Merchant network access qualification auditing method, device and equipment and medium

Abstract

The invention provides a merchant network access qualification auditing method and device, equipment and a medium, and relates to the technical field of merchant network access. The method comprises the steps of obtaining merchant network access application data; wherein the merchant network access application data at least comprises structured field information and an unstructured qualification material file in a network access protocol; based on the structured field information, performing material content verification and material compliance verification on the unstructured qualification material file to obtain a binary verification result set; and based on the binary verification result set, generating a target auditing conclusion for the merchant network access application. According to the embodiment of the invention, the weight is dynamically configured for the binary verification result generated by automatic verification, and the target auditing conclusion is comprehensively generated, so that intelligent and integrated auditing of the multi-modal merchant qualification material is realized, and the automation level and decision accuracy of auditing processing are remarkably improved.

Description

Merchant onboarding qualification verification methods, devices, equipment and media Technical Field

[0001] This invention relates to the field of merchant onboarding technology, and in particular to a method, apparatus, equipment and medium for verifying merchant onboarding qualifications. Background Technology

[0002] In the merchant onboarding process, reviewing the various qualification materials submitted by merchants (such as business licenses, ID cards, storefront photos, etc.) is a crucial step. The purpose is to verify the authenticity, completeness, and consistency of the materials with the contract information.

[0003] Currently, the most common review methods rely heavily on manual operation. Reviewers need to manually compare the information on paper or electronic materials with the contract fields entered into the system, and rely on personal experience to judge whether the materials are compliant (such as whether they are stamped, whether the photos are authentic). This method is not only time-consuming and labor-intensive, but also prone to inconsistencies in review standards due to staff fatigue or differences in experience, leading to omissions or misjudgments and increasing business risks.

[0004] Existing auxiliary technologies, such as optical character recognition (OCR) for extracting text information, can alleviate the burden of data entry to some extent. However, they typically only digitize single pieces of information and do not form an automated review loop that covers both content and compliance perspectives and can comprehensively and intelligently judge multiple types and modal materials. The generation of review conclusions still heavily relies on manual subjective weighting and final decision-making of all scattered verification results, failing to fundamentally solve the efficiency bottlenecks and accuracy challenges of the review process. Summary of the Invention

[0005] This invention provides a method, apparatus, equipment, and medium for merchant onboarding qualification review. By dynamically configuring weights for the binary verification results generated by automated verification and comprehensively generating the target review conclusion, it realizes intelligent and integrated review of multimodal merchant qualification materials, significantly improving the automation level and decision accuracy of the review process.

[0006] In a first aspect, the present invention provides a method for verifying merchant network access qualifications, comprising the following steps: obtaining merchant network access application data; wherein the merchant network access application data includes at least structured field information and unstructured qualification material documents in the network access agreement; based on the structured field information, verifying the content and compliance of the unstructured qualification material documents to obtain a binary verification result set; and generating a target review conclusion for the merchant network access application based on the binary verification result set; wherein each verification result in the binary verification result set is configured with a dynamic weight coefficient, the dynamic weight coefficient being dynamically calculated based on the importance of the material type and the verification confidence level of the unstructured qualification material documents.

[0007] Preferably, according to the merchant access qualification verification method provided by the present invention, the step of verifying the content and compliance of the unstructured qualification materials based on the structured field information to obtain a binary verification result set includes: verifying the content of the unstructured qualification materials based on the structured field information to obtain a content verification result; verifying the compliance of the unstructured qualification materials based on the structured field information to obtain a compliance verification result; and combining the content verification result and the compliance verification result to obtain the binary verification result set.

[0008] Preferably, according to the merchant access qualification verification method provided by the present invention, the unstructured qualification material document includes at least text-based material images; the step of verifying the content of the unstructured qualification material document to obtain the content verification result includes: using a scene text detection model based on differentiable binarization to perform pixel-level text region segmentation on the text-based material image, generating text region bounding boxes for each text line; using a sequence recognition model based on a combination of convolutional neural networks and recurrent neural networks to perform feature sequence extraction processing on the image region within each text region bounding box, outputting an initial recognition text sequence; using a rule template or named entity recognition model pre-configured for different types of materials to structurally extract a first type of key information that matches the structured field information from the initial recognition text sequence; and matching the first type of key information with the structured field information to obtain the content verification result.

[0009] Preferably, according to the merchant access qualification verification method provided by the present invention, the unstructured qualification material document includes at least a first image and a second image; the step of verifying the compliance of the unstructured qualification material document to obtain a material compliance verification result includes: performing authenticity verification processing on the first image to obtain an image authenticity verification result; performing identity verification processing on the second image to obtain an identity verification result; and determining the material compliance verification result based on the image authenticity verification result and the identity verification result.

[0010] Preferably, according to the merchant access qualification verification method provided by the present invention, the step of performing authenticity verification processing on the first image to obtain the image authenticity verification result includes: performing multi-scale feature extraction processing on the first image to obtain depth feature maps at different convolutional levels; inputting the depth feature maps into a deep learning classification model trained by a joint loss function to output the image authenticity probability; determining whether the first image is a forged image based on the image authenticity probability, and generating the corresponding image authenticity verification result; wherein, the joint loss function is composed of a weighted sum of classification loss and triplet loss, the classification loss is used to classify the first image and the forged image, and the triplet loss is used to narrow the feature distance between real images of the same type and widen the feature distance between the real image and the forged image.

[0011] Preferably, according to the merchant on-network qualification verification method provided by the present invention, the step of performing identity verification processing on the second image to obtain an identity verification result includes: performing face detection on the second image and cropping out the face region image containing the face; using a pre-trained face feature encoder, encoding each detected face region image into a fixed-dimensional face feature vector; calculating the cosine similarity between the face feature vector to be verified and the baseline feature vector pre-stored in the authorized personnel feature library; if the cosine similarity is greater than or equal to a preset similarity threshold, then the image identity verification is determined to be passed, and the identity verification result containing the matching personnel identifier is generated; if the cosine similarity is less than the preset similarity threshold, then the image identity verification is determined to be failed, and the identity verification result that failed verification is generated.

[0012] Preferably, according to the merchant onboarding qualification verification method provided by the present invention, the step of generating a target verification conclusion for the merchant onboarding application based on the binary verification result set includes: obtaining a corresponding first weight coefficient according to the predefined type of the unstructured qualification material document associated with each verification result in the binary verification result set; calculating a second weight coefficient through a preset mapping function according to the confidence score output by the corresponding verification model that generates each verification result; multiplying the first weight coefficient by the second weight coefficient to obtain the dynamic weight coefficient used for that verification result; weighting the binary verification result set using the dynamic weight coefficient, and inputting the weighted binary verification result set into a preset verification rule engine for logical operation to generate the target verification conclusion.

[0013] Secondly, the present invention also provides a merchant network access qualification verification device, comprising the following modules: an acquisition module, used to acquire merchant network access application data; wherein the merchant network access application data includes at least structured field information and unstructured qualification material documents in the network access agreement; a verification module, used to perform material content verification and material compliance verification on the unstructured qualification material documents based on the structured field information, to obtain a binary verification result set; and a generation module, used to generate a target review conclusion for the merchant network access application based on the binary verification result set; wherein each verification result in the binary verification result set is configured with a dynamic weight coefficient, the dynamic weight coefficient being dynamically calculated based on the material type importance and verification confidence level of the unstructured qualification material documents.

[0014] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the merchant access qualification verification method described above.

[0015] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the merchant access qualification verification method as described above.

[0016] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the merchant access qualification verification method described above.

[0017] This invention provides a method, apparatus, device, and medium for merchant onboarding qualification verification. It acquires merchant onboarding application data, which includes at least structured field information and unstructured qualification material documents from the onboarding agreement. Based on the structured field information, it verifies the content and compliance of the unstructured qualification material documents, obtaining a binary verification result set. Based on the binary verification result set, it generates a target verification conclusion for the merchant onboarding application. Each verification result in the binary verification result set is configured with a dynamic weight coefficient, which is dynamically calculated based on the importance of the material type and the verification confidence level of the unstructured qualification material documents. By dynamically configuring weights for the binary verification results generated by automated verification and comprehensively generating the target verification conclusion, it achieves intelligent and integrated verification of multimodal merchant qualification materials, significantly improving the automation level and decision accuracy of the verification process. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 is one of the flowcharts of the merchant on-network qualification verification method provided by the present invention.

[0020] Figure 2 is a second flowchart of the merchant on-network qualification verification method provided by the present invention.

[0021] Figure 3 is a structural schematic diagram of the merchant access qualification verification device provided by the present invention.

[0022] Figure 4 is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] All actions involving the acquisition of signal information or data in this invention are carried out in compliance with the relevant data protection laws and policies of the country where the device is located, and with the authorization granted by the owner of the device.

[0025] The following description, in conjunction with Figures 1-4, describes a merchant access qualification review method, apparatus, equipment, and medium of the present invention. By dynamically configuring weights for the binary verification results generated by automated verification and comprehensively generating the target review conclusion, intelligent and integrated review of multimodal merchant qualification materials is achieved, significantly improving the automation level and decision accuracy of the review process.

[0026] Figure 1 is a flowchart illustrating one aspect of a merchant on-network qualification verification method provided by the present invention. As shown in Figure 1, the method may include, but is not limited to, steps S100 to S300: S100, obtaining merchant on-network application data; wherein, the merchant on-network application data includes at least structured field information and unstructured qualification material documents in the on-network agreement; S200, based on the structured field information, verifying the content and compliance of the unstructured qualification material documents to obtain a binary verification result set; S300, based on the binary verification result set, generating a target verification conclusion for the merchant on-network application; wherein, each verification result in the binary verification result set is configured with a dynamic weight coefficient, and the dynamic weight coefficient is dynamically calculated based on the importance of the material type and the verification confidence level of the unstructured qualification material documents.

[0027] Referring to Figures 1 and 2, the execution entity of the merchant onboarding qualification verification method provided in this embodiment is typically deployed on a server or cloud computing platform. This merchant onboarding qualification verification system interacts with the merchant signing terminal (such as the mobile application "Yinshangpai" used by account managers), the internal review workflow platform, and the merchant file management system. The overall method follows a closed-loop logic of "data collection -> intelligent verification -> weighted decision -> process-driven".

[0028] Specifically, in step S100 of some embodiments, merchant network access application data is obtained; wherein, the merchant network access application data includes at least structured field information and unstructured qualification material files in the network access agreement.

[0029] Specifically, after the merchant signing process is completed, the merchant's onboarding application data is pushed to the intelligent review system. This data package includes two core parts: 1) Structured field information, which is the standardized data extracted from the electronic agreement and filled into the form, such as the merchant name, unified social credit code, legal representative's name, business scope, etc.; 2) Unstructured qualification materials, which are various uploaded images or scans, such as a copy of the business license (stamped with the official seal), photos of the front and back of the legal representative's ID card, a photo of the storefront, and a photo of the account manager on site.

[0030] In step S200 of some embodiments, based on the structured field information, the unstructured qualification material document is subjected to material content verification and material compliance verification to obtain a binary verification result set.

[0031] Specifically, the system initiates two parallel verification tasks: material content verification and material compliance verification. This step calls upon multiple dedicated AI processing units to analyze unstructured qualification material documents and compares the analysis results with structured field information. Ultimately, it generates a verification result for each document that includes two dimensions: "content consistency" and "compliance." All results constitute a binary verification result set.

[0032] In step S300 of some embodiments, a target review conclusion for merchant network access application is generated based on the binary verification result set; wherein, each verification result in the binary verification result set is configured with a dynamic weight coefficient, which is dynamically calculated based on the material type importance and verification confidence of the unstructured qualification material document.

[0033] Understandably, the system doesn't simply count all verification results as "yes" or "no." Instead, it assigns a dynamic weighting coefficient to each item in the result set. This dynamic weighting coefficient comprehensively considers both the importance of the material itself and the reliability of the verification. The weighted binary verification result set is then fed into a pre-defined audit rule engine for comprehensive logical judgment, outputting the final target audit conclusion, such as "automatic pass," "automatic rejection," or "transfer to manual review."

[0034] In some embodiments, based on the target review conclusion, the system automatically performs corresponding operations in the business process, such as pushing to the next stage, returning for modification, or creating a manual review task, thereby achieving intelligent closed-loop management of the review process.

[0035] This method, through the collaborative processing of structured and unstructured data and combined with a dynamically weighted intelligent decision-making mechanism, has transformed merchant qualification review from being driven by human experience to being driven by data and algorithms, significantly improving review efficiency, accuracy, and standardization.

[0036] In some embodiments of the present invention, the step of verifying the content and compliance of the unstructured qualification material documents based on the structured field information to obtain a binary verification result set includes: verifying the content of the unstructured qualification material documents based on the structured field information to obtain a content verification result; verifying the compliance of the unstructured qualification material documents based on the structured field information to obtain a compliance verification result; and combining the content verification result and the compliance verification result to obtain the binary verification result set.

[0037] Understandably, the specific steps of the material content verification sub-process are as follows: the system traverses all unstructured qualification material files, and for materials identified as text (such as business licenses and ID cards), it initiates the OCR recognition and information extraction process, accurately matches the extracted text information with the corresponding fields in the agreement, and generates material content verification results such as "Business License Name: Consistent" and "Legal Person ID Number: Inconsistent".

[0038] The steps of the material compliance verification sub-process are as follows: For materials that require verification of formal requirements and authenticity, the system initiates an image analysis and model judgment process. For example, it checks whether the business license copy has an official seal, whether the storefront photo is suspected of being forged, and whether the person in the group photo is the account manager, generating material compliance verification results such as "Official Seal Validity: Compliant", "Image Authenticity: Questionable", and "Account Manager Identity: Verified".

[0039] The system associates and encapsulates the two types of results mentioned above. For example, for a "copy of a business license," the final record in the result set might be: {Material ID: 'Biz_License_01', Content Verification: {'Credit Code': 'Consistent'}, Compliance Verification: {'Official Seal Validity': 'Compliant'}}. All such records for all materials together constitute a complete binary verification result set.

[0040] This embodiment breaks down the verification dimensions into "content" and "compliance," making the review logic clearer, facilitating targeted optimization of the technology stack for different verification paths, and providing a structured foundation for subsequent issue localization and weight configuration.

[0041] In some embodiments of the present invention, the unstructured qualification material file includes at least text-based material images; the step of verifying the content of the unstructured qualification material file to obtain the content verification result includes: using a scene text detection model based on differentiable binarization to perform pixel-level text region segmentation on the text-based material image, generating text region bounding boxes for each text line; using a sequence recognition model based on a combination of convolutional neural networks and recurrent neural networks to perform feature sequence extraction processing on the image region within each text region bounding box, outputting an initial recognition text sequence; using a rule template or named entity recognition model pre-configured for different types of materials to structurally extract a first type of key information that matches the structured field information from the initial recognition text sequence; and matching the first type of key information with the structured field information to obtain the content verification result.

[0042] Understandably, the steps for text region detection using a scene text detection model are as follows: The system receives a scanned image of a business license. First, a scene text detection model based on differentiable binarization (such as DBNet) is used for processing. This scene text detection model transforms the text detection task into a probability prediction problem of whether each pixel is text through a differentiable binarization process. It can locate text lines of arbitrary shapes (horizontal, tilted, curved) in the image with extremely high accuracy, outputting text region bounding boxes that closely fit the text regions. For example, it can accurately locate the text regions of labels such as "Unified Social Credit Code," "Name," and "Legal Representative" and their corresponding values ​​in the image.

[0043] The text sequence recognition step involves, for each image slice within a bounding box obtained in the previous step, invoking a sequence recognition model based on a combination of convolutional neural networks and recurrent neural networks (CNN+RNN model). The convolutional neural network (CNN) extracts visual feature sequences from the image slices; the recurrent neural network (RNN, commonly LSTM or GRU) models the context of these feature sequences, understanding the dependencies between characters; finally, a connection to a connectionist temporal classification (CTC) decoding layer maps the feature sequences to the final initial recognized text sequence. For example, the area following "legal representative" is identified as "Zhang Moumou".

[0044] As can be understood, a sequence recognition model is an end-to-end deep neural network, whose structure includes, in sequence, a convolutional feature extraction module, a sequence modeling module, and a sequence transcription module. The following example, using the recognition of an image slice containing the text "9123010MA1BUFLG4X", illustrates the data flow process.

[0045] 1. Structure of the Convolutional Feature Extraction Module (corresponding to the CNN part): This module consists of multiple convolutional layers, batch normalization layers, and non-linear activation layers (such as ReLU) stacked alternately, usually connected to a pooling layer at the end. Its function is to convert the input text region image (such as a grayscale image normalized to 32x256 pixels) into a visual feature sequence.

[0046] The data processing flow specifically includes: Input: A text region image (single-channel grayscale image) with dimensions [H, W, 1].

[0047] Feature transformation: After the image undergoes multiple convolution operations, it outputs a three-dimensional feature tensor with dimensions [H', W', C]. Here, H' is typically 1 (because the height of the text region is compressed to 1 after pooling), W' corresponds to the length of the feature sequence, and C is the dimension (number of channels) of the feature vector.

[0048] Output: This module treats each column of features in the width direction (W' dimension) of the image as input for one time step, and finally outputs a visual feature sequence V = {v1, v2, …, vW'}, where each vi is a C-dimensional vector. For the example image, W' may be 40, meaning a feature sequence of length 40 is generated, where each vector contains local and global visual information of the image at the corresponding horizontal position.

[0049] 2. Structure of the Sequence Modeling Module (corresponding to the RNN part): This module is typically composed of a Bidirectional Long Short-Term Memory (Bi-LSTM) network or a Gated Recurrent Unit (GRU). Its function is to capture the contextual dependencies between various positions in the feature sequence V. For example, when recognizing an ID card number, the recognition of the current character needs to refer to the contextual information of the characters before and after it.

[0050] The data processing flow specifically includes: Input: Visual feature sequence V.

[0051] Context encoding: Bi-LSTM processes the sequence simultaneously in both forward and backward directions. At each time step, it combines the current feature vi with the hidden state of the previous time step to output a context feature vector that incorporates contextual information.

[0052] Output: After processing through multiple Bi-LSTM layers, an enhanced feature sequence H = {h1, h2, …,hW'} is output, which has the same length as the input sequence V (W'), but the dimension of each hi may change (e.g., become 256 dimensions), which contains richer sequence context information.

[0053] 3. Structure of the Sequence Transcription Module (corresponding to the CTC decoding layer): The core of this module is the Connectionist Temporal Classification (CTC) layer. CTC is a loss function and decoding algorithm that allows the model to be trained without requiring strict alignment between the input and output sequences.

[0054] The data processing flow (training / inference difference) specifically includes: Input: Context feature sequence H.

[0055] Linear transformation: First, a fully connected layer maps each hi to a vector of dimension (K+1), where K is the size of the character table (e.g., numbers 0-9, letters A, B, and C, a total of 36 classes), and +1 represents a special whitespace character. Then, the softmax function is used to transform it into a probability distribution for each of the (K+1) classes at each time step, forming a probability matrix.

[0056] During training (CTC Loss calculation): Given an input sequence H and a true label sequence L (e.g., "9123010MA1BUFLG4X"), the CTC loss function uses a dynamic programming algorithm to calculate the sum of probabilities of all possible alignment paths (allowing duplicate characters and whitespace). Its core principle is to maximize the total probability of the correct label sequence without specifying the exact position of each character in the feature sequence.

[0057] During inference (CTC decoding): The probability matrix above is decoded. The most common method is greedy decoding: at each time step, the character with the highest probability is selected, then duplicate characters are merged and whitespace is removed, finally obtaining the initial recognized text sequence. For example, the model might output "99-1-22-333-…-blank-…", which is decoded to "9123…". A better method is to use beam search decoding, which retains multiple paths with higher probabilities to obtain more accurate results.

[0058] Furthermore, the model training method is implemented as follows: 1. Data preparation: Collect massive amounts of text image data, including synthetic text images and text images in real-world scenes (such as images of certificates and documents under various fonts, backgrounds, and lighting conditions). Label each image to generate corresponding real character sequence labels. Perform data augmentation on the images, such as random rotation, scaling, adding noise, blurring, etc., to improve the model's robustness.

[0059] 2. Loss Function: The CTC Loss function is used as the optimization objective for model training. Its mathematical form is negative log-likelihood: Loss = -log(P(L|X)), where P(L|X) is the probability of the output label sequence L given the input features X (obtained from the image through the network), which is obtained by integrating the probability matrix using the CTC forward algorithm.

[0060] 3. Training Process: Initialization: Use pre-trained CNN weights (e.g., pre-trained on ImageNet) or randomly initialize all network parameters. Forward Propagation: Input a batch of text images, sequentially pass them through convolutional feature extraction, sequence modeling, and linear transformation to obtain the probability matrix. Loss Calculation: Call the CTC loss function to calculate the total loss value based on the probability matrix and the corresponding true label sequence of the batch. Backpropagation and Optimization: Utilize stochastic gradient descent (SGD) or its variants (e.g., the Adam optimizer) to calculate the gradient of the loss relative to the parameters of each layer of the model through the backpropagation algorithm, and update the parameters to minimize the CTC loss. Iteration: Repeat the above steps until the model's recognition accuracy on the validation set converges or reaches the preset number of iterations.

[0061] The CNN-BiLSTM-CTC model structure described in this embodiment effectively extracts visual features through CNN, models long-range contextual dependencies between characters through BiLSTM, and finally solves the challenging alignment problem in sequence recognition through the CTC mechanism. This training method enables the model to learn from a large amount of data, ultimately achieving high-precision and robust sequence recognition of text in complex scenes, with multiple fonts and different layouts, providing a solid technical foundation for the automated extraction of key text information from merchant qualification materials.

[0062] The specific steps of information structure extraction are as follows: The original identified text, i.e., the initial identified text sequence, is a continuous string. Based on the material type "business license," the system loads a pre-configured rule template. This rule template defines the positional relationships of key fields (such as "unified social credit code" and "name") in the text, keyword prefixes, and other rules, thereby accurately extracting the corresponding first-category key information from the initial identified text sequence. More advanced implementations use a Named Entity Recognition Model (NER) to automatically identify and classify entities in the text, such as "ORG" (organization) corresponding to the merchant name, and "PER" (person's name) corresponding to the legal representative's name.

[0063] The extracted "Unified Social Credit Code" value is compared with the contract credit code obtained from the structured field information. If they match completely, a "matched" content verification result is generated for that field; otherwise, it is marked as "inconsistent".

[0064] This embodiment achieves high-precision and robust recognition of complex-formatted document text by using the advanced OCR technology stack of DBNet and CNN+RNN+CTC; further, it uses rule-based or NER models for structuring to ensure the accuracy of extracted information, laying a reliable data foundation for subsequent automated comparison.

[0065] In some embodiments of the present invention, the unstructured qualification material document includes at least a first image and a second image; the step of verifying the compliance of the unstructured qualification material document to obtain a material compliance verification result includes: performing authenticity verification processing on the first image to obtain an image authenticity verification result; performing identity verification processing on the second image to obtain an identity verification result; and determining the material compliance verification result based on the image authenticity verification result and the identity verification result.

[0066] Specifically, in some embodiments of the present invention, the step of performing authenticity verification processing on the first image to obtain an image authenticity verification result includes: performing multi-scale feature extraction processing on the first image to obtain depth feature maps at different convolutional levels; inputting the depth feature maps into a deep learning classification model trained by a joint loss function to output the image authenticity probability; determining whether the first image is a fake image based on the image authenticity probability, and generating the corresponding image authenticity verification result; wherein, the joint loss function is composed of a weighted sum of classification loss and triplet loss, the classification loss is used to classify the first image and the fake image, and the triplet loss is used to narrow the feature distance between real images of the same type and widen the feature distance between the real image and the fake image.

[0067] Understandably, taking the anti-counterfeiting verification of a "storefront photo" (the first image) as an example: First, multi-scale feature extraction is performed, that is, the storefront photo is input into a deep convolutional neural network (such as ResNet). Different layers of the network capture image features of different granularities. Shallow layers capture low-level features such as edges and textures; deep layers capture high-level semantic features such as objects and scenes. The system obtains these deep feature maps at different convolutional levels, which together constitute a comprehensive description of the image.

[0068] The classification model trained using the joint loss function, i.e., the deep learning classification model used for the final judgment, is trained using the joint loss function: L_total = α. L_softmax + β L_triplet.

[0069] Where L_softmax is the classification loss, L_triplet is the triplet loss, and α and β are their respective loss weights.

[0070] Classification loss (such as Softmax loss): This is the loss that drives the model to learn to correctly classify an input image into a discrete category, such as "real" or "fake". It calculates the difference between the class probability distribution predicted by the model and the true label.

[0071] Triplet Loss: In the feature space, it brings together all features of "real storefront photos" while widening the gap between features of "real storefront photos" and various "fake storefront photos" (such as photoshopped, AI-generated, and stolen images from the internet). It requires the model to learn a feature embedding space with good intra-class aggregation and inter-class separation properties.

[0072] The joint loss function provided in this embodiment is a weighted sum of α and β, which simultaneously optimizes classification accuracy and feature discriminative power.

[0073] During inference, the multi-scale features extracted in the above steps (i.e., the first image) are input into this trained model, and the model outputs an image authenticity probability, such as "authentic probability is 0.92". The system compares this probability with a preset risk threshold (e.g., 0.7). If the authenticity probability is higher than the threshold, the "image authenticity verification result" is generated as "authentic"; otherwise, it is marked as "suspected forgery" and manual review is triggered.

[0074] This embodiment employs a counterfeit detection model trained by integrating multi-scale features and joint loss, which can effectively detect various forgery methods such as photoshopped images, AI-generated images, and screenshots, greatly enhancing the review system's ability to verify the authenticity of materials.

[0075] In some embodiments of the present invention, the step of performing identity verification processing on the second image to obtain an identity verification result includes: performing face detection on the second image and cropping out a face region image containing a face; using a pre-trained face feature encoder to encode each detected face region image into a fixed-dimensional face feature vector; calculating the cosine similarity between the face feature vector to be verified and a baseline feature vector pre-stored in an authorized personnel feature library; if the cosine similarity is greater than or equal to a preset similarity threshold, the image identity verification is determined to be passed, and the identity verification result containing the matching personnel identifier is generated; if the cosine similarity is less than the preset similarity threshold, the image identity verification is determined to be failed, and the identity verification result that failed verification is generated.

[0076] Understandably, taking the identity verification of the "customer manager and storefront photo" (second image) as an example: This embodiment uses models such as Multi-task Convolutional Neural Network (MTCNN) to perform face detection on the photo, locate and crop out all face regions appearing in the photo.

[0077] The cropped face region image is input into a pre-trained face feature encoder (such as ResNet trained based on ArcFace loss). This encoder maps the face image into a fixed-dimensional face feature vector, such as a 512-dimensional vector, which represents the uniqueness of the face in the feature space.

[0078] Calculate the cosine similarity between each facial feature vector to be verified and all baseline feature vectors in the authorized personnel feature database. The cosine similarity value ranges from -1 to 1. The closer the value is to 1, the more consistent the directions of the two vectors are, i.e., the more similar the faces are.

[0079] Set a preset similarity threshold (e.g., 0.85). If the similarity between a vector to be verified and a baseline vector in the library is greater than or equal to this threshold, the image is determined to be the corresponding authorized account manager, generating an "Identity Verification Passed" result with the employee's employee number attached. If no match is found for any faces, "Identity Verification Failed" is determined.

[0080] This embodiment uses high-precision facial recognition technology to automatically verify the identity of on-site inspection personnel, ensuring the authenticity of the "personal visit and signature" business, effectively preventing risks such as impersonation, and the entire process does not require manual visual comparison.

[0081] All actions involving the acquisition / processing of facial information or data in this invention are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with authorization from the owner of the corresponding device.

[0082] In some embodiments of the present invention, generating a target review conclusion for a merchant's network access application based on the binary verification result set includes: obtaining a first weight coefficient according to a predefined type of the unstructured qualification material document associated with each verification result in the binary verification result set; calculating a second weight coefficient using a preset mapping function based on the confidence score output by the corresponding verification model that generates each verification result; multiplying the first weight coefficient by the second weight coefficient to obtain the dynamic weight coefficient used for that verification result; weighting the binary verification result set using the dynamic weight coefficient; and inputting the weighted binary verification result set into a preset review rule engine for logical operation to generate the target review conclusion.

[0083] Understandably, the first step is to calculate a primary weighting coefficient based on the importance of the material type; that is, the system maintains a material type importance mapping table. For example, the importance of a business license is defined as 1.0, the legal representative's ID card as 0.9, the storefront sign as 0.7, and the agent's ID card as 0.6. When the verification result is associated with a specific material, the corresponding primary weighting coefficient is directly retrieved from this table. This reflects the business logic that the verification results of core certificates have a greater impact on the final conclusion.

[0084] The second weighting coefficient is calculated based on the verification confidence level. This means that the AI ​​model includes a confidence score with the verification result. For example, the confidence level for OCR recognition of a "credit code" is 0.98, and the similarity for face comparison is 0.91. The system uses a pre-defined mapping function (such as the sigmoid function or a piecewise linear function) to convert this confidence score into a second weighting coefficient between 0.5 and 1.5. A higher confidence level results in a larger coefficient (upper limit 1.5), indicating a more reliable verification result; a lower confidence level results in a smaller coefficient (lower limit 0.5), indicating high uncertainty and a need to reduce its influence in decision-making.

[0085] The step of calculating the dynamic weight coefficient is to multiply the coefficients obtained in the first and second steps, i.e., Dynamic Weight Coefficient = First Weight Coefficient × Second Weight Coefficient. For example, a business license (importance 1.0) has a credit code verification confidence level of 0.98 (coefficient after mapping 1.45), and its dynamic weight is 1.0. 1.45 = 1.45. However, a poorly clear storefront photo (importance 0.7) has a confidence level of only 0.65 for authenticity verification (mapped coefficient 0.8), and its dynamic weight is only 0.7. 0.8 = 0.56.

[0086] The review rule engine receives a weighted binary verification result set. Its internal rules may take the form of: "If (Business License.Credit Code == Consistent AND Weight > 1.2) AND (Legal Representative ID Card.Name == Consistent) AND (Storefront Photo.Authenticity == Authentic OR Weight < 0.6), then the conclusion = Automatic Approval". The engine executes all similar rules, considering the Boolean value and weight of each condition, performing comprehensive logical operations, and finally outputting the target review conclusion.

[0087] This embodiment introduces dynamic weights based on business importance and model confidence, enabling the system to simulate the decision-making process of senior auditors. It gives greater importance to key and reliable verification results and reduces the weight of minor or uncertain results, thereby generating more accurate and reasonable automated audit conclusions and reducing reliance on manual review.

[0088] In this embodiment of the invention, the method further includes: Semantic Association Verification (MCC Verification): The system integrates a large language model to perform semantic understanding on the "business scope" text. The business scope description (e.g., "sales of daily necessities and stationery") and the merchant name are input into a finely tuned large language model (LLM) to obtain a semantic vector. This vector is then used to calculate semantic similarity with the description vectors of each category in the UnionPay MCC code library to recommend or verify whether the selected MCC code matches the protocol. This effectively avoids incorrect MCC selection due to broad or ambiguous business scope descriptions, representing a semantic-level deepening of "material content verification."

[0089] The feature extraction and verification of semantic modalities are specifically applied to the correlation verification between merchant category codes and business scopes, including: obtaining a textual description of the merchant's business scope from the structured contract field information; inputting the textual description of the business scope into a large language model fine-tuned by domain text to obtain its last hidden state as a semantic feature representation; calculating the semantic relevance score between the semantic feature representation and multiple standard merchant category code node vectors in a pre-built knowledge graph, wherein the knowledge graph contains the correlation relationships of MCC codes, industry classifications, and typical business descriptions; and verifying the correlation degree of the selected merchant category codes in the merchant's network access application based on the ranking of the semantic relevance scores. If the correlation degree is lower than a preset threshold, a semantic inconsistency judgment is generated.

[0090] Understandably, the system extracts the text description corresponding to the "Business Scope" field from the structured contract field information of the merchant's network access application data. For example, the obtained text might be: "Retail sales of daily necessities, office stationery, and sporting goods; repair of household appliances." (This is the text description of the business scope).

[0091] Semantic features are generated by fine-tuning a large language model, which involves inputting the text description obtained in the previous step into a large language model that has been fine-tuned by the domain text.

[0092] The model's underlying architecture is a Transformer encoder. To enable it to more accurately understand terms and semantic relationships in fields such as commerce, finance, and retail services, we performed supervised fine-tuning on the model using a domain-specific text corpus containing a large amount of merchant registration information, industry classification descriptions, and product and service directories. Fine-tuning tasks were typically designed as next-sentence prediction, text classification, or masked language modeling to enhance its domain semantic representation capabilities.

[0093] After processing the input text, the model extracts the output vector corresponding to the [CLS] tag (or sequence start tag) in the last hidden state. This output vector is an aggregation and condensation of the semantics of the entire input sequence, constituting a semantic feature representation that represents the essence of the merchant's business operations. It is usually a high-dimensional floating-point vector (e.g., 768-dimensional).

[0094] Furthermore, semantic relevance is calculated based on the output vector and the knowledge graph node vectors. That is, the system maintains a pre-built knowledge graph. This graph organizes knowledge in the form of nodes and relationships, where key node types include: MCC code nodes: attributes include the MCC code (e.g., "5411", "5732") and its official category name.

[0095] Industry classification nodes: such as "Retail" and "Service".

[0096] Typical business description node: contains natural language description text of the business covered by the MCC (such as the description of MCC 5411: "large grocery store, supermarket").

[0097] Edges in a knowledge graph represent relationships, such as "belongs to" or "contains". During graph construction or service startup, the system uses a finely tuned large language model to encode the official names corresponding to all standard Merchant Category Codes (MCCs) and their associated typical business descriptions into vectors, which are then attached as attributes to the corresponding MCC nodes, forming MCC node vectors. In this step, the system calculates the semantic relevance score between the merchant's semantic feature representation and all relevant MCC node vectors in the knowledge graph. The calculation typically uses a cosine similarity algorithm, which assesses directional similarity by measuring the cosine of the angle between two vectors in space, with a value range of [-1, 1]. Higher values ​​indicate greater semantic similarity.

[0098] Furthermore, the system verifies and generates a judgment based on relevance ranking, that is, the system sorts all the calculated semantic relevance scores in descending order. At the same time, it obtains the selected merchant category codes (such as the selected MCC "5999") selected by the account manager or merchant in the merchant's network access application, as well as their positions and scores in the ranking.

[0099] Furthermore, the system checks whether the semantic relevance score between the node vector corresponding to the selected merchant category code and the merchant's semantic feature representation is higher than a preset threshold (e.g., the score must be greater than 0.7), and whether its ranking is among the top (e.g., ranked in the top 3). If the conditions are met, the MCC selection is deemed reasonable, and the verification passes.

[0100] If the relevance score is lower than a preset threshold (e.g., a score below 0.7 or a ranking outside the top 5), the system generates a semantic inconsistency judgment. This judgment can be included as an independent "semantic verification" result in the "material content verification result" set, and its status may be marked as "MCC suspected mismatch," along with a list of the top 3 MCC suggestions recommended by the system based on the highest relevance score.

[0101] This embodiment transforms unstructured business scope text into deep semantic features and performs vectorized matching with a structured industry knowledge graph, achieving intelligent verification based on semantic understanding that surpasses keyword matching. It effectively addresses the issues of diverse and colloquial descriptions of "business scope," automatically and accurately recommending or verifying the MCC code that best matches the merchant's actual business operations. This significantly reduces merchant classification risks and subsequent transaction disputes caused by incorrect MCC selection, improving the intelligence and accuracy of the verification process.

[0102] Furthermore, the system generates immutable logs with timestamps, operator (system or human), input data fingerprints, and output results for each verification, weight calculation, rule engine trigger, and manual review operation. All logs are linked in a chain with the original application materials and intermediate verification data and stored in a dedicated database. This makes the entire review process completely transparent, auditable, and reproducible, meeting compliance requirements.

[0103] The system monitors the discrepancies between the final human decision and the system's initial conclusion in cases where the review process is transferred to human review. These cases, along with complete chains of evidence, are automatically collected to form an incremental training dataset. This dataset is used to periodically train or fine-tune models such as OCR, anti-counterfeiting, and facial recognition online, enabling the entire intelligent review system to continuously optimize and adapt to changes in business needs.

[0104] The merchant access qualification verification device provided by the present invention is described below. The merchant access qualification verification device described below and the merchant access qualification verification method described above can be referred to in correspondence.

[0105] Figure 3 shows a schematic diagram of the merchant access qualification verification device provided by the present invention. The device includes the following modules: an acquisition module 310, used to acquire merchant access application data; wherein the merchant access application data includes at least structured field information and unstructured qualification material documents in the access agreement; a verification module 320, used to perform material content verification and material compliance verification on the unstructured qualification material documents based on the structured field information, obtaining a binary verification result set; and a generation module 330, used to generate a target verification conclusion for the merchant access application based on the binary verification result set; wherein each verification result in the binary verification result set is configured with a dynamic weight coefficient, the dynamic weight coefficient being dynamically calculated based on the importance of the material type and the verification confidence level of the unstructured qualification material documents.

[0106] Preferably, the merchant access qualification verification device provided by the present invention is specifically used to verify the content of the unstructured qualification material document based on the structured field information to obtain a content verification result; to verify the compliance of the unstructured qualification material document based on the structured field information to obtain a compliance verification result; and to combine the content verification result and the compliance verification result to obtain the binary verification result set.

[0107] Preferably, the merchant access qualification verification device provided by the present invention is specifically used for the unstructured qualification material document to include at least text-based material images; employing a scene text detection model based on differentiable binarization to perform pixel-level text region segmentation on the text-based material images, generating text region bounding boxes for each text line; utilizing a sequence recognition model based on a combination of convolutional neural networks and recurrent neural networks to perform feature sequence extraction processing on the image region within each text region bounding box, outputting an initial recognition text sequence; using a rule template or named entity recognition model pre-configured for different types of materials, structurally extracting a first type of key information matching the structured field information from the initial recognition text sequence; matching the first type of key information with the structured field information to obtain the material content verification result.

[0108] Preferably, the merchant access qualification verification device provided by the present invention is specifically used for the unstructured qualification material document to include at least a first image and a second image; to perform authenticity verification processing on the first image to obtain an image authenticity verification result; to perform identity verification processing on the second image to obtain an identity verification result; and to determine the material compliance verification result based on the image authenticity verification result and the identity verification result.

[0109] Preferably, the merchant access qualification verification device provided by the present invention is specifically used to perform multi-scale feature extraction processing on the first image to obtain depth feature maps at different convolutional levels; input the depth feature maps into a deep learning classification model trained by a joint loss function, and output the image authenticity probability; determine whether the first image is a fake image based on the image authenticity probability, and generate the corresponding image authenticity verification result; wherein, the joint loss function is composed of a weighted sum of classification loss and triplet loss, the classification loss is used to classify the first image and the fake image, and the triplet loss is used to narrow the feature distance between real images of the same type and widen the feature distance between the real image and the fake image.

[0110] Preferably, the merchant access qualification verification device provided by the present invention is specifically used to perform face detection on the second image and crop out the face region image containing the face; using a pre-trained face feature encoder, each detected face region image is encoded into a fixed-dimensional face feature vector; the cosine similarity between the face feature vector to be verified and the baseline feature vector pre-stored in the authorized personnel feature library is calculated; if the cosine similarity is greater than or equal to a preset similarity threshold, the image identity verification is determined to be passed, and the identity verification result containing the matching personnel identifier is generated; if the cosine similarity is less than the preset similarity threshold, the image identity verification is determined to be failed, and the identity verification result that failed verification is generated.

[0111] Preferably, the merchant access qualification verification device provided by the present invention is specifically used to obtain a first weight coefficient according to the predefined type of the unstructured qualification material file associated with each verification result in the binary verification result set; calculate a second weight coefficient through a preset mapping function according to the confidence score output by the corresponding verification model that generates each verification result; multiply the first weight coefficient by the second weight coefficient to obtain the dynamic weight coefficient used for the verification result; perform weighted processing on the binary verification result set using the dynamic weight coefficient, and input the weighted binary verification result set into a preset review rule engine for logical operation to generate the target review conclusion.

[0112] Figure 4 illustrates a schematic diagram of the physical structure of an electronic device. As shown in Figure 4, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440. The processor 410, communication interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a merchant network access qualification review method. This method includes: acquiring merchant network access application data; wherein the merchant network access application data includes at least structured field information and unstructured qualification material documents in the network access agreement; based on the structured field information, verifying the material content and compliance of the unstructured qualification material documents to obtain a binary verification result set; and generating a target review conclusion for the merchant network access application based on the binary verification result set. Each verification result in the binary verification result set is configured with a dynamic weight coefficient, which is dynamically calculated based on the importance of the material type and the verification confidence level of the unstructured qualification material documents.

[0113] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0114] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the merchant access qualification review method provided by the above methods. The method includes: acquiring merchant access application data; wherein the merchant access application data includes at least structured field information and unstructured qualification material files in the access agreement; based on the structured field information, performing material content verification and material compliance verification on the unstructured qualification material files to obtain a binary verification result set; and generating a target review conclusion for the merchant access application based on the binary verification result set; wherein each verification result in the binary verification result set is configured with a dynamic weight coefficient, and the dynamic weight coefficient is dynamically calculated based on the material type importance and verification confidence of the unstructured qualification material files.

[0115] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the merchant access qualification verification method provided by the above methods. This method includes: acquiring merchant access application data; wherein the merchant access application data includes at least structured field information and unstructured qualification material files in the access agreement; based on the structured field information, performing material content verification and material compliance verification on the unstructured qualification material files to obtain a binary verification result set; and based on the binary verification result set, generating a target verification conclusion for the merchant access application; wherein each verification result in the binary verification result set is configured with a dynamic weight coefficient, the dynamic weight coefficient being dynamically calculated based on the material type importance and verification confidence level of the unstructured qualification material files.

[0116] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for verifying the qualifications of merchants joining the network, characterized in that, include: Obtain merchant onboarding application data; wherein the merchant onboarding application data includes at least structured field information and unstructured qualification material documents in the onboarding agreement; based on the structured field information, verify the content and compliance of the unstructured qualification material documents to obtain a binary verification result set; based on the binary verification result set, generate a target review conclusion for the merchant onboarding application; wherein each verification result in the binary verification result set is configured with a dynamic weight coefficient, the dynamic weight coefficient being dynamically calculated based on the importance of the material type and the verification confidence level of the unstructured qualification material documents.

2. The merchant onboarding qualification verification method according to claim 1, characterized in that, The step of verifying the content and compliance of the unstructured qualification materials based on the structured field information to obtain a binary verification result set includes: verifying the content of the unstructured qualification materials based on the structured field information to obtain a content verification result; verifying the compliance of the unstructured qualification materials based on the structured field information to obtain a compliance verification result; and combining the content verification result and the compliance verification result to obtain the binary verification result set.

3. The merchant onboarding qualification verification method according to claim 2, characterized in that, The unstructured qualification material file includes at least text-based material images; the process of verifying the content of the unstructured qualification material file to obtain the content verification result includes: using a scene text detection model based on differentiable binarization to perform pixel-level text region segmentation on the text-based material images, generating text region bounding boxes for each text line; using a sequence recognition model based on a combination of convolutional neural networks and recurrent neural networks to perform feature sequence extraction processing on the image regions within each text region bounding box, outputting an initial recognition text sequence; using a rule template or named entity recognition model pre-configured for different types of materials to structurally extract a first type of key information that matches the structured field information from the initial recognition text sequence; and matching the first type of key information with the structured field information to obtain the content verification result.

4. The merchant onboarding qualification verification method according to claim 2, characterized in that, The unstructured qualification document includes at least a first image and a second image; the process of verifying the compliance of the unstructured qualification document to obtain a compliance verification result includes: performing authenticity verification processing on the first image to obtain an image authenticity verification result; performing identity verification processing on the second image to obtain an identity verification result; and determining the compliance verification result based on the image authenticity verification result and the identity verification result.

5. The merchant onboarding qualification verification method according to claim 4, characterized in that, The step of performing authenticity verification processing on the first image to obtain the image authenticity verification result includes: performing multi-scale feature extraction processing on the first image to obtain deep feature maps at different convolutional levels; inputting the deep feature maps into a deep learning classification model trained by a joint loss function to output the image authenticity probability; determining whether the first image is a fake image based on the image authenticity probability, and generating the corresponding image authenticity verification result; wherein, the joint loss function is composed of a weighted sum of classification loss and triplet loss, the classification loss is used to classify the first image and the fake image, and the triplet loss is used to narrow the feature distance between real images of the same type and widen the feature distance between the real image and the fake image.

6. The merchant onboarding qualification verification method according to claim 4, characterized in that, The step of performing identity verification processing on the second image to obtain an identity verification result includes: performing face detection on the second image and cropping out the face region image containing the face; using a pre-trained face feature encoder, encoding each detected face region image into a fixed-dimensional face feature vector; calculating the cosine similarity between the face feature vector to be verified and the baseline feature vector pre-stored in the authorized personnel feature library; if the cosine similarity is greater than or equal to a preset similarity threshold, the image identity verification is determined to be passed, and the identity verification result containing the matching personnel identifier is generated; if the cosine similarity is less than the preset similarity threshold, the image identity verification is determined to be failed, and the identity verification result that failed verification is generated.

7. The merchant onboarding qualification verification method according to any one of claims 1 to 6, characterized in that, The step of generating a target review conclusion for a merchant's onboarding application based on the binary verification result set includes: obtaining a first weight coefficient according to the predefined type of the unstructured qualification material document associated with each verification result in the binary verification result set; calculating a second weight coefficient using a pre-set mapping function based on the confidence score output by the corresponding verification model for each verification result; multiplying the first weight coefficient by the second weight coefficient to obtain the dynamic weight coefficient used for that verification result; weighting the binary verification result set using the dynamic weight coefficient; and inputting the weighted binary verification result set into a pre-set review rule engine for logical operation to generate the target review conclusion.

8. A merchant onboarding qualification verification device, characterized in that, include: The acquisition module is used to acquire merchant onboarding application data; wherein the merchant onboarding application data includes at least structured field information and unstructured qualification material documents in the onboarding agreement; the verification module is used to verify the content and compliance of the unstructured qualification material documents based on the structured field information, and obtain a binary verification result set; the generation module is used to generate a target review conclusion for the merchant onboarding application based on the binary verification result set; wherein each verification result in the binary verification result set is configured with a dynamic weight coefficient, and the dynamic weight coefficient is dynamically calculated based on the importance of the material type and the verification confidence level of the unstructured qualification material documents.

9. 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 merchant on-network qualification verification method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the merchant on-network qualification verification method as described in any one of claims 1 to 7.

Images