Multi-country compliance intelligent auditing method, device and equipment for cross-border e-commerce and medium

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a multi-country compliance intelligent auditing method, device and equipment for cross-border e-commerce and a medium, and the method comprises the steps: obtaining the law and regulation revision frequency of each specified trade country; setting a regulation capturing period for a corresponding specified trade country according to the regulation revision frequency; based on the law and regulation capturing period of each specified trade country, obtaining a law and regulation text of each specified trade country; analyzing the law and regulation text by utilizing the semantic analysis capability of a preset large language model to obtain analysis content; constructing a structured knowledge graph based on the analysis content; and obtaining a trade text, and checking the compliance of the trade text based on the structured knowledge graph. The method has the beneficial effects that the method is helpful to quickly adapt to the regulation change of the global market, the violation risk is reduced, and the competitiveness and market adaptability of enterprises are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-country compliance intelligent auditing of cross-border e-commerce, and in particular to a multi-country compliance intelligent auditing method, device, equipment and medium for cross-border e-commerce. BACKGROUND

[0002] Under the background of globalization trade, the frequent changes of laws and regulations of various countries have a significant impact on international business activities. Especially in the field of commodity compliance and energy efficiency standards, timely tracking and understanding the dynamic adjustment of these regulations is crucial for market access and compliance of enterprises. However, the traditional manually maintained regulation database has a long update cycle of 3 to 6 months, making it difficult for enterprises to keep abreast of the latest regulations of various countries in real time. This lag leads to a high compliance risk for enterprises operating in the global market. For example, due to the implementation of the new EU energy efficiency label regulation in 2023, a large number of imported goods were returned due to inconsistent information. This not only affects the economic benefits of enterprises, but also causes damage to brand reputation, further exacerbating the disadvantage of enterprises in market competition. At the same time, the complexity and diversity of regulations force enterprises to invest more manpower and resources to deal with compliance audits, increasing operating costs. SUMMARY

[0003] Therefore, it is necessary to propose a multi-country compliance intelligent auditing method, device, equipment and medium for cross-border e-commerce to solve the existing problems of multi-country compliance intelligent auditing of cross-border e-commerce.

[0004] A multi-country compliance intelligent auditing method for cross-border e-commerce, the method comprising:

[0005] obtaining the regulation amendment frequency of each specified trade country;

[0006] setting a regulation crawling period for the corresponding specified trade country according to the regulation amendment frequency;

[0007] obtaining the regulation text of each specified trade country based on the regulation crawling period of each specified trade country;

[0008] using the semantic parsing capability of a preset large language model to parse the regulation text and obtain parsed content;

[0009] constructing a structured knowledge graph based on the parsed content;

[0010] obtaining trade text and auditing the compliance of the trade text based on the structured knowledge graph.

[0011] Further, the step of obtaining trade text and auditing the compliance of the trade text based on the structured knowledge graph comprises:

[0012] Obtaining image information of each to-be-traded commodity;

[0013] Extracting features of each image information through a preset image encoder to obtain a label feature map corresponding to each image information respectively;

[0014] Text recognition is performed on each label feature map based on a preset text encoder to obtain a text feature corresponding to each label feature map;

[0015] The trade text is generated based on all the text features;

[0016] The compliance of the trade text is audited based on the structured knowledge graph.

[0017] Further, the step of generating the trade text based on all the text features comprises:

[0018] Extracting an image feature corresponding to each label feature map;

[0019] Processing the image feature and the text feature corresponding to each label feature map through a double-channel attention mechanism to obtain standard text information corresponding to each label feature map;

[0020] The standard text information corresponding to each label feature map is integrated to obtain the trade text.

[0021] Further, before the step of extracting features of each image information through a preset image encoder to obtain a label feature map corresponding to each image information respectively, the method further comprises:

[0022] Obtaining multiple cross-border commodity real object images;

[0023] Recognizing a label area in the commodity real object image;

[0024] Marking a cross-border commodity real object image with a label area smaller than a preset value as a target cross-border commodity real object image;

[0025] Performing domain adaptation training on the target cross-border commodity real object image to obtain the preset image encoder.

[0026] Further, the step of auditing the compliance of the trade text based on the structured knowledge graph comprises:

[0027] Extracting element information in the trade text;

[0028] Mapping the element information with nodes in the structured knowledge graph to extract corresponding compliance standards;

[0029] The compliance standards and the element information are input into a preset audit model to audit the compliance of the trade text.

[0030] Furthermore, before the step of inputting the compliance standards and the element information into a preset audit model to audit the compliance of the trade document, the step further includes:

[0031] Extracting first regional feature information from the element information;

[0032] Calculating similarity between the first regional feature information and a plurality of pre-stored second regional feature information;

[0033] Recording the second regional feature information with the greatest similarity as the target second regional feature information;

[0034] A corresponding preset audit model is obtained based on the target second regional characteristic information; wherein each second regional characteristic information corresponds to an audit model, and each audit model is trained in a labeled training method through corresponding element information and its compliance.

[0035] Furthermore, after the step of inputting the compliance standards and the element information into a preset audit model to audit the compliance of the trade document, the following steps are further included:

[0036] Obtain the predicted audit results and actual audit results of multiple elements of information;

[0037] Calculating Mahalanobis distance based on the predicted audit result and the actual audit result;

[0038] The feature information whose Mahalanobis distance is greater than the set threshold is recorded as differentiated feature information;

[0039] Based on the differentiated factor information and the corresponding actual audit results, the preset audit model is incrementally learned and trained using an elastic weight solidification algorithm to obtain a new preset audit model.

[0040] A multi-national compliance intelligent audit device for cross-border e-commerce, comprising:

[0041] The first acquisition module is used to obtain the frequency of regulatory revisions of each designated trading country;

[0042] A setting module is used to set a regulatory crawling cycle for a corresponding designated trading country according to the regulatory revision frequency;

[0043] A second acquisition module is configured to acquire the regulatory text of each designated trading country based on the regulatory capture cycle of each designated trading country;

[0044] The parsing module is configured to parse the regulation text by using semantic parsing capability of a preset large language model to obtain parsed content.

[0045] The construction module is configured to construct a structured knowledge graph based on the parsed content.

[0046] The auditing module is configured to obtain trade text and audit compliance of the trade text based on the structured knowledge graph.

[0047] A computer device includes a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the following steps:

[0048] Obtain the regulation amendment frequency of each specified trade country;

[0049] Set a regulation crawling period for the corresponding specified trade country according to the regulation amendment frequency;

[0050] Obtain the regulation text of each specified trade country based on the regulation crawling period of each specified trade country;

[0051] Parse the regulation text by using semantic parsing capability of a preset large language model to obtain parsed content;

[0052] Construct a structured knowledge graph based on the parsed content;

[0053] Obtain trade text and audit compliance of the trade text based on the structured knowledge graph.

[0054] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to make the processor execute the following steps:

[0055] Obtain the regulation amendment frequency of each specified trade country;

[0056] Set a regulation crawling period for the corresponding specified trade country according to the regulation amendment frequency;

[0057] Obtain the regulation text of each specified trade country based on the regulation crawling period of each specified trade country;

[0058] Parse the regulation text by using semantic parsing capability of a preset large language model to obtain parsed content;

[0059] Construct a structured knowledge graph based on the parsed content;

[0060] Obtain trade text and audit compliance of the trade text based on the structured knowledge graph.

[0061] The beneficial effects of the present application: by acquiring the frequency of regulations revision of each designated trading country in real time, and setting a dynamic regulation scraping cycle according to this information, enterprises can timely grasp the changes in regulations, enabling enterprises to realize real-time tracking and intelligent management of compliance information, providing a strong support for cross-border e-commerce enterprises, helping to quickly adapt to the changes in global market regulations, reducing the risk of violation, and improving the competitiveness and market adaptability of enterprises. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0063] Among them:

[0064] Figure 1 It is an application environment diagram of the multi-country compliance intelligent auditing method for cross-border e-commerce in an embodiment.

[0065] Figure 2 It is a flowchart of the multi-country compliance intelligent auditing method for cross-border e-commerce in an embodiment.

[0066] Figure 3 It is a structural block diagram of the multi-country compliance intelligent auditing device for cross-border e-commerce in an embodiment.

[0067] Figure 4 It is a structural block diagram of the computer device in an embodiment. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0069] Figure 1 It is an application environment diagram of the multi-country compliance intelligent auditing for cross-border e-commerce in an embodiment. Refer to Figure 1The cross-border e-commerce multi-country compliance intelligent auditing method is applied to a cross-border e-commerce multi-country compliance intelligent auditing system. The cross-border e-commerce multi-country compliance intelligent auditing system comprises a terminal 110 and a server 120. The terminal 110 and the server 120 are connected through a network. The terminal 110 can be a desktop terminal or a mobile terminal, and the mobile terminal can be at least one of a mobile phone, a tablet computer, a notebook computer, etc. The server 120 can be implemented by an independent server or a server cluster composed of multiple servers. The terminal 110 is used to collect the regulation revision frequency of each specified trade country, and the server 120 is used for compliance auditing.

[0070] As shown in Figure 2 In one embodiment, a cross-border e-commerce multi-country compliance intelligent auditing method is provided. The method can be applied to a terminal or a server. In this embodiment, the method is applied to a server. The cross-border e-commerce multi-country compliance intelligent auditing method specifically comprises the following steps:

[0071] S1: Obtain the regulation revision frequency of each specified trade country;

[0072] S2: Set a regulation crawling period for the corresponding specified trade country according to the regulation revision frequency;

[0073] S3: Obtain the regulation text of each specified trade country based on the regulation crawling period of each specified trade country;

[0074] S4: Analyze the regulation text using the semantic analysis capability of a preset large language model to obtain analysis content;

[0075] S5: Construct a structured knowledge graph based on the analysis content;

[0076] S6: Obtain a trade text, and audit the compliance of the trade text based on the structured knowledge graph.

[0077] As described in step S1 above, the frequency of regulations revision of each designated trading country is obtained. In the global cross-border e-commerce environment, frequent updates of regulations directly affect the compliance of products. Without sufficient understanding of the changes in regulations of each country, enterprises may face the risk of compliance failure. Therefore, first of all, it is necessary to establish an effective data collection mechanism to regularly monitor and obtain the regulation revision information of each designated trading country. This includes accessing the official website of the regulatory agency of each country, industry announcements, regulation change documents and other sources, collecting relevant information and compiling the historical revision record of the regulations and its frequency. It should be noted that the revision frequency here is a general frequency, for example, the average period of modification can be obtained from multiple modifications, so as to obtain the corresponding regulation revision frequency. This can help enterprises identify which countries have a high regulation update frequency, providing data support for the development of enterprise compliance strategies. Through the analysis of the regulation revision frequency, enterprises can optimize compliance plans and allocate resources to ensure timely response to regulations.

[0078] As described in step S2 above, the regulation scraping cycle is set for the corresponding designated trading country according to the regulation revision frequency. According to the previously obtained regulation revision frequency, the enterprise needs to set the corresponding regulation scraping cycle for each designated trading country. The regulation revision frequency of different countries is significantly different, so it is necessary to set a suitable scraping cycle. In countries with high frequency of regulation revision, the enterprise may need to adopt a daily or weekly scraping strategy to ensure timely access to the latest regulation information; while for countries with low regulation revision frequency, the enterprise can choose to scrape every month or every quarter to save resources and optimize efficiency. This differentiated scraping cycle setting enables enterprises to flexibly respond to the compliance requirements of different markets and reduce the risks caused by information lag. At the same time, enterprises can also use data analysis tools to regularly evaluate and adjust these scraping cycles to adapt to market and regulation changes. This flexible and dynamic regulation scraping strategy provides a more accurate basis for subsequent audit work, promoting the intelligent process of enterprise compliance management.

[0079] As described in step S3 above, the regulations text of each designated trade country is obtained based on the regulation crawling cycle of each designated trade country. After setting the regulation crawling cycle, the regulation text of each designated trade country is obtained according to the set cycle within a specified time interval. This process usually involves automated web crawler technology. Enterprises can configure the crawler program to automatically access the official websites of regulatory agencies and other relevant information sources in different countries, and periodically download and update the regulation text. For regulation texts in different countries, there may be differences in format and language. Enterprises need to design the text processing process in advance to uniformly process regulation texts from different sources. In addition, attention should be paid to the crawling policy of each country's website during the crawling process to avoid legal risks caused by frequent access. By centrally storing and managing these regulation text data, enterprises can build a dynamic regulation database, which lays a solid foundation for subsequent compliance audits.

[0080] As described in step S4 above, the regulation text is parsed using the semantic parsing capability of a pre-set large language model to obtain parsed content. After obtaining the regulation text, a pre-set large language model (such as BERT, GPT, etc.) is used for semantic parsing. These models have powerful natural language processing capabilities and can understand and analyze text in depth. Through the parsing of the regulation text, the model can extract key compliance elements such as applicable regulation clauses, compliance requirements, and restrictive clauses. The present application preferably uses an LLM model (Large Language Model). The specific parsing process includes inputting the regulation text into the large language model for processing. The model will generate corresponding structured output, including the definition of legal terms, clauses, and elements, and convert unstructured regulation text into the following element information: regulatory subject (country / region), applicable commodity category (HS code mapping), mandatory compliance elements (such as CE certification mark), and associated constraints (such as battery products must comply with transportation safety and environmental protection regulations).

[0081] As described in step S5 above, a structured knowledge graph is constructed based on the parsed content. After parsing the regulation text, the parsed content is converted into a structured knowledge graph. The knowledge graph is a dynamic and rich knowledge base that structures information such as legal clauses, compliance elements, and regulatory units through nodes (representing entities) and edges (representing relationships). Node types include the following types:

[0082] Commodity node: represents a specific commodity category or model, such as electronic products, food, cosmetics, etc.

[0083] Regulation node: represents regulations and standards in different countries, such as CE certification, FDA standards, REACH regulations, etc.

[0084] Country / Region Node: Represents the relevant country or region of a transaction, such as the United States, the European Union, China, Southeast Asia, etc.

[0085] Compliance Element Node: Represents specific requirements needed for the compliance of goods, such as test reports (e.g. UN38.3), compliance documents (e.g. MSDS), and certificates of origin, etc.

[0086] Regulatory Agency Node: Represents relevant regulatory agencies or organizations, such as the FDA (United States), CE Marking Agency (European Union), and customs, etc.

[0087] Market Node: Represents different market characteristics or transaction environments, such as the Middle East market, the North American market, etc.

[0088] Edge Types include the following types:

[0089] "Applicable" Relationship: Represents the relationship between regulations and specific goods, for example, a certain regulation applies to a specific category of goods (such as "CE certification" - "electronic products").

[0090] "Requirement" Relationship: Represents the compliance elements that goods or regulations need to meet, for example, "lithium batteries" - "need UN38.3 test report".

[0091] "Regulation" Relationship: Represents the relationship between goods or specific compliance elements and regulatory agencies, for example, "FDA" - "regulates food and drugs".

[0092] "Located" Relationship: Represents the relationship between the origin of goods and the country, for example, "a certain brand of food" - "originated in China".

[0093] "Correlation" Relationship: Represents the correlation between different regulations, for example, "REACH regulation" - "correlated to environmental compliance requirements".

[0094] "Belongs to" Relationship: Represents the belonging relationship of goods to the market, for example, "a certain electronic product" - "belongs to the North American market".

[0095] When building a knowledge graph, you can use a graph database (such as Neo4j) to store the parsed data, and design appropriate query mechanisms to help compliance auditors quickly obtain the required information through specific queries. In addition, the knowledge graph can be dynamically updated according to usage, adding new knowledge and marking the update history of legal provisions in a timely manner, so as to maintain the timeliness and accuracy of the knowledge graph.

[0096] As described in step S6 above, the trade text is obtained and its compliance is audited based on the structured knowledge graph. The trade text to be audited is obtained and its compliance is checked based on the structured knowledge graph constructed. First, the trade file to be audited is extracted from the data source, which may include relevant texts such as commodity description, business contract, certificate of origin, etc. Next, through the associated query with the knowledge graph, the auditing system will focus on checking the compliance between these trade texts and the corresponding regulations. The auditing process involves checking against the compliance standards and related requirements in the knowledge graph, including applicable regulatory provisions, mandatory compliance documents, and restrictions, etc. For example, the system will verify whether the specific commodity meets the safety standards and document requirements of the importing country, ensuring that all compliance elements are fully met. Based on the results of the audit, the system will generate a compliance report and point out potential compliance risks and recommended improvement measures. This process not only improves the efficiency of compliance auditing, but also significantly reduces the market risks caused by compliance problems, promoting the healthy development of cross-border e-commerce.

[0097] In one embodiment, the step S6 of obtaining trade text and auditing the compliance of the trade text based on the structured knowledge graph comprises:

[0098] S601: Obtain image information of each trade commodity;

[0099] S602: Extract features from each image information through a preset image encoder to obtain a label feature map corresponding to each image information;

[0100] S603: Perform text recognition on each label feature map based on a preset text encoder to obtain a text feature corresponding to each label feature map;

[0101] S604: Generate the trade text based on all the text features;

[0102] S605: Audit the compliance of the trade text based on the structured knowledge graph.

[0103] As described in steps S601-S605 above, during the process of auditing the compliance of trade commodities, trade texts are obtained by scanning various products. Specifically, image information of each trade commodity is obtained by taking pictures or scanning the physical images of the goods. These images can be product display images, including front, side, back and detailed close-up shots, to ensure that all product information and label content are captured. In actual operation, enterprises may use automated equipment or mobile devices to obtain images of goods, which can improve the speed and consistency of image acquisition. During the acquisition process, reasonable lighting and shooting angles should be paid attention to to ensure that all information on the label (such as barcodes, compliance logos, product descriptions, etc.) is clearly visible.

[0104] After obtaining the image information of the goods to be audited, the image encoder is used to extract the features of each image. The image encoder is usually a deep learning model based on convolutional neural network (CNN), such as EfficientNet, ResNet, etc. These models have been trained to effectively extract important features from images. The present application preferably uses EfficientNet-V2 network for feature extraction. The process of feature extraction involves inputting the original image into the model, which will gradually refine the high-level features of the image through multiple convolutional layers and pooling layers. These extracted features usually include color, shape, texture, etc. information, and finally generate a label feature map corresponding to each trade commodity. These label feature maps not only retain the key information of the commodity image, but also lay the foundation for subsequent text recognition and compliance audit.

[0105] The text encoder is trained through structured neural networks such as long short-term memory network (LSTM), Transformer, etc., aiming to identify and extract text information from the label feature map, including product name, specifications, compliance logo, bar code, etc. Preferably, DeBERTa-v3 model is used to enhance multilingual understanding through adversarial training, so that different language text information can be identified.

[0106] Specifically, after the label feature map is input into the text encoder, the network will identify each character, word, and phrase in the image to generate corresponding text features. This process usually involves multiple steps such as character segmentation, feature extraction, and sequence modeling. Since the information in the label is usually presented in different fonts, sizes, and directions, the text encoder needs to have strong robustness to ensure accurate text recognition under different environmental conditions (such as light, background noise, etc.).

[0107] After completing text recognition, the final trade text is generated based on all extracted text features. This process usually involves organizing and structuring text features extracted from multiple commodity labels to form a complete written document. The process of generating trade text may include multiple sub-steps. First, normalize the text features to ensure that text information from different sources is unified into the same format. Then, the system will organize the text features according to the pre-set template or structure in a specific logical order, such as grouping and organizing according to product information, compliance logo, applicable regulations, etc. modules. Thus, it provides the necessary information basis for subsequent compliance audit. At the same time, the generated text can also be used as a basis for document archiving and future analysis, playing a core role in the entire process.

[0108] Based on the previously constructed structured knowledge graph, the generated trade text is audited for compliance. Relevant regulatory information and compliance elements are queried from the knowledge graph. This can be achieved through structured query language (such as SPARQL), and the system retrieves all legal provisions and standards related to the type of goods, market, and compliance requirements. Next, the audit system compares the generated trade text and identifies compliance issues, such as missing necessary compliance proof documents, non-compliance with regulatory requirements, and incorrect country of origin information. Through the audit process, a compliance audit result is generated, including a compliance score, potential risk identification, and improvement suggestions. This intelligent audit method based on knowledge graph not only improves the accuracy and efficiency of compliance, but also reduces the risk of errors caused by manual auditing, helping enterprises quickly adapt to market demand and regulatory changes. Ultimately, it provides strong compliance protection for cross-border trade of enterprises, promoting their healthy and sustainable development.

[0109] In one embodiment, the step S604 of generating the trade text based on all the text features comprises:

[0110] S6041: extracting the corresponding image features from each label feature map;

[0111] S6042: processing the image features and text features corresponding to each label feature map through a dual-channel attention mechanism to obtain standard text information corresponding to each label feature map;

[0112] S6043: synthesizing the standard text information corresponding to each label feature map to obtain the trade text.

[0113] As described in steps S6041-S6043 above, first, the corresponding image features are extracted from each label feature map. Image features are obtained through the previous image encoder, which contains key information reflecting the layout characteristics of the text in the product label. Then, the dual-channel attention mechanism is used to process the previously extracted image features and text features. The dual-channel attention mechanism is a reinforcement learning model that helps to strengthen the relevance between image data and text data. It mainly consists of two independent channels: one channel processes image features, and the other channel processes text features to achieve forced semantic space alignment of visual features and text descriptions, obtaining standard text information corresponding to each label feature map. Depending on the preset text template or standard format, the previously obtained standard text information is filled into the template to form the final trade text.

[0114] In one embodiment, before the step S602 of extracting features from each image information through a preset image encoder to obtain a label feature map corresponding to each image information, the method further comprises:

[0115] S6011: Obtain multiple cross-border commodity real photos;

[0116] S6012: Identify the label area in the commodity real photo;

[0117] S6013: Mark the cross-border commodity real photo with a label area smaller than the preset value as the target cross-border commodity real photo;

[0118] S6014: Domain adaptation training on the target cross-border commodity real photo to obtain the preset image encoder.

[0119] As described in steps S6011-S6014 above, first, multiple cross-border commodity real photos need to be obtained. In actual operation, these image information can be obtained in various ways, such as using photographic equipment to take real objects, downloading commodity pictures from e-commerce platforms, or using third-party data providers to obtain high-quality commodity pictures. The obtained commodity real photos should contain various styles and types to cover different commodity categories and market environments. In order to ensure the effectiveness and diversity of image data, enterprises should pay attention to image quality, including clarity, lighting conditions and shooting angles, etc., to ensure that each image can clearly display the identification, label and other information of the commodity. These images should also cover commodities in different countries and regions in order to better reflect the diversity of regulations and compliance requirements. Since the label area contains important compliance information that directly affects subsequent compliance review and text recognition, the label area usually includes product name, certification logo, barcode and other legally prescribed information. Specifically, computer vision techniques such as edge detection, contour extraction and region segmentation algorithms can be used. By applying deep learning models such as convolutional neural networks for image analysis, the system can automatically identify and mark the areas containing labels in the picture.

[0120] The identified label area is further screened, especially focusing on those cross-border commodity real photos with a label area smaller than a preset value. This preset value can be a pre-set value, for example, it can be that the label area accounts for 5% of all image areas, that is, less than 5% is considered a small label area. Mark the commodity real photo with a label area smaller than the preset value as the target cross-border commodity real photo, mainly to ensure that the model focuses on those image samples that may have difficulty in actual application during model training. Such screening helps to enhance the representativeness of the model in the learning process, and improves its robustness and accuracy in processing small size labels or complex backgrounds.

[0121] The target cross-border commodity real object image selected is subjected to domain adaptation training to obtain a preset image encoder. Domain adaptation training is a technique specially used to improve the performance of a model in a specific application scenario, and its purpose is to enable the model to adapt to the distribution changes of input data, especially when facing diversified and heterogeneous data. In this process, the target commodity real object image is input into the pre-designed image encoder as training data. During the domain adaptation training process, the system will update the model parameters through iteration to gradually optimize the model's performance on the target data set, especially in terms of small-size label recognition ability.

[0122] In one embodiment, the step S6 of auditing the compliance of the trade text based on the structured knowledge graph comprises:

[0123] S611: extracting element information in the trade text;

[0124] S612: mapping the element information with nodes in the structured knowledge graph to extract corresponding compliance standards;

[0125] S613: inputting the compliance standards and the element information into a preset auditing model to audit the compliance of the trade text.

[0126] As described in steps S611-S613 above, element information is extracted from the generated trade text. This process is the basis for compliance auditing, because only by accurately identifying and analyzing the key information in the text can its compliance be effectively judged. Element information usually includes various data related to commodities, such as commodity name, model number, specifications, applicable compliance standards, manufacturer information, supply chain information, and other key compliance data.

[0127] To perform this extraction process, natural language processing (NLP) techniques are typically employed. The extracted information is then organized into a structured format for subsequent processing, with the extracted element information mapped to nodes in a structured knowledge graph. The structured knowledge graph typically stores relationships between specific commodities, industry regulations, and compliance standards. For example, if the extracted element information includes the name of a certain type of commodity and the corresponding market, the system can quickly locate the regulations that the commodity needs to follow in the specific market using the relationships present in the knowledge graph. Specifically, a query statement using a graph database is used to retrieve the regulation nodes that match the extracted element information. The accuracy of the data mapping directly affects the results of the subsequent audit, so it is necessary to ensure that the mapping process is fine enough to handle various possible changes, such as language differences, regulation updates, and industry dynamics. Through the mapping of element information to knowledge graph nodes, the audit system can generate a clear list of compliance standards, clearly indicating the specific compliance requirements that each commodity needs to follow, thereby providing information support for the next step of compliance audit, ensuring that the compliance assessment of trade texts is comprehensive and accurate.

[0128] The pre-set audit model is typically based on machine learning or deep learning algorithms and can process input compliance standards and element information. The audit model analyzes this data to identify potential compliance issues, such as whether the necessary compliance documents are complete, whether the identification is in accordance with the specification, and whether the applicable regulations are followed. The design of the model may include algorithms such as decision trees, support vector machines (SVM), random forests, or use neural network models for more complex compliance judgments.

[0129] In one embodiment, before the step S613 of inputting the compliance standards and the element information into a pre-set audit model to audit the compliance of the trade text, the method further comprises:

[0130] S6121: Extracting first regional feature information in the element information;

[0131] S6122: Calculating the similarity of the first regional feature information with pre-stored second regional feature information respectively;

[0132] S6123: The second regional feature information with the largest similarity is recorded as the target second regional feature information;

[0133] S6124: Obtain the corresponding pre-set audit model based on the target second regional feature information; wherein each second regional feature information corresponds to an audit model, and each audit model is trained by corresponding element information and its compliance in a labeled training manner.

[0134] As described in steps S6121-S6124 above, the first regional characteristic information is identified and extracted from the extracted element information. The first regional characteristic information is usually related to factors such as the market of the commodity, the country of supply, regulatory requirements, and geographical location. During the extraction process, the system can use natural language processing (NLP) techniques to analyze the trade text to identify key terms and information related to regional characteristics. For example, specific countries, region names, applicable regulations or market clauses for products, etc. are extracted from the text.

[0135] The extracted first regional characteristic information is subjected to similarity calculation with a plurality of second regional characteristic information pre-stored. This process aims to quantify the similarity between different regional characteristics to help the system determine the best matching target second regional characteristic information. To achieve similarity calculation, the system can use various methods such as cosine similarity, Euclidean distance, or Jaccard similarity techniques to compare the similarity between the first regional characteristic information and each second regional characteristic information.

[0136] According to the similarity calculation result of the previous step, the second regional characteristic information with the largest similarity is marked as the target second regional characteristic information, which can effectively link to the audit model corresponding to this regional characteristic information, and prepare for the subsequent steps of the audit process. The corresponding pre-set audit model is obtained according to the target second regional characteristic information determined previously. Each second regional characteristic information is pre-linked to a specific audit model, which is trained through labeled training of relevant element information and its compliance. According to the selected target second regional characteristic information, the pre-set audit model associated with it is retrieved from the stored audit model library. These audit models have fully considered the specific regulations, market characteristics and compliance requirements of the target region during training, and thus can better adapt to the approval process. In this way, the system can ensure that the selected audit model has the most relevant compliance assessment capability and contains rules and standards for specific markets, thereby effectively improving the efficiency and accuracy of compliance audit and reducing potential compliance risks.

[0137] In one embodiment, after the step S613 of inputting the compliance standards and the element information into the pre-set audit model to audit the compliance of the trade text, the method further comprises:

[0138] S6141: obtaining the predicted audit results and the actual audit results of a plurality of element information;

[0139] S6142: calculating the Mahalanobis distance according to the predicted audit results and the actual audit results;

[0140] S6143: recording the element information with a Mahalanobis distance greater than a set threshold as differential element information;

[0141] S6144: Based on the differentiated element information and the corresponding actual audit result, the preset audit model is trained by using the elastic weight consolidation algorithm for incremental learning, so as to obtain a new preset audit model.

[0142] As described in steps S6141-S6144, the predicted audit results of multiple element information and the corresponding actual audit results are obtained. The predicted audit result is the output obtained by previously performing compliance audit on the trade text by the preset audit model. The actual audit result is obtained from the feedback of the regulatory agency.

[0143] The Mahalanobis distance will be used to measure the difference between the predicted audit result and the actual audit result. Mahalanobis distance is a statistical method that can calculate the distance between different variables in multi-dimensional space. Compared with Euclidean distance, Mahalanobis distance considers the correlation between variables, which can more accurately reflect the difference between the two groups of data. The process of calculating Mahalanobis distance involves extracting the feature vector of multiple element information, which usually needs to use the covariance matrix to standardize the difference between different features. The specific steps are as follows: first, express the predicted audit result and the actual audit result as a feature vector, and then calculate the Mahalanobis distance between them through the covariance matrix. This process will enable the model to better identify which element information has significant differences in prediction. By calculating the Mahalanobis distance, the cases where the model prediction result and the actual customs release have significant differences are screened out, that is, the element information with a Mahalanobis distance greater than a certain threshold is recorded as the differentiated element information. Based on the identified differentiated element information and its corresponding actual audit result, the preset audit model is trained by incremental learning to generate a new and optimized audit model. The significance of using the elastic weight consolidation (EWC) algorithm is that by fine-tuning the model parameters, it ensures that the existing knowledge is not forgotten when learning new information. Specifically, the EWC algorithm calculates the Fisher information matrix to evaluate which model parameters have a significant impact on the performance of the model. During model training, EWC will increase the protection of these important parameters, thereby avoiding the problem of performance degradation when updating the model. In this way, the system can train the model with differentiated element information while retaining existing knowledge, improving the model's adaptability in new contexts. Through this incremental learning method, the accuracy and reliability of the audit model can be quickly improved without retraining the entire model. This process not only improves the model's ability to judge complex scenarios, but also enhances its adaptability to future regulatory changes. Finally, the new preset audit model generated will have higher compliance audit capability, which helps enterprises operate stably in a changing market environment and effectively control compliance risks.

[0144] In one embodiment, the automatic update of the audit model can be set, that is, every preset time interval, the predicted audit result and the actual audit result of the plurality of element information are automatically triggered to update the audit model.

[0145] Referring Figure 3 The application further provides a multi-country compliance intelligent audit device for cross-border e-commerce, which comprises:

[0146] A first acquisition module 902 is configured to acquire the regulation amendment frequency of each specified trade country.

[0147] A setting module 904 is configured to set a regulation crawling period for a corresponding specified trade country according to the regulation amendment frequency.

[0148] A second acquisition module 906 is configured to acquire the regulation text of each specified trade country based on the regulation crawling period of each specified trade country.

[0149] An analysis module 908 is configured to analyze the regulation text by using the semantic analysis capability of a preset large language model to obtain analysis content.

[0150] A construction module 910 is configured to construct a structured knowledge graph based on the analysis content.

[0151] An audit module 912 is configured to acquire trade text and audit the compliance of the trade text based on the structured knowledge graph.

[0152] In one embodiment, the audit module 912 comprises:

[0153] An image information acquisition sub-module is configured to acquire image information of each trade commodity.

[0154] A feature extraction sub-module is configured to extract features of each image information by using a preset image encoder to obtain a label feature map corresponding to each image information.

[0155] A text recognition sub-module is configured to recognize text based on a preset text encoder to obtain a text feature corresponding to each label feature map.

[0156] A trade text generation sub-module is configured to generate the trade text based on all the text features.

[0157] A compliance audit sub-module is configured to audit the compliance of the trade text based on the structured knowledge graph.

[0158] In one embodiment, the trade text generation sub-module comprises:

[0159] An extraction unit is configured to extract corresponding image features in each label feature map;

[0160] A processing unit is configured to process the image features and the text features corresponding to each label feature map through a double-channel attention mechanism to obtain standard text information corresponding to each label feature map;

[0161] A synthesis unit is configured to synthesize the standard text information corresponding to each label feature map to obtain the trade text.

[0162] In an embodiment, the auditing module 912 further includes:

[0163] A cross-border commodity real object image acquisition submodule is configured to acquire multiple cross-border commodity real object images;

[0164] A label region identification submodule is configured to identify a label region in the commodity real object image;

[0165] A cross-border commodity real object image marking submodule is configured to mark a cross-border commodity real object image with a label region smaller than a preset value as a target cross-border commodity real object image;

[0166] A domain adaptation training submodule is configured to perform domain adaptation training on the target cross-border commodity real object image to obtain the preset image encoder.

[0167] In an embodiment, the auditing module 912 includes:

[0168] An element information extraction submodule is configured to extract element information in the trade text;

[0169] A mapping submodule is configured to map the element information to nodes in the structured knowledge graph to extract corresponding compliance standards;

[0170] An element information input submodule is configured to input the compliance standards and the element information into a preset auditing model to audit the compliance of the trade text.

[0171] In an embodiment, the auditing module 912 further includes:

[0172] A first regional feature information extraction submodule is configured to extract first regional feature information in the element information;

[0173] A similarity calculation submodule is configured to calculate the similarity of the first regional feature information with multiple second regional feature information stored in advance, respectively;

[0174] A target second regional feature information marking submodule is configured to mark the second regional feature information with the largest similarity as target second regional feature information;

[0175] The preset audit model acquisition submodule is used to obtain the corresponding preset audit model based on the target second regional characteristic information; wherein each second regional characteristic information corresponds to an audit model, and each audit model is trained in a labeled training method through the corresponding element information and its compliance.

[0176] In one embodiment, the audit module 912 further includes:

[0177] The actual audit result acquisition submodule is used to obtain the predicted audit results and actual audit results of multiple element information;

[0178] A Mahalanobis distance calculation submodule, configured to calculate the Mahalanobis distance based on the predicted audit result and the actual audit result;

[0179] The differential element information marking submodule is used to record the element information with a Mahalanobis distance greater than a set threshold as differential element information;

[0180] The incremental learning training submodule is used to perform incremental learning training on the preset audit model using an elastic weight solidification algorithm based on the differentiated factor information and the corresponding actual audit results, thereby obtaining a new preset audit model.

[0181] Figure 4 FIG1 shows an internal structure diagram of a computer device in an embodiment. The computer device can be a terminal or a server. Figure 4 As shown, the computer device includes a processor, a memory and a network interface connected via a system bus. Among them, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement a multi-country intelligent compliance audit method for cross-border e-commerce. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor can implement a multi-country intelligent compliance audit method for cross-border e-commerce. Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0182] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0183] acquire the regulation amendment frequency of each designated trade country;

[0184] set a regulation crawling period for the corresponding designated trade country according to the regulation amendment frequency;

[0185] acquire the regulation text of each designated trade country based on the regulation crawling period of each designated trade country;

[0186] analyze the regulation text by using the semantic analysis capability of a preset large language model to obtain analysis content;

[0187] construct a structured knowledge graph based on the analysis content;

[0188] acquire trade text and audit the compliance of the trade text based on the structured knowledge graph.

[0189] By acquiring the regulation amendment frequency of each designated trade country in real time and setting a dynamic regulation crawling period according to this information, an enterprise can timely grasp the regulation changes, so that the enterprise can realize real-time tracking and intelligent management of compliance information, which provides strong support for cross-border e-commerce enterprises and helps to quickly adapt to the regulation changes of the global market, reduce the risk of violation, and improve the competitiveness and market adaptability of the enterprise.

[0190] In one embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0191] acquire the regulation amendment frequency of each designated trade country;

[0192] set a regulation crawling period for the corresponding designated trade country according to the regulation amendment frequency;

[0193] acquire the regulation text of each designated trade country based on the regulation crawling period of each designated trade country;

[0194] analyze the regulation text by using the semantic analysis capability of a preset large language model to obtain analysis content;

[0195] construct a structured knowledge graph based on the analysis content;

[0196] acquire trade text and audit the compliance of the trade text based on the structured knowledge graph.

[0197] By obtaining the frequency of regulations revision of each designated trading country in real time and setting a dynamic regulation scraping cycle according to this information, enterprises can timely grasp the regulation changes, so that the enterprises can realize real-time tracking and intelligent management of compliance information, and provide a strong support for cross-border e-commerce enterprises, which helps to quickly adapt to the changes of global market regulations, reduces the risk of violation, and improves the competitiveness and market adaptability of enterprises.

[0198] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The program can be stored in a non-volatile computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments of each method. In the embodiments provided in the present application, any reference to memory, storage, database or other medium can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchl ink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0199] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application.

[0200] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A multi-national intelligent compliance audit method for cross-border e-commerce, characterized by: The method comprises: Obtain the frequency of regulatory revisions for each designated trading country; Set a regulatory crawling cycle for the designated trading country based on the regulatory revision frequency; Obtaining the regulatory text of each designated trading country based on the regulatory crawling cycle of each designated trading country; Parsing the regulatory text using the semantic parsing capability of a preset large language model to obtain parsed content; Constructing a structured knowledge graph based on the parsed content; Obtain trade texts, and review the compliance of the trade texts based on the structured knowledge graph.

2. The multi-national intelligent compliance audit method for cross-border e-commerce according to claim 1 is characterized in that: The step of obtaining a trade text and reviewing the compliance of the trade text based on the structured knowledge graph includes: Obtain image information of each commodity to be traded; Extract features from each of the image information using a preset image encoder to obtain a label feature map corresponding to each of the image information; Performing text recognition on each of the label feature maps based on a preset text encoder to obtain text features corresponding to each of the label feature maps; generating the trade text based on all the text features; The compliance of the trade text is reviewed based on the structured knowledge graph.

3. The multi-national intelligent compliance audit method for cross-border e-commerce according to claim 2 is characterized in that: The step of generating the trade text based on all the text features comprises: Extracting corresponding image features from each of the label feature maps; The image features and the text features corresponding to each of the label feature maps are processed through a dual-channel attention mechanism to obtain standard text information corresponding to each of the label feature maps; The standard text information corresponding to each of the label feature images is integrated to obtain the trade text.

4. The multi-national intelligent compliance audit method for cross-border e-commerce according to claim 2 is characterized in that: Before the step of extracting features from each of the image information using a preset image encoder to obtain a label feature map corresponding to each of the image information, the method further includes: Obtain multiple physical images of cross-border products; Identifying the label area in the physical image of the product; The cross-border commodity images whose label areas are smaller than the preset value are recorded as target cross-border commodity images; Domain adaptation training is performed on the target cross-border commodity physical image to obtain the preset image encoder.

5. The multi-national intelligent compliance audit method for cross-border e-commerce according to claim 1 is characterized in that: The step of reviewing the compliance of the trade text based on the structured knowledge graph includes: extracting element information from the trade text; Mapping the element information with nodes in the structured knowledge graph to extract corresponding compliance standards; The compliance standards and the element information are input into a preset audit model to audit the compliance of the trade text.

6. The multi-national intelligent compliance audit method for cross-border e-commerce according to claim 5 is characterized in that: Before the step of inputting the compliance standards and the element information into a preset audit model to audit the compliance of the trade document, the method further includes: Extracting first regional feature information from the element information; Calculating similarity between the first regional feature information and a plurality of pre-stored second regional feature information; Recording the second regional feature information with the greatest similarity as the target second regional feature information; A corresponding preset audit model is obtained based on the target second regional characteristic information; wherein each second regional characteristic information corresponds to an audit model, and each audit model is trained in a labeled training method through corresponding element information and its compliance.

7. The multi-national intelligent compliance audit method for cross-border e-commerce according to claim 5 is characterized in that: After the step of inputting the compliance standards and the element information into a preset audit model to audit the compliance of the trade document, the method further includes: Obtain the predicted audit results and actual audit results of multiple elements of information; Calculating Mahalanobis distance based on the predicted audit result and the actual audit result; The feature information whose Mahalanobis distance is greater than the set threshold is recorded as differentiated feature information; Based on the differentiated factor information and the corresponding actual audit results, the preset audit model is incrementally learned and trained using an elastic weight solidification algorithm to obtain a new preset audit model.

8. A multi-national compliance intelligent audit device for cross-border e-commerce, characterized by: The device comprises: The first acquisition module is used to obtain the frequency of regulatory revisions of each designated trading country; A setting module is used to set a regulatory crawling cycle for a corresponding designated trading country according to the regulatory revision frequency; A second acquisition module is configured to acquire the regulatory text of each designated trading country based on the regulatory capture cycle of each designated trading country; A parsing module, configured to parse the regulatory text using the semantic parsing capability of a preset large language model to obtain parsed content; A construction module, configured to construct a structured knowledge graph based on the parsed content; The audit module is used to obtain trade texts and audit the compliance of the trade texts based on the structured knowledge graph.

9. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed by a processor, the processor executes the steps of the multi-country compliance intelligent audit method for cross-border e-commerce as described in any one of claims 1 to 7.

10. A computer device, characterized in that: The device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the multi-country compliance intelligent audit method for cross-border e-commerce as described in any one of claims 1 to 7.

Images