Credit trade background review methods, apparatuses, and computer program products

Abstract

The application discloses a credit trade background examination method and device and a computer program product, relates to the technical field of electric digital data processing, and the method comprises the following steps: type identification is performed on each page of an original file of a credit trade background to be examined to obtain an initial file type, and the original file is divided according to the initial file type to obtain file segments; a classifier is used to classify the file segments to obtain a first classification result, and a large language model is used to classify the file segments to obtain a second classification result; the target file type of each file segment is determined according to the first classification result and the second classification result; structured examination element information is extracted from the file segments according to the target file type, and the examination element information is examined based on a preset examination rule to generate an examination result of the original file. The application improves the accuracy of the examination result of the credit trade background.

Description

Technical Field

[0001] This application relates to the field of electronic digital data processing technology, and in particular to a method, equipment, and computer program product for reviewing credit and trade background. Background Technology

[0002] Credit trade background review is a crucial step for financial institutions to examine the transaction documents provided by borrowers before issuing loans. Transaction documents, such as contracts, invoices, orders, and waybills, are intended to prevent credit risks such as fraudulent trade and double financing.

[0003] The current review methods mainly rely on manual verification of documents page by page, comparison of subject information, calculation of amounts, and checking of date logic. This is not only labor-intensive, but also prone to omissions and misjudgments due to fatigue or lack of experience, resulting in low accuracy of credit and trade background reviews. Summary of the Invention

[0004] The main purpose of this application is to provide a method, equipment, and computer program product for credit and trade background review, which aims to solve the technical problem of low accuracy in credit and trade background review.

[0005] To achieve the above objectives, this application proposes a method for reviewing credit and trade background, the method comprising: Each page of the original document containing the credit and trade background to be examined is type-identified to obtain an initial document type. The original document is then divided into document fragments based on the initial document type. The file fragments are classified using a classifier to obtain a first classification result, and the file fragments are classified using a large language model to obtain a second classification result; The target file type of each file fragment is determined based on the first classification result and the second classification result; Structured review element information is extracted from the file fragment according to the target file type, and the review element information is reviewed based on preset review rules to generate the review result of the original file.

[0006] In one embodiment, the steps of classifying the file fragment using a classifier to obtain a first classification result and classifying the file fragment using a large language model to obtain a second classification result include: The file fragments are classified using a classifier to obtain a first classification result and a first classification confidence score; When the confidence score of the first classification is less than a preset first threshold, the file fragment is classified using a large language model to obtain a second classification result and a second classification confidence score. The step of determining the target file type of each file fragment based on the first classification result and the second classification result includes: When the confidence level of the first classification is greater than or equal to a preset first threshold, the target file type of the file fragment is determined based on the first classification result. When the confidence level of the second classification is greater than or equal to the preset second threshold, the target file type of the file fragment is determined based on the second classification result.

[0007] In one embodiment, the step of classifying the file fragment using a classifier to obtain a first classification result and a first classification confidence score includes: The fragment features of the file fragments are extracted using a classifier, and the fragment features include file name features, text features, and image features; The file fragments are classified based on the fragment features, and the first classification result and the first classification confidence score are output. The first classification result includes the probability distribution of each file type.

[0008] In one embodiment, the step of determining the target file type of the file fragment based on the first classification result includes: Based on a preset exclusion word list and preset priority rules, the first classification result is corrected; The target file type of the file fragment is determined based on the corrected first classification result.

[0009] In one embodiment, the step of type identification for each page of the original document containing the credit trade background to be examined to obtain the initial document type includes: Based on a preset invoice recognition model and preset contract recognition rules, the initial file type is obtained by performing type recognition on each page of the original document; Specifically, the invoice recognition model is used to identify invoice feature elements in the current page, and the current page is determined to be an invoice file based on the invoice feature elements. The keywords and layout features of the current page are determined, and the contract identification rules, keywords, and layout features are used to determine whether the current page is a contract document.

[0010] In one embodiment, the step of extracting structured review element information from the file fragment based on the target file type includes: When the target file type is a standard format file fragment, prompt words are configured based on the target file type, and the review element information is extracted based on the prompt words using a large language model. When the target file type is not a standard format file fragment, the review element information is extracted by combining the optical character recognition results and image features of the file fragment with a visual language model; When a file fragment of the target file type contains preset elements, the preset elements are extracted using the extraction model corresponding to the preset elements to obtain the review element information. The preset elements include at least one of the following: seal, handwritten signature, and invoice-specific field.

[0011] In one embodiment, the step of reviewing the review element information based on preset review rules and generating the review result of the original document further includes: The review elements are reviewed based on preset review rules to generate the review results of file fragments of the original file; The review results of the original file are generated by using the review results of each file segment from the large language model.

[0012] In one embodiment, prior to the step of type identification of each page of the original document containing the credit trade background to be examined to obtain the initial document type, the method further includes: If the original file is a compressed file, the original file is decompressed while retaining its hierarchical structure, and the format of all decompressed files is identified. The tilt angle of each page of the original document is detected using a rotation recognition model, and rotation correction is performed on each page according to the corresponding angle.

[0013] In addition, to achieve the above objectives, this application also proposes a credit and trade background review device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the credit and trade background review method as described above.

[0014] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the credit and trade background review method described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: The system identifies the type of each page of the original document and divides it into file segments accordingly, accurately breaking down multi-page documents into semantically independent review units, thus avoiding type misjudgment caused by cross-page information interference. Then, a classifier and a large language model are used to perform parallel dual-track classification of the file segments, fully leveraging the efficiency of the classifier on samples with well-defined features and the depth of understanding of the large language model on semantically complex samples. By fusing the first and second classification results, the target file type is determined, effectively overcoming the limitations of a single model in recognizing heterogeneous and non-standardized documents. Furthermore, based on the determined target file type, structured review element information is extracted, and automated review is performed based on preset review rules. This transforms the traditional model of relying on manual review, comparison, and verification of each document into an end-to-end intelligent review process, shortening the review cycle for credit and trade background information and improving the efficiency and accuracy of the review results. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the credit and trade background review method for this application. Figure 2 This is a flowchart illustrating the second embodiment of the credit and trade background review method for this application. Figure 3 This is a flowchart illustrating the third embodiment of the method for examining the credit and trade background of this application. Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the credit trade background review method in this application embodiment.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] The main solution of this application embodiment is as follows: Type identification is performed on each page of the original document containing the credit and trade background to be reviewed to obtain an initial document type; the original document is then divided into document fragments based on the initial document type; a classifier is used to classify the document fragments to obtain a first classification result, and a large language model is used to classify the document fragments to obtain a second classification result; the target document type of each document fragment is determined based on the first and second classification results; structured review element information is extracted from the document fragments based on the target document type, and the review element information is reviewed based on preset review rules to generate the review result of the original document.

[0023] In this embodiment, for ease of description, the following description uses a credit trade background review device as the implementing entity.

[0024] This application provides a solution that identifies the type of each page of the original document and divides it into file segments accordingly, accurately breaking down mixed multi-page documents into semantically independent review units, avoiding type misjudgment caused by cross-page information interference. Then, a classifier and a large language model are used to perform parallel dual-track classification of the file segments, fully leveraging the efficiency of the classifier on samples with well-defined features and the depth of understanding of the large language model on semantically complex samples. By fusing the first and second classification results, the target file type is determined, effectively overcoming the limitations of a single model in the identification of heterogeneous and non-standardized documents. Furthermore, based on the determined target file type, structured review element information is extracted, and automated review is performed based on preset review rules. This transforms the traditional model of relying on manual review, comparison, and verification of each document into an end-to-end intelligent review process, shortening the review cycle for credit and trade background information and improving the efficiency and accuracy of the review results.

[0025] It should be noted that the credit and trade background review device, the executing entity in this embodiment, can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions. The following description uses a credit and trade background review device as an example to illustrate this embodiment and the subsequent embodiments.

[0026] Based on this, the embodiments of this application provide a method for reviewing credit and trade backgrounds, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the credit and trade background examination method for this application.

[0027] In this embodiment, the credit trade background review method includes steps S10 to S40: Step S10: Type identification is performed on each page of the original document containing the credit trade background to be examined to obtain the initial document type. The original document is then divided into document fragments based on the initial document type.

[0028] In this embodiment, the original documents refer to the electronic supporting documents uploaded by the user to the server via the client when initiating the credit and trade background review process. Original documents may include: transaction vouchers, such as VAT invoices, VAT general invoices, sales contracts, purchase orders, delivery notes, commercial drafts, customs declarations, etc.; entity qualification documents, such as copies of business licenses, legal representative identification documents, bank account opening permits, etc.; and supporting documents, such as screenshots of bank transfer statements, account statements, receipts, and delivery confirmation letters. In actual business scenarios, the original documents uploaded by users have various file formats. Supported file formats include document formats, such as PDF, Word, Excel, etc.; image formats, such as JPG, JPEG, PNG, BMP, TIFF, and other static images; and compressed archive formats, such as ZIP, RAR, 7z, and other compressed files.

[0029] In this embodiment, the original file is divided into file fragments based on the initial file type. A file fragment is a preprocessed single file or single-page document unit that can be processed independently later. Multiple pages of the original file can be merged into one file fragment. For example, if a contract has 10 pages, and all 10 pages are identified as belonging to the contract type, they are merged into one file fragment. In this case, 1 file fragment = 10 pages. This ensures that during subsequent element extraction, the large language model can read the complete contract terms, rather than just reading fragments of a single page. A single page can also be a separate file fragment. If the uploaded file is a single-page invoice, then after splitting, 1 file fragment = 1 page. A page can be split into multiple fragments. If an image contains both an invoice and a contract, that page is cropped into two independent file fragments. In this case, 1 page = multiple file fragments.

[0030] The original documents may contain a mix of documents. In practice, users often merge multiple documents of different types into a single PDF file for convenience, or take photos containing multiple invoices in the same image, resulting in a single digital file containing multiple business entity types. The original documents may also exhibit diverse formats and origins, including standard invoices exported from standard tax systems, non-standard contracts with custom templates, and handwritten simplified agreements, each with different structures. Furthermore, the original documents may have image quality defects. Many documents originate from mobile phone photography, rapid scanning with a document scanner, or faxes, and may suffer from issues such as page tilt or rotation, blurry images, uneven lighting, low resolution, a mixture of handwritten and printed text, and overlapping or obscuring seals and text.

[0031] Because the original files are unstructured, they need to be preprocessed. This preprocessing involves standardizing the user-uploaded original files to make them standardized inputs that can be used for subsequent classification and feature extraction.

[0032] In an optional embodiment, before step S10, the method further includes: if the original file is a compressed package, decompressing the original file and retaining the hierarchical structure of the original file, and identifying the format of all files after decompression; using a rotation recognition model to detect the tilt angle of each page of the original file, and performing rotation correction for each page at the corresponding angle.

[0033] Preprocessing of the original files includes decompressing compressed files, automatically traversing and decompressing all files, preserving the original file hierarchy, and performing format recognition and classification on the decompressed files. Compressed file decompression supports ZIP, RAR, 7z, and other compression formats.

[0034] Preprocessing operations for the original files include page rotation correction, detecting and correcting page tilt caused by mobile phone photography, scanners, cameras, etc., and supporting automatic rotation at preset angles, such as 0°, 90°, 180°, and 270°.

[0035] By preserving the hierarchical structure through decompression and multi-format compatibility recognition, combined with a rotation correction mechanism, standardized file preprocessing is achieved. When the original file is a compressed package, it is decompressed while fully preserving the original file's hierarchical structure, ensuring that contextual information such as nested directories and associated attachments is not lost. Simultaneously, the format of all decompressed files is automatically identified, achieving unified compatibility processing for various types of heterogeneous files. Furthermore, a rotation recognition model is used to detect the tilt angle of each page of the original file and perform corresponding rotation correction, resolving page tilting and inversion issues caused by improper device placement or transmission in scanned or photographed documents. This transforms scattered, heterogeneous, and differently oriented original files into standardized, structurally clear, and correctly oriented file fragments, providing a high-quality input foundation for subsequent feature extraction by classifiers and semantic understanding by large language models. This improves the automation and recognition accuracy of the entire credit and trade background review process, reducing the risk of classification errors and information extraction omissions due to improper preprocessing.

[0036] This application provides end-to-end configurable management, supporting the definition of rules for each stage of classification, extraction, and review through configuration files. Optionally, before step S10, it also includes: configuring classification categories, priorities, keywords, exclusion words, and classification model parameters; configuring the fields of elements to be extracted, prompt word templates, and corresponding extraction model types for each document type; configuring single-document rules, cross-document association rules, and rule activation conditions; configuring the activation status of large model types and dedicated small models, as well as model call parameters. Model parameters include thresholds, fusion strategies, etc. Model types used for each element include LLM (Large Language Model), multimodal, dedicated small models, etc. Single-document rules include element value ranges, format validation, etc. Cross-document rules include association relationships, matching logic, etc. Rule activation conditions include document type combinations, business scenarios, etc. Model call parameters include temperature, maximum token, etc. This achieves end-to-end configurability in the trade background review field, supporting hot-swappable configuration of classification categories, extracted elements, review rules, and model selection, enabling rapid addition, modification, and reuse of custom scenario rules without code modification.

[0037] In an optional embodiment, step S10, which involves identifying the type of each page of the original document containing the credit and trade background to be examined to obtain the initial document type, includes: identifying the type of each page of the original document based on a preset invoice recognition model and preset contract recognition rules to obtain the initial document type; wherein, the invoice recognition model is used to identify invoice feature elements in the current page, and the invoice feature elements are used to determine whether the current page is an invoice document; the keywords and layout features of the current page are determined, and the contract recognition rules, the keywords, and the layout features are used to determine whether the current page is a contract document.

[0038] Based on invoice recognition models and contract recognition rules, each page is pre-identified to accurately segment it into different types. The invoice recognition model is a specialized model trained for VAT invoices, general invoices, and special invoices, identifying invoice characteristic elements such as invoice code, invoice number, tax rate, and amount. Contract recognition rules are based on keywords and layout features. Keywords include contract number, Party A / Party B, and signing date; layout features include paragraph structure and signature / signature area layout.

[0039] Based on a pre-defined invoice recognition model, the system automatically detects invoice-related features on the current page, quickly and accurately identifying invoice-type documents. This overcomes the limitations of general classifiers in recognizing invoices with diverse formats and complex anti-counterfeiting features. Simultaneously, by combining contract recognition rules and comprehensively analyzing keywords and format features of the current page, the system identifies contract-type documents. Dedicated recognition strategies are configured for invoices and contracts—two core, high-frequency document types—in credit and trade background reviews. This leverages the high accuracy of the dedicated model in specific domains while utilizing the interpretability of the rule engine to reduce the risk of misjudgment.

[0040] Step S20: Classify the file fragment using a classifier to obtain a first classification result, and classify the file fragment using a large language model to obtain a second classification result.

[0041] Optionally, the classifier can be an XGBoost (Extreme Gradient Boosting) classifier, which is an optimization machine learning algorithm based on gradient boosting decision trees. The XGBoost classifier constructs a series of weak learners and combines them to form a strong learner. XGBoost adds a regularization term to the loss function to prevent overfitting and uses second-order gradient information to accelerate convergence. Optionally, the XGBoost classifier is trained using historical labeled data, automatically learning keyword weights and supporting incremental learning to adapt to new categories. The trained classifier is then used to classify the file fragments to obtain the first classification result.

[0042] XGBoost can be replaced with LightGBM or CatBoost, depending on actual business needs. Both belong to the gradient boosting decision tree family and have their own advantages in training speed and memory usage. LightGBM uses a histogram-based decision tree splitting algorithm and one-sided gradient sampling technique, resulting in faster training speeds on large-scale datasets; CatBoost, on the other hand, has better native support for categorical features and can automatically handle categorical variables without manual coding. Alternatively, traditional machine learning algorithms such as Random Forest, Support Vector Machine, or Logistic Regression can also be used as alternatives.

[0043] Large language models refer to deep learning models that have been pre-trained on large-scale corpora and possess powerful semantic understanding capabilities. Classification prompts and the original file are input into a large language model, which then performs classification based on semantic understanding, outputting a secondary classification result. The prompt engineering involves carefully designed input hints to guide the model in generating specific outputs.

[0044] In this application, both the first classification result and the second classification result are probability distributions of the file type to which the file fragment belongs.

[0045] Step S30: Determine the target file type for each file fragment based on the first classification result and the second classification result.

[0046] Optionally, the target document types may include trade background documents such as invoices, contracts, and orders, as well as international trade documents such as customs declarations, bills of lading, and waybills; financial documents such as bank statements and credit reports; and qualification certificates such as ID cards and business licenses.

[0047] In this embodiment, the first classification result and the second classification result are combined according to a preset strategy to obtain the final file type. Optionally, the two models are combined after weighting according to their confidence levels. Alternatively, when the two are inconsistent, the priority preset by the rule takes precedence. Or, more models are introduced for voting. Fusion of the two classification results refers to combining the first classification result of the classifier with the second classification result output by the large language model according to a preset strategy. The classifier is finely trained based on business-annotated data, which has high accuracy and speed in classifying standard format files; the general large language model is pre-trained based on massive amounts of general knowledge, which has strong understanding of non-standard formats and complex semantics, but is costly. The fusion of the two can achieve complementary advantages.

[0048] Step S40: Extract structured review element information from the file fragment according to the target file type, and review the review element information based on preset review rules to generate the review result of the original file.

[0049] In this embodiment, the corresponding information extraction model is invoked based on the target file type to extract structured review element information from the file fragments. Based on the extracted structured element information, logical judgments are performed according to preset review rules, and a structured review result is finally output. The review result can be a structured JSON report, containing the pass / fail determination for each review rule, a risk point explanation, and a visual display. In addition to the JSON report, a visual dashboard can be added, and the review result can be pushed to downstream systems such as credit approval systems and risk control systems. It can also generate a PDF review report for archiving and printing.

[0050] In an optional embodiment, the step S40 of reviewing the review element information based on preset review rules and generating the review result of the original document further includes: reviewing the review element information based on preset review rules and generating the review result of the file fragments of the original document; and using a large language model to review the review result of each file fragment and generate the review result of the original document.

[0051] Based on preset review rules, each document fragment's review elements are independently reviewed, generating review results for each fragment. This fully leverages the rule engine's advantages in the certainty, interpretability, and consistency of structured element verification, ensuring rigorous verification of compliance in key dimensions such as subject, amount, quantity, and serial number. Furthermore, a large language model is used to perform global semantic understanding and logical integration of the dispersed review results from each document fragment, automatically identifying implicit relationships across fragments, temporal contradictions, or integrity defects in the overall trade chain, generating a comprehensive review conclusion covering the overall risk assessment of the original documents. This organic combination of the rigid constraints of the rule engine and the flexible reasoning capabilities of the large language model overcomes the limitations of pure rule-based review in handling complex logical relationships across documents, while avoiding the illusions and computational waste that can easily occur when the large language model directly processes the original documents. This achieves a systematic leap from partial compliance to global authenticity in review, enhancing the comprehensiveness and accuracy of credit and trade background review conclusions.

[0052] In an optional embodiment, step S40 involves reviewing the review element information based on preset review rules to generate the review result of the original document. This includes: comparing the review element information of file fragments from different original documents based on consistency review rules, wherein the consistency review rules include at least one of subject consistency, amount consistency, quantity consistency, and number consistency; and performing logical verification on the review element information of file fragments from the original document based on validity review rules, wherein the validity review rules include at least one of time sequence validity, time limit validity, completeness, seal validity, and document authenticity.

[0053] The review rules include single-document rules, which validate elements within a single document, such as whether the amount format is correct and the date is valid. The review rules also include cross-document rules, which compare and validate elements across multiple documents. The review content includes consistency checks, such as matching the seller's name on the contract with the seller's name on the invoice, matching the contract amount with the total invoice amount, matching the order quantity with the invoice quantity, and verifying the correspondence between the contract number and the invoice remarks number. The review content also includes validity checks, such as ensuring the invoice date is not earlier than the contract signing date, the contract is within its validity period, page numbers are consecutive without missing pages, and the seal is complete and unsmeared.

[0054] The consistency review rules are shown in the table below:

[0055] The validity review rules are shown in the table below:

[0056] Based on consistency review rules, the system compares and correlates information from document fragments across different original documents, covering key dimensions such as consistency of subject, amount, quantity, and serial number. This effectively identifies information inconsistencies and data mismatches between multi-source documents such as contracts, invoices, and logistics documents, preventing fictitious trade backgrounds and circular invoicing. Based on validity review rules, the system logically verifies information from review elements, including temporal validity, time limit validity, completeness, seal validity, and document authenticity, ensuring a reasonable transaction timeline, compliant business terms, and complete and authentic signature elements. By integrating scattered and isolated document fragment review elements into a complete chain of trade background evidence, the system achieves a systematic review from the compliance of individual documents to the overall authenticity of the transaction, improving the comprehensiveness and accuracy of credit risk identification.

[0057] In this embodiment, the original document is type-identified for each page and divided into file segments accordingly. This accurately breaks down multi-page documents into semantically independent review units, avoiding type misjudgments caused by cross-page information interference. Then, a classifier and a large language model are used for parallel dual-track classification of the file segments. This fully leverages the efficiency of the classifier on samples with well-defined features and the depth of understanding of the large language model on semantically complex samples. By fusing the first and second classification results, the target file type is determined, effectively overcoming the limitations of a single model in recognizing heterogeneous and non-standardized documents. Furthermore, structured review element information is extracted based on the determined target file type, and automated review is performed based on preset review rules. This transforms the traditional model of manually reviewing, comparing, and verifying each document one by one into an end-to-end intelligent review process, shortening the credit and trade background review cycle and improving the efficiency and accuracy of the review results.

[0058] Based on the first embodiment of this application, in the second embodiment of this application, the same or similar content as the above embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Step S20 includes: Step S21: Use a classifier to classify the file fragments to obtain a first classification result and a first classification confidence level; Step S22: When the confidence score of the first classification is less than the preset first threshold, the file fragment is classified using a large language model to obtain a second classification result and a second classification confidence score. Step S30 includes: Step S31: When the confidence level of the first classification is greater than or equal to a preset first threshold, the target file type of the file fragment is determined according to the first classification result; Step S32: When the confidence level of the second classification is greater than or equal to the preset second threshold, the target file type of the file fragment is determined according to the second classification result.

[0059] When the confidence level of the first classification is greater than or equal to a preset first threshold, the first classification result is directly adopted. When the confidence level of the first classification output by the classifier is less than the preset threshold, the first classification result is not sufficiently confident, triggering a second-level classification. The first confidence level refers to the classifier's degree of confidence in a particular classification, expressed as a probability value. The preset first threshold is a configurable critical value used to determine the credibility of the model's output.

[0060] The preset first threshold is a configurable hyperparameter, ranging from 0 to 1. A higher preset first threshold results in stricter conditions for triggering the large language model, allowing more files to be classified independently by the classifier, leading to faster system response but potentially sacrificing a small amount of accuracy. Conversely, a lower preset first threshold results in more lenient conditions for triggering the large language model, allowing more files to receive assistance from the model, leading to higher classification accuracy but also increasing system response speed and computational cost. In a preferred embodiment, the preset threshold is set to 0.7. This value was determined through a balance test of classification accuracy and response latency under different business scenarios, ensuring that most standard documents can be quickly processed while also ensuring that difficult documents receive supplementary judgment from the large language model.

[0061] In an optional embodiment, step S21 includes: extracting fragment features of the file fragment using a classifier, the fragment features including filename features, text features and image features; classifying the file fragment based on the fragment features, and outputting a first classification result and a first classification confidence score, the first classification result including the probability distribution of each file type.

[0062] Filename features include textual information extracted from the stored names of file fragments. These features include keywords in the filename, such as invoices, contracts, orders, etc. In real-world business scenarios, user-uploaded file names often contain semantic information, such as purchase contracts from XX company, VAT invoices from March 2024, and order details. By extracting keywords from filenames and performing word segmentation, statistical analysis, and encoding, valuable prior clues can be provided for classification.

[0063] Text features include textual statistical information extracted from the main text of a document fragment. Optionally, textual features include word frequency statistics, keyword frequency, and the presence or absence of specific terms. For example, a contract typically contains frequently occurring terms such as "Party A," "Party B," "contract number," and "signing date," while an invoice frequently contains terms such as "invoice code," "invoice number," "tax rate," and "total price including tax." By statistically analyzing and encoding the frequency distribution of these terms, textual feature vectors that distinguish different document types can be constructed.

[0064] Image features include layout and graphic element features extracted from the visual appearance of document fragments. For scanned or photographed documents, document types can often be distinguished by their layout structure. For example, VAT invoices typically have fixed red borders and specific table layouts, while contracts usually feature centered titles and clearly defined paragraphs. Image features can be extracted by calculating text density distribution, table area proportions, and image edge gradient histograms, or they can be automatically learned using deep learning methods such as convolutional neural networks.

[0065] Other feature dimensions can be added or replaced according to the actual scenario. Among them, fragment features can also include: file metadata features, such as creation time, modification time, file size, file format suffix, etc.; OCR confidence features, such as the confidence distribution of each character in the recognition result; and page number features.

[0066] If N file types are supported, such as invoices, contracts, orders, delivery notes, customs declarations, business licenses, and ID cards, the classifier output is an N-dimensional vector, for example, [0.02, 0.85, 0.05, 0.03, 0.02, 0.01, 0.02]. The largest component, 0.85, corresponds to the contract category, indicating that the classifier judges that the file fragment has an 85% probability of belonging to the contract category.

[0067] By simultaneously extracting filename, text, and image features using a classifier, a multi-dimensional fragment feature vector is constructed from three dimensions: file naming rules, text semantic content, and page visual layout. This overcomes the information blind spots of single text or image features in heterogeneous document recognition and enhances the generalization ability to recognize file fragments from different sources, such as scanned documents, photographs, and electronic documents. Based on the fused fragment features, classification is performed, and the output includes the probability distribution of each category and the confidence score of the first classification. This not only provides a quantitative basis for subsequent fusion decisions with the second classification results of the large language model, but also enables the measurability of classification certainty through the confidence score index.

[0068] In an optional embodiment, the step of determining the target file type of the file fragment based on the first classification result includes: modifying the first classification result based on a preset exclusion vocabulary and preset priority rules; and determining the target file type of the file fragment based on the modified first classification result.

[0069] Priority rules are a set of pre-configured logical judgment conditions. Each rule specifies whether a particular category should be favored or excluded when a specific keyword appears in the content of a document fragment. For example, a rule can be configured to prioritize classifying a document as an invoice when both an invoice code and a tax rate appear in the document body; another rule can be configured to prioritize classifying a document as a contract when both "Party A" and "Party B" appear in the document body. Priority rules can be set with different priorities, and when multiple rules are triggered simultaneously, they are decided according to the preset priority order.

[0070] An exclusion thesaurus is a list of specific words. When a word from the exclusion thesaurus appears in a document fragment, the corresponding classification result should be excluded or downgraded. For example, for contract classification, exclusion words such as "for reference only," "draft," "void," and "invalid" can be configured. When a document is marked as a draft or void, even if its content meets the characteristics of a contract, it should not be considered a valid contract and proceed to the subsequent element extraction and review process. The purpose of the exclusion thesaurus is to filter out invalid, void, or internally-only documents, preventing these documents from being mistakenly identified as valid business documents.

[0071] Based on a pre-defined exclusion list and pre-defined priority rules, the first classification result is modified. Optionally, this can be done by increasing the probability value of the category corresponding to the matching keyword, decreasing the probability value of the excluded word hitting the category, directly overriding the first classification result, or applying a weighted bias term to the first classification result. The modified first classification result still maintains a normalized state, that is, the sum of the probabilities of all categories equals 1.

[0072] The first classification result is filtered based on a pre-defined exclusion vocabulary to eliminate obvious misclassifications caused by ambiguous filenames, OCR misidentification, or image interference, reducing the model's sensitivity to noisy features. The corrected classification result is then adjudicated using pre-defined priority rules, ranking and optimizing different file types based on their criticality, frequency of occurrence, or risk weight in the business scenario. This ensures that higher-priority types are matched first when probability distributions are similar or ambiguity exists. This deep integration of purely data-driven classification results with domain knowledge rules compensates for classifier biases when training sample coverage is insufficient or boundary samples are ambiguous, improving the interpretability and reliability of the classification results in practical applications.

[0073] In this embodiment, a lightweight classifier is used to quickly perform initial classification of document fragments and output a first classification confidence score, fully leveraging its advantages of low latency and low computational consumption in processing conventional samples. Only when the first classification confidence score falls below a preset first threshold is a computationally expensive general-purpose language model invoked for secondary classification. This provides deep semantic understanding support for difficult boundary samples, avoiding resource waste and response delays caused by indiscriminately calling large models on all samples. While ensuring overall classification accuracy, this reduces the frequency of large model calls and computational costs, achieving a dynamic balance between review efficiency and accuracy. It solves the problem of high efficiency but low accuracy or high accuracy but low efficiency faced by single models in credit and trade background review scenarios.

[0074] Based on any embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 The content of extracting structured review element information from the file fragment according to the target file type in step S40 includes: Step S41: When the target file type is a standard format file fragment, configure prompt words based on the target file type, and use a large language model to extract the review element information based on the prompt words; Step S42: When the target file type is not a standard format file fragment, the review element information is extracted by combining the optical character recognition results and image features of the file fragment with a visual language model; Step S43: When the file fragment of the target file type has preset elements, the preset elements are extracted using the extraction model corresponding to the preset elements to obtain the review element information. The preset elements include at least one of the following: seal, handwritten signature, and invoice-specific field.

[0075] Large-scale language model extraction is suitable for standard-format text files and structured tables, extracting data through configurable prompt word templates. Visual language model extraction is suitable for documents with complex layouts and mixed text and images, combining optical character recognition results and image features for joint extraction. Unstructured document content is transformed into structured data with clearly defined field names and values, such as JSON format, for example: {Contract Number: HT2024001, Party A: XX Company, Amount: 1000000}. Dedicated small-scale model extraction is used, deploying specialized models for preset elements such as seal recognition, handwriting recognition, and invoice recognition.

[0076] The review elements are key information fields required for conducting trade background reviews. For example, the review elements include: contract number, name of Party A, name of Party B, contract amount, signing date, invoice code, invoice number, name of seller, name of buyer, invoice date, invoice amount, order number, product name, quantity, unit price, etc.

[0077] In this embodiment, when the target file type is a standard format type, the large language model is called first; when it is a non-standard image and text mixed type, the visual language model is called first; when a special physical verification is detected, a special small model is dynamically interspersed.

[0078] Standard format document fragments can be standard format text and structured tables. Standard format text refers to electronic documents with standardized layout, clear text, primarily linear reading order, and where complete semantic information can be obtained without relying on complex page layouts. Examples include electronic invoices directly exported from tax systems, standard contract templates generated from enterprise information systems, Word-formatted agreement documents, and clearly scanned PDF documents with correct text orientation. Structured tables refer to document content that organizes data in rows and columns, with clearly defined fields and a clear correspondence between values ​​and labels. Examples include Excel-formatted order details tables, CSV-formatted financial statements, and tables with regular gridlines in PDFs.

[0079] In this solution, when the large language model extracts elements, the system starts the streaming output mode, and the extraction results are presented incrementally. Users can see that some elements have been extracted within a few seconds, which improves the interactive experience.

[0080] Non-standard format document fragments can be documents with complex layouts and a mix of text and images. Complex layouts refer to documents with irregular page layouts, where information is distributed across multiple areas of the page, and the logical relationship between these areas does not depend on linear reading order but on visual positional correspondence. Examples include multi-column documents with newspaper-style layouts, financial statements with complex nested tables, and scanned copies of invoices and contracts mixed on the same page. A mix of text and images refers to documents that contain numerous image elements in addition to text content, and these image elements themselves carry crucial review information. Examples include scanned copies of contracts with company logos and seals, application forms with handwritten annotations, and logistics documents containing barcodes or QR codes.

[0081] Optical character recognition (OCR) results are the output of an OCR system after detecting and recognizing text in document images. The system detects the location of text regions on the page, then performs character recognition on the text images in each region, outputting the recognized text content and its bounding box coordinates on the page. In this scheme, the OCR results received by the visual language model include not only the recognized text strings, but also the position coordinates of each text line, the recognition confidence score, and overall page orientation information.

[0082] Image features refer to high-dimensional feature vectors extracted from the visual images of document pages, used to characterize the visual appearance attributes of the page. These features can be deep semantic features extracted through convolutional neural networks, such as edge distribution, texture patterns, color distribution, and page block segmentation. Image features capture information that the text itself cannot convey, such as the color and shape of a seal, the handwriting style of a handwritten signature, the texture and pattern of the page background, and the visual separation between different content blocks.

[0083] Joint extraction refers to the visual language model's simultaneous use of optical character recognition results and image features for comprehensive judgment during the inference process, rather than treating them as independent information sources and processing them separately. The visual language model inputs the recognized text content, the text's position coordinates, and the overall visual image of the page into a unified network architecture, and uses an attention mechanism to allow the model to autonomously learn the relationships between different modalities of information.

[0084] This refers to scenarios where file fragments of the target file type contain pre-defined elements. These pre-defined elements are information elements that are difficult to extract stably using large language models or general visual language models, requiring specialized training and processing. Seal information is typically presented on a document as a red circular or elliptical image. The text may be obscured by the seal's border, overlap with other text or lines, or even become blurred due to uneven pressure. Seal recognition not only needs to detect the presence of a seal but also assess its integrity, clarity, and logical consistency with the document's claims. Handwritten signatures are highly personalized; different people have vastly different writing styles, and even the same person's signature may vary at different times. Furthermore, handwritten signatures are usually in image form rather than text form, making handwritten recognition far less accurate than printed signatures. Handwritten signature recognition not only needs to confirm the signature's presence in a specified location but sometimes also requires a preliminary assessment of its authenticity. Invoice-specific fields refer to key information items in VAT invoices that have specific format requirements, such as invoice code, invoice number, tax control code, and verification code.

[0085] The extraction model includes a seal recognition model, which is a deep learning model specifically designed for detecting and recognizing seal information in documents. Seal recognition typically comprises three sub-tasks: the first is seal detection, which involves locating the seal in the document image; the second is seal segmentation, which separates the seal area from the background; and the third is seal text recognition, which reads the text content from the seal. The text in seals is usually arranged in a circular or vertical pattern, requiring specialized text correction and recognition algorithms for processing.

[0086] The extraction model includes a handwritten signature recognition model, which is specifically designed to detect and recognize handwritten signature regions in documents. Handwritten signature recognition typically comprises two sub-tasks: the first is signature detection, which involves locating the signature region in the document image; the second is signature verification, which determines whether the signature is authentic and matches a pre-reserved signature sample. Signature verification usually uses dynamic features such as stroke characteristics, pressure distribution characteristics, and writing rhythm characteristics, or static image features for comparison.

[0087] Extraction models include invoice-specific field recognition models, which are specifically designed to extract key tax-related fields from invoice images. These models are typically optimized for fixed invoice formats, utilizing the relative positions of various fields within the invoice to improve extraction accuracy. Some invoice-specific field recognition models also integrate format validation logic, such as automatically verifying whether the invoice number is 8 digits and whether the invoice code conforms to the coding rules of the State Taxation Administration.

[0088] In this embodiment, when the target document is a standard format document fragment, targeted prompts are configured based on the document type. A large language model is used to extract review element information based on these prompts, fully leveraging the model's semantic understanding capabilities of structured text to quickly and accurately extract common elements such as contract terms, amounts, and dates. When the target document is not a standard format document fragment, a visual language model is used in conjunction with optical character recognition results and image features for extraction. This effectively addresses scanned or photographed documents with chaotic layouts and complex formatting, ensuring no information is missed through multimodal fusion. When the document fragment contains preset elements such as seals, handwritten signatures, or invoice-specific fields, the corresponding dedicated extraction model is invoked for accurate identification, overcoming the shortcomings of general models in recognizing special elements. Ultimately, this achieves full coverage extraction of review elements for credit and trade documents from different sources, in different formats, and with varying degrees of complexity, improving the accuracy of structured information extraction.

[0089] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the credit and trade background examination method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0090] This application provides a credit and trade background review device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the credit and trade background review method in Embodiment 1 above.

[0091] The following is for reference. Figure 4 The diagram illustrates a structural schematic suitable for implementing the credit and trade background review device in the embodiments of this application. The credit and trade background review device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, personal digital assistants (PDAs), tablet computers (PADs), portable multimedia players (PMPs), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The credit and trade background review equipment shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0092] like Figure 4As shown, the credit and trade background review equipment may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the credit and trade background review equipment. The processing device 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the credit and trade background review equipment to communicate wirelessly or wiredly with other equipment to exchange data. While the figures show credit and trade background review equipment with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0093] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0094] The credit and trade background review equipment provided in this application, employing the credit and trade background review method described in the above embodiments, can solve the technical problem of low accuracy in credit and trade background reviews. Compared with the prior art, the beneficial effects of the credit and trade background review equipment provided in this application are the same as those of the credit and trade background review method provided in the above embodiments, and other technical features of this credit and trade background review equipment are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0095] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0096] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0097] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to perform the credit trade background review method in the above embodiments.

[0098] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), etc., or any suitable combination thereof.

[0099] The aforementioned computer-readable storage medium may be included in the credit and trade background review equipment; or it may exist independently and not be assembled into the credit and trade background review equipment.

[0100] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the credit and trade background review equipment, the equipment performs the following actions: It identifies the type of each page of the original document and divides it into file segments accordingly, accurately breaking down multi-page documents into semantically independent review units, avoiding type misjudgments caused by cross-page information interference; it then uses a classifier and a large language model to perform parallel dual-track classification of the file segments, fully leveraging the efficiency of the classifier on samples with well-defined features and the depth of understanding of the large language model on semantically complex samples. By fusing the first and second classification results, it determines the target file type, effectively overcoming the limitations of a single model in recognizing heterogeneous and non-standardized documents; furthermore, it extracts structured review element information based on the determined target file type and performs automated review based on preset review rules, transforming the traditional model of relying on manual review, comparison, and verification of each document into an end-to-end intelligent review process, shortening the credit and trade background review cycle and improving the efficiency and accuracy of the review results.

[0101] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0102] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0103] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0104] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described credit and trade background review method, thereby solving the technical problem of low accuracy in credit and trade background reviews. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the credit and trade background review method provided in the above embodiments, and will not be repeated here.

[0105] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the credit and trade background review method described above.

[0106] The computer program product provided in this application can solve the technical problem of low accuracy in credit and trade background review. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the credit and trade background review method provided in the above embodiments, and will not be repeated here.

[0107] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for reviewing credit and trade background, characterized in that, The methods for reviewing credit and trade background include: Each page of the original document containing the credit and trade background to be examined is type-identified to obtain an initial document type. The original document is then divided into document fragments based on the initial document type. The file fragments are classified using a classifier to obtain a first classification result, and the file fragments are classified using a large language model to obtain a second classification result; The target file type of each file fragment is determined based on the first classification result and the second classification result; Structured review element information is extracted from the file fragment according to the target file type, and the review element information is reviewed based on preset review rules to generate the review result of the original file.

2. The credit and trade background review method as described in claim 1, characterized in that, The steps of classifying the file fragment using a classifier to obtain a first classification result, and classifying the file fragment using a large language model to obtain a second classification result include: The file fragments are classified using a classifier to obtain a first classification result and a first classification confidence score; When the confidence score of the first classification is less than a preset first threshold, the file fragment is classified using a large language model to obtain a second classification result and a second classification confidence score. The step of determining the target file type of each file fragment based on the first classification result and the second classification result includes: When the confidence level of the first classification is greater than or equal to a preset first threshold, the target file type of the file fragment is determined based on the first classification result. When the confidence level of the second classification is greater than or equal to the preset second threshold, the target file type of the file fragment is determined based on the second classification result.

3. The credit and trade background review method as described in claim 2, characterized in that, The step of classifying the file fragment using a classifier to obtain a first classification result and a first classification confidence score includes: The fragment features of the file fragments are extracted using a classifier, and the fragment features include file name features, text features, and image features; The file fragments are classified based on the fragment features, and the first classification result and the first classification confidence score are output. The first classification result includes the probability distribution of each file type.

4. The credit and trade background review method as described in claim 3, characterized in that, The step of determining the target file type of the file fragment based on the first classification result includes: Based on a preset exclusion word list and preset priority rules, the first classification result is corrected; The target file type of the file fragment is determined based on the corrected first classification result.

5. The credit and trade background review method as described in claim 1, characterized in that, The step of identifying the initial document type for each page of the original document containing the credit and trade background to be examined includes: Based on a preset invoice recognition model and preset contract recognition rules, the initial file type is obtained by performing type recognition on each page of the original document; Specifically, the invoice recognition model is used to identify invoice feature elements in the current page, and the current page is determined to be an invoice file based on the invoice feature elements. The keywords and layout features of the current page are determined, and the contract identification rules, keywords, and layout features are used to determine whether the current page is a contract document.

6. The credit and trade background review method as described in claim 1, characterized in that, The step of extracting structured review element information from the file fragment based on the target file type includes: When the target file type is a standard format file fragment, prompt words are configured based on the target file type, and the review element information is extracted based on the prompt words using a large language model. When the target file type is not a standard format file fragment, the review element information is extracted by combining the optical character recognition results and image features of the file fragment with a visual language model; When a file fragment of the target file type contains preset elements, the preset elements are extracted using the extraction model corresponding to the preset elements to obtain the review element information. The preset elements include at least one of the following: seal, handwritten signature, and invoice-specific field.

7. The credit and trade background review method as described in claim 1, characterized in that, The step of reviewing the review element information based on preset review rules and generating the review result of the original document further includes: The review elements are reviewed based on preset review rules to generate the review results of file fragments of the original file; The review results of the original file are generated by using the review results of each file segment from the large language model.

8. The credit and trade background review method as described in claim 1, characterized in that, Before the step of type identification of each page of the original document containing the credit trade background to be examined to obtain the initial document type, the method further includes: If the original file is a compressed file, the original file is decompressed while retaining its hierarchical structure, and the format of all decompressed files is identified. The tilt angle of each page of the original document is detected using a rotation recognition model, and rotation correction is performed on each page according to the corresponding angle.

9. A credit and trade background review device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the credit trade background review method as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the credit trade background review method as described in any one of claims 1 to 8.

Images