A financial management system and a method for automatically generating financial data

Abstract

This invention provides a financial management system and an automatic financial data generation method, belonging to the field of financial management and generation technology. The system includes modules for financial semantic parsing, intelligent model training, content processing, intensity determination, and financial data generation. The method extracts the semantics and features of multi-dimensional historical financial information and labels them, constructing a financially adapted convolutional neural network model to generate a financial discrimination model. Based on historical data entropy and business semantic processing of the financial business sequence to be processed, it dynamically adjusts the management intensity by combining feature clustering, coding efficiency, and risk level. Finally, it integrates the management intensity, model discrimination process, and feature processing results to automatically generate financial data that complies with accounting standards. This invention solves the problems of low efficiency, poor model adaptation, and fixed control in traditional systems, improving the automation and compliance level of financial data generation and adapting to the needs of enterprises in multiple industries.

Description

Technical Field

[0001] This invention relates to the field of financial management and data generation technology, and in particular to a financial management system and a method for automatically generating financial data. Background Technology

[0002] In practical applications of corporate financial management, the processing and automatic generation of financial data are core aspects of financial work. Traditional financial management systems rely heavily on manual extraction of business features and labeling of compliance data. This is not only inefficient but also prone to incomplete feature extraction and incorrect labeling due to subjective human judgment, making it difficult to adapt to the multi-dimensional and multi-type historical financial information processing needs of enterprises. Regarding financial model training, traditional systems' neural network models are not specifically adapted to financial scenarios, using only general models for training. This results in the model's inability to accurately learn the mapping relationship between business data and financial compliance rules and automatic accounting logic, leading to insufficient model accuracy and adaptability.

[0003] Traditional methods for processing financial transaction sequences lack scientific coding and segmentation approaches. They fail to incorporate historical data entropy for calculating optimal continuous segment lengths and simply truncate or pad feature strings, easily leading to loss of business semantics or feature redundancy. This fails to guarantee the completeness and accuracy of the financial transaction sequence's feature representation. Furthermore, traditional systems often employ fixed control standards for setting the management intensity of financial transactions, unable to dynamically adjust based on feature redundancy, coding efficiency, and risk levels. This results in a mismatch between control intensity and actual business risk, leading to overly strict control for low-risk transactions and insufficient control for high-risk transactions.

[0004] In the process of automatically generating financial data, traditional systems rely solely on model judgment results to generate data without taking into account the dynamic adjustment of management intensity and feature processing results for comprehensive judgment. Furthermore, they lack quantitative calculation of confidence metrics for model output paths and rule inference paths. As a result, the generated financial accounting data is prone to problems such as inconsistency with accounting standards and mismatch with actual business semantics, failing to meet the needs of enterprises for automated and compliant generation of financial data.

[0005] Therefore, this invention proposes a financial management system and a method for automatically generating financial data. Summary of the Invention

[0006] This invention provides a financial management system and a method for automatically generating financial data to solve the aforementioned technical problems.

[0007] This invention provides a financial management system, comprising:

[0008] The financial semantic parsing module is used to extract the core business semantic information of multi-dimensional historical financial information and associate it with the source of traceable data to generate a structured semantic element set. At the same time, it extracts financial field features and accounting element features from the multi-dimensional historical financial information, and combines them with a preset accounting standard rule base to label the financial field features and accounting element features with financial compliance labels and accounting processing mapping labels respectively.

[0009] The intelligent model training module is used to build a deep learning framework and configure a convolutional neural network model adapted to financial scenarios. The financial field features, accounting element features and corresponding labels are input into the convolutional neural network model for iterative training. The model learns the mapping relationship between business data and financial compliance rules and automatic accounting logic, and generates a financial discrimination model.

[0010] The content processing module is used to perform continuous feature string encoding on the financial business sequence to be processed, and calculate the m optimal continuous segment lengths based on historical data entropy as the benchmark continuous segment lengths. It performs feature restoration decoding on the content of continuous segments whose length is less than all other benchmark continuous segment lengths in the continuous feature string encoding result, and performs segmentation and splitting processing based on business semantics on the content of continuous segments whose length is greater than any benchmark continuous segment length.

[0011] The intensity determination module is used to update the continuous feature string encoding sequence based on the feature restoration decoding result and the segmentation and splitting processing result to obtain multiple sets of candidate encoding sequences, and dynamically adjust the management intensity of the financial business sequence to be processed based on the multiple sets of candidate encoding sequences to eliminate redundant features based on feature clustering, optimize encoding efficiency based on entropy value, and adjust the risk level output by the financial discrimination model.

[0012] The financial data generation module is used to automatically generate financial accounting data that meets the requirements of accounting standards by invoking the discrimination process of the financial discrimination model based on management intensity and combining the feature restoration and decoding results of the content processing module.

[0013] Preferably, the financial semantic parsing module is configured with a correction network, including:

[0014] The offset calculation unit is used to set the rule conflict feature vector and the structured semantic element vector corresponding to the current label to be corrected, and calculates the correction offset based on the preset weight matrix, bias vector and activation function.

[0015] The correction vector unit is used to correct the structured semantic element set by taking the original label vector as the basis, combining the learning rate factor dynamically adjusted according to the conflict confidence, and superimposing the product of the learning rate factor and the correction offset onto the original label vector.

[0016] Preferably, the content processing module includes:

[0017] Define a unit, used to set the data points of the historical financial transaction sequence as... Where N is the total number of data points, and the data points are located between positions i and j. This forms a continuous segment S(i,j), and the internal entropy of the continuous segment S(i,j) is defined as H(i,j), with information density weights. ;

[0018] Finding cells are used to find the optimal sequence of split points. ,satisfy , and Let m be the preset number of segments, and define the total cost function F as: ,in, This is the inter-segment segmentation penalty coefficient;

[0019] The solution element is used to minimize the total cost function F, with the following objective: ,in, The optimal segmentation point sequence directly determines the lengths of the m optimal continuous segments. .

[0020] Preferably, the content processing module further includes:

[0021] The probability judgment unit is used to analyze the financial business segment. A sliding window with a fixed length of L is set at position i. Based on the sliding window, the semantic coherence score of the corresponding position i is calculated to determine the probability that position i can become a semantic boundary. The probability that position i is a semantic boundary is positively correlated with the absolute value of the difference in coherence score and negatively correlated with the sum of local standard deviation and smoothing constant.

[0022] In the process of recursively segmenting the financial business sequence, points with a probability exceeding a preset threshold and a corresponding probability at position i being a local maximum are selected as semantic splitting points of the historical financial business sequence, thereby obtaining N data points, where N is a variable.

[0023] Preferably, the intensity determination module includes:

[0024] The splicing and fusion unit is used to perform feature enhancement processing on the feature restoration and decoding results for financial business semantics, supplement the decoding results with feature identifiers related to accounting elements, and perform multi-scale encoding length adaptation on the segmentation and splitting processing results, so that the length of each segmented encoding segment forms a gradient match with the length of the baseline continuous segment, and splicing and fusing the feature-enhanced restoration and decoding encoding segments with the length-adapted splitting processing encoding segments in the financial business dimension.

[0025] The sequence generation unit is used to combine the dimensional consistency check of financial field features to achieve compliant splicing of encoded fragments, and to perform differentiated feature weight allocation and encoding order adjustment on the fused encoded sequence, generating multiple sets of candidate encoded sequences that retain the core accounting element features and complete business semantic information of the financial business sequence to be processed, and the feature dimensions and encoding rules of each set of candidate encoded sequences are adapted to the input requirements of the financial discrimination model.

[0026] Preferably, the intensity determination module further includes:

[0027] A redundancy elimination unit is used to determine the feature vector of each candidate coding sequence and calculate the weighted cluster similarity of any two feature vectors. When the weighted cluster similarity is greater than a preset similarity, it is determined that the two feature vectors are fused and fused to obtain the coding sequence after redundancy removal.

[0028] Calculate the coding efficiency of the deredundant coded sequence set based on information entropy value. ;

[0029] The financial discrimination model is invoked to output the risk level of the financial business sequence to be processed. The clustering effectiveness coefficient after redundancy removal is fused. Coding efficiency Determine the intensity of management ;

[0030] ,in, , For the profile coefficient, , These are the maximum and minimum values ​​of the profile coefficient, respectively. , , These are the weighting coefficients; The feature vector of the encoded sequence after removing redundancy. ,in, The maximum theoretical information entropy of financial characteristics; for Information entropy; for The quantity.

[0031] Preferably, the financial data generation module includes:

[0032] The first calculation unit is used to determine the intermediate results generated by the output path of the financial discrimination model and to calculate the confidence level. ;

[0033] ,in, Maximize the score of the financial discrimination model's output; This refers to the management intensity level coefficient. For the complexity factor of financial transactions;

[0034] The second calculation unit is used to determine the intermediate results generated by the rule-based reasoning path of the financial discrimination model, and its confidence level is [value missing]. ;

[0035] ,in, For rule coverage; Score the semantic matching.

[0036] Among them, confidence level With confidence level As a parameter of the discrimination result in the discrimination process.

[0037] This invention provides a method for automatically generating financial data in a financial management system, comprising:

[0038] Step 1: Extract the core business semantic information of multi-dimensional historical financial information and associate it with the source of traceability markers to generate a structured semantic element set. At the same time, extract financial field features and accounting element features from the multi-dimensional historical financial information, and combine them with the preset accounting standard rule base to mark the financial field features and accounting element features with financial compliance labels and accounting processing mapping labels respectively.

[0039] Step 2: Construct a deep learning framework and configure a convolutional neural network model adapted to the financial scenario. Input the financial field features, accounting element features and corresponding labels into the convolutional neural network model for iterative training, learn the mapping relationship between business data and financial compliance rules and automatic accounting logic, and generate a financial discrimination model.

[0040] Step 3: Perform continuous feature string encoding on the financial business sequence to be processed, and calculate the lengths of m optimal continuous segments based on historical data entropy as the benchmark continuous segment lengths. Perform feature restoration decoding on the content of continuous segments whose length is less than all other benchmark continuous segment lengths in the continuous feature string encoding result, and perform segmentation and splitting processing based on business semantics on the content of continuous segments whose length is greater than any benchmark continuous segment length.

[0041] Step 4: Based on the feature restoration decoding results and the segmentation and splitting processing results, update the continuous feature string encoding sequence to obtain multiple sets of candidate encoding sequences. Then, based on feature clustering to eliminate redundant features, based on entropy to optimize encoding efficiency, and based on the risk level output by the financial discrimination model, dynamically adjust the management intensity of the financial business sequence to be processed.

[0042] Step 5: Based on the management intensity, invoke the discrimination process of the financial discrimination model, and combine the feature restoration and decoding results to automatically generate financial accounting data that meets the requirements of accounting standards.

[0043] Compared with the prior art, the beneficial effects of this application are as follows:

[0044] Through the collaborative work of five major functional modules, accurate analysis and full-chain traceability of multi-dimensional historical financial information are achieved. A convolutional neural network model adapted to financial scenarios is constructed and trained to obtain a high-precision financial discrimination model. Based on historical data entropy, scientific encoding and segmentation of financial business sequences are realized. By combining feature redundancy, encoding efficiency and business risk, dynamic adjustment of management intensity is realized. Finally, financial accounting data that meets the requirements of accounting standards is generated. This effectively solves the problems of low processing efficiency, poor model adaptability, unscientific processing of business sequences, fixed management intensity and insufficient compliance of data generation in traditional financial management systems, and significantly improves the automation, intelligence and compliance level of enterprise financial management.

[0045] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0047] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0048] Figure 1 This is a structural diagram of a financial management system according to an embodiment of the present invention;

[0049] Figure 2 This is a flowchart of a method for automatically generating financial data in a financial management system according to an embodiment of the present invention. Detailed Implementation

[0050] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0051] This invention provides a financial management system, such as... Figure 1 As shown, it includes:

[0052] The financial semantic parsing module is used to extract the core business semantic information of multi-dimensional historical financial information and associate it with the source of traceable data to generate a structured semantic element set. At the same time, it extracts financial field features and accounting element features from the multi-dimensional historical financial information, and combines them with a preset accounting standard rule base to label the financial field features and accounting element features with financial compliance labels and accounting processing mapping labels respectively.

[0053] The intelligent model training module is used to build a deep learning framework and configure a convolutional neural network model adapted to financial scenarios. The financial field features, accounting element features and corresponding labels are input into the convolutional neural network model for iterative training. The model learns the mapping relationship between business data and financial compliance rules and automatic accounting logic, and generates a financial discrimination model.

[0054] The content processing module is used to perform continuous feature string encoding on the financial business sequence to be processed, and calculate the m optimal continuous segment lengths based on historical data entropy as the benchmark continuous segment lengths. It performs feature restoration decoding on the content of continuous segments whose length is less than all other benchmark continuous segment lengths in the continuous feature string encoding result, and performs segmentation and splitting processing based on business semantics on the content of continuous segments whose length is greater than any benchmark continuous segment length.

[0055] The intensity determination module is used to update the continuous feature string encoding sequence based on the feature restoration decoding result and the segmentation and splitting processing result to obtain multiple sets of candidate encoding sequences, and dynamically adjust the management intensity of the financial business sequence to be processed based on the multiple sets of candidate encoding sequences to eliminate redundant features based on feature clustering, optimize encoding efficiency based on entropy value, and adjust the risk level output by the financial discrimination model.

[0056] The financial data generation module is used to automatically generate financial accounting data that meets the requirements of accounting standards by invoking the discrimination process of the financial discrimination model based on management intensity and combining the feature restoration and decoding results of the content processing module.

[0057] In this embodiment, multi-dimensional historical financial information refers to all financially related information generated by an enterprise during its operations, covering different dimensions and time periods. Dimensions include time, business type, department, financial account, and capital, while information types include financial vouchers, accounting books, financial statements, business documents, cash flow records, and tax information. For example, the monthly profit and loss statements, VAT invoices for raw material purchases, bank statements of sales revenue from the sales department, expense reimbursement forms from the R&D department, and corporate income tax returns of an electronics manufacturing company from 2022 to 2024 together constitute the company's multi-dimensional historical financial information.

[0058] Core business semantic information is the core semantic content extracted from multi-dimensional historical financial information after stripping away non-critical and redundant information. It can represent the essence of financial business and reflects the core elements such as the subject of the financial business, business type, transaction amount, time of occurrence, accounting attribution, and fund flow.

[0059] The source of traceability is marked by adding a unique identifier to the extracted core business semantic information. This identifier enables full-chain traceability of financial information and includes key information such as the storage medium, storage location, entry time, entry personnel, and original number of the original financial information.

[0060] A structured semantic element set is a collection of elements formed by organizing the extracted core business semantic information according to a pre-defined financial business structured framework. The framework includes elements such as business entity, business type, transaction object, amount, tax amount, payment method, accounting subject, and occurrence time. Each element corresponds to unique semantic content. For example, the structured semantic element set for R&D equipment procurement is: {Business entity: R&D department, Business type: Equipment procurement, Transaction object: Supplier B, Amount excluding tax: 200,000 yuan, Input tax: 26,000 yuan, Payment method: Bank transfer, Primary account 1: Fixed assets, Primary account 2: Accounts payable, Occurrence time: 2024-06-15}.

[0061] Financial field features are characteristic information extracted from various financial fields of multi-dimensional historical financial information that can characterize the field attributes. Financial fields include amount fields, tax fields, account fields, time fields, invoice number fields, etc. The characteristic information includes the data type of the field, the range of values, the field relationship, the field compliance requirements, etc.

[0062] Accounting element characteristics are feature information extracted from the accounting processing elements of financial transactions that can characterize the attributes of accounting processing. Accounting processing elements include accounting subjects, posting direction, debit and credit amounts, accounting items, accounting processing methods, etc., and feature information includes the category of elements, the correspondence between elements, and the posting rules of elements.

[0063] The Accounting Standards and Rules Repository is a pre-built database that stores various accounting standards and financial accounting rules, including specific standards, tax accounting rules such as value-added tax and corporate income tax, and internal financial management systems of enterprises. The repository supports real-time updates and rule retrieval.

[0064] Financial compliance labels are labels that are combined with the accounting standards rule base to mark the characteristics of financial fields and are used to characterize whether the financial fields meet the compliance requirements. The labels are divided into three categories: "compliant", "pending verification" and "non-compliant". The labeling is based on the degree of matching between the characteristics of the financial fields and the corresponding rules in the accounting standards rule base.

[0065] The accounting processing mapping label is a label that combines the accounting standards rule base to mark the characteristics of accounting elements and represents the standard accounting processing method corresponding to the accounting elements. The label contains mapping information such as accounting subject, posting direction, and accounting processing flow, which provides a basis for subsequent automatic bookkeeping.

[0066] The deep learning framework is a basic framework for model training based on deep learning algorithms. Mature deep learning frameworks such as TensorFlow and PyTorch are selected. This invention customizes the financial scenario on the basis of the general framework, adds dedicated network layers such as financial feature processing layer and rule matching layer, selects PyTorch framework as the basic framework, removes the image feature processing layer that is not related to the financial scenario, and adds financial field feature extraction layer, accounting standard rule matching layer and accounting processing logic mapping layer to build a deep learning framework adapted to this invention.

[0067] The convolutional neural network (CNN) model adapted for financial scenarios is a model that optimizes the structure of a CNN based on the characteristics and needs of financial business. By adjusting the convolutional kernel size, increasing the pooling layer type, and adding fully connected layer dimensions, the model can accurately learn the correlation between financial field features, accounting element features, compliance rules, and accounting logic. The convolutional kernel size is adapted to the dimension of the financial feature vector, and the pooling layer uses a combination of mean pooling and max pooling. The constructed financial scenario-adapted CNN model includes an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The input layer has a dimension of 256 (adapting to the dimension of the financial feature vector), the convolutional layer has three convolutional kernels with sizes of 3×3, 5×5, and 7×7, the pooling layer uses a hybrid pooling method of mean pooling and max pooling, there are two fully connected layers, and the output layer outputs the financial compliance judgment results and accounting processing mapping results.

[0068] Iterative training involves inputting financial field features, accounting element features, and corresponding labels into a convolutional neural network model. The model output is then calculated via forward propagation, compared to the true labels, and the loss value is calculated. The model's weights and biases are then adjusted via backpropagation. This process is repeated until the model loss value falls below a preset threshold and the model accuracy reaches a preset requirement. For example, if the features and labels of 100,000 historical financial data points of an enterprise are input into the model for iterative training, with the cross-entropy loss function, the Adam optimizer, a learning rate of 0.001, and a batch size of 32, after 200 iterations, the model loss value drops below 0.02, and the compliance discrimination accuracy reaches 98.5%, at which point training is stopped.

[0069] The mapping relationship between business data and financial compliance rules and automatic accounting logic is the correspondence between the characteristic information of financial business data and the compliance rules in accounting standards and the automatic accounting logic of enterprises. This relationship is the core content learned by the model through iterative training, which is manifested as specific financial characteristics corresponding to specific compliance rules and accounting methods.

[0070] The financial discrimination model, obtained through iterative training, is a model capable of discriminating the compliance of financial transactions and mapping accounting logic. This model receives feature vectors of the financial transactions to be processed and outputs the compliance level, risk level, and corresponding standard accounting treatment method. For example, after training, the financial discrimination model, when inputting the financial feature vector of R&D equipment procurement, can output the discrimination result: "Compliance Level: A (Compliant), Risk Level: Low Risk, Accounting Treatment Method: Debit Fixed Assets - R&D Equipment, Credit Accounts Payable - Supplier".

[0071] A pending financial transaction sequence is a collection of financial transactions for an enterprise, arranged chronologically and logically, that require data processing. It consists of multiple consecutive sub-items, each containing corresponding financial characteristic information. For example, a series of financial transactions that occurred on June 15, 2024, for an enterprise: R&D department purchasing R&D equipment → Production department purchasing raw materials → Sales department collecting sales proceeds → Administration department reimbursing office expenses. This series of transactions, arranged in chronological order, forms the pending financial transaction sequence.

[0072] Continuous feature string encoding is the process of converting the feature information of each financial business sub-item in the financial business sequence to be processed into a computer-recognizable digital code, and then concatenating all codes in the order of the business sequence to form a continuous digital feature string. The encoding method adopts a combination of one-hot encoding and numerical encoding to ensure the complete representation of the feature information. For example, the code for "R&D department purchasing R&D equipment" in the above financial business sequence to be processed is [1,0,0,0,200000,26000,1], and the code for "Production department purchasing raw materials" is [0,1,0,0,50000,6500,1]. After concatenating the codes of all business sub-items, a continuous feature string is formed: [1,0,0,0,200000,26000,1,0,1,0,0,50000,6500,1,……].

[0073] Historical data entropy is the entropy value obtained by calculating the information entropy of the characteristic information of historical financial business sequences. Information entropy is a quantitative indicator that characterizes the degree of information disorder and information richness. Historical data entropy reflects the characteristic distribution pattern and information complexity of historical financial business sequences.

[0074] The m optimal continuous segment lengths are calculated using an entropy optimization algorithm based on historical data entropy, which yields the longest continuous segment lengths that maximize the representation efficiency of financial business sequence feature information and minimize semantic loss. m is a preset positive integer, set according to the complexity of the company's financial business, and is generally between 2 and 5. For example, if m=2 optimal continuous segment lengths are obtained based on historical data entropy, which are 8 and 12 respectively, then the feature representation efficiency is highest when the financial business sequence feature string of the company is segmented into 8 or 12 code bits.

[0075] The baseline continuous segment length is the length of the m optimal continuous segments calculated as the baseline for the content processing module to process the continuous feature string encoding result. It is the basis for subsequent feature restoration decoding and segmentation processing. For example, the calculated lengths of 8 and 12 are the baseline continuous segment lengths, and the continuous feature string encoding result will be processed based on these lengths.

[0076] Feature restoration decoding is the process of restoring the numerical codes of continuous segments whose length is less than the length of all other baseline continuous segments in the continuous feature string encoding result to the corresponding financial business feature information and semantic content. The decoding process is the reverse of the encoding process, ensuring the complete restoration of the semantic content. For example, if the baseline continuous segment lengths are 8 and 12, and there is a continuous segment with a length of 6 in the continuous feature string encoding result, which is less than 8 and 12, feature restoration decoding is performed on the numerical code of this continuous segment [0,1,0,0,50000,6500] to obtain the corresponding semantic content: The production department purchases raw materials, with a tax-exclusive amount of 50,000 yuan and an input tax of 6,500 yuan.

[0077] Segmentation based on business semantics involves dividing a continuous feature string encoding result into segments whose length exceeds that of any baseline continuous segment, according to the semantic relevance of financial business. Each resulting sub-segment is approximately the length of the baseline continuous segment, ensuring the completeness of the business semantics within each sub-segment without semantic fragmentation. For example, if the baseline continuous segment lengths are 8 and 12, and a continuous feature string encoding result contains a segment with a length of 18, which is greater than 8, this segment is segmented based on business semantics into two sub-segments of lengths 10 and 8. Each sub-segment corresponds to complete financial business semantics, without semantic fragmentation.

[0078] The feature restoration decoding result and the segmentation splitting result are two types of results obtained by the content processing module after processing the continuous feature string encoding result. The feature restoration decoding result is the financial business feature information and semantic content after digital encoding restoration, and the segmentation splitting result is multiple semantically complete digital feature string sub-segments obtained after splitting a long continuous segment. For example, the feature restoration decoding result is the semantic content and corresponding feature information of the production department purchasing raw materials, with a tax-exclusive amount of 50,000 yuan and an input tax of 6,500 yuan; the segmentation splitting result is splitting a continuous segment of length 18 into two digital feature string sub-segments of lengths 10 and 8: [1,0,0,0,200000,26000,1,0,1,0] and [0,0,50000,6500,1,0,1,0].

[0079] The continuous feature string encoding sequence is an ordered numerical encoding sequence formed by encoding the financial business sequence to be processed with continuous feature strings. This sequence is arranged according to the order in which the financial business occurs and represents all the feature information of the financial business to be processed. For example, the continuous feature string encoding sequence of the financial business sequence to be processed is [1,0,0,0,200000,26000,1,0,1,0,0,50000,6500,1,0,0,1,100000,13000,0].

[0080] Multiple candidate coding sequences are obtained by re-encoding the feature restoration decoding results and updating the feature string segments of the segmented split processing results to the original continuous feature string coding sequence according to different business orders and feature combination methods. Each sequence retains the core feature information of the financial business sequence to be processed. If the feature restoration decoding result is re-encoded as [0,1,0,0,50000,6500,1], and combined with the two sub-segments of the segmentation and splitting result in different combinations, three sets of candidate encoding sequences are obtained, namely [0,1,0,0,50000,6500,1,1,0,0,0,200000,26000,1,0,1,0], [1,0,0,0,200000,26000,1,0,1,0,0,1,0,0,50000,6500,1], and [1,0,0,0,200000,26000,1,0,1,0,0,50000,6500,1,0,1,0].

[0081] Feature clustering involves using a clustering algorithm to analyze the feature vectors of multiple candidate coding sequences, grouping sequences with high feature similarity into one class to aggregate feature information. The K-means clustering algorithm is selected, and the clustering is based on the similarity of feature vectors. The K-means clustering algorithm is used to cluster the feature vectors of the three candidate coding sequences mentioned above, grouping sequences with a similarity greater than 0.9 into one class, thus achieving feature aggregation.

[0082] Redundant features are feature terms that repeatedly represent the same financial business characteristics in multiple candidate coding sequences. Redundant features reduce coding efficiency and increase the complexity of model processing. For example, if two candidate coding sequences both contain the feature term [0,1,0,0,50000,6500,1], this feature term is a redundant feature.

[0083] Entropy-optimized coding efficiency uses information entropy as a quantitative indicator. It calculates the entropy of the coded sequence after eliminating redundant features, and by adjusting the feature items and length of the coded sequence, the information entropy is brought closer to the optimal value, thereby improving the efficiency of the coded sequence in representing financial business characteristics. Coding efficiency is an indicator of the coded sequence's ability and compactness in representing financial business characteristics. For example, if the information entropy of the coded sequence after eliminating redundant features is 0.72, by adjusting the order of feature items and removing non-critical feature items, the information entropy can be optimized to 0.78, and the coding efficiency can be improved from 85% to 95%.

[0084] The risk level output by the financial discrimination model is the probability level of financial risk occurring in a given financial transaction sequence after its feature vector is input into the model. The risk level is divided into four categories: low risk, medium risk, high risk, and very high risk, determined by the model based on the compliance and feature matching degree of the transaction. For example, if the feature vector of a financial transaction sequence involving the procurement of R&D equipment is input into the financial discrimination model, the model will output a risk level of "low risk".

[0085] The management intensity is dynamically adjusted based on the redundancy elimination results of feature clustering, the coding efficiency optimized by entropy, and the risk level output by the financial discrimination model. A comprehensive calculation yields the management intensity value for the financial business sequence to be processed, categorized into four levels: low, medium, high, and very high. Different management intensity levels are matched to different financial business sequences based on the calculation results, achieving dynamic adaptation of management intensity. For example, a low-risk financial business sequence with 95% coding efficiency and no redundant features is matched with a "low" management intensity; a high-risk financial business sequence with 70% coding efficiency and many redundant features is matched with a "very high" management intensity.

[0086] Management intensity is a dynamically set level indicator by the intensity determination module for the financial business sequence to be processed, representing the strength of financial control. It is an important basis for the generation of financial data. The higher the management intensity, the stricter the compliance requirements for the generation of financial data, and the more verification steps in the generation process. For example, financial business sequences with extremely high management intensity need to undergo 3 compliance verifications during the data generation process; financial business sequences with low management intensity only need 1 compliance verification.

[0087] The discrimination process of a financial discrimination model involves inputting the feature vectors of the financial business sequence to be processed into the model, and then the model performs compliance discrimination, risk level determination, and accounting logic mapping. This process includes steps such as feature vector input, model forward propagation, rule matching, and result output. For example, after the feature vectors of the financial business sequence to be processed are input into the financial discrimination model, the model first performs convolution and pooling processing on the feature vectors, then performs rule matching with the accounting standard rule base, and finally outputs the compliance level, risk level, and accounting logic. This entire process constitutes the discrimination process.

[0088] Financial accounting data that meets accounting standards is generated based on the discrimination process of management intensity and financial discrimination model, combined with the result of feature restoration decoding, and is consistent with the requirements of accounting standards and financial system. Data types include accounting vouchers, journal vouchers, trial balances, etc., and data content includes accounting subjects, debit and credit directions, amounts, summaries, accounting items, etc.

[0089] The beneficial effects of the above technical solution are as follows: through the collaborative work of the five major functional modules, accurate analysis and full-chain traceability of multi-dimensional historical financial information are achieved; a convolutional neural network model adapted to financial scenarios is constructed and trained to obtain a high-precision financial discrimination model; scientific encoding and segmentation processing of financial business sequences are realized based on historical data entropy; dynamic adjustment of management intensity is achieved by combining feature redundancy, encoding efficiency and business risk; and finally, financial accounting data that meets the requirements of accounting standards is generated. This effectively solves the problems of low processing efficiency, poor model adaptability, unscientific processing of business sequences, fixed management intensity and insufficient compliance of data generation in traditional financial management systems, and significantly improves the automation, intelligence and compliance level of enterprise financial management.

[0090] This invention provides a financial management system, wherein the financial semantic parsing module is configured with a correction network, including:

[0091] The offset calculation unit is used to set the rule conflict feature vector and the structured semantic element vector corresponding to the current label to be corrected, and calculates the correction offset based on the preset weight matrix, bias vector and activation function.

[0092] The correction vector unit is used to correct the structured semantic element set by taking the original label vector as the basis, combining the learning rate factor dynamically adjusted according to the conflict confidence, and superimposing the product of the learning rate factor and the correction offset onto the original label vector.

[0093] In this embodiment, the rule conflict feature vector is a numerical vector obtained by vectorizing the feature information of the conflict when the label marked by the financial semantic parsing module conflicts with the rule in the accounting standards rule base. This vector includes feature dimensions such as conflict type, financial account involved in the conflict, accounting standard clause corresponding to the conflict, and importance of the conflict feature. Each dimension corresponds to a unique numerical value. For example, if the accounting standards rule base stipulates that "the input tax on raw material purchases should be included in the taxes payable - VAT payable (input tax)," and if the label of a certain raw material purchase business includes the input tax in the cost of raw materials, the conflict information is vectorized to obtain the rule conflict feature vector. Let the vector dimension be 10, and the value be [1,0,0,3,6,0,2,0,0,1], where 1 represents the account belonging to the conflict, 3 represents the raw material account, 6 represents the taxes payable - VAT payable (input tax), and 2 represents the corresponding clause number.

[0094] The structured semantic element vector corresponding to the label to be corrected is a numerical vector obtained by vectorizing the structured semantic element set corresponding to the label to be corrected that currently has rule conflicts. The vector dimension is consistent with the element items of the structured semantic element set, and each element item corresponds to one dimension of the vector and is assigned a value. For example, the structured semantic element set of the above raw material procurement business is {Business Entity: Production Department, Business Type: Raw Material Procurement, Transaction Object: Supplier C, Amount Excluding Tax: 80,000 RMB, Input Tax: 10,400 RMB, Funding Method: Bank Transfer, First-Level Account 1: Raw Materials, First-Level Account 2: Raw Materials}. Vectorizing it yields a structured semantic element vector with 8 dimensions and values ​​of [2, 1, 5, 80000, 10400, 1, 3, 3], where 2 represents the production department, 1 represents the procurement business, 5 represents Supplier C, and 3 represents the raw material account.

[0095] The preset weight matrix is ​​a pre-defined weight matrix used to calculate the correction offset. The row dimension of the matrix is ​​consistent with the dimension of the rule conflict feature vector, and the column dimension is consistent with the dimension of the structured semantic element vector. Each element in the matrix is ​​a weight value, set according to the importance of the financial feature, with a value range of 0-1. For example, the preset weight matrix is ​​set to a 10-row, 8-column matrix, where the weight value for the subject dimension is 0.8, the weight value for the conflict type dimension is 0.7, the weight value for the amount dimension is 0.5, and the weight value for the other dimensions is 0.3.

[0096] The bias vector is a pre-defined auxiliary vector used to calculate the correction offset. Its dimension is consistent with that of the structured semantic element vector, and the value of each element is a fixed constant set according to the type of financial transaction. For example, if the bias vector is set to an 8-dimensional vector with values ​​[0.1, 0.1, 0.05, 0.05, 0.05, 0.1, 0.2, 0.2], it is suitable for feature calculations in procurement transactions.

[0097] The activation function is used to introduce nonlinear factors and improve the accuracy of the correction offset calculation. This invention selects the ReLU activation function, which can effectively avoid the gradient vanishing problem and is suitable for calculating financial feature vectors. The expression of the ReLU activation function is y=max(0,x). When the input x is a positive number, the output is x; when the input x is a negative number, the output is 0.

[0098] The correction offset is a numerical offset calculated by inputting the rule conflict feature vector and the structured semantic element vector into a preset computational model, and combining it with a preset weight matrix, bias vector, and activation function. This offset is a vector with the same dimension as the structured semantic element vector and is used to correct the original label vector. For example, by combining the rule conflict feature vector, the structured semantic element vector, the preset weight matrix, and the bias vector, and obtaining intermediate results through matrix operations, and then processing them with the ReLU activation function, the correction offset is obtained as [0,0,0,0,0,0,-0.2,0.2].

[0099] The original label vector is a numerical vector obtained by vectorizing the original labels that were initially labeled by the financial semantic parsing module and have rule conflicts. The vector dimension is consistent with the feature terms of the label and represents all the feature information of the original label. For example, the original label of the above raw material procurement business is "Debit: Raw Materials (810400), Credit: Accounts Payable - Supplier C (810400)", which is vectorized to obtain the original label vector [3,810400,7,810400], where 3 represents the raw materials account, 7 represents the accounts payable account, and 810400 is the corresponding amount.

[0100] Conflict confidence is a quantitative indicator that characterizes the degree of certainty of conflict between a label and the accounting standards rule base. Its value ranges from 0 to 1; the closer the value is to 1, the more certain the conflict is, and the higher the necessity for correction. It is calculated by the rule matching algorithm based on the degree of matching of conflict features. For example, the conflict matching degree between the label of the raw material procurement business and the accounting standards rule base is 99%, and the calculated conflict confidence is 0.99.

[0101] The dynamically adjusted learning rate factor is a factor that is dynamically set according to the conflict confidence level and is used to control the correction magnitude of the correction offset on the original label vector. The learning rate factor is positively correlated with the conflict confidence level. The higher the conflict confidence level, the larger the learning rate factor and the greater the correction magnitude. The value range is 0-1.

[0102] The corrected label vector is obtained by multiplying the learning rate factor by the correction offset to obtain the correction adjustment amount, and then superimposing the correction adjustment amount onto the original label vector. This new label vector eliminates rule conflicts and meets the requirements of the accounting standard rule base. For example, multiplying the learning rate factor 0.099 by the correction offset [0,0,0,0,0,0,-0.2,0.2] yields the correction adjustment amount [0,0,0,0,0,0,-0.0198,0.0198]. Superimposing this onto the original label vector [3,810400,7,810400] yields the corrected label vector [3,80000,6,26000,7,826000]. Here, 6 represents the "Taxes Payable - Value Added Tax (Input Tax)" account, with corresponding amounts of 80000, 10400, and 826000, respectively.

[0103] The beneficial effects of the above technical solution are as follows: through the collaborative work of the offset calculation unit and the correction vector unit, the rule conflict between financial labels and the accounting standard rule base can be accurately identified, the correction offset can be quantitatively calculated and the learning rate factor can be dynamically adjusted to scientifically correct the original label vector. This effectively solves the rule conflict problem of label annotation in the financial semantic parsing module, improves the accuracy of the structured semantic element set and label annotation, and provides high-quality feature and label data for subsequent model training and financial data generation.

[0104] This invention provides a financial management system, wherein the content processing module includes:

[0105] Define a unit, used to set the data points of the historical financial transaction sequence as... Where N is the total number of data points, and the data points are located between positions i and j. This forms a continuous segment S(i,j), and the internal entropy of the continuous segment S(i,j) is defined as H(i,j), with information density weights. ;

[0106] Finding cells are used to find the optimal sequence of split points. ,satisfy , and Let m be the preset number of segments, and define the total cost function F as: ,in, This is the inter-segment segmentation penalty coefficient;

[0107] The solution element is used to minimize the total cost function F, with the following objective: ,in, The optimal segmentation point sequence directly determines the lengths of the m optimal continuous segments. .

[0108] In this embodiment, the data points of the historical financial transaction sequence To obtain the smallest unit of financial business representation after discretizing the historical financial business sequence according to the logic and time sequence of business occurrence, each data point It contains single and indivisible financial business characteristic information, where N is the total number of data points obtained after discretizing the historical financial business sequence, and i is the sequence number of the data point.

[0109] The continuous segment S(i,j) is selected from the set of data points in the historical financial transaction sequence, where the sequence number satisfies... The set of continuous data points, where i is the starting data point index and j is the ending data point index, and the continuous segment S(i,j) contains data points from... arrive All continuous data points. For example, selecting i=1 and j=4 yields the continuous segment S(1,4)={ , , , }; Choosing i=5 and j=8, we obtain the continuous segment S(5,8)={ , , , }

[0110] Internal entropy H(i,j) is a quantitative indicator characterizing the degree of information disorder and information richness of data points within a continuous segment S(i,j), and is calculated using the information entropy formula. Using the Shannon information entropy formula, the internal entropy of the continuous segment S(1,4) is calculated to be H(1,4) = 0.82, and the internal entropy of the continuous segment S(5,8) is H(5,8) = 0.35, indicating that the information richness of S(1,4) is higher than that of S(5,8).

[0111] The optimal segmentation sequence is an ordered sequence of segmentation points that divide N data points of a historical financial business sequence into m consecutive segments. Each segmentation point in the sequence corresponds to the index of a data point and is used to define the start and end positions of each consecutive segment. It is the core basis for achieving scientific segmentation of the financial business sequence. For example, if N=10 in the historical financial business sequence and the preset number of segments m=3, the optimal segmentation sequence s={0,3,7,10}, where 0 is the starting virtual segmentation point, 3 and 7 are the actual segmentation points, and 10 is the ending virtual segmentation point, dividing the data points into three consecutive segments: S(0,3), S(3,7), and S(7,10).

[0112] The preset number of segments, m, is a positive integer pre-set number of segments based on the complexity of the company's financial transactions and the volume of historical financial data. It is generally set in conjunction with the average business chain length of the company's daily financial transactions, and the value ranges from 2 to 6. It can be flexibly adjusted according to actual business needs. For example, small and medium-sized manufacturing enterprises have shorter daily financial business chains, so the preset number of segments m=2; large group enterprises have complex financial business chains and many data points, so the preset number of segments m=4.

[0113] The total cost function F is a function used to quantitatively evaluate the merits of financial business sequence segmentation schemes. It consists of the sum of the products of the information density weights and internal entropies of each continuous segment, plus the segmentation penalty term. The smaller the function value, the better the overall effect of the segmentation scheme, ensuring the information representation efficiency of each continuous segment while avoiding over-segmentation.

[0114] Inter-segment splitting penalty coefficient A pre-set penalty coefficient, ranging from 0.5 to 1.3, is used to limit excessive segmentation of financial transaction sequences. For example, when m=2, set =0.5; when m=4, set =1.2.

[0115] The optimization process of minimizing the total cost function F is to select the optimal solution from all candidate segmentation point sequences with the minimum value of the total cost function F as the optimization objective. This is the core optimization criterion for determining the optimal length of the continuous segment, ensuring that the segmented continuous segment achieves the optimal balance in terms of information representation and semantic integrity.

[0116] Optimal split point sequence The segmentation sequence obtained by minimizing the total cost function F is the sequence with the smallest total cost function F among all candidate segmentation sequence, and the segmentation scheme corresponding to this sequence is the optimal segmentation scheme of the historical financial business sequence.

[0117] Optimal continuous segment length For the optimal split point sequence The lengths of each continuous segment, obtained directly from the calculation, are the differences in the indices of two adjacent dividing points. The m dividing point sequences correspond to the lengths of the m optimal continuous segments, which serve as the baseline length for subsequent processing of the financial transaction sequences to be processed. For example, the optimal dividing point sequence... Given the interval {0, 3, 7, 10}, calculate the differences between adjacent dividing points: 3-0=3, 7-3=4, and 10-7=3 to obtain the three optimal continuous segment lengths. ={3,4,3}, this set of values ​​is the baseline continuous segment length of the content processing module.

[0118] The beneficial effects of the above technical solution are as follows: By configuring definition units, search units, and solution units for the content processing module, the standardized definition of historical financial business sequence data points and the quantitative calculation of information entropy-related indicators are realized. A total cost function that takes into account both information representation efficiency and segmentation rationality is constructed. The optimal segmentation point sequence and the optimal continuous segment length are obtained by solving the dynamic programming algorithm. This provides a scientific and quantitative benchmark for the encoding and segmentation of the financial business sequence to be processed, avoids the subjectivity and arbitrariness of traditional segmentation methods, ensures the accuracy and consistency of financial business sequence segmentation processing, and effectively improves the processing accuracy of the content processing module.

[0119] This invention provides a financial management system, wherein the content processing module further includes:

[0120] The probability judgment unit is used to analyze the financial business segment. A sliding window with a fixed length of L is set at position i. Based on the sliding window, the semantic coherence score of the corresponding position i is calculated to determine the probability that position i can become a semantic boundary. The probability that position i is a semantic boundary is positively correlated with the absolute value of the difference in coherence score and negatively correlated with the sum of local standard deviation and smoothing constant.

[0121] In the process of recursively segmenting the financial business sequence, points with a probability exceeding a preset threshold and corresponding probability being a local maximum at position i are selected as semantic splitting points of the historical financial business sequence, thereby obtaining N data points, where N is a variable.

[0122] In this embodiment, the financial business segment to be analyzed is a number of financial business fragments obtained after preliminary segmentation of the historical financial business sequence. It is the basic analysis unit for semantic boundary judgment. Each segment contains several continuous financial data points and retains some independent business semantics. For example, the historical financial business sequence of N=10 is initially divided into two segments [0,5] and [5,10], and each segment is the financial business segment to be analyzed.

[0123] A sliding window of fixed length L is a window tool used to extract local data features from financial business segments. The window length L is a pre-set positive integer, and ,in, This is the rounding function.

[0124] In this embodiment, semantic coherence score = ,in, The feature weights for the k-th data point are: 0.3 for subject dimension, 0.3 for amount dimension, 0.2 for business type dimension, 0.1 for payment method dimension, and 0.1 for time dimension. Let L be the cosine similarity of the feature vectors of the k-th and k+1-th data points; L is the sliding window length (even number). If i is the boundary position, the mean of the valid data points within the window is calculated.

[0125] The probability of a semantic boundary is the probability value that a certain position i in the financial business sub-segment becomes the semantic boundary of different business segments. The value ranges from 0 to 1. The closer the value is to 1, the more suitable the position is as a semantic boundary, splitting the sub-segment into two semantically independent continuous segments.

[0126] The absolute value of the coherence score difference is the absolute value obtained by taking the difference between the semantic coherence scores of the sliding windows on the left and right sides of a certain position i. For example, if the left window score is 0.95 and the right window score is 0.32 at position i=3, the absolute value of the difference is |0.95-0.32|=0.63.

[0127] The local standard deviation is the standard deviation calculated by taking the semantic coherence scores of the sliding window within a preset range centered at position i. For example, taking the scores of 5 adjacent windows centered at position i=3 {0.92, 0.95, 0.32, 0.35, 0.38}, the local standard deviation is 0.31.

[0128] The smoothing constant is a pre-set constant used to avoid meaningless calculations when the local standard deviation is 0. It is a minimum value greater than 0 and is generally set to 0.001.

[0129] The recursive segmentation is an iterative segmentation process that first initially segments the financial business sequence, then performs semantic boundary judgment and splitting on each segment. If the split sub-segments still have cross-semantic situations, then the semantic boundary judgment and splitting are continued until all sub-segments are semantically complete independent segments.

[0130] The preset threshold is a pre-defined critical value for determining the semantic boundary probability, ranging from 0.5 to 0.8. When the semantic boundary probability of a certain position is greater than this value, it is included in the semantic splitting point candidate set. For example, if the preset threshold is set to 0.6, the semantic boundary probability of position i=3 is 0.85, so it is included in the candidate set; the probability of position i=2 is 0.45, so it is removed.

[0131] A local maximum is a position in the sequence of positions of a financial business segment where the probability of the semantic boundary at a certain position is greater than the probability of its left and right adjacent positions. This position is the local maximum point, representing the most suitable semantic boundary in the surrounding area.

[0132] Semantic split points are selected from the candidate set that simultaneously satisfy the condition that the probability exceeds a preset threshold and is a local maximum. They are the final semantic segmentation points of the historical financial business sequence, used to determine the division boundary of data points, and the total number of data points N will dynamically change with the number of semantic split points. For example, if positions 3 and 7 are selected from the candidate set as local maxima with a probability exceeding the threshold, these two positions are semantic split points, and the original sequence of N=10 is split into 3 semantic segments. If additional subdivided data points are added after the split, N is updated to 12.

[0133] In this embodiment, the semantic boundary probability = the absolute value of the coherence score difference / (local standard deviation + smoothing constant).

[0134] The beneficial effects of the above technical solution are as follows: by combining a sliding window with semantic coherence scoring, the semantic boundary of the financial business sequence is quantitatively determined; the introduction of recursive segmentation ensures that all segmented sub-segments have complete business semantics; at the same time, the total number of data points N is dynamically adjusted, which solves the semantic fragmentation problem caused by fixed data point division in traditional segmentation methods, and makes the calculated optimal continuous segment length more in line with the actual semantic features of financial business, further improving the accuracy and fit of the content processing module in processing financial business sequences.

[0135] This invention provides a financial management system, wherein the strength determination module includes:

[0136] The splicing and fusion unit is used to perform feature enhancement processing on the feature restoration and decoding results for financial business semantics, supplement the decoding results with feature identifiers related to accounting elements, and perform multi-scale encoding length adaptation on the segmentation and splitting processing results, so that the length of each segmented encoding segment forms a gradient match with the length of the baseline continuous segment, and splicing and fusing the feature-enhanced restoration and decoding encoding segments with the length-adapted splitting processing encoding segments in the financial business dimension.

[0137] The sequence generation unit is used to combine the dimensional consistency check of financial field features to achieve compliant splicing of encoded fragments, and to perform differentiated feature weight allocation and encoding order adjustment on the fused encoded sequence, generating multiple sets of candidate encoded sequences that retain the core accounting element features and complete business semantic information of the financial business sequence to be processed, and the feature dimensions and encoding rules of each set of candidate encoded sequences are adapted to the input requirements of the financial discrimination model.

[0138] In this embodiment, feature enhancement processing is a process of supplementing and strengthening the financial business feature information of the feature restoration decoding result. By adding business type identifiers, financial account association information, compliance verification results, etc., the richness and recognizability of the feature information are improved, making the features more suitable for subsequent model input requirements. For example, feature enhancement is performed on the restoration decoding result of "production department purchasing raw materials" by supplementing feature information such as "business type: raw material purchase, primary account: raw materials, compliance: compliance, tax type: input tax".

[0139] The feature identifier associated with accounting elements is a unique identifier added to the feature restoration decoding result, corresponding one-to-one with the accounting processing element. This identifier contains accounting element information such as accounting subject, posting direction, and accounting item, realizing the precise association between feature information and accounting processing elements. For example, the feature identifier "KM003-J-002" is added to the decoding result of raw material procurement, where KM003 represents the raw material subject, J represents the debit side, and 002 represents the production department accounting item.

[0140] Multi-scale coding length adaptation is a process of adjusting the length of each coding segment resulting from the segmentation process based on different scales of the baseline continuous segment length. The adjustment methods include feature completion (for short coding segments) and feature simplification (for long coding segments). Both completion and simplification are based on business semantics and do not lose core features. For example, if the baseline continuous segment lengths are 8 and 12, and the length of the split coding segment is 10, it will be adapted to the scale of 12 through feature completion; if the length is 6, it will be adapted to the scale of 8 through core feature simplification.

[0141] Gradient matching aims to ensure that the lengths of the split coded segments correspond to the lengths of the baseline continuous segments in a gradient relationship, either from large to small or from small to large. This avoids excessive differences between the coded segment lengths and the baseline length, thus guaranteeing the regularity of the coded sequence. For example, if the baseline continuous segment lengths are 8 and 12, and the lengths of the split coded segments are 7, 9, 11, and 13, then after length adaptation, they are adjusted to 8, 8, 12, and 12, forming a gradient match with the baseline length.

[0142] The fusion of financial transaction dimensions involves combining feature-enhanced restored decoding encoded segments with length-adapted split encoded segments based on dimensions such as the type, timing, and accounting category of the financial transaction. The fusion preserves the semantic relationships between the segments, and the arrangement of the encoded segments conforms to the logical sequence of financial transactions. For example, based on the transaction timing dimension, the restored decoding encoded segment of raw material procurement and the split encoded segment of R&D equipment procurement are combined into [Raw Material Procurement Code, R&D Equipment Procurement Code], which aligns with the actual transaction sequence.

[0143] In this embodiment, dimensionality consistency verification is a process of verifying the feature dimensions of each encoded segment to be spliced, ensuring that the feature dimensions of the spliced ​​encoded sequence are consistent with the input feature dimensions of the financial discrimination model, thus avoiding the model's inability to recognize due to dimensional mismatch. The verification content includes the number of feature items, the data type of the feature items, and the range of feature values. For example, if the input feature dimension of the financial discrimination model is 256 dimensions, the verification of the encoded segments to be spliced ​​will complete the 128-dimensional segment to 256 dimensions and remove invalid feature items that exceed the dimensionality limit.

[0144] Compliance splicing involves splicing coded segments according to the financial and business dimensions after passing the dimensional consistency check. The spliced ​​coded sequence has no dimensional conflicts and no missing feature items, and conforms to the input specifications of the financial discrimination model.

[0145] Differentiated feature weight allocation is the process of assigning different weight values ​​to different feature items in the fused coding sequence based on the importance of financial transactions and the core needs of accounting. Core feature items (such as amount, account, and tax) are assigned high weights, while secondary feature items (such as entry time and entry person) are assigned low weights, making the model focus more on core features. For example, the weight of the amount feature item is assigned 0.3, the account feature item 0.25, the tax feature item 0.2, and the entry person feature item 0.05.

[0146] The encoding order adjustment is a process of adjusting the order of feature items in the fused encoding sequence while preserving the core semantics of financial transactions. The adjustment is based on the input feature order requirements of the financial discrimination model, ensuring that the encoding sequence fully matches the model's input rules. For example, if the input order of the financial discrimination model is [business type, account, amount, tax, payment method], the feature items in the encoding sequence will be adjusted to this order to ensure normal model recognition.

[0147] The beneficial effects of the above technical solution are as follows: the information richness and regularity of the encoded fragments are improved through feature enhancement and length adaptation; the splicing based on the financial business dimension ensures the coherence of business semantics; and the combination of dimension consistency verification, differentiated weight allocation and encoding order adjustment generates multiple sets of candidate encoding sequences that are fully adapted to the financial discrimination model. This provides high-quality and standardized basic data for the subsequent dynamic adjustment of management intensity and avoids the deviation in management intensity judgment caused by non-standard encoding sequences.

[0148] This invention provides a financial management system, wherein the strength determination module further includes:

[0149] A redundancy elimination unit is used to determine the feature vector of each candidate coding sequence and calculate the weighted cluster similarity of any two feature vectors. When the weighted cluster similarity is greater than a preset similarity, it is determined that the two feature vectors are fused and fused to obtain the coding sequence after redundancy removal.

[0150] Calculate the coding efficiency of the deredundant coded sequence set based on information entropy value. ;

[0151] The financial discrimination model is invoked to output the risk level of the financial business sequence to be processed. The clustering effectiveness coefficient after redundancy removal is fused. Coding efficiency Determine the intensity of management ;

[0152] ,in, , For the profile coefficient, , These are the maximum and minimum values ​​of the profile coefficient, respectively. , , These are the weighting coefficients; The feature vector of the encoded sequence after removing redundancy. ,in, The maximum theoretical information entropy of financial characteristics; for Information entropy; for The quantity.

[0153] In this embodiment, ,in, for The weighted average distance to all other feature vectors in its cluster is calculated using Euclidean distance combined with financial feature weights. for The weighted average distance to all feature vectors in the nearest neighbor cluster.

[0154] ,in, The weight of the t-th type of financial feature; categorical feature: , The number of value categories for a feature; numerical features: , The standard deviation is the characteristic.

[0155] Preferably, the financial data generation module includes:

[0156] The first calculation unit is used to determine the intermediate results generated by the output path of the financial discrimination model and to calculate the confidence level. ;

[0157] ,in, Maximize the score of the financial discrimination model's output; This refers to the management intensity level coefficient. For the complexity factor of financial transactions;

[0158] The second calculation unit is used to determine the intermediate results generated by the rule-based reasoning path of the financial discrimination model, and its confidence level is [value missing]. ;

[0159] ,in, For rule coverage; Score the semantic matching.

[0160] Among them, confidence level With confidence level As a parameter of the discrimination result in the discrimination process.

[0161] In this embodiment, the value of Le is: when the management intensity level is low, the value of Le is 0.1;

[0162] The management intensity level is medium, and Le is 0.5.

[0163] The management intensity level is high, and Le is 0.9.

[0164] The value ranges from 0 to 1, and is quantified by the number of accounting elements involved in the business / accounting standards and rules;

[0165] The value range is from 0 to 1; Where M3 represents the number of accounting standard rules matched with the pending business; and M4 represents the total number of accounting standard rules that should be matched with the pending business.

[0166] In this embodiment, the feature vector is a numerical vector converted from the numerical encoding of the candidate coding sequence for cluster analysis. The vector dimension is consistent with the feature dimension of the candidate coding sequence. Each feature item corresponds to one dimension of the vector and is assigned a value. This is the core form for quantitatively representing the features of the candidate coding sequence. For example, a 256-dimensional candidate coding sequence is converted into a 256-dimensional feature vector, where the business type dimension is assigned a value of 1, the subject dimension is assigned a value of 3, and the amount dimension is assigned a value of 200000.

[0167] Weighted cluster similarity = W is a feature weight diagonal matrix, where the diagonal elements are the weights of each financial feature dimension. The weight allocation is consistent with the semantic coherence score (subject 0.3, amount 0.3, business type 0.2, funding method 0.1, time 0.1), and the off-diagonal elements are 0. , These are the feature vectors of two candidate encoded sequences.

[0168] The preset similarity is a pre-defined threshold for redundancy feature detection, ranging from 0.8 to 0.9. Where D is the number of financial feature dimensions. When the weighted cluster similarity of two feature vectors is greater than this value, it is determined that there is redundancy between the two and they need to be merged and deduplicated.

[0169] The clustering effectiveness coefficient is an indicator used to quantitatively evaluate the clustering effect of features. It is calculated from the silhouette coefficient and ranges from 0 to 1. The closer the value is to 1, the better the clustering effect and the higher the feature discrimination of the coded sequence after redundancy removal.

[0170] If the silhouette coefficient of the feature vector of the coded sequence after redundancy removal is 0.85, =0.9, =0.2, calculated as follows =(0.85-0.2) / (0.9-0.2)=0.93.

[0171] Weighting coefficient , , The pre-defined coefficients used to adjust the impact of clustering effectiveness, coding efficiency, and risk level on management intensity Q are all positive numbers. It can be adjusted according to the company's financial control priorities. If the company focuses on risk control, it can improve... The value to be set. For example, if a company focuses on risk management, it sets... =0.2、 =0.2、 =0.6; Focusing on encoding efficiency, set =0.2、 =0.6、 =0.2.

[0172] Maximum theoretical information entropy of financial features The theoretical maximum information entropy value, calculated based on the range and distribution patterns of financial characteristics, serves as a benchmark for quantitatively evaluating coding efficiency. It is a fixed constant determined by the type of financial business characteristics of the enterprise. For example, if an enterprise's financial characteristics combine numerical and categorical types, the calculated value would be... =1.2.

[0173] Coding efficiency It is an index that characterizes the efficiency and compactness of the redundancy-removed encoded sequence set in representing financial business characteristics. The value ranges from 0 to 1, and the closer the value is to 1, the better.

[0174] The beneficial effects of the above technical solution are as follows: weighted clustering enables the accurate removal of redundant features from candidate coding sequences; the combination of silhouette coefficient and information entropy enables the quantitative evaluation of clustering effect and coding efficiency; a quantitative calculation model for management intensity that integrates clustering effectiveness, coding efficiency and business risk is constructed, enabling dynamic and scientific adjustment of management intensity; and solving the problem of mismatch between the fixed control intensity of the traditional financial management system and the actual business risk, making the financial control strength more in line with the actual business situation.

[0175] In this embodiment, the output path is a computational path in which the financial discrimination model, based on deep learning algorithms, performs feature extraction, convolution, pooling, and other operations on the input encoded sequence and directly outputs the discrimination result. This path depends on the training results of the model and is suitable for the rapid discrimination of routine financial transactions.

[0176] Confidence This is a quantitative indicator that characterizes the reliability of the discrimination results of the output path of the financial discrimination model. The value ranges from 0 to 1. The closer the value is to 1, the more reliable the discrimination results of the output path are.

[0177] The maximum output score is the largest probability value among all the discrimination results calculated by the softmax function in the output layer of the financial discrimination model. It represents the confidence level of the model in a certain discrimination result, and its value ranges from 0 to 1. For example, if the model outputs the probabilities of compliance, pending verification, and non-compliance as 0.95, 0.03, and 0.02, respectively, the maximum output score is 0.95.

[0178] Management Intensity Level Coefficient The coefficient obtained by classifying the management intensity value Q into levels is assigned the values ​​1, 2, 3, and 4 according to the management intensity levels of low, medium, high, and extremely high. It is the adjustment factor of management intensity on confidence. The higher the management intensity, the greater the adjustment range on the confidence of the output path.

[0179] Financial business complexity factor This is a quantitative indicator characterizing the complexity of a sequence of financial transactions to be processed. It is determined based on the number of financial items involved, transaction stages, and the amount involved. The value is a positive number; the more complex the transaction, the higher the complexity. The larger the value.

[0180] In this embodiment, the rule-based reasoning path is the computational path by which the financial discrimination model performs rule matching and logical reasoning on the input encoded sequence based on a preset accounting standard rule base, and then outputs the discrimination result. This path depends on accounting standard rules and is suitable for the accurate discrimination of special and complex financial transactions.

[0181] Confidence This is a quantitative indicator that characterizes the reliability of the judgment results of the rule-based reasoning path in the financial judgment model. The value ranges from 0 to 1. The closer the value is to 1, the more reliable the judgment results of the rule-based reasoning path are.

[0182] rule coverage To characterize the degree of matching coverage between the characteristics of the financial business sequence to be processed and the relevant rules in the accounting standards rule base, the value ranges from 0 to 1. The closer the value is to 1, the more comprehensive the rule coverage and the more reliable the inference result.

[0183] Semantic matching score To characterize the degree of matching between the business semantics of the financial business sequence to be processed and the semantics of the rule descriptions in the accounting standards rule base, the value ranges from 0 to 1. The closer the value is to 1, the higher the semantic matching degree and the more accurate the reasoning result.

[0184] The beneficial effects of the above technical solution are as follows: By configuring a first calculation unit and a second calculation unit into the financial data generation module, the confidence of the output path and rule reasoning path of the financial discrimination model is quantitatively calculated. The management intensity level coefficient and the financial business complexity factor are introduced to dynamically adjust the confidence level, making the confidence level judgment more in line with the control requirements and complexity of the business. This provides a dual-path credibility reference for financial data generation, solves the problem of low credibility of the traditional single model discrimination results, improves the accuracy and reliability of the financial discrimination process, and lays a solid foundation for the compliant generation of subsequent financial data.

[0185] This invention provides a method for automatically generating financial data in a financial management system, such as... Figure 2 As shown, it includes:

[0186] Step 1: Extract the core business semantic information of multi-dimensional historical financial information and associate it with the source of traceability markers to generate a structured semantic element set. At the same time, extract financial field features and accounting element features from the multi-dimensional historical financial information, and combine them with the preset accounting standard rule base to mark the financial field features and accounting element features with financial compliance labels and accounting processing mapping labels respectively.

[0187] Step 2: Construct a deep learning framework and configure a convolutional neural network model adapted to the financial scenario. Input the financial field features, accounting element features and corresponding labels into the convolutional neural network model for iterative training, learn the mapping relationship between business data and financial compliance rules and automatic accounting logic, and generate a financial discrimination model.

[0188] Step 3: Perform continuous feature string encoding on the financial business sequence to be processed, and calculate the lengths of m optimal continuous segments based on historical data entropy as the benchmark continuous segment lengths. Perform feature restoration decoding on the content of continuous segments whose length is less than all other benchmark continuous segment lengths in the continuous feature string encoding result, and perform segmentation and splitting processing based on business semantics on the content of continuous segments whose length is greater than any benchmark continuous segment length.

[0189] Step 4: Based on the feature restoration decoding results and the segmentation and splitting processing results, update the continuous feature string encoding sequence to obtain multiple sets of candidate encoding sequences. Then, based on feature clustering to eliminate redundant features, based on entropy to optimize encoding efficiency, and based on the risk level output by the financial discrimination model, dynamically adjust the management intensity of the financial business sequence to be processed.

[0190] Step 5: Based on the management intensity, invoke the discrimination process of the financial discrimination model, and combine the feature restoration and decoding results to automatically generate financial accounting data that meets the requirements of accounting standards.

[0191] The beneficial effects of the above technical solution are: it integrates financial semantic parsing, model training, business sequence processing, dynamic adjustment of management intensity, and financial data generation into a standardized and automated process. Each step is progressive and data is linked, realizing the fully automated processing from the extraction of financial business information to the generation of compliant accounting data. This significantly reduces manual operation costs and improves the efficiency and compliance of financial data generation. At the same time, the method is highly flexible and can flexibly adjust parameters according to the financial business characteristics and control priorities of enterprises, adapting to the financial management needs of enterprises of different sizes and industries.

[0192] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A financial management system, characterized in that, include: The financial semantic parsing module is used to extract the core business semantic information of multi-dimensional historical financial information and associate it with the source of traceable data to generate a structured semantic element set. At the same time, it extracts financial field features and accounting element features from the multi-dimensional historical financial information, and combines them with a preset accounting standard rule base to label the financial field features and accounting element features with financial compliance labels and accounting processing mapping labels respectively. The intelligent model training module is used to build a deep learning framework and configure a convolutional neural network model adapted to financial scenarios. The financial field features, accounting element features and corresponding labels are input into the convolutional neural network model for iterative training. The model learns the mapping relationship between business data and financial compliance rules and automatic accounting logic, and generates a financial discrimination model. The content processing module is used to perform continuous feature string encoding on the financial business sequence to be processed, and calculate the m optimal continuous segment lengths based on historical data entropy as the benchmark continuous segment lengths. It performs feature restoration decoding on the content of continuous segments whose length is less than all other benchmark continuous segment lengths in the continuous feature string encoding result, and performs segmentation and splitting processing based on business semantics on the content of continuous segments whose length is greater than any benchmark continuous segment length. The intensity determination module is used to update the continuous feature string encoding sequence based on the feature restoration decoding result and the segmentation and splitting processing result to obtain multiple sets of candidate encoding sequences, and dynamically adjust the management intensity of the financial business sequence to be processed based on the multiple sets of candidate encoding sequences to eliminate redundant features based on feature clustering, optimize encoding efficiency based on entropy value, and adjust the risk level output by the financial discrimination model. The financial data generation module is used to automatically generate financial accounting data that meets the requirements of accounting standards by invoking the discrimination process of the financial discrimination model based on management intensity and combining the feature restoration and decoding results of the content processing module.

2. The financial management system according to claim 1, characterized in that, The financial semantic parsing module is configured with a correction network, including: The offset calculation unit is used to set the rule conflict feature vector and the structured semantic element vector corresponding to the current label to be corrected, and calculates the correction offset based on the preset weight matrix, bias vector and activation function. The correction vector unit is used to correct the structured semantic element set by taking the original label vector as the basis, combining the learning rate factor dynamically adjusted according to the conflict confidence, and superimposing the product of the learning rate factor and the correction offset onto the original label vector.

3. The financial management system according to claim 1, characterized in that, The content processing module includes: Define a unit, used to set the data points of the historical financial transaction sequence as... Where N is the total number of data points, and the data points are located between positions i and j. This forms a continuous segment S(i,j), and the internal entropy of the continuous segment S(i,j) is defined as H(i,j), with information density weights. ; Finding cells are used to find the optimal sequence of split points. ,satisfy , and Let m be the preset number of segments, and define the total cost function F as: ,in, This is the inter-segment segmentation penalty coefficient; The solution element is used to minimize the total cost function F, with the following objective: ,in, The optimal segmentation point sequence directly determines the lengths of the m optimal continuous segments. .

4. The financial management system according to claim 3, characterized in that, The content processing module further includes: The probability judgment unit is used to analyze the financial business segment. A sliding window with a fixed length of L is set at position i. Based on the sliding window, the semantic coherence score of the corresponding position i is calculated to determine the probability that position i can become a semantic boundary. The probability that position i is a semantic boundary is positively correlated with the absolute value of the difference in coherence score and negatively correlated with the sum of local standard deviation and smoothing constant. In the process of recursively segmenting the financial business sequence, points with a probability exceeding a preset threshold and a corresponding probability at position i being a local maximum are selected as semantic splitting points of the historical financial business sequence, thereby obtaining N data points, where N is a variable.

5. The financial management system according to claim 1, characterized in that, The intensity determination module includes: The splicing and fusion unit is used to perform feature enhancement processing on the feature restoration and decoding results for financial business semantics, supplement the decoding results with feature identifiers related to accounting elements, and perform multi-scale encoding length adaptation on the segmentation and splitting processing results, so that the length of each segmented encoding segment forms a gradient match with the length of the baseline continuous segment, and splicing and fusing the feature-enhanced restoration and decoding encoding segments with the length-adapted splitting processing encoding segments in the financial business dimension. The sequence generation unit is used to combine the dimensional consistency check of financial field features to achieve compliant splicing of encoded fragments, and to perform differentiated feature weight allocation and encoding order adjustment on the fused encoded sequence, generating multiple sets of candidate encoded sequences that retain the core accounting element features and complete business semantic information of the financial business sequence to be processed, and the feature dimensions and encoding rules of each set of candidate encoded sequences are adapted to the input requirements of the financial discrimination model.

6. The financial management system according to claim 5, characterized in that, The intensity determination module further includes: A redundancy elimination unit is used to determine the feature vector of each candidate coding sequence and calculate the weighted cluster similarity of any two feature vectors. When the weighted cluster similarity is greater than a preset similarity, it is determined that the two feature vectors are fused and fused to obtain the coding sequence after redundancy removal. Calculate the coding efficiency of the deredundant coded sequence set based on information entropy value. ; The financial discrimination model is invoked to output the risk level of the financial business sequence to be processed. The clustering effectiveness coefficient after redundancy removal is fused. Coding efficiency Determine the intensity of management ; ,in, , For the profile coefficient, , These are the maximum and minimum values ​​of the profile coefficient, respectively. , , These are the weighting coefficients; The feature vector of the encoded sequence after removing redundancy. ,in, The maximum theoretical information entropy of financial characteristics; for Information entropy; for The quantity.

7. The financial management system according to claim 1, characterized in that, The financial data generation module includes: The first calculation unit is used to determine the intermediate results generated by the output path of the financial discrimination model and to calculate the confidence level. ; ,in, Maximize the score of the financial discrimination model's output; This refers to the management intensity level coefficient. For the complexity factor of financial transactions; The second calculation unit is used to determine the intermediate results generated by the rule-based reasoning path of the financial discrimination model, and its confidence level is [value missing]. ; ,in, For rule coverage; Score the semantic matching. Among them, confidence level With confidence level As a parameter of the discrimination result in the discrimination process.

8. A method for automatically generating financial data in a financial management system, characterized in that, include: Step 1: Extract the core business semantic information of multi-dimensional historical financial information and associate it with the source of traceability markers to generate a structured semantic element set. At the same time, extract financial field features and accounting element features from the multi-dimensional historical financial information, and combine them with the preset accounting standard rule base to mark the financial field features and accounting element features with financial compliance labels and accounting processing mapping labels respectively. Step 2: Construct a deep learning framework and configure a convolutional neural network model adapted to the financial scenario. Input the financial field features, accounting element features and corresponding labels into the convolutional neural network model for iterative training, learn the mapping relationship between business data and financial compliance rules and automatic accounting logic, and generate a financial discrimination model. Step 3: Perform continuous feature string encoding on the financial business sequence to be processed, and calculate the lengths of m optimal continuous segments based on historical data entropy as the benchmark continuous segment lengths. Perform feature restoration decoding on the content of continuous segments whose length is less than all other benchmark continuous segment lengths in the continuous feature string encoding result, and perform segmentation and splitting processing based on business semantics on the content of continuous segments whose length is greater than any benchmark continuous segment length. Step 4: Based on the feature restoration decoding results and the segmentation and splitting processing results, update the continuous feature string encoding sequence to obtain multiple sets of candidate encoding sequences. Then, based on feature clustering to eliminate redundant features, based on entropy to optimize encoding efficiency, and based on the risk level output by the financial discrimination model, dynamically adjust the management intensity of the financial business sequence to be processed. Step 5: Based on the management intensity, invoke the discrimination process of the financial discrimination model, and combine the feature restoration and decoding results to automatically generate financial accounting data that meets the requirements of accounting standards.

Images