Enterprise credit evaluation method and system based on fusion of principal modal guidance and dynamic gating

Abstract

The application provides a kind of main modal guidance and dynamic door control fusion enterprise credit evaluation method and system, it is related to enterprise credit evaluation technical field, the evaluation method includes: obtaining original credit data containing text mode and structured financial mode, carries out feature coding processing, obtains text feature and financial feature;Further input to dynamic confidence prediction module, obtain the confidence weight of each mode;With text feature as the leading, cross-modal attention fusion processing is carried out to financial feature and text feature, and preliminary fusion credit feature is obtained;Further input distance perception contrast learning module, based on the continuous value of credit risk label, construct positive and negative sample pairs, and calculate contrast loss;Preliminary fusion credit feature and financial feature are input into mode binding module, through modal cross attention binding and door control enhancement fusion, obtain enhanced multi-modal credit representation;Input credit score module, output enterprise credit grade, credit score or default risk prediction result.

Description

Technical Field

[0001] This invention relates to the field of corporate credit assessment technology, and in particular to a corporate credit assessment method and system that integrates master mode guidance and dynamic gating. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] With the rapid development of digital finance, corporate credit risk assessment is playing an increasingly important role in scenarios such as commercial banks, credit reporting agencies, insurance companies, and supply chain finance. Traditional credit reporting systems mainly rely on structured financial data of enterprises, such as financial statements, tax records, and bank statements. While these data have a certain degree of objectivity, they are significantly lagging and have relatively limited dimensions, making it difficult to comprehensively and timely reflect the true operating conditions and potential risks of enterprises.

[0004] In recent years, unstructured text data such as internet public opinion, news reports, corporate announcements, and bidding information have become important supplementary sources of credit signals, reflecting a company's reputation, operational activity, and potential public opinion risks. However, existing credit analysis methods still face the following technical challenges when processing multi-source heterogeneous information:

[0005] (1) Insufficient fusion of multi-source information: Existing methods mostly use single-modal machine learning models, such as logistic regression based on statistical features or independent text analysis models, which cannot effectively integrate the deep relationship between text public opinion and financial data, and fail to form information complementarity;

[0006] (2) Fixed intermodal weights: Traditional methods cannot dynamically adjust the contribution of text and financial data according to the data quality and information importance of specific samples. For example, in samples with abnormal financial data, the weight of public opinion data cannot be automatically increased.

[0007] (3) Weak model interpretability: Existing models are difficult to support the needs of financial supervision and risk tracing, and cannot clearly show whether the risk warning was triggered by textual risk terms or specific financial indicators, which affects the credibility and application value of the model in financial practice.

[0008] Therefore, there is an urgent need for a credit analysis system that can deeply integrate textual and financial features and has the ability to guide the main modality and dynamically gating, in order to improve the accuracy, robustness and interpretability of credit assessment. Summary of the Invention

[0009] To address at least one of the technical problems mentioned above, this invention introduces a master modality guidance mechanism and a dynamic gating fusion network to achieve sample-level adaptive fusion of textual and financial dual modal features, thus solving the technical problems of insufficient multi-source information fusion, fixed modality weights, and weak model interpretability in the prior art.

[0010] The first aspect of this invention provides a corporate credit assessment method that integrates master mode guidance and dynamic gating, comprising:

[0011] Obtain raw credit data containing textual and structured financial modalities;

[0012] The text modality and the structured financial modality are respectively subjected to feature encoding processing to obtain text features and financial features;

[0013] The text features and financial features are input into the dynamic confidence prediction module for processing to obtain the confidence weights of each modality;

[0014] Using the text features as the primary focus, the financial features and the text features are subjected to cross-modal attention fusion processing to obtain preliminary fused credit features;

[0015] The preliminary fused credit features are input into the distance-aware contrastive learning module for processing. Positive and negative sample pairs are constructed based on the continuous values ​​of credit risk labels, and the contrastive loss is calculated to optimize the semantic consistency and discriminability of the fused features.

[0016] The preliminary fused credit features and the financial features are input into the modal binding module, and enhanced multimodal credit representation is obtained through modal cross-attention binding and gating enhancement fusion.

[0017] The enhanced multimodal credit representation is input into the credit scoring module or risk predictor, and the enterprise credit rating, credit score or default risk prediction result is output.

[0018] Furthermore, the feature encoding process includes:

[0019] For text modalities, a pre-trained language model is used to extract sentence-level semantic representations as text features;

[0020] Extract corporate financial indicators, solvency indicators, transaction indicators, and credit indicators from structured financial models;

[0021] The aggregation of the text features is calculated using the following formula:

[0022] ;

[0023] in, Representing text features, Represents the [CLS] tag vector of the pre-trained language model, { } represents the set of sentence vectors, MeanPool represents the average pooling operation, and AttPool represents the attention pooling operation. , and Let be the weight coefficient, and satisfy... + + =1.

[0024] Furthermore, the processing of the dynamic confidence prediction module includes:

[0025] Each modal feature is input into the fully connected layer and a nonlinear activation function is applied;

[0026] The unnormalized confidence score is output after the second fully connected layer;

[0027] The confidence weights of each modality are obtained by normalization using the Softmax function. These weights are non-negative and sum to 1, and are used to dynamically reflect the confidence level of each modality.

[0028] The confidence weight is calculated using the following formula:

[0029] ;

[0030] in, and These represent the confidence weights for the text modality and the financial modality, respectively. This represents the concatenated vector of textual and financial features. This is the weight matrix of the first fully connected layer. This is the bias vector of the first fully connected layer. This is the weight matrix of the second fully connected layer. This is the bias vector for the second fully connected layer. This represents the ReLU activation function.

[0031] Furthermore, the cross-modal attention fusion processing includes:

[0032] Using financial features as the query and text features as the key and value, calculate the cross-modal attention matrix;

[0033] Based on the weighted summation of the attention matrix and context modeling using a Transformer encoder, preliminary fused credit features are obtained.

[0034] Furthermore, the processing of the distance-aware contrastive learning module includes:

[0035] Obtain the credit risk labels of the samples and calculate the label distance between sample pairs;

[0036] Based on a preset threshold, positive and negative sample pairs are divided;

[0037] The contrast loss is calculated using cosine similarity with temperature parameters, and the contrast loss is jointly optimized with the main task credit prediction loss to improve the model's ability to distinguish risky samples.

[0038] The contrast loss is calculated using the following formula:

[0039] ;

[0040] in, Indicates sample The comparative loss, , , They represent the first i , j , k The fusion credit feature vector obtained from each sample during the fusion stage Indicates sample The set of positive sample pairs For temperature parameters, This represents the cosine similarity function.

[0041] Furthermore, the processing of the modal binding module includes:

[0042] Using the initial fused features as the query and the financial features as the key and value, we calculate the cross-modal attention matrix and perform residual connections to obtain the bound features;

[0043] The bound features are input into the gated enhancement fusion module, where global pooling and channel weight generation are performed to achieve adaptive enhancement and noise suppression of important credit factors.

[0044] Output the enhanced multimodal credit representation.

[0045] A second aspect of the present invention provides an enterprise credit assessment system that integrates master mode guidance and dynamic gating, comprising:

[0046] The data acquisition module is used to acquire raw credit data containing text modalities and structured financial modalities;

[0047] The feature encoding module is used to perform feature encoding processing on the text modality and the structured financial modality respectively to obtain text features and financial features;

[0048] The dynamic confidence prediction module is used to process the text features and financial features input to obtain the confidence weights of each modality.

[0049] The main modality guidance module is used to perform cross-modal attention fusion processing on the financial features and the text features, with the text features as the main focus, to obtain preliminary fused credit features;

[0050] The distance-aware contrastive learning module is used to process the preliminary fused credit features, construct positive and negative sample pairs based on the continuous values ​​of credit risk labels, and calculate the contrastive loss to optimize the semantic consistency and discriminability of the fused features.

[0051] The modal binding module is used to input the preliminary fused credit features and the financial features, and obtain the enhanced multimodal credit representation through modal cross-attention binding and gating enhancement fusion;

[0052] The risk prediction output module is used to input the enhanced multimodal credit representation into the credit scoring module or risk predictor, and output the enterprise credit rating, credit score or default risk prediction result.

[0053] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the enterprise credit assessment method of master mode guidance and dynamic gating fusion as described in the first aspect of the present invention.

[0054] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the enterprise credit assessment method that integrates master mode guidance and dynamic gating as described in the first aspect of the present invention.

[0055] A fifth aspect of the present invention provides a computer program product comprising software code, wherein the program in the software code performs the steps of the enterprise credit assessment method that integrates master mode guidance and dynamic gating as described in the first aspect of the present invention.

[0056] Compared with existing technologies, the enterprise credit assessment method and system that integrates master mode guidance and dynamic gating provided by the present invention has the following beneficial effects:

[0057] (1) Based on the dynamic confidence prediction module, this invention can automatically evaluate the data quality and reliability of each modality through the sample-level adaptive weight allocation mechanism, increase the weight of the text modality when the financial data is abnormal, and enhance the contribution of the financial data when the text quality is low, thereby suppressing noise modality interference and improving the accuracy of credit risk identification and model robustness.

[0058] (2) Based on the main modality guided fusion, this invention uses corporate text credit information as semantic anchor point and guides the financial data modality to align and fuse in the semantic space through cross-modal attention mechanism, establishes a deep association between text and financial features, ensures the semantic consistency of credit information, and solves the problem of simple splicing of multi-source information and lack of semantic interaction in traditional methods;

[0059] (3) Based on the technical features of the distance-aware contrastive learning module, this invention introduces continuous distance between credit risk labels to construct high-quality positive and negative sample pairs, so that the model can maintain discriminativeness and consistency in the credit score space, and enhance the model's ability to aggregate enterprises with similar risks and distinguish enterprises with different risks.

[0060] (4) In order to further enhance the feature expression capability, this invention is based on the technical features of modality binding and gating enhancement module. Through high-order cross-modal dependency modeling and feature channel enhancement, it ensures that key credit risk factors are fully expressed, while suppressing redundant and noisy information, thereby improving the quality of credit representation.

[0061] (5) In response to the technical problem of weak interpretability of existing models, the collaborative work of the various functional modules provided by this invention provides financial institutions with an intelligent credit risk assessment framework that integrates public opinion text and financial data. Through modular design, the functions and outputs of each processing stage are clearly defined. Combined with optional interpretation and visualization modules, the basis for risk judgment can be clearly displayed, meeting the compliance and interpretability requirements of multiple scenarios such as credit approval, post-loan monitoring, investment due diligence and supply chain finance risk control. Attached Figure Description

[0062] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0063] Figure 1 This is a flowchart of a corporate credit assessment method that integrates master mode guidance and dynamic gating, provided in Embodiment 1 of the present invention;

[0064] Figure 2 This is a schematic diagram of the enterprise credit assessment system that integrates main modality guidance and dynamic gating provided in Embodiment 2 of the present invention. Detailed Implementation

[0065] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0066] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0067] The parameter symbols used in this specification are defined as follows: W1, b1, W2, and b2 are the weight matrices and bias terms of the two fully connected layers in the dynamic confidence prediction module, respectively; F fuse,i F fuse,j F fuse,k These represent the credit feature vectors obtained from different enterprise samples during the feature fusion stage; Δ ij The distance between samples represents the credit risk label distance; τ is the temperature parameter, and σ(·) is the ReLU activation function. Unless otherwise stated, the parameter symbols in this invention have their conventional mathematical meanings in deep learning models, used to characterize network weights, biases, feature vectors, similarity, and loss function weights, etc.

[0068] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0069] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0070] Example 1

[0071] like Figure 1 This embodiment provides a corporate credit assessment method that integrates master mode guidance and dynamic gating, including:

[0072] Obtain raw credit data containing textual and structured financial modalities;

[0073] The text modality and the structured financial modality are respectively subjected to feature encoding processing to obtain text features and financial features;

[0074] The text features and financial features are input into the dynamic confidence prediction module for processing to obtain the confidence weights of each modality;

[0075] Using the text features as the primary focus, the financial features and the text features are subjected to cross-modal attention fusion processing to obtain preliminary fused credit features;

[0076] The preliminary fused credit features are input into the distance-aware contrastive learning module for processing. Positive and negative sample pairs are constructed based on the continuous values ​​of credit risk labels, and the contrastive loss is calculated to optimize the semantic consistency and discriminability of the fused features.

[0077] The preliminary fused credit features and the financial features are input into the modal binding module, and enhanced multimodal credit representation is obtained through modal cross-attention binding and gating enhancement fusion.

[0078] The enhanced multimodal credit representation is input into the credit scoring module or risk predictor, and the enterprise credit rating, credit score or default risk prediction result is output.

[0079] This invention provides a complete bimodal enterprise credit assessment process, achieving deep integration of text and financial data through multi-stage fusion processing. Its core innovation lies in the introduction of a master modality guidance mechanism and a dynamic gating fusion strategy, solving the problem of traditional methods that simply splice together multi-source information without semantic interaction. In practical applications, this method can adaptively adjust the processing strategy according to the characteristics of different enterprise samples, significantly improving the accuracy and robustness of credit risk assessment.

[0080] Specifically, the feature encoding process includes:

[0081] For text modalities, a pre-trained language model is used to extract sentence-level semantic representations as text features;

[0082] Extract corporate financial indicators, solvency indicators, transaction indicators, and credit indicators from structured financial models;

[0083] The aggregation of the text features is calculated using the following formula:

[0084] ;

[0085] in, Representing text features, Represents the [CLS] tag vector of the pre-trained language model, { } represents the set of sentence vectors, MeanPool represents the average pooling operation, and AttPool represents the attention pooling operation. , and Let be the weight coefficient, and satisfy... + + =1.

[0086] This invention addresses the problem of insufficient information utilization in traditional text feature extraction through a specific feature encoding strategy, particularly a hybrid aggregation formula for text features. By combining three different features—[CLS] tag vectors, average pooling, and attention pooling—it preserves global semantic information while enhancing the contribution of key sentences, thereby improving the comprehensiveness and accuracy of text feature representation. This technique ensures that the text modality can fully capture key credit signals from corporate disclosure documents and public opinion information, providing high-quality input features for subsequent multimodal fusion.

[0087] Specifically, the processing of the dynamic confidence prediction module includes:

[0088] Each modal feature is input into the fully connected layer and a nonlinear activation function is applied;

[0089] The unnormalized confidence score is output after the second fully connected layer;

[0090] The confidence weights of each modality are obtained by normalization using the Softmax function. These weights are non-negative and sum to 1, and are used to dynamically reflect the confidence level of each modality.

[0091] The confidence weight is calculated using the following formula:

[0092] ;

[0093] in, and These represent the confidence weights for the text modality and the financial modality, respectively. This represents the concatenated vector of textual and financial features. , These are the weight matrix and bias vector for the first fully connected layer (FC Layer 1). They are used to perform linear mapping and dimensionality reduction on the input features (concatenated text features and financial features) to capture preliminary modal correlations. and This represents the weight matrix and bias vector of the second fully connected layer (FC Layer 2). It is used to further perform nonlinear transformations on the features, outputting unnormalized confidence scores (i.e., the confidence scores of each modality in the current sample). This represents the ReLU activation function.

[0094] The dynamic confidence prediction module provided by this invention achieves adaptive allocation of modality weights at the sample level through a learnable parameterized network. This feature directly solves the problem of fixed weights between modalities in the background technology by dynamically adjusting the contribution of each modality feature through analysis of its intrinsic quality and consistency. For example, when abnormal financial data or significant risk events appear in the text are detected, the module automatically adjusts the weight allocation to suppress the influence of unreliable modalities. This adaptive mechanism significantly improves the robustness of the model in the face of data quality issues, ensuring the reliability of credit assessment results.

[0095] Specifically, the cross-modal attention fusion process includes:

[0096] Using financial features as the query and text features as the key and value, calculate the cross-modal attention matrix;

[0097] Based on the weighted summation of the attention matrix and context modeling using a Transformer encoder, preliminary fused credit features are obtained.

[0098] The cross-modal attention fusion processing provided by this invention achieves deep semantic alignment by establishing an interaction mechanism between text and financial features. The design, using financial features as the query and text features as the key and value, allows financial indicators to find corresponding explanations and support in the text semantic space, forming a semantic bridge between financial data and textual descriptions. This fusion approach not only enhances the expressive power of features but also improves the interpretability of the model, as attention weights can be used to analyze which text content is most closely related to specific financial indicators.

[0099] Specifically, the processing of the distance-aware contrastive learning module includes:

[0100] Obtain the credit risk labels of the samples and calculate the label distance between sample pairs;

[0101] Based on a preset threshold, positive and negative sample pairs are divided;

[0102] The contrast loss is calculated using cosine similarity with temperature parameters, and the contrast loss is jointly optimized with the main task credit prediction loss to improve the model's ability to distinguish risky samples.

[0103] The contrast loss is calculated using the following formula:

[0104] ;

[0105] in, Indicates sample The comparative loss, , , They represent the first i , j , k The fused credit feature vector obtained from each sample during the fusion phase. This feature, after being fused using master modality guidance and dynamic gating, is used for comparative learning or credit risk prediction. Indicates sample The set of positive sample pairs For temperature parameters, This represents the cosine similarity function.

[0106] The distance-aware contrastive learning module provided in this invention constructs more reasonable positive and negative sample pairs by utilizing the continuous characteristics of credit risk labels. Unlike traditional contrastive learning that uses discrete labels, this technique divides sample pairs based on label distance thresholds, making companies with similar credit risks closer together in the feature space, while companies with large risk differences are kept apart. This mechanism significantly improves the model's discriminative ability in the credit semantic space, especially enabling more refined differentiation between companies with similar credit risks, thus solving the problem of insufficient modeling of continuous risk changes in traditional methods.

[0107] Specifically, the processing of the modal binding module includes:

[0108] Using the initial fused features as the query and the financial features as the key and value, we calculate the cross-modal attention matrix and perform residual connections to obtain the bound features;

[0109] The bound features are input into the gated enhancement fusion module, where global pooling and channel weight generation are performed to achieve adaptive enhancement and noise suppression of important credit factors.

[0110] Output the enhanced multimodal credit representation.

[0111] The modality binding module provided in this invention achieves high-order interaction and refinement of multimodal features through a secondary attention mechanism and a gating enhancement strategy. By using the initially fused features as a query for further attention interaction with financial features, deeper cross-modal dependencies can be captured. Meanwhile, the gating enhancement fusion strengthens feature channels important for credit assessment and suppresses irrelevant or noisy information through adaptive generation of channel weights. This dual enhancement mechanism ensures that the final credit representation contains rich cross-modal information while focusing on key risk factors, improving the model's expressive power and generalization performance.

[0112] In one specific embodiment, the enterprise credit assessment method includes:

[0113] 1. Data Acquisition and Text Preprocessing

[0114] We collect textual data sources related to corporate credit, including but not limited to: annual / quarterly reports, audit opinions, regulatory announcements and penalty notices, prospectuses and bond offering memorandums, announcements of major events, court judgments and enforcement information, news and public opinion and research report summaries, bidding and winning announcements, supply chain and accounts receivable descriptions, pledge / guarantee / litigation disclosures, etc.

[0115] Perform cleaning and standardization on the original text: remove HTML / footnote / table noise, standardize currency and time expressions, perform regular expression extraction on monetary ratios and financial items (such as "operating revenue", "net operating cash flow", "interest coverage ratio"), and add structured annotation tags (such as...) to the text.<F_CF_OP> ...< / F_CF_OP> Sentence segmentation and paragraph breaking, filtering of very short sentences, and retention of key information disclosure paragraphs.

[0116] 2. Text Feature Coding (Finance-PLM)

[0117] Domain-adaptive pre-trained language models (such as FinBERT / RoBERTa-Finance) are used to encode sentence-level / paragraph-level text to obtain enterprise text representations. (Optimal d=768). Paragraph aggregation employs a hybrid strategy of "CLSCLS + average pooling + attention pooling":

[0118]

[0119] Where {h s} is the set of sentence vectors. + + =1.

[0120] 3. Enhanced risk semantics

[0121] By combining a financial dictionary with event templates, the system performs entity recognition (enterprises, shareholders, upstream and downstream partners, guarantors), event extraction (default / litigation / pledge / guarantee / merger and acquisition / senior management changes, etc.), and metric extraction (amount, proportion, time window) on the text. Event and entity labels are injected into the encoder (prompt tuning / adapters) through learnable word vectors or cue vectors to improve the ability to perceive risk semantics.

[0122] 4. Distance-aware contrastive learning (continuous risk labeling)

[0123] Given the text representation zi=Norm(Ft) of sample i and the continuous credit risk label yi (which can be a standardized credit score or PD), define the label distance between sample pairs. .

[0124] Set a threshold c (preferably c=0.4): when Classify as positive, otherwise as negative. Construct an InfoNCE-type loss using cosine similarity and temperature τ (preferably τ=0.07):

[0125] ;

[0126] Where P(i) is the set of pairs of samples i. This mechanism utilizes the fine-grained differences in continuous labels to make textual representations of similar credit risks more aggregated and those that are far apart more separated.

[0127] 5. Primary task prediction head and joint loss

[0128] Send Ft into the task header output:

[0129] 5.1 Credit rating classification (e.g., AAA–D): Fully connected + Softmax, cross-entropy loss ;

[0130] 5.2 Probability of Default Regression (PD): Fully Connected + Sigmoid / Linear, MAE / MSE Loss Joint optimization objective:

[0131]

[0132] Optimal hyperparameters: dropout 0.1, learning rate Batch size 16–32, training epochs 5–10.

[0133] 6. Explanation and Compliance Output

[0134] Generate an interpretable report on the final results: output key sentences / paragraphs that trigger high-risk judgments, related risk events and financial terms, and provide confidence scores and text change trends over the past three periods to meet audit and compliance requirements.

[0135] Example 2

[0136] like Figure 2 This embodiment provides an enterprise credit assessment system that integrates master mode guidance and dynamic gating, including:

[0137] The data acquisition module is used to acquire raw credit data containing text modalities and structured financial modalities;

[0138] The feature encoding module is used to perform feature encoding processing on the text modality and the structured financial modality respectively to obtain text features and financial features;

[0139] The dynamic confidence prediction module is used to process the text features and financial features input to obtain the confidence weights of each modality.

[0140] The main modality guidance module is used to perform cross-modal attention fusion processing on the financial features and the text features, with the text features as the main focus, to obtain preliminary fused credit features;

[0141] The distance-aware contrastive learning module is used to process the preliminary fused credit features, construct positive and negative sample pairs based on the continuous values ​​of credit risk labels, and calculate the contrastive loss to optimize the semantic consistency and discriminability of the fused features.

[0142] The modal binding module is used to input the preliminary fused credit features and the financial features, and obtain the enhanced multimodal credit representation through modal cross-attention binding and gating enhancement fusion;

[0143] The risk prediction output module is used to input the enhanced multimodal credit representation into the credit scoring module or risk predictor, and output the enterprise credit rating, credit score or default risk prediction result.

[0144] In one specific embodiment, the enterprise credit assessment system includes the following functional modules:

[0145] 1. Feature Encoding Module

[0146] Used to acquire and preprocess enterprise text modal and structured financial modal data:

[0147] Text-based modal data includes corporate annual reports, financial reports, regulatory announcements, news and public opinion, judicial information, credit information disclosures, and prospectuses, etc.

[0148] Structured financial modal data includes balance sheets, income statements, cash flow statements, solvency indicators, revenue growth rates, credit rating records, public credit scores, and tax information.

[0149] The module performs feature encoding on the two types of data respectively:

[0150] The text modality uses pre-trained financial language models (such as FinBERT or RoBERTa-Finance) to extract semantic vector representations;

[0151] Structured financial modalities are vectorized through standardization, feature selection, and multilayer perceptron (MLP) network.

[0152] Finally, we obtain the text features Ft and the financial features Fs.

[0153] 2. Dynamic Gating Fusion Module

[0154] Used for dynamic weighting and information filtering between bimodal features.

[0155] This module adaptively assigns weight parameters based on the modal quality, autocorrelation, and consistency of the samples using a gating network. and The calculation method is as follows:

[0156]

[0157] in, It is the ReLU activation function. This indicates a splicing operation.

[0158] Through a dynamic gating mechanism, the system increases the weight of text with high quality and automatically suppresses its impact when financial data is abnormal or missing, thereby achieving adaptive fusion between modalities.

[0159] 3. Main Modal Guidance Module

[0160] The text modality is the dominant modality, guiding the alignment of financial modality features within the semantic space.

[0161] This module calculates the interaction correlation between text and financial features through a dominant modality attention mechanism:

[0162]

[0163]

[0164] in , , These are queries and key-value matrices representing textual and financial features, respectively. For feature dimensions.

[0165] Features after fusion Text features By performing splicing and linear mapping, a preliminary fused representation is obtained. .

[0166] 4. Distance-aware contrast learning module

[0167] Discriminative features for optimizing bimodal fusion features under the guidance of continuous credit labels (such as credit scores or probability of default, PD).

[0168] This module calculates the distance between credit tags of different samples. Positive and negative sample pairs are distinguished based on a threshold c (preferably 0.4), and a temperature parameter is used. The cosine similarity is used to construct the InfoNCE contrast loss:

[0169]

[0170] By using a contrastive learning mechanism, enterprise samples with similar risks are more closely clustered in the semantic space, while samples with large risk differences are kept at a greater interval, thereby improving the discriminativeness and continuity of credit risk prediction.

[0171] 5. Credit Risk Prediction Module

[0172] It is used to input the fused credit information into the predictor and output corporate credit ratings or risk indicators.

[0173] Classification task: Output credit rating (AAA–D), using a Softmax structure and cross-entropy loss;

[0174] Regression task: Output the probability of default (PD) or credit score, using mean squared error (MSE) loss.

[0175] The total loss function is:

[0176]

[0177] in The main task loss is γ, and the weight of the contrastive learning loss is γ (preferably 0.3–0.5).

[0178] 6. Results Interpretation and Visualization Module

[0179] Interpretable outputs used to generate model predictions include:

[0180] Visualize the risk weight of key text paragraphs (such as statements that trigger "abnormal operations" or "legal risks" clauses);

[0181] Sensitivity ranking of key financial indicators (e.g., the impact of "cash flow fluctuations" and "abnormal debt-to-equity ratio" on PD);

[0182] An optional visualization interface is available to display risk assessment results, credit trends, and traceability reports, meeting the audit and compliance needs of financial institutions.

[0183] System Operation Instructions

[0184] This system can be deployed on bank credit reporting servers, supply chain finance risk control platforms, or cloud environments of third-party credit reporting agencies, supporting batch processing of enterprise credit data and online credit risk assessment.

[0185] Through a dynamic gating fusion mechanism of primary modality (text) and secondary modality (financial), this system can maintain stable credit prediction performance and interpretability even when faced with missing financial statements, incomplete text disclosures, or high information heterogeneity.

[0186] Advantages of this embodiment:

[0187] (1) Only text and financial data sources are used, avoiding the collection and privacy costs caused by unnecessary modalities such as audio and images;

[0188] (2) The main modality guidance and dynamic gating mechanism can adaptively adjust the contribution of text and finance, thereby enhancing the robustness of the model;

[0189] (3) Distance-aware contrastive learning ensures the continuity and separability of credit risk prediction;

[0190] (4) The interpretable module meets the needs of credit investigation, compliance supervision and internal risk control analysis;

[0191] (5) It has good scalability and can be directly connected to the bank's existing credit data warehouse or financial cloud platform.

[0192] Example 3

[0193] Embodiment 3 of the present invention provides an electronic device.

[0194] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the enterprise credit assessment method that integrates master mode guidance and dynamic gating as described in Embodiment 1 of the present invention.

[0195] The detailed steps are the same as those of the enterprise credit assessment method that integrates master mode guidance and dynamic gating provided in Example 1, and will not be repeated here.

[0196] Example 4

[0197] Embodiment 4 of the present invention provides a computer-readable storage medium.

[0198] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the enterprise credit assessment method that integrates master mode guidance and dynamic gating as described in Embodiment 1 of the present invention.

[0199] The detailed steps are the same as those of the enterprise credit assessment method that integrates master mode guidance and dynamic gating provided in Example 1, and will not be repeated here.

[0200] Example 5

[0201] Embodiment 5 of the present invention provides a computer program product.

[0202] A computer program product includes software code, wherein the program in the software code performs the steps of the enterprise credit assessment method that integrates master mode guidance and dynamic gating as described in Embodiment 1 of the present invention.

[0203] The detailed steps are the same as those of the enterprise credit assessment method that integrates master mode guidance and dynamic gating provided in Example 1, and will not be repeated here.

[0204] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0205] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0206] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0207] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0208] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Clearly, those skilled in the art can make various alterations and variations to the invention without departing from its spirit and scope.

[0209] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A corporate credit assessment method integrating master mode guidance and dynamic gating, characterized in that, include: Obtain raw credit data containing textual and structured financial modalities; The text modality and the structured financial modality are respectively subjected to feature encoding processing to obtain text features and financial features; The text features and financial features are input into the dynamic confidence prediction module for processing to obtain the confidence weights of each modality; The dynamic confidence prediction module's processing includes: inputting each modality's features into a fully connected layer and applying a non-linear activation function; outputting an unnormalized confidence score through a second fully connected layer; and obtaining the confidence weights for each modality through normalization using a Softmax function. These weights are non-negative and sum to 1, dynamically reflecting the confidence level of each modality. The confidence weights are calculated using the following formula: ； in, and These represent the confidence weights for the text modality and the financial modality, respectively. This represents the concatenated vector of textual and financial features. This is the weight matrix of the first fully connected layer. This is the bias vector of the first fully connected layer. This is the weight matrix of the second fully connected layer. This is the bias vector for the second fully connected layer. Represents the ReLU activation function; Using the text features as the primary focus, the financial features and text features are fused together via cross-modal attention processing. The financial features are used as the query, and the text features are used as the key and value to calculate the cross-modal attention matrix. The attention matrix is ​​then weighted and summed, and contextual modeling is performed using a Transformer encoder to obtain preliminary fused credit features. The preliminary fused credit features are then input into a distance-aware contrastive learning module for processing. Positive and negative sample pairs are constructed based on the continuous values ​​of credit risk labels, and contrastive loss is calculated to optimize the semantic consistency and discriminativeness of the fused features. The preliminary fused credit features and the financial features are input into the modal binding module, and enhanced multimodal credit representation is obtained through modal cross-attention binding and gating enhancement fusion. The enhanced multimodal credit representation is input into the credit scoring module or risk predictor, and the enterprise credit rating, credit score or default risk prediction result is output.

2. The method as described in claim 1, characterized in that, The feature encoding process includes: For text modalities, a pre-trained language model is used to extract sentence-level semantic representations as text features; Extract corporate financial indicators, solvency indicators, transaction indicators, and credit indicators from structured financial models; The aggregation of the text features is calculated using the following formula: ； in, Representing text features, Represents the [CLS] tag vector of the pre-trained language model, { } represents the set of sentence vectors, MeanPool represents the average pooling operation, and AttPool represents the attention pooling operation. , and Let be the weight coefficient, and satisfy... + + =1.

3. The method as described in claim 1, characterized in that, The processing of the distance-aware contrastive learning module includes: Obtain the credit risk labels of the samples and calculate the label distance between sample pairs; Based on a preset threshold, positive and negative sample pairs are divided; The contrast loss is calculated using cosine similarity with temperature parameters, and the contrast loss is jointly optimized with the main task credit prediction loss to improve the model's ability to distinguish risky samples. The contrast loss is calculated using the following formula: ； in, Indicates sample The comparative loss, , , They represent the first i , j , k The fusion credit feature vector obtained from each sample during the fusion stage Indicates sample The set of positive sample pairs For temperature parameters, This represents the cosine similarity function.

4. The method as described in claim 1, characterized in that, The processing of the modal binding module includes: Using the initial fused features as the query and the financial features as the key and value, we calculate the cross-modal attention matrix and perform residual connections to obtain the bound features; The bound features are input into the gated enhancement fusion module, where global pooling and channel weight generation are performed to achieve adaptive enhancement and noise suppression of important credit factors. Output the enhanced multimodal credit representation.

5. A corporate credit assessment system integrating master mode guidance and dynamic gating, characterized in that, include: The data acquisition module is used to acquire raw credit data containing text modalities and structured financial modalities; The feature encoding module is used to perform feature encoding processing on the text modality and the structured financial modality respectively to obtain text features and financial features; The dynamic confidence prediction module is used to process the text features and financial features input to obtain the confidence weights of each modality. The dynamic confidence prediction module's processing includes: inputting each modality's features into a fully connected layer and applying a non-linear activation function; outputting an unnormalized confidence score through a second fully connected layer; and obtaining the confidence weights for each modality through normalization using a Softmax function. These weights are non-negative and sum to 1, dynamically reflecting the confidence level of each modality. The confidence weights are calculated using the following formula: ； in, and These represent the confidence weights for the text modality and the financial modality, respectively. This represents the concatenated vector of textual and financial features. This is the weight matrix of the first fully connected layer. This is the bias vector of the first fully connected layer. This is the weight matrix of the second fully connected layer. This is the bias vector for the second fully connected layer. Represents the ReLU activation function; The main modality guidance module is used to perform cross-modal attention fusion processing on the financial features and the text features, with the financial features as the query and the text features as the key and value, to calculate the cross-modal attention matrix; based on the attention matrix, the weighted sum is calculated and the context is modeled by the Transformer encoder to obtain the preliminary fused credit features; The distance-aware contrastive learning module is used to process the preliminary fused credit features, construct positive and negative sample pairs based on the continuous values ​​of credit risk labels, and calculate the contrastive loss to optimize the semantic consistency and discriminability of the fused features. The modal binding module is used to input the preliminary fused credit features and the financial features, and obtain the enhanced multimodal credit representation through modal cross-attention binding and gating enhancement fusion; The risk prediction output module is used to input the enhanced multimodal credit representation into the credit scoring module or risk predictor, and output the enterprise credit rating, credit score or default risk prediction result.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the enterprise credit assessment method that integrates master mode guidance and dynamic gating as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the enterprise credit assessment method that combines master mode guidance and dynamic gating as described in any one of claims 1 to 4.

8. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps of the enterprise credit assessment method that integrates master mode guidance and dynamic gating as described in any one of claims 1 to 4.

Images