Cross-border transaction auditing method and device, equipment and storage medium
By combining multimodal large models and large language models, the problems of low efficiency and accuracy in processing multi-format files in cross-border transaction audits are solved, and efficient and accurate cross-border transaction compliance audits are achieved.
Patent Information
- Application Number
- CN202510791600.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-13
- Publication Date
- 2025-09-12
AI Technical Summary
Traditional OCR technology is inefficient and error-prone when processing complex multi-format, multi-modal documents in cross-border transaction reviews, and it is difficult to accurately assess the correlation and consistency between fields.
Use multimodal big models and data analysis tools to extract fields, combine with big language models to conduct rule review, process transaction materials in formats such as images, PDF, and Excel through multimodal big models, and use big language models for semantic understanding and logical reasoning.
It improves the automation level, accuracy and efficiency of cross-border transaction audits, ensures the accuracy and consistency of data processing, and generates audit conclusions that meet the requirements.
Smart Images
Figure CN120634729A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing technology, and specifically to a method, apparatus, device, and storage medium for reviewing cross-border transactions. Background Art
[0002] Compliance reviews for cross-border transactions primarily cover two key aspects. First, the transaction instructions submitted by merchants on cross-border platforms contain crucial transaction details, such as the payer, payee, amount, currency, and transaction object. This directly reveals the basic transaction data and capital flow. Second, transaction authenticity documents such as contracts, invoices, and logistics documents are used to verify the authenticity and legality of transactions. Comprehensive and accurate review of this information is crucial to ensuring compliance in cross-border transactions.
[0003] However, traditional auditing methods primarily rely on optical character recognition (OCR) technology to extract text content from a variety of file types. While this step completes the digital conversion of data, the diverse file formats and complex and dispersed content make the audit process time-consuming and prone to omissions and misjudgments. Furthermore, relying solely on OCR to generate text often lacks semantic understanding, making it difficult to accurately assess the association and consistency between fields, thus affecting the accuracy and efficiency of audits.
[0004] Therefore, how to efficiently and accurately review the complex and diverse document data of cross-border transactions to overcome the inefficiency and high error rate of traditional OCR methods when processing files with various formats, complex and scattered content is a technical problem that technicians in this field urgently need to solve. Summary of the Invention
[0005] Based on the above problems, this application provides a cross-border transaction audit method, device, equipment and storage medium, which can break through the limitations of traditional OCR that is limited to text extraction when processing multi-format, multi-modal complex transaction materials, and achieve deep semantic understanding and precise field extraction of cross-border transaction instructions and authenticity materials, thereby improving the automation level, accuracy and efficiency of the audit.
[0006] The embodiments of this application disclose the following technical solutions:
[0007] A cross-border transaction review method, comprising:
[0008] Obtain transaction-related materials for the cross-border transaction to be reviewed; the transaction-related materials include transaction information and transaction authenticity materials; the transaction information includes transaction element information between the buyer and the seller during the cross-border transaction; the transaction authenticity materials include trade certificates generated by the buyer and the seller during the cross-border transaction;
[0009] Based on the data format of the data in the transaction-related materials, using a multimodal large model and / or a data analysis tool to perform field extraction processing on the transaction-related materials to obtain multiple field information;
[0010] Merging and merging the same type of fields on the multiple fields to obtain information to be reviewed;
[0011] The information to be reviewed, the review rules and the first prompt word are input into the large language model to conduct rule review of the information to be reviewed, and the review result of the information to be reviewed is obtained; the first prompt word is used to guide the large language model to accurately understand the review rules, strictly execute the rule logic, and output a review conclusion that meets the requirements.
[0012] In a possible implementation, the data format of the data in the transaction-related materials includes at least one of a picture format, a portable document PDF format, and an Excel format.
[0013] In one possible implementation, based on the data format of the data in the transaction-related materials, a multimodal large model and / or a data analysis tool is used to perform field extraction processing on the transaction-related materials to obtain multiple field information, including:
[0014] If the data format of the data in the transaction-related materials is a picture format, the transaction-related materials and the second prompt word are input into the multimodal large model for field extraction processing to obtain the multiple field information; the second prompt word is used to guide the multimodal large model to accurately extract key information from the transaction-related materials.
[0015] In one possible implementation, based on the data format of the data in the transaction-related materials, a multimodal large model and / or a data analysis tool is used to perform field extraction processing on the transaction-related materials to obtain multiple field information, including:
[0016] If the data format of the transaction-related materials is PDF format, splitting the transaction-related materials into multiple images to obtain multiple image files;
[0017] The multiple image files and the second prompt word are input into the multimodal large model for field extraction processing to obtain the multiple field information; the second prompt word is used to guide the multimodal large model to accurately extract key information from the transaction-related materials.
[0018] In one possible implementation, based on the data format of the data in the transaction-related materials, a multimodal large model and / or a data analysis tool is used to perform field extraction processing on the transaction-related materials to obtain multiple field information, including:
[0019] If the data format of the data in the transaction-related materials is Excel format, the transaction-related materials are subjected to field extraction processing by the data analysis tool to obtain the plurality of field information.
[0020] In a possible implementation, when the data format of the transaction-related materials includes an image format and / or a PDF format, merging and merging the multiple fields of information of the same type to obtain the information to be reviewed includes:
[0021] The multiple field information and the third prompt word are input into the large language model for structured integration processing to obtain the information to be reviewed; the third prompt word is used to guide the multimodal large model to accurately merge field information of the same type in the multiple field information.
[0022] In a possible implementation, when the data format of the data in the transaction-related materials includes image format + Excel format, image format + PDF format + Excel format, or PDF format + Excel format, merging and consolidating the multiple fields of information of the same type to obtain the information to be reviewed includes:
[0023] Inputting multiple fields of information extracted from the transaction-related materials in image format and / or PDF format and the third prompt word into the large language model for structured integration processing to obtain first preliminary information to be reviewed, and merging and merging multiple fields of information extracted from the transaction-related materials in Excel format for fields of the same type to obtain second preliminary information to be reviewed; the third prompt word is used to guide the multimodal large model to accurately merge field information of the same type among the multiple fields of information;
[0024] The first preliminary information to be reviewed and the second preliminary information to be reviewed are merged and processed by merging the same type of fields to obtain the information to be reviewed.
[0025] A cross-border transaction audit device, comprising:
[0026] An acquisition unit, configured to acquire transaction-related materials for a cross-border transaction to be reviewed; the transaction-related materials include transaction information and transaction authenticity materials; the transaction information includes transaction element information between the buyer and the seller during the cross-border transaction; the transaction authenticity materials include trade vouchers generated by the buyer and the seller during the cross-border transaction;
[0027] a field extraction unit configured to extract fields from the transaction-related materials using a multimodal large model and / or a data analysis tool based on the data format of the data in the transaction-related materials to obtain a plurality of field information;
[0028] A merging unit, configured to merge the same type of fields on the plurality of field information to obtain information to be reviewed;
[0029] The input unit is used to input the information to be reviewed, the review rules and the first prompt word into the large language model to perform rule review of the information to be reviewed, and obtain the review result of the information to be reviewed; the first prompt word is used to guide the large language model to accurately understand the review rules, strictly execute the rule logic, and output a review conclusion that meets the requirements.
[0030] A cross-border transaction audit device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the cross-border transaction audit method described above is implemented.
[0031] A computer-readable storage medium stores instructions, which, when executed on a terminal device, enable the terminal device to execute the cross-border transaction audit method as described above.
[0032] Compared with the existing technology, this application has the following beneficial effects:
[0033] The present application provides a method, apparatus, device, and storage medium for auditing cross-border transactions. Specifically, when executing the cross-border transaction audit method provided in the embodiments of the present application, first, materials related to the transaction to be audited can be obtained. These materials include transaction information and transaction authenticity materials. The transaction information covers the key elements of the buyer and seller in the cross-border transaction, while the transaction authenticity materials include the trade documents generated by both parties. For transaction-related materials in different data formats, a multimodal large model and / or data analysis tools are used to extract fields to obtain multiple structured field information. Subsequently, the extracted field information is merged according to the same type to form complete and consistent information to be audited. Finally, the information to be audited, the preset audit rules, and the first prompt word are input into the large language model. Leveraging its powerful semantic understanding and logical reasoning capabilities, the audit rules are accurately executed and a satisfactory audit conclusion is generated, thereby achieving efficient and accurate cross-border transaction compliance audits. By utilizing a multimodal large model and data analysis tools, the present application can more efficiently process and extract field information from transaction-related materials in different formats, improving the accuracy and speed of data processing. The extracted information is then merged and consolidated for fields of the same type to form structured information for review, enhancing data consistency and integrity. Furthermore, the information for review, review rules, and the first prompt word are fed into a large language model for rule review. Leveraging the model's semantic understanding and logical reasoning capabilities, the accuracy and efficiency of the review results are ensured. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0035] Figure 1 A flowchart of a cross-border transaction review method provided in an embodiment of the present application;
[0036] Figure 2 A schematic diagram of the structure of a cross-border transaction audit device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0037] To facilitate understanding of the technical solutions provided by the embodiments of the present application, the background technology involved in the embodiments of the present application will be described below.
[0038] Compliance audits for cross-border transactions primarily consist of two key components: first, the transaction instructions submitted by merchants on cross-border platforms. These instructions detail key elements such as the payer, payee, amount, currency, and transaction subject, directly reflecting the basic transaction information and capital flow; second, transaction authenticity documents, such as contracts, invoices, and logistics documents, which are used to verify the authenticity and legality of the transaction. Comprehensive and accurate review of this information is crucial for ensuring the compliance of cross-border transactions. However, traditional audit methods primarily rely on optical character recognition (OCR) technology to extract text content from various file types. While this step achieves digital data conversion, the diverse file formats (including images, PDFs, spreadsheets, etc.), complex content structures, and dispersed information make the audit process time-consuming and prone to omissions and misjudgments. Furthermore, text generated solely by OCR often lacks semantic understanding, making it difficult to effectively determine the association and consistency between fields, thus affecting the accuracy and efficiency of the audit.
[0039] To solve this problem, an embodiment of the present application provides a cross-border transaction audit method, device, equipment and storage medium. First, relevant materials of the transaction to be audited are obtained. These materials include transaction information and transaction authenticity materials. The transaction information covers the key elements of the buyer and seller in the cross-border transaction process, and the transaction authenticity materials are the trade documents generated by both parties. Subsequently, based on the data format of the transaction-related materials, a multimodal large model or data analysis tool is used to extract fields from these materials to extract multiple key information fields. Then, fields of the same type are merged to form structured information to be audited. Finally, the information to be audited, the preset audit rules and the first prompt word are input into the large language model. With the help of the prompt word, the model is guided to accurately understand and strictly implement the audit rules, thereby generating audit results that meet the requirements and achieving effective audit of cross-border transactions. By utilizing a multimodal large model and data analysis tools, the present application can more efficiently process and extract field information in transaction-related materials in different formats (such as pictures, portable documents (PDF), Excel), thereby significantly improving the accuracy and speed of data processing. In addition, this application merges and combines the same type of fields in the extracted multiple fields to form structured information to be reviewed, thereby enhancing the consistency and integrity of the data. Furthermore, these information to be reviewed, review rules, and first prompt words are input into the large language model for review. With the help of the large language model's powerful semantic understanding and logical reasoning capabilities, the accuracy and efficiency of the review results are ensured. This application not only overcomes the limitations of traditional OCR technology in processing complex documents, but also improves the efficiency and reliability of the overall review process through intelligent means.
[0040] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.
[0041] See also Figure 1 , which is a flow chart of a method for reviewing cross-border transactions provided by an embodiment of the present application, such as Figure 1 As shown, the cross-border transaction review method may include steps S101-S104:
[0042] S101: Obtain transaction-related materials for cross-border transactions to be reviewed.
[0043] The review of cross-border transactions first requires obtaining transaction-related materials to be reviewed. These materials are mainly divided into two categories: transaction information and transaction authenticity materials. Transaction information refers to the key transaction element information involved in the cross-border transaction process between the buyer and the seller, such as the name and account number of the payee and payee, the remittance amount, the currency and the transaction subject, etc. This type of data usually exists in a structured format. Common data formats include JSON, XML, CSV or database fields, which are convenient for storing and transmitting clear fields and values. In contrast, transaction authenticity materials cover various trade documents generated by both parties during the transaction process, such as contracts, receipts, customs declarations, invoices and verification forms. These materials mostly exist in the form of unstructured or semi-structured files, including image formats (such as PNG, JPG scans) and PDF files, etc., which may also include image files scanned from paper documents.
[0044] S102: Based on the data format of the data in the transaction-related materials, use a multimodal large model and / or a data analysis tool to perform field extraction processing on the transaction-related materials to obtain multiple field information.
[0045] Typically, text extraction relies heavily on traditional OCR recognition technology. However, due to the complex structure of some image materials, the content extracted by traditional OCR often has structural problems, which seriously affects the subsequent data processing and audit results. To this end, in response to the diverse data formats in transaction-related materials, we first use multimodal large models and / or data analysis tools to perform field extraction processing based on their specific format characteristics. This method is not only compatible with multiple data types such as images, PDFs, and Excel, but can also accurately identify and extract multiple structured field information, ensuring that key transaction elements and information in authenticity certificates are completely and efficiently digitized, thereby providing a solid and reliable data foundation for subsequent compliance audits.
[0046] It should be noted that multimodal large models are advanced AI models capable of processing and understanding multiple data formats (such as images, PDFs, etc.). They can integrate information from different modalities to achieve deep semantic understanding and associative reasoning across formats. Data analysis tools, on the other hand, focus on cleaning, converting, and summarizing structured and semi-structured data (such as Excel), helping to improve data quality and utilization efficiency. The combination of the two not only enhances the processing capabilities of complex heterogeneous data, but also significantly improves the accuracy of field extraction and the automation level of overall review.
[0047] In one possible implementation, the multimodal large-scale model used is Qwen2.5-VL-72B, where "Qwen" represents the model name, "2.5" denotes the version number, "VL" refers to its visual language fusion capabilities, and "72B" reflects its 72 billion parameter size. This model possesses powerful multimodal understanding and information processing capabilities, enabling efficient parsing of diverse data formats within complex transaction materials. It should be noted that while the Qwen2.5-VL-72B model is used in this example, other multimodal large-scale models with similar functionality can be substituted in practice based on specific needs.
[0048] In one possible implementation, data analysis tools may include, but are not limited to, Python's Pandas. Pandas, with its powerful data processing and analysis capabilities, is widely used for cleaning, transforming, and performing statistical calculations on structured data. Pandas enables efficient organization and merging of extracted field information, improving data consistency and operability. Furthermore, other applicable data analysis tools can be flexibly selected based on specific business needs to meet the diverse requirements of different types of data processing.
[0049] S103: Merge and combine the same type of fields on the plurality of field information to obtain information to be reviewed.
[0050] The extracted multiple fields are categorized and merged by field type, and related data within the same category is consolidated to eliminate redundancy and duplication, improving data consistency and integrity. This consolidation and merging operation integrates scattered and diverse field information, creating a clearly structured and comprehensive set of auditable information. This provides an accurate and unified data foundation for subsequent rule audits, thereby improving the efficiency and accuracy of the entire audit process.
[0051] S104: Inputting the information to be reviewed, the review rules and the first prompt word into a large language model to perform rule review on the information to be reviewed, and obtaining a review result of the information to be reviewed.
[0052] In the final stage of cross-border transaction review, the organized information to be reviewed, the preset review rules, and the specially designed first prompt words are all input into the large language model. Leveraging the large language model's powerful semantic understanding and logical reasoning capabilities, a systematic rule-based review of the information to be reviewed is performed. The first prompt word plays a key role. It not only helps the large language model accurately understand the complex and specific review rule requirements, but also guides the model to strictly follow the rule logic for reasoning and judgment, avoiding review biases caused by ambiguity or misunderstanding. In this way, the large language model can comprehensively and accurately analyze whether the information to be reviewed complies with relevant compliance regulations and requirements, and then output review conclusions that meet business needs and regulatory standards. The entire process achieves automated and intelligent rule execution, significantly improving the efficiency and accuracy of reviews while ensuring the credibility and consistency of review results, providing solid and reliable technical support for compliance management of cross-border transactions.
[0053] In one possible implementation, the large language model that can be used is Qwen3-32B, where "Qwen" represents the name of the large language model, "3" represents the version number of the model, reflecting its iterative upgrade in architecture and performance, and "32B" means that the model has 32 billion parameters, indicating that it has strong computing power and deep learning capabilities, and can effectively handle complex natural language understanding and reasoning tasks. Through training on a large amount of diverse data, this model has excellent semantic understanding, logical reasoning and context association capabilities, and is particularly suitable for application scenarios such as rule review that require precise judgment and inference. Although this implementation plan uses the Qwen3-32B large language model, it can also be flexibly replaced with other advanced large language models with similar functions and performance according to actual needs and resource conditions to ensure the overall adaptability and scalability of the system and meet the intelligent review needs in different business environments.
[0054] For example, the first prompt is: Please check the transaction information provided item by item according to the following review rules:
[0055] 1. The payment amount shall not exceed the amount agreed in the contract;
[0056] 2. The remittance currency must be the same as the transaction currency;
[0057] 3. The identities of the payee and payee must match the contract information;
[0058] 4. All required fields must not be empty and must be in the correct format.
[0059] Please perform the audit operation strictly according to the rule logic. If any non-conformity is found, the specific reason should be indicated and the final audit conclusion (compliant or non-compliant) should be given.
[0060] In a possible implementation, the method further includes: converting the audit rules into rules in natural language to obtain structured audit instructions.
[0061] By converting complex audit rules into natural language, not only is the difficulty of coding the rules reduced, but the model's adaptability to rule logic is also enhanced, thereby ensuring the rationality and compliance of the final audit results.
[0062] In a possible implementation, inputting the information to be reviewed, the review rules, and the first prompt word into a large language model to perform rule review on the information to be reviewed, and obtaining a review result of the information to be reviewed, includes:
[0063] Input the information to be reviewed, the structured review instructions and the first prompt word into the large language model to conduct rule review of the information to be reviewed, and obtain the review result of the information to be reviewed
[0064] Based on the content of S101-S104, it can be seen that the relevant materials for the cross-border transaction to be reviewed are first obtained, including transaction information and transaction authenticity materials, covering the transaction elements of both parties and trade documents. Next, with the support of a multimodal large model and / or data analysis tools, these materials are subjected to field extraction processing to obtain multiple fields of information. Subsequently, the multiple fields of information are merged and combined with the same type of fields to form structured information for review. Then, the information for review, the review rules, and the first prompt word are input into the large language model for rule review to ensure the accuracy of the review results. The first prompt word not only guides the large language model to understand the review rules, but also guides the strict implementation of the rule logic and outputs a review conclusion that meets the requirements. This application uses a multimodal large model and data analysis tools to efficiently process transaction-related materials from different formats, achieve accurate field information extraction, and significantly improve the speed and accuracy of data processing. By merging and combining the extracted multiple fields of information with the same type, a structured and consistent information for review is constructed, ensuring the integrity and uniformity of the data. In addition, the information to be reviewed, together with the review rules and the first prompt word, is input into the large language model. Utilizing its powerful semantic understanding and logical reasoning capabilities, the rules are strictly enforced and verified to ensure that the final review results are both accurate and efficient, effectively improving the intelligence level of cross-border transaction compliance review.
[0065] In a possible implementation, the data format of the data in the transaction-related materials includes at least one of a picture format, a PDF format, and an Excel format.
[0066] In one possible implementation, step S102 performs field extraction processing on the transaction-related materials based on the data format of the data in the transaction-related materials using a multimodal large model and / or a data analysis tool to obtain multiple field information, including:
[0067] If the data format of the data in the transaction-related materials is a picture format, the transaction-related materials and the second prompt word are input into the multimodal large model for field extraction processing to obtain the multiple field information.
[0068] Specifically, when the data format of transaction-related materials is in image format, the image data can be input into the multimodal large model together with a specially designed second prompt word for field extraction processing. Since the image format usually contains rich and complex visual and textual information, the role of the second prompt word is particularly critical. It guides the multimodal large model to focus on the key information points in the image, accurately identifying and extracting field content that is closely related to cross-border transaction review. In this way, the multimodal large model can not only overcome the limitations of traditional OCR in processing complex structural images, but also combine visual and language comprehension capabilities to achieve in-depth analysis of implicit information in image materials. Ultimately, the model outputs multiple structured field information, providing a reliable foundation for subsequent data integration and rule review, thereby significantly improving the efficiency and accuracy of the entire cross-border transaction review process.
[0069] For example, the second prompt reads: Please focus on identifying and extracting the following fields in the image: payee and payee name, account number, remittance amount, currency, transaction subject, contract number, invoice number, date, and signatures of both parties. Ignore irrelevant background and watermarks, and ensure that the fields are complete and formatted correctly.
[0070] In one possible implementation, step S102 performs field extraction processing on the transaction-related materials based on the data format of the data in the transaction-related materials using a multimodal large model and / or a data analysis tool to obtain multiple field information, including:
[0071] If the data format of the transaction-related materials is PDF, the transaction-related materials are split into multiple images to obtain multiple image files. The multiple image files and the second prompt word are then input into the multimodal large model for field extraction to obtain the multiple field information.
[0072] Specifically, when transaction-related materials are in PDF format, the system first splits the PDF document into multiple single-page image files to better accommodate the multimodal model's image data processing capabilities. Each split image file contains complete visual information about the corresponding page in the PDF. These image files, along with a specially designed second prompt, are then fed into the multimodal model for field extraction. This second prompt plays a crucial guiding role, enabling the model to focus on key information in the image that is closely related to cross-border transaction audits, such as the contract number, payee and payee names, remittance amount, and transaction subject matter, thereby accurately identifying and extracting multiple structured fields. By first converting the PDF into images and then incorporating the second prompt, the multimodal model effectively overcomes the challenges of complex and diverse text layouts in PDFs, achieving a deep understanding and precise extraction of the rich content within transaction authenticity materials. This ensures the integrity and accuracy of the data required for subsequent audits, significantly improving the efficiency and quality of the entire cross-border transaction compliance audit.
[0073] For example, the second prompt reads: Please carefully identify and extract the following key information fields from each image: contract number, payee and payee name and account number, remittance amount, currency, transaction subject, invoice number, date, and signatures of both parties. Ignore watermarks, borders, and irrelevant background information, ensuring the extracted content is complete and formatted correctly.
[0074] In one possible implementation, step S102 performs field extraction processing on the transaction-related materials based on the data format of the data in the transaction-related materials using a multimodal large model and / or a data analysis tool to obtain multiple field information, including:
[0075] If the data format of the data in the transaction-related materials is Excel format, the transaction-related materials are subjected to field extraction processing by the data analysis tool to obtain the plurality of field information.
[0076] Specifically, when the data format of transaction-related materials is in Excel format, since Excel files themselves have good structural characteristics and tabular formats, data analysis tools can be used directly to extract fields from them. Through powerful data analysis libraries such as Python's Pandas, it is possible to efficiently read various types of tabular data in Excel, accurately locate and extract field information containing key transaction elements, such as the name of the payee and payee, account number, remittance amount, currency, and transaction subject, etc. This process can not only automatically identify the table header and the corresponding data columns, but also flexibly process different worksheets or multiple data structures, thereby ensuring that the extracted field information is complete and formatted in a standardized manner. Compared with the complex parsing of unstructured data, the use of data analysis tools to directly operate Excel format materials greatly improves the efficiency and accuracy of field extraction, providing a solid and reliable data foundation for subsequent review links.
[0077] In a possible implementation, when the data format of the transaction-related materials includes an image format and / or a PDF format, step S103 performs a merging process on the multiple fields of information of the same type to obtain information to be reviewed, including:
[0078] The multiple field information and the third prompt word are input into the large language model for structured integration processing to obtain the information to be reviewed; the third prompt word is used to guide the multimodal large model to accurately merge field information of the same type in the multiple field information.
[0079] Specifically, when the data format of transaction-related materials includes image format and / or PDF format, since multiple fields of information extracted from these unstructured or semi-structured data often have redundancy, duplication, or even format inconsistencies, these fields must be merged and combined of the same type to form a unified and clearly structured information to be reviewed. Therefore, in step S103, the extracted multiple field information can be directly input into the large language model along with a specially designed third prompt word. The model's powerful semantic understanding and information integration capabilities can be utilized to accurately identify and reasonably merge fields of the same type. The third prompt word plays a key role in this process, guiding the model to focus on the relevance and consistency between field types, avoiding information omissions or incorrect merging, thereby ensuring that the merged field data is both complete and standardized. Through this structured integration process, not only are the problems of fragmented and duplicated original field information resolved, but the consistency and availability of the data are also greatly improved, providing a reliable and high-quality basis for the information to be reviewed for subsequent rule review.
[0080] For example, the third prompt is: Please merge the same type of fields in the input, such as the same payee and payee name, account number or transaction amount, etc., to ensure that the merged field content is complete and non-repetitive; for different data, please mark all different versions and explain the possible reasons; for non-existent fields, please do not make up and directly return empty; the result is returned in json format.
[0081] In a possible implementation, when the data format of the transaction-related materials includes image format + Excel format, image format + PDF format + Excel format, or PDF format + Excel format, step S103 merges and combines the same type of fields on the multiple fields to obtain information to be reviewed, including:
[0082] Multiple fields of information extracted from the transaction-related materials in image format and / or PDF format and the third prompt word are input into a large language model for structural integration to obtain first preliminary information to be reviewed. Multiple fields of information extracted from the transaction-related materials in Excel format are merged and combined for fields of the same type to obtain second preliminary information to be reviewed. The first preliminary information to be reviewed and the second preliminary information to be reviewed are then merged and combined for fields of the same type to obtain the information to be reviewed.
[0083] Specifically, when the data format combinations of transaction-related materials include image and Excel formats, image format plus PDF format plus Excel format, or a combination of PDF and Excel formats, step S103 requires merging and consolidating the multiple fields extracted from the different formats for fields of the same type to obtain a unified and structured information to be reviewed. In this case, the multiple fields extracted from the image and / or PDF format transaction materials, along with the third prompt word, are first input into the large language model. Leveraging its powerful semantic understanding and integration capabilities, this unstructured or semi-structured data is structured and consolidated, thereby generating the first preliminary information to be reviewed. Simultaneously, for transaction-related materials in Excel format, due to their inherently well-structured nature, data analysis tools can be used to directly merge and consolidate the extracted fields for fields of the same type, generating the second preliminary information to be reviewed. Subsequently, these two parts of preliminary information to be reviewed are further merged for fields of the same type to eliminate duplication and inconsistencies, unify the data format and content, and ultimately generate a complete and high-quality information to be reviewed. The third prompt word in this process plays a key role in guiding the large language model to accurately identify and merge fields of the same type, ensuring that the merging operation is both comprehensive and detailed, avoiding information omissions or incorrect merging, and thus providing a solid data foundation for subsequent review steps.
[0084] For example, the third prompt is: Please merge the same type of fields in the input, such as the same or equivalent payee and payee names, account numbers, transaction amounts and dates, etc., to ensure that the merged field content is complete and non-repetitive; if there are differences in field content, please keep all different versions and indicate the source format (such as picture, PDF or Excel); please unify the field naming conventions and data formats to ensure that the output results are clearly structured, standardized and suitable for subsequent review; exclude obvious errors or invalid information; please do not make up for non-existent fields and return empty directly; the results are returned in json format.
[0085] See also Figure 2 , Figure 2 This is a schematic diagram of the structure of a cross-border transaction audit device provided in an embodiment of the present application. Figure 2 As shown, the cross-border transaction audit device includes:
[0086] An acquisition unit 201 is configured to acquire transaction-related materials for a cross-border transaction to be reviewed; the transaction-related materials include transaction information and transaction authenticity materials; the transaction information includes transaction element information between the buyer and the seller during the cross-border transaction; the transaction authenticity materials include trade vouchers generated by the buyer and the seller during the cross-border transaction;
[0087] A field extraction unit 202 is configured to extract fields from the transaction-related materials using a multimodal large model and / or a data analysis tool based on the data format of the data in the transaction-related materials to obtain a plurality of field information;
[0088] A merging unit 203 is configured to merge the same type of fields on the plurality of field information to obtain information to be reviewed;
[0089] The input unit 204 is used to input the information to be reviewed, the review rules and the first prompt word into the large language model to perform rule review of the information to be reviewed, and obtain the review result of the information to be reviewed; the first prompt word is used to guide the large language model to accurately understand the review rules, strictly execute the rule logic, and output a review conclusion that meets the requirements.
[0090] In a possible implementation, the data format of the data in the transaction-related materials includes at least one of a picture format, a portable document PDF format, and an Excel format.
[0091] In a possible implementation, the field extraction unit 202 is specifically configured to:
[0092] If the data format of the data in the transaction-related materials is a picture format, the transaction-related materials and the second prompt word are input into the multimodal large model for field extraction processing to obtain the multiple field information; the second prompt word is used to guide the multimodal large model to accurately extract key information from the transaction-related materials.
[0093] In a possible implementation, the field extraction unit 202 is specifically configured to:
[0094] If the data format of the transaction-related materials is PDF format, splitting the transaction-related materials into multiple images to obtain multiple image files;
[0095] The multiple image files and the second prompt word are input into the multimodal large model for field extraction processing to obtain the multiple field information; the second prompt word is used to guide the multimodal large model to accurately extract key information from the transaction-related materials.
[0096] In a possible implementation, the field extraction unit 202 is specifically configured to:
[0097] If the data format of the data in the transaction-related materials is Excel format, the transaction-related materials are subjected to field extraction processing by the data analysis tool to obtain the plurality of field information.
[0098] In a possible implementation, when the data format of the transaction-related materials includes an image format and / or a PDF format, the merging unit 203 is specifically configured to:
[0099] The multiple field information and the third prompt word are input into the large language model for structured integration processing to obtain the information to be reviewed; the third prompt word is used to guide the multimodal large model to accurately merge field information of the same type in the multiple field information.
[0100] In a possible implementation, when the data format of the transaction-related materials includes picture format + Excel format, picture format + PDF format + Excel format, or PDF format + Excel format, the merging unit 203 is specifically configured to:
[0101] Inputting multiple fields of information extracted from the transaction-related materials in image format and / or PDF format and the third prompt word into the large language model for structured integration processing to obtain first preliminary information to be reviewed, and merging and merging multiple fields of information extracted from the transaction-related materials in Excel format for fields of the same type to obtain second preliminary information to be reviewed; the third prompt word is used to guide the multimodal large model to accurately merge field information of the same type among the multiple fields of information;
[0102] The first preliminary information to be reviewed and the second preliminary information to be reviewed are merged and processed by merging the same type of fields to obtain the information to be reviewed.
[0103] In addition, an embodiment of the present application also provides an audit device for cross-border transactions, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the cross-border transaction audit method described above is implemented.
[0104] In addition, an embodiment of the present application also provides a computer-readable storage medium, which stores instructions. When the instructions are executed on a terminal device, the terminal device executes the cross-border transaction review method as described above.
[0105] The present embodiment provides a cross-border transaction audit device. First, an acquisition unit 201 acquires transaction-related materials for a cross-border transaction to be audited. A field extraction unit 202, based on the data format of the transaction-related materials, uses a multimodal large model and / or data analysis tools to perform field extraction on the transaction-related materials, obtaining multiple fields of information. A merging unit 203 then merges and combines the multiple fields of information into similar categories to obtain information to be audited. This allows an input unit 204 to input the information to be audited, audit rules, and a first prompt word into a large language model for rule audit of the information to be audited, thereby obtaining an audit result for the information to be audited. This application utilizes a multimodal large model and data analysis tools to effectively process and extract field information from transaction-related materials in different formats, improving data processing efficiency and accuracy. Simultaneously, the extracted multiple fields of information are integrated and similar fields are merged to generate structured information to be audited, enhancing data consistency and integrity. Furthermore, by inputting the information to be audited, audit rules, and first prompt word into the large language model for rule audit, the semantic understanding and logical reasoning capabilities of the large language model are leveraged to ensure the accuracy and efficiency of the audit results.
[0106] The above is a detailed introduction to the cross-border transaction audit method, device, equipment and storage medium provided by the present application. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of this application, several improvements and modifications can be made to this application, and these improvements and modifications also fall within the scope of protection of the claims of this application.
[0107] It should be understood that in this application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.
[0108] It should also be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.
Claims
1. A cross-border transaction audit method, characterized in that: The method comprises: Obtain transaction-related materials for the cross-border transaction to be reviewed; the transaction-related materials include transaction information and transaction authenticity materials; the transaction information includes transaction element information between the buyer and the seller during the cross-border transaction; the transaction authenticity materials include trade certificates generated by the buyer and the seller during the cross-border transaction; Based on the data format of the data in the transaction-related materials, using a multimodal large model and / or a data analysis tool to perform field extraction processing on the transaction-related materials to obtain multiple field information; Merging and merging the same type of fields on the multiple fields to obtain information to be reviewed; The information to be reviewed, the review rules and the first prompt word are input into the large language model to conduct rule review of the information to be reviewed, and the review result of the information to be reviewed is obtained; the first prompt word is used to guide the large language model to accurately understand the review rules, strictly execute the rule logic, and output a review conclusion that meets the requirements.
2. The method according to claim 1, characterized in that The data format of the data in the transaction-related materials includes at least one of a picture format, a portable document PDF format and an Excel format.
3. The method according to claim 2, characterized in that Based on the data format of the data in the transaction-related materials, the multimodal large model and / or data analysis tool is used to perform field extraction processing on the transaction-related materials to obtain multiple field information, including: If the data format of the data in the transaction-related materials is a picture format, the transaction-related materials and the second prompt word are input into the multimodal large model for field extraction processing to obtain the multiple field information; the second prompt word is used to guide the multimodal large model to accurately extract key information from the transaction-related materials.
4. The method according to claim 2, characterized in that Based on the data format of the data in the transaction-related materials, the multimodal large model and / or data analysis tool is used to perform field extraction processing on the transaction-related materials to obtain multiple field information, including: If the data format of the transaction-related materials is PDF format, splitting the transaction-related materials into multiple images to obtain multiple image files; The multiple image files and the second prompt word are input into the multimodal large model for field extraction processing to obtain the multiple field information; the second prompt word is used to guide the multimodal large model to accurately extract key information from the transaction-related materials.
5. The method according to claim 2, characterized in that Based on the data format of the data in the transaction-related materials, the multimodal large model and / or data analysis tool is used to perform field extraction processing on the transaction-related materials to obtain multiple field information, including: If the data format of the data in the transaction-related materials is Excel format, the transaction-related materials are subjected to field extraction processing by the data analysis tool to obtain the plurality of field information.
6. The method according to claim 2, characterized in that When the data format of the transaction-related materials includes image format and / or PDF format, the multiple fields of information are merged and combined into the same type of fields to obtain the information to be reviewed, including: The multiple field information and the third prompt word are input into the large language model for structured integration processing to obtain the information to be reviewed; the third prompt word is used to guide the multimodal large model to accurately merge field information of the same type in the multiple field information.
7. The method according to claim 2, characterized in that When the data format of the transaction-related materials includes picture format + Excel format, picture format + PDF format + Excel format, or PDF format + Excel format, the multiple fields of information are merged and combined into the same type of fields to obtain the information to be reviewed, including: Inputting multiple fields of information extracted from the transaction-related materials in image format and / or PDF format and the third prompt word into the large language model for structured integration processing to obtain first preliminary information to be reviewed, and merging and merging multiple fields of information extracted from the transaction-related materials in Excel format for fields of the same type to obtain second preliminary information to be reviewed; the third prompt word is used to guide the multimodal large model to accurately merge field information of the same type among the multiple fields of information; The first preliminary information to be reviewed and the second preliminary information to be reviewed are merged and processed by merging the same type of fields to obtain the information to be reviewed.
8. A cross-border transaction audit device, characterized in that: The device comprises: An acquisition unit, configured to acquire transaction-related materials for a cross-border transaction to be reviewed; the transaction-related materials include transaction information and transaction authenticity materials; the transaction information includes transaction element information between the buyer and the seller during the cross-border transaction; the transaction authenticity materials include trade vouchers generated by the buyer and the seller during the cross-border transaction; a field extraction unit configured to extract fields from the transaction-related materials using a multimodal large model and / or a data analysis tool based on the data format of the data in the transaction-related materials to obtain a plurality of field information; A merging unit, configured to merge the same type of fields on the plurality of field information to obtain information to be reviewed; The input unit is used to input the information to be reviewed, the review rules and the first prompt word into the large language model to perform rule review of the information to be reviewed, and obtain the review result of the information to be reviewed; the first prompt word is used to guide the large language model to accurately understand the review rules, strictly execute the rule logic, and output a review conclusion that meets the requirements.
9. A cross-border transaction audit device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the cross-border transaction audit method according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed on the terminal device, the terminal device executes the cross-border transaction review method according to any one of claims 1 to 7.