Automatic verification and intelligent error correction method and system for enterprise tax declaration data

Abstract

The invention provides an enterprise tax declaration data automatic verification and intelligent error correction method and system, and relates to the technical field of tax information processing, and the method comprises the steps: receiving original voucher data, constructing a voucher association diagram, recognizing an abnormal transaction chain, and generating financial accounting data; comparing the financial accounting data with the tax payment rules, identifying violation data items and distributing risk weights; and constructing a data balance constraint matrix based on the risk weight, forming a data correction cluster, and executing grouping iteration correction. According to the invention, declaration data errors can be automatically identified, intelligent correction is provided, declaration accuracy is improved, and manual intervention is reduced.

Description

Technical Field

[0001] The present invention relates to tax information processing technology, and in particular to a method and system for automatic verification and intelligent error correction of enterprise tax declaration data. Background Art

[0002] With the rapid growth of corporate tax declaration data, traditional manual verification methods are no longer able to meet the tax authorities' needs for efficient audits. Existing automated verification systems, which primarily rely on rule-based matching, are unable to effectively identify complex abnormal transaction chains and hidden violations, resulting in a large amount of abnormal data being unable to be discovered and processed in a timely manner.

[0003] Current intelligent error correction methods mostly rely on single-dimensional data verification, lacking in-depth analysis of the business relationships between documents. Furthermore, the correction process fails to consider the logical connections between data items, which can easily lead to new data inconsistencies caused by local corrections. Furthermore, when processing multi-source, heterogeneous data, existing methods struggle to ensure that correction results are consistent with real-time tax rules.

[0004] Existing data correction methods often use fixed correction strategies, unable to dynamically adjust correction schemes based on data characteristics. They also lack effective stability assessment mechanisms, making correction results prone to oscillation. Therefore, a tax declaration data verification method is urgently needed that can automatically identify abnormal transaction chains, fully consider the relevance of data items, and implement intelligent dynamic correction. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for automatic verification and intelligent error correction of enterprise tax declaration data, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for automatic verification and intelligent error correction of corporate tax declaration data, comprising: Receive original voucher data input by the enterprise and perform feature extraction, construct a feature vector, generate a voucher digital fingerprint based on the feature vector, establish a voucher association graph based on the voucher digital fingerprint, analyze the business association relationship between the vouchers in the voucher association graph, generate an abnormal transaction chain identification result, and convert the original voucher data into financial accounting data based on the abnormal transaction chain identification result and the voucher association graph; Acquire real-time tax rule data, compare financial accounting data with tax rule data, calculate the logical correlation between each data item in the financial accounting data, perform forward traversal and reverse verification, identify data items that violate tax rules, and assign risk weight values ​​to these data items; A data balance constraint matrix is ​​constructed based on the risk weight value. The degree of deviation of data items that violate tax rules in the data balance constraint matrix is ​​analyzed. Data items with the same degree of deviation are grouped into data correction clusters. Group iterative correction is performed on the data correction clusters. When the correction results meet the balance constraints, tax declaration data is generated.

[0007] In an optional embodiment, Receive the original credential data input by the enterprise and perform feature extraction to construct a feature vector. Generate the credential digital fingerprint based on the feature vector, including: Receiving original voucher data input by the enterprise side, and extracting transaction information and document information from the original voucher data; The transaction amount in the transaction information is segmented according to preset amount intervals, a normalized ratio of each interval is calculated to obtain an amount characteristic value, the transaction time is discretized according to preset time units to obtain a time characteristic value, and the amount characteristic value and the time characteristic value are combined to obtain a first characteristic value; Constructing a co-occurrence matrix for the transaction type and account attributes in the document information, calculating the correlation between the attributes based on the co-occurrence matrix, performing a dimensionality reduction transformation on the document information to obtain a second eigenvalue; normalizing the first and second eigenvalues ​​and concatenating them to obtain an eigenvector; Calculating the mutual information value between each dimension feature in the feature vector and the preset abnormal transaction identifier, and normalizing it to obtain the importance weight value of each dimension feature, and performing weighted summation on the importance weight value and the feature value of the corresponding dimension in the feature vector to obtain a weighted feature vector; The weighted feature vector is divided into multiple sub-vectors, the binary hash value of each sub-vector is calculated and combined to obtain a hash signature, the semantic features representing the transaction type and account attributes in the feature vector are extracted, and the semantic features are combined with the hash signature to generate a digital fingerprint of the credential.

[0008] In an optional embodiment, Based on the digital fingerprint of the voucher, a voucher association graph is established, and the business association relationship between the vouchers in the voucher association graph is analyzed to generate the abnormal transaction chain identification results, including: Construct association nodes based on the digital fingerprint of the voucher, calculate the cosine similarity between adjacent nodes as the association strength, set edge weights based on the association strength, extract transaction time information from the original voucher data to determine the direction of the edge, calculate a dynamic filtering threshold based on the number of connections and distribution density of the nodes, and use the dynamic filtering threshold to filter the edge weights to obtain a voucher association graph; Calculating the degree of association between adjacent nodes in the credential association graph to obtain a node weight coefficient, performing weighted fusion on the business features of the nodes according to the node weight coefficient to obtain a business association feature, combining the business association feature with the credential digital fingerprint to generate node representation information; Arranging the node representation information in chronological order to form a feature sequence, extracting the dependency relationship between the nodes to obtain a transaction chain feature, calculating the similarity between the transaction chain feature and the preset abnormal feature, and marking the continuous transaction sequence with a similarity higher than a first preset threshold as a candidate abnormal transaction chain; The time interval and amount change rate of adjacent transactions in the candidate abnormal transaction chain are calculated, and the correlation between the time interval and amount change rate and the transaction type is used as the transaction correlation degree. The candidate abnormal transaction chain is scored according to the transaction correlation degree, and the candidate abnormal transaction chain with a score exceeding a second preset threshold is determined as an abnormal transaction chain, and an abnormal transaction chain identification result is generated.

[0009] In an optional embodiment, Based on the abnormal transaction chain identification result and the voucher association graph, converting the original voucher data into financial accounting data includes: Calculate the amount change rate of adjacent nodes in the abnormal transaction chain to obtain an amount change feature, calculate the transaction time difference of adjacent nodes to obtain a time interval feature, and extract the debit and credit account combinations of adjacent nodes to obtain an account combination feature; determine the abnormal propagation direction based on the amount change feature, time interval feature, and account combination feature, and generate an abnormal propagation path; mark the original vouchers corresponding to the associated nodes located on the abnormal propagation path in the voucher association graph as affected vouchers; In the voucher association diagram, select the node sequence without abnormality as the normal transaction chain, extract the matching relationship and corresponding relationship of the debit and credit accounts in the normal transaction chain, and generate the accounting rule template; Based on the connections of the abnormal propagation paths, adjacent affected vouchers are grouped into the same accounting unit. Within each accounting unit, the debit and credit accounts of the affected vouchers are paired and verified according to the accounting rule template. The accounting results are compared with historical normal data to identify accounting deviations. Based on the business association strength between nodes in the voucher association graph, subgraphs are divided, and parallel accounting processing is performed on the original vouchers in each subgraph to generate subgraph accounting results; the subgraph accounting results are merged according to the association relationship between subgraphs to generate financial accounting data.

[0010] In an optional embodiment, Acquire real-time tax rule data, compare financial accounting data with tax rule data, calculate the logical correlation between each data item in the financial accounting data, perform forward traversal and reverse verification, identify data items that violate tax rules, and assign risk weight values ​​to data items that violate tax rules, including: Obtain tax rule data through the rule data interface, extract the judgment conditions and constraints in the tax rule data, construct a rule description vector, calculate the similarity between the rule description vector and the rule vector in the local rule library, identify new and changed rule content, perform structural processing, and convert it into a rule expression containing trigger conditions, judgment thresholds, and constraint parameters; The data items in the financial accounting data are constructed into an association graph, with the data items as nodes and the logical dependencies between the data items as edges. The numerical association degree and the subject association degree between adjacent nodes are calculated, and the weighted values ​​of the two association degrees are used as edge weights. Based on the edge weights, the shortest path between the nodes is calculated to obtain the association strength of the data items. Traverse the nodes in the association graph in the order of business processes, compare the attribute values ​​of data items with the rule expressions, and mark data items as violation nodes when they do not meet the rule expressions; Starting from each violation node, trace back along the association graph to identify the associated nodes that affect the violation node; mark the path from the associated node to the violation node as the violation propagation path; calculate the impact range of the node based on the violation propagation path; A basic score is determined based on the number of rule violations, a propagation coefficient is calculated based on the scope of influence, and an aging coefficient is calculated based on the update time of the rule expression; the product of the basic score, the propagation coefficient and the aging coefficient is used as the risk weight value.

[0011] In an optional embodiment, Based on the risk weight values, a data balance constraint matrix is ​​constructed. The degree of deviation of data items that violate tax rules in the data balance constraint matrix is ​​analyzed. Data items with the same degree of deviation are grouped into data correction clusters. Grouped iterative corrections are performed on the data correction clusters. When the correction results meet the balance constraints, the tax declaration data generated includes: A self-organizing multi-state equilibrium network is constructed based on risk weight values. The initial state of each data item is determined by the risk weight value. Data items influence each other through state transition rules. The state transition rules include two modes: weight decay and weight enhancement, generating a state evolution sequence of data items. Identifying key state transition points based on the state evolution sequence, wherein the key state transition points reflect the moment when the state of the data item suddenly changes, performing comparative analysis on the data items before and after the key state transition points, and extracting change characteristics of the data items; Calculating a state change value of the data item according to the change characteristics, dividing the data items whose state change values ​​are in the same preset interval into the same data correction cluster, and calculating an ideal state value of the data item in each data correction cluster based on the state transition rule; Data correction is performed alternately based on the weight decay and weight enhancement modes in the state transition rules, and the degree of difference between the data items after each round of correction and the ideal state value is calculated. When the degree of difference is less than the preset difference threshold, the data correction cluster is determined to have reached a stable state. When all data correction clusters have reached a stable state, tax return data is generated.

[0012] In an optional embodiment, Data correction is performed alternately based on the weight decay and weight enhancement modes in the state transition rule. The difference between the data item after each round of correction and the ideal state value is calculated. When the difference is less than the preset difference threshold, it is determined that the data correction cluster has reached a stable state. The following includes: Obtaining a difference sequence between a current state value and an ideal state value of a data item, establishing a state offset curve based on the difference sequence, and calculating a fluctuation trend and a fluctuation amplitude of the state offset curve; Alternately switching between a weight decay mode and a weight enhancement mode based on the positive and negative changes in the difference sequence, generating an adaptive adjustment coefficient based on the fluctuation trend and the current correction round, executing the weight decay mode when the difference sequence is positive, and executing the weight enhancement mode when the difference sequence is negative, dynamically updating the current state value of the data item based on the adaptive adjustment coefficient, updating the current state value and recalculating the difference sequence after each round of execution; Among them, the weight decay mode realizes downward adjustment of data items based on a multi-level decreasing function, and the weight enhancement mode realizes upward adjustment of data items based on a piecewise enhancement function; The degree of difference of all data items in the data correction cluster is calculated based on the current state value and the ideal state value, a stability evaluation index is generated according to the degree of difference, and the trend of change of the stability evaluation index over multiple consecutive rounds is analyzed. When the degree of difference is less than a preset difference threshold and is maintained for a specified number of rounds, it is determined that the data correction cluster has reached a stable state.

[0013] A second aspect of an embodiment of the present invention provides an automatic verification and intelligent error correction system for corporate tax declaration data, comprising: The first unit is configured to receive original voucher data input by the enterprise end and perform feature extraction, construct a feature vector, generate a voucher digital fingerprint based on the feature vector, establish a voucher association graph based on the voucher digital fingerprint, analyze the business association relationships between the vouchers in the voucher association graph, generate an abnormal transaction chain identification result, and convert the original voucher data into financial accounting data based on the abnormal transaction chain identification result and the voucher association graph; The second unit is used to obtain real-time tax rule data, compare financial accounting data with the tax rule data, calculate the logical correlation between each data item in the financial accounting data, perform forward traversal and reverse verification, identify data items that violate tax rules, and assign risk weight values ​​to the data items that violate tax rules; The third unit is used to construct a data balance constraint matrix based on the risk weight value, analyze the degree of deviation of data items that violate tax rules in the data balance constraint matrix, group data items with the same degree of deviation into data correction clusters, perform group iterative correction on the data correction clusters, and generate tax declaration data when the correction results meet the balance constraints.

[0014] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In this embodiment, by receiving original voucher data and constructing a voucher association graph, intelligent analysis of corporate transaction data is achieved, which can accurately identify abnormal transaction chains, ensure the accuracy and compliance of financial accounting data, and improve the quality of tax declaration data from the source. Combined with real-time tax rule data for comparison and verification, through forward traversal and reverse verification, data items that violate tax rules are effectively identified, and quantitative assessment is made through risk weight values, providing a more accurate risk management mechanism and reducing human errors in the tax declaration process. The data balance constraint matrix and group iterative correction method are used to intelligently correct illegal data to ensure that the corrected data meets the balance constraint relationship. While ensuring data consistency, it improves tax declaration efficiency, reduces corporate tax risks, and provides companies with a more convenient and accurate tax declaration solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a method for automatic verification and intelligent error correction of corporate tax declaration data according to an embodiment of the present invention; Figure 2 The flowchart of abnormal transaction chain identification based on credential digital fingerprint. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0019] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0020] Figure 1 Schematic diagram of the process of automatic verification and intelligent error correction method of enterprise tax declaration data according to an embodiment of the present invention. Figure 1 As shown, the method includes: Receive original voucher data input by the enterprise and perform feature extraction, construct a feature vector, generate a voucher digital fingerprint based on the feature vector, establish a voucher association graph based on the voucher digital fingerprint, analyze the business association relationship between the vouchers in the voucher association graph, generate an abnormal transaction chain identification result, and convert the original voucher data into financial accounting data based on the abnormal transaction chain identification result and the voucher association graph; Acquire real-time tax rule data, compare financial accounting data with tax rule data, calculate the logical correlation between each data item in the financial accounting data, perform forward traversal and reverse verification, identify data items that violate tax rules, and assign risk weight values ​​to these data items; A data balance constraint matrix is ​​constructed based on the risk weight value. The degree of deviation of data items that violate tax rules in the data balance constraint matrix is ​​analyzed. Data items with the same degree of deviation are grouped into data correction clusters. Group iterative correction is performed on the data correction clusters. When the correction results meet the balance constraints, tax declaration data is generated.

[0021] In an optional embodiment, receiving original credential data inputted by the enterprise and performing feature extraction to construct a feature vector, and generating a credential digital fingerprint based on the feature vector includes: Receiving original voucher data input by the enterprise side, and extracting transaction information and document information from the original voucher data; The transaction amount in the transaction information is segmented according to preset amount intervals, a normalized ratio of each interval is calculated to obtain an amount characteristic value, the transaction time is discretized according to preset time units to obtain a time characteristic value, and the amount characteristic value and the time characteristic value are combined to obtain a first characteristic value; Constructing a co-occurrence matrix for the transaction type and account attributes in the document information, calculating the correlation between the attributes based on the co-occurrence matrix, performing a dimensionality reduction transformation on the document information to obtain a second eigenvalue; normalizing the first and second eigenvalues ​​and concatenating them to obtain an eigenvector; Calculating the mutual information value between each dimension feature in the feature vector and the preset abnormal transaction identifier, and normalizing it to obtain the importance weight value of each dimension feature, and performing weighted summation on the importance weight value and the feature value of the corresponding dimension in the feature vector to obtain a weighted feature vector; The weighted feature vector is divided into multiple sub-vectors, the binary hash value of each sub-vector is calculated and combined to obtain a hash signature, the semantic features representing the transaction type and account attributes in the feature vector are extracted, and the semantic features are combined with the hash signature to generate a digital fingerprint of the credential.

[0022] In this implementation, the original voucher data input by the enterprise is first received, and transaction information and document information are extracted from it. Transaction information primarily includes data such as the transaction amount and transaction time, while document information primarily includes data such as the transaction type and account attributes. The original voucher data can be a variety of financial documents, such as value-added tax invoices, ordinary invoices, and bank statements.

[0023] For the extracted transaction information, the transaction amounts are segmented according to preset amount ranges. For example, the amount ranges can be set as 0-1,000 yuan, 1,000-5,000 yuan, 5,000-10,000 yuan, 10,000-50,000 yuan, and above 50,000 yuan. The number of transactions within each range is counted, and the proportion of transactions in each range to the total number of transactions is calculated to obtain a normalized amount characteristic value. For example, if a company has 100 transactions in a month, of which 30 are in the 0-1,000 yuan range, the normalized proportion for this range is 0.3. Furthermore, the transaction time is discretized according to preset time units to obtain a time characteristic value. The time unit can be set to day, week, or month, and the transaction frequency distribution within each time unit is calculated. For example, if a month is divided into four weeks, the proportion of transactions in each week is calculated, such as 0.2 for the first week, 0.3 for the second week, 0.4 for the third week, and 0.1 for the fourth week. The above-mentioned amount characteristic value and time characteristic value are combined to form a first characteristic value.

[0024] For document information, a co-occurrence matrix is ​​constructed for transaction types and account attributes. This matrix reflects the degree of correlation between different transaction types and account attributes. For example, the transaction type "Sales of Goods" may co-occur more frequently with the account "Accounts Receivable" but less frequently with the account "Fixed Assets." By counting the number of co-occurrences of each transaction type and each account attribute, a co-occurrence matrix is ​​generated. Based on this co-occurrence matrix, the correlation between attributes is calculated. For example, the pointwise mutual information method can be used to calculate the strength of the correlation between transaction types and account attributes. Next, a dimensionality reduction transformation is performed on the document information, mapping the high-dimensional transaction type and account attribute information into a low-dimensional space to obtain the second eigenvalue. This dimensionality reduction process preserves the key information of the original data while reducing computational complexity.

[0025] The first and second eigenvalues ​​are normalized and then concatenated to form an eigenvector. Normalization ensures that features of different dimensions fall within the same numerical range, preventing certain features from dominating the model due to their large values. Normalized eigenvalues ​​are typically distributed between 0 and 1, or within a range with a mean of 0 and a standard deviation of 1.

[0026] To enhance the ability to identify abnormal transactions, the mutual information (MI) value is calculated between each dimension in the feature vector and a preset abnormal transaction identifier. This identifier can be a manually labeled known abnormal transaction sample or a rule-based abnormal transaction feature. By calculating the MI value between each dimensional feature and the abnormal transaction identifier, the contribution of that feature to identifying abnormal transactions is assessed. A higher MI value indicates a greater contribution to identifying abnormal transactions. The calculated MI values ​​are normalized to obtain an importance weight for each dimensional feature. For example, if the MI value for the transaction amount feature is 0.8 and the MI value for the transaction time feature is 0.5, the normalized weights might be 0.615 and 0.385, respectively. The importance weights are then weighted and summed with the eigenvalues ​​for the corresponding dimensions in the feature vector to produce a weighted feature vector. This weighting process emphasizes features that are more important for identifying abnormal transactions, improving the accuracy of the model.

[0027] The weighted feature vector is divided into multiple subvectors. This division can be based on the attribute categories of the features, for example, dividing amount-related features, time-related features, and transaction type-related features into different subvectors. For each subvector, a binary hash value is calculated. The hash algorithm can be SHA-256 or MD5, which converts subvectors of varying lengths into hash values ​​of fixed length. The hash values ​​of each subvector are combined to generate a hash signature. The hash signature uniquely identifies a specific transaction pattern, facilitating rapid comparison and retrieval.

[0028] Semantic features representing transaction types and account attributes are extracted from the feature vector. Semantic features capture the meaning and association of transaction types and account attributes within the business context. For example, while "selling goods" and "purchasing raw materials" are both transactions, they have different semantic meanings within the business context. Semantic features are extracted by analyzing the textual descriptions of transaction types and account attributes. These semantic features are combined with the hash signature to generate the final voucher digital fingerprint. The voucher digital fingerprint uniquely identifies the original voucher data and is used for subsequent data verification and anomaly detection. When an enterprise submits tax declaration data, the system can quickly retrieve and compare relevant vouchers based on the voucher digital fingerprint to verify the accuracy and completeness of the declared data.

[0029] Based on the above technical solution, it is possible to achieve efficient feature extraction and accurate representation of the original voucher data of the enterprise, construct a digital fingerprint through multi-dimensional features such as amount, time, transaction type and subject attributes, and improve the uniqueness and identifiability of the voucher data. By introducing normalization, discretization, co-occurrence matrix analysis and dimensionality reduction processing, the expressiveness and distinguishing ability of features are enhanced; further combining mutual information calculation and weighting mechanism, it effectively highlights the feature dimensions that have a significant contribution to the identification of abnormal transactions, and improves the sensitivity and accuracy of fingerprint generation. The final voucher digital fingerprint combines structural features and semantic information, has good comparability and tamper resistance, and provides a reliable foundation for voucher anti-counterfeiting verification, duplicate detection and intelligent auditing.

[0030] In an optional embodiment, establishing a credential association graph based on the credential digital fingerprint, analyzing the business association relationships between the credentials in the credential association graph, and generating an abnormal transaction chain identification result includes: Construct association nodes based on the digital fingerprint of the voucher, calculate the cosine similarity between adjacent nodes as the association strength, set edge weights based on the association strength, extract transaction time information from the original voucher data to determine the direction of the edge, calculate a dynamic filtering threshold based on the number of connections and distribution density of the nodes, and use the dynamic filtering threshold to filter the edge weights to obtain a voucher association graph; Calculating the degree of association between adjacent nodes in the credential association graph to obtain a node weight coefficient, performing weighted fusion on the business features of the nodes according to the node weight coefficient to obtain a business association feature, combining the business association feature with the credential digital fingerprint to generate node representation information; Arranging the node representation information in chronological order to form a feature sequence, extracting the dependency relationship between the nodes to obtain a transaction chain feature, calculating the similarity between the transaction chain feature and the preset abnormal feature, and marking the continuous transaction sequence with a similarity higher than a first preset threshold as a candidate abnormal transaction chain; The time interval and amount change rate of adjacent transactions in the candidate abnormal transaction chain are calculated, and the correlation between the time interval and amount change rate and the transaction type is used as the transaction correlation degree. The candidate abnormal transaction chain is scored according to the transaction correlation degree, and the candidate abnormal transaction chain with a score exceeding a second preset threshold is determined as an abnormal transaction chain, and an abnormal transaction chain identification result is generated.

[0031] Figure 2 This is a flowchart for identifying abnormal transaction chains based on digital fingerprints of vouchers. For example, first, associated nodes are constructed based on the digital fingerprints of the vouchers. Each digital fingerprint of the voucher corresponds to a node, and the node contains the core feature information of the voucher. For any two adjacent voucher nodes, the cosine similarity between them is calculated as the association strength. During the cosine similarity calculation process, the two voucher digital fingerprints are regarded as multidimensional vectors, and the cosine value of the angle between the vectors is calculated. The closer the value is to 1, the more similar the two vouchers are. For example, the digital fingerprint vectors of two invoice vouchers are [0.25, 0.32, 0.18, 0.45] and [0.28, 0.30, 0.20, 0.42] respectively. The calculated cosine similarity is 0.997, indicating that the two vouchers are highly similar.

[0032] Edge weights are set based on the calculated association strength; higher association strengths result in larger edge weights. Transaction time information is extracted from the original voucher data to determine the direction of edges. Typically, earlier vouchers point to later vouchers, indicating the direction of business flow. For example, a raw material purchase voucher might point to a subsequent production and processing voucher, which then points to a product sales voucher, forming a complete business process chain.

[0033] In order to filter out truly relevant voucher connections, this method calculates a dynamic filtering threshold based on the number of connections and distribution density of the nodes. In actual business, different types of vouchers may have different connection characteristics. For example, daily small reimbursement vouchers may have fewer connections with other vouchers, while core transaction vouchers for large projects may be associated with multiple vouchers. The calculation of the dynamic filtering threshold takes into account the average number of connections of the nodes and the standard deviation of the connection distribution, so that the threshold can adapt to different business scenarios. For example, in a company's voucher network, if the average number of connections of the nodes is 5 and the standard deviation is 2, the dynamic threshold may be set to 0.3; in the case of dense connections, such as the average number of connections is 15 and the standard deviation is 4, the threshold may be increased to 0.5 to ensure that truly closely related connections are filtered out.

[0034] The calculated dynamic filtering threshold is used to filter edge weights, retaining edges with weights greater than the threshold and removing edges with weights less than the threshold to obtain a voucher association graph. This processed voucher association graph clearly demonstrates the substantive business connections between vouchers while filtering out accidental, non-business-driven connections.

[0035] After constructing the voucher association graph, the degree of association between adjacent nodes in the graph is calculated to obtain the node weight coefficient. This calculation takes into account the node's position in the graph, its connection pattern with other nodes, and its importance within the business context. For example, a voucher node located at the intersection of multiple transaction chains may have a higher weight coefficient, while a voucher node at the end of a transaction chain may have a lower weight coefficient. Specifically, the weight coefficient can be calculated by comprehensively considering factors such as the node's degree centrality (how many other nodes it is directly connected to) and betweenness centrality (how often the node acts as a bridge to other nodes).

[0036] Based on the calculated node weight coefficient, the business features of the nodes are weighted and integrated to generate business-related features. Business features can include transaction type, amount range, counterparty type, and more. During this weighted integration process, features with higher importance are given greater weight. For example, for a special VAT invoice, the tax amount may be given a higher weight; for an expense reimbursement form, the reimbursement type and amount may be given higher weights. Business-related features are combined with the voucher's digital fingerprint to generate node representation information, which comprehensively describes the voucher's characteristics and position in the business network.

[0037] Arrange node representation information in chronological order to form a feature sequence. This chronological arrangement reflects the natural evolution of business processes and helps identify unusual transactions that deviate from normal business processes. Dependencies between nodes are extracted from the feature sequence to generate transaction chain features. Dependencies can be extracted based on chronological order, amount flow, logical associations between business types, and more. For example, a raw material procurement transaction may have a clear dependency with subsequent production and processing transactions. This dependency is reflected in the temporal order, amount matching (taking into account processing value-added), and consistency in business types.

[0038] Calculate the similarity between transaction chain features and pre-defined anomaly features, which are a set of pre-defined abnormal transaction patterns based on historical anomaly cases and professional knowledge. Anomaly features may include cyclical transaction patterns, unusual time intervals, and fluctuations in amounts that do not conform to industry norms. For example, frequent capital flows between related companies within a short period of time may indicate fictitious transactions; a significant deviation in the ratio of sales revenue to corresponding VAT output tax from industry standards may indicate tax evasion. Similarity calculations can use vector distance metrics to assess the degree of closeness between the actual transaction chain and the anomaly pattern.

[0039] Continuous transaction sequences with a similarity above a first preset threshold are marked as candidate abnormal transaction chains. The first preset threshold needs to balance the detection rate and false positive rate, and can be adjusted based on historical data and business characteristics. For example, for risk-sensitive enterprises, the threshold can be set at 0.75; for general enterprises, the threshold can be set at 0.85 to reduce the interference caused by false positives.

[0040] For the identified candidate abnormal transaction chains, this method further calculates the time interval and amount change rate between adjacent transactions in the candidate abnormal transaction chain. The time interval refers to the time difference between adjacent transactions, and the amount change rate refers to the proportional change in the amounts of adjacent transactions. The correlation between the time interval and amount change rate and the transaction type is used as the transaction correlation. Transaction correlation assesses the logical rationality between transactions and takes into account the unique characteristics of different transaction types. For example, a very short time interval between raw material procurement and finished product sales may be unreasonable, while a very long time interval between daily office supply purchases and expense reimbursements may also be problematic.

[0041] Candidate abnormal transaction chains are scored based on the calculated transaction correlation. This scoring mechanism comprehensively considers the correlation between all adjacent transaction pairs within the transaction chain, as well as the integrity and consistency of the entire transaction chain. Candidate abnormal transaction chains with scores exceeding a second preset threshold are identified as abnormal transaction chains, and an abnormal transaction chain identification result is generated. The second preset threshold is typically set higher than the first preset threshold to ensure a high degree of credibility for the ultimately identified abnormal transaction chain.

[0042] In a real-world application, when a company filed its VAT return, the system received nearly a year's worth of VAT special invoice data. Based on the digital fingerprints generated from this data, a voucher association graph consisting of approximately 5,000 nodes was constructed. Analysis of this association graph revealed an unusual transaction chain consisting of seven nodes. These transactions, which recurred over a three-month period, involved business transactions between the company and three of its suppliers. The time intervals between these transactions were unusually short (an average of only two days), and the transaction amounts were nearly identical (approximately 1 million yuan), with highly similar transaction descriptions. The system calculated a similarity of 0.92 between this transaction chain and the pre-defined "false VAT invoice" anomaly pattern, exceeding the first preset threshold of 0.85, and therefore marked it as a candidate abnormal transaction chain.

[0043] Further analysis revealed that the rate of change in the transaction amounts of these transactions was close to zero, the time intervals were significantly shorter than the industry's normal business cycle, and the transaction descriptions did not align with actual business operations. The system calculated a transaction correlation score of 0.94, exceeding the second preset threshold of 0.9, ultimately identifying this transaction chain as abnormal. The system's identification of abnormal transaction chains indicated that the company may have engaged in illegal VAT invoicing to inflate input tax. Tax officials were advised to focus on verifying the authenticity and compliance of the relevant transactions and to make appropriate corrections to the company's VAT declaration data.

[0044] Based on the above technical solution, it is possible to achieve in-depth modeling of business relationships between corporate credentials and intelligent identification of abnormal transaction chains. By constructing an association graph through the digital fingerprint of the credential and combining cosine similarity, transaction time and connection density to dynamically filter edge weights, irrelevant or weakly associated nodes are effectively eliminated, improving the accuracy and structural clarity of the association graph. Combining the weighted fusion of node weight coefficients and business characteristics, the ability to express potential business relationships between nodes is enhanced; through time series chain building and dependency analysis, the logical evolution path between consecutive transactions can be accurately portrayed. With the help of similarity calculation and transaction feature scoring mechanism, the accurate screening and marking of transaction chains with potential abnormal characteristics can be further achieved, improving the automation level and accuracy of abnormal transaction detection.

[0045] In an optional embodiment, based on the abnormal transaction chain identification result and the voucher association graph, converting the original voucher data into financial accounting data includes: Calculate the amount change rate of adjacent nodes in the abnormal transaction chain to obtain an amount change feature, calculate the transaction time difference of adjacent nodes to obtain a time interval feature, and extract the debit and credit account combinations of adjacent nodes to obtain an account combination feature; determine the abnormal propagation direction based on the amount change feature, time interval feature, and account combination feature, and generate an abnormal propagation path; mark the original vouchers corresponding to the associated nodes located on the abnormal propagation path in the voucher association graph as affected vouchers; In the voucher association diagram, select the node sequence without abnormality as the normal transaction chain, extract the matching relationship and corresponding relationship of the debit and credit accounts in the normal transaction chain, and generate the accounting rule template; Based on the connections of the abnormal propagation paths, adjacent affected vouchers are grouped into the same accounting unit. Within each accounting unit, the debit and credit accounts of the affected vouchers are paired and verified according to the accounting rule template. The accounting results are compared with historical normal data to identify accounting deviations. Based on the business association strength between nodes in the voucher association graph, subgraphs are divided, and parallel accounting processing is performed on the original vouchers in each subgraph to generate subgraph accounting results; the subgraph accounting results are merged according to the association relationship between subgraphs to generate financial accounting data.

[0046] After obtaining original voucher data, identifying abnormal transaction chains, and constructing a voucher association graph, the original voucher data needs to be converted into financial accounting data. First, the amount change rate of adjacent nodes in the abnormal transaction chain is calculated to obtain the amount change feature. For example, for adjacent nodes A and B, if the transaction amount at node A is 10,000 yuan and the transaction amount at node B is 12,000 yuan, the amount change rate is (12,000 - 10,000) / 10,000 = 0.2, indicating a 20% increase in the amount. The transaction time difference between adjacent nodes is also calculated to obtain the time interval feature. For example, if the transaction time at node A is 10:00 on May 10, 2023, and the transaction time at node B is 14:30 on May 10, 2023, the time interval is 4.5 hours. The system also extracts the debit and credit account combinations of adjacent nodes as the account combination feature. For example, if node A is "Debit: Accounts Receivable, Credit: Principal Operating Income" and node B is "Debit: Inventory, Credit: Accounts Payable," the account combination feature is the combination of these two sets of debit and credit accounts.

[0047] The direction of anomaly propagation is determined based on the extracted features. When the amount change rate is abnormal (e.g., exceeding the normal business threshold of 0.15), the time interval characteristics are abnormal (e.g., less than the system-defined minimum normal business processing time of 0.5 hours), and the account combination characteristics do not conform to normal business processes (e.g., not meeting the preset account matching rules), the system determines that the anomaly has propagated from A to B. By connecting these nodes with propagation relationships, the system generates an anomaly propagation path, such as A→B→C→D. In the voucher association diagram, the system marks the original vouchers corresponding to the associated nodes on the anomaly propagation path as affected vouchers. That is, the original vouchers corresponding to A, B, C, and D are marked as affected vouchers.

[0048] In the voucher association diagram, select node sequences that are not marked as abnormal as normal transaction chains. For example, the node sequence E→F→G is not marked as abnormal and is considered a normal transaction chain. From this normal transaction chain, extract the matching relationships between debit and credit accounts and the corresponding amounts to generate an accounting rule template. For example, if the "Debit: Bank Deposits, Credit: Accounts Receivable" link at node E and the "Debit: Accounts Receivable, Credit: Principal Operating Income" link at node F are related to the account "Accounts Receivable" and the amounts are the same, this matching relationship is recorded as part of the accounting rule template.

[0049] Based on the connections along the exception propagation path, adjacent affected vouchers are grouped into the same accounting unit. For example, if A→B and C→D are adjacent in the exception propagation path, A and B are grouped into one accounting unit, while C and D are grouped into another. Within each accounting unit, the system verifies the matching of the debit and credit accounts of the affected vouchers based on the accounting rule template. For example, within accounting units A and B, the system checks whether the credit account "Main Business Income" in A matches the debit account in B as defined in the rule template. The accounting results are compared with historical normal data. If the credit account "Main Business Income" in A is 10,000 yuan, but there is no corresponding debit account in B, or the matching amounts are inconsistent, an accounting discrepancy of 10,000 yuan is identified.

[0050] The subgraph is divided based on the business association strength between nodes in the voucher association graph. The business association strength can be calculated by the transaction frequency, amount similarity and account association between nodes. For example, the business association strength between the node set {A, B, C, D} is high, forming a subgraph. Figure 1 ; The business association strength between the node set {E, F, G, H} is high, forming a sub Figure 2 The system performs parallel accounting processing on the original vouchers in each sub-graph. Figure 1 According to the above method, the calculation relationship between A, B, C, and D is analyzed to generate sub Figure 1 The accounting results, including account balances, accounting deviations and other information. Figure 2 Generate its accounting results.

[0051] Finally, the subgraph accounting results are merged based on the relationship between the subgraphs to generate the final financial accounting data. The relationship between subgraphs is mainly determined by shared nodes or cross-subgraph business processes. For example, if a subgraph Figure 1 Kazuko Figure 2 Shared account "Accounts Receivable" and sub Figure 1 The ending balance of this account is RMB 50,000. Figure 2 The opening balance of this account should be RMB 50,000. Verify whether this relationship is established. Through consolidation, generate complete financial accounting data including account balance sheet, income statement, cash flow statement, etc.

[0052] In this embodiment, by combining the results of abnormal transaction chain identification with a voucher association graph, a structured conversion of original voucher data into financial accounting data is achieved. Compared with the existing technology that relies on static mapping based on a rule base or accounting judgment based only on a single voucher data, it has higher accuracy and flexibility. The existing technology mostly relies on manually formulated fixed rule templates, which cannot dynamically identify the complex propagation relationships between vouchers. In processing business scenarios involving abnormal propagation, important information is often omitted, resulting in distorted accounting results or missed abnormalities. This application extracts key features such as amount changes, time intervals, and account combinations to construct abnormal propagation paths and mark affected vouchers, thereby achieving dynamic identification and attribution of abnormal chains. It further generates accounting rule templates based on normal transaction chains, replacing static rule configuration, improving rule adaptability and business coverage capabilities. It then combines the association graph to divide accounting units and parallelize processing, solving the efficiency bottleneck in the process of processing large-scale voucher data. The starting point of the improvement is to improve the abnormality perception and business recovery capabilities in the voucher accounting process, ultimately achieving the technical effects of improved accounting accuracy, more contextual abnormality identification, and more efficient overall processing.

[0053] In an optional embodiment, real-time tax rule data is obtained, financial accounting data is compared with the tax rule data, and the degree of logical association between data items in the financial accounting data is calculated, forward traversal and reverse verification are performed to identify data items that violate tax rules, and risk weight values ​​are assigned to the data items that violate tax rules. Obtain tax rule data through the rule data interface, extract the judgment conditions and constraints in the tax rule data, construct a rule description vector, calculate the similarity between the rule description vector and the rule vector in the local rule library, identify new and changed rule content, perform structural processing, and convert it into a rule expression containing trigger conditions, judgment thresholds, and constraint parameters; The data items in the financial accounting data are constructed into an association graph, with the data items as nodes and the logical dependencies between the data items as edges. The numerical association degree and the subject association degree between adjacent nodes are calculated, and the weighted values ​​of the two association degrees are used as edge weights. Based on the edge weights, the shortest path between the nodes is calculated to obtain the association strength of the data items. Traverse the nodes in the association graph in the order of business processes, compare the attribute values ​​of data items with the rule expressions, and mark data items as violation nodes when they do not meet the rule expressions; Starting from each violation node, trace back along the association graph to identify the associated nodes that affect the violation node; mark the path from the associated node to the violation node as the violation propagation path; calculate the impact range of the node based on the violation propagation path; A basic score is determined based on the number of rule violations, a propagation coefficient is calculated based on the scope of influence, and an aging coefficient is calculated based on the update time of the rule expression; the product of the basic score, the propagation coefficient and the aging coefficient is used as the risk weight value.

[0054] For example, tax rule data is first obtained through the rule data interface. This interface can connect to a database or authorized third-party data source to regularly synchronize the latest tax rules. Tax rule data includes, but is not limited to, calculation rules, reporting requirements, and deduction conditions for various taxes, such as value-added tax, corporate income tax, and personal income tax. Decision conditions and constraints are extracted from the obtained tax rule data. Decision conditions determine the scope of application of the rule, such as "general taxpayers with annual sales exceeding 5 million yuan." Constraints specify specific compliance requirements, such as "VAT special invoice deductions must be certified within 180 days." Based on the extracted decision conditions and constraints, a rule description vector is constructed. The rule description vector is a numerical representation of the tax rule, containing feature values ​​for multiple dimensions, such as rule type, scope of application, decision threshold, and timeliness. For example, for the rule "75% additional deduction for enterprise R&D expenses," the rule description vector might include features such as rule type (income tax preferential treatment), applicable entity (general enterprises), deduction ratio (75%), and time range (current fiscal year).

[0055] The rule description vector is compared with the rule vectors in the local rule base to identify new and changed rules. Similarity is calculated using the vector cosine similarity method, which calculates the cosine of the angle between the two rule vectors. The closer this value is to 1, the more similar the rules are. This comparison identifies new rules (similar rules do not exist in the local rule base), changed rules (highly similar to local rules but with modified parameters), and discontinued rules (local rules no longer exist in the new rules). The identified new and changed rules are structured and converted into rule expressions containing trigger conditions, decision thresholds, and constraint parameters. Rule expressions use an "if-then" logic structure, for example, "If enterprise type = technology-based SME and R&D expense ratio > 5%, then the R&D expense deduction ratio is 100%." ​​This structured rule expression facilitates automated comparison and judgment by the system.

[0056] Data items in financial accounting data are constructed as an association graph, with data items as nodes and logical dependencies between data items as edges. Data items can be accounting accounts, voucher entries, specific values ​​in declaration forms, etc. Logical dependencies represent computational relationships or business connections between data items. For example, "Taxable Income" depends on "Total Revenue" and "Allowable Deductions." The numerical and account correlations are calculated between adjacent nodes in the association graph. The numerical correlation reflects the correlation between numerical changes between nodes and can be calculated by analyzing the numerical change patterns of two data items in historical data. For example, analyzing the numerical ratio of "Sales Revenue" and "VAT Payable - Output Tax" over multiple accounting periods. If the ratio remains stable at around 17%, it indicates that the two data items have a high numerical correlation. The account correlation reflects the correlation between business attributes between nodes and is determined based on the attribute classification and business significance of accounting accounts. For example, "Fixed Assets" and "Accumulated Depreciation" have a high account correlation, while "Fixed Assets" and "Wages Payable" have a low account correlation. The numerical and account correlations are weighted together to obtain edge weights. Typically, you can set the numeric association weight to 0.6 and the subject association weight to 0.4, but the specific weights can be adjusted based on the company's actual situation. Based on the calculated edge weights, a graph algorithm is applied to calculate the shortest path between nodes to determine the association strength of the data items. A higher association strength indicates a closer dependency between the data items.

[0057] The nodes in the association graph are traversed in the order of the business process. This business process sequence typically follows the basic accounting process, from source documents, accounting vouchers, accounting books, to financial statements and tax returns. For each node, the attribute values ​​of its data items are compared with the rule expression. Attribute values ​​include the data item's value, type, and time attributes. The comparison process determines whether the data item meets the rule requirements based on the logical structure of the rule expression. If a data item does not meet the rule expression, it is marked as a violation node. For example, a company's "R&D expense super deduction" data item shows a deduction rate of 100%. However, according to the rule expression, the company is not a technology-based SME and its R&D expense ratio is less than 5%, which does not meet the conditions for a 100% super deduction. Therefore, the data item is marked as a violation node.

[0058] Starting from each violation node, trace back along the association graph to identify the associated nodes that affect the violation node. The depth of reverse tracing can be capped, such as tracing back three layers of associated nodes, to avoid inefficient analysis due to excessive tracing scope. Associated nodes refer to other data items that directly or indirectly affect the violation node through business logic or calculation relationships. The path from the associated node to the violation node is marked as the violation propagation path. The influence range of the node is calculated based on the violation propagation path. The influence range takes into account the position of the node in the association graph, the number of connections with other nodes, and the strength of the connections. The influence range of core business data is generally larger, while the influence range of edge auxiliary data is smaller.

[0059] The base score is determined based on the number of rule violations. The more rules violated, the higher the base score. Grading standards can be set, such as 10 points for violating 1-2 rules, 20 points for violating 3-5 rules, and 30 points for violating 6 or more rules. The propagation coefficient is calculated based on the scope of influence. The propagation coefficient reflects the impact of the violation data on the overall tax declaration. The larger the scope of influence, the higher the propagation coefficient. For example, the propagation coefficient for an impact scope involving 1-3 associated nodes is 1.2, the propagation coefficient for an impact scope involving 4-6 associated nodes is 1.5, and the propagation coefficient for an impact scope involving 7 or more associated nodes is 2.0. The timeliness coefficient is calculated based on the update time of the rule expression. Newly introduced or recently updated rule violations have a higher timeliness coefficient; long-standing and stable rule violations have a lower timeliness coefficient. For example, the timeliness coefficient for rules updated within 30 days is 1.5, for rules updated between 30 and 90 days it is 1.3, for rules updated between 90 and 180 days it is 1.1, and for rules updated over 180 days it is 1.0. The risk weight is the product of the base score, the propagation coefficient, and the timeliness coefficient. This value intuitively reflects the risk level of the offending data item and provides a basis for subsequent intelligent error correction.

[0060] In this embodiment, by introducing a real-time access and structured parsing mechanism for tax rules, the problems of delayed rule updates, limited expressive power, and inability to adapt to complex financial and tax scenarios in the existing technology are solved. Based on the similarity comparison between the rule description vector and the local rule vector, the newly added or changed tax rules are automatically identified and parsed, and a structured rule expression is constructed to achieve dynamic updating and precise matching of the rules. By constructing a data item association graph and integrating the numerical correlation and subject correlation, the modeling capability of the internal logical relationship of financial data is improved. On this basis, combined with the two-way verification mechanism of forward traversal and reverse tracing, not only can the data items that directly violate the rules be identified, but also their impact paths can be traced and potential risk sources can be identified. Finally, by introducing the propagation coefficient and the timeliness coefficient, a quantitative assessment of the risk weight of the violation node is achieved, which improves the real-time and intelligent level of risk management.

[0061] In an optional embodiment, a data balance constraint matrix is ​​constructed based on risk weight values, the degree of deviation of data items violating tax rules in the data balance constraint matrix is ​​analyzed, data items with the same degree of deviation are grouped into data correction clusters, grouped iterative correction is performed on the data correction clusters, and when the correction results meet the balance constraints, the tax return data is generated, including: A self-organizing multi-state equilibrium network is constructed based on risk weight values. The initial state of each data item is determined by the risk weight value. Data items influence each other through state transition rules. The state transition rules include two modes: weight decay and weight enhancement, generating a state evolution sequence of data items. Identifying key state transition points based on the state evolution sequence, wherein the key state transition points reflect the moment when the state of the data item suddenly changes, performing comparative analysis on the data items before and after the key state transition points, and extracting change characteristics of the data items; Calculating a state change value of the data item according to the change characteristics, dividing the data items whose state change values ​​are in the same preset interval into the same data correction cluster, and calculating an ideal state value of the data item in each data correction cluster based on the state transition rule; Data correction is performed alternately based on the weight decay and weight enhancement modes in the state transition rules, and the degree of difference between the data items after each round of correction and the ideal state value is calculated. When the degree of difference is less than the preset difference threshold, the data correction cluster is determined to have reached a stable state. When all data correction clusters have reached a stable state, tax return data is generated.

[0062] During implementation, a self-organizing multi-state equilibrium network is first constructed based on risk weight values. A self-organizing multi-state equilibrium network is a special data structure used to simulate the mutual influence and equilibrium relationship between data items. In this network, each data item is represented as a node with a state attribute, and the node's initial state is determined by its risk weight value. A higher risk weight indicates a more serious violation of tax rules by the data item, and the further its initial state deviates from the equilibrium point. For example, for a data item with a risk weight of 25, its initial state might be set to -5 (assuming the equilibrium point is 0); for a data item with a risk weight of 10, its initial state might be set to -2. The sign of the state value indicates the direction of deviation: a negative value indicates a lower-than-expected value, and a positive value indicates a higher-than-expected value.

[0063] Data items influence each other through state transition rules. State transition rules include weight decay and weight boosting. Weight decay applies to data items with excessively high state values. By transferring some of their state value to related data items, their state value is brought closer to equilibrium. For example, if the state value of the "Taxable Sales" data item in a VAT return is +4, weight decay might transfer 1 unit of state value to the related "Output Tax Amount" data item, reducing its own state value to +3. Weight boosting applies to data items with excessively low state values. By acquiring state value from related data items, their own state value is increased. For example, if the state value of the "Input Tax Amount" data item is -3, weight boosting might acquire 1 unit of state value from the related "Input Tax Amount Transferred Out" data item, raising its own state value to -2.

[0064] By applying the state transition rules over multiple rounds of iteration, a state evolution sequence for each data item is generated. This state evolution sequence records the trajectory of each data item's state value during the iteration process. For example, the state evolution sequence for a data item might be [-5, -4, -3, -2, -1.5, -1.2, -1.0, -0.9, -0.85], indicating that after eight rounds of iteration, the state value of the data item gradually approaches -0.85 from -5.

[0065] Identify key state transition points based on state evolution sequences. Key state transition points refer to moments when the state value of a data item changes significantly, reflecting the key nodes where the data item undergoes mutations due to external influences. The identification of key state transition points can be based on the state value change rate. When the change rate exceeds a preset threshold, the corresponding moment is marked as a key state transition point. For example, if the change rate threshold is set to 50%, when the state value changes from -4 to -2, the change rate is 50%, and this moment is marked as a key state transition point. Compare and analyze the data items before and after the key state transition point to extract the change characteristics of the data items. Change characteristics include the magnitude of the state value change, the direction of change, the rate of change, etc.

[0066] The state change value of the data item is calculated based on the change characteristics. The state change value reflects the degree of change of the data item from the initial state to the current state, and can be calculated by the difference between the initial state and the current state. Data items whose state change values ​​are in the same preset interval are divided into the same data correction cluster. For example, data items with state change values ​​in the interval [2, 3] may be divided into the same data correction cluster, indicating that these data items have similar change characteristics and may require similar correction strategies. The ideal state value of the data items in each data correction cluster is calculated based on the state transition rule. The ideal state value refers to the state that the data item should reach under equilibrium conditions, usually close to the equilibrium point.

[0067] Data correction is performed alternately using the weight decay and weight boost modes in the state transition rules. This alternating process prevents the correction process from falling into local optima and improves the stability of the correction results. After each round of correction, the difference between the data item and the ideal state value is calculated. This difference can be calculated using the Euclidean distance between the current state value and the ideal state value, or other distance metrics. When the difference falls below a preset difference threshold, the data correction cluster is considered to have reached a stable state. When all data correction clusters have reached a stable state, the final tax return data is generated.

[0068] In this embodiment, by constructing a data balance constraint mechanism driven by risk weights, intelligent correction of data that violates tax rules is achieved. Compared to existing methods that rely on manual judgment or static rule backfilling, this approach improves the adaptability and overall consistency of the correction process. Existing technologies often lack modeling of dynamic relationships between data, resulting in isolated correction processes, which can easily lead to local adjustments that disrupt overall balance. To address this issue, this solution introduces a self-organizing multi-state balance network, which drives state evolution using risk weights and implements state transition rules to implement the mutual influence between data items, accurately identifying key nodes of data mutation. By extracting change characteristics, data items with similar states are clustered into correction clusters, avoiding ineffective correction of isolated data. An iterative approach, combined with weight decay and enhancement strategies, guides data convergence towards the ideal state, ensuring the directionality and stability of the correction. Ultimately, through global convergence judgment, tax return data that meets the balance constraints is generated. This achieves batch repair and standardized output of non-compliant data while maintaining business logic consistency, significantly improving tax data quality and compliance.

[0069] In an optional embodiment, data correction is performed alternately based on the weight decay and weight enhancement modes in the state transition rule, the difference between the data item after each round of correction and the ideal state value is calculated, and when the difference is less than a preset difference threshold, it is determined that the data correction cluster has reached a stable state, including: Obtaining a difference sequence between a current state value and an ideal state value of a data item, establishing a state offset curve based on the difference sequence, and calculating a fluctuation trend and a fluctuation amplitude of the state offset curve; Alternately switching between a weight decay mode and a weight enhancement mode based on the positive and negative changes in the difference sequence, generating an adaptive adjustment coefficient based on the fluctuation trend and the current correction round, executing the weight decay mode when the difference sequence is positive, and executing the weight enhancement mode when the difference sequence is negative, dynamically updating the current state value of the data item based on the adaptive adjustment coefficient, updating the current state value and recalculating the difference sequence after each round of execution; Among them, the weight decay mode realizes downward adjustment of data items based on a multi-level decreasing function, and the weight enhancement mode realizes upward adjustment of data items based on a piecewise enhancement function; The degree of difference of all data items in the data correction cluster is calculated based on the current state value and the ideal state value, a stability evaluation index is generated according to the degree of difference, and the trend of change of the stability evaluation index over multiple consecutive rounds is analyzed. When the degree of difference is less than a preset difference threshold and is maintained for a specified number of rounds, it is determined that the data correction cluster has reached a stable state.

[0070] In this embodiment, a state shift curve is first constructed by obtaining a sequence of differences between the current state value and the ideal state value of a data item. The state shift curve's fluctuation trend and amplitude are then calculated. A difference sequence is the sequence of differences between the current state value and the ideal state value of a data item during each iteration of data correction. For example, the difference sequence for the "Tax Payable" data item in a VAT return over five consecutive correction rounds might be [+2.5, +1.8, +1.2, +0.7, +0.4], indicating that the state value of the data item is gradually approaching its ideal state value. The state shift curve is a graphical representation of the difference sequence, reflecting the trend of the data item's state value over the correction rounds. The fluctuation trend is determined by analyzing the slope of the state shift curve and can be categorized as rapidly converging, slowly converging, oscillating, and diverging. The amplitude is determined by calculating the change in the difference between adjacent rounds, reflecting the severity of the state adjustment.

[0071] Switching alternately between weight decay mode and weight enhancement mode is done based on the positive and negative changes in the difference sequence. When the difference sequence is positive, it means that the current state value of the data item is higher than the ideal state value, and the weight decay mode needs to be executed for downward adjustment; when the difference sequence is negative, it means that the current state value of the data item is lower than the ideal state value, and the weight enhancement mode needs to be executed for upward adjustment. An adaptive adjustment coefficient is generated based on the fluctuation trend and the current correction round. The adaptive adjustment coefficient is a dynamically changing parameter used to control the intensity of each round of correction. In the early stages of correction, if the fluctuation trend shows rapid convergence, a larger adjustment coefficient can be set to accelerate convergence; in the later stages of correction, if it is close to a stable state, a smaller adjustment coefficient is set to avoid over-adjustment and resulting in oscillation. For example, for data items that show a rapid convergence trend and are in the early stages of correction, the adjustment coefficient may be set to 0.8; while for data items close to a stable state, the adjustment coefficient may be reduced to 0.2.

[0072] The current state value of the data item is dynamically updated based on the adaptive adjustment coefficient. The weight decay mode implements downward adjustment of the data item based on a multi-level decreasing function. The multi-level decreasing function calculates the state amount to be reduced based on the current state value, ideal state value, and adaptive adjustment coefficient of the data item. This function can be designed into multiple intervals, each using a different decay strategy. For example, when the difference is greater than 2, linear decay is used; when the difference is between 0.5 and 2, quadratic decay is used; and when the difference is less than 0.5, cubic decay is used. This segmented design can provide appropriate decay strength at different correction stages, ensuring convergence speed while avoiding over-adjustment.

[0073] The weighted boosting mode uses a piecewise boosting function to adjust data items upward. Similar to a multi-level decreasing function, but in the opposite direction, a piecewise boosting function is used to increase the status value of a data item. This function can also be designed with multiple intervals to accommodate different boosting requirements. For example, when the difference is less than -2, exponential boosting is used; when the difference is between -2 and -0.5, linear boosting is used; and when the difference is between -0.5 and 0, logarithmic boosting is used. After each round of execution, the current status value of the data item is updated and the difference sequence is recalculated to provide a basis for the next round of corrections.

[0074] The degree of difference of all data items in the data correction cluster is calculated based on the current state value and the ideal state value. The degree of difference can be obtained by calculating statistical indicators such as the average absolute value or standard deviation of the state difference values ​​of all data items. A stability evaluation index is generated based on the degree of difference. The stability evaluation index not only considers the degree of difference in the current round, but also the trend of change over multiple consecutive rounds. It is an indicator that comprehensively reflects the convergence state of the data correction cluster. The trend of change of the stability evaluation index over multiple consecutive rounds is analyzed. When the degree of difference is less than the preset difference threshold and maintained for the specified rounds, the data correction cluster is judged to have reached a stable state. For example, the difference threshold can be set to 0.2 and the specified rounds can be set to 3. That is, when the degree of difference of the data correction cluster is less than 0.2 for three consecutive rounds, the cluster is judged to have reached a stable state.

[0075] This application achieves dynamic and progressive correction of illegal data items by introducing an alternating mechanism of weight decay and enhancement modes driven by a difference sequence. Compared with the existing static correction methods that rely on fixed thresholds or one-way adjustment strategies, it has stronger sensitivity and convergence. Existing technologies usually ignore the volatility of state changes during the correction of data items, resulting in mismatched or stagnant correction rhythms. This solution constructs a state offset curve to analyze the fluctuation trend and amplitude of the data item state in real time, automatically switches between decay and enhancement modes based on the difference sign, and introduces an adaptive adjustment coefficient to dynamically adjust the intensity of each round of correction, making the correction path more directional and flexible. The decay and enhancement functions control the speed of data regression to the ideal state in a hierarchical and segmented manner, respectively, to avoid over-correction or oscillation. The stability assessment mechanism determines whether the correction effect has converged through multiple rounds of difference change trends, ensuring that the data correction cluster reaches a balanced state before generating the final result, thereby improving the stability, accuracy and intelligence of the data correction, and laying a solid foundation for the compliance output of subsequent declared data.

[0076] A second aspect of an embodiment of the present invention provides a system for automatic verification and intelligent error correction of corporate tax declaration data, the system comprising: The first unit is configured to receive original voucher data input by the enterprise end and perform feature extraction, construct a feature vector, generate a voucher digital fingerprint based on the feature vector, establish a voucher association graph based on the voucher digital fingerprint, analyze the business association relationships between the vouchers in the voucher association graph, generate an abnormal transaction chain identification result, and convert the original voucher data into financial accounting data based on the abnormal transaction chain identification result and the voucher association graph; The second unit is used to obtain real-time tax rule data, compare financial accounting data with the tax rule data, calculate the logical correlation between each data item in the financial accounting data, perform forward traversal and reverse verification, identify data items that violate tax rules, and assign risk weight values ​​to the data items that violate tax rules; The third unit is used to construct a data balance constraint matrix based on the risk weight value, analyze the degree of deviation of data items that violate tax rules in the data balance constraint matrix, group data items with the same degree of deviation into data correction clusters, perform group iterative correction on the data correction clusters, and generate tax declaration data when the correction results meet the balance constraints.

[0077] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0078] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0079] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The method for automatic verification and intelligent error correction of corporate tax declaration data is characterized by: include: Receive original voucher data input by the enterprise and perform feature extraction, construct a feature vector, generate a voucher digital fingerprint based on the feature vector, establish a voucher association graph based on the voucher digital fingerprint, analyze the business association relationship between the vouchers in the voucher association graph, generate an abnormal transaction chain identification result, and convert the original voucher data into financial accounting data based on the abnormal transaction chain identification result and the voucher association graph; Acquire real-time tax rule data, compare financial accounting data with tax rule data, calculate the logical correlation between each data item in the financial accounting data, perform forward traversal and reverse verification, identify data items that violate tax rules, and assign risk weight values ​​to these data items; A data balance constraint matrix is ​​constructed based on the risk weight value. The degree of deviation of data items that violate tax rules in the data balance constraint matrix is ​​analyzed. Data items with the same degree of deviation are grouped into data correction clusters. Group iterative correction is performed on the data correction clusters. When the correction results meet the balance constraints, tax declaration data is generated.

2. The method according to claim 1, characterized in that Receive the original credential data input by the enterprise and perform feature extraction to construct a feature vector. Generate the credential digital fingerprint based on the feature vector, including: Receiving original voucher data input by the enterprise side, and extracting transaction information and document information from the original voucher data; The transaction amount in the transaction information is segmented according to preset amount intervals, a normalized ratio of each interval is calculated to obtain an amount characteristic value, the transaction time is discretized according to preset time units to obtain a time characteristic value, and the amount characteristic value and the time characteristic value are combined to obtain a first characteristic value; Constructing a co-occurrence matrix for the transaction type and account attributes in the document information, calculating the correlation between the attributes based on the co-occurrence matrix, performing a dimensionality reduction transformation on the document information to obtain a second eigenvalue; normalizing the first and second eigenvalues ​​and concatenating them to obtain an eigenvector; Calculating the mutual information value between each dimension feature in the feature vector and the preset abnormal transaction identifier, and normalizing it to obtain the importance weight value of each dimension feature, and performing weighted summation on the importance weight value and the feature value of the corresponding dimension in the feature vector to obtain a weighted feature vector; The weighted feature vector is divided into multiple sub-vectors, the binary hash value of each sub-vector is calculated and combined to obtain a hash signature, the semantic features representing the transaction type and account attributes in the feature vector are extracted, and the semantic features are combined with the hash signature to generate a digital fingerprint of the credential.

3. The method according to claim 1, characterized in that Based on the digital fingerprint of the voucher, a voucher association graph is established, and the business association relationship between the vouchers in the voucher association graph is analyzed to generate abnormal transaction chain identification results including: Construct association nodes based on the digital fingerprint of the voucher, calculate the cosine similarity between adjacent nodes as the association strength, set edge weights based on the association strength, extract transaction time information from the original voucher data to determine the direction of the edge, calculate a dynamic filtering threshold based on the number of connections and distribution density of the nodes, and use the dynamic filtering threshold to filter the edge weights to obtain a voucher association graph; Calculating the degree of association between adjacent nodes in the credential association graph to obtain a node weight coefficient, performing weighted fusion on the business features of the nodes according to the node weight coefficient to obtain a business association feature, combining the business association feature with the credential digital fingerprint to generate node representation information; Arranging the node representation information in chronological order to form a feature sequence, extracting the dependency relationship between the nodes to obtain a transaction chain feature, calculating the similarity between the transaction chain feature and the preset abnormal feature, and marking the continuous transaction sequence with a similarity higher than a first preset threshold as a candidate abnormal transaction chain; The time interval and amount change rate of adjacent transactions in the candidate abnormal transaction chain are calculated, and the correlation between the time interval and amount change rate and the transaction type is used as the transaction correlation degree. The candidate abnormal transaction chain is scored according to the transaction correlation degree, and the candidate abnormal transaction chain with a score exceeding a second preset threshold is determined as an abnormal transaction chain, and an abnormal transaction chain identification result is generated.

4. The method according to claim 1, wherein Based on the abnormal transaction chain identification result and the voucher association graph, converting the original voucher data into financial accounting data includes: Calculate the amount change rate of adjacent nodes in the abnormal transaction chain to obtain an amount change feature, calculate the transaction time difference of adjacent nodes to obtain a time interval feature, and extract the debit and credit account combinations of adjacent nodes to obtain an account combination feature; determine the abnormal propagation direction based on the amount change feature, time interval feature, and account combination feature, and generate an abnormal propagation path; mark the original vouchers corresponding to the associated nodes located on the abnormal propagation path in the voucher association graph as affected vouchers; In the voucher association diagram, select the node sequence without abnormality as the normal transaction chain, extract the matching relationship and corresponding relationship of the debit and credit accounts in the normal transaction chain, and generate the accounting rule template; Based on the connections of the abnormal propagation paths, adjacent affected vouchers are grouped into the same accounting unit. Within each accounting unit, the debit and credit accounts of the affected vouchers are paired and verified according to the accounting rule template. The accounting results are compared with historical normal data to identify accounting deviations. Based on the business association strength between nodes in the voucher association graph, subgraphs are divided, and parallel accounting processing is performed on the original vouchers in each subgraph to generate subgraph accounting results; the subgraph accounting results are merged according to the association relationship between subgraphs to generate financial accounting data.

5. The method according to claim 1, wherein Acquire real-time tax rule data, compare financial accounting data with tax rule data, calculate the logical correlation between each data item in the financial accounting data, perform forward traversal and reverse verification, identify data items that violate tax rules, and assign risk weight values ​​to data items that violate tax rules, including: Obtain tax rule data through the rule data interface, extract the judgment conditions and constraints in the tax rule data, construct a rule description vector, calculate the similarity between the rule description vector and the rule vector in the local rule library, identify new and changed rule content, perform structural processing, and convert it into a rule expression containing trigger conditions, judgment thresholds, and constraint parameters; The data items in the financial accounting data are constructed into an association graph, with the data items as nodes and the logical dependencies between the data items as edges. The numerical association degree and the subject association degree between adjacent nodes are calculated, and the weighted values ​​of the two association degrees are used as edge weights. Based on the edge weights, the shortest path between the nodes is calculated to obtain the association strength of the data items. Traverse the nodes in the association graph in the order of business processes, compare the attribute values ​​of data items with the rule expressions, and mark data items as violation nodes when they do not meet the rule expressions; Starting from each violation node, trace back along the association graph to identify the associated nodes that affect the violation node; mark the path from the associated node to the violation node as the violation propagation path; calculate the impact range of the node based on the violation propagation path; A basic score is determined based on the number of rule violations, a propagation coefficient is calculated based on the scope of influence, and an aging coefficient is calculated based on the update time of the rule expression; the product of the basic score, the propagation coefficient and the aging coefficient is used as the risk weight value.

6. The method according to claim 1, characterized in that Based on the risk weight values, a data balance constraint matrix is ​​constructed. The degree of deviation of data items that violate tax rules in the data balance constraint matrix is ​​analyzed. Data items with the same degree of deviation are grouped into data correction clusters. Grouped iterative corrections are performed on the data correction clusters. When the correction results meet the balance constraints, the tax declaration data generated includes: A self-organizing multi-state equilibrium network is constructed based on risk weight values. The initial state of each data item is determined by the risk weight value. Data items influence each other through state transition rules. The state transition rules include two modes: weight decay and weight enhancement, generating a state evolution sequence of data items. Identifying key state transition points based on the state evolution sequence, wherein the key state transition points reflect the moment when the state of the data item suddenly changes, performing comparative analysis on the data items before and after the key state transition points, and extracting change characteristics of the data items; Calculating a state change value of the data item according to the change characteristics, dividing the data items whose state change values ​​are in the same preset interval into the same data correction cluster, and calculating an ideal state value of the data item in each data correction cluster based on the state transition rule; Data correction is performed alternately based on the weight decay and weight enhancement modes in the state transition rules, and the degree of difference between the data items after each round of correction and the ideal state value is calculated. When the degree of difference is less than the preset difference threshold, the data correction cluster is determined to have reached a stable state. When all data correction clusters have reached a stable state, tax return data is generated.

7. The method according to claim 6, characterized in that Data correction is performed alternately based on the weight decay and weight enhancement modes in the state transition rule. The difference between the data item after each round of correction and the ideal state value is calculated. When the difference is less than the preset difference threshold, it is determined that the data correction cluster has reached a stable state. The following includes: Obtaining a difference sequence between a current state value and an ideal state value of a data item, establishing a state offset curve based on the difference sequence, and calculating a fluctuation trend and a fluctuation amplitude of the state offset curve; Alternately switching between a weight decay mode and a weight enhancement mode based on the positive and negative changes in the difference sequence, generating an adaptive adjustment coefficient based on the fluctuation trend and the current correction round, executing the weight decay mode when the difference sequence is positive, and executing the weight enhancement mode when the difference sequence is negative, dynamically updating the current state value of the data item based on the adaptive adjustment coefficient, updating the current state value and recalculating the difference sequence after each round of execution; Among them, the weight decay mode realizes downward adjustment of data items based on a multi-level decreasing function, and the weight enhancement mode realizes upward adjustment of data items based on a piecewise enhancement function; The degree of difference of all data items in the data correction cluster is calculated based on the current state value and the ideal state value, a stability evaluation index is generated according to the degree of difference, and the trend of change of the stability evaluation index over multiple consecutive rounds is analyzed. When the degree of difference is less than a preset difference threshold and is maintained for a specified number of rounds, it is determined that the data correction cluster has reached a stable state.

8. An automatic verification and intelligent error correction system for corporate tax declaration data, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is configured to receive original voucher data input by the enterprise end and perform feature extraction, construct a feature vector, generate a voucher digital fingerprint based on the feature vector, establish a voucher association graph based on the voucher digital fingerprint, analyze the business association relationships between the vouchers in the voucher association graph, generate an abnormal transaction chain identification result, and convert the original voucher data into financial accounting data based on the abnormal transaction chain identification result and the voucher association graph; The second unit is used to obtain real-time tax rule data, compare financial accounting data with the tax rule data, calculate the logical correlation between each data item in the financial accounting data, perform forward traversal and reverse verification, identify data items that violate tax rules, and assign risk weight values ​​to the data items that violate tax rules; The third unit is used to construct a data balance constraint matrix based on the risk weight value, analyze the degree of deviation of data items that violate tax rules in the data balance constraint matrix, group data items with the same degree of deviation into data correction clusters, perform group iterative correction on the data correction clusters, and generate tax declaration data when the correction results meet the balance constraints.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images