An intelligent single auditing method for cross-border trade finance based on multi-modal feature fusion

Abstract

The application discloses a kind of based on multi-modal feature fusion's cross-border trade finance intelligent single method for examining, it is related to computer processing technical field, including following steps: S01, acquisition in cross-border trade finance business is to be examined single data, the single data to be examined includes text modal single data, image modal single data and numerical modal single data;S02, the data set is obtained by feature extraction to the single data to be examined, then again based on the first level check rule of confidence score cross check is carried out to the data set, generates first check confidence.The application to high-risk instant interception and flexible processing to slight risk, effectively overcome the defect that traditional single system only has pass / reject binary conclusion, easy to cause injury or miss detection, significantly improve the fine degree of single decision and business adaptability.

Description

Technical Field

[0001] This invention relates to the field of computer processing technology, and in particular to an intelligent document review method for cross-border trade finance based on multimodal feature fusion. Background Technology

[0002] In recent years, some financial institutions have introduced automated document review systems based on single-modal data. Existing technologies use optical character recognition (OCR) technology to extract key fields from invoices or bills of lading and then compare them with preset rules, or use natural language processing models to identify risk clauses in the text of trade contracts.

[0003] Referring to a published Chinese patent, publication number CN120047235A, publication date May 27, 2025, a method, apparatus, device, and storage medium for verifying trade background information are disclosed, which are applied in the field of financial technology. The method includes: acquiring multimodal trade background data of customers, wherein the multimodal trade background data includes text trade background data, image trade background data, and video trade background data; and extracting different parameters from the trade background data according to the modality using a trade background information extraction model.

[0004] The shortcomings of existing technologies are that although they have achieved the extraction and fusion of multimodal features, they have not quantitatively evaluated the quality of each modality feature, namely OCR recognition confidence, image clarity and text semantic reliability. Furthermore, they have not triggered a closed-loop mechanism to supplement collection or re-verification when feature confidence is insufficient, which leads to the risk of misjudgment in the fusion model when the input quality is uneven. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention adopts a cross-border trade finance intelligent document review method based on multimodal feature fusion, comprising the following steps: S01. Collect pending document data in cross-border trade finance business, including text modal document data, image modal document data and numerical modal document data; S02. Extract features from the data of the documents to be reviewed to obtain a dataset, and then perform cross-validation on the dataset based on the first-level validation rules of the confidence score to generate the first validation confidence. When the first verification confidence of any feature contained in the dataset is lower than the preset confidence threshold, a supplementary acquisition signal is triggered to obtain supplementary document data and re-execute feature extraction and verification. S03. Input the feature data contained in the dataset with the first verification confidence level higher than the confidence level threshold into the preset multimodal feature fusion model, and align and fuse the features of different modalities through the multimodal feature fusion model to generate a feature set; S04. Compare the feature set with the preset risk feature library, calculate the matching degree between the feature set and the risk template in the risk feature library that matches the corresponding feature, and identify one or more potential risk points in the data of the document to be reviewed based on the matching degree. S05. Based on the identified potential risk points and in conjunction with the preset document review decision rule base, generate the final document review conclusion, and simultaneously locate and output the specific location of the potential risk points in the document data to be reviewed and the corresponding risk reasons.

[0006] Preferably, the text modal document data includes unstructured text from trade contracts, commercial invoices, and packing lists; The image modal document data includes scanned copies of the bill of lading and certificate of origin with a stamp. The numerical modal document data includes invoice amount, quantity, unit price, and tax rate.

[0007] Preferably, in step S02, text features are extracted from the text modal document data, image features are extracted from the image modal document data, and numerical features are extracted from the numerical modal document data. The extracted text features, image features, and numerical features are then cross-validated using a first-level verification rule based on confidence scoring to generate a first verification confidence level. When the first verification confidence level of any feature is lower than the preset confidence threshold, a supplementary acquisition signal is triggered to obtain supplementary document data, and feature extraction and verification are re-executed, including the following steps: S21. A bidirectional encoder representation model based on a converter is used to extract semantic features from text modal document data to obtain text feature vectors; S22. Use a convolutional neural network to extract features of seals, signatures and key field areas from image modal document data to obtain image feature vectors; S23. Process the numerical modal data to obtain numerical feature vectors; S24. Compare the text feature vector, image feature vector, and numerical feature vector with the corresponding preset business rules, which include document consistency rules, logical self-consistency rules, and historical data verification rules, and generate preliminary verification results.

[0008] Preferably, the multimodal feature fusion model in S03 dynamically assigns weights to features of different modalities through an attention mechanism, and concatenates the multimodal features after weight assignment and maps them to a fully connected layer to generate the feature set. The aforementioned multimodal feature fusion model is constructed by collecting historical cross-border trade financial document data and corresponding historical risk labeling data as training samples. Through supervised learning, with the goal of minimizing risk identification error, the network parameters of the initial fusion model are iteratively optimized to complete the training.

[0009] Preferably, the preset risk feature library in S04 includes multiple risk categories, each risk category corresponds to a risk template, and the risk template is composed of a set of typical features extracted from historical risk events; The matching degree is the cosine similarity or Euclidean distance between the feature set and the risk template.

[0010] Preferably, in step S04, when the matching degree between the feature set and any risk template exceeds a preset first threshold, it is determined that there is a corresponding potential risk point. When all matching scores are below the preset second threshold, the system is deemed risk-free. The first threshold is greater than the second threshold.

[0011] Preferably, the order review decision rule base in S05 includes multiple decision paths, which are combined according to the type, number and severity level of potential risk points; When no potential risks are identified, the review conclusion is "approved". When one or more potential risk points of the first pre-defined type exist, the review conclusion is rejection; When there are one or more potential risk points of the second preset type and there are no potential risk points of the first preset type, the review conclusion is manual review.

[0012] Preferably, the location output of the potential risk point in S05 includes outputting the starting character index of the risk text when the risk point originates from a text modality; When the risk point originates from the image modality, a heatmap or bounding box is overlaid on the image for annotation; when the risk point originates from the numerical modality, the cell or field name where the abnormal value is located is output.

[0013] Preferably, the order review decision rule base in step S05 includes decision paths based on the type, number, and severity level of potential risk points, and the output of the decision paths is dynamically corrected by the second verification confidence level. The second verification confidence score is obtained by performing a reverse confidence score evaluation on the generation process of the feature set, and is used to characterize the reliability of the output results of the multimodal feature fusion model. When the confidence level of the second verification is lower than the decision threshold, the review conclusion will be forcibly switched to manual review and a low confidence level warning label will be attached.

[0014] Preferably, the dataset includes text features extracted from the text modal document data, image features extracted from the image modal document data, and numerical features extracted from the numerical modal document data.

[0015] The present invention has at least the following beneficial effects: Cross-validation is performed on text, image, and numerical features to generate an initial confidence score. When the confidence score of any feature falls below a threshold, supplementary signal acquisition is automatically triggered, and extraction and validation are re-executed. This ensures the input quality of multimodal feature fusion, fundamentally avoiding the risk of misjudgment due to abnormal or low-quality single-modal data, and significantly improving the robustness and reliability of the system in real and complex cross-border trade scenarios.

[0016] The first confidence level is used to screen the input quality before feature fusion. The second confidence level quantifies the reliability of the multimodal feature fusion model's output by inversely evaluating the generation process of the comprehensive feature vector. When the second confidence level falls below the decision threshold, the review conclusion is forcibly switched to manual review and an early warning label is added. This achieves intelligent integration between automated review and manual review, maximizing review efficiency while ensuring accuracy, and providing a traceable basis for decision-making reliability for subsequent regulatory audits.

[0017] High matching scores (exceeding the first threshold) trigger risk warnings, low matching scores (all below the second threshold) are deemed risk-free, and scores in between are entered into the decision-making process based on risk type and quantity. Combined with cross-validation of multimodal features and dynamic confidence level correction, this effectively solves the problems of coarse risk identification granularity, high false alarm rates, and inability to adaptively handle complex edge cases in traditional document review methods, significantly improving the accuracy and business adaptability of intelligent document review in cross-border trade finance. Attached Figure Description

[0018] 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.

[0019] Figure 1 The flowchart of a cross-border trade finance intelligent document review method based on multimodal feature fusion provided in Embodiment 1 of the present invention is shown below. Figure 2 The flowchart of S02 provided in Embodiment 1 of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0021] 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 embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "including," "having," and any variations are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0022] Example 1

[0023] This embodiment provides a cross-border trade finance intelligent document review method based on multimodal feature fusion. The method includes the following steps: Figure 1 As shown: S01. Collect pending document data in cross-border trade finance transactions. The pending document data includes: Text-based modal document data, which includes unstructured text in trade contracts, commercial invoices, and packing lists; Image modal document data, including scanned copies of stamped bills of lading and certificates of origin; Numerical modal document data, which includes invoice amount, quantity, unit price and tax rate.

[0024] Specifically, the system connects to enterprise resource planning systems, bank documentation centers, third-party e-commerce platforms, and logistics and warehousing systems through API interfaces, SFTP file transfer protocol, message queues, or RPA (Robotic Process Automation) systems. Collected data is categorized, stored, and preprocessed according to modality. Text-based document data: This mainly includes unstructured or semi-structured text content from trade contracts, commercial invoices, packing lists, and certificates of origin. These texts are typically in PDF, Word, Excel, or plain text format.

[0025] Image-based document data primarily includes scanned copies of paper bills of lading bearing physical or electronic seals, stamped certificates of origin, and customs declarations with handwritten signatures. These documents are typically in JPEG, PNG, or TIFF formats, or are scanned pages embedded in PDFs. The system uses image preprocessing algorithms (such as adaptive binarization, noise reduction, and tilt correction) to improve image quality and leverages deep learning-based layout analysis models (such as Mask R-CNN or YOLO series) to automatically locate the seal areas, signature areas, and bounding boxes containing key fields within the images.

[0026] Numerical modal document data: This mainly includes structured or semi-structured numerical information such as invoice amount, quantity, unit price, tax rate, total weight, and container quantity. It includes plain numbers in table cells or numbers recognized from image documents via OCR. The system extracts numerical values ​​and their corresponding units of measurement, currency, and contextual labels from text and image parsing results using regular expression matching, named entity recognition (NER), or rule-based numerical extractors.

[0027] S02. Perform feature extraction on the data of the documents to be reviewed to obtain a dataset (the dataset includes text features extracted from text modal documents, image features extracted from image modal documents, and numerical features extracted from numerical modal documents). Then, perform cross-validation on the dataset based on the first-level validation rules of the confidence score to generate the first validation confidence score. When the first verification confidence level of any feature in the dataset falls below a preset confidence threshold, a supplementary acquisition signal is triggered to obtain supplementary document data, and feature extraction and verification are re-executed, including the following steps (such as...). Figure 2 (as shown) S21. A bidirectional encoder representation model based on a converter is used to extract semantic features from text modal document data to obtain text feature vectors; S22. Use a convolutional neural network to extract features of seals, signatures and key field areas from image modal document data to obtain image feature vectors; S23. Process the numerical modal data to obtain numerical feature vectors; S24. Compare the text feature vector, image feature vector, and numerical feature vector with the corresponding preset business rules. The preset business rules include document consistency rules, logical self-consistency rules, and historical data verification rules, and generate preliminary verification results.

[0028] Specifically, for text-based document data (such as unstructured text in trade contracts, commercial invoices, and packing lists), the Transformer-based Bidirectional Encoder Representation Model (BERT) is used for semantic feature extraction. The BERT model, pre-trained on massive general corpora, can capture deep semantic relationships between words in context. Its operating principle involves segmenting the text into tokens, generating an embedding vector containing positional encoding for each token, and then performing self-attention computation through a multi-layer bidirectional Transformer encoder. This allows each token to focus on information from all other tokens in the sequence, resulting in a token-level feature vector that incorporates the contextual information. In the document review scenario of this invention, the BERT model outputs a sentence-level feature vector for the entire text (usually labeled with [CLS]), which encodes key semantic information such as product names in invoices, trade terms in contracts, and specifications in packing lists.

[0029] For image-based document data (such as scanned copies of stamped bills of lading and certificates of origin), convolutional neural networks (CNNs) are used for feature extraction. CNNs automatically learn local to global features of an image through a hierarchical structure of convolutional layers, pooling layers, and fully connected layers. The operating principle is as follows: first, the image is input as a three-channel (RGB) pixel matrix; convolutional kernels (such as 3×3 or 5×5) slide across the image to perform convolution operations, extracting low-level features such as edges, corners, and textures. As the network deepens, higher-level convolutional kernels can combine low-level features to form more complex structures, such as the circular outline of a stamp, the stroke direction of a signature, and the text borders of key field areas. This invention particularly focuses on the detection and feature extraction of stamps, signatures, and key field areas. Therefore, a Region Proposal Network (RPN) can be introduced into the CNN, or an attention-based object detection module can be used to locate sensitive areas in the image and output corresponding feature vectors.

[0030] Outlier detection is performed on numerical fields (such as removing negative values ​​or values ​​that are outside the reasonable range), and then Z-score standardization (subtracting the mean and dividing by the standard deviation) or Min-Max normalization (scaling to the [0,1] interval) is used to make all numerical features comparable in scale.

[0031] After extracting the three types of features, this invention enters the first-level verification stage. This stage employs cross-validation rules based on confidence scores to perform multi-dimensional logical consistency checks on text feature vectors, image feature vectors, and numerical feature vectors, generating a first verification confidence score. The verification rules include three main categories: document consistency rules (checking whether corresponding fields match between different documents, such as whether the buyer's name in the contract matches the consignee on the bill of lading), logical self-consistency rules (checking whether there are logical contradictions within the same document, such as whether the quantity × unit price on the invoice equals the total price, and whether the numerical value matches the text description), and historical data verification rules (comparing the current document features with historical normal data of the same trading party and similar goods to determine if there are any abnormal deviations). Each rule outputs a sub-confidence score (a score between 0 and 1) after comparison, and the final first verification confidence score can be the weighted average or minimum value of all sub-confidence scores.

[0032] S03. Input the feature data contained in the dataset with the first verification confidence level higher than the confidence level threshold into the preset multimodal feature fusion model, and align and fuse the features of different modalities through the multimodal feature fusion model to generate a feature set.

[0033] Furthermore, the aforementioned multimodal feature fusion model dynamically assigns weights to features of different modalities through an attention mechanism, and concatenates the weighted multimodal features and maps them to a fully connected layer to generate a feature set. The aforementioned multimodal feature fusion model is constructed by collecting historical cross-border trade financial document data and corresponding historical risk labeling data as training samples. Through supervised learning, with the goal of minimizing risk identification error, the network parameters of the initial fusion model are iteratively optimized to complete the training.

[0034] Specifically, after completing the first-level confidence screening, for all text feature vectors, image feature vectors, and numerical feature vectors with confidence scores higher than a set threshold, the system inputs them into a pre-trained multimodal feature fusion model. In its operation, the model first receives outputs from three feature extraction branches: text feature vectors (e.g., 768-dimensional semantic vectors generated by the BERT model), image feature vectors (e.g., 512-dimensional spatial structure features extracted by a convolutional neural network), and numerical feature vectors (e.g., vectors composed of normalized scalar values ​​such as amount, quantity, and unit price). Because the dimensions and distributions of different modal features differ, the model first maps each modal feature to the same dimensional space through independent linear transformation layers, making subsequent attention calculations comparable.

[0035] Furthermore, the model generates corresponding query vectors, key vectors, and value vectors for each modality. Taking the text modality as an example, its query vector is used to evaluate the importance of the current text features in explaining other modal features (such as whether the seal in the image matches the seller in the text); while the text key vector is used to be matched with queries from other modalities. By calculating the dot product similarity between each modal query and all modal keys, and then normalizing it using Softmax, a set of dynamic weight coefficients is obtained. These coefficients reflect the relative contribution of different modal features to the fusion task within the current trade document context.

[0036] The model generates corresponding query vectors, key vectors, and value vectors for each modality. Taking the text modality as an example, its query vector is used to evaluate the importance of the current text features in explaining the features of other modalities (such as whether the seal in the image matches the seller in the text); while the text key vector is used to match queries from other modalities. By calculating the dot product similarity between each modal query and all modal keys, and then normalizing it using Softmax, a set of dynamic weight coefficients is obtained.

[0037] S04. Compare the feature set with the preset risk feature library, calculate the matching degree between the feature set and the risk template in the risk feature library that matches the corresponding feature, and identify one or more potential risk points in the data of the document to be reviewed based on the matching degree.

[0038] The preset risk feature library in the above embodiments includes multiple risk categories, each risk category corresponds to a risk template, and the risk template is composed of a set of typical features extracted from historical risk events; The matching degree is the cosine similarity or Euclidean distance between the feature set and the risk template.

[0039] Specifically, the system compares the comprehensive feature vector output by the multimodal feature fusion model with a pre-set risk feature library to identify potential risk points in the documents under review. The construction of the risk feature library is the core foundation of this step. This feature library is pre-built offline and includes: collecting a large number of documents related to confirmed risk events in historical cross-border trade finance transactions (such as fraudulent trade, document discrepancies, duplicate financing, and amount tampering); extracting and fusing the multimodal document data for each type of risk event in the same manner as steps S01 to S03 to generate a comprehensive feature vector; these comprehensive feature vectors constitute the risk template for the corresponding risk category.

[0040] S05. Based on the identified potential risk points and combined with the preset document review decision rule base, generate the final document review conclusion, and simultaneously locate and output the specific location of the potential risk points in the document data to be reviewed and the corresponding risk reasons.

[0041] Furthermore, the order review decision rule base in the embodiment includes multiple decision paths, which are combined according to the type, number, and severity level of potential risk points; When no potential risks are identified, the review conclusion is "approved". When one or more potential risk points of the first pre-defined type exist, the review conclusion is rejection; When there are one or more potential risk points of the second preset type and there are no potential risk points of the first preset type, the review conclusion is manual review.

[0042] Specifically, the document review decision rule base essentially encodes business expert experience and regulatory compliance requirements into several decision paths that can be executed by the system. Each decision path is triggered by a combination of conditions across three dimensions: the type, quantity, and severity level of the risk points, and corresponds to a unique review conclusion (approval, manual review, or rejection). During runtime, all identified potential risk points are first structurally analyzed: the category label for each risk point is extracted, the frequency of occurrence of each category of risk points is counted, and a severity level is assigned to each risk point according to a preset risk level mapping table.

[0043] The system first iterates through all risk points to determine if any belong to the first preset risk type. These risk points typically correspond to absolute rejections such as fraud, forgery, or major compliance violations. If at least one risk point of the first preset type exists, regardless of the other risk points, the review conclusion is directly set as rejection, and no further path checks are performed. If no risk point of the first preset type exists, the system continues to determine if there are risk points of the second preset type. These risk points correspond to anomalies requiring manual verification, such as inconsistencies in document elements or logical contradictions. If one or more risk points of the second preset type exist, the review conclusion is manual review. If there are neither first nor second preset type risk points, the review conclusion is approval.

[0044] Example 2

[0045] Based on the above embodiment 1, in this embodiment, when the matching degree between the feature set and any risk template exceeds a preset first threshold, it is determined that there is a corresponding potential risk point; When all matching scores are below the preset second threshold, the system is deemed risk-free. The first threshold is greater than the second threshold.

[0046] Specifically, the system calculates the matching degree between the comprehensive feature vector of the document to be reviewed and all risk templates in the risk feature library, and identifies risk points based on preset judgment thresholds (such as the first threshold and the second threshold). When the matching degree exceeds the first threshold, it is determined that there is a potential risk point corresponding to the risk template; if all matching degrees are lower than the second threshold, it is determined that there is no risk; if they are in between, a comprehensive judgment is made in combination with other rules.

[0047] Example 3

[0048] Based on the above embodiment 1, the order review decision rule base in this embodiment includes multiple decision paths, which are combined according to the type, number and severity level of potential risk points; When no potential risks are identified, the review conclusion is "approved". When one or more potential risk points of the first pre-defined type exist, the review conclusion is rejection; When there are one or more potential risk points of the second preset type and there are no potential risk points of the first preset type, the review conclusion is manual review.

[0049] Furthermore, it locates and outputs potential risk points, including outputting the starting character index of the risk text when the risk point originates from a text modality; When the risk point originates from the image modality, a heatmap or bounding box is overlaid on the image for annotation; when the risk point originates from the numerical modality, the cell or field name where the abnormal value is located is output.

[0050] Furthermore, the order review decision rule base in the above embodiments includes decision paths based on the type, number, and severity level of potential risk points, and the output of the decision paths is dynamically corrected by the second verification confidence level. The second verification confidence score is obtained by performing a reverse confidence assessment on the feature set generation process, and is used to characterize the reliability of the output results of the multimodal feature fusion model. When the confidence level of the second verification is lower than the decision threshold, the review conclusion will be forcibly switched to manual review and a low confidence level warning label will be attached.

[0051] Specifically, the system first categorizes and grades each potential risk point matched in the risk feature database: A first predefined type (such as forged seals, serious contradictions in monetary logic, and blacklisted trading entities—high-risk risks that cannot be tolerated) and a second predefined type (such as inconsistent document date formats, differences in product description terminology, and non-standard quantity unit writing—minor or verifiable risks). Based on the matching results between the comprehensive feature vector and the risk template, the system counts the presence and quantity of each type of risk point.

[0052] When there are no potential risk points, the review conclusion is directly determined as approved; when there are one or more risk points of the first preset type, the conclusion is rejection regardless of whether there are risk points of the second preset type; when there are one or more risk points of the second preset type and there are no risk points of the first preset type, the conclusion is manual review.

[0053] Furthermore, for text-based modal risks (such as discrepancies between payment terms in a contract and payment conditions on an invoice), the system not only outputs the risk type but also utilizes a converter-based bidirectional encoder representation model to extract the risk text location information, outputting the start and end character indices of the risk text segment, thus highlighting the specific risk statement in the display interface. For image-based modal risks (such as altered stamps on bills of lading or abnormal signatures on certificates of origin), the system uses a convolutional neural network to extract features, combines a gradient-weighted activation mapping algorithm to calculate the contribution of each region of the image to risk decision-making, and then overlays a semi-transparent heatmap on the original image (the darker the color, the greater the contribution of the region to risk assessment), while accurately outlining the abnormal stamp or signature area with a red bounding box to achieve visual risk labeling. For numerical modal risks (such as logical errors in the product of invoice amount, quantity, and unit price), the system directly outputs the specific cell name or field name where the abnormal value is located and indicates the correct value range.

[0054] Furthermore, this invention introduces a second verification confidence level as a reverse evaluation index for the reliability of the multimodal feature fusion model's output, and uses this to dynamically correct the output of the decision path. The second verification confidence level is not directly generated from the original data, but is obtained through a reverse confidence evaluation of the generation process of the comprehensive feature vector. After the last fully connected layer of the multimodal feature fusion model, an additional confidence evaluation branch is set. This branch receives the feature vectors of each modality before fusion and the attention weight distribution, and quantifies the model's confidence in the current output result through gradient information and feature reconstruction errors during backpropagation.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A cross-border trade finance intelligent document review method based on multimodal feature fusion, characterized in that, Includes the following steps: S01. Collect pending document data in cross-border trade finance business, including text modal document data, image modal document data and numerical modal document data; S02. Extract features from the data of the documents to be reviewed to obtain a dataset, and then perform cross-validation on the dataset based on the first-level validation rules of the confidence score to generate the first validation confidence. When the first verification confidence of any feature contained in the dataset is lower than the preset confidence threshold, a supplementary acquisition signal is triggered to obtain supplementary document data and re-execute feature extraction and verification. S03. Input the feature data contained in the dataset with the first verification confidence level higher than the confidence level threshold into the preset multimodal feature fusion model, and align and fuse the features of different modalities through the multimodal feature fusion model to generate a feature set; S04. Compare the feature set with the preset risk feature library, calculate the matching degree between the feature set and the risk template in the risk feature library that matches the corresponding feature, and identify one or more potential risk points in the data of the document to be reviewed based on the matching degree. S05. Based on the identified potential risk points and in conjunction with the preset document review decision rule base, generate the final document review conclusion, and simultaneously locate and output the specific location of the potential risk points in the document data to be reviewed and the corresponding risk reasons.

2. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 1, characterized in that, The text modal document data includes unstructured text from trade contracts, commercial invoices, and packing lists; The image modal document data includes scanned copies of the bill of lading and certificate of origin with a stamp. The numerical modal document data includes invoice amount, quantity, unit price, and tax rate.

3. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 1, characterized in that, In step S02, text features are extracted from the text modal document data, image features are extracted from the image modal document data, and numerical features are extracted from the numerical modal document data. The extracted text features, image features, and numerical features are then cross-validated using a first-level verification rule based on confidence scoring to generate a first verification confidence level. When the first verification confidence level of any feature is lower than the preset confidence threshold, a supplementary acquisition signal is triggered to obtain supplementary document data, and feature extraction and verification are re-executed, including the following steps: S21. A bidirectional encoder representation model based on a converter is used to extract semantic features from text modal document data to obtain text feature vectors; S22. Use a convolutional neural network to extract features of seals, signatures and key field areas from image modal document data to obtain image feature vectors; S23. Process the numerical modal data to obtain numerical feature vectors; S24. Compare the text feature vector, image feature vector, and numerical feature vector with the corresponding preset business rules, which include document consistency rules, logical self-consistency rules, and historical data verification rules, and generate preliminary verification results.

4. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 1, characterized in that, The multimodal feature fusion model in S03 dynamically assigns weights to features of different modalities through an attention mechanism, and concatenates the weighted multimodal features and maps them to a fully connected layer to generate the feature set. The aforementioned multimodal feature fusion model is constructed by collecting historical cross-border trade financial document data and corresponding historical risk labeling data as training samples. Through supervised learning, with the goal of minimizing risk identification error, the network parameters of the initial fusion model are iteratively optimized to complete the training.

5. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 1, characterized in that, The preset risk feature library in S04 includes multiple risk categories, each risk category corresponds to a risk template, and the risk template is composed of a set of typical features extracted from historical risk events; The matching degree is the cosine similarity or Euclidean distance between the feature set and the risk template.

6. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 5, characterized in that, In step S04, when the matching degree between the feature set and any risk template exceeds a preset first threshold, it is determined that there is a corresponding potential risk point. When all matching scores are below the preset second threshold, the system is deemed risk-free. The first threshold is greater than the second threshold.

7. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 1, characterized in that, The order review decision rule base mentioned in S05 includes multiple decision paths, which are combined according to the type, number, and severity level of potential risk points; When no potential risks are identified, the review conclusion is "approved". When one or more potential risk points of the first pre-defined type exist, the review conclusion is rejection; When there are one or more potential risk points of the second preset type and there are no potential risk points of the first preset type, the review conclusion is manual review.

8. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 7, characterized in that, The positioning and output of the potential risk point in S05 includes outputting the starting character index of the risk text when the risk point originates from a text modality. When the risk point originates from the image modality, a heatmap or bounding box is overlaid on the image for annotation; when the risk point originates from the numerical modality, the cell or field name where the abnormal value is located is output.

9. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 1, characterized in that, The order review decision rule base in step S05 includes decision paths based on the type, number, and severity level of potential risk points, and the output of the decision paths is dynamically corrected by the second verification confidence level. The second verification confidence score is obtained by performing a reverse confidence score evaluation on the generation process of the feature set, and is used to characterize the reliability of the output results of the multimodal feature fusion model. When the confidence level of the second verification is lower than the decision threshold, the review conclusion will be forcibly switched to manual review and a low confidence level warning label will be attached.

10. The intelligent document review method for cross-border trade finance based on multimodal feature fusion according to claim 1, characterized in that, The dataset includes text features extracted from the text modal document data, image features extracted from the image modal document data, and numerical features extracted from the numerical modal document data.

Images