Enterprise tax data processing method, device and equipment, storage medium and program product

Abstract

The invention discloses an enterprise tax data processing method. The method comprises the steps of obtaining a tax sample of a target enterprise; performing feature extraction on the tax samples to obtain a tax feature set of the target enterprise; according to the tax feature set, first prediction tax data of a target month of the target enterprise is predicted by using a tax data processing model, an encoder and a decoder in the tax data processing model comprise self-attention modules, and the self-attention modules are used for analyzing mutual dependency relationships among different tax feature groups in the tax feature set; obtaining first actual tax data of a target month of the target enterprise; determining a tax difference value according to the first predicted tax data and the first actual tax data; and when the tax difference value satisfies a warning condition, reporting financial risk early warning for the target enterprise. Therefore, the accuracy of financial risk early warning for the enterprise based on the tax data report is effectively improved, and the situations of misjudgment and missed judgment when the financial risk occurs in the enterprise are reduced.

Description

Technical Field

[0001] The present application relates to the financial field, and in particular to a method, device, equipment, storage medium and program product for processing corporate tax data. Background Art

[0002] At present, for the financial services provided by financial institutions to enterprises, customer access management and the determination of customer credit limits are important links before and during financial services. Generally speaking, in credit loan services for enterprises, the tax data of enterprises is an important factor in measuring the credit status and repayment ability of enterprises, and is usually an important basis for whether enterprises pass loan approval and determine the credit limit of enterprises.

[0003] In actual application, the tax data of enterprises is used in the pre-loan access and credit granting process of enterprise credit loan services. However, for the post-loan management of enterprise credit loan services, the tax data of enterprises is often easily overlooked.

[0004] Normally, the account managers of financial institutions can monitor the tax data of enterprises during the post-loan management process. For example, the account managers can conduct physical visits to enterprises to verify their qualifications. However, when enterprises encounter operating risks, it is often difficult for account managers to find changes in the credit status and repayment ability of enterprises from tax data. In addition, due to the complexity of tax data itself, it is very important to accurately determine whether an enterprise has operating risks through tax data. Summary of the invention

[0005] The embodiments of the present application provide a method, apparatus, device, storage medium and program product for processing enterprise tax data to improve the accuracy of financial institutions in determining whether an enterprise has operating risks through tax data.

[0006] In a first aspect, an embodiment of the present application provides a method for processing corporate tax data, comprising: obtaining a tax sample of a target enterprise, the tax sample comprising N tax data corresponding to N consecutive months of the history of the target enterprise, wherein N is a positive integer greater than 1; performing feature extraction on the tax sample to obtain a tax feature set of the target enterprise, the tax feature set comprising N tax feature groups, the N tax feature groups corresponding one-to-one to the N tax data, each of the N tax feature groups comprising M tax features, wherein M is a positive integer greater than 1; predicting the target enterprise's tax rate using a tax data processing model based on the tax feature set; The first predicted tax data for the target month, the tax data processing model includes an encoder and a decoder, the encoder and the decoder include a self-attention module, the self-attention module is used to parse the mutual dependencies between different tax feature groups in the tax feature set; obtain the first actual tax data of the target enterprise for the target month; determine a tax difference value based on the first predicted tax data and the first actual tax data, the tax difference value is used to indicate the degree of data difference between the first actual tax data and the first predicted tax data; when the tax difference value meets the warning condition, report a financial risk warning for the target enterprise.

[0007] In a possible implementation, the method further includes: obtaining a plurality of training tax samples, each of the plurality of training tax samples comprising N training tax data corresponding to the N months in the history of an enterprise; performing feature extraction on the plurality of training tax samples to obtain a plurality of the tax feature sets; and training an initial model according to each tax feature set in the plurality of the tax feature sets to obtain the tax data processing model.

[0008] In one possible implementation, the decoder includes a mask module, which is used to process the output data of the self-attention module, and the mask module is used to mask the dependency of the tax feature group corresponding to the first month in the tax feature set on the tax feature group corresponding to the second month, where the first month is earlier than the second month in the time domain.

[0009] In one possible implementation, the multiple tax feature sets include a first tax feature set, and the training of the initial model based on the first tax feature set includes: inputting the first tax feature set into the encoder to obtain a first tax feature vector; using the decoder to predict, based on the first tax feature vector, second predicted tax data corresponding to the third month of the enterprise to which the first tax feature set belongs; obtaining second actual tax data corresponding to the third month of the enterprise to which the first tax feature set belongs; calculating a loss value using a mean square error function based on the second predicted tax data and the actual tax data; and updating the encoder and the decoder based on the loss value.

[0010] In a possible implementation, the tax data includes basic information, comprehensive declaration information, declaration details, illegal and regulatory violation information, audit case information, invoice information, balance sheet and income statement of the target enterprise in the month corresponding to the tax data.

[0011] In a possible implementation, obtaining a plurality of training tax samples includes: confirming a plurality of identifiers corresponding to the plurality of training tax samples, each of the plurality of identifiers being used to indicate that the training tax sample corresponding to each identifier satisfies a training condition, the training condition including that the credit level of the enterprise to which the training tax sample belongs meets the training requirements; obtaining the plurality of training tax samples according to the plurality of identifiers.

[0012] In a second aspect, an embodiment of the present application provides a corporate tax data processing device, characterized in that the device comprises:

[0013] An acquisition module, the acquisition module is used to acquire a tax sample of a target enterprise, the tax sample includes N tax data corresponding to N consecutive months of the history of the target enterprise, where N is a positive integer greater than 1; the acquisition module is also used to acquire the first actual tax data of the target month of the target enterprise;

[0014] A feature extraction module, the feature extraction module is used to extract features from the tax sample to obtain a tax feature set of the target enterprise, the tax feature set includes N tax feature groups, the N tax feature groups correspond to the N tax data one by one, each of the N tax feature groups includes M tax features, and M is a positive integer greater than 1;

[0015] a prediction module, the prediction module being used to predict first predicted tax data of a target month of the target enterprise based on the tax feature set using a tax data processing model, the tax data processing model comprising an encoder and a decoder, the encoder and the decoder comprising a self-attention module, the self-attention module being used to parse mutual dependencies between different tax feature groups in the tax feature set;

[0016] a determination module, the determination module being used to determine a tax difference value according to the first predicted tax data and the first actual tax data, the tax difference value being used to indicate a degree of data difference between the first actual tax data and the first predicted tax data;

[0017] A reporting module, wherein the reporting module is used to report a financial risk warning for the target enterprise when the tax difference value meets the warning condition.

[0018] In a possible implementation, the acquisition module is also used to acquire multiple training tax samples, each of the multiple training tax samples includes N training tax data corresponding to the N months in the history of an enterprise; the feature extraction module is also used to extract features from the multiple training tax samples to obtain multiple tax feature sets; the prediction module is also used to train the initial model according to each tax feature set in the multiple tax feature sets to obtain the tax data processing model.

[0019] In one possible implementation, the decoder includes a mask module, which is used to process the output data of the self-attention module, and the mask module is used to mask the dependency of the tax feature group corresponding to the first month in the tax feature set on the tax feature group corresponding to the second month, where the first month is earlier than the second month in the time domain.

[0020] In a possible implementation, the multiple tax feature sets include a first tax feature set, and the prediction module is specifically used to: input the first tax feature set into the encoder to obtain a first tax feature vector; use the decoder to predict, based on the first tax feature vector, second predicted tax data corresponding to the third month of the enterprise to which the first tax feature set belongs; obtain second actual tax data corresponding to the third month of the enterprise to which the first tax feature set belongs; use the mean square error function to calculate the loss value based on the second predicted tax data and the actual tax data; and update the encoder and the decoder based on the loss value.

[0021] In a possible implementation, the tax data includes basic information, comprehensive declaration information, declaration details, illegal and regulatory violation information, audit case information, invoice information, balance sheet and income statement of the target enterprise in the month corresponding to the tax data.

[0022] In a possible implementation, the acquisition module is specifically used to: confirm multiple identifiers corresponding to the multiple training tax samples, each of the multiple identifiers is used to indicate that the training tax sample corresponding to each identifier meets the training conditions, and the training conditions include that the credit level of the enterprise to which the training tax sample belongs meets the training requirements; and acquire the multiple training tax samples according to the multiple identifiers.

[0023] In a third aspect, an embodiment of the present application further provides a computing device, which may include a processor and a memory:

[0024] The memory is used to store computer programs;

[0025] The processor is used to execute the method described in the first aspect and any one of the embodiments of the first aspect according to the computer program.

[0026] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program, and the computer program is used to execute the method described in the first aspect and any one of the implementation methods of the first aspect.

[0027] In a fifth aspect, an embodiment of the present application further provides a computer program product comprising instructions, which, when executed on a computing device, enables the computing device to execute the method described in the first aspect and any one of the implementations of the first aspect.

[0028] In the above implementation method of the embodiment of the present application, a tax sample of the target enterprise is obtained, and the tax sample includes N tax data corresponding to N consecutive months of the history of the target enterprise, where N is a positive integer greater than 1; feature extraction is performed on the tax sample to obtain a tax feature set of the target enterprise, and the tax feature set includes N tax feature groups, and the N tax feature groups correspond to the N tax data one by one, and each tax feature group in the N tax feature groups includes M tax features, where M is a positive integer greater than 1; based on the tax feature set, a tax data processing model is used to predict the first predicted tax data of the target month of the target enterprise, and the tax data processing model includes an encoder and a decoder, and the encoder and the decoder include a self-attention module, and the self-attention module is used to parse the mutual dependencies between different tax feature groups in the tax feature set; the first actual tax data of the target month of the target enterprise is obtained; based on the first predicted tax data and the first actual tax data, a tax difference value is determined, and the tax difference value is used to indicate the degree of data difference between the first actual tax data and the first predicted tax data; when the tax difference value meets the warning condition, a tax warning for the target enterprise is reported.

[0029] In this way, through the historical tax data of the enterprise, the tax data processing model can be used to estimate the tax data of the enterprise in the future target month. When there is a deviation between the actual tax data of the target month and the estimated tax data, and the tax difference value of the deviation meets the warning condition, a tax warning for the enterprise can be reported. Then, in the process of post-loan management of the enterprise, it can be quantitatively determined whether the enterprise has operating risks through the changes in tax data, and an early warning can be issued when the enterprise has operating risks, avoiding the credit risk caused by the decline in the repayment ability of the enterprise, and improving the risk prevention and control ability of financial institutions facing the credit loan services of the enterprise. In addition, since in the tax data processing model, both the encoder and the decoder include a self-attention module, and the self-attention module can parse the dependency between the tax feature groups of different months in the tax feature set, compared with a single tax feature, the dependency between the tax feature groups composed of multiple tax features can better reflect the changes in tax data. When predicting the tax data of the target month, the model can calculate the correlation between the tax data of the target month and each month in the N consecutive historical months, effectively improving the accuracy of the financial risk early warning for the enterprise based on the tax data report, and reducing the misjudgment and omission of the enterprise when the financial risk occurs. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0031] Figure 1 This is a schematic diagram of the architecture of an enterprise tax data processing system in an embodiment of the present application;

[0032] Figure 2 This is a flow chart of a method for processing corporate tax data in an embodiment of the present application;

[0033] Figure 3 This is a schematic diagram of the architecture of an encoder in an embodiment of the present application;

[0034] Figure 4 This is a schematic diagram of the structure of an enterprise tax data processing device in an embodiment of the present application;

[0035] Figure 5 A schematic diagram of the hardware structure of a computing device in an embodiment of the present application. DETAILED DESCRIPTION

[0036] See also Figure 1 , is a schematic diagram of an enterprise tax data processing system provided in this application. Figure 1 As shown, the enterprise tax data processing system 10 includes a processor and a memory. Figure 1 The example includes processor 100, memory 201 and memory 202. In actual application scenarios, the enterprise tax data processing system 10 may include a larger number of memories.

[0037] The processor 100 and the memory 201, and the processor 100 and the memory 202 may be connected via a bus, and the bus may be, for example, a peripheral component interconnect express (PCIE) bus, etc. Alternatively, the processor 100 and the memory 201, and the processor 100 and the memory 202 may also be connected via a network, and the network may be, for example, a local area network (LAN), a wide area network (WAN), etc. in terms of coverage, and may be, for example, a wired network or a wireless network, etc. in terms of connection mode.

[0038] The processor 100 refers to a device having data processing and control capabilities. For example, the processor 100 may be a graphics processing unit (GPU), a central processing unit (CPU), etc., and this is not limited.

[0039] The memory 201 (similar to the memory 202) refers to a device with file storage capability, such as a semiconductor memory, a magnetic memory, an optical memory, etc., which is not limited. In addition, the memory 201 may store a database for data management and data query, which is not limited.

[0040] Normally, in credit loan services for enterprises, the tax data of the enterprise is an important factor in measuring the credit status and repayment ability of the enterprise. It is usually an important basis for whether the enterprise passes the loan approval and determines the credit limit of the enterprise. Among them, the service clients of credit loan services can be small and micro enterprises.

[0041] In actual application, the tax data of enterprises can be applied to the pre-loan access and credit granting process in the enterprise credit loan service. However, for the post-loan management of the enterprise credit loan service, the tax data of enterprises is often easily overlooked. Usually, the account manager of the financial institution can monitor the tax data of the enterprise in the process of post-loan management. For example, the account manager can visit the enterprise in person to verify the qualifications of the enterprise. However, when the enterprise encounters operating risks, it is often difficult for the account manager to find the changes in the credit status and repayment ability of the enterprise from the tax data. This is because the complete tax data of the enterprise has the characteristics of large data volume, many data features and wide data range. In the process of post-loan management, even if the tax data can be used as an important basis for assessing the credit and financial risks of the enterprise, there are still a lot of misjudgments or omissions. In addition, due to the large number of tax features of tax data, in actual application, the financial risk problems of enterprises are often related to a large number of tax features. It is usually difficult to accurately reflect the risk problems of the enterprise through the assessment of a single tax feature, which will also cause a large number of misjudgments or omissions. Therefore, it is very important to determine whether the enterprise has operating risks through tax data with accuracy and efficiency.

[0042] Based on this, a tax sample of the target enterprise is obtained, and the tax sample includes N tax data corresponding to N consecutive months of the target enterprise's history, where N is a positive integer greater than 1; feature extraction is performed on the tax sample to obtain a tax feature set of the target enterprise, and the tax feature set includes N tax feature groups, and the N tax feature groups correspond to the N tax data one by one, and each of the N tax feature groups includes M tax features, where M is a positive integer greater than 1; according to the tax feature set, a tax data processing model is used to predict the first predicted tax data of the target month of the target enterprise, and the tax data processing model includes an encoder and a decoder, and the encoder and the decoder include a self-attention module, and the self-attention module is used to parse the mutual dependencies between different tax feature groups in the tax feature set; the first actual tax data of the target month of the target enterprise is obtained; according to the first predicted tax data and the first actual tax data, a tax difference value is determined, and the tax difference value is used to indicate the degree of data difference between the first actual tax data and the first predicted tax data; when the tax difference value meets the warning condition, a tax warning for the target enterprise is reported.

[0043] In this way, through the historical tax data of the enterprise, the tax data processing model can be used to estimate the tax data of the enterprise in the future target month. When there is a deviation between the actual tax data of the target month and the estimated tax data, and the tax difference value of the deviation meets the warning condition, a tax warning for the enterprise can be reported. Then, in the process of post-loan management of the enterprise, it can be quantitatively determined whether the enterprise has operating risks through the changes in tax data, and an early warning can be issued when the enterprise has operating risks, avoiding the credit risk caused by the decline in the repayment ability of the enterprise, and improving the risk prevention and control ability of financial institutions facing the credit loan services of the enterprise. In addition, since in the tax data processing model, both the encoder and the decoder include a self-attention module, and the self-attention module can parse the dependency between the tax feature groups of different months in the tax feature set, compared with a single tax feature, the dependency between the tax feature groups composed of multiple tax features can better reflect the changes in tax data. When predicting the tax data of the target month, the model can calculate the correlation between the tax data of the target month and each month in the N consecutive historical months, effectively improving the accuracy of the financial risk early warning for the enterprise based on the tax data report, and reducing the misjudgment and omission of the enterprise when the financial risk occurs.

[0044] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, various non-limiting implementation methods in the embodiments of the present application are exemplarily described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0045] See also Figure 2 , Figure 2 A schematic diagram of a process flow of a method for processing corporate tax data in an embodiment of the present application is shown. The method can be applied to Figure 1 The enterprise tax data processing system 10 shown in FIG. 1 may be applied to other applicable enterprise tax data processing systems. Figure 1 The enterprise tax data processing system 10 shown in the figure is used as an example to illustrate the enterprise tax data processing. Figure 2 As shown, the method may specifically include the following steps.

[0046] S201: The processor obtains a tax sample of a target enterprise, where the tax sample includes N tax data corresponding to N consecutive months of the history of the target enterprise, where N is a positive integer greater than 1.

[0047] In order to predict the tax data of the target month of the target enterprise, the processor can obtain the tax data of the target enterprise for a period of time in history. Generally, the tax sample of the target enterprise obtained by the processor may include N tax data corresponding to N consecutive months in the history of the target enterprise, where N is a positive integer greater than 1. The processor can obtain the tax sample of the target enterprise from the memory 201 (or the memory 202), or the tax sample of the target enterprise can be obtained from the database of the memory 201. In actual application, since the credit loan service for enterprises usually lasts for one year, the value of N can be 12, that is, the processor obtains the tax data of the target enterprise for 12 months in history throughout the year.

[0048] In actual application, for example, each of the N tax data of the target enterprise obtained by the processor may include, in the month corresponding to the tax data, the basic information, comprehensive declaration information, declaration details, violation information, audit case information, invoice information, balance sheet and income statement of the target enterprise.

[0049] It should be noted that the above tax data may contain more specific data contents, which are not limited.

[0050] S202: The processor extracts features from the tax sample to obtain a tax feature set of the target enterprise, where the tax feature set includes N tax feature groups, where the N tax feature groups correspond one to one with the N tax data, and each of the N tax feature groups includes M tax features, where M is a positive integer greater than 1.

[0051] Before the processor predicts the tax data of the target enterprise in the target month based on the tax sample, the processor needs to pre-process the tax sample to refine the tax sample into specific tax features so that the model can perform analysis and processing based on the tax features.

[0052] Specifically, the processor may extract N tax feature groups corresponding to the N tax data included in the tax sample, wherein each tax feature group may include M tax features, and M is a positive integer greater than 1. Each tax feature may indicate a feature of one dimension of the tax status of the enterprise in the month corresponding to the tax feature, and the M tax features directly reflect the overall tax status of the enterprise in the month corresponding to the tax feature.

[0053] Exemplarily, the processor performs feature extraction on the first tax data among N tax data as an example for explanation, wherein the first tax data includes, in the month corresponding to the first tax data, the basic information, comprehensive declaration information, declaration details, illegal and regulatory information, audit case information, invoice information, balance sheet and income statement of the target enterprise. For example, the processor can extract tax features such as taxpayer type, registration type code, taxpayer qualification type code, taxpayer status code, registered capital, competent tax bureau code, administrative division dictionary of registered address, business scope, opening date, enterprise scale code, industry code, etc. from the basic information in the first tax data; the processor can also extract tax features such as collection item code, total sales revenue, taxable sales revenue, taxable amount, collection item code, prepaid tax amount, tax refund amount, tax exemption amount, etc. from the comprehensive declaration information in the first tax data; the processor can also extract tax type code, The processor can also extract tax characteristics such as actual payment amount, tax rate, and tax basis; the processor can also extract tax characteristics such as main illegal and illegal means code, illegal and illegal status code from the illegal and illegal information in the first tax data; the processor can also extract tax characteristics such as case type code, major case mark, major case mark, and handling status in advance from the audit case information in the first tax data; the processor can also extract tax characteristics such as amount and tax amount from the invoice information in the first tax data; the processor can also extract tax characteristics such as asset type and asset amount from the balance sheet in the first tax data; the processor can also extract tax characteristics such as current period amount from the income statement in the first tax data.

[0054] It should be noted that the above processor extracts the tax features from the first tax data for exemplary purposes only. In actual application, the processor may also extract other tax features.

[0055] Furthermore, the processor may perform other preprocessing operations on the extracted tax features. Exemplarily, the processor may normalize the extracted tax features of numerical type, that is, transform the numerical value of the tax features to a specific numerical range, which may be, for example, 0 to 1, or -1 to 1. Alternatively, the processor may perform one-hot encoding on the extracted tax features with predetermined category values, for example, the tax features may be taxpayer types, registration type codes, tax item codes, processing conditions, or major case flags.

[0056] In addition, some tax features of the first tax data may not be successfully extracted, that is, the values ​​of the tax features are missing. The processor may fill the missing tax features with zeros, or the processor may add masks to the missing tax features. Then, when the processor makes a prediction based on the tax feature set later, the processor may mask the missing tax features to prevent the missing values ​​from affecting the prediction.

[0057] In actual application, the processor extracts M tax features from the first tax data, and the value of M may be 1642, for example, that is, the first tax feature set corresponding to the first tax data contains tax features of 1×1642 dimensions.

[0058] S203: The processor predicts the first predicted tax data of the target month of the target enterprise based on the tax feature set using the tax data processing model, the tax data processing model includes an encoder and a decoder, the encoder and the decoder include a self-attention module, and the self-attention module is used to parse the mutual dependencies between different tax feature groups in the tax feature set.

[0059] The processor can use the tax data processing model to predict the first predicted tax data of the target enterprise in the target month according to the extracted tax feature set. The first predicted tax data is the estimated tax data that the target enterprise should achieve to maintain a stable credit status.

[0060] Among them, the tax data processing model includes an encoder and a decoder, and the encoder and the decoder include a self-attention module, and the self-attention module is used to analyze the mutual dependencies between different tax feature groups in the tax feature set. In this way, since in the tax data processing model, both the encoder and the decoder include a self-attention module, and the self-attention module can analyze the dependencies between tax feature groups in different months in the tax feature set, compared with a single tax feature, the dependencies between tax feature groups composed of multiple tax features can better reflect the changes in tax data. When predicting the tax data of the target month, the model can calculate the correlation between the tax data of the target month and each month in the N consecutive historical months, which effectively improves the accuracy of the financial risk warning for enterprises based on the tax data report and reduces the misjudgment and omission of enterprises when financial risks occur.

[0061] In practical applications, the tax data processing model can be a transformer model. Compared with the Recurrent Neural Network (RNN) model, the residual connection structure in the transformer model can avoid the problem of gradient explosion or gradient disappearance when processing long sequence data, thereby improving the accuracy of prediction. In addition, compared with the Long Short-Term Memory (LSTM) model, the transformer model can perform parallel computing, thereby improving the efficiency of model prediction and model training.

[0062] In specific implementation, the processor can also train the above-mentioned tax data processing model. Exemplarily, the processor can first obtain multiple training tax samples, each of which includes N training tax data corresponding to N consecutive months of history of an enterprise. Among them, the processor can obtain multiple training tax samples from the memory 201 (or the memory 202), or can obtain multiple training tax samples from the database of the memory 201. Then, the processor performs the same feature extraction method as step S202 on the multiple training tax samples to obtain multiple tax feature sets. Subsequently, the processor trains the initial model according to each tax feature set in the multiple tax feature sets to obtain the tax data processing model. In actual applications, the processor can obtain training tax samples of different categories of enterprises (i.e., different categories of taxpayer types, registration type codes, and tax item codes) for training, so that the predicted tax data of the target month of the target enterprise generated by the processor is more relevant to the category of the target enterprise.

[0063] It should be noted that before the processor obtains multiple training tax samples, the processor may first confirm multiple identifiers corresponding to the multiple training tax samples, and each of the multiple identifiers is used to indicate that the training tax sample corresponding to each identifier meets the training conditions, wherein the training conditions may include that the credit level of the enterprise to which the training tax sample belongs meets the training requirements. Furthermore, according to the multiple identifiers, the processor may obtain multiple training tax samples. In this way, there will be no tax samples of enterprises whose credit levels do not meet the training requirements in the training tax samples used for training, avoiding dirty data from polluting the model.

[0064] Furthermore, an example is given for explaining the training of an initial model based on a first tax feature set among multiple tax feature sets.

[0065] As a first step, the processor may input the first tax feature set into the encoder of the initial model to obtain a first tax feature vector. In actual application, the tax data processing model may include multiple encoders, and the processor inputs the first tax feature set into multiple encoders connected end to end in the initial model to obtain the first tax feature vector. The number of encoders may be, for example, 6. In addition, the last encoder of the multiple encoders connected end to end may be connected to a fully connected layer and output the result. For example, the first tax feature set may include a historical 12-month tax feature group of the enterprise corresponding to the first tax feature set, that is, the first tax feature set may include tax features of 12×1642 dimensions. The first tax feature vector finally obtained by the processor may include tax features of 1×143 dimensions.

[0066] Exemplarily, the first encoder of the plurality of encoders connected end to end is as follows Figure 3 As shown in FIG. 1 , the self-attention module, the mask module, the first summation and normalization, the feedforward neural network module, and the second summation and normalization in the encoder are connected end to end in sequence.

[0067] Specifically, the self-attention module can perform the calculation z=softmax(QK T )V, where Q is the query matrix, K T is the transpose of the key matrix. The softmax function is used to normalize the dot product between the query matrix and the transpose of the key matrix to obtain the attention weights, which are finally multiplied by the value matrix V to obtain the output z of the self-attention layer.

[0068] The mask module can specifically perform the calculation z=mask(z), where the size of the mask matrix is ​​the same as the size of the z matrix, and the value of the element located in the upper right corner of the mask matrix (excluding the elements on the diagonal of the mask matrix) is 0, and the value of the element located in the lower left corner of the mask matrix (including the elements on the diagonal of the mask matrix) is 1. Specifically, in the actual tax work of the enterprise, the tax data of the earlier month will not depend on the tax data of the later month, or it can be said that the tax data of the later month will not affect the tax data of the earlier month. Therefore, the mask module can mask the dependency of the tax feature group corresponding to the first month on the tax feature group corresponding to the second month in the tax feature set by processing the output data z in the self-attention module, where the first month is earlier than the second month in the time domain. For example, for the matrix z, the element z[i,j] in the matrix z is used to represent the dependency of the tax data of the i-th month on the tax data of the j-th month. Therefore, when the value of i is less than j, the value of the element z[i,j] in the matrix z after processing by the mask matrix is ​​0, that is, the tax data of the i-th month will not depend on the tax data of the j-th month, or the tax data of the j-th month will not affect the tax data of the i-th month.

[0069] The first summation and normalization can be specifically performed by calculating LayerNorm(X+Sublayer(X)), where X is the input matrix of the previous module, and Sublayer(X) is the output matrix after the previous module processes X. Summing X and Sublayer(X) can reduce the gradient vanishing problem of the model, that is, residual connection is performed on the output data of the mask module. Furthermore, the LayerNorm function can normalize the result of the residual connection.

[0070] The feedforward neural network module can specifically perform the calculation FFN(X)=max(0,XW1+b1)W2+b2, where X is the input matrix of the feedforward neural network module, that is, the matrix obtained by the first summation and normalization. W1 is the weight matrix of the first linear transformation, b1 is the bias of the first linear transformation, W2 is the weight matrix of the second linear transformation, and b2 is the bias of the second linear transformation. In addition, max(0,XW1+b1) is the result of applying the activation function ReLU (Rectified Linear Unit) to the first linear transformation, which represents the nonlinear transformation in the feedforward neural network. The final FFN(X) is the output matrix of the feedforward neural network module.

[0071] The second summation and normalization can be specifically performed by calculating LayerNorm(O+FFN(O))), where O is the input matrix of the feedforward neural network module, and FFN(O) is the output matrix of the feedforward neural network module, that is, residual connection is performed on the output data of the feedforward neural network module. Furthermore, the LayerNorm function can normalize the result of the residual connection.

[0072] It should be noted that the structure of the above encoder and the execution content of each module in the encoder are all exemplary descriptions. In actual application, the above encoder can also include more modules, or the above modules can also execute other execution contents, which are not limited.

[0073] As a second step, the processor can use the decoder and, based on the first tax feature vector obtained in the above steps, predict the second predicted tax data corresponding to the third month of the enterprise to which the first tax feature set belongs, wherein the third month is later than the N consecutive months in the training tax sample in the time domain. Exemplarily, the training tax sample includes the tax data of an enterprise from the 1st month to the 12th month, and the processor can use the decoder to predict the predicted tax data of the enterprise for the 13th month based on the first tax feature vector. Further, the processor can use the decoder to predict the predicted tax data of the enterprise for the 14th month based on the first tax feature vector and the actual tax data of the enterprise for the 13th month. Alternatively, the processor can use the decoder to predict the predicted tax data of the enterprise for the 14th month based on the first tax feature vector and the tax data of the enterprise for the 13th month predicted by the decoder.

[0074] As a third step, the processor may obtain the second actual tax data corresponding to the third month of the enterprise to which the first tax feature set belongs, that is, the obtained second actual tax data is the actual tax data of the enterprise in the third month.

[0075] As a fourth step, the processor calculates the loss value using a mean square error function based on the second predicted tax data and the second actual tax data. Specifically, the second predicted tax data and the second actual tax data both contain multi-dimensional tax features, and then the processor can calculate the mean square error value of each tax feature of the second predicted tax data and the second actual tax data, and the mean square error value is the loss value of the second predicted tax data predicted by the processor. In actual application, the processor can predict the tax data of multiple months of the enterprise to which the first tax feature set belongs, and calculate the loss value.

[0076] As a fifth step, the processor may update the encoder and the decoder according to the loss value.

[0077] In actual application, the processor can perform the above steps for each tax feature set in multiple tax feature sets, and perform multiple iterations to continuously update the encoder and decoder. During the iterative training of the model, the processor can use the optimization method of stochastic gradient descent for training. After the model reaches convergence, the processor can keep the weight parameters of the encoder and decoder fixed, and the final enterprise tax data processing model can be used to more accurately predict the first predicted tax data of the target month of the target enterprise.

[0078] S204: The processor obtains the first actual tax data of the target enterprise in the target month.

[0079] S205: Determine a tax difference value based on the first predicted tax data and the first actual tax data, where the tax difference value is used to indicate the extent of data difference between the first actual tax data and the first predicted tax data.

[0080] The processor may compare the tax data predicted for the target month of the target enterprise with the actual tax data, thereby determining a tax difference value, which may indicate the extent of the data difference between the predicted tax data and the actual tax data. The processor may obtain the first actual tax data from the memory 201 (or the memory 202), or may obtain the first actual tax data from a database in the memory 201.

[0081] In actual application, both the first predicted tax data and the first actual tax data contain multi-dimensional tax features, and then the processor can calculate the mean square difference value of each tax feature of the first predicted tax data and the first actual tax data, and the mean square difference value is the above-mentioned tax difference value. In addition, the processor can also determine the tax difference value based on other methods, which is not limited.

[0082] S206: When the tax difference value meets the warning conditions, a financial risk warning for the target enterprise is reported.

[0083] According to the determined tax difference value, the processor may determine whether the tax difference value satisfies a warning condition, wherein the warning condition may be that the tax difference value is higher than a preset tax change threshold.

[0084] In this way, through the company's historical tax data, the tax data of the company's future target month can be estimated using the tax data processing model. When there is a deviation between the actual tax data of the target month and the estimated tax data, and the tax difference value of the deviation meets the warning conditions, a tax warning for the company can be reported. Furthermore, in the process of post-loan management of the company, it is possible to quantitatively determine whether the company has operating risks through changes in tax data, and issue a warning when the company has operating risks, thereby avoiding credit risks caused by the decline in the company's repayment ability and improving the risk prevention and control capabilities of financial institutions in facing credit loan services to enterprises.

[0085] In addition, the present application embodiment also provides a corporate tax data processing device. Figure 4 , Figure 4 A schematic diagram of the structure of a corporate tax data processing device in an embodiment of the present application is shown. Figure 4 The enterprise tax data processing device 400 shown includes:

[0086] An acquisition module 401 is used to acquire a tax sample of a target enterprise, wherein the tax sample includes N tax data corresponding to N consecutive months of the history of the target enterprise, where N is a positive integer greater than 1; the acquisition module 401 is also used to acquire the first actual tax data of the target month of the target enterprise;

[0087] A feature extraction module 402, the feature extraction module 402 is used to extract features from the tax sample to obtain a tax feature set of the target enterprise, the tax feature set includes N tax feature groups, the N tax feature groups correspond to the N tax data one by one, each of the N tax feature groups includes M tax features, and M is a positive integer greater than 1;

[0088] A prediction module 403, the prediction module 403 is used to predict the first predicted tax data of the target month of the target enterprise according to the tax feature set using a tax data processing model, the tax data processing model includes an encoder and a decoder, the encoder and the decoder include a self-attention module, the self-attention module is used to analyze the mutual dependencies between different tax feature groups in the tax feature set;

[0089] a determination module 404, the determination module 404 being used to determine a tax difference value according to the first predicted tax data and the first actual tax data, the tax difference value being used to indicate a degree of data difference between the first actual tax data and the first predicted tax data;

[0090] The reporting module 405 is used to report a financial risk warning for the target enterprise when the tax difference value meets the warning condition.

[0091] In a possible implementation, the acquisition module 401 is also used to acquire multiple training tax samples, each of the multiple training tax samples includes N training tax data corresponding to the N months in the history of an enterprise; the feature extraction module 402 is also used to extract features from the multiple training tax samples to obtain multiple tax feature sets; the prediction module 403 is also used to train the initial model according to each tax feature set in the multiple tax feature sets to obtain the tax data processing model.

[0092] In one possible implementation, the decoder includes a mask module, which is used to process the output data of the self-attention module, and the mask module is used to mask the dependency of the tax feature group corresponding to the first month in the tax feature set on the tax feature group corresponding to the second month, where the first month is earlier than the second month in the time domain.

[0093] In a possible implementation, the multiple tax feature sets include a first tax feature set, and the prediction module 403 is specifically used to: input the first tax feature set into the encoder to obtain a first tax feature vector; use the decoder to predict, based on the first tax feature vector, second predicted tax data corresponding to the third month of the enterprise to which the first tax feature set belongs; obtain second actual tax data corresponding to the third month of the enterprise to which the first tax feature set belongs; use the mean square error function to calculate the loss value based on the second predicted tax data and the actual tax data; and update the encoder and the decoder based on the loss value.

[0094] In a possible implementation, the tax data includes basic information, comprehensive declaration information, declaration details, illegal and regulatory violation information, audit case information, invoice information, balance sheet and income statement of the target enterprise in the month corresponding to the tax data.

[0095] In a possible implementation, the acquisition module 401 is specifically used to: confirm multiple identifiers corresponding to the multiple training tax samples, each of the multiple identifiers is used to indicate that the training tax sample corresponding to each identifier meets the training conditions, and the training conditions include that the credit level of the enterprise to which the training tax sample belongs meets the training requirements; and acquire the multiple training tax samples according to the multiple identifiers.

[0096] It should be noted that the information interaction, execution process, etc. between the modules and units of the above-mentioned device are based on the same concept as the method embodiment in the embodiments of the present application, and the technical effects they bring are the same as those of the method embodiment in the embodiments of the present application. For specific contents, please refer to the description in the method embodiment shown in the aforementioned embodiments of the present application, and will not be repeated here.

[0097] In addition, the present application embodiment also provides a computing device. Figure 5 , Figure 5 FIG. 1 is a schematic diagram showing the hardware structure of a computing device in an embodiment of the present application. Figure 5 As shown, computing device 400 may include a processor 401 and a memory 402 .

[0098] Wherein, the memory 402 is used to store computer programs;

[0099] The processor 401 is used to execute the enterprise tax data processing method described in the above method embodiment according to the computer program.

[0100] In addition, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store a computer program, and the computer program is used to execute the corporate tax data processing method described in the above method embodiment.

[0101] In addition, an embodiment of the present application also provides a computer program product containing instructions, which, when executed on a computing device, enables the computing device to execute the corporate tax data processing method described in the above method embodiment.

[0102] In the examples of this application, the "first" in the names such as "first month" and "first tax feature set" is only used as a name identifier and does not mean the first in order. The same rule applies to "second", "third", etc.

[0103] Through the description of the above implementation methods, it can be known that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a read-only memory (ROM) / RAM, a magnetic disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in each embodiment of the present application or some parts of the embodiments.

[0104] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, in which the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0105] The above description is merely an exemplary embodiment of the present application and is not intended to limit the protection scope of the present application.

Claims

1. A method for processing corporate tax data, characterized in that: The method comprises: Obtaining a tax sample of a target enterprise, wherein the tax sample includes N tax data corresponding to N consecutive months of history of the target enterprise, where N is a positive integer greater than 1; Extracting features from the tax sample to obtain a tax feature set of the target enterprise, the tax feature set comprising N tax feature groups, the N tax feature groups corresponding to the N tax data one by one, each of the N tax feature groups comprising M tax features, where M is a positive integer greater than 1; According to the tax feature set, a tax data processing model is used to predict first predicted tax data of a target month of the target enterprise, wherein the tax data processing model includes an encoder and a decoder, wherein the encoder and the decoder include a self-attention module, and the self-attention module is used to analyze the mutual dependencies between different tax feature groups in the tax feature set; Acquire the first actual tax data of the target enterprise in the target month; determining a tax difference value according to the first predicted tax data and the first actual tax data, wherein the tax difference value is used to indicate the extent of data difference between the first actual tax data and the first predicted tax data; When the tax difference value meets the warning condition, a financial risk warning for the target enterprise is reported.

2. The method according to claim 1, characterized in that: The method further comprises: Acquire a plurality of training tax samples, each of the plurality of training tax samples comprising N training tax data corresponding to the N months in the history of an enterprise; Performing feature extraction on the plurality of training tax samples to obtain a plurality of tax feature sets; The initial model is trained according to each tax feature set in the multiple tax feature sets to obtain the tax data processing model.

3. The method according to claim 1 or 2, characterized in that: The decoder includes a mask module, which is used to process the output data of the self-attention module. The mask module is used to mask the dependency of the tax feature group corresponding to the first month in the tax feature set on the tax feature group corresponding to the second month, where the first month is earlier than the second month in the time domain.

4. The method according to claim 2, characterized in that: The plurality of tax feature sets include a first tax feature set, and training the initial model according to the first tax feature set includes: Inputting the first tax feature set into the encoder to obtain a first tax feature vector; Using the decoder, based on the first tax feature vector, predict second predicted tax data corresponding to the third month of the enterprise to which the first tax feature set belongs; Acquire second actual tax data corresponding to the third month of the enterprise to which the first tax feature set belongs; According to the second predicted tax data and the actual tax data, a loss value is calculated using a mean square error function; According to the loss value, the encoder and the decoder are updated.

5. The method according to claim 1, characterized in that The tax data includes the basic information, comprehensive declaration information, declaration details, illegal and regulatory violation information, audit case information, invoice information, balance sheet and income statement of the target enterprise in the month corresponding to the tax data.

6. The method according to claim 2, characterized in that The obtaining of multiple training tax samples comprises: Confirm that the multiple training tax samples correspond to multiple identifiers, each identifier in the multiple identifiers is used to indicate that the training tax sample corresponding to each identifier meets the training condition, and the training condition includes that the credit level of the enterprise to which the training tax sample belongs meets the training requirements; According to the multiple identifiers, the multiple training tax samples are obtained.

7. An enterprise tax data processing device, characterized in that: The device shown comprises: An acquisition module, the acquisition module is used to acquire a tax sample of a target enterprise, the tax sample includes N tax data corresponding to N consecutive months of the history of the target enterprise, where N is a positive integer greater than 1; the acquisition module is also used to acquire the first actual tax data of the target month of the target enterprise; A feature extraction module, the feature extraction module is used to extract features from the tax sample to obtain a tax feature set of the target enterprise, the tax feature set includes N tax feature groups, the N tax feature groups correspond to the N tax data one by one, each of the N tax feature groups includes M tax features, and M is a positive integer greater than 1; a prediction module, the prediction module being used to predict first predicted tax data of a target month of the target enterprise based on the tax feature set using a tax data processing model, the tax data processing model comprising an encoder and a decoder, the encoder and the decoder comprising a self-attention module, the self-attention module being used to parse mutual dependencies between different tax feature groups in the tax feature set; a determination module, the determination module being used to determine a tax difference value according to the first predicted tax data and the first actual tax data, the tax difference value being used to indicate a degree of data difference between the first actual tax data and the first predicted tax data; A reporting module, wherein the reporting module is used to report a financial risk warning for the target enterprise when the tax difference value meets the warning condition.

8. A computing device, characterized in that The computing device comprises a processor and a memory: The memory is used to store computer programs; The processor is configured to execute the method of any one of claims 1 to 6 according to the computer program.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store a computer program, and the computer program is used to execute the method according to any one of claims 1 to 6.

10. A computer program product comprising instructions, characterized in that When the method is executed on a computing device, the computing device is caused to execute the method according to any one of claims 1 to 6.

Images