster is used to store core layer risk pattern vectors, and Apache Ignite is used to achieve full memory residency of hot data. Implement incremental updates for newly added risk patterns: only recalculate the affected historical pattern subgraph and update node embeddings through the graph neural network local propagation algorithm.

[0036] The effects of the above technical solution are: By acquiring multimodal data such as structured business data, unstructured text description data, business scenario images, and multiple rounds of voice interactions from customers, and comprehensively processing and analyzing them, it breaks the limitation of traditional methods that rely only on a single or a small amount of data, and achieves 360-degree coverage of customer information without blind spots. It can explore the risk characteristics of customers from multiple dimensions, present a complete risk picture of customers to financial institutions and other enterprises, avoid risk omissions, and make risk assessment more comprehensive and accurate.

[0037] Leveraging the powerful semantic understanding and deep learning capabilities of large language models, combined with the adaptive evolution of a layered risk knowledge system, the construction of a scientific negative sample database, and a collaborative training mechanism using strong supervision, weak supervision, and reinforcement learning, this method can accurately locate subtle risk features within text, such as obscure logical loopholes, subtle contradictions, unusual semantics, and various types of logical fallacies. By processing multimodal data through a multi-level attention network, it achieves intra-modal enhancement and cross-modal fusion, further improving the ability to identify risk features. Its risk identification accuracy is significantly improved compared to traditional methods, enabling enterprises to make more informed decisions.

[0038] The constructed layered risk knowledge system is adaptively evolving through graph neural networks and generative models. Combined with the integration and updating of multi-source data, it enables the risk knowledge system to closely track market dynamics, industry trends, and evolving customer behaviors. Furthermore, through online learning and dynamic optimization of model parameters, the system ensures real-time updates of risk assessment criteria, accurately adapting to evolving business risk characteristics. This provides enterprises with up-to-date risk assessment services, enabling them to better address various risk challenges in a complex and volatile business environment.

[0039] By classifying customers and optimizing the large language model using different training mechanisms, the model can effectively learn the characteristics of different customer types, including new customers, existing customers with no risks, existing customers with a single conflict, and existing customers with multiple conflicts, thereby enhancing the model's generalization capabilities. Furthermore, by constructing a negative sample database and employing multiple training methods, the model's ability to identify various risk and non-risk patterns is improved, enabling the model to maintain high assessment accuracy and stability across different business scenarios and customer groups.

[0040] The generated positioning report contains detailed information such as conflict type coding, business impact dimension assessment, and risk level quantification. It uses a combination of intuitive visual charts and clear text descriptions to provide companies with comprehensive and accurate customer risk profiles and decision-making recommendations.

[0041] In a possible implementation, S2 includes: Core-layer risk patterns (such as fundamental risk types like guarantee chain rupture) are defined as basic risk nodes, with pattern coding and typical feature vector attributes configured. Derivative-layer risk patterns (such as industry scenario-based risks and new business conflicts) are defined as scenario-based risk nodes, with industry labels and time series association rule attributes configured. The typical feature vectors are adapted to the enhanced features within the multimodal data modality, including: Typical feature vectors of core layer basic risk nodes, associated with temporal fluctuation characteristics of structured business data and semantic features of key risk expressions in unstructured text data; The derived layer scenario-based risk node attributes, associated business scenario image local-global abnormal features (improved SwinTransformer extraction), and voice interaction data emotion-semantic joint embedding vector.

[0042] The core risk model node builds a hierarchical indexing system when storing, which supports the rapid location of the feature vectors that are most relevant to the semantic association vector: First-level classification index: classified by risk type, each type corresponds to a set of basic risk nodes.

[0043] Index fields: risk type code, industry label (such as finance, manufacturing), and risk level (high / medium / low).

[0044] Secondary feature index: For the typical feature vector of each risk node, extract the key feature dimensions and construct an inverted index: Structural characteristics: threshold conditions such as revenue growth rate > 30% and debt-to-asset ratio > 80%; Semantic features: keyword vectors extracted by BERT, such as revenue recognition and inflated profits.

[0045] Refer to the attached Figure 2 In a possible implementation, S2 includes: Perform cluster analysis on unlabeled text to obtain new risk patterns; Calculate the risk confidence of the new risk model based on the maximum similarity between the new risk model and each historical risk model in the core layer and the frequency of occurrence of the new risk model, and determine the storage level of the new risk model data; Integrate multi-source data, including expert feedback, model discovery, policy analysis, and public opinion monitoring, to automatically trigger updates through NLP and time series analysis. Integrate external knowledge bases through semantic fusion and transfer learning, and combine them with generative large language models to expand and optimize the knowledge base content.

[0046] In one possible implementation, the step of calculating the risk confidence of the new risk model based on the maximum similarity between the new risk model and each historical risk model in the core layer and the frequency of occurrence of the new risk model, and determining the storage level of the new risk model data, includes: If the maximum similarity is less than the similarity threshold, then calculating the abnormality score of the new risk pattern; The abnormality score of the new risk pattern is calculated by using the new risk pattern vector, the mean and standard deviation of multiple historical pattern vectors in the core layer, and the frequency of occurrence of the new risk pattern. The abnormality score is obtained by the following formula:

[0047] in, is the characteristic vector of the new risk pattern; is the mean of multiple historical pattern vectors in the core layer; is the standard deviation of the historical pattern eigenvector; To quantify the difference between the new pattern feature vector and the mean of the historical pattern vector by the square of the Euclidean distance; is the risk pattern frequency; is the frequency reference value; If the abnormality score is greater than or equal to the first score threshold, it is pushed to the expert annotation queue; the risk pattern data annotated by the expert is stored in the core layer, and the non-risk pattern data is stored in the negative sample library; If the abnormality score is less than the first score threshold, the new risk pattern data is stored in the derivative layer. The distribution of multiple historical abnormality scores (OTS) is calculated based on a rolling time window (e.g., 30 days). The third quartile Q3 and interquartile range (IQR) are taken. The first score threshold can be Q3 + a1 × IQR. a1 is the first coefficient, a1∈[0.5,1.5]. For example, a1=0.5 for high risk warning and a1=1.5 for low risk. If the maximum similarity is greater than or equal to the similarity threshold, the corresponding risk knowledge system layer is selected and stored according to the risk confidence level; including: If the risk confidence of the new risk model is greater than or equal to the first confidence threshold, the data of the risk model is stored in the core layer; If the risk confidence of the new risk model is greater than or equal to the second confidence threshold and less than the first confidence threshold, the data of the risk model is stored in the scenario layer; If the risk confidence of the new risk model is lower than the second confidence threshold, the data of the new risk model is stored in the derivative layer; When the frequency of risk patterns in the derivative layer or scenario layer continues to increase for multiple statistical periods, its abnormality score and risk confidence level are recalculated and migrated to the core layer or scenario layer after meeting the standards.

[0048] In one possible implementation, the method for obtaining the risk confidence level includes: Extract historical risk pattern feature vectors and new risk pattern feature vectors through pre-trained semantic encoders; calculate the maximum similarity between the new risk pattern and each historical risk pattern feature vector as the basic confidence level; Determine the final risk confidence level through basic confidence level, risk pattern frequency and policy correlation coefficient; Among them, the final risk confidence is:

[0049] in, is the final risk confidence level, is the basic confidence level, ∈[0,1], is the risk pattern frequency; is the frequency reference value; is the policy relevance coefficient, ∈[0.7, 1.6]; min() is the minimum value; is the risk confidence level of the historical risk model that is most similar to the new risk model; In a possible implementation, the policy relevance coefficient is obtained by the following steps: The policies are divided into 3 levels (adjustable according to industry): Level 1 (strong correlation): If there are major national / industry policies, Take 1.3-1.5; Level 2 (medium related): If local / special rectification policies, then Take 1.0-1.3; Level 3 (weakly relevant): If it is a general guidance document / initiative, Take 0.8-1 Adjustment based on the policy of "whether to directly constrain risk model": If the policy explicitly mentions the risk model keyword, Add 0.1 to the upper limit of the corresponding level; If the policy is only mentioned in general terms, Subtract 0.1 from the lower limit of the corresponding level; We can use the historical “actual impact of risk patterns after policy releases” to infer , to achieve dynamic calibration.

[0050] In another possible implementation, the policy relevance coefficient is obtained by the following steps: Construct a financial regulatory knowledge graph (nodes: policy keywords, risk contradiction types; edges: association strength); For the keywords extracted from the policy text, the associated risk contradiction types are searched through the knowledge graph; Calculate the semantic relevance score:

[0051] : The core contradiction has a high weight, and the derivative contradiction has a low weight; the number of all activated associated edges, that is, the total number of times that the keywords extracted from the policy text have effective associations with the risk contradiction types; Establish a scenario-based keyword library (including negative keywords in credit scenarios and compliance scenarios); Filter the corresponding keyword library by policy scenario (identified through text classification models, such as "Commercial Bank Loan Risk Management Notice" → Credit scenario); Get sentiment polarity score:

[0052] is the semantic association weight; ∈[0,1], used to balance word frequency and knowledge association, determined comprehensively based on business scenario characteristics, data distribution, and expert experience; Based on historical policy and risk correlation data (such as risk trigger rate and business fluctuation range after policy release), Divide the sentiment polarity score into intervals and map the initial policy relevance coefficient; For example: Strong regulatory policies: >0.3 (negative keywords significantly outnumber positive ones) → Initial policy relevance coefficient equals 1.5 (high policy relevance, enhanced risk detection); Conventional policy: -0.3≤ ≤0.3→The initial policy correlation coefficient is equal to 1.5 (basic policy correlation); Encouraging policies: <-0.3 (positive keywords significantly outnumber negative ones) → initial alpha = 0.7 (low policy relevance, weakening conventional detection); Set different attenuation periods according to policy types (emergency policies, regular policies, incentive policies), and dynamically adjust the policy correlation coefficient based on changes in related business risks.

[0053] When the contradiction type associated with the policy keyword is triggered frequently during detection (for example, ≥3 times per week), the attenuation rate will be automatically halved; if there is no associated risk trigger for two consecutive weeks, the original attenuation rate will be restored, forming a "policy-risk-attenuation" business closed loop.

[0054] The working principle of the above technical solution is: Unlabeled text may contain new risk patterns that have not yet been included in the risk knowledge system. Cluster analysis can group text fragments with similar semantics and structures into one category, thereby identifying new risk patterns with common characteristics. The core of this method is to use the spatial distribution pattern of the semantic feature vectors of the text to cluster texts with similar characteristics, with each cluster representing a potential risk pattern. Use pre-trained semantic encoders (such as BERT) to extract feature vectors of unlabeled text and convert the text into points in a high-dimensional semantic space.

[0055] Clustering algorithms such as DBSCAN and K-means are used to cluster feature vectors and divide text into different clusters.

[0056] Semantic analysis is performed on the text of each cluster to extract common features and form a preliminary description of the new risk pattern.

[0057] By measuring the degree of difference between a new risk pattern and the historical risk patterns of the core layer and its frequency of occurrence, we assess the abnormality and potential risk level of the pattern. The greater the difference between the new risk pattern and the higher the frequency of occurrence, the higher the abnormality score and the greater the potential risk.

[0058] If the maximum similarity is less than the similarity threshold, it indicates that the new pattern differs significantly from the historical pattern in the feature space, which is a potential risk. The anomaly score should be used to assess the degree of deviation from the historical pattern and its frequency of occurrence. The squared Euclidean distance measures the morphological difference between the new pattern and the historical mean, and dividing it by the standard deviation eliminates the scale effect of the historical pattern distribution. The frequency of occurrence is converted into abnormal contribution (the higher the frequency, the greater the contribution); the final OTS is the product of the degree of morphological abnormality and the degree of frequency abnormality. The higher the score, the more likely it is a new risk worthy of attention.

[0059] If the maximum similarity is greater than or equal to the similarity threshold, it means that the new model and the historical model are similar in form and are "variants / continuations of known risks". Confidence needs to be used to assess the inheritance of historical risks plus incremental factors such as frequency / policy.

[0060] The fundamental parameter for calculating risk confidence (i.e., maximum similarity) is derived directly from the feature matching results of the new model and the historical model. Subsequent adjustments to frequency and policy factors will be based on this basis through linear or nonlinear transformations to prevent patterns with similar morphology but no significant increase in actual risk from being repeatedly included in the core layer.

[0061] Even if the similarity with the historical pattern is high, if the frequency or policy relevance of the new pattern increases significantly, the final risk confidence level may reach the core layer standard; in this case, the essence of storing in the core layer is to update the risk weight of the historical pattern rather than to store it repeatedly (for example, expanding the historical pattern from "credit application" to "credit card application" scenarios to form a more general core risk model); if the first confidence threshold > ≥ the second confidence threshold (e.g. 0.6), indicating that although the new model is similar to the historical model, the risk weight does not meet the core layer standard (e.g. the frequency of contradictions in income certificates in the credit card scenario does not exceed that of the historical credit scenario). ), it is then stored in the scenario layer as a risk reference for specific business scenarios to avoid duplication with historical core models. If the risk confidence level is less than the second confidence threshold, the risk weight of the new model is significantly lower than that of the historical core model (for example, an income discrepancy in a credit card application is a rare event), and it is stored in the derivative layer for verification to avoid occupying core layer resources.

[0062] When the frequency of risk patterns in the derivative layer or scenario layer continues to increase for multiple statistical periods, its abnormality score and risk confidence level are recalculated and migrated to the core layer or scenario layer after meeting the standards.

[0063] Multi-source data such as expert feedback, model discovery, policy analysis and public opinion monitoring reflect changes in risk characteristics from different dimensions. By integrating these data through NLP and time series analysis, we can timely discover content that needs to be updated in the risk knowledge system and realize the dynamic evolution of the knowledge system.

[0064] Semantic fusion technology is used to map and align concepts and terms in the external knowledge base with the semantic space of the internal risk knowledge system to ensure the consistency and coherence of knowledge.

[0065] Leveraging transfer learning technology, we migrate knowledge from external knowledge bases to our internal risk knowledge system, enriching the description and classification of risk patterns. Incorporating generative large language models, we generate new risk pattern descriptions and risk assessment rules based on the internal risk knowledge system and the external knowledge base, expanding and optimizing the knowledge base content.

[0066] The effects of the above technical solution are: The risk confidence calculation comprehensively considers multiple factors, including the similarity between new risk patterns and historical patterns, frequency of occurrence, and policy relevance. Compared to single-factor assessments, this method can more accurately assess the risk level and significance of new risk patterns, thereby correctly storing them at the appropriate level and providing more accurate knowledge support for subsequent risk feature mining. Similarity calculations can accurately identify seemingly novel patterns that are actually highly similar to historical risk patterns, preventing them from being misidentified as new risk patterns.

[0067] Multi-source data fusion and automatic update mechanisms enable the risk knowledge system to respond in real time to changes in the internal and external environment, including expert experience, model discovery, policy changes, and market sentiment. This allows for timely adjustments to risk assessment standards and knowledge content, ensuring that risk assessments can adapt to evolving business scenarios and risk characteristics, improving both timeliness and adaptability. For example, when the government adjusts its credit policies, the system can promptly update relevant risk models through policy analysis, ensuring that risk assessments comply with the latest policy requirements.

[0068] The hierarchical storage mechanism (core layer, scenario layer, derivative layer, and archive layer) enables refined management of risk models. Risk models at different levels correspond to different processing methods and application scenarios, enabling risk assessment to provide more targeted risk assessment results based on different business needs and risk characteristics, thereby improving the level of refined risk management. For example, key risk models at the core layer have higher weights in risk assessment, and risk models at the scenario layer can be flexibly applied according to specific business scenarios.

[0069] By accurately identifying and storing new risk patterns and continuously improving the negative sample database, we provide richer and more accurate risk and non-risk pattern data for large language model training, continuously optimizing the quality of model training data, thereby enhancing the large language model's ability to identify risk characteristics and the accuracy of risk assessment. For example, after verification, risk patterns at the derivative layer can be upgraded to core or scenario-level patterns, providing more reliable positive samples for model training.

[0070] By integrating external knowledge bases through semantic fusion and transfer learning, and combining them with generative large language models to expand and optimize knowledge base content, the risk knowledge system can fully leverage the rich external knowledge, compensate for the deficiencies in internal knowledge, and enhance the comprehensiveness and accuracy of the knowledge system. For example, integrating external industry knowledge and policies and regulations can enable the risk knowledge system to more comprehensively cover the risk characteristics and policy requirements of various industries, thereby improving the accuracy of risk assessments.

[0071] In a possible implementation, the S4 includes: Conduct main task training based on manually annotated core layer samples. Filter manually annotated risky (positive) and non-risky (negative) samples in the core layer and divide them into training and validation sets proportionally. Use CNN (TextCNN for text scenarios) or Transformer (such as BERT) as the base model, with the input being the feature vector / text sequence of the risk pattern. Use the cross-entropy loss function to optimize the model's risk identification ability for "clear core layer samples" to obtain the feature representation of the strong supervision branch. The weakly supervised training includes positive samples and negative samples, including: The new risk pattern samples in the core layer are taken as positive samples, and positive and negative sample pairs for comparative learning are constructed based on the positive samples; from the new risk pattern samples in the core layer, "different data enhancement versions" of the same sample (such as text synonym replacement, vector noise addition) are randomly selected as positive sample pairs; let the positive sample be , through the data enhancement function generate , construct the comparison pair ,The goal is to make the model close the feature distance between the two; The non-risk patterns annotated by experts are used as negative samples; the non-risk pattern samples annotated by experts are used as negative samples , build Yes; let the model push its feature distance; use InfoNCE loss to calculate feature similarity and optimize.

[0072]

[0073] is the feature vector of the i-th positive sample; for Positive sample pair features generated by data enhancement; sim() is similarity calculation, usually cosine similarity; is the temperature coefficient (hyperparameter, usually 0.5-1.0), which is used to adjust the smoothness of the similarity distribution: The smaller it is, the similarity difference is amplified, and the distinction between positive and negative samples is stricter; is the feature vector of the j-th negative sample (such as the non-risk pattern features marked by experts); N is the total number of samples; is the contrast loss; Generate adversarial samples based on the derivative layer risk patterns and expand the training set of multi-contradictory old customers. Use methods such as FGSM (Fast Gradient Signed Method) and PGD (Projected Gradient Descent) to impose small perturbations on the derivative layer risk pattern samples to generate adversarial samples. Add these to the training set and mix them with the core layer samples in a ratio of 3:7 (adjusted based on the amount of data) to enhance the model's robustness to potential risks. Convert scenario-level business models into regularized constraints; analyze the rules of scenario-level business models (e.g., "In a credit scenario, high-quality customers must have an income / debt ratio greater than 1.5") and convert them into mathematical constraints; incorporate the constraints into the loss function (weighted sum of the constraints and the main loss (cross entropy)); The feature representation learned by the strong supervision branch is distilled to the weak supervision branch to construct a hybrid attention guidance mechanism; the intermediate layer features of the strong supervision branch are distilled to the weak supervision branch through L2 loss; in the Transformer layer of the weak supervision branch, the attention weight of the strong supervision is integrated with the self-attention, guiding the model to focus on key risk features.

[0074] The effects of the above technical solution are: Feature distillation allows weakly supervised branches to learn on the shoulders of strong supervision, reducing their reliance on massive samples. Hybrid attention integrates experience to accelerate model convergence and reduce training resource consumption. Through a layered, multi-branch collaborative training approach, risk assessment models evolve from relying on manual labeling to autonomously learning business logic and covering the full range of risk profiles, ultimately improving the accuracy and practicality of intelligent risk assessments for customers.

[0075] In a possible implementation, the intra-modality enhancement includes: A temporal self-attention mechanism is used on structured business data to capture the cyclical fluctuation patterns of historical business indicators; Implement BERT-based semantic enhancement on unstructured text data and reconstruct key risk description fragments through masked language models; An improved SwinTransformer is used to extract local-global features from business scene images, strengthening the focus on digital areas in financial voucher images. A voiceprint feature encoder is constructed for voice interaction data to separate emotional features from semantic features and generate an emotion-semantic joint embedding vector.

[0076] The working principle and effects of the above technical solution are as follows: The self-attention mechanism is used to capture the dependencies of structured business data (such as financial indicators, transaction records, etc.) in time series, focusing on identifying patterns such as cyclical fluctuations and abnormal mutations, thereby discovering potential risks. For example, if a company's revenue data shows regular seasonal fluctuations, it is normal, but if it suddenly declines sharply for multiple consecutive cycles, it may indicate business risks.

[0077] Arrange structured business data in chronological order and form a time series after normalization; for each time point, calculate its attention weight relative to other time points; aggregate historical information through attention weights to obtain a time series feature representation; use Fourier transform or sliding window methods to analyze the cyclical components of time series features, and identify abnormal fluctuations through preset thresholds; capture early risk signals such as a sudden drop in revenue and a surge in costs, such as a company's net profit falling by more than 30% quarter-on-quarter for two consecutive quarters, triggering a risk warning; and distinguish between normal seasonal fluctuations and abnormal risks to avoid misjudgment.

[0078] Leveraging the Masked Language Model (MLM) capabilities of the pre-trained BERT model, by randomly masking key tokens in text, the model learns to reconstruct masked risk statements. This strengthens its understanding of semantics and helps identify implicit risk information, such as disclaimers in contracts and contradictory statements in customer communications. After tokenizing unstructured text (such as credit applications or customer service records), tokens are randomly masked at a certain ratio (e.g., 15%). The masked text is fed into the pre-trained BERT model to obtain contextual representations of each token. The model is trained using a cross-entropy loss function to predict the masked tokens. After training, BERT's [CLS] vectors or outputs from specific layers are used as text features, focusing on the contextual representation of risk-related words. This approach can capture the implicit semantics of the text.

[0079] SwinTransformer is a hierarchical visual transformer that processes local-global image features through a moving window mechanism. The improved SwinTransformer is specifically designed for images of financial documents (such as invoices and bank statements), focusing on numerical regions. By adding a numerical feature enhancement module or adjusting attention weights, it improves the recognition accuracy of key figures such as amounts and dates, thereby detecting risks such as tampering and forgery in images.

[0080] First, use OCR or target detection models (such as YOLO) to locate the digital areas in the image (such as the amount column and date column); in the window attention calculation of SwinTransformer, additional weights are applied to the feature maps corresponding to the digital areas.

[0081] SwinTransformer extracts features of different scales (such as 4×4 and 8×8 pixel windows) in layers; it fuses shallow (details) and deep (semantic) features through jump connections, especially retaining the edge details and global position information of the digits.

[0082] Add a fully connected layer after the model output, input the digital area features, and predict whether there is tampering (output 0-1 probability).

[0083] Combine cross entropy with feature distance loss (e.g. comparing the feature differences between tampered and normal numbers).

[0084] The voiceprint feature encoder decomposes the speech signal into emotional features (such as tension and hesitation reflected by intonation and speaking speed) and semantic features (the textual semantics of the speech content), generating a joint embedding vector to simultaneously capture the emotional tendency and semantic information in the speech.

[0085] The two-stream network can be used to process acoustic features separately, including: Emotional branch: LSTM+CNN extracts emotional features such as intonation and speaking speed, with an output dimension of, for example, 128 dimensions; Semantic branch: First, convert speech to text (ASR), then use BERT to extract semantic features, and output, for example, 768 dimensions.

[0086] Concatenate or weightedly fuse emotion and semantic features (e.g., emotion feature weight 0.3, semantic weight 0.7) to obtain an emotion-semantic joint vector. Use contrastive learning to make the joint vectors of similar emotion-semantic pairs (e.g., true overdue explanations) closer and those of different pairs farther apart.

[0087] In one possible implementation, cross-modal fusion is achieved by constructing a fourth-order modal interaction matrix, including: First-order fusion uses a parameter-sharing cross-modal Transformer to achieve semantic alignment between text and speech and generate a semantic association vector. Second-order fusion introduces a graph attention network to transform the time series features of structured data into dynamic knowledge graph nodes; it associates dynamic nodes with core risk nodes; Third-order fusion builds a joint latent space of image and text through variational autoencoders; The fourth-order fusion design gated fusion unit adaptively adjusts the weight coefficient of each modality according to the business scenario type.

[0088] The working principle of the above technical solution is: First-order fusion uses parameter-sharing Transformer to align text semantic features (after BERT enhancement) with speech semantic features (after voiceprint separation) to generate a semantic association vector; the cosine similarity between the semantic association vector and the typical feature vector of the core-layer risk pattern node is calculated: if the similarity is greater than the second similarity threshold, the core-layer risk node activation is triggered; based on the keywords in the semantic association vector, the risk node set containing relevant feature labels is quickly located through the secondary feature index; if the semantic association vector contains industry information, the nodes matching the industry label are further filtered out; for the filtered node set, the cosine similarity between the semantic association vector and the typical feature vector of each node is calculated in parallel; nodes with similarity greater than the second threshold are retained and sorted in descending order of similarity; if the highest similarity node is greater than the second similarity threshold, the node activation is triggered; otherwise, it is determined to be "no matching core-layer risk" and transferred to the derivative layer detection; the second similarity threshold is set by empirical method or cross-validation.

[0089] Second-order fusion uses the Graph Attention Network (GAT) to transform the temporal fluctuation characteristics of structured data into dynamic nodes (such as "monthly revenue node" and "asset-liability node"). Based on the temporal association rules of the derivative layer, the dynamic nodes are associated with the basic risk nodes of the core layer. For example, if the revenue fluctuation exceeds 30% and the asset-liability ratio exceeds 80% for three consecutive months, the core layer "cash flow interruption" risk node is associated; Third-order fusion constructs a joint image-text latent space, uses a variational autoencoder (VAE) to extract and cluster image-text anomaly features, mines new anomaly patterns, generates derivative-layer scenario-based risk nodes, and archives them into a hierarchical risk knowledge system; The third-order fusion adaptively adjusts the weights of the core layer and derivative layer risk nodes according to the business scenario type (such as credit, compliance, etc.). For example, in the compliance scenario, the core layer node weight = 0.7, and the derivative layer node weight = 0.3, and vice versa in the innovative business scenario.

[0090] The effects of the above technical solution are: By aligning text-speech semantics and constructing an image-text latent space, we break down single-modal information barriers (e.g., unusual speech expressions and implicit risk cues in images can be identified through text semantic association), addressing the problem of missed cross-media risks in traditional single-modal assessments. For example, by combining textual anomalies in financial reports with semantic inconsistencies in contemporaneous meeting recordings, we can accurately identify financial risks.

[0091] The fourth-order fusion gating unit dynamically adjusts modal weights based on the business scenario (credit, compliance, or innovative business). For example, compliance scenarios prioritize core nodes (weight 0.7) to ensure fundamental risk management, while innovative business scenarios prioritize the derivative layer (weight 0.7) to uncover potential patterns. This enables a single system to adapt to risk assessments across multiple scenarios, reducing model rebuilding costs when switching between scenarios.

[0092] Second-order fusion transforms structured data into dynamic knowledge graph nodes and associates core-derivative layer risks, upgrading risk assessment from single-point feature matching to graph-based association reasoning. For example, by associating revenue fluctuations with the time series of asset-liability data, it automatically triggers the activation of cash flow interruption risk nodes, enabling intelligent mining of risk transmission relationships and resolving the problem that traditional rule engines cannot dynamically associate complex risk chains.

[0093] The third-order fusion uses VAE (variational autoencoder) to construct an image-text joint latent space, which can mine new cross-modal anomalies (such as the hidden risks of inconsistencies between invoice image anomalies and text descriptions), generate scenario-based derivative layer nodes and archive the knowledge system, supporting the self-evolution of the risk knowledge system.

[0094] First-order fusion uses keyword / industry label secondary indexing of semantic association vectors to rapidly filter risk node sets and calculate similarities in parallel, improving matching efficiency. If a core-level match is not found, the process automatically transitions to the derivative layer, achieving a closed-loop tiered detection process and reducing redundant computation. Intermediate results from each fusion order (semantic association vectors, dynamic graph nodes, and cross-modal clustering features) are traceable and, combined with interpretable metrics such as cosine similarity and temporal association rules, address the black-box nature of AI models. For example, risk node activation can be traced back to speech keywords and image anomaly clustering results, meeting the interpretability requirements of regulatory scenarios.

[0095] In one possible implementation, conflict detection by building a dual-channel conflict detection engine includes: In the first channel, a lightweight detector based on knowledge distillation loads the core layer risk pattern vector library and matches the cross-modal fusion features with the risk pattern vector library through a graph neural network. Based on the matching results, it outputs the core layer conflict type code and preliminary location. Pattern matching verification is also performed on the negative sample database. In the second channel, the multimodal fusion features are mapped into a time series risk map, and the abnormality of the map nodes is calculated using the spatiotemporal graph convolutional network. According to the abnormality of the map nodes, the historical cases in the archive layer are triggered for verification, and the contradiction type code and precise positioning of the derivative layer are output.

[0096] When the dual-channel detection results conflict, the reinforcement learning arbitration module is triggered to dynamically allocate confidence weights based on the risk level.

[0097] The working principle of the above technical solution is: In the first channel, the core layer risk pattern vector library is loaded, and the cross-modal fusion features are matched with the risk pattern vector library using a graph neural network (GNN). The matching score is calculated. Based on the matching results, if the matching score is greater than the matching threshold (for example, 0.7), the core layer contradiction type code and preliminary positioning are output; pattern matching verification is simultaneously performed in the negative sample database to verify the consistency of the contradiction features; the matching threshold is based on a combination of statistical analysis of historical data and model performance evaluation.

[0098] In the second channel, the cross-modal fusion features are mapped into a temporal risk map, and the abnormality of the map nodes is calculated using the spatiotemporal graph convolutional network (ST-GCN);

[0099] is the i-th risk graph node, For nodes Correlation derivative layer pattern; WT is the time window; is the abnormality of the i-th risk map node; when the abnormality of the map node is greater than the node abnormality threshold, the archive layer historical case retrieval is triggered, and the derivative layer contradiction type code and precise positioning are output; wherein, over a period of time, such as 30 days, the abnormality score of the new risk pattern is continuously collected and calculated to form time series data; for this time series, its stationarity is first determined by methods such as the ADF test; if the stationarity condition is met, the abnormality score is smoothed using the moving average method or exponential smoothing method to eliminate the impact of short-term fluctuations. Then, based on the statistical characteristics of historical data, such as the mean and standard deviation, the abnormality score values ​​at different quantiles are calculated; at the same time, combined with the experience of business experts and the risk sensitivity requirements in different business scenarios, the appropriate quantile is determined as a threshold reference.

[0100] When the dual-channel detection results conflict, the reinforcement learning arbitration module is triggered to dynamically allocate confidence weights based on the risk level.

[0101] For example, when determining the risk of the core layer, the core layer detection confidence weight is 0.6, and the derivative layer detection confidence weight is 0.4; when determining the risk of the derivative layer, the core layer detection confidence weight is 0.3, and the derivative layer detection confidence weight is 0.7; By combining the core layer arbitration reward function with the layered knowledge system verification set performance, the arbitration reward function is:

[0102] is the core layer accuracy weight; the weight distribution is dynamically optimized according to the reward value R, and the detection result with higher confidence is selected as the final contradiction judgment.

[0103] The effects of the above technical solution are: The first channel uses a lightweight detector based on knowledge distillation combined with graph neural network matching to achieve second-level matching and preliminary positioning of core-layer risk patterns (such as a 0.7 threshold to quickly screen out high-confidence risks). Combined with negative sample database verification, it solves the "high false alarm rate" problem of traditional single-channel detection, reduces the false alarm rate of core-layer contradictions, and ensures the accuracy of basic risk prevention and control.

[0104] The second channel constructs a time-series risk map through the spatiotemporal graph convolutional network (ST-GCN) to explore derivative risk associations across time and nodes (such as the transmission of revenue fluctuations and asset-liability anomalies); combined with historical case verification and anomaly smoothing, it can identify implicit derivative risks missed by traditional rule engines (such as new compliance risks under innovative business models) and improve the detection rate of derivative layer risks.

[0105] Rapid screening of the core layer and in-depth excavation of the derivative layer form a layered detection closed loop. When the core layer is not matched, it will automatically transfer to the derivative layer to avoid risk omission.

[0106] The reinforcement learning arbitration module dynamically allocates confidence weights based on the risk level (core / derivative) (for example, the core channel weight is 0.6 when determining the core layer), and optimizes the weights based on the reward function (precision + recall dual dimensions).

[0107] The second channel processes abnormality scores through ADF tests and moving average / exponential smoothing, and dynamically sets thresholds by combining historical statistical characteristics (mean, standard deviation) with the experience of business experts. This solves the problem of traditional "fixed thresholds being unsuitable for business fluctuations." For example, in innovative business scenarios, the abnormality threshold can be automatically lowered to capture potential risks and improve business scenario adaptability.

[0108] The first channel, a lightweight detector (a knowledge distillation compression model), ensures detection speed and is suitable for high-frequency basic detection. The second channel, deep graph analysis, focuses on complex risks and triggers them on demand. Compared to full-depth detection, this reduces system computing power consumption while ensuring that complex risk detection is not degraded.

[0109] By continuously collecting the abnormality scores of new risk patterns and updating the thresholds, new risks in business changes (such as compliance risks of financial innovation products) can be adaptively captured; the dual-channel detection results (core / derivative contradiction coding) continuously feed back into the layered risk knowledge system, and combined with historical cases in the archive layer, self-iteration of the risk pattern library is achieved.

[0110] This dual-channel conflict detection solution upgrades risk prevention and control from "single-dimensional rule matching" to multi-dimensional intelligent penetration through layered precision detection, intelligent dynamic arbitration, and risk knowledge self-iteration. While ensuring core-layer risk accuracy, it deeply explores hidden risks at the derivative layer and adapts to dynamic changes in business scenarios. This provides a "smarter, more adaptable, and more sustainable" risk control technology foundation for fields such as finance and compliance, significantly reducing risk omission rates and operating costs, and supporting a balanced approach between business innovation and regulatory compliance.

[0111] Refer to the attached Figure 3 ,In a possible implementation, the customer is automatically routed to the corresponding ,detection channel based on the customer’s historical risk level and customer ,type; If the customer type is a new customer, the default initial risk level is 0.5, and the first channel detection is enabled; the new historical risk level is determined based on the first channel detection results; If the customer type is an old customer with no risk and the historical risk level is less than the level threshold, the first channel is activated; If the customer type is a single-conflict old customer or a multiple-conflict old customer, or the historical risk level ≥ the level threshold, the first channel and the second channel are enabled at the same time; the risk knowledge system is dynamically updated based on the conflict detection results to optimize the detection strategy.

[0112] In one possible implementation, the historical risk level of a customer is determined based on the number of times different types of conflicts of the customer have been triggered, the risk pattern weight, severity, and duration;

[0113] k is the contradiction type index, is the hierarchical weight of the kth contradiction type, which is determined by the contradiction source level (core / derivative) and can be determined through historical data and expert experience; ∈[0.3,0.8]; is the severity coefficient of the k-th type of contradiction, ∈[0.5,1], calibrated by expert evaluation and historical data; is the periodic coefficient, ∈[0.05,0.2]; is the duration period of the k-th type of contradiction; is the time attenuation coefficient; ∈[0.05,0.2]; is the interval between the current evaluation of the k-th contradiction and the time when the contradiction is triggered; Among them, the level threshold is determined based on the historical risk level distribution; for example, based on the historical risk level data of the same customer type, the distribution curve is fitted through kernel density estimation, and the 80% quantile is taken as the warning threshold; when the calculated historical risk level is greater than or equal to the warning threshold of the same customer type, dual-channel deep risk detection is triggered, otherwise lightweight detection is performed.

[0114] The lightweight detection channel (first channel) reduces computing power and manpower consumption (reducing overall detection costs when the proportion of new / low-risk customers is high); dual-channel deep detection focuses on high-risk customers, reducing ineffective deep detection. This achieves a balance between risk control cost and effectiveness. Through dynamic threshold adaptation (based on the same customer type quantile) and knowledge system iteration, the solution can quickly adapt to new business scenarios (such as cross-border e-commerce and cryptocurrency trading). For example, when a new business is launched, adaptive thresholds are automatically generated based on risk distribution fit based on historical data, supporting risk control agility for business innovation. Through multi-dimensional historical risk quantification, intelligent customer stratification routing, and closed-loop knowledge system iteration, risk control is upgraded from "experience-driven" to "data-driven intelligence." While ensuring accurate risk identification, it optimizes detection resource allocation and supports the dynamic expansion of business scenarios. This provides a "smarter, more efficient, and more adaptable" customer stratification risk control technology foundation for industries such as finance and retail, significantly reducing risk missed detection rates and operating costs, and improving the scientific nature and explainability of risk control decisions.

[0115] The present application also provides a customer risk intelligent assessment system based on a large language model, which is used to implement the customer risk intelligent assessment method based on a large language model in the embodiment of the present application. The system includes: The data acquisition module is used to obtain and pre-process the customer's multimodal data to extract structured and unstructured features; the multimodal data includes structured business data, unstructured text description data, business scenario images, and multi-round voice interaction data; A system construction module is used to construct a layered risk knowledge system and achieve adaptive evolution of the knowledge system through graph neural networks and generative models. The layered risk knowledge system includes a core layer, a scenario layer, a derivative layer, and an archive layer. A risk pattern vector library is established based on multiple anomaly types, including logical fallacies, contradictory expression paradigms, and abnormal semantics. The negative sample library construction module is used to randomly sample multimodal data that meets the preset conditions to build an initial negative sample library. It combines expert-annotated non-risk patterns and derivative layer analysis to build a negative sample database. The model training module is used to classify and label risk text data by customer type, optimizing the large language model using a collaborative training mechanism of strong supervision, weak supervision, and reinforcement learning for each classification. The customer types include new customers, existing customers without risks, existing customers with a single conflict, and existing customers with multiple conflicts. The output module is used to use the optimized large language model to mine risk features in text, process multimodal data through a multi-level attention network, locate logical contradictions through intra-modal enhancement, cross-modal fusion and contradiction detection, and generate a positioning report that includes contradiction type coding, business impact dimension assessment and risk level quantification.

[0116] The working principle and effect of the above technical solution are the same as those in the embodiment of the method of this application, and will not be repeated here.

[0117] The present application also provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of any method described in the present application are implemented.

[0118] This application is explained from the perspectives of purpose of use, effectiveness, progress and novelty, and has complied with the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings of this application are only preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc. that are similar or identical to those of this application, that is, all equivalent replacements or modifications made in accordance with the scope of the patent application of this application, should fall within the scope of protection of the patent application of this application.

Claims

1. A customer risk intelligent assessment method based on a large language model, characterized by: The method comprises: S1. Obtain and preprocess the customer's multimodal data to extract structured and unstructured features; the multimodal data includes structured business data, unstructured text description data, business scenario images, and multiple rounds of voice interaction data; S2. Construct a layered risk knowledge system and implement adaptive evolution of the knowledge system through graph neural networks and generative models. The layered risk knowledge system includes a core layer, a scenario layer, a derivative layer, and an archive layer. A risk pattern vector library is established based on multiple anomaly types, including logical fallacies, contradictory expression paradigms, and abnormal semantics. S3. Randomly sample multimodal data that meets the preset conditions to build an initial negative sample library. Combined with expert-annotated non-risk patterns and derivative layer analysis, a negative sample database is constructed. S4. Classify and label risk text data by customer type, and optimize the large language model using a collaborative training mechanism of strong supervision, weak supervision, and reinforcement learning for each category; S5. Use the optimized large language model to mine risk features in text, process multimodal data through a multi-level attention network, locate logical contradictions through intra-modal enhancement, cross-modal fusion, and contradiction detection, and generate a positioning report that includes contradiction type coding, business impact dimension assessment, and risk level quantification.

2. The customer risk intelligent assessment method according to claim 1, characterized in that: The S2 includes: Perform cluster analysis on unlabeled text to obtain new risk patterns; Calculate the risk confidence of the new risk model based on the maximum similarity between the new risk model and each historical risk model in the core layer and the frequency of occurrence of the new risk model, and determine the storage level of the new risk model data; Integrate multi-source data, including expert feedback, model discovery, policy analysis, and public opinion monitoring, to automatically trigger updates through NLP and time series analysis. Integrate external knowledge bases through semantic fusion and transfer learning, and combine them with generative large language models to expand and optimize the knowledge base content.

3. The customer risk intelligent assessment method according to claim 2, characterized in that: The method of calculating the risk confidence of the new risk model based on the maximum similarity between the new risk model and each historical risk model of the core layer and the frequency of occurrence of the new risk model, and determining the storage level of the new risk model data, includes: If the maximum similarity is less than the similarity threshold, then calculating the abnormality score of the new risk pattern; The abnormality score of the new risk pattern is calculated by using the new risk pattern vector, the mean and standard deviation of multiple historical pattern vectors in the core layer, and the frequency of occurrence of the new risk pattern. The abnormality score is obtained by the following formula: ； in, is the characteristic vector of the new risk pattern; is the mean of multiple historical pattern vectors in the core layer; is the standard deviation of the historical pattern eigenvector; To quantify the difference between the new pattern feature vector and the mean of the historical pattern vector by the square of the Euclidean distance; is the risk pattern frequency; is the frequency reference value; If the abnormality score is greater than or equal to the first score threshold, it is pushed to the expert annotation queue; the risk pattern data annotated by the expert is stored in the core layer, and the non-risk pattern data is stored in the negative sample library; If the abnormality score is less than the first score threshold, the new risk pattern data is stored in the derivative layer; If the maximum similarity is greater than or equal to the similarity threshold, the corresponding risk knowledge system layer is selected and stored according to the risk confidence level; including: If the risk confidence of the new risk model is greater than or equal to the first confidence threshold, the data of the risk model is stored in the core layer; If the risk confidence of the new risk model is greater than or equal to the second confidence threshold and less than the first confidence threshold, the data of the risk model is stored in the scenario layer; If the risk confidence of the new risk model is lower than the second confidence threshold, the data of the new risk model is stored in the derivative layer; When the frequency of risk patterns in the derivative layer or scenario layer continues to increase for multiple statistical periods, its abnormality score and risk confidence level are recalculated and migrated to the core layer or scenario layer after meeting the standards.

4. The customer risk intelligent assessment method according to claim 2, characterized in that: The method for obtaining the risk confidence level includes: Extract historical risk pattern feature vectors and new risk pattern feature vectors through pre-trained semantic encoders; calculate the maximum similarity between the new risk pattern and each historical risk pattern feature vector as the basic confidence level; Determine the final risk confidence level through basic confidence level, risk pattern frequency and policy correlation coefficient; Among them, the final risk confidence is: ； in, is the final risk confidence level, is the basic confidence level, ∈[0,1], is the risk pattern frequency; is the frequency reference value; is the policy relevance coefficient, ∈[0.7, 1.6]; min() is the minimum value; is the risk confidence of the historical risk pattern that is most similar to the new risk pattern.

5. The customer risk intelligent assessment method according to claim 1, characterized in that: The S4 includes: Perform main task training based on manually labeled samples in the core layer; The weakly supervised training includes positive samples and negative samples, including: The new risk pattern samples in the core layer are used as positive samples, and positive and negative sample pairs for comparative learning are constructed based on the positive samples; Use non-risk patterns annotated by experts as negative samples; Generate adversarial samples based on the derivative layer risk model and expand the training set of old customers with multiple conflicts; Convert the scenario-level business model into regularization constraints; The feature representation learned by the strong supervision branch is distilled to the weak supervision branch to construct a hybrid attention guidance mechanism.

6. The customer risk intelligent assessment method according to claim 1, characterized in that: The intra-modal enhancement includes: A temporal self-attention mechanism is used on structured business data to capture the cyclical fluctuation patterns of historical business indicators; Implement BERT-based semantic enhancement on unstructured text data and reconstruct key risk description fragments through masked language models; An improved SwinTransformer is used to extract local-global features from business scene images, strengthening the focus on digital areas in financial voucher images. A voiceprint feature encoder is constructed for voice interaction data to separate emotional features from semantic features and generate an emotion-semantic joint embedding vector.

7. The customer risk intelligent assessment method according to claim 1, characterized in that: By constructing a fourth-order modal interaction matrix, cross-modal fusion is achieved, including: First-order fusion uses a parameter-sharing cross-modal Transformer to achieve semantic alignment between text and speech; Second-order fusion introduces graph attention network to transform the time series features of structured data into dynamic knowledge graph nodes; Third-order fusion builds a joint latent space of image and text through variational autoencoders; The fourth-order fusion design gated fusion unit adaptively adjusts the weight coefficient of each modality according to the business scenario type.

8. The customer risk intelligent assessment method according to claim 1, characterized in that: Contradiction detection is performed by building a dual-channel conflict detection engine, including: In the first channel, a lightweight detector based on knowledge distillation loads the core layer risk pattern vector library and matches the cross-modal fusion features with the risk pattern vector library through a graph neural network. Based on the matching results, it outputs the core layer conflict type code and preliminary location. Pattern matching verification is also performed on the negative sample database. The second channel maps the multimodal fusion features into a time-series risk map. The spatiotemporal graph convolutional network is used to calculate the abnormality of the map nodes. Based on the abnormality of the map nodes, the archive layer historical cases are triggered for verification, and the derivative layer contradiction type code and precise positioning are output. When the dual-channel detection results conflict, the reinforcement learning arbitration module is triggered to dynamically allocate confidence weights based on the risk level.

9. The customer risk intelligent assessment method according to claim 8, characterized in that: Automatically route customers to the corresponding testing channels based on their historical risk level and customer type: If the customer type is a new customer, an initial risk level is assigned by default and the first channel detection is enabled; a new risk level is determined based on the first channel detection results; If the customer type is an old customer with no risk and the historical risk level is less than the level threshold, the first channel is activated; If the customer type is a single conflict old customer or a multiple conflict old customer, or the historical risk level is greater than or equal to the level threshold, both the first and second channels are activated; Dynamically update the risk knowledge system based on contradiction detection results.

10. A customer risk intelligent assessment system based on a large language model, used to implement the customer risk intelligent assessment method based on a large language model as described in any one of claims 1 to 9, characterized in that: The system comprises: The data acquisition module is used to obtain and pre-process the customer's multimodal data to extract structured and unstructured features; the multimodal data includes structured business data, unstructured text description data, business scenario images, and multi-round voice interaction data; A system construction module is used to construct a layered risk knowledge system and achieve adaptive evolution of the knowledge system through graph neural networks and generative models. The layered risk knowledge system includes a core layer, a scenario layer, a derivative layer, and an archive layer. A risk pattern vector library is established based on multiple anomaly types, including logical fallacies, contradictory expression paradigms, and abnormal semantics. The negative sample library construction module is used to randomly sample multimodal data that meets the preset conditions to build an initial negative sample library. It combines expert-annotated non-risk patterns and derivative layer analysis to build a negative sample database. The model training module is used to classify and label risk text data by customer type, optimizing the large language model using a collaborative training mechanism of strong supervision, weak supervision, and reinforcement learning for each classification. The customer types include new customers, existing customers without risks, existing customers with a single conflict, and existing customers with multiple conflicts. The output module is used to use the optimized large language model to mine risk features in text, process multimodal data through a multi-level attention network, locate logical contradictions through intra-modal enhancement, cross-modal fusion and contradiction detection, and generate a positioning report that includes contradiction type coding, business impact dimension assessment and risk level quantification.

Images