CN-121998784-A - Tax meeting difference automatic adjustment method based on knowledge distillation

CN121998784ACN 121998784 ACN121998784 ACN 121998784ACN-121998784-A

Abstract

The invention relates to the technical field of tax meeting data processing, in particular to a tax meeting difference automatic adjustment method based on knowledge distillation, which comprises the steps of docking an enterprise financial system through an interface, collecting accounting entry data and converting the accounting entry data into structured data; the method comprises the steps of inputting the tax meeting difference adjustment accounting records and the corresponding difference types thereof into a pre-training sequence labeling model, inputting a knowledge distillation model, calculating and outputting a preliminary tax payment adjustment result, inputting the preliminary tax payment adjustment result into an adjustment strategy optimization model, generating an optimal adjustment scheme, and converting the optimal adjustment scheme into standard tax payment adjustment records and declaration attached table data according to a declaration form specification of a tax system. Through the scheme, the technical problems of high omission risk, low efficiency, limited accuracy and poor consistency in the existing manual adjustment technology for tax meeting difference are solved.

Inventors

ZHANG WENDOU
ZHANG BIN
NIE YANG
Ji Yunmei
ZHAO KAIPENG
WU HAN

Assignees

中软云企信息服务(南京)有限公司

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. The automatic tax meeting difference adjusting method based on knowledge distillation is characterized by comprising the following steps of: S1, docking an enterprise financial system through an interface, collecting accounting entry data, cleaning and standardizing the accounting entry data, and converting the accounting entry data into structured data; S2, inputting the structured data into a pre-trained sequence labeling model, identifying accounting entry to be subjected to tax meeting difference adjustment and corresponding difference types and confidence degrees, and outputting the accounting entry in a JSON array form; S3, inputting the identified accounting entry, the difference type and the confidence coefficient into a knowledge distillation model, and calculating and outputting a preliminary tax payment adjustment result; S4, inputting a preliminary tax adjustment result and business state data of the enterprise into an adjustment strategy optimization model to generate an optimal adjustment scheme; S5, converting the optimal adjustment scheme into standard tax payment adjustment entry and declaration attached table data according to a declaration form specification of the tax system.
2. The automated tax variance adjustment method of claim 1, wherein the accounting entry data comprises a subject code, subject name, amount, time of occurrence, type of business, credential ID core field.
3. The method according to claim 1, wherein in S1, the system interfaces with the enterprise financial system via a real-time interface or a batch interface, the real-time interface comprises RESTfulAPI and JDBC direct connections, and the batch interface comprises file transfer protocol and standard financial data exchange protocol.
4. The automatic adjustment method for tax variance based on knowledge distillation according to claim 2, wherein the collected raw accounting entry data is cleaned including repeated data cleaning, format error correction and outlier processing; The repeated data cleaning comprises the steps of keeping a first record and deleting a subsequent repeated item when the code, the name, the amount, the occurrence time, the service type and the certificate ID of the code, the business name and the amount are completely consistent in the same accounting period; the format error correction comprises the steps of uniformly converting the date format into the YYYY-MM-DD format, extracting pure numbers and reserving two decimal places for the non-numeric characters of the amount, and extracting the pure numbers containing letters/special symbols for the subject codes; The abnormal value processing comprises the steps of marking that the amount is negative, the service type is 'income type', prompting manual rechecking after the abnormality, automatically eliminating the inter-period data and recording logs when the occurrence time exceeds the period of the current period accounting.
5. The method for automatically adjusting tax variance based on knowledge distillation according to claim 1, wherein said step S2 comprises, S21, constructing a sequence annotation model based on BiLSTM networks and CRF layers; S22, constructing a training set of a labeling sample, wherein the labeling result is 'to be adjusted-difference type' or 'to be adjusted', inputting the training set into BiLSTM networks, capturing context association information of an accounting entry sequence through bidirectional propagation, learning constraint relations among labels through a CRF layer, and optimizing sequence labeling model parameters through reverse propagation; S23, the structured data are disassembled into four-dimensional feature vectors of 'subject code-amount-business type-occurrence period', the four-dimensional feature vectors are input into a trained sequence labeling model, wherein the subject code is converted into numerical features by adopting single-heat coding, the amount is normalized, the business type and the occurrence period are converted into vector representation through word embedding technology, and the sequence labeling model outputs labeling results corresponding to each meeting score record.
6. The method for automatically adjusting tax variance based on knowledge distillation according to claim 2, wherein said step S3 comprises, S31, constructing a teacher model by taking a tax expert rule base as a core, converting tax rules into a structured logic judgment module, and outputting accurate adjustment direction, adjustment proportion and adjustment amount by the teacher model after inputting accounting entry data; S32, constructing a lightweight convolutional neural network as a student model; S33, inputting training samples into a teacher model and a student model for distillation training, wherein the output of the teacher model is used as a soft label, the actual tax adjustment result is used as a hard label, and the trained student model is used as a knowledge distillation model.
7. The method for automatically adjusting tax variance based on knowledge distillation of claim 6, wherein step S31 comprises, S311, acquiring tax law rules which are regulated about tax meeting difference through a carding official document, and classifying the tax law rules into 5 major categories according to income categories, cost categories, expense categories, asset categories and special service categories; s312, adopting a four-component decomposition method to convert each tax rule into { trigger conditions, adjusting logic, outputting a result and carrying out exception processing }; s313, constructing a rule base based on the Drools frame, and storing the disassembled tax rule in the rule base based on the Drools frame.
8. The method for automatically adjusting tax variance of knowledge based distillation of claim 6, wherein in distillation training, distillation loss function is constructed comprising burden of taxation compliance constraint , wherein, For the cross entropy penalty of the student model and the hard tag, C is the total number of categories, N is the number of samples in batches, The one-hot value for the real label of the nth sample in class i, Predicting the probability of the nth sample as a class i for the student model; KL divergence loss for soft labels of student models and teacher models, , Is a temperature parameter, is used for softening probability distribution, As a function of the softmax of the sample, / Non-normalized output of teacher/student model, respectively; for the burden of taxation compliance constraint terms, K is the number of tax compliance items to be checked, As the weight coefficient of the kth rule, Alpha and beta are weight coefficients, alpha epsilon [0.2,0.4], beta epsilon [0.3,0.7] for the penalty function for the kth rule.
9. The method for automatically adjusting tax variance based on knowledge distillation according to claim 2, wherein said step S4 comprises, S41, defining tax meeting difference adjustment scenes as environments, defining an adjustment strategy optimization module as an intelligent agent, wherein the actions of the intelligent agent comprise adjustment amount distribution and adjustment time sequence selection; s42, adjusting a reward function of the strategy optimization model, wherein the reward function consists of two parts, namely dynamic compliance rewards and burden of taxation optimization rewards, and the formula of the reward function is as follows: , wherein, Indicating a dynamic compliance benefit is to be awarded, The optimized prize is indicated burden of taxation, =0.7, =0.3, If the adjustment result meets tax regulation, obtaining positive dynamic compliance rewards, if the adjusted enterprise burden of taxation is lower than the reference value and within the compliance range, obtaining positive burden of taxation optimized rewards, and if the compliance problem occurs, obtaining negative rewards, wherein the rewards amount is positively correlated with the burden of taxation reduction range; And S43, continuously interacting the intelligent agent with the environment, continuously updating the network parameters of the adjustment strategy optimization model to maximize the accumulated rewards, and outputting the optimal adjustment strategy by the adjustment strategy optimization model according to the actual conditions of different enterprises after training.
10. The method for automatically adjusting tax variance based on knowledge distillation according to claim 9, wherein step S41 comprises, S411, defining a 6-dimensional standardized state space and representing enterprise operation state data; S412, designing discrete actions and continuous actions, wherein the discrete actions comprise development cost plus-sum deduction proportion and depreciation method selection for adapting policy-allowed discrete options, and the continuous actions comprise fixed asset accelerated depreciation year allocation proportion for adapting flexible adjustment requirements; s413, designing a strategy network and a value network, wherein the strategy network is used for inputting a 6-dimensional state space vector, outputting discrete action probability distribution and continuous action values, and is responsible for generating an adjustment strategy, and the value network is used for inputting the 6-dimensional state space vector, outputting the state value and evaluating the advantages and disadvantages of the current strategy.

Description

Tax meeting difference automatic adjustment method based on knowledge distillation Technical Field The invention relates to the technical field of tax meeting data processing, and particularly provides a tax meeting difference automatic adjustment method based on knowledge distillation. Background At present, the adjustment of the difference between the accounting profit and the tax payment amount is mainly realized by a manual adjustment mode. Firstly, the financial staff needs to collect all accounting documents, account book records and related tax related data of the enterprise in the current period, including but not limited to various accounting records and original documents such as income types, cost types, expense types and the like, secondly, the financial staff needs to master the regulations about tax meeting difference in enterprise accounting rules, tax laws obtained by the enterprise of the people's republic of China and related regulations, and check the difference between accounting processing and tax law requirements one by comparing 200 surplus known difference items such as business recruitment fees, asset depreciation, employee benefit fees, development fee addition deduction and the like, and finally calculate the adjustment amount of the tax-paying income amount according to the check result, and manually compile tax adjustment records and tax-paying statement related attached sheets. However, the conventional manual adjustment method has the following drawbacks. Since tax can differ by up to 200 remainder items, and some of the difference items are hidden, e.g., revenue validation time point differences for certain special businesses, billing and pre-tax deduction differences for asset reduction preparations, etc. In the manual checking process one by one, the difference item is easy to miss due to human negligence, tax declaration errors are further caused, and the tax risk of enterprises and the possibility of subsequent tax check additional tax payment and fine are increased. The manual processing needs financial staff to spend a large amount of time to comb certificates, contrast rules and calculate adjustment amount, and especially for enterprises with large scale and complex ad business, the difference adjustment work before tax declaration in each period often needs a plurality of financial staff to spend a plurality of days to finish, and tax declaration efficiency is seriously affected. Tax meeting discrepancy adjustments involve complex tax term interpretation linked to accounting processes, with extremely high requirements on the professional literacy of the financial staff. The new staff financial staff is easy to cause adjustment errors due to the problems of incomplete understanding of tax law clauses, unfamiliar special business difference items and the like, and even the senior financial staff can influence the adjustment accuracy due to factors such as fatigue, subjective judgment deviation and the like. When different financial staff deal with the tax meeting difference of the same type, the consistency of adjustment results is possibly lacking due to the fact that different adjustment logics and calculation modes are adopted, meanwhile, key steps such as checking, calculation and the like in the manual adjustment process lack clear record tracks, and quick tracing of adjustment basis is difficult when tax disputes occur later. Disclosure of Invention The invention is provided for overcoming the defects, and solves or at least partially solves the technical problems of high omission risk, low efficiency, limited accuracy and poor consistency of the existing manual adjustment technology for tax meeting difference. The invention provides a tax meeting difference automatic adjustment method based on knowledge distillation, which comprises the following steps: S1, docking an enterprise financial system through an interface, collecting accounting entry data, cleaning and standardizing the accounting entry data, and converting the accounting entry data into structured data; S2, inputting the structured data into a pre-trained sequence labeling model, identifying accounting entry to be subjected to tax meeting difference adjustment and corresponding difference types and confidence degrees, and outputting the accounting entry in a JSON array form; S3, inputting the identified accounting entry, the difference type and the confidence coefficient into a knowledge distillation model, and calculating and outputting a preliminary tax payment adjustment result; S4, inputting a preliminary tax adjustment result and business state data of the enterprise into an adjustment strategy optimization model to generate an optimal adjustment scheme; S5, converting the optimal adjustment scheme into standard tax payment adjustment entry and declaration attached table data according to a declaration form specification of the tax system. Further, the accounting entry data includes a subject cod