A review method for a financial teller processing system

Abstract

The application discloses a review method for a financial counter service processing system, and belongs to the technical field of bank risk control review, and comprises four stages of data acquisition and standardized processing, core BERT model construction and optimization, real-time intelligent review judgment and review feedback output; the application collects business specifications, counter audio and video, voucher images, historical risk control and system input data from multiple sources, respectively extracts feature vectors, fuses input fine-tuning BERT models, realizes business rule guided cross-modal semantic alignment verification relying on an attention mechanism, and dynamically optimizes the model in combination with historical risk data; the application realizes automatic intelligent review of bank full-scene counter services, solves the problems of high cost of traditional manual review, difficulty in checking multiple source vouchers and poor universality of fixed rules, greatly reduces human input, improves review efficiency and risk identification accuracy, the model has continuous and independent iteration capability, and is suitable for various bank branch counter transaction risk control review scenes.

Description

Technical Field

[0001] This invention relates to the field of risk control review technology for bank counter business, specifically a review method for a financial counter business processing system. Background Technology

[0002] The traditional review process for bank branch counter services, which combines manual review with fixed system rule verification, has multiple drawbacks: 1. High labor costs: Dedicated review positions are required to review each counter transaction, resulting in a backlog of reviews during peak periods and low efficiency in handling business. 2. High development and maintenance costs: Specific verification rules need to be developed for different counter transaction scenarios, and the system verification logic needs to be iterated and updated when new business specifications are added, resulting in large development investment; 3. Difficulty in coordinating and verifying multi-source data: It is impossible for manual verification to quickly and synchronously check the consistency of counter audio recordings, on-site video recordings, paper business vouchers, and system-entered data, which makes it very easy to miss or mis-verify. 4. Lagging risk identification: Traditional fixed rules can only identify known risk scenarios and cannot predict new and hidden operational risks based on historical violation cases; The existing patent CN113515792B discloses an intelligent image verification solution that relies solely on single image data for verification. It cannot integrate voice, business rules, historical risk control data, and real-time data input to achieve cross-modal unified verification, resulting in insufficient adaptability and completeness of risk identification.

[0003] In summary, existing review technologies struggle to balance review efficiency, labor costs, multi-source data consistency verification, and dynamic risk identification requirements. There is an urgent need for a universal, iterative, and cross-modal integrated intelligent review method for bank counters. Summary of the Invention

[0004] The purpose of this invention is to provide a verification method for a financial counter service processing system to solve the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a verification method for a financial counter service processing system, comprising the following steps: S1. Data Acquisition and Standardization Processing Stage: The data acquisition module collects business specification data from multiple sources, including counter interaction audio and video recordings, voucher document image data, historical records and risk control data, and real-time system data. Feature vectors are extracted from each type of source data and then concatenated and fused to obtain a unified high-dimensional feature vector. ,in ; For business specification feature vectors, For interactive audio and video feature vectors, The feature vector of the voucher image. For historical experience feature vectors, This represents the feature vector of the risk pattern. S2. Core Model Construction and Optimization Stage: Based on the BERT large model as the basic architecture, multi-dimensional feature fusion and attention mechanism initialization, cross-modal alignment logic training, business knowledge graph embedding, iterative optimization based on historical risk data are completed based on the unified high-dimensional feature vector. Lightweight model deployment is completed through model pruning and edge computing, and a dynamic continuous learning mechanism is configured. S3. Real-time processing and judgment stage: The unified high-dimensional feature vector is input into the optimized BERT verification model to perform multi-dimensional semantic consistency judgment, risk-driven dynamic priority scheduling, multi-task parallel anomaly identification, and simultaneously generate an interpretable evidence chain. S4. Review and Feedback Output Stage: Output review conclusions, risk level labels, and visually interpretable analytical evidence to complete the intelligent review of counter services.

[0006] Preferably, the data collected from multiple sources in step S1 specifically includes: Business compliance data: The bank's latest business operation procedures, compliance requirements, and institutional guidelines; Counter interaction audio and video recording data: audio recordings of conversations between tellers and customers, and on-site video of the counter; Image data of voucher documents: images of historical business documents, customer certificates, and business vouchers; Historical records and risk control data: over-the-counter transaction records, manually reviewed and annotated data, historical abnormal events, and post-risk control supervision records; The system inputs data in real time: transaction elements such as account number, amount, and loan identifier entered by tellers in the business system, page restoration data, and real-time data extracted from the database.

[0007] Preferably, the specific methods for processing and extracting feature vectors from the source data in step S1 include: Business Specification Features The business text is parsed using Natural Language Processing (NLP) and converted into semantic features for the model, which are then used as the query vector for the attention mechanism. Interactive audio and video features Speech recognition transcribes recorded text, semantic analysis extracts customer's spoken business type, transaction amount, and business intent, and generates audio and video semantic features; Voucher Image Features Visual features of seals, signatures, and document anti-counterfeiting points are extracted through image recognition and convolutional visual encoder and quantized to serve as key vectors and value vectors for the attention mechanism. Historical experience and risk characteristics Complete missing information related to historical transaction records, correct manually reviewed and labeled samples, and extract characteristics of historical risk events; The system inputs features in real time: standardized conversion of real-time input JSON files and extracted data from the database generates real-time transaction features.

[0008] Preferably, step S2, the multidimensional feature fusion and attention mechanism initialization step, includes: Will As a query, , Using these as Key and Value respectively, the semantic similarity between Query and Key is calculated through dot product, generating a visual representation guided by the semantics of business specifications, and locating the verification area in the voucher image that matches the business rules.

[0009] Preferably, the verification objects for the cross-modal alignment logic training in step S2 are three types of semantic data: from Extracted customer voice service intent, by Extracted voucher text information, by The system records transaction data in real time; when the semantic space values ​​of the three are aligned, the business is marked as compliant; when the semantic difference exceeds a preset threshold, the business is marked as abnormal.

[0010] Preferably, step S2, the business knowledge graph embedding step, includes: embedding business specifications... The business knowledge graph and risk control rules are embedded into the semantic space of the BERT model. When the model outputs the review conclusion, it binds the corresponding business specification clauses to make the review results interpretable.

[0011] Preferably, the dynamic continuous learning mechanism in step S2 is as follows: collect newly added business cases at the counter, manually review newly labeled samples, and add risk event data, continuously fine-tune the BERT review model parameters, dynamically update the risk identification and judgment criteria, and realize the model's autonomous iterative optimization.

[0012] Preferably, step S3, risk-driven dynamic priority scheduling, specifically involves: based on , For current transactions, a risk profile is created, and high-risk transactions such as large cash withdrawals, high-frequency abnormal transactions, and transactions that match historical violation characteristics are assigned a higher review priority, with high-risk transactions being reviewed intelligently first.

[0013] Preferably, the multi-task parallel computing and anomaly identification logic in step S3 is as follows: parallel execution of feature vector conversion for multiple transactions and model inference; when , , The semantic alignment score is below a preset threshold, or the visual features of the voucher signature or seal do not match. When implementing business procedures, abnormal transactions are identified and corresponding risk tags are assigned.

[0014] Preferably, the method for generating the interpretable evidence chain in step S3 is as follows: map the model attention weights to the original data source, associate them with the corresponding business specification clauses, voucher image verification areas, and audio text fragments, retain the complete review and judgment basis, and output them synchronously to the review feedback module.

[0015] Compared with the prior art, the beneficial effects of the present invention are: the verification method for a financial counter service processing system: 1. Free up review manpower and reduce labor costs: All routine counter transactions are reviewed automatically, and human intervention is only required for complex and difficult transactions that the model cannot identify. This eliminates dedicated review positions or reduces the number of review personnel, thereby reducing labor costs in the long term. 2. Solves the challenge of multi-source data collaborative verification: Automatically completes cross-modal alignment verification of voice, video, paper vouchers, and system-entered data, with efficiency and accuracy far exceeding that of manual verification; 3. Standardized modules reduce development and maintenance costs: One set of intelligent verification modules can be adapted to all counter transaction scenarios, eliminating the need to develop separate verification rules for each type of business and reducing investment in system iteration development; 4. Enhance review throughput and risk management capabilities: Parallel computing + dynamic high-risk priority scheduling significantly improves review processing speed and quickly intercepts large-value, high-frequency and other high-risk transactions; 5. Model autonomous iteration reduces human intervention: The continuous learning mechanism automatically optimizes the recognition capability as business updates, and the number of transactions requiring manual review continues to decrease; The review results are traceable and explainable: each abnormal transaction is linked to corresponding business specifications, voucher image areas, and audio recording text evidence, allowing managers to quickly locate the root cause of business errors. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the present invention; Figure 2 This is a schematic diagram of the verification steps of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention have been clearly and completely described. 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.

[0018] Example 1

[0019] Please see Figures 1-2 This invention provides a technical solution: a verification method for a financial counter service processing system, comprising the following steps: S1. Data Acquisition and Standardization Processing Stage: The data acquisition module collects business specification data from multiple sources, including counter interaction audio and video recordings, voucher document image data, historical records and risk control data, and real-time system data. Feature vectors are extracted from each type of source data and then concatenated and fused to obtain a unified high-dimensional feature vector. ,in ; For business specification feature vectors, For interactive audio and video feature vectors, The feature vector of the voucher image. For historical experience feature vectors, This represents the feature vector of the risk pattern. S2. Core Model Construction and Optimization Stage: Based on the BERT large model as the basic architecture, multi-dimensional feature fusion and attention mechanism initialization, cross-modal alignment logic training, business knowledge graph embedding, iterative optimization based on historical risk data are completed based on the unified high-dimensional feature vector. Lightweight model deployment is completed through model pruning and edge computing, and a dynamic continuous learning mechanism is configured. S3. Real-time processing and judgment stage: The unified high-dimensional feature vector is input into the optimized BERT verification model to perform multi-dimensional semantic consistency judgment, risk-driven dynamic priority scheduling, multi-task parallel anomaly identification, and simultaneously generate an interpretable evidence chain. S4. Review and Feedback Output Stage: Output review conclusions, risk level labels, and visually interpretable analytical evidence to complete the intelligent review of counter services.

[0020] The data collected from multiple sources in step S1 specifically includes: Business compliance data: The bank's latest business operation procedures, compliance requirements, and institutional guidelines; Counter interaction audio and video recording data: audio recordings of conversations between tellers and customers, and on-site video of the counter; Image data of voucher documents: images of historical business documents, customer certificates, and business vouchers; Historical records and risk control data: over-the-counter transaction records, manually reviewed and annotated data, historical abnormal events, and post-risk control supervision records; The system inputs data in real time: transaction elements such as account number, amount, and loan identifier entered by tellers in the business system, page restoration data, and real-time data extracted from the database.

[0021] The specific methods for processing and extracting feature vectors from the source data in step S1 include: Business Specification Features The business text is parsed using Natural Language Processing (NLP) and converted into semantic features for the model, which are then used as the query vector for the attention mechanism. Interactive audio and video features Speech recognition transcribes recorded text, semantic analysis extracts customer's spoken business type, transaction amount, and business intent, and generates audio and video semantic features; Voucher Image Features Visual features of seals, signatures, and document anti-counterfeiting points are extracted through image recognition and convolutional visual encoder and quantized to serve as key vectors and value vectors for the attention mechanism. Historical experience and risk characteristics Complete missing information related to historical transaction records, correct manually reviewed and labeled samples, and extract characteristics of historical risk events; The system inputs features in real time: standardized conversion of real-time input JSON files and extracted data from the database generates real-time transaction features.

[0022] Step S2, the multidimensional feature fusion and attention mechanism initialization steps, include: Will As a query, , Using these as Key and Value respectively, the semantic similarity between Query and Key is calculated through dot product, generating a visual representation guided by the semantics of business specifications, and locating the verification area in the voucher image that matches the business rules.

[0023] Step S2, the verification objects for cross-modal alignment logic training, are three types of semantic data: ... Extracted customer voice service intent, by Extracted voucher text information, by The system records transaction data in real time; when the semantic space values ​​of the three are aligned, the business is marked as compliant; when the semantic difference exceeds a preset threshold, the business is marked as abnormal.

[0024] Step S2, embedding the business knowledge graph, includes: embedding business specifications... The business knowledge graph and risk control rules are embedded into the semantic space of the BERT model. When the model outputs the review conclusion, it binds the corresponding business specification clauses to make the review results interpretable.

[0025] The dynamic continuous learning mechanism in step S2 is as follows: collect new business cases at the counter, manually review newly labeled samples, and new risk event data, continuously fine-tune the BERT review model parameters, dynamically update the risk identification and judgment criteria, and realize the model's autonomous iterative optimization.

[0026] Step S3, risk-driven dynamic priority scheduling, specifically involves: based on , For current transactions, a risk profile is created, and high-risk transactions such as large cash withdrawals, high-frequency abnormal transactions, and transactions that match historical violation characteristics are assigned a higher review priority, with high-risk transactions being reviewed intelligently first.

[0027] Step S3, multi-task parallel computing and anomaly identification logic, involves parallel execution of feature vector transformation for multiple transactions and model inference; when , , The semantic alignment score is below a preset threshold, or the visual features of the voucher signature or seal do not match. When implementing business procedures, abnormal transactions are identified and corresponding risk tags are assigned.

[0028] Step S3, the method for generating the interpretable evidence chain, is as follows: map the model attention weights to the original data source, associate them with the corresponding business specification clauses, voucher image verification areas, and audio text fragments, retain the complete review and judgment basis, and output them synchronously to the review feedback module.

[0029] Example 2

[0030] Example 2 introduces hardware deployment and business implementation scenarios (large-amount cash withdrawal over-the-counter transactions).

[0031] Hardware deployment environment

[0032] Front-end counter terminals: standard bank counter PC, high-definition recording equipment, and high-definition business document scanner; Edge front-end server: 8-core CPU, 32GB memory, GPU inference accelerator card, used for local deployment of lightweight BERT models; Backend storage server: Distributed database, storing business specifications, audio and video files, voucher images, and historical transaction risk control records; Internal network transmission link: Gigabit dedicated line for the bank's internal network, ensuring real-time transmission of multi-source data without delay.

[0033] Business Implementation Scenarios (Large Amount Cash Withdrawal Over-the-Counter Transactions)

[0034] 1. Data Acquisition Phase

[0035] Generate data based on the current period's large withdrawal business standard that "withdrawals exceeding 50,000 yuan require the customer's signature for confirmation". ; Collect audio recordings of the teller and customer (customer verbally stating a withdrawal of 50,000 yuan), and generate counter surveillance video. Scan the withdrawal receipt, use OCR to recognize the handwritten amount of 50,000 and the blank signature field, extract the visual features of the seal and signature to generate the document. ; Retrieve the customer's transaction history for the past three months and past large withdrawal verification records to generate The system collects real-time data such as transaction amount of 50,000 yuan and customer account number entered by the teller to generate the data. The five types of vectors are fused to obtain a unified high-dimensional feature vector.

[0036] 2. Model Inference Stage

[0037] The BERT model uses large withdrawal rules For the query, retrieve voucher images. In the signature area, cross-modal comparison shows that the semantics of 50,000 voice records, 50,000 vouchers, and 50,000 system entries are consistent; the blank signature field of the voucher is identified, and a violation is determined by matching it with the business specification vector; this transaction belongs to a high-risk scenario of large cash withdrawal, and is assigned the highest review priority for parallel inference.

[0038] 3. Results Output

[0039] The model outputs a review conclusion of "failed" and a risk label of "large withdrawal lacking customer signature". The evidence chain includes: business specification clauses, audio transcript, and screenshot of the blank signature area of ​​the voucher. The data is pushed to the post-supervision review terminal. The reviewer only needs to confirm the abnormality and return the transaction to the counter for signature. Other small normal withdrawal transactions are automatically reviewed and approved without manual intervention.

[0040] Implementation effect data

[0041] In a scenario with 1,000 counter transactions per day, traditional manual review requires two dedicated reviewers to complete the task in 8 hours. The intelligent review module of this invention can process 100 concurrent transactions simultaneously on a single edge server. 95% of regular transactions are automatically reviewed and approved, while only 50 complex and abnormal transactions require manual processing, reducing the manpower required for review by 75%. The review time for a single transaction is reduced from an average of 90 seconds by manual review to 0.5 seconds by model inference, significantly improving the efficiency of counter business processing.

Claims

1. A verification method for a financial counter service processing system, characterized in that, Includes the following steps: S1. Data Acquisition and Standardization Processing Stage: The data acquisition module collects business specification data from multiple sources, including counter interaction audio and video recordings, voucher document image data, historical records and risk control data, and real-time system data. Feature vectors are extracted from each type of source data and then concatenated and fused to obtain a unified high-dimensional feature vector. ,in ; For business specification feature vectors, For interactive audio and video feature vectors, The feature vector of the voucher image. For historical experience feature vectors, This represents the feature vector of the risk pattern. S2. Core Model Construction and Optimization Stage: Based on the BERT large model as the basic architecture, multi-dimensional feature fusion and attention mechanism initialization, cross-modal alignment logic training, business knowledge graph embedding, iterative optimization based on historical risk data are completed based on the unified high-dimensional feature vector. Lightweight model deployment is completed through model pruning and edge computing, and a dynamic continuous learning mechanism is configured. S3. Real-time processing and judgment stage: The unified high-dimensional feature vector is input into the optimized BERT verification model to perform multi-dimensional semantic consistency judgment, risk-driven dynamic priority scheduling, multi-task parallel anomaly identification, and simultaneously generate an interpretable evidence chain. S4. Review and Feedback Output Stage: Output review conclusions, risk level labels, and visually interpretable analytical evidence to complete the intelligent review of counter services.

2. The verification method for a financial counter service processing system according to claim 1, characterized in that, The data collected from multiple sources in step S1 specifically includes: Business compliance data: The bank's latest business operation procedures, compliance requirements, and institutional guidelines; Counter interaction audio and video recording data: audio recordings of conversations between tellers and customers, and on-site video of the counter; Image data of voucher documents: images of historical business documents, customer certificates, and business vouchers; Historical records and risk control data: over-the-counter transaction records, manually reviewed and annotated data, historical abnormal events, and post-risk control supervision records; The system inputs data in real time: transaction elements such as account number, amount, and loan identifier entered by tellers in the business system, page restoration data, and real-time data extracted from the database.

3. The verification method for a financial counter service processing system according to claim 1, characterized in that, The specific methods for processing and extracting feature vectors from the source data in step S1 include: Business Specification Features The business text is parsed using Natural Language Processing (NLP) and converted into semantic features for the model, which are then used as the query vector for the attention mechanism. Interactive audio and video features Speech recognition transcribes recorded text, semantic analysis extracts customer's spoken business type, transaction amount, and business intent, and generates audio and video semantic features; Voucher Image Features Visual features of seals, signatures, and document anti-counterfeiting points are extracted through image recognition and convolutional visual encoder and quantized to serve as key vectors and value vectors for the attention mechanism. Historical experience and risk characteristics Complete missing information related to historical transaction records, correct manually reviewed and labeled samples, and extract characteristics of historical risk events; The system inputs features in real time: standardized conversion of real-time input JSON files and extracted data from the database generates real-time transaction features.

4. The verification method for a financial counter service processing system according to claim 1, characterized in that, Step S2, the multidimensional feature fusion and attention mechanism initialization steps, include: Will As a query, , Using these as Key and Value respectively, the semantic similarity between Query and Key is calculated through dot product, generating a visual representation guided by the semantics of business specifications, and locating the verification area in the voucher image that matches the business rules.

5. The verification method for a financial counter service processing system according to claim 1, characterized in that, Step S2, the verification objects for cross-modal alignment logic training, are three types of semantic data: ... Extracted customer voice service intent, by Extracted voucher text information, by The system records transaction data in real time; when the semantic space values ​​of the three are aligned, the business is marked as compliant; when the semantic difference exceeds a preset threshold, the business is marked as abnormal.

6. The verification method for a financial counter service processing system according to claim 1, characterized in that, Step S2, embedding the business knowledge graph, includes: embedding business specifications... The business knowledge graph and risk control rules are embedded into the semantic space of the BERT model. When the model outputs the review conclusion, it binds the corresponding business specification clauses to make the review results interpretable.

7. The verification method for a financial counter service processing system according to claim 1, characterized in that, The dynamic continuous learning mechanism in step S2 is as follows: collect new business cases at the counter, manually review newly labeled samples, and new risk event data, continuously fine-tune the BERT review model parameters, dynamically update the risk identification and judgment criteria, and realize the model's autonomous iterative optimization.

8. The verification method for a financial counter service processing system according to claim 1, characterized in that, Step S3, risk-driven dynamic priority scheduling, specifically involves: based on , For current transactions, a risk profile is created, and high-risk transactions such as large cash withdrawals, high-frequency abnormal transactions, and transactions that match historical violation characteristics are assigned a higher review priority, with high-risk transactions being reviewed intelligently first.

9. The verification method for a financial counter service processing system according to claim 1, characterized in that, Step S3, multi-task parallel computing and anomaly identification logic, involves parallel execution of feature vector transformation for multiple transactions and model inference; when , , The semantic alignment score is below a preset threshold, or the visual features of the voucher signature or seal do not match. When implementing business procedures, abnormal transactions are identified and corresponding risk tags are assigned.

10. The verification method for a financial counter service processing system according to claim 1, characterized in that, Step S3, the method for generating the interpretable evidence chain, is as follows: map the model attention weights to the original data source, associate them with the corresponding business specification clauses, voucher image verification areas, and audio text fragments, retain the complete review and judgment basis, and output them synchronously to the review feedback module.

Images