Data information processing verification system and method

Abstract

The invention provides a data information processing verification system and method, and relates to the technical field of artificial intelligence. Preset rules are abandoned, and a first data set representing document physical and interaction attributes and a second data set representing business logic and format positions are obtained in parallel; and carrying out fusion reasoning on the multi-dimensional features by utilizing a probability graph model, so as to realize high adaptability to unknown formats and unstructured contents. According to the fusion mechanism, through cross validation of a microscopic inconsistency index from vision and a macroscopic deviation index from business, an information barrier is broken, and verification depth and accuracy are improved. Furthermore, the system also utilizes a final processing result to carry out feedback updating on model parameters through an incremental learning algorithm, so that self-adaptive evolution is realized; and meanwhile, the generated traceability information also ensures that the whole verification process is transparent and releasable.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a data information processing and verification system and method. Background Technology

[0002] In the application of industrial vision intelligence technology to the automation process in the financial field, the mainstream technical solution for data processing and verification of key financial documents such as invoices and vouchers usually adopts a two-level pipeline mode of "Optical Character Recognition (OCR) + rule engine verification". Specifically, the existing technology, with publication number CN109711145A, entitled "A Data Verification Method, Apparatus, Storage Medium, and Computer Equipment," effectively solves the problem of increased code size and bloat in controllers when implementing data verification. The method includes: creating a specific object for data verification; binding the specific object to an object to be verified; identifying the verification type of the data to be verified corresponding to the object to be verified; if it is determined that the data to be verified contains sub-data of multiple verification types, then using the specific object to perform data verification on the sub-data of each verification type together according to the preset verification rules corresponding to the verification types, wherein each verification type has its own corresponding preset verification rules.

[0003] However, the above technical solutions have the following technical drawbacks in real-world, complex financial scenarios: The limitations of the data model lead to fragile generalization capabilities: the core validation logic of the above solution heavily relies on "preset rules" and "predefined keywords." This strongly template-based approach makes the system extremely fragile, with very poor adaptability to changes in the format of financial documents, differences in language style, or unstructured layouts (such as payment instructions in contracts). Once an upstream company changes its invoice template, or processes a supplier receipt with a format never seen before, the system will fail to extract key information or make validation errors due to the inability to match the rules. This causes the system's accuracy to plummet when faced with diverse and dynamically changing real business documents, requiring a significant investment of manpower to continuously maintain and update the rule base, thus defeating the purpose of automation.

[0004] Information silos create bottlenecks in scenario adaptability: In terms of technical processes, the OCR module and the verification module are separate. OCR is only responsible for "understanding" the text and outputting it, while the verification engine makes logical judgments entirely based on these text results. This separation causes the verification process to lose valuable original visual context information. For example, when the OCR engine has low confidence in recognizing a blurry number "8", it may incorrectly output "3". The backend rule engine receives "3", and if this incorrect number happens to be within the logical rules such as the numerical range, it will be judged as valid data, causing serious financial risks. The verification module cannot use visual features such as the stroke shape of numbers in the image, the degree of damage, or their relative position in the table to assist in judgment, missing the key opportunity for cross-validation and error correction. Summary of the Invention

[0005] The purpose of this invention is to provide a data information processing and verification system and method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a data information processing and verification method, the specific steps of which include: S1: For the financial document to be verified, obtain the first dataset and the second dataset in parallel; wherein, the first dataset represents the inherent attributes of the financial document to be verified, which are independent of the business content it contains, and the inherent attributes are selected from physical attributes and interaction attributes. Furthermore, the second dataset represents the positional attribute of the business content carried by the financial document to be verified within a preset macro-logic or group statistical model; S2: Based on a preset probabilistic graphical model, perform fusion reasoning on the first dataset and the second dataset to generate a risk qualitative signal, wherein the risk qualitative signal not only includes the posterior probability of the financial document to be verified corresponding to at least two different risk types, but also includes traceability information for indicating the key features on which the posterior probability derivation is based. S3: Determine the target response strategy based on the posterior probability distribution of different risk types in the qualitative risk signal; S4: Execute the target response strategy to complete the verification process of the financial document to be verified.

[0007] Preferably, S1 specifically includes: preprocessing the financial document to be verified, including affine transformation correction of the image and optical character recognition, to extract standardized image data and structured text data; and simultaneously initiating microscopic trace analysis and macroscopic logical positioning. The microscopic trace analysis is used to generate the first dataset, which quantifies the risk of physical tampering by parsing the pixel-level statistical features of the financial documents to be verified, and analyzes the human-computer interaction time sequence patterns associated with the financial documents to be verified to assess the risk of abnormal operation.

[0008] Preferably, S1 further includes: the macro-logic positioning is used to generate a second dataset, by placing the transaction entities extracted from the financial documents to be verified into a preset business knowledge graph to calculate their semantic distance, so as to quantify their business logic rationality, and projecting their layout features into the historical layout feature space for cluster analysis to evaluate their layout compliance.

[0009] Preferably, S2 specifically includes: using a preset Bayesian network as the probabilistic graphical model, inputting the micro-inconsistency index in the first dataset and the macro-deviation index in the second dataset as observation evidence nodes into the Bayesian network; By executing the confidence propagation algorithm, the posterior probability distribution of hidden state nodes in the network that represent different risk types is calculated to generate the risk qualitative signal.

[0010] Preferably, S2 further includes: using the causal relationship structure built into the Bayesian network to trace back the evidence node path that contributes the most to each high posterior probability risk type, and using the information of the evidence node path as the source information; The structure and conditional probability table of the Bayesian network are pre-generated by performing structure learning and parameter learning on a sample set containing historical financial documents and their corresponding risk labels. The structure learning uses a constraint-based learning algorithm to determine the dependencies between nodes.

[0011] Preferably, S3 specifically includes: calculating a comprehensive risk score based on the posterior probability distribution of different risk types in the qualitative risk signal; projecting the comprehensive risk score onto a preset decision matrix that nonlinearly maps the risk score interval to the response strategy level to determine the target response strategy; The response strategy steps preset by the decision matrix include: a first automatic pass step, a second manual review step, and a third interception alarm step. The manual review level is further associated with the source information in the risk qualitative signal to provide review guidance to the reviewers. The risk score interval boundary in the decision matrix is ​​dynamically optimized based on receiver operation characteristic curve analysis of historical handling cases, and is used to balance the false alarm rate and false negative rate under different response strategy levels.

[0012] Preferably, S3 further includes: determining the target response strategy based on the posterior probability distribution of different risk types in the risk qualitative signal; The steps for determining the target response strategy specifically include: A comprehensive risk score is generated by multiplying the posterior probability of each risk type in the qualitative risk signal with a preset risk cost coefficient that characterizes the degree of impact of the current risk type on the business, and then summing all the product results. The comprehensive risk score is compared with a preset decision matrix; the decision matrix has a first threshold and a second threshold, the second threshold being greater than the first threshold; if the comprehensive risk score is not greater than the first threshold, the target response strategy is determined to be automatic pass; if the comprehensive risk score is greater than the first threshold but not greater than the second threshold, the target response strategy is determined to be manual review, and the source tracing information in the risk qualitative signal is invoked; if the comprehensive risk score is greater than the second threshold, the target response strategy is determined to be interception and alarm.

[0013] Preferably, S4 specifically includes: executing the target response strategy; wherein, if the target response strategy is automatic approval, a verification approval identifier is written to the financial document to be verified; if it is manual review, the traceability information is rendered as an interactive highlighted annotation layer and pushed to the manual review workflow along with the financial document to be verified; if it is interception and alarm, the subsequent processing flow of the financial document to be verified is stopped and a structured alarm event is generated.

[0014] Preferably, S4 further includes: after executing the manual review or interception strategy, receiving the final processing result, and updating the parameters of the probabilistic graphical model based on the final processing result and the first and second datasets corresponding to the financial documents to be verified; The feedback update specifically involves using the final processing result as the target label, the first dataset and the second dataset as training features, and using an incremental learning algorithm to adjust the conditional probability table in the probabilistic graphical model.

[0015] A data information processing and verification system, comprising: The multidimensional feature extraction module acquires a first dataset and a second dataset in parallel for the financial document to be verified; wherein, the first dataset represents the intrinsic attributes of the financial document to be verified, which are independent of the business content it contains, and the intrinsic attributes are selected from physical attributes and interaction attributes. Furthermore, the second dataset represents the positional attribute of the business content carried by the financial document to be verified within a preset macro-logic or group statistical model; The fusion inference engine module performs fusion inference on the first dataset and the second dataset based on a preset probabilistic graphical model to generate a risk qualitative signal. The risk qualitative signal includes not only the posterior probability of the financial document to be verified corresponding to at least two different risk types, but also traceability information indicating the key features on which the posterior probability derivation is based. The response strategy decision module determines the target response strategy based on the posterior probability distribution of different risk types in the qualitative risk signal. The closed-loop processing module executes the target response strategy to complete the verification process of the financial document to be verified.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the weakness of traditional solutions in generalization due to reliance on preset rules by constructing a multidimensional and adaptive data model. In S1, a first dataset representing physical and interactive attributes and a second dataset representing business logic and layout position are acquired in parallel. Specifically, this is achieved by analyzing pixel-level statistical features, human-computer interaction temporal patterns, semantic distances in business knowledge graphs, and layout features in historical layout feature spaces, fundamentally eliminating reliance on specific keywords or fixed layouts. Furthermore, in S2, a Bayesian network probabilistic graphical model is used to fuse and infer these abstract features, enabling the system to understand the inherent risks of documents rather than just surface text, thus improving its adaptability to unknown layouts and unstructured text. S4 introduces a feedback update mechanism based on the final processing results, continuously adjusting the conditional probability table of the probabilistic graphical model through incremental learning algorithms. This gives the system self-evolution capabilities, allowing it to dynamically adapt to business changes, thereby reducing manual maintenance costs while maintaining high accuracy.

[0017] This invention also breaks down the information silos between optical character recognition and business verification by designing an information fusion and traceability mechanism, solving the adaptability bottleneck caused by scenario fragmentation. Its core lies in the fact that the first and second datasets acquired in parallel by S1 are not processed independently, but are simultaneously input into a probabilistic graphical model for fusion inference in S2. The system can cross-validate the micro-inconsistency index derived from the physical properties of the image with the macro-deviation index derived from the business content. For example, a seemingly compliant number may still be identified as risky if its pixel features or layout are abnormal, effectively compensating for the blind spots of relying solely on text verification. Furthermore, the risk qualitative signal generated by S2 contains clear traceability information and can be rendered as an interactive highlighted annotation layer in S4, directly linking abstract risk judgments with specific visual evidence on the financial documents to be verified. This provides precise guidance for subsequent manual review, achieving a deep closed loop between visual information and logical verification throughout the entire processing flow. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the overall application of the method of the present invention; Figure 2 This is a schematic diagram of the overall method steps of the present invention; Figure 3 This is a schematic diagram of the S3 process framework of the present invention; Figure 4 This is a schematic diagram of the overall system framework structure of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Example 1: Please see Figure 1 and Figure 2 This invention provides a technical solution: a data information processing and verification method, the specific steps of which include: S1: For the financial document to be verified, obtain the first dataset and the second dataset in parallel; wherein, the first dataset represents the inherent attributes of the financial document to be verified, which are independent of the business content it contains, and the inherent attributes are selected from physical attributes and interaction attributes. Furthermore, the second dataset represents the positional attribute of the business content carried by the financial document to be verified within a preset macro-logic or group statistical model; S2: Based on a preset probabilistic graphical model, perform fusion reasoning on the first dataset and the second dataset to generate a risk qualitative signal, wherein the risk qualitative signal not only includes the posterior probability of the financial document to be verified corresponding to at least two different risk types, but also includes traceability information for indicating the key features on which the posterior probability derivation is based. S3: Determine the target response strategy based on the posterior probability distribution of different risk types in the qualitative risk signal; S4: Execute the target response strategy to complete the verification process of the financial document to be verified.

[0022] In this embodiment, Figure 1 The multidimensional feature extraction representation S1 in the text, Figure 1 Risk quantification and decision representation in S2 Figure 1The fusion reasoning and source tracing representation in S3, Figure 1 The strategy execution and closed-loop feedback in S4 are represented.

[0023] Example 2: Please see Figure 3 S1 specifically includes: preprocessing the financial document to be verified, including affine transformation correction of the image and optical character recognition, to extract standardized image data and structured text data; and simultaneously initiating microscopic trace analysis and macroscopic logical positioning. The microscopic trace analysis is used to generate the first dataset, which quantifies the risk of physical tampering by parsing the pixel-level statistical features of the financial documents to be verified, and analyzes the human-computer interaction time sequence patterns associated with the financial documents to be verified to assess the risk of abnormal operation.

[0024] S1 further includes: the macro-logic positioning is used to generate a second dataset, by placing the transaction entities extracted from the financial documents to be verified into a preset business knowledge graph to calculate their semantic distance, so as to quantify their business logic rationality, and projecting their layout features into the historical layout feature space for cluster analysis to evaluate their layout compliance.

[0025] Furthermore, a transaction entity refers to the key information that can be identified and extracted from financial documents and constitutes the business logic of a transaction. In S1, after the transaction entity is extracted, it is placed into the business knowledge graph as an independent node. The algorithm calculates the distance between these entity nodes in the graph, that is, the distance between the payee company, approver, payer and bank in the graph, to quantify the logical rationality of the transaction in the real business world. When the relationship between entities is distant or does not exist in the graph, it indicates that the transaction may have logical fallacies or fraud risks. Furthermore, the step of "quantifying the risk of physical tampering by parsing the pixel-level statistical features of the financial documents to be verified" is achieved through an error level analysis-based method; standardized image data is used as input and the core calculation process is executed. The first step in the core calculation process is to recompress the input image in memory with a preset, lossy JPEG compression quality factor and save it; in this embodiment, the compression quality factor is preferably 90; the purpose of the first step is to use the characteristics of the JPEG algorithm to amplify the differences between different compression history areas in the image. The second step in the core calculation process is to calculate the difference between the original input image and the aforementioned recompressed image at the pixel level to generate a difference image; The third step of the process is to traverse all pixels of the differential image and count the total number of pixels whose brightness values ​​are higher than a preset brightness threshold. The preset brightness threshold is set based on experience in the field. In this embodiment, the preferred value of the preset brightness threshold is 20 (based on an 8-bit grayscale image, the range is 0-255). Divide the total number of high-brightness pixels counted above by the total number of pixels in the image to obtain a ratio value between 0 and 1. This ratio value is defined as the final output physical tampering risk quantification value. The higher the value, the greater the possibility that the image has unnatural editing traces. Furthermore, the step of "analyzing the human-computer interaction time sequence pattern associated with the financial document to be verified to assess its operational anomaly risk" involves analyzing the user's keyboard input behavior when entering key fields of the document. During the user's data entry, the system captures a series of keyboard key events in the background, forming a timestamp sequence as input. Next, the core calculation process is executed. The first step of the process is to calculate the time interval between each two consecutive key presses based on the timestamp sequence, forming a time interval sequence. The second step of the process is to statistically analyze the numerical distribution of the time interval sequence and calculate the information entropy of the sequence based on the Shannon entropy calculation logic. This information entropy value quantifies the randomness or pattern of user input rhythm. The third step is to compare the currently calculated information entropy value with a "historical baseline information entropy" and calculate the absolute value of the difference between the two. The "historical baseline information entropy" is obtained by collecting a large amount of keyboard input data from different operators entering the same type of document under normal working conditions and then statistically averaging the information entropy calculated from this data. This baseline value is pre-stored in the system. Finally, the absolute value of the aforementioned calculated difference is defined as the final output quantification value of the operational anomaly risk. The larger this value, the more the current user's input pattern deviates from the group's normal behavior.

[0026] Furthermore, the aforementioned "pre-defined business knowledge graph" is a graph database stored in the Resource Description Framework (RDF) format. Its data sources primarily include publicly available corporate annual reports, supply chain directories, and standard commodity and service classification codes. The core entity types in the graph include "company," "goods," and "services," while the core relationship types include "provide," "purchase," and "belong to an industry." Before using this graph, a graph embedding model based on translation principles is used to train all entities and relationships in the graph, mapping them to a low-dimensional continuous vector space. The training process of the graph embedding model aims to minimize the energy of triples that satisfy this relation, i.e., head entity vector + relation vector ≈ tail entity vector; specifically, the steps for "calculating its semantic distance" are as follows: First, core transaction triples are extracted from the document to be verified using optical character recognition (OCR) and named entity recognition (NAME). Second, the vector representations corresponding to the three elements "Company A," "Purchase," and "Company B's Product C" are found in a pre-trained graph embedding model. Finally, the energy score of the triples is calculated based on the energy function of the TransE model. The calculation logic is as follows: the head entity vector plus the relation vector is used to calculate the square of the Euclidean distance with the tail entity vector. This energy score is used as the inverse "semantic distance," and the smaller the value, the more reasonable the transaction is within the business logic contained in the knowledge graph. This score is normalized and output as a quantitative value of the business logic's reasonableness.

[0027] Furthermore, specifically, the "layout features" are extracted using a pre-trained convolutional neural network. In this embodiment, a ResNet-50 model pre-trained on the ImageNet dataset and fine-tuned on a dataset containing multiple financial document images of various types is used. The standardized image data of the document to be verified is input into this model, and the output of its last global average pooling layer is extracted as a 2048-dimensional floating-point vector, which is the aforementioned layout feature vector. The "historical layout feature space" is constructed after model training through the following steps: all the aforementioned financial document images are processed through the fine-tuned ResNet-50 model to extract features, resulting in multiple 2048-dimensional feature vectors. Then, the K-means clustering algorithm is used to cluster these multiple vectors. The number of clusters, K, is set to 50, representing the 50 most common financial document layouts. After clustering, the vector coordinates of these 50 cluster centers collectively define the historical layout feature space. Finally, the steps for "assessing its layout compliance" are as follows: For a new document to be verified, extract its layout feature vector. Calculate the Euclidean distance between this vector and all 50 cluster centers in the feature space. Select the minimum value among these 50 distance values ​​as the document's layout anomaly score; the smaller the score, the closer its layout is to a mainstream layout, and the higher its compliance.

[0028] S2 specifically includes: using a preset Bayesian network as the probabilistic graphical model, and inputting the micro-inconsistency index in the first dataset and the macro-deviation index in the second dataset as observation evidence nodes into the Bayesian network; By executing the confidence propagation algorithm, the posterior probability distribution of hidden state nodes in the network that represent different risk types is calculated to generate the risk qualitative signal.

[0029] S2 further includes: using the causal relationship structure built into the Bayesian network to trace back the evidence node path that contributes the most to each high posterior probability risk type, and using the information of the evidence node path as the source information; The structure and conditional probability table of the Bayesian network are pre-generated by performing structure learning and parameter learning on a sample set containing historical financial documents and their corresponding risk labels. The structure learning uses a constraint-based learning algorithm to determine the dependencies between nodes.

[0030] Furthermore, before performing the fusion inference in S2, the four quantitative indicators generated in S1 need to be aggregated to generate the input node evidence required by the Bayesian network. Specifically, the physical tampering risk quantification value generated in S1 is multiplied by the first weight coefficient, and the operational anomaly risk quantification value generated in S1 is multiplied by the second weight coefficient. Finally, the two products are added together to obtain the micro-inconsistency index. The first and second weight coefficients are determined by performing principal component analysis on the historical sample dataset containing the above two risk indicators, and distributing them proportionally according to the loading of each original indicator in the first principal component. In this embodiment, the first weight coefficient is preferably 0.65, and the second weight coefficient is preferably 0.35. Secondly, the quantitative value of "business logic rationality" generated by S1 is multiplied by the third weight coefficient, and then the quantitative value of "format compliance" generated by S1 is multiplied by the fourth weight coefficient. Finally, the two products are added together to obtain the macro deviation index. Similarly, these two weight coefficients are also determined by principal component analysis. In this embodiment, the third weight coefficient is preferably set to 0.70 and the fourth weight coefficient is preferably set to 0.30. Finally, the two index values ​​obtained are normalized so that their range falls within the interval [0, 1], which serves as observation evidence for the subsequent Bayesian network. Furthermore, the aforementioned reverse tracing function is implemented through a sensitivity analysis method based on evidence removal, specifically as follows: A high posterior probability threshold is set; in this embodiment, the high posterior probability threshold is set to 0.8. After generating the risk qualitative signal in S2, all risk types are traversed, and risk types with posterior probabilities greater than the high posterior probability threshold are selected as targets for source tracing analysis. For each selected target risk type, contribution quantification calculation is performed. Taking "forgery risk" as an example, its original posterior probability is denoted as P-full. The first step of the calculation process is to temporarily remove the observation evidence of "micro-inconsistency index" from the Bayesian network, and under this condition, the confidence propagation algorithm is re-executed to obtain a new posterior probability P-no-micro. The contribution of the micro-index is calculated, and its calculation logic is the difference between the original posterior probability P-full and the new posterior probability P-no-micro. The second step of the calculation process is to restore the micro-index evidence and remove the "macro-deviation index" evidence instead, and similarly calculate the contribution of the macro-index. Finally, the contribution values ​​of the micro-index and the macro-index are compared. The index with the larger contribution value is determined as key evidence. The generated traceability information is a JSON-formatted text containing three fields: the "key-feature" field records the name of the key evidence (i.e., the "micro-inconsistency index"), the "value" field records the original value of the index, and the "source" field records the name of the original S1 indicators that constitute the index (i.e., "physical tampering risk" and "operational anomaly risk"). Furthermore, the Bayesian network model was generated through the following offline training process. First, a training dataset was prepared, containing multiple historical financial document samples, all of which had four quantitative indicators extracted using the S1 process, and were doubly labeled by a senior audit team as "forgery risk" (yes / no) and "compliance risk" (yes / no). Before training, the four continuous quantitative indicators were discretized using the equal-frequency binning method, and each indicator was divided into three levels: "low," "medium," and "high." The PC algorithm is used to perform structure learning. During the algorithm execution, the chi-square test is used to test conditional independence, and the significance level α of the chi-square test is set to 0.05. This step automatically discovers the inter-node dependencies implied in the data, thereby determining the network topology. Next, parameter learning is performed. After the network structure is fixed, the maximum likelihood estimation algorithm with Laplace smoothing is used to learn the conditional probability table of each node in the network. The pseudo-count of Laplace smoothing is set to 1 to prevent the zero probability problem caused by sample sparsity. Finally, the generated network structure and complete conditional probability table are fixed and saved as a preset model used for online inference. In this embodiment, the conditional probability table of the learned "forgery risk" node may contain the following entries: under the condition that the "micro-inconsistency index" is "high" and the "macro-deviation index" is "high", the probability that "forgery risk" is "yes" is 0.92.

[0031] S3 specifically includes: calculating a comprehensive risk score based on the posterior probability distribution of different risk types in the qualitative risk signal; projecting the comprehensive risk score onto a preset decision matrix that nonlinearly maps the risk score interval to the response strategy level to determine the target response strategy; The response strategy steps preset by the decision matrix include: a first automatic pass step, a second manual review step, and a third interception alarm step. The manual review level is further associated with the source information in the risk qualitative signal to provide review guidance to the reviewers. The risk score interval boundary in the decision matrix is ​​dynamically optimized based on receiver operation characteristic curve analysis of historical handling cases, and is used to balance the false alarm rate and false negative rate under different response strategy levels.

[0032] S3 further includes: determining the target response strategy based on the posterior probability distribution of different risk types in the qualitative risk signal; The steps for determining the target response strategy specifically include: A comprehensive risk score is generated by multiplying the posterior probability of each risk type in the qualitative risk signal with a preset risk cost coefficient that characterizes the degree of impact of the current risk type on the business, and then summing all the product results. The comprehensive risk score is compared with a preset decision matrix; the decision matrix has a first threshold and a second threshold, the second threshold being greater than the first threshold; if the comprehensive risk score is not greater than the first threshold, the target response strategy is determined to be automatic pass; if the comprehensive risk score is greater than the first threshold but not greater than the second threshold, the target response strategy is determined to be manual review, and the source tracing information in the risk qualitative signal is invoked; if the comprehensive risk score is greater than the second threshold, the target response strategy is determined to be interception and alarm.

[0033] The aforementioned "risk cost coefficient" is used to weight the posterior probabilities of different risk types, ensuring that their contribution to the overall risk score is proportional to their potential business losses. This coefficient is determined through quantitative analysis of historical business data. The specific determination process is as follows: From the historical case database, at least several confirmed "counterfeiting risk" cases and several confirmed "compliance risk" cases were randomly selected; For each type of case, the average direct economic loss caused is calculated. In this embodiment, the average loss caused by the "forgery risk" case (including forgery of invoices for reimbursement) is set at 5,000 yuan, while the average potential loss or correction cost caused by the "compliance risk" case (including tax penalty risk caused by non-standard invoice filling) is 1,000 yuan. The average loss of each type of risk is normalized to determine its relative risk cost coefficient. In this embodiment, the cost of "compliance risk" is taken as the baseline of 1, and the cost coefficient of "forgery risk" is 5000 divided by 1000, which equals 5. Therefore, when calculating the comprehensive risk score, the risk cost coefficient multiplied by the posterior probability of "forgery risk" is set to 5.0, and the risk cost coefficient multiplied by the posterior probability of "compliance risk" is set to 1.0. The aforementioned "first threshold" and "second threshold" were determined on a dataset containing at least 10,000 historical handling cases through a two-stage optimization process. Each case in this dataset contains a calculated comprehensive risk score and a true label indicating whether it was ultimately identified as "automatic pass," "manual review," or "interception and alert." First, a second threshold was determined to distinguish "interception and alert." The labels for "automatic pass" and "manual review" in the dataset were temporarily merged into the "non-interception" class, thus transforming the original three-class classification problem into a binary classification problem of "interception" and "non-interception." For this binary classification problem, by traversing all possible comprehensive risk scores as cutoff points, the true positive rate (representing the proportion of correctly identifying "interception") and false positive rate (representing the proportion of "non-interception" cases incorrectly identified as "interception") were calculated at each cutoff point, thereby plotting the Receiver Operational Characteristic (ROC) curve. Subsequently, the Youden index was calculated for each point on the curve, which is calculated by subtracting the false positive rate from the true positive rate at that point. The point that maximizes the Youden index is selected, and its corresponding comprehensive risk score is determined as the second threshold. Next, a first threshold is determined to distinguish between "automatic pass" and "manual review." From the original dataset, all cases with comprehensive risk scores lower than the second threshold determined in the previous step are selected. A second binary classification analysis is performed on this subset, categorizing them as "automatic pass" and "manual review." The same process as before is used—plotting the ROC curve and finding the point that maximizes the Youden index—to determine the first threshold. This two-stage process ensures that the determination of both thresholds achieves an optimal balance within their respective decision intervals. By executing the above two-stage ROC analysis optimization process, a set of preferred thresholds was obtained on a typical dataset containing real business data. In this embodiment, the value range of the comprehensive risk score was normalized to between 0 and 10. After optimization calculation, the first threshold was determined to be 2.5, and the second threshold was determined to be 7.0. Therefore, the specific rules of the decision matrix are as follows: if the comprehensive risk score of a document is not greater than 2.5, the "automatic pass" strategy is executed; if its score is greater than 2.5 but not greater than 7.0, the "manual review" strategy is executed; if its score is greater than 7.0, the "interception and alarm" strategy is executed. This setting aims to automate the processing of approximately 70% of low-risk documents, submit 25% of medium-risk documents to manual review, and immediately intercept the most suspicious 5% of documents, thereby achieving a balance between processing efficiency and risk control. S4 specifically includes: executing the target response strategy; wherein, if the target response strategy is automatic approval, a verification approval identifier is written to the financial document to be verified; if it is manual review, the traceability information is rendered as an interactive highlighted annotation layer and pushed to the manual review workflow along with the financial document to be verified; if it is interception and alarm, the subsequent processing flow of the financial document to be verified is stopped and a structured alarm event is generated.

[0034] S4 further includes: after executing the manual review or interception strategy, receiving the final processing result, and updating the parameters of the probabilistic graphical model based on the final processing result and the first and second datasets corresponding to the financial documents to be verified; The feedback update specifically involves using the final processing result as the target label, the first dataset and the second dataset as training features, and using an incremental learning algorithm to adjust the conditional probability table in the probabilistic graphical model.

[0035] Additional example for "Interactive highlight annotation layer rendering": Specifically, the source information generated by S2 is parsed to extract the coordinates of the risk area rectangle and the corresponding detailed text description contained therein; based on the coordinates of the risk area rectangle, a visually highlighted annotation layer is dynamically generated at the corresponding position of the document image to be verified; an interactive event is bound to the annotation layer, and when the reviewer triggers the interactive event, the detailed text description in the source information is automatically retrieved and displayed, thereby providing accurate and timely review guidance for manual review; Furthermore, generating structured alarm events in S4 involves creating a JSON object that follows a predefined data pattern and broadcasting it via the enterprise message bus; the structure of the JSON object specifically includes the following fields: An "event-id" field conforming to the UUIDv4 specification is used to uniquely identify the alarm; a "timestamp" field conforming to the ISO8601 format records the precise time the alarm was generated; a "document-id" field stores the internal unique number of the blocked document; a "risk-score" field records the comprehensive risk score value calculated by S3 that led to the blocking; an array field named "risk-types" contains the name string of the risk type with the highest posterior probability output by S2; and a nested JSON object named "trace-info" contains the complete trace information generated by S2. After generating the JSON object, the system will publish the JSON object as a message body to a message exchange called "financial-risk-alerts" through an AMQP (Advanced Message Queuing Protocol) client. Downstream monitoring systems and consumers of security information and event management platforms can subscribe to this exchange to receive and process these high-risk alert events in real time. Furthermore, the incremental learning algorithm used in S4 to adjust the conditional probability table is specifically implemented through an online Bayesian parameter update method with a forgetting factor. First, the system receives the final handling result provided by the human reviewer, including confirmation of forgery risk, and maps the final handling result to the real state of the corresponding risk node in the Bayesian network. At the same time, it obtains the discretized observation evidence used in step S2 of the document. Second, it executes the conditional probability table update process. The first step in the conditional probability table update process is to locate a unique row in the conditional probability table directly associated with the risk node based on the state combinations of the aforementioned observational evidence. Each cell in this row stores a pseudo-count, representing the number of times that situation occurred in the historical data; The second step in the conditional probability table update process is to multiply all pseudo-count values ​​in the row by a preset "forgetting factor". This forgetting factor gradually reduces the weight of historical data over time, and its value ranges from 0.9 to 1. In this embodiment, the preferred value for the forgetting factor is 0.995. The third step in the conditional probability table update process is to find the column in the row that corresponds to the actual status of this manual review, and add 1 to the estimated value after multiplying it by the forgetting factor. The fourth step in the conditional probability table update process is to sum all the updated pseudo-count values ​​for the row, then divide each pseudo-count value in the row by this sum, and re-normalize to obtain the updated conditional probability distribution. Through this method, the model can continuously and incrementally absorb new human review experience, dynamically adapting to potential changes in data distribution without requiring expensive batch retraining. When the final conditional probability Pup approaches 1, it means that the final pseudo-count value Nfinal dominates the total pseudo-count value Nrow. According to the calculation logic, to achieve this, there must have been a historically high current pseudo-count value Ncurrent, and there must be continuous feedback of new processing results with the same label. Each feedback will amplify the growth advantage of the corresponding final pseudo-count value Nfinal after "multiplying by the preset forgetting factor aforget and adding 1" compared to other pseudo-count values ​​that are only multiplied by the preset forgetting factor aforget. As the final conditional probability Pup approaches 0, it means that the final pseudo-count value Nfinal becomes negligible relative to the total pseudo-count values ​​Nrow. This occurs because, although the current pseudo-count value Ncurrent of the entry may be updating, the pseudo-count values ​​of other entries in the same row grow faster (i.e., under the same observation characteristics, other disposal results occur more frequently). At the same time, the forgetting factor presupposes that the continuous effect of the forgetting factor aforget will steadily decay all historical counts. If an event no longer occurs, its corresponding current pseudo-count value Ncurrent will decay exponentially over time. In a specific embodiment of the present invention, in order to adapt to the limited precision of the digital processing system and improve the computational efficiency, the system further includes a processing step: comparing the calculated updated final conditional probability Pup with a preset probability lower limit threshold emin; if the final conditional probability Pup satisfies a specific logical condition, that is, it is less than the probability lower limit threshold emin, then in subsequent storage and inference calculations, the value of the final conditional probability Pup is forcibly set to 0. The effectiveness of the above processing steps requires that the following conditions be met: 1) The features of the first and second datasets have been correctly discretized into states consistent with the conditional probability table index; 2) The final processing result labels provided manually are considered true values, and their accuracy directly determines the direction of Bayesian network model optimization; online Bayesian parameter learning algorithms are suitable for business scenarios that require Bayesian network models to run online and be continuously optimized, and where computing resources are limited. In the initial settings, the pseudo-count values ​​are: high risk (current pseudo-count value Ncurrent = 5.0), medium risk (current pseudo-count value Ncurrent = 10.0), and low risk (current pseudo-count value Ncurrent = 10.0); see Table 1 below for details: Table 1: Evolution of the posterior probability Pupdated for the target state "High Risk" Test subject name Financial Document A Financial Document B Financial Document C Financial Document D Final handling results High risk High risk Medium risk High risk aforget 0.995 0.995 0.995 0.995 Ncurrent 5.000 5.975 6.945 6.910 Nupdated′ 4.975 5.945 6.910 6.875 Nfinal 5.975 6.945 6.910 7.875 Nrow 25.875 26.745 27.612 28.474 Pupdated 0.231 0.260 0.250 0.277 Table 1 quantifies and illustrates the core content of the feedback update mechanism in S4 through four consecutive cases: From financial document A to financial document B, after receiving the final handling result of "high risk" twice in a row, the final conditional probability of "high risk" in the model increased from 0.231 to 0.260; this shows that the model effectively learns from consistent human confirmation feedback and gradually enhances its confidence in judging specific risk patterns. In financial document C, when a different “medium risk” outcome is received, the model immediately responds to the final conditional probability Pupdated for the “high risk” outcome, decreasing from 0.260 to 0.250. This demonstrates that the feedback mechanism is not rigidly cumulative, but rather possesses a high degree of dynamic adjustment capability, revising existing knowledge based on new, and conflicting, evidence to prevent the model from becoming rigid. After adjustments to financial document C, when financial document D again returned to "high risk," the model's final conditional probability Pupdated for "high risk" resumed its upward trend, rising from 0.250 to 0.277. This demonstrates the resilience of the model's learning process, showing that it can quickly recover and continue to strengthen its learning of mainstream data patterns after being disturbed by a single negative example. Furthermore, the comprehensive risk score in S3 uses the following steps to divide the risk into intervals, including low risk, medium risk, and high risk: 1) Based on the experience of experts in the field of financial auditing, a preliminary qualitative description of the risk level in different numerical ranges is made, including when experts believe that a score below 2.0 corresponds to routine risk-free business; through statistical analysis of the comprehensive risk scores of historical cases, including drawing histograms and calculating quantiles, the data distribution is understood, and combined with the performance of the Bayesian network model in S2 in different score ranges, quantitative information on the effectiveness of the scoring is obtained.

[0036] 2) Based on the expert experience and data analysis results mentioned above, preliminary interval division criteria are formulated, including using the principle of maximizing the Youden index, the optimal split point found on the ROC curve can be used as the preliminary boundary between high risk and non-high risk; in the non-high risk data subset, a similar method is applied again or the boundary between low risk and medium risk is set according to the business-acceptable false positive rate; the preliminary criteria are tested using a set of independent validation datasets to evaluate the accuracy, coverage and business logic rationality in the classification task.

[0037] 3) Based on the verification results and feedback from actual applications collected during the trial operation phase, the initial standards were fine-tuned. For example, if it was found that almost all documents in the 2.5-3.0 range were ultimately confirmed as risk-free, the first threshold could be raised to 3.0 to improve the automation pass rate. Taking into account expert opinions, data verification performance, and business efficiency goals, the final range division standards were determined and solidified into the S3 decision matrix to ensure their practicality and effectiveness in actual applications.

[0038] Example 3: Please see Figure 4 A data information processing and verification system, comprising: The multidimensional feature extraction module acquires a first dataset and a second dataset in parallel for the financial document to be verified; wherein, the first dataset represents the intrinsic attributes of the financial document to be verified, which are independent of the business content it contains, and the intrinsic attributes are selected from physical attributes and interaction attributes. Furthermore, the second dataset represents the positional attribute of the business content carried by the financial document to be verified within a preset macro-logic or group statistical model; The fusion inference engine module performs fusion inference on the first dataset and the second dataset based on a preset probabilistic graphical model to generate a risk qualitative signal. The risk qualitative signal includes not only the posterior probability of the financial document to be verified corresponding to at least two different risk types, but also traceability information indicating the key features on which the posterior probability derivation is based. The response strategy decision module determines the target response strategy based on the posterior probability distribution of different risk types in the qualitative risk signal. The closed-loop processing module executes the target response strategy to complete the verification process of the financial document to be verified.

[0039] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization. The technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random-access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of this invention.

[0040] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0041] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A data information processing and verification method, characterized in that, The specific steps include: S1: For the financial document to be verified, obtain the first dataset and the second dataset in parallel; wherein, the first dataset represents the inherent attributes of the financial document to be verified, which are independent of the business content it contains, and the inherent attributes are selected from physical attributes and interaction attributes. Furthermore, the second dataset represents the positional attribute of the business content carried by the financial document to be verified within a preset macro-logic or group statistical model; S2: Based on a preset probabilistic graphical model, perform fusion reasoning on the first dataset and the second dataset to generate a risk qualitative signal, wherein the risk qualitative signal not only includes the posterior probability of the financial document to be verified corresponding to at least two different risk types, but also includes traceability information for indicating the key features on which the posterior probability derivation is based; S3: Determine the target response strategy based on the posterior probability distribution of different risk types in the qualitative risk signal; S4: Execute the target response strategy to complete the verification process of the financial document to be verified.

2. The data information processing and verification system and method according to claim 1, characterized in that: S1 specifically includes: preprocessing the financial document to be verified, including affine transformation correction of the image and optical character recognition, to extract standardized image data and structured text data; and simultaneously initiating microscopic trace analysis and macroscopic logical positioning. The microscopic trace analysis is used to generate the first dataset, which quantifies the risk of physical tampering by parsing the pixel-level statistical features of the financial documents to be verified, and analyzes the human-computer interaction time sequence patterns associated with the financial documents to be verified to assess the risk of abnormal operation.

3. The data information processing and verification system and method according to claim 2, characterized in that: S1 further includes: the macro-logic positioning is used to generate a second dataset, by placing the transaction entities extracted from the financial documents to be verified into a preset business knowledge graph to calculate their semantic distance, so as to quantify the rationality of the business logic of the financial documents to be verified, and projecting the layout features of the financial documents to be verified into the historical layout feature space for cluster analysis to evaluate their layout compliance.

4. The data information processing and verification system and method according to claim 3, characterized in that: S2 specifically includes: using a preset Bayesian network as the probabilistic graphical model, and inputting the micro-inconsistency index in the first dataset and the macro-deviation index in the second dataset as observation evidence nodes into the Bayesian network; By executing the confidence propagation algorithm, the posterior probability distribution of hidden state nodes in the network that represent different risk types is calculated to generate the risk qualitative signal.

5. The data information processing and verification system and method according to claim 4, characterized in that: S2 further includes: using the causal relationship structure built into the Bayesian network to trace back the evidence node path that contributes the most to each high posterior probability risk type, and using the information of the evidence node path as the source information; The structure and conditional probability table of the Bayesian network are pre-generated by performing structure learning and parameter learning on a sample set containing historical financial documents and their corresponding risk labels. The structure learning uses a constraint-based learning algorithm to determine the dependencies between nodes.

6. The data information processing and verification system and method according to claim 5, characterized in that: S3 specifically includes: calculating a comprehensive risk score based on the posterior probability distribution of different risk types in the qualitative risk signal; projecting the comprehensive risk score onto a preset decision matrix that nonlinearly maps the risk score interval to the response strategy level to determine the target response strategy; The response strategy steps preset by the decision matrix include: a first automatic pass step, a second manual review step, and a third interception alarm step. The manual review level is further associated with the source information in the risk qualitative signal to provide review guidance to the reviewers. The risk score interval boundary in the decision matrix is ​​dynamically optimized based on receiver operation characteristic curve analysis of historical handling cases, and is used to balance the false alarm rate and false negative rate under different response strategy levels.

7. The data information processing and verification system and method according to claim 6, characterized in that: S3 further includes: determining the target response strategy based on the posterior probability distribution of different risk types in the qualitative risk signal; The steps to determine the target response strategy specifically include: A comprehensive risk score is generated by multiplying the posterior probability of each risk type in the qualitative risk signal with a preset risk cost coefficient that characterizes the degree of impact of the current risk type on the business, and then summing all the product results. The comprehensive risk score is compared with a preset decision matrix; the decision matrix has a first threshold and a second threshold, the second threshold being greater than the first threshold; if the comprehensive risk score is not greater than the first threshold, the target response strategy is determined to be automatic pass; if the comprehensive risk score is greater than the first threshold but not greater than the second threshold, the target response strategy is determined to be manual review, and the source tracing information in the risk qualitative signal is invoked; if the comprehensive risk score is greater than the second threshold, the target response strategy is determined to be interception and alarm.

8. The data information processing and verification system and method according to claim 7, characterized in that: S4 specifically includes: executing the target response strategy; wherein, if the target response strategy is automatic approval, a verification approval identifier is written to the financial document to be verified; if it is manual review, the traceability information is rendered as an interactive highlighted annotation layer and pushed to the manual review workflow along with the financial document to be verified; if it is interception and alarm, the subsequent processing of the financial document to be verified is stopped and a structured alarm event is generated.

9. The data information processing and verification system and method according to claim 8, characterized in that: S4 further includes: after executing the manual review or interception strategy, receiving the final processing result, and updating the parameters of the probabilistic graphical model based on the final processing result and the first and second datasets corresponding to the financial documents to be verified; The feedback update specifically involves using the final processing result as the target label, the first dataset and the second dataset as training features, and using an incremental learning algorithm to adjust the conditional probability table in the probabilistic graphical model.

10. A data information processing and verification system, characterized in that: The system is used to execute a data information processing and verification method according to any one of claims 1-9, including: The multidimensional feature extraction module acquires a first dataset and a second dataset in parallel for the financial document to be verified; wherein, the first dataset represents the intrinsic attributes of the financial document to be verified, which are independent of the business content it contains, and the intrinsic attributes are selected from physical attributes and interaction attributes. Furthermore, the second dataset represents the positional attribute of the business content carried by the financial document to be verified within a preset macro-logic or group statistical model; The fusion inference engine module performs fusion inference on the first dataset and the second dataset based on a preset probabilistic graphical model to generate a risk qualitative signal. The risk qualitative signal includes not only the posterior probability of the financial document to be verified corresponding to at least two different risk types, but also traceability information indicating the key features on which the posterior probability derivation is based. The response strategy decision module determines the target response strategy based on the posterior probability distribution of different risk types in the qualitative risk signal. The closed-loop processing module executes the target response strategy to complete the verification process of the financial document to be verified.

Images