A foreign exchange trade document screening method, device, medium and product

Abstract

The application discloses a foreign exchange trade document screening method and device, medium and product, relates to the technical field of data analysis and financial technology, and comprises the following steps: based on document image recognition and analysis, trade business data is extracted as a query basis; through multi-dimensional page adaptation, external equity system automatic stable crawling of equity key fields is realized; based on business data collection, a main body list is formed, compliance list screening is completed in batches, a screening summary is obtained, equity information, screening results and bank stock transaction data are fused to build an equity and transaction two-dimensional correlation graph, and random walk algorithm is combined to quantitatively measure and calculate the risk score and transmission link of each main body, so that full-process automatic compliance due diligence of foreign exchange trade documents is realized, the risk source and transmission path are accurately positioned, finally, standardized compliance review files are automatically integrated from full-dimensional data, cross-border trade compliance review working hours are reduced, and the risk control identification precision and the standardization degree of review files are improved.

Description

Technical Field

[0001] This invention relates to the fields of data analysis and financial technology, and in particular to a method, device, medium and product for screening foreign exchange trade documents. Background Technology

[0002] In the context of international trade business of financial institutions, anti-money laundering and sensitive due diligence are key tasks for compliance management. Traditional sensitive verification requires business personnel to frequently switch between multiple heterogeneous systems such as equity inquiries, compliance screening, and bank intranets, and manually complete repetitive tasks such as page access, information retrieval, and data organization.

[0003] In the existing processing method, staff manually identify document images and extract trade entity information. They rely on a simple recording script to simulate keyboard and mouse operations according to fixed page positions to achieve simple single-system queries. After all external data is acquired, information is manually summarized and compared item by item to complete the sensitive data assessment.

[0004] However, this type of automated script, which relies on fixed page positioning, has poor adaptability. Page changes such as website redesign and pop-ups can easily cause the automated process to be interrupted. Equity, screening, and internal transaction data are scattered and isolated and cannot be linked for analysis, making it difficult to identify hidden risks in multiple nested layers. The reliance on manual operation in many steps reduces screening efficiency and increases the risk of false positives and false negatives. Summary of the Invention

[0005] This invention provides a method, device, medium, and product for screening foreign exchange trade documents, in order to solve the problems of poor adaptability of foreign exchange trade document screening methods, easy interruption of automated processes, difficulty in identifying hidden risks of multi-layered nesting, low screening efficiency, and high risk of false positives and false negatives.

[0006] According to one aspect of the present invention, a method for screening foreign exchange trade documents is provided, comprising:

[0007] Acquire image files of foreign exchange trade documents and extract trade business data through recognition and semantic analysis;

[0008] Using trade business data as the source data for the query, and after performing page adaptation operations for the equity information query system, the key equity field information is retrieved from the equity information query system.

[0009] A list of entities to be screened is constructed based on trade business data. A sensitive list query is performed based on the list of entities to be screened to retrieve the screening summary information corresponding to the entities to be screened.

[0010] Based on key equity field information and screening summary information, combined with the bank's internal historical transaction data, a fusion correlation graph is constructed, and a random walk algorithm is used to perform risk calculation to calculate the sensitive risk score and risk transmission path of all entities to be screened, which serves as the risk assessment result.

[0011] Based on key equity field information, screening summary information, and risk assessment results, a compliance review document matching the image file of foreign exchange trade documents is generated.

[0012] According to another aspect of the present invention, a foreign exchange trade document screening device is provided, comprising:

[0013] The image recognition module is used to acquire image files of foreign exchange trade documents and extract trade business data through recognition and semantic analysis.

[0014] The equity capture module is used to use trade business data as the source data for queries. After performing page adaptation operations for the equity information query system, it captures key equity field information from the equity information query system.

[0015] The screening summary module is used to construct a list of entities to be screened based on trade business data, perform a sensitive list query based on the list of entities to be screened, and retrieve the screening summary information corresponding to the entities to be screened.

[0016] The risk assessment module is used to construct a fusion correlation graph based on key equity field information and screening summary information, combined with the bank's internal historical transaction data, and to perform risk calculation using a random walk algorithm to calculate the sensitive risk score and risk transmission path of all entities to be screened, which serves as the risk assessment result.

[0017] The document review module is used to generate compliance review documents that match the image files of foreign exchange trade documents, based on key equity field information, screening summary information, and risk assessment results.

[0018] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0019] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the foreign exchange trade document screening method according to any embodiment of the present invention.

[0020] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, the computer instructions being configured to cause a processor to execute and implement the foreign exchange trade document screening method according to any embodiment of the present invention.

[0021] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the method as described in any embodiment of the present invention.

[0022] The technical solution of this invention extracts trade business data based on document image recognition and analysis as the basis for querying. Through multi-dimensional page adaptation, it achieves automated and stable crawling of key equity fields from external equity systems, avoiding automation task failures caused by website redesigns and pop-up interference. Based on the collection of business data, it forms a list of entities to complete batch compliance list screening and obtain screening summaries. It integrates equity information, screening results, and existing bank transaction data to build a two-dimensional correlation graph of equity and transactions, and combines a random walk algorithm to quantitatively calculate the risk scores and transmission links of each entity. This achieves fully automated compliance due diligence of foreign exchange trade documents, avoiding the inefficiency and human error caused by manual data entry and verification. It can also penetrate multiple layers of equity relationships and fund transaction links to uncover hidden risks, accurately locate the source of risk and transmission path, and finally automatically integrate all-dimensional data to generate standardized compliance review documents. This reduces the working time of cross-border trade compliance review, improves the accuracy of risk control identification, and enhances the standardization of review documents.

[0023] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of a foreign exchange trade document screening method provided in Embodiment 1 of the present invention;

[0026] Figure 2 This is a schematic diagram illustrating a heterogeneous adaptive login domain applicable to an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of a menu topology drift fuzzy matching method applicable to an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram illustrating the evolution of intercepting unexpected pop-up windows applicable to an embodiment of the present invention;

[0029] Figure 5 This is a flowchart of another foreign exchange trade document screening method provided in Embodiment 2 of the present invention;

[0030] Figure 6 This is a flowchart of another foreign exchange trade document screening method provided in Embodiment 3 of the present invention;

[0031] Figure 7 This is a schematic diagram of dynamic risk reasoning applicable to an embodiment of the present invention;

[0032] Figure 8 This is a schematic diagram of a foreign exchange trade document screening system applicable to an embodiment of the present invention;

[0033] Figure 9 This is a schematic diagram of an adaptive process for screening foreign exchange trade documents applicable to an embodiment of the present invention;

[0034] Figure 10 This is a schematic diagram of the structure of a foreign exchange trade document screening device according to Embodiment 4 of the present invention;

[0035] Figure 11 This is a schematic diagram of the structure of an electronic device that implements the foreign exchange trade document screening method of this invention. Detailed Implementation

[0036] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0037] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0038] Example 1

[0039] Figure 1 This is a flowchart of a foreign exchange trade document screening method provided in Embodiment 1 of the present invention. This embodiment is applicable to the screening of foreign exchange trade documents. The method can be executed by a foreign exchange trade document screening device, which can be implemented in hardware and / or software and is generally configured in an electronic device. Figure 1 As shown, the method includes:

[0040] S110. Obtain image files of foreign exchange trade documents and extract trade business data through recognition and semantic analysis.

[0041] In this embodiment of the invention, the foreign exchange trade document image file can be specifically understood as: electronic image data converted from paper documents of foreign exchange trade business retained by the bank.

[0042] Specifically, after business personnel upload foreign exchange trade document image files to the bank's internal system, the system automatically generates a corresponding sensitive screening record and marks it as a "subject to be identified" business status tag. The background scheduler uses OCR (Optical Character Recognition) to parse the foreign exchange trade document image and extract all character text within the image. Then, through semantic analysis, the system performs field classification, extracts trade business data such as the trade entity name, country of registration, and category of traded goods, and completes the standardization process according to the preset database field specifications. The data is then written into the corresponding data table one by one, realizing the structured storage of trade business data.

[0043] After data extraction, the system automatically changes the status of this sensitive screening record to "Pending Equity Penetration." Specifically, a sensitive screening record can be understood as a business ledger entry used by the bank system for compliance verification of individual foreign trade documents. It is automatically generated based on document uploads and its status changes with business milestones. The status marker can be understood as a system identifier that distinguishes the processing stage of a business transaction based on field parameters, including different business statuses such as "subject to be identified" and "pending equity penetration." Structured storage can be understood as breaking down unstructured document text into structured fields such as subject name, country of origin, and transaction category, and storing them in the database for easy retrieval as source data by subsequent programs.

[0044] S120. Using trade business data as the source data for querying, after performing page adaptation operations for the equity information query system, retrieve key equity field information from the equity information query system.

[0045] In this embodiment of the invention, the key equity field information can be specifically understood as: a structured equity information, which serves as the basic data source for subsequently constructing the relationship graph, and may include: shareholder name, registered region, entity type, direct shareholding data, and other entity information. The equity information query system can be specifically understood as: a database platform storing commercial data such as enterprise registration information, equity structure, shareholder identity, and shareholding ratio from around the world; it is the target business system for RPA-automated collection of key equity field information, and its page layout and menu structure will be iteratively changed periodically.

[0046] Specifically, by invoking a Robotic Process Automation (RPA) robot, sensitive screening records marked as pending equity penetration are obtained, and the trade business data bound to these records is extracted as the source data for the query. After the login page coordinates are located using the webpage adaptive verification mechanism, the user is automatically logged into the equity information query system. For example, the RPA obtains the page DOM (Document Object Model) node traversal program to locate the page block to which the login form belongs. The account and password input controls are located by jointly matching the tag attributes and control types. After automatically filling in the preset login credentials, the login submission command is triggered. In conjunction with the background pop-up monitoring program, various interference pop-ups on the page are intercepted. After the verification is passed, the user is redirected to the homepage of the equity information query system.

[0047] The system retrieves and parses the HTML (HyperText Markup Language) source code of web pages to extract key equity fields such as shareholder name, registered location, entity type, and direct shareholding, and stores them. After completing the full collection of equity information for a single record, the system automatically changes the status of the corresponding sensitive screening record to "awaiting global compliance list screening" to initiate the subsequent compliance screening process.

[0048] In addition to retrieving fields through HTML page parsing, it can also connect to the open interface of the equity information query system to directly pull equity data with structured messages, avoiding the risk of parsing failure caused by page redesign.

[0049] Specifically, Robotic Process Automation (RPA) can be understood as an automated program that automatically completes cross-system access, data querying, and information retrieval according to preset business logic. Webpage adaptive verification mechanism can be understood as a page adaptation technology used to adapt to changes in system page redesigns, pop-up window obstructions, and other page changes. Global compliance list screening can be understood as a business node identifier that changes after equity data collection and archiving are completed, indicating that the process has moved to the compliance and sensitive information screening process.

[0050] Optionally, based on the above embodiments, after performing page adaptation operations for the equity information query system, retrieving key equity field information from the equity information query system may include:

[0051] The login page of the equity information query system is scanned using full-screen optical character recognition (OCR) to retrieve relevant keywords. The input box position is determined based on the relative coordinate offset of the keywords to complete the account and password input. The system is then connected to the equity information query system to extract key equity fields from the query source data. These key fields include at least the following: shareholder name, registered region, entity type, and direct shareholding data. After cleaning and standardizing the key equity fields, the standardized information is sent back to the bank's internal system.

[0052] In this embodiment of the invention, the relative coordinate offset can be specifically understood as: taking the page reference point as a reference, the pixel offset parameters of elements such as the account box and password box are calculated from the keyword text position and used for automatic parameter filling.

[0053] Specifically, after completing the page adaptation work for the equity information query system, the RPA launches the page capture component to capture the full-screen image of the login page of the equity information query system. It calls the OCR engine to recognize characters in each region of the entire page image and outputs the text and corresponding pixel coordinates. In the recognition results, it filters preset login keywords such as account, password, and login. Based on the horizontal and vertical offset thresholds of the pixel coordinates of the keywords, it calculates the control landing coordinates of the account input box and password input box. It controls the simulated keyboard and mouse component to jump to the corresponding coordinate points and fill in the pre-stored account and password characters in sequence. It locates the keyword coordinates of the login button and triggers the click operation. After completing the identity verification, it jumps to the system homepage.

[0054] RPA retrieves source data from trade business queries, filling it into the search box group by group to initiate a main entity query. After the page loads, it parses the HTML source code of the webpage and extracts target fields such as shareholder name, registered region, entity type, and direct shareholding data from the returned shareholder information block, temporarily caching the original text data. Following the bank's preset data specifications, it removes garbled characters, invalid null values, and redundant special symbols, standardizes the translation of country names, enterprise type classification formats, and shareholding value measurement standards. After all fields are standardized and organized, they are pushed in batches to the bank's internal risk control database for storage via a data interface.

[0055] By employing full-screen OCR scanning to retrieve login keywords and relying on text relative coordinate offset positioning controls to achieve automated login, and then extracting multiple key equity fields from the query source data, cleaning and standardizing them before storing them back into the bank's internal system, the drawbacks of relying on fixed DOM element positioning being easily affected by page redesigns are avoided. Image-text matching enables adaptive adaptation of the login process, ensuring stable and reliable automated data collection from external equity systems. After unified standardization, the raw data avoids problems such as messy field formats and information errors, achieving data consistency between external heterogeneous data sources and the bank's internal database. This not only reduces the manpower required for manually entering and organizing equity information and reduces the risk of errors and omissions in manual entry, but also provides a basic equity data source for subsequent association graph construction and risk quantification calculations.

[0056] Furthermore, based on the above embodiments, before retrieving key equity field information from the query source data, the process may further include:

[0057] Using a natural language semantic fuzzy matching algorithm, the semantics of menu text in the equity information query system are analyzed, and the preset menu keywords are fuzzy compared with the existing menu content on the page to construct at least one menu access path.

[0058] Accordingly, based on the above embodiments, extracting key equity field information from the source query data may include:

[0059] The user is redirected to the equity information query page in the equity information query system via the menu access path; the element structure of the equity information query page is traversed layer by layer through the document object model, and floating and pop-up-type obstructing controls are identified and closed or removed; the query source data is entered, an equity information retrieval request is initiated, and key equity field information is extracted from the retrieval results returned by the page.

[0060] In this embodiment of the invention, the menu access path can be specifically understood as: an operation chain generated after matching and positioning, which jumps step by step from the system homepage to the equity query page. The obscuring control can be specifically understood as: pop-ups, floating windows, and other redundant page components that obscure the query input area or interfere with RPA's automatic input of search conditions.

[0061] Natural language semantic fuzzy matching algorithm can be understood as: relying on semantic similarity operation rules, it is unaffected by minor adjustments or word order changes to menu text, and compares preset keywords with the page menu text. DOM layer-by-layer traversal can be understood as: parsing the webpage document object model node tree from top to bottom, and identifying page parsing methods for non-business controls such as page pop-ups and floating ads item by item.

[0062] Specifically, before retrieving equity information, RPA extracts all menu text content and preset standard menu keywords from the equity information query system page. It then calls a semantic fuzzy matching algorithm to calculate the semantic similarity between the two sets of text one by one. Combining layout data such as page coordinates and block distribution, it determines whether the menu bar has shifted position. It filters out matching results with similarity reaching a set threshold and determines the page area where the target menu is located. After completing the adaptive adaptation of the menu position, it sorts out the order of clicks from top to bottom according to the webpage hierarchy, plans and generates multiple independent menu jump operation links that can directly reach each equity query function page, and temporarily stores all jump paths in the program cache for later use.

[0063] RPA sequentially selects individual paths from multiple cached menu access paths, simulating mouse clicks on menus at each level according to a predetermined hierarchy to navigate to the corresponding equity query page. It then initiates a DOM node traversal process to scan the page element structure layer by layer, identifying various obstructing controls such as pop-ups and floating windows, and performing batch closing and removal operations to clear page interference. It locates the page search input area and automatically fills in pre-acquired query source data, triggers the search command, and waits for the results page to load. Finally, it parses the page content and extracts key equity fields such as shareholder name, registered region, entity type, and direct shareholding data.

[0064] In addition, a page element snapshot comparison mechanism can be configured to automatically update the menu keyword database periodically to adapt to system menu iteration changes. If a database interface is available, it can also be connected to bypass front-end page redirection and directly call structured equity data.

[0065] In a specific example, this embodiment of the invention designs a webpage structure change prediction rule library. This library is a set of rules that pre-include typical redesign features of external webpages and corresponding handling logic. It mainly formulates special response rules for three high-frequency page change scenarios: heterogeneous login domains, menu topology drift, and unexpected pop-up evolution. This allows automated programs to identify the types of webpage changes in advance and adaptively adjust the operation logic, ensuring that RPA can still stably complete page interaction and data crawling after webpage iteration and upgrade.

[0066] Rule 1 is a heterogeneous adaptive login domain rule. This rule is specifically designed to address the login area changes that occur after the external equity information query system is upgraded. It covers changes to the hash values ​​of attributes such as the login box HTML tag ID (Identifier), name, and category, as well as scenarios where the system adds login verification methods such as slider verification and dynamic token verification. This solves the problem that traditional element positioning methods become ineffective due to changes in the underlying code of the webpage.

[0067] This rule abandons the traditional method of locating login controls solely based on HTML tag attributes. Instead, it adopts a dual positioning mechanism that combines visual anchors with semantic tags. On the one hand, it relies on visual features such as fixed icons and partition outlines on the page to lock the approximate range of the login area. On the other hand, it combines text semantic tags such as account, password, and login to accurately identify various input controls and function buttons. This dual approach ensures reliable positioning of login controls and adapts to various changes in login area codes and verification methods.

[0068] Figure 2 This is a schematic diagram of a heterogeneous adaptive login domain applicable to an embodiment of the present invention. As shown in the figure, after accessing the login page, the system first searches for the original identifier ID input box. If it exists, it enters the standard input process; if it does not exist, it triggers rule one: heterogeneous login domain, and sequentially performs full-screen OCR scanning, locates the keyword "username / User", calculates the relative coordinate offset, clicks and inputs the credential, detects the verification code features, and if the verification code exists, it calls the verification code recognition model. Finally, it clicks the login button to complete the login process.

[0069] Rule 2 is a fuzzy matching rule for menu topology drift. This rule is specifically designed to address menu structure changes that occur after the upgrade and redesign of the equity information query system. It covers topology changes such as adjusting the "Query" menu from a first-level to a second-level level, as well as synonym replacements of menu text such as changing "Screening" to "Sanction Check". It is used to solve the problem that traditional fixed path jump methods become ineffective due to changes in menu level or text.

[0070] This rule no longer relies on fixed menu paths or text matching. Instead, it first constructs a semantic index based on the Document Object Model (DOM) tree, containing text and hierarchical information of each menu node. Then, it calls Natural Language Processing (NLP) algorithms to calculate the semantic similarity between the preset target menu text and the text of each menu element on the page. At the same time, it combines the menu hierarchy depth with weighted path scoring. For example, it calculates the DOM hierarchy depth penalty value by adding a fixed penalty value for each level of hierarchy. Then, it performs a weighted calculation on the semantic similarity score obtained by the NLP algorithm and the hierarchy depth penalty value to obtain a comprehensive score. Finally, it selects the menu element with the highest score to generate a jump path and execute the click operation, thereby adapting to various changes in menu hierarchy and copy.

[0071] Figure 3 This is a schematic diagram of a menu topology drift fuzzy matching applicable to an embodiment of the present invention. As shown in the figure, when navigation to the target menu is required, the system first determines whether the original XML (Extensible Markup Language) path (XPath) path is valid. If valid, a standard click is executed; if invalid, rule two is triggered: menu topology drift. First, all clickable elements on the page are traversed and the element text and attributes are extracted. Then, semantic similarity is calculated through NLP. Next, the DOM level depth penalty value is calculated. After comprehensive scoring and sorting, the element with the highest score is selected to perform a simulated click. Then, it is determined whether the target page is entered. If so, the navigation is successful. Otherwise, a suboptimal solution is tried, and the simulated click operation is re-executed.

[0072] Rule 3 is the unexpected pop-up evolution interception rule. This rule is specifically designed to deal with various unexpected pop-up interference scenarios added after external system upgrades, including system announcement pop-ups, authorization prompts, and advertising overlays. These pop-ups will obscure core operation buttons, directly causing RPA's automatic click operation to fail. Therefore, an independent prediction and interception mechanism is required.

[0073] This rule is based on the DOM stacking context for prediction. Before executing critical operations, it first detects the element with the highest Z-Index (stackability value) and an element area exceeding a preset threshold. Then, it combines text keywords (such as "close" or "skip") or features to determine whether it is a non-business-essential element. When it is determined to be an interfering pop-up, it forcibly removes or hides the pop-up to eliminate obstruction and ensure that subsequent automated operations are not affected.

[0074] Figure 4This is a schematic diagram of an unexpected pop-up evolution interception applicable to an embodiment of the present invention. As shown in the figure, when preparing to execute a critical operation, the system first injects a front-end interception script to scan the top-level element of the DOM tree and determine whether there is a high Z-Index occlusion. If not, the operation is executed normally. If it exists, rule three is triggered: pop-up evolution interception. First, the occlusion layer features are analyzed to determine whether the keyword "close / skip" is contained. If it is contained, a click to close is simulated. If it is not contained, it is further determined whether the occlusion area is greater than a threshold. If it is greater, the operation of forcibly hiding display:none in Cascading Style Sheets (CSS) is executed. If it is not greater, the operation is executed normally. Finally, all paths lead to the normal execution operation until the operation is completed.

[0075] Before capturing key equity field information, a natural language semantic fuzzy matching algorithm is used to compare preset keywords with page menu text to construct a menu access path. Then, the user jumps along this path and uses a document object model to traverse and identify obstructing controls before entering the query source data to extract field information. This avoids the strong dependence of traditional fixed path jumps on menu hierarchy and copywriting, and can adapt to menu topology drift caused by system redesign. At the same time, by actively removing interfering elements such as floating windows and pop-ups, it avoids failures caused by obstructing key operations. This not only improves the stability and robustness of the automated process, but also eliminates the need for manual maintenance of menu paths, reducing operation and maintenance costs. Ultimately, it ensures the continuity and reliability of key equity field information collection, providing stable and high-quality data support for subsequent association graph construction and risk quantification analysis.

[0076] S130. Construct a list of entities to be screened based on trade business data, and perform a sensitive list query based on the list of entities to be screened to retrieve the screening summary information corresponding to the entities to be screened.

[0077] In this embodiment of the invention, the list of entities to be screened can be specifically understood as: a collection of all trade entities whose sensitive information verification is to be carried out, extracted and compiled from trade business data. The screening summary information can be specifically understood as: the results from the global compliance list screening system that summarize the sensitive information risk status of the entities.

[0078] Specifically, the RPA robot accesses the bank's internal business database, filters out all business records marked as pending global compliance list screening based on data status fields, extracts and integrates information on trading companies, partners, and other entities from each record to generate a standardized list of entities to be screened, and initiates login operations to access the global compliance list system according to a preset process. Throughout the entire process of login, menu navigation, data entry, and result retrieval, it continuously calls the webpage structure change prediction rule library. Relying on three types of rules—login domain heterogeneity, menu topology drift, and unexpected pop-up blocking—it adaptively responds to redesign issues such as page label changes, menu position offsets, and pop-up obstruction. After the screening instruction is executed and the result page is loaded, it parses the document object model structure layer by layer, locates the result display area, and extracts screening summary information such as risk descriptions and verification conclusions corresponding to each entity in the list and stores them temporarily.

[0079] S140. Based on key equity field information and screening summary information, combined with the bank's internal historical transaction data, a fusion association graph is constructed, and a random walk algorithm is used to perform risk calculation to obtain the sensitive risk score and risk transmission path of all entities to be screened, which serves as the risk assessment result.

[0080] In this embodiment of the invention, existing historical transaction data can be specifically understood as: historical business data such as past business dealings, fund transfers, and cooperation records stored internally by the bank. The integrated relational graph can be specifically understood as: integrating multi-dimensional data on equity, screening, and transactions to build a visual network graph that reflects the relationships between entities.

[0081] The random walk algorithm can be understood as a data analysis algorithm that simulates path walking between nodes in a relational graph, deduces risk propagation, and calculates the risk level of each node. Sensitive risk scoring can be understood as a numerical result that quantifies the level of sensitivity risk of an entity, calculated based on multi-dimensional data. Risk transmission paths can be understood as the relationship links formed by the gradual transmission and diffusion of risk among related entities.

[0082] Specifically, the process involves obtaining key equity field information and screening summary information for each entity, retrieving corresponding customer historical transaction records from the bank's internal database, and classifying the data according to business rules.

[0083] The correspondence between various business information and records is as follows: each business document is matched as a single sensitive screening record, and one or more screening entities contained in a single record are uniformly summarized into the same list of entities to be screened. All entities in the list share the equity data and screening basic data of the current group for subsequent risk control calculations. At the same time, according to the correspondence between the list and the screening summary information, an independent query is initiated for each screening entity in the list and a unique screening summary information is generated.

[0084] The collected equity key fields, screening summaries, and customer historical transaction data are cleaned, formatted, and mapped. Each entity to be screened is used as an independent node in the graph. Based on the equity data, equity relationship links such as shareholding and controlling stakes are built between entities. Based on the historical transaction data, transaction relationship links such as fund transfers and business cooperation are constructed. Risk relationship identifiers are added in conjunction with the screening summary information. Then, the link attributes and weights are defined according to the node type and relationship category. All nodes and various relationship links are integrated and connected to complete the construction and initialization of the fused relationship graph.

[0085] Using each subject to be screened in the fused association graph as an independent computing node, the Random Walk (RWR) algorithm is initialized with the starting point and transition probability of each association link based on the preset weight coefficients. According to the algorithm rules, multiple rounds of iterative walk operations are continuously performed between the graph nodes and association links. Each round of iteration updates the accumulated risk feature values ​​of each node. After the iteration converges, the quantitative sensitive risk score of all subjects in the list is obtained based on the final value. At the same time, the node jump trajectory, link direction and propagation direction of each walk are tracked throughout the process. The risk transmission order and association relationship between nodes are sorted out one by one. The link direction of risk spreading from the source subject to the surrounding associated subjects is fully identified and recorded to form the risk transmission path.

[0086] Finally, the risk scores of all subjects and their corresponding risk transmission paths are integrated and summarized to form a complete risk assessment result, which is then stored and output.

[0087] In addition, the scoring results can be graded and labeled according to preset risk level thresholds, and high-risk entities can be marked according to the transmission path. The assessment results can also be linked with the bank's risk control rule base to automatically generate corresponding risk disposal suggestions and early warning prompts.

[0088] S150: Based on key equity field information, screening summary information, and risk assessment results, generate compliance review documents that match the image files of foreign exchange trade documents.

[0089] In this embodiment of the invention, the compliance review document can be specifically understood as: a formal document formed by integrating multiple types of verification data and risk results, used for compliance verification and retention of foreign exchange trade documents.

[0090] Specifically, the key equity field information is sorted and broken down to form complete equity penetration data. Then, the screening summary information is summarized to generate a global compliance list screening report. Dynamic risk reasoning scoring data such as the sensitive risk score and risk transmission path of each entity are retrieved. The above data is formatted and integrated, and bound to the corresponding foreign exchange trade document image file according to the unique document number to generate a compliance review document with a standardized format and complete content.

[0091] Specifically, equity penetration data can be understood as: data content such as the equity hierarchy and shareholding relationships of the main entity obtained by sorting out key equity field information.

[0092] Bank staff can log in to the internal business system with their account permissions, access the dedicated verification result display page, find the corresponding document item, and click the download option to obtain the complete set of compliance review documents locally and complete the sensitive verification of this foreign exchange trade document.

[0093] Meanwhile, the system automatically performs archiving operations in the background, classifying compliance review documents, original foreign exchange trade document images, and various screening reports into the electronic archive according to preset classification rules. It configures retrieval index fields such as document number, business date, and entity name for the archives to enable multi-dimensional and rapid querying. It can also automatically push the entire set of documents to downstream online business review nodes according to internal business flow rules, so that reviewers can carry out online review work.

[0094] The technical solution of this invention extracts trade business data based on document image recognition and analysis as the basis for querying. Through multi-dimensional page adaptation, it achieves automated and stable crawling of key equity fields from external equity systems, avoiding automation task failures caused by website redesigns and pop-up interference. Based on the collection of business data, it forms a list of entities to complete batch compliance list screening and obtain screening summaries. It integrates equity information, screening results, and existing bank transaction data to build a two-dimensional correlation graph of equity and transactions, and combines a random walk algorithm to quantitatively calculate the risk scores and transmission links of each entity. This achieves fully automated compliance due diligence of foreign exchange trade documents, avoiding the inefficiency and human error caused by manual data entry and verification. It can also penetrate multiple layers of equity relationships and fund transaction links to uncover hidden risks, accurately locate the source of risk and transmission path, and finally automatically integrate all-dimensional data to generate standardized compliance review documents. This reduces the working time of cross-border trade compliance review, improves the accuracy of risk control identification, and enhances the standardization of review documents.

[0095] Example 2

[0096] Figure 5 This is a flowchart of another foreign exchange trade document screening method provided in Embodiment 2 of the present invention. This embodiment is a refinement of the above embodiment's "constructing a fusion correlation graph based on key equity field information and screening summary information, combined with existing historical transaction data within the bank." Figure 5 As shown, the method includes:

[0097] S510. Obtain image files of foreign exchange trade documents and extract trade business data through recognition and semantic analysis.

[0098] S520. Using trade business data as the source data for the query, after performing page adaptation operations for the equity information query system, retrieve key equity field information from the equity information query system.

[0099] S530. Construct a list of entities to be screened based on trade business data, and perform a sensitive list query based on the list of entities to be screened to retrieve the screening summary information corresponding to the entities to be screened.

[0100] S540. Using each entity to be screened as the root node, recursively retrieve direct and indirect shareholder information at a preset level and depth based on key equity field information to construct an equity subgraph.

[0101] In this embodiment of the invention, the root node can be specifically understood as: the node of the subject to be screened, serving as the starting point for graph construction. The equity subgraph can be specifically understood as: a local graph built upon equity data, reflecting the equity relationships between the subject to be screened and shareholders at various levels. Recursive retrieval can be specifically understood as: a cyclical retrieval method that traces shareholder information upwards layer by layer.

[0102] S550. Obtain the bank's internal historical transaction data, identify the counterparties corresponding to each entity to be screened, and when the counterparty does not exist in the matching equity subgraph, add a counterparty node and establish a transaction association edge between the root node and the counterparty node. Use a random walk algorithm to perform risk calculation and calculate the sensitive risk score and risk transmission path of all entities to be screened as the risk assessment result.

[0103] In this embodiment of the invention, a counterparty can be specifically understood as a cooperative entity that has financial or business dealings with the entity to be screened. A transaction-related edge can be specifically understood as a connection link created between graph nodes to represent the transactional cooperation relationship between the two parties.

[0104] Specifically, when constructing the fusion association graph, all entities to be screened are traversed and uniformly set as the root node of the fusion association graph. The collected key equity fields are decomposed, cleaned, and standardized in format. Starting from each root node, the corresponding direct and indirect shareholder information is recursively retrieved layer by layer. At the same time, a recursion depth threshold is pre-configured (e.g., the maximum number of layers does not exceed 5). When the retrieval level reaches the threshold, the recursion process is terminated to avoid infinite recursion and effectively avoid problems such as excessive consumption of computing resources and program lag. During the process of retrieving shareholder information layer by layer, node attributes such as shareholding ratio can be bound to each graph node simultaneously. Node creation, relationship building, and attribute assignment are completed in sequence, and finally, a complete equity subgraph with complete structure and attribute information is generated.

[0105] Complete historical transaction records of customers are extracted in batches from the bank's internal database. Each record is analyzed to extract key information such as the counterparty's name, transaction amount, and frequency of cooperation. The extracted counterparties are then matched against all nodes in the existing equity sub-graph for verification based on dimensions such as name and entity code. If the matching result shows that the counterparty does not exist in the existing graph nodes, a new independent node is created within the graph with its basic information improved. A transaction relationship edge is then created between the corresponding root node and the newly created node. Each transaction relationship edge is assigned a unique relationship weight value based on preset rules to quantify the strength of the business relationship between the two parties. For example, transaction amount, frequency of cooperation, and business duration are used as calculation factors, and corresponding score coefficients are configured for each. First, the corresponding data in the transaction records is extracted. After calculating each item using the coefficients, a base score is accumulated. Then, weight levels are divided according to the score range and mapped to standardized weight values. Finally, this value is bound to the corresponding transaction relationship edge to quantify the closeness and strength of the business relationship between the entity to be screened and its counterparty.

[0106] After the basic graph is built, staff can add additional attributes such as sensitive tags and risk levels to all nodes according to business needs. They can also dynamically and differentiate the weights of transaction relationship edges based on different indicators such as the size of a single transaction and the frequency of long-term cooperation. By expanding the dimensions of graph information and optimizing the weight assignment rules, the graph data can be made more in line with actual business scenarios, effectively improving the calculation accuracy and reliability of results when the random walk algorithm is used for risk calculation.

[0107] Specifically, the recursion depth threshold can be understood as a manually set upper limit on the maximum level of equity tracing, used to limit the search scope. Node attributes can be understood as various feature data attached to the graph nodes, primarily the shareholding ratio here. Relationship weights can be understood as numerical parameters used to quantify the degree of correlation between nodes.

[0108] A random walk algorithm was used to perform risk calculations, and the sensitivity risk scores and risk transmission paths of all subjects to be screened were calculated as the risk assessment results.

[0109] Optionally, based on the above embodiments, establishing a transaction association edge between the root node and the counterparty node may include:

[0110] The association weight value is determined based on the ratio of the counterparty's transaction amount to the total transaction volume of the corresponding root node entity; a transaction association edge is established between the root node and the counterparty node, and an association weight value is set for the transaction association edge.

[0111] In this embodiment of the invention, the correlation weight value can be specifically understood as a numerical indicator that quantifies the degree of transaction correlation and the probability of risk transmission between two nodes.

[0112] Specifically, when constructing a transaction association edge between the root node and the counterparty node, the association weight is calculated using the formula of dividing the counterparty's transaction amount by the total transaction volume of the customer (i.e., the total transaction volume of the root node entity). The larger the weight value, the more concentrated the fund flow of the entity to be screened is in the corresponding counterparty, and the higher the probability of risk transmission along this link. After completing the weight assignment, the transaction association edge is set up.

[0113] In addition to this calculation method, factors such as transaction frequency and business cooperation duration can be added to adjust the weights. The calculation results can also be normalized to map the weights to a fixed range of values, making the weight indicators more compatible with the operation rules of the random walk algorithm and optimizing the risk transmission and deduction effect.

[0114] Specifically, the counterparty transaction amount can be understood as the total cumulative transaction amount between the entity to be screened and a single counterparty. The total customer transaction volume can be understood as the total cumulative transaction amount of all transactions between the entity to be screened.

[0115] By determining the weight of transaction-related edges based on the ratio of a single counterparty's transaction amount to the total transaction volume of the root node entity and completing the link construction, the concentration of fund associations between different counterparties and the screened entities can be accurately quantified. The weight values ​​can intuitively reflect the distribution characteristics of fund flows, enabling the fused association graph to objectively reflect the strength of business associations between entities. In the subsequent random walk risk calculation process, the algorithm can distinguish the risk transmission probability of each link based on the differentiated weights, making the calculation results of risk propagation path deduction and sensitive risk scoring more in line with real business scenarios, effectively improving the accuracy and reference value of overall risk assessment.

[0116] S560: Based on key equity field information, screening summary information, and risk assessment results, generate compliance review documents that match the image files of foreign exchange trade documents.

[0117] The technical solution of this invention extracts trade business data based on document image recognition and analysis as the basis for querying. Through multi-dimensional page adaptation, it achieves automated and stable crawling of key equity fields from external equity systems, avoiding automation task failures caused by website redesigns and pop-up interference. Based on business data aggregation, it forms a main entity list to batch complete compliance list screening and obtain screening summaries. By constructing an equity subgraph with the entity to be screened as the root node and limiting the hierarchical recursive retrieval of shareholders, and by introducing existing bank transaction data to supplement counterparty nodes and constructing transaction-related edges, it integrates multiple types of data to construct a relational graph. This solves the data silo problem between equity information and internal transaction data, visually associating and aggregating equity ownership relationships and actual fund transfer relationships. It avoids data redundancy and computational waste caused by infinite equity traversal by limiting the recursive level, and also supplements the fund transfer dimension beyond the equity graph. This provides a complete topological network foundation for subsequent risk walkthrough calculations, facilitating the connection of related entities from the two dimensions of equity nesting and fund transfers, achieving complete mining of hidden related entities, and improving the comprehensiveness and accuracy of subsequent sensitive risk tracing. By combining random walk algorithms to quantitatively calculate the risk scores and transmission links of each entity, the system achieves fully automated compliance due diligence of foreign exchange trade documents. This avoids the inefficiency and human error caused by manual data entry and verification, while also penetrating multiple layers of equity relationships and fund transaction links to uncover hidden risks, accurately locate the source of risk and transmission path, and finally automatically integrate all-dimensional data to generate standardized compliance review documents. This reduces the working time for cross-border trade compliance review, improves the accuracy of risk control identification, and enhances the standardization of review documents.

[0118] Example 3

[0119] Figure 6 This is a flowchart of another foreign exchange trade document screening method provided in Embodiment 3 of the present invention. This embodiment is a refinement of the above embodiment's step of "using a random walk algorithm to perform risk calculation, and calculating the sensitivity risk score and risk transmission path of all entities to be screened." Figure 6 As shown, the method includes:

[0120] S610. Obtain image files of foreign exchange trade documents and extract trade business data through recognition and semantic analysis.

[0121] S620. Using trade business data as the source data for the query, after performing page adaptation operations for the equity information query system, retrieve key equity field information from the equity information query system.

[0122] S630. Construct a list of entities to be screened based on trade business data, and perform a sensitive list query based on the list of entities to be screened to retrieve the screening summary information corresponding to the entities to be screened.

[0123] Furthermore, based on the above embodiments, while retrieving key equity field information, it may also include:

[0124] Capture screenshots of accessed web pages as electronic credentials;

[0125] Accordingly, based on the above embodiments, after retrieving the screening summary information corresponding to the subject to be screened, the process may further include:

[0126] Save the screening summary information as a screening summary form file and download the screening report file simultaneously;

[0127] The screening summary form, screening report, and electronic vouchers are sent back to the bank's internal system for archiving and storage.

[0128] In this embodiment of the invention, the electronic credential can be specifically understood as: a screenshot of a webpage, serving as evidence for online data capture and system operations. The screening summary table file can be specifically understood as: an electronic document that organizes and stores screening summary information in tabular form. The screening report file can be specifically understood as: a complete screening result file issued by an external compliance system, which can be a PDF (Portable Document Format).

[0129] Specifically, when capturing key equity field information, the RPA robot first enters the equity data into a spreadsheet file for storage, and at the same time captures the operation page screen as electronic evidence. The relevant data is cleaned and standardized through the data aggregation node and then sent back to the bank's internal system. After all equity penetration processing is completed, the system automatically updates the business status to pending global compliance list screening.

[0130] The RPA robot then filters business records in the corresponding status and retrieves the list of entities to be screened. It automatically logs into the global compliance list system to perform screening queries, parses the page, saves a summary of the screening results, downloads a detailed screening report, and then organizes the screening summary into a table file. Together with the screening report and electronic vouchers such as webpage screenshots, it is sent back to the bank's internal system to complete the data archiving.

[0131] In addition to the above procedures, timestamps and operation account identifiers can be added to electronic vouchers to enhance their evidentiary validity. Archived files can also be classified and partitioned according to business type, and access permissions can be configured to meet multiple management requirements of bank compliance audits, post-event traceability, and risk control verification.

[0132] Specifically, the data aggregation node can be understood as a transit processing module used to centrally receive, clean, and standardize various externally crawled data. For example, it performs cleaning operations to remove blank fields, garbled characters, duplicate and redundant entries, and invalid abnormal data; corrects errors such as incorrect field formats and inconsistent abbreviations of entity names; standardizes the writing format and unit definitions of all entity names, equity values, business numbers, and dates; and maps heterogeneous data output from different external systems to the bank's preset standard database field templates, generating a standardized, structured dataset with uniform format, well-organized content, and no dirty data. This provides compliant and usable basic data for subsequent graph construction, screening and matching, and risk calculation.

[0133] By simultaneously capturing page screenshots and storing electronic evidence while scraping equity information, generating tabular files after scraping screening summaries, and downloading complete screening reports, and then uniformly uploading all materials back to the internal system for archiving, a complete electronic evidence preservation chain is formed, from original page operations and data storage to document archiving. Webpage screenshots can corroborate the authenticity and validity of data scraping from external systems, while tables and reports retain complete verification values ​​and original compliance documents. The unified archiving model enables one-to-one binding and correspondence of documents, data, and vouchers, which not only meets the financial industry's regulatory requirements for traceable business operations and audit traceability, but also allows for rapid retrieval of the complete set of original materials during subsequent risk reviews and regulatory inspections, reducing the workload of manually organizing paper archives. At the same time, the complete retention of multi-dimensional documents can also provide verifiable original data sources for equity penetration and risk scoring calculations, improving the credibility and rebuttal effectiveness of risk control calculation results.

[0134] S640. Based on equity key field information and screening summary information, combined with the bank's internal historical transaction data, a fusion association graph is constructed. The sanctioned entities in the fusion association graph are marked as risk source nodes. The risk probability propagation is simulated along the equity association edge and the transaction association edge. The operation stops after the probability distribution of all nodes in the graph meets the stability condition. The steady-state probability corresponding to each entity to be screened, which is the root node, is extracted as the sensitive risk score.

[0135] In this embodiment of the invention, the risk source node can be specifically understood as: in the fused association graph, the entity node that is included in the sanctions list and has sensitive violation records is the starting point of risk propagation. The stability condition can be specifically understood as: in the random walk iterative operation, the probability value of each node changes less than a preset minimum threshold in each round, and the overall probability distribution no longer fluctuates significantly, reaching a convergence state.

[0136] Steady-state probability can be understood as: after the operation converges and stabilizes, the fixed probability value obtained by each graph node directly reflects the level of sensitivity risk of the subject.

[0137] S650. Select a preset number of target links according to the sorting rules of sensitive risk scores from high to low, organize the connection relationship of the target links from the risk source node to the corresponding subject node to be screened, and generate a preset number of risk transmission paths as the risk assessment result.

[0138] Specifically, based on key equity field information and screening summary information, combined with the bank's internal historical transaction data, a fusion and correlation graph is constructed.

[0139] The algorithm employs a random walk with restart to perform risk propagation calculations. Entities on the external sanctions list are marked as risk source nodes. Risks spread from the sanctioned entity source node along equity-related edges and transaction-related edges between graph nodes. In each round of walking, a walk transfer probability α is set to move and spread the risk along the related edges, and a restart probability 1−α is set to jump back to the sanctioned entity source node. This constrains the risk score to always radiate outward from the sanctioned entity as the core. The calculation continues until the change in the probability values ​​of all nodes is less than a set threshold and the overall probability distribution tends to stabilize. The steady-state probability value of each screening root node is extracted as the subject's sensitive risk score. Then, the top N propagation links with the highest risk scores are selected and output. The complete connected link from the sanctioned entity through intermediate enterprises and trading counterparties to the screened subject is obtained as the risk transmission path.

[0140] Specifically, the wandering transition probability α can be understood as the probability of risk spreading to adjacent nodes along the equity and transaction association edges in each step of the operation; 1−α is the restart probability of jumping back to the risk source node.

[0141] In a specific example Figure 7This is a schematic diagram of dynamic risk reasoning applicable to an embodiment of the present invention. As shown in the figure, data transmitted by the data aggregation node is received, equity data in the equity information query system is retrieved, and a dynamic relationship graph is constructed based on this data. The entity to be screened is set as the root node, and its indirect controlling shareholder is recursively searched. It is determined whether the recursion depth is less than or equal to 5 levels. If the result is yes, the recursive search for indirect controlling shareholders continues. If the result is no or all shareholders have been recursively searched, a basic equity subgraph is generated. At the same time, historical bank transaction data containing counterparty, amount, and frequency information is introduced, and the counterparty and graph nodes are matched. If the matching node does not exist, the counterparty is added to the graph as a new node and a transaction relationship edge is established. Then, according to... The transaction amount is divided by the total transaction volume of the customer to calculate the transaction edge weight. If a matching judgment node exists, the process directly proceeds to the next step of calculating the transaction edge weight. Then, a complete graph integrating equity and transactions is generated, the RWR algorithm is initialized, and the sanctioned entity is marked as the risk source node. The RWR algorithm is executed to simulate the risk walking along the associated edge. During the process, the risk moves along the edge with probability α and jumps back to the source node with probability 1-α. After updating the node probability distribution in each round, it is judged whether the probability distribution tends to be stable. If it does not tend to be stable, the random walk step with restart is continued iteratively. If it tends to be stable, the calculation is stopped and the steady-state distribution is output. Finally, the steady-state probability value of the root node is extracted to generate a sensitive risk score. The risk propagation path is traced back and the top N high-risk paths are output.

[0142] S660: Based on key equity field information, screening summary information, and risk assessment results, generate compliance review documents that match the image files of foreign exchange trade documents.

[0143] Figure 8 This is a schematic diagram of a foreign exchange trade document screening system applicable to an embodiment of the present invention. As shown in the figure, the bank's internal system layer completes the image upload operation of the foreign exchange trade document. After the document image is uploaded, it enters the OCR stage to complete the extraction and recognition of image text information. The recognized data is sent to the RPA scheduling center inside the intelligent screening engine layer. The RPA scheduling center is connected to the webpage structure change prediction rule library. This rule library is divided into three types of adaptation rules: rule 1, login domain heterogeneity, corresponds to the external system layer query module; rule 2, menu topology drift, corresponds to the navigation and query module; and rule 3, pop-up evolution interception, corresponds to the data capture and download module. The external system layer query module, navigation and query module, and data capture and download module are all connected to the equity information query system and the global compliance list system in the external system layer. The three work together to complete the equity data acquisition. The acquired equity data is sent to the dynamic risk inference engine to carry out risk calculation and analysis. Finally, the dynamic risk inference engine outputs a sensitive risk report. The core links of the entire data interaction, automated query, and risk calculation are all encapsulated and run within the intelligent screening engine layer.

[0144] The entire system is divided into three layers: the internal bank system layer, the intelligent screening engine layer, and the external system layer. The layers rely on secure data interfaces to complete information exchange and transmission. The internal bank system layer is the origin of the entire business process. It includes a document image upload module and an OCR recognition module. Business personnel upload electronic image files of trade background documents through the document image upload module. The document image data is sent to the OCR recognition module for parsing, extracting key content such as the subject name and product information, and generating structured data, which is then passed down to the intelligent screening engine layer.

[0145] The intelligent screening engine layer integrates RPA's automated execution capabilities and the intelligent decision-making capabilities of the predictive rule base. Internally, it sets up an RPA scheduling center and a webpage structure change prediction rule base. The RPA scheduling center acts as an execution controller, receiving structured data from the upper-layer OCR recognition module and uniformly coordinating various automated tasks. The webpage structure change prediction rule base and the RPA scheduling center are electrically connected, enabling the RPA scheduling center to predict the webpage environment status in advance. When structural changes such as page ID changes or menu hierarchy shifts are detected, the element positioning strategy is switched immediately. If no structural changes are detected but the task execution fails, a degradation detection mechanism is activated to collect new page features and write them back to update the rule base. This achieves the effect of the rule base learning autonomously and self-correcting operational faults.

[0146] Figure 9 This is a schematic diagram of an adaptive process for screening foreign exchange trade documents applicable to an embodiment of the present invention. As shown in the figure, the RPA robot starts executing the task, loads the prediction rule base, performs a pre-scan of the page, and enters the rule base matching and judgment stage. If the matching result is yes, it means that the web page structure has been predicted to change, and the adaptive strategy is called. If the matching result is no, it is determined that the page structure is stable, and the standard RPA script is directly executed to build at least one menu access path. After the two branches are processed, they enter the step of executing interactive actions. After the interactive actions are completed, the execution result is verified. If the verification result is successful, it directly advances to the next business step. If the verification result is unsuccessful, the system first captures the exception and saves the exception snapshot, starts the degradation detection logic, updates the rule base feature information according to the detected new page features, and then returns to the step of executing interactive actions to retry the entire operation. For example, after the system triggers the downgrade detection logic, it automatically scans and identifies new page features such as new element identifiers, hierarchical layout, and control positions on the current webpage. The collected new feature data is formatted and organized and written into the webpage structure change prediction rule base. The feature information inside the rule base is iteratively updated. After the entire update process is completed, the program jumps back to the interactive action steps and runs the entire set of interactive operations such as login, query, and crawling again for retry execution.

[0147] The execution module is divided into a login authentication module, a menu navigation and positioning module, and a data capture and report download module. All three work separately under the guidance of the positioning strategy output by the webpage structure change prediction rule library, and sequentially complete three basic automated operations: external system account login verification, page menu element positioning and redirection, and target business data capture and compliance screening report download.

[0148] The raw data collected by the data capture and report download module is uniformly aggregated and sent to the data aggregation node for equity data acquisition. The node will clean and remove dirty data from the multi-source messy raw data, standardize and regulate the field format, and output a unified and standardized structured equity and screening dataset to provide qualified input data for downstream risk calculation.

[0149] The dynamic risk inference engine receives structured equity and screening business data transmitted from the data aggregation node, constructs a dynamic relationship graph, reads shareholder names, shareholding ratios, and equity levels from the equity information query system, and retrieves historical transaction records of customers within the bank, including counterparties, transaction amounts, and transaction frequencies. The entity to be screened is set as the root node of the graph, and direct and indirect shareholders are recursively retrieved. A maximum recursion depth of 5 levels is set to avoid infinite recursion causing computational overload, and a basic equity subgraph with node attributes bound to shareholding ratios is generated.

[0150] Set the weight of the graph transaction relationship, match the name of the counterparty with the node of the equity subgraph one by one. If no match is found, add a new counterparty node and build a transaction relationship edge between the root node and the new node. The weight of the transaction edge is calculated by dividing the counterparty's transaction amount by the total transaction volume of the customer. The higher the weight, the higher the degree of concentration of the screening entity's funds towards that counterparty and the greater the probability of risk transmission.

[0151] The risk propagation RWR algorithm is executed, marking entities on the external sanctions list as risk source nodes. The RWR algorithm is run to simulate risk diffusion. The risk travels from the sanction source node along the equity edge and the transaction edge. At each step, there is an α probability of spreading to the adjacent node and a 1−α probability of jumping back to the risk source node, ensuring that the risk score radiates outward around the sanctioned entity. The calculation stops when the probability change of each node is less than the preset threshold and the probability distribution is stable.

[0152] Extract the output risk results, take the steady-state probability value of each screening root node after convergence as the corresponding subject's quantitative sensitive risk score, and select the top N risk propagation links with the highest output scores.

[0153] The results output module receives the risk score and transmission path analysis results calculated by the dynamic risk reasoning engine, integrates various data such as equity penetration data, screening reports, and risk scores, and generates a complete and standardized formal sensitive risk report after unified formatting.

[0154] The method of this invention solves the industry problem of RPA instability, achieving cross-system adaptive and self-healing operation. Traditional RPA is prone to crashing when encountering changes to external web pages. This invention builds a web page structure change prediction rule base, combined with a dual positioning mode of visual anchors and semantic tags, NLP menu fuzzy matching technology, and DOM overlay context interception mechanism. It can intelligently adapt to and handle three types of page anomaly scenarios: heterogeneous login domains, menu topology drift, and sudden pop-ups. This gives RPA the ability to autonomously cope with web page structure changes and self-heal, replacing the traditional mode of relying on manual intervention to restart the program after a web page anomaly with the system's autonomous degradation detection. The automatic switching of operating strategies significantly enhances the stability and uninterrupted operation of automated processes involving multiple external systems. In traditional screening methods, equity data and transaction data are separated and independent, making it difficult to uncover hidden transmission risks. This invention sets up a dynamic risk inference engine module to accurately calculate the sensitive risk score of the subject and trace the complete risk propagation chain. It upgrades the original shallow screening mode, which only makes static comparisons with blacklists, to a dynamic deep inference mode that integrates multiple data. This mode can accurately identify hidden sensitive risks in multi-layered equity nesting structures, making the bank's risk control judgment more accurate, and at the same time, it can predict potential risks in advance and improve the foresight of risk control work.

[0155] The technical solution of this invention uses trade business data extracted from document image recognition and analysis as the basis for querying. Through multi-dimensional page adaptation, it achieves automated and stable crawling of key equity fields from external equity systems, avoiding automation task failures caused by website redesigns and pop-up interference. Based on business data aggregation, it forms a main entity list to batch complete compliance list screening and obtain screening summaries. It integrates equity information, screening results, and existing bank transaction data to construct a two-dimensional correlation graph of equity and transactions. By setting sanctioned entities in the graph as risk source nodes, it simulates risk probability transmission based on equity and transaction correlation edges. After node probabilities converge, it extracts the steady-state probability of the root node to quantify risk scores, and selects the best links according to scores to form risk transmission paths. It uses the existing ownership and capital relationships in the graph to achieve quantitative risk transmission deduction, overcoming the limitations of traditional manual subjective judgment of risk levels based on experience. It quantifies the sensitivity level of the entity with objective probability values, and simultaneously selects key connected links to intuitively sort the risk flow from the source of sanctions to the screened entity, locating hidden cross-entity risk transmission chains, and improving the objectivity of sensitive risk classification and the targeting of risk tracing. Ultimately, it automatically integrates data from all dimensions to generate standardized compliance review documents, reducing the time required for cross-border trade compliance reviews and improving the accuracy of risk control identification and the standardization of review documents.

[0156] Example 4

[0157] Figure 10 This is a schematic diagram of a foreign exchange trade document screening device provided in Embodiment 4 of the present invention. Figure 10 As shown, the device includes: an image recognition module 1010, a shareholding capture module 1020, a screening summary module 1030, a risk assessment module 1040, and a document review module 1050, wherein:

[0158] The image recognition module 1010 is used to acquire image files of foreign exchange trade documents and extract trade business data through recognition and semantic analysis.

[0159] The equity capture module 1020 is used to use trade business data as the source data for querying. After performing page adaptation operations for the equity information query system, it captures key equity field information from the equity information query system.

[0160] The screening summary module 1030 is used to construct a list of entities to be screened based on trade business data, perform a sensitive list query based on the list of entities to be screened, and retrieve the screening summary information corresponding to the entities to be screened.

[0161] The risk assessment module 1040 is used to construct a fusion correlation graph based on equity key field information and screening summary information, combined with the bank's internal stock of historical transaction data, and to perform risk calculation using a random walk algorithm to calculate the sensitive risk score and risk transmission path of all entities to be screened, as the risk assessment result.

[0162] The document review module 1050 is used to generate compliance review documents that match the image files of foreign exchange trade documents based on key equity fields, screening summary information, and risk assessment results.

[0163] The technical solution of this invention extracts trade business data based on document image recognition and analysis as the basis for querying. Through multi-dimensional page adaptation, it achieves automated and stable crawling of key equity fields from external equity systems, avoiding automation task failures caused by website redesigns and pop-up interference. Based on the collection of business data, it forms a list of entities to complete batch compliance list screening and obtain screening summaries. It integrates equity information, screening results, and existing bank transaction data to build a two-dimensional correlation graph of equity and transactions, and combines a random walk algorithm to quantitatively calculate the risk scores and transmission links of each entity. This achieves fully automated compliance due diligence of foreign exchange trade documents, avoiding the inefficiency and human error caused by manual data entry and verification. It can also penetrate multiple layers of equity relationships and fund transaction links to uncover hidden risks, accurately locate the source of risk and transmission path, and finally automatically integrate all-dimensional data to generate standardized compliance review documents. This reduces the working time of cross-border trade compliance review, improves the accuracy of risk control identification, and enhances the standardization of review documents.

[0164] Based on the above embodiments, the equity capture module 1020 is specifically used for:

[0165] The login page of the equity information query system is scanned using full-screen optical character recognition (OCR) to retrieve relevant keywords. The input box position is determined based on the relative coordinate offset of the keywords to complete the account and password input. The system is then connected to the equity information query system to extract key equity fields from the query source data. These key fields include at least the following: shareholder name, registered region, entity type, and direct shareholding data. After cleaning and standardizing the key equity fields, the standardized information is sent back to the bank's internal system.

[0166] Optionally, based on the above embodiments, the equity capture module 1020 may include: a fuzzy comparison unit, wherein:

[0167] The fuzzy comparison unit is used to analyze the semantics of the menu text in the equity information query system using a natural language semantic fuzzy matching algorithm before retrieving key equity field information based on the query source data. It performs a fuzzy comparison between preset menu keywords and existing menu content on the page to construct at least one menu access path.

[0168] Accordingly, based on the above embodiments, the equity capture module 1020 is further used for:

[0169] The user is redirected to the equity information query page in the equity information query system via the menu access path; the element structure of the equity information query page is traversed layer by layer through the document object model, and floating and pop-up-type obstructing controls are identified and closed or removed; the query source data is entered, an equity information retrieval request is initiated, and key equity field information is extracted from the retrieval results returned by the page.

[0170] Based on the above embodiments, the risk assessment module 1040 is specifically used for:

[0171] Using each entity to be screened as the root node, the system recursively retrieves direct and indirect shareholder information at a preset depth based on key equity fields to construct an equity subgraph. It also obtains existing historical transaction data from within the bank to identify the counterparties corresponding to each entity to be screened. When a counterparty does not exist in the matching equity subgraph, a new counterparty node is added, and a transaction association edge is established between the root node and the counterparty node.

[0172] Based on the above embodiments, the risk assessment module 1040 is further used for:

[0173] The association weight value is determined based on the ratio of the counterparty's transaction amount to the total transaction volume of the corresponding root node entity; a transaction association edge is established between the root node and the counterparty node, and an association weight value is set for the transaction association edge.

[0174] Based on the above embodiments, the risk assessment module 1040 is further used for:

[0175] The sanctioned entities within the fusion association graph are marked as risk source nodes. Risk probability propagation is simulated along equity association edges and transaction association edges. The calculation stops after the probability distribution of all nodes in the graph meets the stability condition. The steady-state probability corresponding to each entity to be screened, which is the root node, is extracted as the sensitive risk score. A preset number of target links are selected according to the sorting rule of sensitive risk scores from high to low. The connectivity of the target links from the risk source nodes to the corresponding entity nodes to be screened is sorted out to generate a preset number of risk transmission paths.

[0176] Furthermore, based on the above embodiments, the foreign exchange trade document screening device may further include: an electronic voucher module, wherein:

[0177] The electronic credentials module is used to capture the access page screen as electronic credentials while capturing key equity field information;

[0178] Accordingly, based on the above embodiments, the foreign exchange trade document screening device may further include: a report file module and an archive storage module, wherein:

[0179] The report file module is used to save the screening summary information as a screening summary table file after retrieving the screening summary information corresponding to the subject to be screened, and to download the screening report file at the same time;

[0180] The archiving and storage module is used to send screening summary form files, screening report files, and electronic vouchers back to the bank's internal system for archiving and storage.

[0181] The foreign exchange trade document screening device provided in this embodiment of the invention can execute the foreign exchange trade document screening method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0182] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0183] In the technical solution disclosed herein, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with the relevant laws, regulations and standards of the relevant countries and regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse.

[0184] In the technical solution disclosed herein, if automated decision-making is involved, a corresponding operation entry will be provided to the user, allowing the user to choose to agree to or reject the automated decision result; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0185] Example 5

[0186] Figure 11 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0187] like Figure 11 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0188] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0189] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the foreign exchange trade document screening method, i.e.:

[0190] Acquire image files of foreign exchange trade documents and extract trade business data through recognition and semantic analysis;

[0191] Using trade business data as the source data for the query, and after performing page adaptation operations for the equity information query system, the key equity field information is retrieved from the equity information query system.

[0192] A list of entities to be screened is constructed based on trade business data. A sensitive list query is performed based on the list of entities to be screened to retrieve the screening summary information corresponding to the entities to be screened.

[0193] Based on key equity field information and screening summary information, combined with the bank's internal historical transaction data, a fusion correlation graph is constructed, and a random walk algorithm is used to perform risk calculation to calculate the sensitive risk score and risk transmission path of all entities to be screened, which serves as the risk assessment result.

[0194] Based on key equity field information, screening summary information, and risk assessment results, a compliance review document matching the image file of foreign exchange trade documents is generated.

[0195] In some embodiments, the foreign exchange trade document screening method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the foreign exchange trade document screening method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the foreign exchange trade document screening method by any other suitable means (e.g., by means of firmware).

[0196] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0197] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0198] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0199] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0200] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0201] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0202] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0203] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for screening foreign exchange trade documents, characterized in that, include: Acquire image files of foreign exchange trade documents and extract trade business data through recognition and semantic analysis; Using trade business data as the source data for the query, and after performing page adaptation operations for the equity information query system, the key equity field information is retrieved from the equity information query system. A list of entities to be screened is constructed based on trade business data. A sensitive list query is performed based on the list of entities to be screened to retrieve the screening summary information corresponding to the entities to be screened. Based on key equity field information and screening summary information, combined with the bank's internal historical transaction data, a fusion correlation graph is constructed, and a random walk algorithm is used to perform risk calculation to calculate the sensitive risk score and risk transmission path of all entities to be screened, which serves as the risk assessment result. Based on key equity field information, screening summary information, and risk assessment results, a compliance review document matching the image file of foreign exchange trade documents is generated.

2. The method according to claim 1, characterized in that, After performing page adaptation operations for the equity information query system, key equity field information is retrieved from the equity information query system, including: The login page of the equity information query system is subjected to full-screen optical character recognition scanning to retrieve login-related text keywords. The position of the input box is determined based on the relative coordinate offset of the keywords to complete the input of account and password. Connect to the equity information query system and extract key equity field information based on the query source data; Among them, the key equity information fields should include at least: shareholder name, registered region, entity type, and direct shareholding data; After the key equity field information is cleaned and standardized, the standardized key equity field information is sent back to the bank's internal system.

3. The method according to claim 2, characterized in that, Before extracting key equity fields from the source data, the process also includes: Using a natural language semantic fuzzy matching algorithm, the semantics of menu text in the equity information query system are analyzed, and the preset menu keywords are fuzzy compared with the existing menu content on the page to construct at least one menu access path. Accordingly, based on the source data, key equity field information is extracted, including: Accessing the equity information query page of the equity information query system via the menu path; By traversing the element structure of the equity information query page layer by layer using the document object model, we can identify floating and pop-up-type obstructing controls and perform closing and removal operations. Enter the source data for the query, initiate an equity information retrieval request, and extract key equity field information from the retrieval results returned by the page.

4. The method according to claim 1, characterized in that, Based on key equity field information and screening summary information, combined with the bank's internal historical transaction data, a fusion correlation graph is constructed, including: Using each entity to be screened as the root node, the direct and indirect shareholder information at a preset level and depth is recursively retrieved based on the key equity field information to construct an equity subgraph; Obtain the bank's internal historical transaction data, identify the counterparties corresponding to each entity to be screened, and when a counterparty does not exist in the matching equity subgraph, add a counterparty node and establish a transaction association edge between the root node and the counterparty node.

5. The method according to claim 4, characterized in that, Establish transaction association edges between the root node and the counterpart node, including: The association weight value is determined based on the ratio of the counterparty's transaction amount to the total transaction volume of the corresponding root node entity; Establish a transaction association edge between the root node and the counterpart node, and set the association weight value for the transaction association edge.

6. The method according to claim 1, characterized in that, A random walk algorithm was used for risk calculation to determine the sensitivity risk scores and risk transmission paths for all entities to be screened, including: The sanctioned entities within the fusion association graph are marked as risk source nodes. The risk probability propagation is simulated along the equity association edge and the transaction association edge. The operation stops after the probability distribution of all nodes in the graph meets the stability condition. The steady-state probability corresponding to each entity to be screened, which is the root node, is extracted as the sensitive risk score. Based on the ranking rules of sensitive risk scores from high to low, a predetermined number of target links are selected, and the connectivity of the target links from the risk source node to the corresponding subject node to be screened is sorted out to generate a predetermined number of risk transmission paths.

7. The method according to claim 1, characterized in that, In addition to extracting key equity field information, it also includes: Capture screenshots of accessed web pages as electronic credentials; Accordingly, after retrieving the screening summary information corresponding to the subject to be screened, the process also includes: Save the screening summary information as a screening summary form file and download the screening report file simultaneously; The screening summary form, screening report, and electronic vouchers are sent back to the bank's internal system for archiving and storage.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the foreign exchange trade document screening method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the foreign exchange trade document screening method according to any one of claims 1-7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the foreign exchange trade document screening method according to any one of claims 1-7.

Images