Risk assessment method and device based on large language model, equipment and storage medium

Abstract

The application discloses a risk assessment method and device based on a large language model, equipment and a storage medium, relates to the field of data processing and financial technology, and comprises the following steps: inputting structured business data, credit investigation and text report data and unstructured data into a pre-trained large language model; processing the structured business data and the credit investigation and text report data through a semantic understanding layer of the large language model to generate first feature representation; processing the unstructured data through a multi-modal encoding layer of the large language model to generate second feature representation; fusing and processing the first feature representation and the second feature representation through a cross-modal fusion layer of the large language model to generate comprehensive feature representation; and outputting a risk assessment conclusion for a target financial service application based on the comprehensive feature representation through a decision output layer of the large language model. The accuracy and reliability of risk assessment are improved.

Description

Technical Field

[0001] This invention relates to the fields of data processing and financial technology, and in particular to a risk assessment method, apparatus, device and storage medium based on a large language model. Background Technology

[0002] In existing technologies, risk assessment methods typically rely solely on structured business data and credit and text report data, employing traditional algorithmic models for evaluation. However, materials generated during financial transactions often contain a large amount of unstructured data, such as online interaction records between customers and business personnel, audio and video materials from remote face-to-face interviews, and images of submitted contracts or supporting documents. This unstructured data may carry crucial information reflecting the customer's true intentions, behavioral characteristics, or the authenticity of the materials. Existing methods cannot effectively integrate and understand such multi-source heterogeneous data, thus limiting the accuracy and reliability of risk assessment conclusions and failing to meet the increasingly sophisticated risk control needs in various financial scenarios such as credit approval, insurance underwriting, and investment suitability assessment.

[0003] Therefore, improving the accuracy and reliability of risk assessment has become an urgent technical problem to be solved. Summary of the Invention

[0004] This invention provides a risk assessment method, apparatus, device, and storage medium based on a large language model to address the problem that existing risk assessment methods suffer from insufficient accuracy and reliability in their output risk assessment conclusions.

[0005] Firstly, a risk assessment method based on a large language model is provided, including: Acquire multi-source heterogeneous data associated with the target financial business application, including structured business data, credit and text report data, and unstructured data; The structured business data, the credit reporting and text report data, and the unstructured data are input into a pre-trained large language model; The structured business data and the credit reporting and text report data are processed through the semantic understanding layer of the large language model to generate a first feature representation; The unstructured data is processed through the multimodal coding layer of the large language model to generate a second feature representation; The first feature representation and the second feature representation are fused through the cross-modal fusion layer of the large language model to generate a comprehensive feature representation. Based on the comprehensive feature representation, the decision output layer of the large language model outputs a risk assessment conclusion for the target financial business application.

[0006] Secondly, a risk assessment device based on a large language model is provided, including: The acquisition module is used to acquire multi-source heterogeneous data associated with the target financial business application. The multi-source heterogeneous data includes structured business data, credit and text report data, and unstructured data. The model processing module is used to input the structured business data, the credit reporting and text report data, and the unstructured data into a pre-trained large language model; process the structured business data and the credit reporting and text report data through the semantic understanding layer of the large language model to generate a first feature representation; process the unstructured data through the multimodal coding layer of the large language model to generate a second feature representation; fuse the first feature representation and the second feature representation through the cross-modal fusion layer of the large language model to generate a comprehensive feature representation; and output a risk assessment conclusion for the target financial business application based on the comprehensive feature representation through the decision output layer of the large language model.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned risk assessment method based on a large language model.

[0008] Fourthly, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-described risk assessment method based on a large language model.

[0009] The beneficial effects of a technical solution provided by this invention are as follows: By acquiring multi-source heterogeneous data associated with a target financial business application, including structured business data, credit and text report data, and unstructured data, and inputting this data into a pre-trained large language model, the semantic understanding layer within the large language model performs semantic parsing and logical reasoning on the structured business data and credit and text report data to generate a first feature representation. A multimodal coding layer extracts features from the unstructured data to generate a second feature representation. Then, a cross-modal fusion layer fuses the first and second feature representations to generate a comprehensive feature representation. Finally, the decision output layer outputs a risk assessment conclusion based on the comprehensive feature representation. This not only fully utilizes the credit information in the structured data but also incorporates potential behavioral and intentional features from unstructured materials into the assessment, thereby achieving a more comprehensive and accurate assessment of the target financial business application and significantly improving the accuracy and reliability of the output risk assessment conclusion. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an application environment for a risk assessment method based on a large language model according to an embodiment of the present invention; Figure 2 This is a flowchart of a risk assessment method based on a large language model in one embodiment of the present invention; Figure 3 This is a schematic diagram of a framework of a large language model in one embodiment of the present invention; Figure 4 This is a schematic diagram of a risk assessment device based on a large language model in one embodiment of the present invention; Figure 5 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] The risk assessment method based on a large language model provided in this invention can be applied to, for example... Figure 1 The application environment is shown. Specifically, this risk assessment method based on large language models is applied in a risk assessment system based on large language models, which includes, for example, […]. Figure 1 The diagram illustrates a client and server that communicate over a network to implement steps in a risk assessment method based on a large language model. The client, also known as the user terminal, refers to the program that provides local services to the client, corresponding to the server. Clients can be, but are not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0014] In one embodiment, such as Figure 2 and 3 As shown, a risk assessment method based on a large language model is provided, which is then applied to... Figure 1Taking the server in the example, the following steps are included: S101: Obtain multi-source heterogeneous data associated with the target financial business application. Multi-source heterogeneous data includes structured business data, credit and text report data, and unstructured data.

[0015] It is understood that in this embodiment, the target financial business application refers to the financial business instance to be intelligently evaluated, such as, but not limited to, loan applications and insurance underwriting requests; multi-source heterogeneous data refers to the data set associated with the target financial business application, which may include, but is not limited to, structured business data, credit and text report data, and unstructured data.

[0016] Structured business data may include, but is not limited to, basic information of the applicant / policyholder / investor, details of business requests, historical financial behavior records and transaction flow data; credit and text report data may include, but is not limited to, credit reports, insurance medical examination reports, corporate audit reports or financial report texts; unstructured data may include, but is not limited to, customer and business personnel interaction data, business processing related images, contract document images and call recording data.

[0017] As an example, when executing step S101 in the financial business approval system, the system can obtain the associated structured business data and unstructured data from the internal business database through a pre-set interface based on the unique identifier of the target financial business request, and obtain the associated credit and text report data from an authorized credit reporting service system, insurance appraisal agency, or third-party data service provider.

[0018] S102: Input multi-source heterogeneous data into a pre-trained large language model; through the semantic understanding layer of the large language model, perform semantic parsing and logical reasoning on structured business data and credit reporting and text report data to generate a first feature representation; through the multimodal coding layer of the large language model, extract features from unstructured data to generate a second feature representation; through the cross-modal fusion layer of the large language model, fuse the first and second feature representations to generate a comprehensive feature representation; through the decision output layer of the large language model, output a risk assessment conclusion for the target financial business application based on the comprehensive feature representation.

[0019] It is understood that in this embodiment, the pre-trained large language model is a multi-module cognitive computing model designed specifically for loan approval scenarios. Its internal layers work together to achieve deep understanding and decision-making on multi-source heterogeneous data.

[0020] The pre-trained large language model can include a risk assessment large language model. Furthermore, the large language model can include specific examples such as... Figure 3The diagram shows a semantic understanding layer, a multimodal coding layer, a cross-modal fusion layer, and a decision output layer; the cross-modal fusion layer is connected to the semantic understanding layer, the multimodal coding layer, and the decision output layer, respectively.

[0021] As an example, in step S102, the structured business data, credit report data, and unstructured application materials obtained in step S101 are input into a pre-built large language model. The semantic understanding layer first parses the structured business data and credit report data, identifying key credit indicators and potential risk points through logical reasoning, generating a first feature representation. The multimodal encoding layer encodes customer-business personnel interaction corpora, business images, contract images, and call recordings, generating a second feature representation. The cross-modal fusion layer correlates and fuses the first and second feature representations to obtain a comprehensive feature representation. Finally, the decision output layer automatically generates a risk assessment conclusion based on the comprehensive feature representation. This risk assessment conclusion can include a comprehensive risk quantification value and the text supporting the risk assessment. This approach not only fully utilizes the credit information in the structured data but also incorporates potential behavioral and intentional characteristics from unstructured materials into the assessment, thereby achieving a more comprehensive and accurate evaluation of the target financial business application and significantly improving the comprehensiveness, accuracy, and reliability of the output results.

[0022] The following is a detailed explanation of steps S101 to S102, using Zhang XX's personal auto loan application as an example: Suppose the business approval system receives a car loan application submitted by Zhang XX. In step S101, based on the identifier of the business request, all related data are concurrently retrieved through a pre-defined data interface. First, the internal business database is accessed to obtain local data bound to the request, including: the online application form (stating occupation "engineer" and monthly income "30,000 yuan"), historical credit records (including a paid-off mortgage, a currently repayable personal loan, and a short-term overdue record from four years ago), and structured business data such as bank account transaction records for the past year; as well as unstructured data such as the complete chat history between Zhang XX and the account manager in WeChat Work, and interactive video files recorded during the remote video interview. Second, a pre-defined secure dedicated line is used to connect to an authorized credit reporting service system to obtain the credit report text matching the request. All data, after being anonymized, is integrated into a unified data object.

[0023] In step S102, the data object is input into a pre-trained financial risk assessment big language model. The structured business data and credit report text are sent to the semantic understanding layer; the unstructured data (chat logs and video files) are sent to the multimodal encoding layer. The semantic understanding layer performs semantic parsing and logical reasoning on the received structured business data and credit report text to generate a first feature representation reflecting Zhang XX's objective credit status. Simultaneously, the multimodal encoding layer extracts features from the received chat logs and video data to generate a second feature representation encoding his behavior and communication intentions. Subsequently, the cross-modal fusion layer receives and fuses the first and second feature representations to generate a comprehensive feature representation that fully characterizes Zhang XX's credit risk. Finally, the decision output layer, based on this comprehensive feature representation, outputs a risk assessment conclusion for the auto loan application, including a comprehensive risk quantification value and the risk assessment basis text.

[0024] It should be noted that the loan application mentioned above is only a specific application example of financial risk assessment. This method is also applicable to other financial risk control scenarios such as insurance underwriting (assessing the risk of the insured and the insured object) and investment suitability assessment (assessing the investor's risk tolerance and product matching). Its core process and data integration logic are consistent, with only the specific data sources and assessment dimensions being emphasized according to business characteristics.

[0025] In summary, the beneficial effects of the technical solution provided by this embodiment of the invention are as follows: By acquiring multi-source heterogeneous data associated with the target financial business application, including structured business data, credit and text report data, and unstructured data, and inputting this data into a pre-trained large language model, the semantic understanding layer within the large language model performs semantic parsing and logical reasoning on the structured business data and credit and text report data to generate a first feature representation. A multimodal coding layer extracts features from the unstructured data to generate a second feature representation. Then, a cross-modal fusion layer fuses the first and second feature representations to generate a comprehensive feature representation. Finally, the decision output layer outputs a risk assessment conclusion based on the comprehensive feature representation. This not only fully utilizes the credit information in the structured data but also incorporates potential behavioral and intentional features from unstructured materials into the assessment, thereby achieving a more comprehensive and accurate assessment of the target financial business application and significantly improving the comprehensiveness, accuracy, and reliability of the output results.

[0026] In one embodiment, specifically in step S102, which involves semantic parsing and logical reasoning of structured business data and credit reporting and text report data through the semantic understanding layer of the large language model to generate a first feature representation, the following steps are included: S121A: Through the semantic encoding sublayer of the semantic understanding layer, structured business data and credit reporting and text report data are mapped to a unified semantic space to generate the first semantic representation; S122A: Through the logical reasoning sublayer of the semantic understanding layer, causal reasoning and / or constraint consistency reasoning are performed on the first semantic representation to generate the first feature representation.

[0027] As an example, the process of generating the first feature representation through the semantic understanding layer includes two sequentially executed sub-steps. First, step S121A is executed: through the semantic encoding sub-layer, structured business data and credit reporting and text report data are mapped to a unified semantic space to generate the first semantic representation. This step achieves standardization and semantic alignment of heterogeneous data, solving the problem that multi-source heterogeneous data cannot be directly compared and fused in a deep semantic way due to differences in format and terminology. Subsequently, step S122A is executed: through the logical reasoning sub-layer, causal relationship reasoning and / or constraint consistency reasoning are performed on the first semantic representation to generate the first feature representation. This step achieves deep logical analysis and credit mining based on unified semantics, overcoming the limitation of traditional methods that can only perform shallow pattern matching. The two steps work together to generate a high-quality feature representation from the original data that contains deep and interpretable credit insights.

[0028] For example, taking the assessment of Zhang XX's loan application as an example, its structured business data includes numerical values ​​such as monthly income and historical liabilities, while the credit reporting and text report data includes credit report text. In step S121A, the semantic encoding sublayer maps these numerical values ​​and text to the same semantic space, generating a comparable first semantic representation. In step S122B, the logical reasoning sublayer analyzes this representation: through causal reasoning, it determines that historical delinquencies have a weak impact on current credit; through constrained consistency reasoning, it verifies the logical rationality between income, liabilities, and cash flow. Finally, a first feature representation is generated, for example, "Stable debt repayment ability, historical flaws have been corrected, and financial data is self-consistent."

[0029] In one embodiment, specifically in step S102, feature extraction of unstructured data is performed through the multimodal coding layer of a large language model to generate a second feature representation, including the following steps: S121B: The multimodal coding layer uses a dedicated multimodal encoder group to encode interactive corpora, business processing related images, contract document images, and call recording data to obtain corresponding modal features; the dedicated multimodal encoder group includes at least a text encoder, a visual encoder, and an audio encoder. S122B: Through the modality normalization sublayer of the multimodal coding layer, the modal features output by different encoders are normalized in dimension, mapped to a unified vector dimension space, and a second feature representation is generated.

[0030] As an example, the process of generating the second feature representation through a multimodal coding layer includes two sequentially executed sub-steps. First, step S121B is executed: the interactive corpus is encoded using a text encoder, the business images and contract images are encoded using a visual encoder, and the call recording data is encoded using an audio encoder, resulting in text modal features, visual modal features, and audio modal features, respectively. This step achieves deep extraction of textual dialogue intent, visual behavioral dynamics, and audio emotional features, overcoming the technical bottleneck of traditional models' inability to effectively utilize such unstructured rich media information. Subsequently, step S122B is executed: through a modality normalization sub-layer, the multiple modal features output by different encoders are mapped to a unified vector dimension space, generating the second feature representation. This step effectively eliminates the heterogeneity between different modal features, achieving feature alignment and unified representation, laying a directly usable foundation for subsequent cross-modal deep fusion.

[0031] For example, taking the assessment of Zhang XX's loan application as an example, its unstructured data includes chat logs and interview videos. In step S121B, the text encoder extracts the text modal feature of "planning to repay the loan with the year-end bonus" from the chat logs, and the visual encoder extracts the visual modal feature of "brief hesitation" when the applicant answers stress questions from the video. In step S122B, the modality normalization sublayer maps these two different dimensional features to the same vector space to generate a second feature representation, for example, a vector encoding semantics such as "clear intention to repay but slight hesitation in response to unexpected events".

[0032] In one embodiment, specifically in step S102, where the first feature representation and the second feature representation are fused through the cross-modal fusion layer of the large language model to generate a comprehensive feature representation, the following steps are included: S121C: Through the attention fusion sublayer of the cross-modal fusion layer, bidirectional attention is calculated based on the first feature representation and the second feature representation to generate bidirectional attention features; S122C: Through the gated fusion unit of the cross-modal fusion layer, gate weights are generated based on bidirectional attention features, and the first feature representation and the second feature representation are fused according to the gate weights to generate a comprehensive feature representation.

[0033] As an example, the process of generating a comprehensive feature representation through a cross-modal fusion layer includes two sequentially executed sub-steps. First, step S121C is executed: a bidirectional attention feature is generated by performing bidirectional attention calculation based on the first and second feature representations through an attention fusion sub-layer. This step models and quantifies the deep semantic association between the two types of heterogeneous features. Subsequently, step S122C is executed: a gating fusion unit generates gating weights based on the bidirectional attention feature, and the first and second feature representations are fused according to these gating weights to generate a comprehensive feature representation. This step achieves adaptive information fusion based on the association strength, generating a fused representation that supports the final decision.

[0034] For example, consider evaluating Zhang XX's auto loan application. The first feature representation encodes "stable income but debt ratio close to the threshold," and the second feature representation encodes "clear repayment intention but slight hesitation in dealing with unexpected events." In the cross-modal fusion layer, firstly, the attention fusion sublayer performs bidirectional attention calculation. This calculation not only finds a strong correlation between the "hesitant behavior" segment in the second feature representation and the "high debt" segment in the first feature representation, but also finds a strong correlation between "clear repayment intention" in the second feature representation and "stable income" in the first feature representation. The bidirectional attention feature generated by this calculation is a matrix quantifying the strength of these two sets of key associations. Next, the gating fusion unit performs calculations based on this bidirectional attention feature. Its gating network may generate a set of dynamic gating weights, which assign higher fusion weights to mutually corroborating risk associations ("high debt" and "hesitant behavior"), thus highlighting risk warnings in the output; while assigning moderate weights to mutually corroborating positive associations ("stable income" and "clear repayment intention") for preservation. Ultimately, the gating fusion unit merges the first feature representation and the second feature representation based on this weight. The resulting comprehensive feature representation is no longer a list of isolated facts, but rather an integrated representation of the fusion judgment that "although the applicant has a stable fundamental and a clear repayment plan, their high debt status and behavioral feedback under stress test jointly reveal the debt repayment flexibility risk that needs careful attention."

[0035] In one embodiment, specifically in step S102, where the decision output layer of the large language model outputs a risk assessment conclusion for the target financial business application based on comprehensive feature representation, the steps include: S121D: Through the risk scoring sub-layer of the decision output layer, the comprehensive feature representation is non-linearly mapped to generate a comprehensive risk quantification value; S122D: Through the decision interpretation sub-layer of the decision output layer, risk assessment basis text is generated based on comprehensive feature representation, comprehensive risk quantification value, and intermediate semantic information extracted from the first feature representation and the second feature representation; The risk assessment conclusion includes a comprehensive risk quantification value and the text on which the risk assessment is based.

[0036] As an example, the process of generating a risk assessment conclusion through the decision output layer includes two parallel sub-steps. First, step S121D is executed: through the risk scoring sub-layer, a nonlinear mapping is performed on the comprehensive feature representation to generate a comprehensive risk quantification value. This step transforms the fused high-dimensional features into an intuitive and comparable quantitative risk assessment indicator. Simultaneously, step S122D is executed: through the decision interpretation sub-layer, based on the comprehensive feature representation, the comprehensive risk quantification value, and intermediate semantic information extracted from the first and second feature representations, a risk assessment basis text is generated. This step transforms the complex model reasoning process and quantification results into a readable, understandable, and evidence-supported natural language explanation, thereby ensuring the transparency and traceability of the decision. The risk assessment conclusion is composed of both the comprehensive risk quantification value and the risk assessment basis text.

[0037] For example, consider the assessment of Zhang XX's loan application. In step S121D, the risk scoring sublayer processes the comprehensive feature representation that incorporates "stable fundamentals but elastic risk," and outputs a specific comprehensive risk quantification value, such as 75 points, through nonlinear function mapping. Simultaneously, in step S122D, the decision interpretation sublayer receives this comprehensive feature representation, the risk value (75 points), and calls upon key intermediate semantic information generated during the generation of the first and second feature representations (e.g., "historical delinquencies have been resolved," "high debt," "correlation between behavioral hesitation"). Based on these inputs, the decision interpretation sublayer generates a risk assessment basis text, such as: "After assessment, the applicant has stable income, a generally good credit history, and a clear willingness to repay. However, their current debt level is high, and their performance in handling stressful scenarios during the video interview suggests a marginal deficiency in their debt repayment elasticity. A comprehensive score of 75 points is recommended, with conditional approval and an adjustment of the loan amount to 85% of the application amount." The final output risk assessment conclusion includes both the quantitative decision of "75 points" and the aforementioned structured interpretation text.

[0038] In one embodiment, such as Figure 2 As shown, after step S102, the loan assessment result includes a comprehensive risk quantification value and an assessment opinion text. That is, after the large language model outputs the loan assessment result for the target loan application, the method includes: S103: If the comprehensive risk quantification value is less than the first preset threshold, a business approval notice will be automatically generated and sent to the applicant's account associated with the target business application. The approval notice shall at least contain the risk assessment basis text. S104: If the comprehensive risk quantification value is greater than or equal to the first preset threshold and less than the second preset threshold, the target business application will be automatically marked as pending manual review and a pending review reminder will be sent to the reviewer's account of the target reviewer. The pending review reminder shall at least contain the risk assessment basis text. S105: If the comprehensive risk quantification value is greater than or equal to the second preset threshold, a rejection notice will be automatically generated and sent to the applicant's account associated with the target business application. The business rejection notice shall at least contain the risk assessment basis text.

[0039] As an example, after generating the risk assessment conclusion, an automated decision-making execution process directly linked to the assessment conclusion is also included. This process specifically comprises three mutually exclusive execution branches, automatically triggered based on a preset risk range to which the comprehensive risk quantification value belongs: If the overall risk quantification value is less than the first preset threshold (e.g., below 60 points), then step S103 is executed: a business approval notification is automatically generated and sent to the applicant's account associated with the target business application (e.g., mobile app, SMS, or email). The approval notification must contain at least the text of the risk assessment basis. This step achieves second-level, automated closed-loop processing and customer outreach for low-risk applications without manual intervention, greatly improving approval efficiency and customer experience, while ensuring decision-making transparency by conveying the assessment basis.

[0040] If the comprehensive risk quantification value is greater than or equal to the first preset threshold and less than the second preset threshold (e.g., between 60 and 80 points), then step S104 is executed: the target business application is automatically marked as pending manual review in the approval system, and a pending review reminder is sent to the pre-designated reviewer's account (such as internal workbench, WeChat Work, or email). The pending review reminder must contain at least the text supporting the risk assessment. This step achieves precise task allocation through human-machine collaboration. The model not only makes judgments but also provides focused risk clues and decision-making basis for manual review through structured opinions, improving the pertinence and efficiency of manual review.

[0041] If the overall risk quantification value is greater than or equal to the second preset threshold (e.g., higher than or equal to 80 points), then step S105 is executed: a rejection notice is automatically generated and sent to the applicant's account associated with the target business application. The rejection notice must at least include the text of the risk assessment basis. This step, while implementing the high-risk business rejection strategy, fulfills the obligation of disclosure by providing an assessment opinion containing specific risk reasons, which helps to improve customer trust and compliance levels and may reduce disputes.

[0042] For example, taking a credit approval scenario, Zhang XX's auto loan application generates a comprehensive risk quantification score of 75. Assume the first threshold is preset to 60 (low-risk upper limit) and the second threshold to 80 (high-risk lower limit). Since 75 falls between 60 and 80, step S104 is automatically executed: First, the application's work order status is updated to "Pending manual review" in the background; simultaneously, a pending review reminder is sent to the workbench account of the account manager responsible for this region or product line, with the message: "Application No. [XXX] (Client: Zhang XX) requires manual review, with a comprehensive risk quantification score of 75. Risk assessment basis text: The applicant has stable income and a generally good credit history, but their current debt level is high, and their performance in handling stressful situations during the video interview suggests a marginal deficiency in their debt repayment flexibility. It is recommended to focus on reviewing their debt-bearing capacity and contingency plans for unexpected expenditures." This allows the account manager to conduct an efficient review based on this clear guidance. If Zhang XX's score is 55, step S103 will be executed, automatically generating an approval notification and sending it to their mobile app. If the score is 85, then proceed to step S105, automatically generate a rejection notification and send it to the user's registered mobile phone number.

[0043] It should be noted that the above example of credit approval illustrates an automated decision-making process. In an insurance underwriting scenario, the corresponding notifications might be "Notification of Successful Underwriting," "Notification of Supplementary Underwriting Materials," or "Notification of Rejection." In an investment suitability assessment scenario, the notifications might be "Notification of Activation of Trading Access," "Notification of Confirmation of Risk Warning," or "Notification of Trading Restriction," etc.

[0044] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0045] In one embodiment, a risk assessment device based on a large language model is provided, which corresponds one-to-one with the risk assessment method based on a large language model described in the above embodiments. For example... Figure 4 As shown, the risk assessment device based on a large language model includes an acquisition module 201 and a model processing module 202. Detailed descriptions of each functional module are as follows: The acquisition module 201 is used to acquire multi-source heterogeneous data associated with the target financial business application. The multi-source heterogeneous data includes structured business data, credit and text report data, and unstructured data. The model processing module 202 is used to input the structured business data, the credit reporting and text report data, and the unstructured data into a pre-trained large language model; process the structured business data and the credit reporting and text report data through the semantic understanding layer of the large language model to generate a first feature representation; process the unstructured data through the multimodal coding layer of the large language model to generate a second feature representation; fuse the first feature representation and the second feature representation through the cross-modal fusion layer of the large language model to generate a comprehensive feature representation; and output a risk assessment conclusion for the target financial business application based on the comprehensive feature representation through the decision output layer of the large language model.

[0046] In one embodiment, the structured business data includes basic information of the applicant / policyholder / investor, details of the business request, historical financial behavior records and transaction flow data; the credit and text report data includes credit reports, corporate audit reports or financial report texts; and the unstructured data includes customer and business personnel interaction data, business processing related images, contract document images and call recording data.

[0047] In one embodiment, the model processing module 202 is further configured to: Through the semantic encoding sublayer of the semantic understanding layer, the structured business data and the credit reporting and text report data are mapped to a unified semantic space to generate a first semantic representation; The first feature representation is generated by performing causal reasoning and / or constraint consistency reasoning on the first semantic representation through the logical reasoning sublayer of the semantic understanding layer.

[0048] In one embodiment, the model processing module 202 is further configured to: The multimodal dedicated encoder group of the multimodal coding layer encodes the interactive corpus, the business processing related images, the contract document images, and the call recording data to obtain corresponding modal features; wherein, the multimodal dedicated encoder group includes at least a text encoder, a visual encoder, and an audio encoder. The modality normalization sublayer of the multimodal coding layer performs dimensionality normalization on the modal features output by different encoders, mapping them to a unified vector dimension space to generate the second feature representation.

[0049] In one embodiment, the model processing module 202 is further configured to: Through the attention fusion sublayer of the cross-modal fusion layer, bidirectional attention calculation is performed based on the first feature representation and the second feature representation to generate bidirectional attention features; The gating fusion unit of the cross-modal fusion layer generates gating weights based on the bidirectional attention features, and fuses the first feature representation and the second feature representation according to the gating weights to generate the comprehensive feature representation.

[0050] In one embodiment, the model processing module 202 is further configured to: The risk scoring sub-layer of the decision output layer performs a non-linear mapping on the comprehensive feature representation to generate a comprehensive risk quantification value. Through the decision interpretation sublayer of the decision output layer, a risk assessment basis text is generated based on the comprehensive feature representation, the comprehensive risk quantification value, and the intermediate semantic information extracted from the first feature representation and the second feature representation; The risk assessment conclusion includes the comprehensive risk quantification value and the risk assessment basis text.

[0051] In one embodiment, the risk assessment conclusion includes the comprehensive risk quantification value and the risk assessment basis text, and the device further includes: The first sending module is used to automatically generate a business approval notice and send it to the applicant's account associated with the target business application if the comprehensive risk quantification value is less than a first preset threshold. The approval notice contains at least the risk assessment basis text. The second sending module is used to automatically mark the target business application as pending manual review if the comprehensive risk quantification value is greater than or equal to the first preset threshold and less than the second preset threshold, and send a pending review reminder to the reviewer's account of the target reviewer. The pending review reminder contains at least the risk assessment basis text. The third sending module is used to automatically generate a rejection notification and send it to the applicant's account associated with the target business application if the comprehensive risk quantification value is greater than or equal to the second preset threshold. The business rejection notification contains at least the risk assessment basis text.

[0052] Specific limitations regarding the risk assessment device based on large language models can be found in the limitations of the risk assessment method based on large language models mentioned above, and will not be repeated here. Each module in the aforementioned risk assessment device based on large language models can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0053] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores the data required to execute the risk assessment method based on a large language model. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a risk assessment method based on a large language model.

[0054] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the risk assessment method based on a large language model described in the above embodiments, for example... Figure 2 S101-S105, as shown, will not be described again here to avoid repetition. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in this embodiment of the risk assessment device based on a large language model, for example... Figure 4 The functions of the risk assessment device based on the large language model shown are not described in detail here to avoid repetition.

[0055] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the risk assessment method based on a large language model described in the above embodiments, for example... Figure 2 S101-S105, as shown, will not be described again here to avoid repetition. Alternatively, when the computer program is executed by the processor, it implements the functions of each module / unit in this embodiment of the risk assessment device based on a large language model, for example... Figure 4 The functions of the risk assessment device based on the large language model shown are not described again here to avoid repetition. The computer-readable storage medium can be non-volatile or volatile.

[0056] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0057] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0058] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A risk assessment method based on a large language model, characterized in that, include: Acquire multi-source heterogeneous data associated with the target financial business application, including structured business data, credit and text report data, and unstructured data; The structured business data, the credit reporting and text report data, and the unstructured data are input into a pre-trained large language model; The structured business data and the credit reporting and text report data are processed through the semantic understanding layer of the large language model to generate a first feature representation; The unstructured data is processed through the multimodal coding layer of the large language model to generate a second feature representation; The first feature representation and the second feature representation are fused through the cross-modal fusion layer of the large language model to generate a comprehensive feature representation. Based on the comprehensive feature representation, the decision output layer of the large language model outputs a risk assessment conclusion for the target financial business application.

2. The method according to claim 1, characterized in that, The structured business data includes applicant basic information, business request details, historical financial behavior records and transaction flow data; the credit and text report data includes credit reports, corporate audit reports or financial report texts; the unstructured data includes customer and business personnel interaction data, business processing related images, contract document images and call recording data.

3. The method according to claim 1, characterized in that, The process of processing the structured business data and the credit reporting and text report data through the semantic understanding layer of the large language model to generate a first feature representation includes: Through the semantic encoding sublayer of the semantic understanding layer, the structured business data and the credit reporting and text report data are mapped to a unified semantic space to generate a first semantic representation; The first feature representation is generated by performing causal reasoning and / or constraint consistency reasoning on the first semantic representation through the logical reasoning sublayer of the semantic understanding layer.

4. The method according to claim 2, characterized in that, The process of processing the unstructured data through the multimodal coding layer of the large language model to generate a second feature representation includes: The multimodal dedicated encoder group of the multimodal coding layer encodes the interactive corpus, the business processing related images, the contract document images, and the call recording data to obtain corresponding modal features; wherein, the multimodal dedicated encoder group includes at least a text encoder, a visual encoder, and an audio encoder. The modality normalization sublayer of the multimodal coding layer performs dimensionality normalization on the modal features output by different encoders, mapping them to a unified vector dimension space to generate the second feature representation.

5. The method according to claim 1, characterized in that, The process of fusing the first feature representation and the second feature representation through the cross-modal fusion layer of the large language model to generate a comprehensive feature representation includes: Through the attention fusion sublayer of the cross-modal fusion layer, bidirectional attention calculation is performed based on the first feature representation and the second feature representation to generate bidirectional attention features; The gating fusion unit of the cross-modal fusion layer generates gating weights based on the bidirectional attention features, and fuses the first feature representation and the second feature representation according to the gating weights to generate the comprehensive feature representation.

6. The method according to claim 1, characterized in that, The decision output layer of the large language model, based on the comprehensive feature representation, outputs a risk assessment conclusion for the target financial business application, including: The risk scoring sub-layer of the decision output layer performs a non-linear mapping on the comprehensive feature representation to generate a comprehensive risk quantification value. Through the decision interpretation sublayer of the decision output layer, a risk assessment basis text is generated based on the comprehensive feature representation, the comprehensive risk quantification value, and the intermediate semantic information extracted from the first feature representation and the second feature representation; The risk assessment conclusion includes the comprehensive risk quantification value and the risk assessment basis text.

7. The method according to claim 1, characterized in that, The risk assessment conclusion includes a comprehensive risk quantification value and a risk assessment basis text. After the large language model outputs the risk assessment conclusion for the target business application, the method includes: If the comprehensive risk quantification value is less than the first preset threshold, a business approval notice will be automatically generated and sent to the applicant's account associated with the target business application. The approval notice shall at least contain the risk assessment basis text. If the comprehensive risk quantification value is greater than or equal to the first preset threshold and less than the second preset threshold, the target business application will be automatically marked as pending manual review, and a pending review reminder will be sent to the reviewer's account. The pending review reminder will at least contain the risk assessment basis text. If the comprehensive risk quantification value is greater than or equal to the second preset threshold, a rejection notice will be automatically generated and sent to the applicant's account associated with the target business application. The business rejection notice shall at least contain the risk assessment basis text.

8. A risk assessment device based on a large language model, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous data associated with the target financial business application. The multi-source heterogeneous data includes structured business data, credit and text report data, and unstructured data. The model processing module is used to input the structured business data, the credit investigation and text report data, and the unstructured data into a pre-trained large language model; and to process the structured business data and the credit investigation and text report data through the semantic understanding layer of the large language model to generate a first feature representation. The unstructured data is processed through the multimodal coding layer of the large language model to generate a second feature representation; the first feature representation and the second feature representation are fused through the cross-modal fusion layer of the large language model to generate a comprehensive feature representation; and the risk assessment conclusion for the target financial business application is output based on the comprehensive feature representation through the decision output layer of the large language model.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the risk assessment method based on a large language model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the risk assessment method based on a large language model as described in any one of claims 1 to 7.

Images