A bank credit risk control large model generation result reliability evaluation technology

Abstract

The present application relates to the technical field of artificial intelligence and financial technology, and in particular to a bank credit risk control large model generation result reliability evaluation technology, which adopts a hierarchical modular architecture, and from bottom to top, there are data layer, knowledge layer, evaluation engine layer, fusion decision layer and application interface layer. The technology constructs a multi-dimensional reliability evaluation framework, establishes a fact verification engine based on a financial knowledge graph, designs a robustness evaluation method based on adversarial samples, develops an illusion detection technology for the financial field, realizes dynamic calibration and uncertainty quantification, and provides systematic evaluation technology support for the safe and reliable application of large models in the credit risk control scenario of commercial banks.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technologies of artificial intelligence and financial technology, and in particular to the technology of output quality assessment and reliability assurance of Large Language Model (LLM) in the context of credit risk management in commercial banks. Background Technology

[0002] With the rapid development of artificial intelligence technology, Large Language Models (LLMs) have been widely applied in the field of credit risk control in commercial banks, including but not limited to: intelligent generation of credit approval opinions, writing of customer credit risk assessment reports, intelligent review of loan contract terms, automatic interpretation of risk warning information, analysis of fraudulent and abnormal behavior, and compliance checks of regulatory policies. Leveraging their powerful natural language understanding and generation capabilities, LLMs can effectively improve the processing efficiency and intelligence level of credit risk control operations.

[0003] In the field of large model reliability assessment, existing technologies have made some progress. Chinese patent application CN117851566A discloses a cross-document question-answering knowledge extraction method, system, and electronic device, which verifies the facts of content generated by large models by constructing a domain knowledge graph. However, this method mainly focuses on verifying the accuracy of facts in general domains and does not specifically address the special needs of financial lending scenarios (such as regulatory compliance and risk sensitivity), resulting in a relatively singular evaluation dimension. Chinese patent application CN118070008A discloses a multi-level time series forecasting method and device, employing a multiple sampling consistency analysis strategy to identify illusory content in the output of large models. However, this method does not systematically integrate with knowledge graph verification, uncertainty quantification, and other technologies, making it difficult to meet the needs of multi-dimensional comprehensive evaluation in financial decision-making scenarios. In academic research, Manakul et al. proposed a self-consistency-based hallucination detection method in their paper "SelfCheckGPT: Zero-Resource Black-Box HallucinationDetection for Generative Large Language Models" (published at the 2023 EMNLP conference). However, this method is mainly aimed at open-domain text generation scenarios and lacks the ability to evaluate professional dimensions such as terminology standardization and policy compliance in the financial field. Furthermore, it does not consider the actual needs of uncertainty quantification and manual review priority ranking in credit risk control scenarios.

[0004] However, the application of large-scale models in the field of credit risk control in commercial banks faces severe reliability challenges, and existing technologies have the following prominent problems:

[0005] First, the "illusion" problem seriously threatens the security of financial decision-making. Large models may generate information that appears reasonable but is actually incorrect, such as fabricating non-existent regulatory provisions, fictitious customer transaction records, and fabricating industry risk data. In critical decision-making scenarios such as credit approval, this could lead to significant economic losses and compliance risks.

[0006] Second, there is a lack of effective means to verify the accuracy of facts. Existing technologies are insufficient to effectively verify the accuracy of key facts such as financial data, regulatory policies, and customer information in the results generated by large models, and there is a lack of specialized fact-checking mechanisms for the financial sector.

[0007] Third, the ability to detect logical consistency is insufficient. When large models deal with complex credit risk analysis scenarios, problems such as logical inconsistencies, causal errors, and broken reasoning chains may occur. Existing technologies lack systematic methods for detecting logical consistency.

[0008] Fourth, there are blind spots in financial compliance assessment. There is a lack of automated compliance assessment technology to determine whether the content generated by the large model complies with the requirements of laws and regulations such as the Commercial Banking Law, the General Rules for Loans, and the regulations of the State Financial Regulatory Commission.

[0009] Fifth, there is a lack of uncertainty quantification mechanisms. Existing technologies cannot effectively assess the confidence level and uncertainty of large models under different credit risk control scenarios, making it difficult to provide a scientific basis for prioritizing subsequent manual reviews.

[0010] Sixth, there is a lack of unified reliability assessment standards and frameworks. Currently, there are no systematic reliability assessment standards in the industry for the application of large-scale models in the field of credit risk control in commercial banks. The assessment methods adopted by various institutions are inconsistent, making it difficult to conduct horizontal comparisons and benchmark assessments.

[0011] In summary, existing general large model evaluation methods suffer from problems such as limited evaluation dimensions, lack of financial domain specificity, and failure to achieve multi-technology synergy and integration. There is an urgent need for a systematic large model generation reliability evaluation technology for the field of commercial bank credit risk control. Summary of the Invention

[0012] The purpose of this invention is to provide a reliability evaluation technology for the generation results of a large-scale bank credit risk control model, in order to solve the problems mentioned in the background art, such as the difficulty in detecting the "illusion" problem of large-scale models, the lack of effective means to verify the accuracy of facts, the insufficient ability to detect logical consistency, the existence of blind spots in financial compliance judgment, and the lack of uncertainty quantification mechanism.

[0013] To achieve the above objectives, the present invention provides the following technical solution: a reliability evaluation system for the generation results of a large-scale bank credit risk control model, comprising:

[0014] The data layer is used to store the original input data, generated result data, and external financial data used as a verification benchmark for the large model. The data layer adopts a unified data model to manage the above-mentioned data in a structured manner and maintains the timeliness and traceability of the data through a data version control mechanism.

[0015] The knowledge layer is used to store and manage financial knowledge graphs, regulatory rule bases, financial terminology dictionaries, and historical case databases. The financial knowledge graph stores financial entities, their relationships, and attributes in the form of triples, and the regulatory rule base stores financial regulations and their applicable conditions in the form of structured rule entries.

[0016] The evaluation engine layer is used to perform multi-dimensional reliability evaluations. It includes a multi-dimensional reliability evaluation framework module, a financial knowledge graph fact-checking engine module, a financial illusion detection technology module, a dynamic calibration and uncertainty quantification module, and an adversarial robustness evaluation module. The evaluation engine layer coordinates the parallel execution of each module through a unified evaluation task scheduler, and the modules pass intermediate results through a message queue. The detection results of the financial illusion detection technology module are used to assist the financial knowledge graph fact-checking engine module in determining the key verification scope, and the quantification results of the dynamic calibration and uncertainty quantification module serve as the input basis for the fusion decision layer to calculate the priority of manual review.

[0017] The fusion decision layer is used to receive and fuse the evaluation results from the evaluation engine layer, weight and fuse the scores of each dimension based on the preset business scenario weights to calculate the comprehensive reliability score, and generate an evaluation report containing review priority suggestions by combining the uncertainty quantification results.

[0018] The application interface layer provides standardized RESTful API interfaces to receive assessment requests, return assessment results, and provide a visual report interface.

[0019] Preferably, the multi-dimensional reliability evaluation framework module is configured to evaluate the large model generation results in at least six dimensions, including:

[0020] Factual Accuracy (FA): This dimension assesses the accuracy of factual content such as financial data, customer information, market indicators, and policies and regulations in the generated results. It adopts a knowledge graph-based fact triple verification method to calculate the fact recall rate (FR) and fact precision rate (FP), and obtains the fact accuracy score by combining them: FA = 2×FR×FP / (FR+FP), where the fact recall rate (FR) is defined as the ratio of the number of successfully verified fact triples to the total number of fact triples that should be included in the generated results, and the fact precision rate (FP) is defined as the ratio of the number of successfully verified fact triples to the total number of fact triples extracted from the generated results.

[0021] Logical Consistency (LC): This dimension detects the logical consistency of the generated results in terms of causal relationships, temporal relationships, and quantitative relationships. It adopts a contradiction detection model based on Natural Language Inference (NLI) to identify logical conflicts within the text and between the text and the input information. The logical consistency score is calculated as: LC = 1 - Σwi×Ci / N, where Ci is the number of logical contradictions of the i-th type detected. The types of logical contradictions include at least three categories: causal contradictions, numerical contradictions, and temporal contradictions. wi is the severity weight corresponding to the i-th type of logical contradiction, and N is a normalization factor, which takes the value of the total number of logical relationship pairs detected.

[0022] Policy Compliance Dimension (RC): Determine whether the generated results comply with relevant financial regulatory requirements, construct a regulatory rule knowledge base, and use a combination of rule matching and semantic similarity calculation to detect potential compliance risk points and calculate the compliance score: RC = Σ(1-vi×si) / M, where vi is the severity coefficient of the violation of the i-th rule, ranging from 0 to 1, si is the confidence level of the violation detection, ranging from 0 to 1, and M is the total number of rules checked;

[0023] Risk Sensitivity (RS): This dimension assesses the sensitivity and accuracy of the generated results in responding to input risk signals. By designing test cases with gradient risk levels, it tests the model's ability to distinguish between different risk levels and calculates the risk sensitivity score: RS = Σrank_correlation(Rpred, Rtrue) / K, where Rpred is the risk level sequence predicted by the model, Rtrue is the manually labeled real risk level sequence, rank_correlation is the Spearman rank correlation coefficient, and K is the number of test case groups.

[0024] Temporal Validity (TV): Validates the validity of time-related information in the generated results, including policy version validity, data expiration date, event timestamps, etc. Establishes a time-sensitive information index and calculates the timeliness score through timestamp comparison and version verification: TV = Σvalid(ti) / T, where valid(ti) is the validity judgment function of the i-th time-sensitive information, which takes a value of 1 when the time information is still valid at the evaluation time point, otherwise takes a value of 0, and T is the total number of time-sensitive information;

[0025] Expression Standardization (ES): This dimension assesses the standardization of the generated results in terms of financial terminology usage, numerical format, and unit conversion. It employs a matching method based on a financial terminology dictionary and expression standardization templates to calculate the expression standardization score: ES = (Σterm_match(ti) + Σformat_match(fi)) / (Nt+Nf), where term_match(ti) is the matching degree of the i-th financial term, which is 1 when the term usage conforms to the financial terminology dictionary standard and 0 otherwise; format_match(fi) is the matching degree of the i-th numerical format, which is 1 when the numerical format conforms to the financial industry expression standard and 0 otherwise; Nt is the total number of financial terms involved in the generated results; and Nf is the total number of numerical formats involved.

[0026] The fusion decision layer performs weighted fusion of the scores of the above dimensions according to the preset business scenario weights. The comprehensive reliability score is calculated as follows: R_total = Σαi×Di, where Di is the score of the i-th dimension (FA, LC, RC, RS, TV, ES), αi is the weight coefficient of the dimension in a specific business scenario, and the sum of all weight coefficients is 1, which can be dynamically configured according to business needs.

[0027] Preferably, the financial knowledge graph fact verification engine module includes:

[0028] Financial Knowledge Graph Construction: This project integrates multi-source data, including regulatory policies, industry standards, market data, and enterprise information, to construct a financial knowledge graph comprising entities, relationships, and attributes. Specifically, it adopts a Neo4j graph database storage solution. Entity types include: institutional entities, product entities, person entities, policy entities, indicator entities, and event entities; relationship types include: hierarchical relationships, business relationships, guarantee relationships, association relationships, and temporal relationships; attribute types include: numerical attributes, time attributes, status attributes, and classification attributes. The knowledge graph construction employs a remotely supervised relationship extraction method combined with manual review and verification to ensure graph quality. The initial graph size includes no fewer than 500,000 entity nodes and 2 million relationship edges.

[0029] Fact Triple Extraction: Named Entity Recognition (NER) and Relation Extraction (RE) are performed on the results generated by the large model to convert unstructured text into a set of fact triples in the form of (subject, predicate, object). A joint extraction framework consisting of a BERT-CRF-based named entity recognition model and a Transformer encoder-based relation classification model is adopted, sharing the underlying encoder parameters to achieve end-to-end fact triple extraction. The entity recognition model uses BERT-base-Chinese as a pre-trained encoder and is fine-tuned on annotated corpora in the financial domain. The relation classification model adds a multi-head attention layer and a classification layer on top of the shared encoder, supporting the recognition of 96 types of financial relations.

[0030] Knowledge Graph Matching and Verification: Extracted fact triples are aligned and verified against the financial knowledge graph. Triples that directly match are marked as "verified"; for partially matched triples (e.g., entity matches but relation doesn't), missing information is filled in using graph reasoning based on TransE embedding before verification; for triples that cannot match, they are marked as "awaiting manual verification" or "suspected hallucination." Entity linking technology is used during the matching process to handle entity referencing disambiguation, with an accuracy rate of no less than 92%.

[0031] Incremental update mechanism: Establish a dynamic update process for the knowledge graph. By accessing external data sources such as regulatory document release systems, financial data terminals (such as Wind, Bloomberg, etc.), and enterprise information disclosure platforms (such as the National Enterprise Credit Information Publicity System, CNINFO, etc.), an incremental synchronization strategy based on change detection is adopted to achieve near real-time updates of the knowledge graph. The update frequency can be configured to real-time, daily, or weekly to ensure the timeliness of fact verification.

[0032] Preferably, the illusion detection technology module in the financial field adopts a multi-model collaborative verification strategy, including:

[0033] Self-Consistency Check: The same input is sampled N times (default N=5) to ensure output diversity by controlling the sampling temperature parameter T between 0.7 and 1.0. Cosine semantic similarity is calculated between each output, and hierarchical clustering is used to cluster the output results. The consistency score is calculated as: SC = 1 - Var(Semantic_Similarity(Oi, Oj)) / (n×(n-1) / 2), where Oi and Oj are the semantic vector representations of the i-th and j-th outputs (encoded using the Sentence-BERT model), n is the number of samples, and Var is the variance calculation function. A low consistency score indicates that the model's answer to the question is uncertain, posing a risk of hallucination. When the SC value is below a preset threshold of 0.6, the deep hallucination detection process is triggered.

[0034] Cross-Model Verification: This involves calling at least two heterogeneous large models (such as the GPT series, Wenxin Yiyan, and Tongyi Qianwen) to answer the same question, and identifying potential illusions through answer consistency analysis based on semantic implication relationships. Specifically, an NLI model is used to determine pairwise implication relationships between key statements output by each model. When the majority of models (≥2 / 3) outputs are consistent but individual models deviate, the deviation results are labeled as illusions, and the degree of deviation is recorded.

[0035] Attribution Verification: This requires large models to provide information sources or reasoning basis when generating results, and then automatically verifies the claimed sources. Specifically, it verifies the authenticity of the source through web retrieval and knowledge base queries, and then calculates the matching degree between the source content and the generated content using semantic similarity. Statements whose sources cannot be verified or whose semantic matching degree between the source and content is below the threshold of 0.7 are marked as suspected hallucinations.

[0036] Adversarial Probing: This involves designing targeted probing strategies, including three probing patterns: (a) detailed probing—requiring the large model to provide specific data sources and computational processes involved in the original answer; (b) reverse probing—posing questions with premises contradicting the original answer and observing whether the model easily changes its stance; and (c) boundary probing—making minor changes to the conditions of the original question and observing whether the answer produces a reasonable corresponding change. Unstable or self-contradictory answer patterns are considered hallucination indicators, specifically defined as: failure to provide reasonable evidence in detailed probing, changes in stance exceeding 50% of the total probing in reverse probing, or disproportionate changes in answer to minor changes in conditions in boundary probing.

[0037] The module integrates the detection results of the above units and outputs a comprehensive hallucination detection score H_score. The comprehensive hallucination detection score is: H_score = 1 - (w1×SC_anomaly + w2×CV_anomaly + w3×AV_anomaly + w4×AP_anomaly), where SC_anomaly, CV_anomaly, AV_anomaly, and AP_anomaly are the anomaly scores of the four detection methods (each ranging from 0 to 1, where 0 indicates no anomaly and 1 indicates a severe anomaly), and wi is the corresponding weight. The default weight configuration is w1=0.25, w2=0.30, w3=0.25, w4=0.20, which can be adjusted according to the needs of the financial scenario. The weights of each detection unit are adaptively updated based on their accuracy in historical evaluations, with an update cycle of weight optimization after every 1000 evaluation tasks.

[0038] Preferably, the dynamic calibration and uncertainty quantification module is configured as follows:

[0039] Token-level uncertainty estimation: Based on the token probability distribution output by the large model, the information entropy H(ti) = -Σp(ti|context)×log2(p(ti|context)) of each token, the average entropy value H_avg =ΣH(ti) / L (L is the sequence length) and the entropy variance Var(H) of the entire sequence are calculated as measures of the inherent uncertainty of the model. When H_avg exceeds a preset threshold (default value is 2.5 bits), the overall uncertainty of the generated result is considered to be high.

[0040] Semantic-level confidence calibration: A dedicated confidence estimation network is trained, employing a 3-layer fully connected neural network structure with hidden layer dimensions of 256 and 128, and the activation function is ReLU. It takes the hidden layer representation of the last layer of the large model (with dimensions of 768 or 1024, depending on the base model) as input to predict the correct probability of the generated result across various reliability dimensions. Temperature scaling calibration eliminates the model's overconfidence or overconservatism biases. Specifically, the optimal temperature parameter T* is learned on the validation set to minimize the expected calibration error (ECE) between the calibrated probability distribution and the actual accuracy, with a target ECE value below 0.05.

[0041] Bayesian uncertainty decomposition: The total uncertainty is decomposed into epistemic uncertainty (due to insufficient model knowledge) and aleatoric uncertainty (due to inherent noise in the input data). The Monte Carlo Dropout method (with a Dropout rate of 0.1 and 20 forward propagations) is used for estimation. Epistemic uncertainty is calculated from the variance of multiple forward propagation results, and aleatoric uncertainty is calculated from the average variance of the model's predicted distribution, providing support for subsequent uncertainty attribution analysis.

[0042] Scenario-adaptive optimization: For different credit risk control business scenarios (such as pre-loan investigation, loan approval, and post-loan management), scenario-specific calibration models are trained on labeled datasets specific to each scenario. Each scenario has at least 5000 labeled samples for training data, improving the business relevance and accuracy of confidence level estimation. For the loan approval scenario, a stricter confidence threshold (0.85) is used for content involving credit limit and risk level judgments, while a relatively lenient confidence threshold (0.70) is used for monitoring content in the post-loan management scenario.

[0043] Output uncertainty quantification results: U = {U_total, U_epistemic, U_aleatoric,Confidence_calibrated, Review_priority}, where U_total is the total uncertainty (ranging from 0 to 1), U_epistemic is the cognitive uncertainty, U_aleatoric is the random uncertainty, Confidence_calibrated is the calibrated confidence level, and Review_priority is the manual review priority calculated based on the uncertainty level and business importance (divided into high, medium, and low levels). The rule for determining the review priority is: when U_epistemic / U_total > 0.6 and the business importance is high, Review_priority is set to high priority.

[0044] Preferably, the robustness evaluation module is configured to perform at least one of the following tests:

[0045] Semantic Preservation Perturbation: Adversarial examples are generated through synonym replacement, sentence structure transformation, and information reordering while preserving the semantics of the input, testing the stability of the model's output. A perturbation word generation method based on the BERT masking language model is used. Words with semantic similarity higher than 0.85 are selected from the candidate replacement words for replacement, and the perturbation rate of each input is controlled to not exceed 15% (i.e., the proportion of perturbed tokens to the total number of tokens), ensuring that the perturbation intensity is within the semantic preservation threshold. At least 10 perturbation variants are generated for each test input, and the consistency of the model's output on the original input and the perturbed input is calculated.

[0046] Boundary condition testing: Construct test cases near the decision boundary, such as customer cases that just meet / do not meet credit conditions (financial indicators deviate from the threshold by ±5%), and policy-mandated threshold cases, to test the accuracy and consistency of the model's boundary judgments. Specifically, a boundary test case library is built based on historical credit approval data, with 50 positive and 50 negative test cases for each decision boundary, and the model's judgment reversal rate and consistency score near the boundary are calculated.

[0047] Induced Attack Testing: Design adversarial inputs that attempt to induce the model to generate incorrect information or violate regulations, including three attack patterns: (a) privilege escalation attacks that fake an emergency request to skip the review process; (b) information extraction attacks that use obfuscated statements to attempt to obtain sensitive information; and (c) logic-induced attacks that guide the model to draw incorrect conclusions by embedding false premises. Design no fewer than 100 test cases for each type of attack and calculate the model's attack defense success rate.

[0048] Data contamination simulation: A certain proportion (5%, 10%, 20%, and 30% at four gradients) of erroneous or contradictory information is injected into the input to test the model's fault tolerance and anomaly detection capabilities against noisy data. Specifically, the erroneous information includes three categories: numerical tampering (e.g., modifying financial data), relational tampering (e.g., modifying equity structure), and time tampering (e.g., modifying date information). The accuracy decay curves of the model's output are calculated under different contamination proportions.

[0049] Robustness score: R_robust = (Σstability(pi) + Σboundary_accuracy(bi) + Σsecurity(si) + Σnoise_tolerance(ni)) / (Np+Nb+Ns+Nn), where stability(pi) is the stability pass determination of the i-th perturbation test case (a value of 1 is determined when the output consistency is higher than 0.90, otherwise a value of 0 is taken), boundary_accuracy(bi) is the accuracy judgment result of the i-th boundary test case, security(si) is the defense success result of the i-th induced attack test case, noise_tolerance(ni) is the fault tolerance pass result of the i-th data pollution test case, and Np, Nb, Ns, and Nn are the total number of test cases in the four types of tests.

[0050] Furthermore, an end-to-end automated assessment pipeline will be constructed, specifically including:

[0051] Test Case Management: Establish a hierarchical and categorized test case library, covering types such as basic capability testing, business scenario testing, extreme case testing, and regression testing. Support version management, automatic expansion, and difficulty rating of test cases. The initial size of the test case library should be no less than 10,000 entries, covering major business scenarios such as credit approval, risk warning, and contract review.

[0052] Batch test execution: Enables parallel scheduling and execution of test tasks, supports efficient processing of large-scale test cases, adopts a message queue-based distributed task scheduling architecture, supports horizontal scaling to meet test needs of different scales, and the single-node test throughput is no less than 500 test cases per hour.

[0053] Results aggregation and analysis: Automatically summarizes the evaluation results of each dimension and generates multi-granularity scoring reports (overall score, dimension score, scenario score, use case score), supporting comparative analysis with historical versions and tracking of change trends;

[0054] Visualized report generation: Automatically generates visualized assessment reports containing radar charts, trend charts, distribution maps, case details, etc., making it easier for business and technical personnel to understand the assessment results;

[0055] Continuous Integration Integration: Provides standardized API interfaces to support continuous integration (CI / CD) with model development pipelines, enabling automated regression testing after model updates.

[0056] A method for evaluating the reliability of the generation results of a large-scale bank credit risk control model, using the evaluation system described in any of the above technical solutions, includes the following steps:

[0057] Step S1: Receive the evaluation request through the application interface layer, obtain the large model generation result to be evaluated and its corresponding original input data, and preprocess the generation result and the original input data, including text segmentation, entity annotation and format standardization;

[0058] Step S2: Through the evaluation task scheduler of the evaluation engine layer, according to the pre-configured evaluation strategy, the modules coordinate to perform multi-dimensional reliability evaluation on the generated results in parallel or sequentially, and the modules exchange intermediate results through a collaborative feedback mechanism.

[0059] Step S3: In the fusion decision layer, the evaluation scores and uncertainty quantification results of each module are received, and weighted fusion is performed according to the weight configuration corresponding to the business scenario to calculate the comprehensive reliability score. The priority of manual review is determined based on the uncertainty level and business importance.

[0060] Step S4: Generate a visual evaluation report containing a comprehensive score, scores for each dimension, anomaly annotations, uncertainty decomposition, and improvement suggestions through the application interface layer, and return it to the requesting party.

[0061] Preferably, in step S2, the multi-dimensional reliability assessment includes at least a factual accuracy assessment, the specific steps of which are as follows:

[0062] The financial knowledge graph fact verification engine uses a joint extraction model based on the Transformer architecture to extract fact triples in the form of subject-predicate-object from the generated results.

[0063] The extracted triples are aligned with the financial knowledge graph for entity alignment and relationship verification, and the matching status and confidence level are output for each triple.

[0064] Based on the verification results, the fact accuracy score is calculated as FA = 2 × FR × FP / (FR + FP), where FR is the fact recall rate, defined as the ratio of the number of successfully verified fact triples to the total number of fact triples that should be included; and FP is the fact precision rate, defined as the ratio of the number of successfully verified fact triples to the total number of extracted triples.

[0065] Preferably, the method is applied to a credit approval opinion generation scenario, and step S2 further includes:

[0066] The policy compliance assessment module is invoked to match the credit limit suggestion with the concentration restriction rules in the regulatory rule base to verify whether the credit limit meets the concentration restriction requirements, and to match the guarantee plan with the collateral ratio regulations to check whether the guarantee plan meets the collateral ratio regulations.

[0067] The logical consistency assessment module is invoked, and a natural language reasoning model is used to judge the implied relationship between the risk point identification conclusion and the credit recommendation, and to detect whether there is a logical contradiction in identifying high-risk points but giving high credit recommendations;

[0068] The illusion detection technology module in the financial field is invoked to generate approval opinions by sampling at least 5 times and calculating the semantic consistency between each version. At the same time, at least 2 heterogeneous large models are invoked for cross-validation to identify false or unreliable content.

[0069] Compared with the prior art, the beneficial effects of the present invention are:

[0070] 1. Systematic and comprehensive: This invention proposes for the first time a multi-dimensional reliability evaluation framework (MREF) covering six core dimensions: factual accuracy, logical consistency, policy compliance, risk sensitivity, timeliness, and expression standardization. It achieves a comprehensive evaluation of the results generated by large models and overcomes the limitations of existing technologies.

[0071] 2. Professionalism and Targeted Approach: This invention addresses the specific needs of commercial bank credit risk control by constructing a financial knowledge graph fact-checking engine and a domain-specific illusion detection technology. Compared to the general SelfCheckGPT method, this invention, by integrating multiple strategies such as financial knowledge graph verification, multi-model cross-validation, and adversarial detection, improves the F1 score for illusion detection in financial loan approval scenarios from 0.62 to 0.81, an improvement of approximately 30%.

[0072] 3. Scientific nature and quantifiability: This invention designs clear quantitative calculation formulas and scoring standards for each evaluation dimension. The variables are clearly defined, and the evaluation results have good comparability and reproducibility, providing a scientific basis for model selection, optimization iteration, and regulatory review.

[0073] 4. Practicality and Operability: This invention constructs an automated assessment pipeline and standardized API interface, supporting batch assessment, continuous integration, and visualization report generation. In actual testing, the complete assessment of a single credit approval opinion takes approximately 45 seconds (including assessment of all six dimensions and report generation), which is about 40 times more efficient than the approximately 30 minutes required for manual evaluation.

[0074] 5. Collaboration and Intelligence: This invention innovatively designs a collaborative feedback mechanism between various assessment modules. The hallucination detection results guide the focus of fact verification, and the uncertainty quantification results affect the weight of the fusion decision, realizing intelligent collaboration among the modules. Compared with the scheme where each module operates independently, the comprehensive assessment accuracy in the collaborative mode is improved by about 8.5%.

[0075] 6. Security and Reliability: This invention, through its robustness assessment module, covers four test scenarios: semantic preservation perturbation, boundary conditions, induced attacks, and data pollution. It can effectively identify the vulnerability of the model and provide a guarantee for the safe and stable operation of the financial risk control system.

[0076] 7. Dynamism and Adaptability: This invention supports dynamic updates of the knowledge graph and scenario-adaptive confidence calibration, enabling timely responses to changes in regulatory policies and adjustments in business needs, and maintaining the timeliness and relevance of the evaluation system.

[0077] 8. Standardization and scalability: The evaluation framework and scoring system proposed in this invention have good universality and can be used as a reference standard for the reliability evaluation of large models in the industry. They can also be extended to other business areas of the banking industry and other financial institutions. Attached Figure Description

[0078] Figure 1 This is the overall architecture diagram of the reliability evaluation system for the generation results of large models in the field of commercial bank credit risk control.

[0079] Figure 2 This is a structural diagram of the Multidimensional Reliability Evaluation Framework (MREF) of this invention;

[0080] Figure 3 This is a flowchart of the Financial Knowledge Graph Fact Verification Engine (FKGV) of this invention;

[0081] Figure 4 This is a schematic diagram illustrating the multi-model collaborative verification of the Financial Illusion Detection Technology (FHDT) of this invention;

[0082] Figure 5 This is a schematic diagram of the Dynamic Calibration and Uncertainty Quantization (DCUQ) module technology of the present invention;

[0083] Figure 6 This is a schematic diagram illustrating the generation of test cases for the Anti-Robustness Assessment Module (ART) of this invention;

[0084] Figure 7 This is a diagram showing the deployment architecture of the automated evaluation pipeline of this invention;

[0085] Figure 8 This is an example diagram of the visualization interface for the reliability evaluation report of this invention. Detailed Implementation

[0086] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0087] like Figure 1-8 As shown, this invention provides a reliability evaluation technology for the generation results of large models in the field of commercial bank credit risk control, providing systematic evaluation technology support for the safe and reliable application of large models in commercial bank credit risk control scenarios.

[0088] Example 1: Reliability Assessment of Credit Approval Opinion Generation Scenarios

[0089] This embodiment uses the automatic generation of commercial bank credit approval opinions as an application scenario to explain in detail the implementation process of the technical solution of the present invention:

[0090] Step 1: Input Data Preparation: Collect sample credit approval opinions generated by the large model to be evaluated, including: basic loan application information, customer qualification analysis, risk point identification, credit limit suggestions, approval conclusions, etc. Simultaneously, prepare the corresponding original input data (customer application materials, credit reports, financial statements, etc.) as a reference benchmark. In this embodiment, the evaluation sample is the approval opinion of a commercial bank for a working capital loan to a manufacturing enterprise, with a loan amount of RMB 50 million and a term of one year.

[0091] Step 2, Fact Accuracy Assessment: The Financial Knowledge Graph Fact Verification Engine (FKGV) is used to verify the factual content in the approval comments.

[0092] (1) Extract the basic enterprise information triples, such as customer name, registered capital, establishment date, and shareholder information, and match and verify them with the enterprise information knowledge graph. Specifically, the BERT-CRF named entity recognition model identified 23 entities from the approval opinions, and the joint extraction model extracted 37 fact triples. After matching with the knowledge graph, 32 of the triples were successfully verified (the matching confidence was higher than 0.90), 3 triples were partially matched (2 were successfully verified after graph reasoning and completion, and 1 required manual verification), and 2 triples could not be matched (marked as suspected hallucinations). The fact precision rate FP = 34 / 37 = 0.919, the fact recall rate FR = 34 / 36 = 0.944 (the total number of baseline triples is 36), and the fact accuracy dimension score FA = 2×0.919×0.944 / (0.919+0.944) = 0.931.

[0093] (2) Extract financial indicators (debt-to-equity ratio, current ratio, net profit, etc.) and verify them against the original financial statement data. In this embodiment, the approval opinion involves 12 financial indicators. After verification, 11 of the data are completely correct, and 1 (current ratio) deviates from the original data by more than 0.5%, which is marked as requiring attention.

[0094] (3) The cited regulatory policy clauses and regulatory rules knowledge base were extracted for version verification and content consistency checks. The approval opinion cited three regulatory provisions, all of which were verified to be current and valid versions with accurate content citations.

[0095] Step 3, Logical Consistency Assessment: Apply a RoBERTa-based fine-tuned natural language reasoning model to detect the internal logical relationships of the approval comments.

[0096] (1) Check the logical consistency between risk point identification and credit recommendation. In this embodiment, the model identified two main risk points (industry downturn risk and high accounts receivable concentration), and the credit limit recommended in the approval opinion was RMB 50 million. The NLI model determined that there was no obvious logical contradiction between risk point identification and credit recommendation (implied relationship confidence level 0.82).

[0097] (2) Verify the consistency between the quantitative analysis conclusions and the qualitative description. One potential inconsistency was detected: the quantitative data showed that the customer's revenue increased by 3.2% year-on-year (slight increase), but the qualitative description used the phrase "good growth trend". The NLI model marked this as a slight inconsistency (confidence level of 0.45, which is lower than the serious inconsistency threshold of 0.70 but higher than the warning threshold of 0.30).

[0098] (3) The rationality of the timeline description was checked. The approval opinion involved 5 time nodes, all of which passed the time sequence consistency check. The logical consistency dimension score LC = 1 - (0.3×1 + 0.0×0 + 0.0×0) / 15 = 0.980, of which 1 minor causal contradiction was detected (weight 0.3), and there were no numerical contradictions or time sequence contradictions. The normalization factor N=15 (the total number of logical relationship pairs detected).

[0099] Step 4, Policy Compliance Assessment: Utilize the regulatory rule matching engine to check the compliance of the approval opinions.

[0100] (1) Verify whether the proposed credit line complies with the relevant concentration limits. After matching and calculation, the customer credit line accounts for 0.8% of the bank's net capital, which does not exceed the single customer concentration limit (10%), and is compliant.

[0101] (2) Check whether the guarantee plan meets the requirements of the mortgage-pledge ratio. The guarantee plan recommended in the approval opinion is a real estate mortgage with a mortgage ratio of 65%, which meets the internal requirement of not exceeding 70% and is compliant.

[0102] (3) Verify whether the risk classification judgment meets the five-level classification standard. The approval opinion classifies the loan as "normal". After rule matching verification, it is consistent with the assessment results of the customer's current financial situation and repayment ability, and is compliant.

[0103] (4) Confirm whether the information disclosure is complete and compliant. Check the 12 information elements that need to be disclosed. The approval opinion covers 11 of them and omits 1 (environmental compliance statement), which is marked as a compliance defect. Policy compliance dimension score RC = (1-0×0+ 1-0×0 + 1-0×0 + 1-0.5×0.90) / (4) = (1+1+1+0.55) / 4 = 0.888.

[0104] Step 5, Illusion Detection: Applying Financial Illusion Detection Technology (FHDT) to identify false or unreliable content in approval opinions:

[0105] (1) The approval opinions were generated five times (sampling temperature T=0.8), and the semantic consistency between the versions was calculated. The average pairwise semantic similarity of the five outputs was 0.87, and the consistency score SC=0.91, which is higher than the threshold of 0.6, indicating good overall self-consistency. However, the consistency score of the "Industry Prospect Analysis" paragraph was only 0.58, and this paragraph was marked as a high-risk area for illusion.

[0106] (2) Use GPT-4 and Wenxin Yiyan to generate approval opinions for the same input and perform cross-validation. The three models are highly consistent in terms of customer fundamentals analysis and financial data (CV=0.93), but there are differences in terms of industry prospects judgment (CV=0.62), which is marked as needing key review.

[0107] (3) Verify the source of data claimed in the opinion. The approval opinion cited three external data sources, including the "China Manufacturing PMI Index". After automated verification, the source and value of two data sources were found to be accurate, while the value of one data source was found to be different from the actual value (claimed as 48.5, actual as 49.2). The anomaly score for source verification was AV_anomaly=0.15.

[0108] (4) Design follow-up questions to verify the consistency of details. Ten follow-up questions were asked on the industry prospect analysis section. The model's response stability score was 0.73, with two instances of inconsistency in stance. The adversarial detection anomaly score was AP_anomaly=0.27.

[0109] The comprehensive hallucination detection score H_score = 1 - (0.25×0.09 + 0.30×0.15 + 0.25×0.15 +0.20×0.27) = 1 - (0.0225 + 0.045 + 0.0375 + 0.054) = 1 - 0.159 = 0.841.

[0110] Step 6, Uncertainty Quantification: Apply the Dynamic Calibration and Uncertainty Quantification (DCUQ) module to assess the confidence level of the approval opinion.

[0111] (1) Calculate the entropy distribution of tokens related to key conclusions (such as credit recommendations and risk assessments). The average entropy of tokens related to credit recommendations is H_avg=1.8 bits (below the threshold of 2.5, with acceptable uncertainty), while the average entropy of tokens related to industry prospect analysis is H_avg=3.2 bits (above the threshold, with higher uncertainty).

[0112] (2) The correctness probability of each conclusion was predicted by a calibrated confidence estimation network (3-layer fully connected network, hidden layer dimension 256-128, input dimension 768). The calibrated confidence level of the credit limit suggestion was 0.88, the calibrated confidence level of the risk point identification was 0.85, and the calibrated confidence level of the industry prospect analysis was 0.61.

[0113] (3) Decomposing cognitive uncertainty and stochastic uncertainty. Monte Carlo Dropout (Dropout rate 0.1, forward propagation 20 times) was used to calculate: the cognitive uncertainty U_epistemic = 0.08 and the stochastic uncertainty U_aleatoric = 0.04 for the credit limit suggestion part; the cognitive uncertainty U_epistemic = 0.28 and the stochastic uncertainty U_aleatoric = 0.11 for the industry prospect analysis part. The high cognitive uncertainty in the industry prospect analysis part indicates that the model has insufficient knowledge in this field.

[0114] (4) Generate manual review priority suggestions. Industry prospect analysis section: U_epistemic / U_total = 0.28 / 0.39 = 0.72 > 0.6, and involves core judgments in credit approval (business importance is high), Review_priority is set to high priority. Credit limit suggestion section: After calibration, confidence level is 0.88 > 0.85 (threshold for loan approval scenario), Review_priority is set to low priority.

[0115] Step 7, Comprehensive Scoring and Report Generation: Summarize the evaluation results of each dimension. In the scenario of generating credit approval opinions, the business weights are configured as follows: α_FA=0.25, α_LC=0.20, α_RC=0.25, α_RS=0.10, α_TV=0.10, α_ES=0.10.

[0116] The overall reliability score R_total = 0.25×0.931 + 0.20×0.980 + 0.25×0.888 +0.10×0.95 + 0.10×1.0 + 0.10×0.92 = 0.233 + 0.196 + 0.222 + 0.095 + 0.100 +0.092 = 0.938.

[0117] (Note: RS=0.95, TV=1.0, and ES=0.92 are the evaluation scores for these three dimensions in this embodiment.)

[0118] Generate a visual assessment report, including a radar chart showing scores across six dimensions, detailed breakdowns, anomaly markers (the industry outlook analysis section is marked as "illusion, high risk, and high uncertainty"), and improvement suggestions (it is recommended that the industry outlook analysis content be reviewed manually and that an environmental compliance statement be added).

[0119] Example 2: Reliability Evaluation of Risk Warning Information Interpretation Scenarios

[0120] This embodiment uses the scenario of automatically interpreting risk warning information (such as corporate public opinion, industry risks, regulatory penalties, etc.) using a large model:

[0121] Step 1: Prepare the raw data for risk warning (news reports, regulatory announcements, industry reports, etc.) and the warning interpretation content generated by the large model. This example selects a debt default warning event of a real estate company. The raw data includes 3 news reports, 1 regulatory announcement, and 1 industry analysis report.

[0122] Step 2, Focus on Timeliness Dimension (TV): Verify the validity of time-sensitive information such as policy versions, data cutoff times, and event occurrence times cited in the interpretation content to ensure that the warning interpretation is based on the latest information. In this embodiment, the warning interpretation involves 8 time-sensitive information points, of which 7 were verified to be valid (the cited policy versions and event dates are both the latest), and 1 data cutoff time was not clearly marked. The timeliness dimension score TV = 7 / 8 = 0.875.

[0123] Step 3: Focus on evaluating the risk sensitivity dimension (RS): Design a gradient risk level test. Construct 5 sets of gradient risk level test cases (from low risk to high risk) to verify the model's ability to distinguish between risk events of different severity and its response accuracy. The Spearman rank correlation coefficient between the model's predicted risk level sequence and the actual risk level sequence is 0.89, and the risk sensitivity dimension score RS = 0.89.

[0124] Step 4: Conduct adversarial robustness assessment:

[0125] (1) Test the model's ability to identify false early warning information: 3 false early warning messages were injected, and the model correctly identified 2 of them, with a defense success rate of 66.7%;

[0126] (2) Testing the stability of the model's response to extreme scenarios: Five extreme scenario variants were constructed, and the model output consistency score was 0.82;

[0127] (3) Test the model’s ability to correctly understand misleading statements: Eight misleading statement test cases were constructed, and the model correctly understood six of them, with an accuracy rate of 75%.

[0128] Step 5: Generate an evaluation report, highlighting timeliness issues (one data cutoff time is unclear), risk misjudgment (the ability to identify false early warning information needs improvement), and potential vulnerabilities.

[0129] Example 3: Reliability Assessment of Loan Contract Terms Review Scenarios

[0130] This embodiment focuses on the application scenario of using large models to assist in reviewing loan contract terms:

[0131] Step 1: Prepare the loan contract text to be reviewed and the review comments generated by the large model (including clause interpretation, risk warnings, modification suggestions, etc.). This example selects a corporate working capital loan contract, which contains 32 clauses, and the large model generates review comments for each clause.

[0132] Step 2, Focus on Policy Compliance Dimension (RC): Verify the accuracy of the review comments' judgments on the compliance of the clauses, whether any important compliance risk points have been omitted, and whether the cited legal basis is correct. The review comments on 32 clauses were examined, correctly identifying 7 out of 8 compliance risk points (recall rate 87.5%), generating 2 false positives (precision rate 77.8%), and omitting 1 compliance issue related to the calculation method of prepayment penalties. The policy compliance dimension score RC = 0.847.

[0133] Step 3: Focus on assessing the standardization of expression (ES): Examine the accuracy of legal terminology used in the review comments, the completeness of clause citations, and the standardization of risk level statements. The review comments involved 45 legal terms, of which 42 were used correctly (term_match score 42 / 45 = 0.933), and 18 numerical formats, of which 16 were correctly formatted (format_match score 16 / 18 = 0.889). The standardization score for the expression dimension ES = (42+16) / (45+18) = 58 / 63 = 0.921.

[0134] Step 4: Conduct boundary condition testing: Use the designed boundary test cases (such as contracts containing complex nested clauses, non-standard formats, or ambiguous expressions) to test the boundaries of the model's review capabilities. Construct 20 boundary test cases; the model correctly judged 15 cases, achieving a boundary condition accuracy of 75%. However, its ability to handle ambiguous expressions was relatively weak (only 2 out of 5 ambiguous clauses were correctly handled).

[0135] Step 5: Generate an evaluation report, outputting key indicators such as the accuracy rate of compliance judgment (87.5%), omission rate (12.5%), and false alarm rate (22.2%), and provide improvement suggestions: It is recommended to strengthen the ability to detect the compliance of early repayment-related clauses and improve the ability to understand and handle ambiguous clauses.

[0136] Example 4: Deployment and Operation of Automated Evaluation Pipeline

[0137] This example illustrates the deployment architecture and operation mechanism of an automated evaluation pipeline:

[0138] Step 1, System Deployment: Deploy the evaluation system in the bank's data center or private cloud environment, including: (1) Financial Knowledge Graph Service Cluster (based on Neo4j graph database, 3-node cluster deployment, storage capacity not less than 500GB); (2) Illusion Detection Service Cluster (not less than 2 GPU servers, used to run Sentence-BERT semantic encoding and NLI inference); (3) Uncertainty Quantification Service Cluster (not less than 1 GPU server, used for Monte Carlo Dropout calculation); (4) Adversarial Evaluation Service Cluster (not less than 4 CPU servers, used for adversarial sample generation and batch test execution); (5) Evaluation Engine Master Node (based on message queue distributed task scheduling, using RabbitMQ or Kafka as message middleware); (6) Report Generation and Visualization Service (front-end display service based on Web framework). Each component communicates through an internal API gateway, using the gRPC protocol to ensure communication efficiency.

[0139] Step 2, Knowledge Base Initialization: Import basic knowledge data such as the regulatory policy database (covering rules and regulations issued by regulatory agencies such as the State Financial Supervision and Administration Bureau and the People's Bank of China, with an initial inclusion of no less than 2,000 legal provisions), the industry standard database (covering national and industry standards related to credit business, with an initial inclusion of no less than 500 standard entries), the financial terminology database (containing no less than 15,000 financial terms and their standard expressions), and the historical case database (containing no less than 3,000 labeled historical credit approval cases), and construct an initial financial knowledge graph (no less than 500,000 entity nodes and 2 million relationship edges).

[0140] Step 3, Test Case Configuration: Configure a set of test cases according to the business scenario to be evaluated, including: (1) Standard test set (covering typical business scenarios, with no less than 5,000 test cases); (2) Boundary test set (covering extreme and abnormal situations, with no less than 2,000 test cases); (3) Adversarial test set (covering security attack scenarios, with no less than 1,500 test cases); (4) Regression test set (used for version iteration comparison, with no less than 1,500 test cases).

[0141] Step 4, Pipeline Execution: The evaluation tasks are automatically executed according to the configured evaluation plan through the scheduling engine: (1) Obtain input data from the test case library; (2) Call the API of the large model to be tested to obtain the generated results; (3) Call the evaluation services of each dimension in parallel for evaluation (distribute tasks through message queues, and each service consumes and processes them independently); (4) Summarize the evaluation results to the scoring engine; (5) Generate the evaluation report and archive it. The complete evaluation of a single test case takes about 45 seconds, and the batch evaluation of 10,000 test cases can be completed within 8 hours (using 4-node parallel processing).

[0142] Step 5, Continuous Integration: Monitor large model version updates via a Webhook mechanism, automatically trigger regression testing tasks, compare the reliability score changes of the new version with the baseline version, and output a version iteration analysis report. An alert notification is automatically triggered when the overall reliability score of the new version is more than 5% lower than the baseline version.

[0143] Step 6, Dynamic Update and Maintenance: Establish a regular knowledge graph update mechanism (daily synchronization of regulatory documents, weekly updates of market data, and monthly full verification) to ensure the timeliness of the evaluation benchmark; continuously collect evaluation feedback to optimize the scoring model parameters (update the weights once every 1000 evaluations); regularly expand and update the test case library (add no less than 500 test cases per quarter).

[0144] Example 5: Experimental Verification and Performance Evaluation

[0145] To verify the effectiveness and advancement of the technical solution of this invention, an experimental environment was constructed to conduct a systematic performance evaluation.

[0146] I. Experimental Dataset

[0147] Construct a dedicated benchmark dataset for financial credit assessment, including:

[0148] (1) Credit Approval Opinion Evaluation Set: This set contains 500 credit approval opinions generated by a large model, covering 10 industry types including manufacturing, real estate, wholesale and retail, and information technology. Each approval opinion is manually annotated by 3 senior credit approval personnel. The annotation content includes scores for dimensions such as factual accuracy, logical consistency, and policy compliance, as well as anomaly annotations. The inter-rater consistency (Cohen's Kappa) of the manually annotated opinions is 0.82, indicating that the annotation quality is reliable.

[0149] (2) Hallucination detection evaluation set: 127 hallucination contents (including 68 factual hallucinations, 35 inference hallucinations, and 24 faithful hallucinations) were manually marked from the above 500 approval comments and used as the evaluation benchmark for the hallucination detection module.

[0150] (3) Adversarial test set: 600 adversarial test cases were constructed, including 200 semantic preservation perturbations, 150 boundary condition tests, 150 inducement attack tests, and 100 data pollution simulations.

[0151] II. Performance Evaluation of Each Module

[0152] (1) Fact accuracy assessment module: On the credit approval opinion assessment set, the precision of the fact triplet extraction was 89.3%, the recall was 85.7%, and the F1 score was 87.5%. The Pearson correlation coefficient with the manually labeled fact accuracy score was 0.91, indicating that the automatic assessment results were highly consistent with the manual assessment.

[0153] (2) Hallucination detection module: The performance on the hallucination detection test set is shown in the table below:

[0154]

[0155] Experimental results show that the multi-strategy fusion hallucination detection method of the present invention achieves an F1 score of 0.81 in the financial credit scenario, which is about 30.6% higher than the SelfCheckGPT method and about 24.6% higher than the general NLI detection method.

[0156] (3) Uncertainty Quantification Module: In terms of calibration performance, the expected calibration error (ECE) of the uncalibrated model was 0.142, which decreased to 0.038 after temperature scaling calibration, meeting the target ECE < 0.05 requirement. The Pearson correlation coefficient between the post-calibration confidence level and the actual accuracy was 0.89. The Spearman rank correlation coefficient between the priority ranking of manual reviewers and the ranking by senior approvers was 0.84, indicating that the uncertainty quantification results provide good guidance for the priority ranking of manual reviewers.

[0157] (4) Adversarial robustness evaluation module: The test results on the adversarial test set are shown in the table below:

[0158]

[0159] III. Evaluation of Module Collaboration Effect

[0160] To verify the effectiveness of the collaborative feedback mechanism between modules, an ablation experiment was conducted:

[0161]

[0162] Experimental results show that the collaborative feedback mechanism among the modules improved the overall assessment accuracy from 84.2% to 91.3%, an increase of approximately 8.4%, validating the effectiveness of the collaborative mechanism. Among them, the collaborative effect of hallucination detection guiding fact-checking was the most significant (contributing approximately 2.9 percentage points), because hallucination detection can effectively narrow the focus of fact-checking, improving the targeting and efficiency of the verification.

[0163] IV. System Performance Evaluation

[0164] Under the above experimental environment (cluster configuration of 4 GPU servers + 8 CPU servers), the system performance indicators are shown in the table below:

[0165]

[0166] Compared to manual evaluation (which takes about 30 minutes for a senior reviewer to evaluate a single approval comment), the automated evaluation efficiency of this invention is about 40 times higher.

[0167] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A reliability evaluation system for the generation results of a large-scale bank credit risk control model, characterized in that, include: The data layer is used to store the original input data, generated result data, and external financial data used as a verification benchmark for the large model. The data layer adopts a unified data model to manage the above-mentioned data in a structured manner and maintains the timeliness and traceability of the data through a data version control mechanism. The knowledge layer is used to store and manage financial knowledge graphs, regulatory rule bases, financial terminology dictionaries, and historical case databases. The financial knowledge graph stores financial entities, their relationships, and attributes in the form of triples, and the regulatory rule base stores financial regulations and their applicable conditions in the form of structured rule entries. The evaluation engine layer is used to perform multi-dimensional reliability evaluation. It includes a multi-dimensional reliability evaluation framework module, a financial knowledge graph fact verification engine module, a financial domain illusion detection technology module, a dynamic calibration and uncertainty quantification module, and an adversarial robustness evaluation module. The evaluation engine layer coordinates the parallel execution of each module through a unified evaluation task scheduler, and the modules pass intermediate results through a message queue. The detection results of the financial domain illusion detection technology module are used to assist the financial knowledge graph fact verification engine module in determining the key verification scope, and the quantification results of the dynamic calibration and uncertainty quantification module serve as the input basis for the fusion decision layer to calculate the priority of manual review. The fusion decision layer is used to receive and fuse the evaluation results from the evaluation engine layer, weight and fuse the scores of each dimension based on the preset business scenario weights to calculate the comprehensive reliability score, and generate an evaluation report containing review priority suggestions by combining the uncertainty quantification results. The application interface layer provides standardized RESTful API interfaces to receive assessment requests, return assessment results, and provide a visual report interface.

2. The reliability evaluation system for the generation results of a large-scale bank credit risk control model according to claim 1, characterized in that, The multi-dimensional reliability evaluation framework module is configured to evaluate the large model generation results in at least six dimensions, including: The factual accuracy dimension is used to assess the correctness of factual content such as financial data, customer information, and policies and regulations in the generated results; The logical consistency dimension is used to detect logical inconsistencies within the generated results and between the generated results and the input information. The policy compliance dimension is used to determine whether the generated results comply with relevant financial regulatory requirements. The risk sensitivity dimension is used to assess the sensitivity and accuracy of the generated results in response to input risk signals. The timeliness dimension is used to verify the validity of time-related information in the generated results; The standardization dimension is used to assess the degree of standardization in the use of financial terminology and numerical format of the generated results. The fusion decision layer performs weighted fusion of the scores of the above-mentioned dimensions according to the preset business scenario weights to obtain a comprehensive reliability score. The comprehensive reliability score is calculated as follows: the comprehensive reliability score is equal to the weighted sum of the scores of each dimension and the corresponding weight coefficients. The scores of each dimension correspond to the quantitative scores of the above six dimensions. Each weight coefficient is dynamically configured according to the needs of a specific business scenario, and the sum of all weight coefficients is 1.

3. The reliability evaluation system for the generation results of a large-scale bank credit risk control model according to claim 1, characterized in that, The financial knowledge graph fact-verification engine module includes: The knowledge graph construction unit is used to integrate multi-source financial data to construct a financial knowledge graph containing entities, relationships, and attributes. The entity types include at least institutional entities, product entities, person entities, policy entities, and indicator entities, and the relationship types include at least subordinate relationships, business relationships, guarantee relationships, and association relationships. The knowledge graph construction unit uses a relationship extraction method based on remote supervision to automatically identify entities and relationships from unstructured financial text. The fact triple extraction unit is used to perform named entity recognition and relation extraction on the generated results of the large model, and convert them into a set of fact triples in the form of subject-predicate-object. The fact triple extraction unit adopts a joint extraction model based on the Transformer architecture, and completes entity recognition and relation classification tasks simultaneously through a shared encoder. The knowledge graph matching and verification unit is used to align and verify the extracted fact triples with the financial knowledge graph. Triples that match directly are judged as verified. Triples that match partially are judged after the missing information is filled in by graph reasoning. Triples that cannot match are marked as pending manual verification or suspected hallucinations. The verification result and confidence level are output. The incremental update unit is used to access external data sources such as regulatory document release systems, financial data terminals, and enterprise information disclosure platforms, and to achieve dynamic updates of the financial knowledge graph through change detection and incremental synchronization mechanisms.

4. The reliability evaluation system for the generation results of a large-scale bank credit risk control model according to claim 1, characterized in that, The aforementioned illusion detection technology module in the financial field employs a multi-model collaborative verification strategy, including: The self-consistency detection unit is used to sample and generate multiple times from the same input. By calculating the semantic similarity between each output and performing cluster analysis, the consistency of the output results is evaluated. Low consistency indicates the risk of hallucination. The cross-validation detection unit is used to call multiple heterogeneous large models to answer the same question. It identifies potential illusions by analyzing the semantic alignment and consistency of the answers between models. For cases where most models are consistent but some models deviate, the deviation results are labeled as illusions. The source tracing and verification detection unit is used to check the authenticity and matching degree of the information sources claimed in the large model generation results. Statements that cannot be verified or whose sources do not match the content are marked as suspected hallucinations. The adversarial detection unit is used to design targeted probing strategies to ask the large model detailed questions and boundary condition questions related to the original answer, and to test the stability and consistency of the answer. Unstable or self-contradictory answer patterns are judged as hallucination indicators. The illusion detection technology module in the financial field integrates the detection results of the above units and calculates the comprehensive score of illusion detection through weighted fusion. The weight of each unit in the weighted fusion is adaptively adjusted according to its historical accuracy in illusion detection in the financial field.

5. The reliability evaluation system for the generation results of a large-scale bank credit risk control model according to claim 1, characterized in that, The dynamic calibration and uncertainty quantification module is configured as follows: Based on the token probability distribution output by the large model, the information entropy of each token is calculated as a token-level uncertainty estimate; By training a dedicated confidence estimation network and combining it with temperature scaling techniques, the confidence of the generated results is semantically calibrated. The confidence estimation network takes the hidden layer representation of a large model as input and predicts the correct probability of the generated results in each reliability dimension. The total uncertainty is decomposed into cognitive uncertainty and random uncertainty by using the Monte Carlo Dropout method or deep ensemble method. Cognitive uncertainty represents the uncertainty caused by insufficient model knowledge, while random uncertainty represents the uncertainty caused by the inherent noise of the input data. For different credit risk control business scenarios, we train scenario-specific calibration models to perform scenario-adaptive confidence estimation and optimization. The output includes the quantification results of uncertainty, including total uncertainty, cognitive uncertainty, random uncertainty, post-calibrated confidence level, and human review priority.

6. The reliability evaluation system for the generation results of a large-scale bank credit risk control model according to claim 1, characterized in that, The adversarial robustness evaluation module is configured to perform at least one of the following tests: Semantic preservation perturbation test involves perturbing the input using at least one of the following methods—synonym substitution, sentence transformation, or information reordering—while maintaining semantic integrity. This generates adversarial examples, and the perturbation intensity is controlled within a preset semantic preservation threshold to test the stability of the model's output. Boundary condition testing involves constructing test cases located near the decision boundary, including customer cases that just meet or do not meet the credit conditions, as well as policy-mandated threshold cases, to test the model's boundary judgment capability. Induced attack testing involves designing adversarial inputs that attempt to induce the model to generate incorrect information or illegal suggestions, thereby testing the model's security protection capabilities. Data contamination simulation test: By injecting a preset proportion of erroneous or contradictory information into the input, the test is conducted to assess the model's noise tolerance and anomaly detection capabilities.

7. The reliability evaluation system for the generation results of a large-scale bank credit risk control model according to claim 1, characterized in that, The evaluation engine layer has a collaborative feedback mechanism among its modules, specifically including: The financial hallucination detection technology module transmits the detected suspected hallucination content and its associated entity and relationship information to the financial knowledge graph fact verification engine module, which guides the fact verification engine to conduct in-depth verification of suspected hallucination areas. The financial knowledge graph fact-checking engine module feeds back the information of unmatched triples to the financial illusion detection technology module to improve the accuracy of illusion detection. The dynamic calibration and uncertainty quantification module transmits the high uncertainty region identifier to the adversarial robustness evaluation module, which guides the adversarial evaluation module to generate more targeted test cases for the high uncertainty region. When calculating the overall reliability score, the fusion decision layer dynamically adjusts the weight coefficients of each dimension based on the uncertainty type distribution output by the dynamic calibration and uncertainty quantification module.

8. The reliability evaluation system for the generation results of a large-scale bank credit risk control model according to claim 1, characterized in that, It also includes an automated assessment pipeline module, configured as follows: Maintain a hierarchical and categorized test case library, covering basic capability testing, business scenario testing, extreme case testing, and regression testing, and support version management and automatic expansion of test cases; The evaluation tasks are executed in parallel through a distributed task scheduling engine, supporting the batch processing of large-scale test cases. Automatically summarizes the evaluation results of each dimension, generates a multi-granularity scoring report including overall score, dimension score, scenario score, and use case score, and supports comparative analysis with historical versions; It provides standardized API interfaces to support continuous integration with model development pipelines and enable automated regression testing after model updates.

9. A method for evaluating the reliability of the generated results of a large-scale bank credit risk control model, as described in any one of claims 1-8, characterized in that... Includes the following steps: Step S1: Receive the evaluation request through the application interface layer, obtain the large model generation result to be evaluated and its corresponding original input data, and preprocess the generation result and the original input data, including text segmentation, entity annotation and format standardization; Step S2: Through the evaluation task scheduler of the evaluation engine layer, according to the pre-configured evaluation strategy, coordinate each module to perform multi-dimensional reliability evaluation on the generated results in parallel or sequentially. The multi-dimensional reliability evaluation includes at least factual accuracy evaluation, logical consistency evaluation, policy compliance evaluation, illusion detection, and uncertainty quantification. The modules exchange intermediate results through a collaborative feedback mechanism to improve the accuracy and completeness of the evaluation. Step S3: In the fusion decision layer, the evaluation scores and uncertainty quantification results of each module are received, and weighted fusion is performed according to the weight configuration corresponding to the business scenario to calculate the comprehensive reliability score. The priority of manual review is determined based on the uncertainty level and business importance. Step S4: Generate a visual evaluation report containing a comprehensive score, scores for each dimension, anomaly annotations, uncertainty decomposition, and improvement suggestions through the application interface layer, and return it to the requesting party; In step S2, the specific steps for the factual accuracy assessment are as follows: The fact triple extraction unit in the financial knowledge graph fact verification engine module uses a joint extraction model based on the Transformer architecture to extract fact triples in the form of subject-predicate-object from the generated results. The knowledge graph matching and verification unit performs entity alignment and relationship verification with the financial knowledge graph, and outputs the matching status and confidence level for each triple. Based on the verification results, the fact recall rate and fact precision rate are calculated respectively, and the fact precision rate is calculated by harmonic mean. The fact recall rate is the ratio of the number of verified triples to the total number of fact triples that should be included in the generated results, and the fact precision rate is the ratio of the number of verified triples to the total number of extracted triples.

10. The method for evaluating the reliability of the generation results of a large-scale bank credit risk control model according to claim 9, characterized in that, When the method is applied to a credit approval opinion generation scenario, step S2 further includes: The relevant functions of the policy compliance assessment dimensions are invoked to match the credit limit suggestion with the concentration restriction rules in the regulatory rule base to verify whether the credit limit meets the concentration restriction requirements, and to match the guarantee plan with the collateral ratio regulations to check whether the guarantee plan meets the collateral ratio regulations. The relevant functions of the aforementioned logical consistency assessment dimension are invoked, and a natural language reasoning model is used to judge the implied relationship between the risk point identification conclusion and the credit recommendation, and to detect whether there is a logical contradiction in identifying high-risk points but giving high credit recommendations; The illusion detection technology module in the financial field is invoked to generate approval opinions by sampling at least 5 times and calculating the semantic consistency between each version. At the same time, at least 2 heterogeneous large models are invoked for cross-validation to identify false or unreliable content.

Images