System and method for stepwise reliability measurement and ai labeling adjustment for automated tax accounting data

Abstract

The present invention relates to a system and method for stepwise reliability measurement and AI labeling adjustment for automated tax accounting data, the system comprising: a data collection unit for collecting and standardizing accounting data; an automatic classification generation unit for automatically classifying account titles and eligibility for value-added tax deduction and generating accounting codes; a stepwise reliability measurement unit for calculating stepwise treatment reliability; a quality inspection unit for performing a multi-step quality inspection by reflecting the level and state of an inspector; an AI labeling adjustment unit for automatically adjusting a learning label according to the result; and a data output unit for storing and providing a result dataset.

Description

System and Method for Step-by-Step Reliability Measurement and AI Labeling Adjustment of Automated Tax Accounting Data

[0001] The present invention relates to a system and method for measuring the reliability of stepwise tax accounting data and adjusting AI labeling, and more specifically, to a system and method for measuring the reliability of stepwise tax accounting data and adjusting AI labeling configured to automatically classify and code data subject to accounting processing, calculate the reliability of stepwise data during the processing of the data, and automatically adjust or update the labels of an AI deep learning training dataset by reflecting the results of a quality inspection, thereby improving the quality and reliability of the training data.

[0002] With the recent advancements in tax and accounting automation and AI-based data analysis, there is a continuously increasing demand for technologies capable of efficiently classifying and processing large-scale accounting data and constructing highly reliable training datasets. In particular, as the accuracy and reliability of tax and accounting data are directly linked to the legal and financial stability of accounting practices, there is a growing need for technologies that can minimize errors in data processing and quantitatively manage the quality of training data.

[0003] Conventional automatic classification and training data generation technologies mostly rely on single classification algorithms or fixed criteria, making it difficult to flexibly respond to changes in data characteristics or exceptional situations. These technologies lack features for measuring reliability stepwise or dynamic label adjustments that reflect quality inspection results, and errors are frequently incorporated into training without modification.

[0004] Furthermore, existing systems have simple or perfunctory quality inspection procedures, making it difficult to conduct precise quality assessments that reflect the inspector's proficiency or status, and limiting functions for re-inspection or history-based reliability correction. Consequently, data quality deviations accumulate during the AI ​​model training process, resulting in a decline in analysis accuracy and predictive performance.

[0005] Accordingly, there is a need for the development of technology that can quantitatively measure reliability at each stage of the tax accounting data processing, automatically adjust training data labels by reflecting quality inspection results, and continuously improve data quality.

[0006] The present invention aims to solve the aforementioned problems by providing an automated tax accounting data processing system and method that can quantitatively calculate the reliability of step-by-step processing results, such as classifying account items, determining eligibility for value-added tax deduction, and generating accounting codes, during the large-scale tax accounting data processing process, and improve the quality of the training dataset based on this.

[0007] Furthermore, the present invention aims to provide a technology capable of flexibly responding to changes in data characteristics or exceptional situations by updating step-by-step reliability based on various conditions, such as transaction amount, type of value-added tax, whether account items have changed, transaction type, and whether predefined keywords appear.

[0008] Furthermore, the present invention aims to provide a technology capable of verifying data quality from various angles by correcting quality inspection scores by reflecting quality influencing factors such as the inspector's proficiency, weighting by specialized field, and physical and mental state during the quality inspection process, and by combining condition-based overall quality inspection with augmented data-based partial quality inspection.

[0009] Furthermore, the present invention aims to provide a technology capable of continuously improving the quality of an AI training dataset by automatically adjusting or updating the labels of the training dataset based on a comprehensive reliability calculated by combining quality inspection results and step-by-step reliability, and by storing the label change history to correct the initial reliability value during reprocessing.

[0010] Furthermore, the present invention aims to provide a technology that can maximize the quality and efficiency of tax accounting automation by integrating and implementing the above functions, thereby increasing the processing accuracy of tax accounting data and ensuring the reliability of AI model training.

[0011] The problem to be solved by this specification is not limited to what is described above and can be extended to various matters that can be derived from the embodiments of the invention described below.

[0012] A system for measuring the reliability of stepwise automatic tax accounting data and adjusting AI labeling according to one embodiment of the present invention may include: a data collection unit (110) for collecting data subject to accounting processing; an automatic classification generation unit (120) for automatically classifying account items and whether value-added tax is deductible based on the collected data subject to accounting processing and generating an accounting processing code; a stepwise reliability measurement unit (130) for calculating the reliability of stepwise by applying a calculation formula that reflects multiple probability elements and weights for each processing step of the automatic classification generation unit (120) based on Bayesian probability theory; a quality inspection unit (140) for evaluating the level and status of an inspector and performing a multi-stage quality inspection on generated data including the measured reliability of stepwise, and for producing a quality inspection result; an AI labeling adjustment unit (150) for automatically adjusting or updating the labels of an AI deep learning training dataset by reflecting the quality inspection result and the reliability of stepwise; and a data output unit (160) for storing or providing externally the generated data with assigned reliability and the adjusted AI training dataset.

[0013] According to one embodiment, the step-by-step reliability measurement unit (130) can calculate an update reliability by combining a previously calculated reliability value with a pre-set weight and reflecting at least one of a transaction amount, a value-added tax type, whether an account item has been changed, a transaction type, or whether a pre-defined keyword has appeared, and if some items of the data are missing or incomplete, it can apply replacement conditions and reflect a correction coefficient calculated by estimating uncertainty due to randomness in the reliability calculation.

[0014] According to one embodiment, the quality inspection unit (140) can calculate the proficiency of the inspector based on the inspection grade or past inspection accuracy, and correct the quality inspection score by applying a weight set for each of the inspector's professional fields and a condition weight according to a predefined classification of physical and mental conditions.

[0015] According to one embodiment, the AI ​​labeling adjustment unit (150) calculates an overall reliability by combining step reliability and quality inspection scores, changes or deletes the labels of data where the overall reliability is less than a preset allowable value, and maintains the labels of data where the reliability is greater than or equal to the allowable value and reflects them in the training dataset.

[0016] A method for measuring the reliability of stepwise tax accounting data and adjusting AI labeling according to another embodiment of the present invention may include: a data collection unit collecting data to be processed for accounting; an automatic classification generation unit automatically classifying account items and whether value-added tax is deductible based on the collected data to be processed for accounting and generating an accounting processing code; a stepwise reliability measurement unit calculating the reliability of stepwise by reflecting a plurality of probability factors and weights; a quality inspection unit evaluating the level and status of an inspector on generated data including the measured reliability of stepwise and performing a multi-stage quality inspection to produce a quality inspection result; an AI labeling adjustment unit automatically adjusting or updating the labels of an AI deep learning training dataset by reflecting the quality inspection result and the reliability of stepwise; and a data output unit storing or providing the generated data with assigned reliability and the adjusted AI training dataset externally.

[0017] According to one embodiment of the present invention, by automatically classifying and coding data subject to accounting processing and calculating the reliability of each processing step, it has the advantage of enabling quantitative data quality evaluation that reflects differences at each step, compared to existing single classification and verification methods.

[0018] In addition, according to the present invention, reliability is updated by reflecting various conditions such as transaction amount, type of value-added tax, whether account items have changed, transaction type, and whether predefined keywords appear, thereby providing the advantage of stable and adaptive data classification and verification even in the event of changes in data characteristics or exceptional situations.

[0019] In addition, according to the present invention, by adjusting the quality inspection score by reflecting the inspector's proficiency, weights by specialized field, and predefined physical and mental conditions during the quality inspection process, there is an advantage of being able to objectively and precisely improve the quality of the inspection.

[0020] In addition, according to the present invention, by performing condition-based full quality inspection and augmented data-based partial quality inspection in parallel, and designating only data where the discrepancy rate exceeds an allowable range as targets for re-inspection, it has the advantage of simultaneously achieving efficient use of inspection resources and homogenization of data quality.

[0021] In addition, according to the present invention, a comprehensive reliability is calculated by combining the quality inspection results and step-by-step reliability, and the labels of the training dataset are automatically adjusted or updated based on this, thereby having the advantage of continuously improving the quality of data used for training AI models.

[0022] In addition, according to the present invention, the label change history is stored in the dataset metadata, and the initial reliability value can be corrected by referring to the history and quality inspection results when reprocessing the same data, thereby having the advantage of maintaining consistent quality and reliability even during repetitive processing.

[0023] In addition, according to the present invention, since all processing from data collection, automatic classification, reliability measurement, quality inspection, label adjustment, and data output is automatically performed within an integrated system, it has the advantage of enabling the rapid and reliable construction of training datasets even in the processing of large-scale tax accounting data.

[0024] It should be understood that the effects of this specification are not limited to the matters described above and can be extended to various contents that can be derived from the detailed description of the embodiments of the invention below.

[0025] FIG. 1 is a schematic diagram illustrating the operational structure of a step-by-step reliability measurement and AI labeling adjustment system (100) for automatic tax accounting data according to one embodiment of the present invention.

[0026] FIG. 2 is a diagram illustrating the overall configuration of a step-by-step reliability measurement and AI labeling adjustment system (100) for automatic tax accounting data according to one embodiment of the present invention.

[0027] FIG. 3 is a schematic diagram illustrating the hardware configuration of an automatic tax accounting data step-by-step reliability measurement and AI labeling adjustment system (100) according to one embodiment of the present invention.

[0028] FIG. 4 is a flowchart illustrating the step-by-step reliability measurement process according to one embodiment of the present invention.

[0029] FIG. 5 is a flowchart illustrating the quality inspection process according to one embodiment of the present invention in steps.

[0030] FIG. 6 is a flowchart illustrating the step-by-step process of calculating and applying recommendation reliability in an AI labeling adjustment process according to an embodiment of the present invention.

[0031] FIG. 7 is a drawing illustrating an example screen of an inspector's level and status evaluation according to an embodiment of the present invention.

[0032] FIG. 8 is a drawing illustrating an example screen displaying inspection results by quality inspection item according to one embodiment of the present invention.

[0033] FIG. 9 is a drawing illustrating a statistical and comparative example screen of quality inspection results according to one embodiment of the present invention.

[0034] FIG. 10 is a flowchart illustrating the overall procedure of a step-by-step reliability measurement and AI labeling adjustment method for automatic tax accounting data according to an embodiment of the present invention.

[0035] A data collection unit (110) that collects data subject to accounting processing;

[0036] An automatic classification generation unit (120) that automatically classifies account items and whether value-added tax is deductible based on the collected accounting processing target data and generates an accounting processing code;

[0037] A step-by-step reliability measurement unit (130) that calculates step-by-step reliability by applying a calculation formula reflecting multiple probability elements and weights to each processing step of the automatic classification generation unit (120) based on Bayesian probability theory;

[0038] A quality inspection unit (140) that performs a multi-stage quality inspection and evaluates the level and status of the inspector on generated data including the measured step-by-step reliability to produce a quality inspection result;

[0039] AI labeling adjustment unit (150) that automatically adjusts or updates the labels of the AI ​​deep learning training dataset by reflecting the above quality inspection results and step-by-step reliability; and

[0040] A system for step-by-step reliability measurement and AI labeling adjustment of automatic tax accounting data, comprising a data output unit (160) that stores or provides externally the generated data with assigned reliability and the adjusted AI training dataset.

[0041] The above step-by-step reliability measurement unit (130) is,

[0042] A step-by-step reliability measurement and AI labeling adjustment system for automatic tax accounting data, which combines a previously calculated reliability value with a pre-set weight, calculates an update reliability by reflecting at least one of the transaction amount, VAT type, whether an account item has changed, the transaction form, or the appearance of a pre-defined keyword, and, if some items of the data are missing or incomplete, applies replacement conditions and reflects a correction coefficient calculated by estimating uncertainty due to randomness into the reliability calculation.

[0043] The above quality inspection department (140) is,

[0044] An automated tax accounting data step-by-step reliability measurement and AI labeling adjustment system that calculates the proficiency of an inspector based on inspection grades or past inspection accuracy, and corrects the quality inspection score by applying weights set for each of the inspector's professional fields together with state weights based on predefined classifications of physical and mental states.

[0045] The above AI labeling adjustment unit (150) is,

[0046] An automatic tax accounting data step-by-step reliability measurement and AI labeling adjustment system that calculates overall reliability by combining step-by-step reliability and quality inspection scores, changes or deletes labels of data where the overall reliability is below a preset allowable value, and maintains labels of data where the reliability is above the allowable value and reflects them in a training dataset.

[0047] A step in which a data collection unit collects data subject to accounting processing;

[0048] The automatic classification generation unit automatically classifies account items and whether value-added tax is deductible based on the collected accounting processing target data, and generates an accounting processing code;

[0049] A method for measuring step-by-step reliability and adjusting AI labeling of automatic tax accounting data, comprising: a step of a step-by-step reliability measurement unit calculating step-by-step reliability by reflecting multiple probability elements and weights; a step of a quality inspection unit evaluating the level and status of an inspector on generated data including the measured step-by-step reliability and performing a multi-step quality inspection to produce a quality inspection result; a step of an AI labeling adjustment unit automatically adjusting or updating the labels of an AI deep learning training dataset by reflecting the quality inspection result and the step-by-step reliability; and a step of a data output unit storing or providing externally the generated data with assigned reliability and the adjusted AI training dataset.

[0050] In describing the embodiments of this specification, if it is determined that a detailed description of known configurations or functions could obscure the essence of the embodiments of this specification, such detailed description is omitted. Additionally, parts of the drawings unrelated to the description of the embodiments of this specification have been omitted, and similar parts are denoted by similar reference numerals.

[0051] In the embodiments of this specification, distinct components are intended to clearly explain their respective features and do not imply that the components are necessarily separated. That is, multiple components may be integrated to form a single hardware or software unit, or a single component may be distributed to form multiple hardware or software units. Therefore, such integrated or distributed embodiments are included within the scope of the embodiments of this specification, even if not otherwise mentioned.

[0052]

[0053] FIG. 1 is a schematic diagram illustrating the operational structure of an automatic tax accounting data step-by-step reliability measurement and AI labeling adjustment system (100) according to one embodiment of the present invention, and FIG. 2 is a diagram illustrating the overall configuration of an automatic tax accounting data step-by-step reliability measurement and AI labeling adjustment system (100) according to one embodiment of the present invention.

[0054] Referring to FIGS. 1 and 2, a step-by-step reliability measurement and AI labeling adjustment system (100) for automatic tax accounting data according to one embodiment of the present invention may be configured to include a data collection unit (110), an automatic classification generation unit (120), a step-by-step reliability measurement unit (130), a quality inspection unit (140), an AI labeling adjustment unit (150), and a data output unit (160).

[0055] The data collection unit (110) may be configured to acquire data subject to accounting processing, normalize it into an internal standard schema, and store or transmit it in a form available for use in subsequent modules. Here, the data subject to accounting processing may include card approval results, bank deposit and withdrawal records, payroll statements, value-added tax related data, business partner masters and industry information, voucher images and scan data, and existing accounting processing history data.

[0056] More specifically, the data collection unit (110) may include a plurality of input adapters. The input adapters may include an application programming interface collection adapter, a message queue collection adapter, a file monitoring collection adapter, an image scan collection adapter, and a batch upload collection adapter, and each adapter may be configured to receive source data based on a predefined protocol and format.

[0057] According to one embodiment, the data collection unit (110) may be configured to perform data extraction from a card receipt image generated from a user terminal or a merchant terminal. In this case, code values ​​that are difficult to visually distinguish included in the receipt image may be interpreted by a decoding logic utilizing flash reflection characteristics or pixel value differences, and a card approval result corresponding to the interpreted code value may be retrieved and combined. Additionally, if character information extraction of the receipt image is required, an optical character recognition module may be applied.

[0058] Additionally, the data collection unit (110) may be operated by separating the raw storage area and the normalized storage area to preserve the original form of the input data. The raw storage area may be configured to store the data at the time of reception and transmission metadata without alteration, and the normalized storage area may be configured to store results to which field mapping, type conversion, unit unification, and code conversion have been applied according to an internal standard schema.

[0059] Additionally, the data collection unit (110) may include a consistency verification function for the received data. The consistency verification may include verification of the existence of essential fields, verification of data types, verification of code system validity, verification of amount consistency, and verification of date and time zone consistency, and integrity verification using a checksum or hash may be performed during the transmission and storage sections.

[0060] In one embodiment, the data collection unit (110) may include a duplicate detection function to prevent duplicate reception. Duplicate detection may be performed using a composite key generated from a combination of a transaction identifier, an approval number, a transaction time, an amount, a card identifier, and a merchant identifier, and if the same key exists, an update policy or a merge policy may be applied.

[0061] Additionally, the data collection unit (110) may include a metadata extension structure to ensure collection quality and traceability. The metadata may include source system identifier, collection adapter type, transmission path, network segment identifier, terminal identifier, image recognition reliability, OCR reliability, decoding code version, schema version, processing pipeline stage number, number of retries, and error code information, and may be shared so that the stage reliability measurement unit (130) can utilize it.

[0062] Additionally, the data collection unit (110) may include a clock management function to maintain time synchronization. The clock management may be configured to synchronize the reference time through a network time synchronization protocol and to distinguish the event time and the processing time by including the reception time and the source creation time for each input event. Time zone conversion may be set to be stored unified into a standardized reference time zone.

[0063] Additionally, the data collection unit (110) may include a processing function for protecting personally identifiable information. The protection processing may be performed by one or more of pseudonymization, partial masking, tokenization, and encryption for a specified field, and whether or not to apply and the method may be set as a policy according to the level of authority and the purpose of processing.

[0064] According to one embodiment, the data collection unit (110) may include a quarantine queue and a reprocessing queue to handle errors and exceptions. Parsing errors, schema mismatches, integrity failures, and authentication failures are isolated and stored in the quarantine queue, and recollection may be performed after automatic retries or operator review according to a reprocessing policy along with the cause code.

[0065] Additionally, the data collection unit (110) can support the combined operation of real-time processing and batch processing. Real-time processing is applied to event-based data such as card approval results, and batch processing is applied to periodic files such as payroll statements, value-added tax data, and previous year accounting processing history. Batch schedules can be managed in work units including cycles, file patterns, receiving paths, and verification rules.

[0066] Additionally, the data collection unit (110) can acquire accounting processing data from the previous year and corporate information from an external server and provide them as source data for creating a database of accounting data classifications by company. At this time, the business name, industry and item, tax service code and the type of industry corresponding to the code can be mapped and stored in a standardized code system.

[0067] In one embodiment, the data collection unit (110) may include an output interface for coupling with a subsequent module. The output interface may include a synchronous call interface and an asynchronous event publishing interface, and normalized records, summary text, keyword candidates, customer characteristics, and industry codes may be transmitted to the automatic classification generation unit (120). Additionally, the collection path, parsing reliability, OCR reliability, decoding success status, and transmission integrity indicator may be transmitted together to the step-by-step reliability measurement unit (130), and source snapshots and voucher image references may be provided to the quality inspection unit (140).

[0068] Additionally, the data collection unit (110) may include an audit logging function for system operation. The audit log may be configured to include a collection request identifier, a requester or system account, the time of collection start and completion, the number of processing cases, the number of successes and failures, the type of error, and the reprocessing result, and may be sequentially recorded in an immutable storage.

[0069] Additionally, the data collection unit (110) can be designed to be scale-out with consideration for performance management and scalability. The input adapter can be horizontally scaled on an instance basis, and for inputs requiring order guarantee, a partition key-based order-keeping queue can be applied. During mass collection, a buffer storage for temporary buffering can be used.

[0070]

[0071] The automatic classification generation unit (120) may be configured to perform account item determination, value-added tax deduction determination, and accounting processing code generation based on accounting processing target data input from the data collection unit (110). Here, the accounting processing code may include an account code, a VAT code, deduction status, tax rate classification, a business partner identifier, a description text standardization result, a voucher type, and auxiliary journal entry attributes.

[0072] More specifically, the automatic classification generation unit (120) may be composed of a preprocessor, a feature extractor, a candidate generator, a classification engine, a post-matching unit, and a code generator. The preprocessor can perform field normalization, unification of numerical units, unification of date formats, unification of currency conversion notation, cleanup of summary text, and removal of special characters. The feature extractor can be configured with input features such as transaction amount, breakdown of supply price and tax amount, keywords within approval results, names of merchants or business partners, industry codes, payment method types, and auxiliary identifiers extracted from voucher images.

[0073] According to one embodiment, the automatic classification generation unit (120) can perform space-based or morpheme-based tokenization after removing particles, conjunctions, and idiomatic stop words from summary text and approval result strings. At this time, it can be configured to generate tokens by recognizing pairs of predefined keywords and keyword values, for example, combinations of company name and amount, payment type and tax classification, and region name and merchant type can be composed of tokens.

[0074] Additionally, the automatic classification generation unit (120) can generate a set of candidate account items by referring to a database of accounting data classifications by company. The candidate generator can extract candidate accounts by applying criteria for matching business names or industry similarity between the previous year's accounting processing history and the current year's data, and can be configured to use a tax service code and an industry type mapping table together.

[0075] In one embodiment, the classification engine may include one or more of a rule-based classifier and a learning-based classifier. The rule-based classifier can evaluate mapping rules based on tax deduction requirements, industry-specific expense classification criteria, and voucher form regulations, while the learning-based classifier may be configured to predict account categories and VAT types through a model trained on past finalized journal entry data. The structure of the learning-based classifier may be any one of a linear model, a decision tree series, or an embedding-based classifier, and the specific structure may be selected according to a pre-set policy.

[0076] Additionally, the automatic classification generation unit (120) may include a post-matching unit for checking the internal consistency of the classification results. The post-matching unit may be configured to check whether the sum of the supply price and the tax amount matches the total amount, whether the combination of the tax type and the tax rate code is valid, and whether the combination of the account item and whether VAT deduction is available conflicts with accounting standards. If a conflict is detected as a result of the check, the candidate ranking may be readjusted by applying a replacement rule or a correction rule.

[0077] According to one embodiment, the automatic classification generation unit (120) can generate multiple candidates and assign priorities. The priority can be calculated based on a combination of rule suitability, frequency of selection from the same past business partner, standard journal entry statistics by industry, and text similarity scores, and can be configured to determine one result by applying a predefined decision rule when priorities are the same.

[0078] Additionally, the automatic classification generation unit (120) can evaluate criteria such as whether a tax invoice has been received, the type of receipt, the nature of the transaction, the item classification, and the limit regulations to determine whether VAT deduction is possible. The evaluation result may be indicated as deductible, non-deductible, or pending, and the pending status may be transmitted as a subject for verification during the subsequent quality inspection stage.

[0079] Additionally, the automatic classification generation unit (120) may be configured to use a mapping table with an account code system, a VAT code system, and a business partner code system to generate accounting processing codes. The code generator may specify the account codes for debit and credit and produce a voucher item including an amount allocation, a tax allocation, subsidiary ledger attributes, a tax classification code, and a supporting document classification. The voucher item may include a reference identifier and input source tracking information.

[0080] In one embodiment, the automatic classification generation unit (120) can perform mutual verification between the code value or OCR result extracted from the scan of the voucher image and the structured data. The mutual verification can be configured to check whether the approval number, transaction time, amount, and merchant identifier match, and to apply correction rules for discrepancies within a certain range.

[0081] Additionally, the automatic classification generation unit (120) may include a processing path for missing or incomplete data. In the missing data processing path, one or more procedures may be applied, such as filling in default values, referencing recent journal entries of the same past business partner, referencing industry average journal entries, and text-based estimation, and the applied procedures and grounds may be configured to be recorded as metadata.

[0082] Additionally, the automatic classification generation unit (120) may include additional metadata for linking to subsequent steps when outputting. The additional metadata may include a classification rule identifier, the number of applied rules, whether a conflict is detected, the number of candidates, the score interval between candidates, a text similarity index, and a verification result between the image and the structured data, and the metadata may be transmitted to the step-by-step reliability measurement unit (130) for reference in calculating the update reliability.

[0083] According to one embodiment, the automatic classification generation unit (120) can automatically generate an initial mapping for a new business partner or a new industry code by periodically synchronizing the accounting data classification database by business. The initial mapping is generated by referring to corporate information and a tax service code table collected from an external server and can be used with limited authority until approved by an operator.

[0084] Additionally, the automatic classification generation unit (120) may be configured to include timeout management, retry policy, partial failure isolation, and rollback policy for performance and error response in an operating environment. The transaction unit may be set to a voucher row unit or a transaction group unit, and journaling may be performed to enable recovery in the event of failure.

[0085] Additionally, the automatic classification generation unit (120) may be configured to preserve a record of input data snapshots, applied rule identifiers, model versions, and mappings with output codes to ensure auditability. The audit record is stored in an immutable repository and can be used for post-verification and quality inspection reproduction.

[0086]

[0087] The step-by-step reliability measurement unit (130) may be configured to receive the calculation results for each processing section of the automatic classification generation unit (120) and metadata added during the data collection stage as input, and to update and store the reliability status value. Here, the reliability status value is a value that quantitatively indicates whether the calculation results for each section are appropriate for reflection in the training dataset, and may be expressed within a predefined range.

[0088] More specifically, the step-by-step reliability measurement unit (130) may include an input integrator, a feature vector generator, a prior information estimator, an uncertainty estimator, an update engine, a corrector, and a state storage unit.

[0089] The input integrator can be configured to combine the classification results provided by the automatic classification generation unit (120), the rule evaluation log, the text similarity indicator, the code matching result, the multiple candidate score, the missing data processing history, and the collection path, parsing reliability, OCR reliability, decoding success status, and transmission integrity indicator provided by the data collection unit (110).

[0090] The feature vector generator can construct a vector that includes business characteristics such as transaction amount, VAT type, whether account category has changed, transaction type, and the occurrence of predefined keywords, collection quality indicators, score intervals between candidates, and the degree of discrepancy between rules and models.

[0091] Additionally, the step-by-step reliability measurement unit (130) may be configured to manage masking variables and randomness estimates together to reflect the effects of missing and incomplete data.

[0092] Masking variables indicate validity by field, and randomness estimates can be used to estimate the range of result variation based on substitution conditions. The prior information estimator can be configured to calculate interval-specific prior confidence from business partners, industry codes, tax service codes, and the distribution of previous year's journal entry history. The uncertainty estimator can be configured to aggregate the entropy of input features, score intervals between candidates, and indicators of discrepancy between rule-based and model-based results.

[0093] According to one embodiment, the step-by-step reliability measurement unit (130) can combine the reliability state value stored in the previous section through a transformation that guarantees monotonic increase or decrease, and calculate the reliability of the current section by considering prior information, the likelihood large-scale indicator, uncertainty penalty, and discrepancy penalty together. The corrector can be configured to stabilize the output range by applying temperature scaling or a calibration function.

[0094] In addition, in one embodiment, the step reliability measurement unit (130) can calculate the reliability of the current section through the following mathematical formula 1.

[0095] [Mathematical Formula 1]

[0096]

[0097] Here, is the current interval confidence, and is the previously stored confidence level. Also, is a logistic function, and is its inverse function. can be a value of a quantity that is pre-set as a temperature correction constant. is the prior confidence level calculated from the business partner, industry code, and previous year's journal entry history. is the feature vector of the current interval, and is a field validation mask. can be used as a term for combining at least one of the rule evaluation result, the correction probability of the learning model, the text similarity score, and the code matching score in the logarithmic domain as a large-scale likelihood indicator for a specific. is a predefined or learned non-negative weight It can be normalized to satisfy the conditions. It can be calculated by combining field-specific entropy and masking information as an uncertainty function due to missing values ​​and incompleteness. is a randomness indicator that estimates result variation based on replacement conditions, and is an indicator representing the degree of discrepancy between rule-based results and model-based results.

[0098] These can be set according to the policy as weighting factors for the previous reliability combination, uncertainty penalty, randomness penalty, and inconsistency penalty, respectively. is as a gate function, It can be defined as such and serves to determine the mixing ratio between the current interval calculation value and the previous confidence level.

[0099] Additionally, Equation 1 represents an example of a pre-set combination function and penalty function, and the log-domain combination can be replaced with one or more of the additive, multiplicative, logistic mixed, or normalized exponential combinations. The likelihood big metric can be defined as the model output to which flat scaling, isotonic correction, and tempering have been applied. The uncertainty function and discrepancy metric can be implemented as one or more of entropy, variance, score intervals between candidates, and prediction consistency metrics. The temperature constant and weighting factors can be adjusted according to the operational policy or training procedure.

[0100] In addition, the step-by-step reliability measurement unit (130) is updated It can be configured to be stored in a state storage unit and provided along with a reliability history so that the quality inspection unit (140) and the AI ​​labeling adjustment unit (150) can refer to it. The history provided may include a range identifier, the version of the combined function and correction function used, a summary indicator of the feature vector, and the contribution of each penalty item.

[0101] According to one embodiment, the step-by-step reliability measurement unit (130) may generate a warning flag when uncertainty or discrepancy exceeding a threshold condition is detected and mark it as a candidate for re-inspection to the quality inspection unit (140). Additionally, if there are many missing values, the influence ratio of the previous reliability may be temporarily amplified through a gate function.

[0102] Additionally, the step-by-step reliability measurement unit (130) can periodically re-evaluate prior information and weighting factors to respond to changes in distribution during long-term operation. The re-evaluation can be performed using statistics of recently confirmed journal entries and verified data, and the re-evaluation results can be configured to be recorded along with the version to ensure traceability.

[0103]

[0104] The quality inspection unit (140) can be configured to receive generated data containing step-by-step reliability as input, perform an inspection procedure, calculate a quality inspection score, and provide it to a subsequent stage. Here, the quality inspection score is a quantitative indicator used for determining whether to reflect the training dataset and for label adjustment criteria, and can be expressed within a predefined range.

[0105] More specifically, the quality inspection unit (140) may include an inspection planner, an inspector profiler, a full inspection engine, a partial inspection engine, a score calibrator, a result integrator, an inconsistency analyzer, a weighting corrector, a re-inspection decision unit, and a history storage unit. The inspection planner may be configured to determine whether to perform full inspection and partial inspection and the weight thereof based on the type and importance of input data, the reliability of each stage, and past inspection history.

[0106] According to one embodiment, an inspector profiler may be configured to generate a set of weighting factors by querying the inspector's proficiency, field of expertise, and predefined status classification. Proficiency may be calculated based on inspection grades or past inspection accuracy, field-specific weights may be set based on field-specific past quality inspection accuracy or difficulty coefficients, and status weights may be mapped according to predefined physical and mental status classifications.

[0107] Additionally, the full verification engine may be configured to evaluate a set of condition-based rules to check for item-specific matching, amount alignment, tax rate code validity, and the validity of combinations of account names and VAT deductions. The partial verification engine may be configured to perform item-level or range-level verification on variation samples generated by a data augmentation method. One or more of text synonym substitution, generation of amount boundary adjacent cases, variation of document image resolution, and keyword masking may be used as the augmentation method.

[0108] Additionally, the result integrator can be configured to collect the distribution of total inspection scores and partial inspection scores and perform scale alignment through a score calibrator. The discrepancy analyzer can be set to calculate the discrepancy ratio between total and partial inspections and record discrepancy patterns by item group. The weighting corrector can be configured to correct the final quality inspection score by combining inspector weighting factors and the step-by-step reliability of the input data.

[0109] In addition, in one embodiment, the quality inspection unit (140) can calculate the quality inspection score for the current data through the following mathematical formula 2.

[0110] [Mathematical Formula 2]

[0111]

[0112] Here, is the quality inspection score. is the condition-based overall inspection score. is the partial verification score for the k-th augmented sample. is a calibration function for partial inspection scores, for example, can be used. and It can be set as a weighting factor based on proficiency, according to the inspection grade or past inspection accuracy. is the current interval reliability provided by the step-by-step reliability measurement unit (130), and is a coefficient representing the bond strength. is a weighting factor by specialized field, and is a state weighting factor based on a predefined state classification. is a function that indicates the degree of discrepancy between the full inspection and the partial inspection, for example, It can be defined as. is the entropy-based uncertainty function of the partial inspection score distribution, and is an outlier indicator such as the number of rule violation items or the ratio of items exceeding the boundary. is a temperature correction constant, is a non-negative coefficient set according to the policy. is the logistic function, is its inverse function. is an operation that limits the output range to a predefined interval, and is a positive constant for numerical stabilization.

[0113] In addition, Equation 2 represents an example of a weighted combination and a penalty combination, and the log-domain combination can be replaced with an additive combination or a normalized exponential combination. The calibration function can be implemented using isotonic correction, flat scaling, or tempering methods, and the specific forms of the discrepancy function and uncertainty function can be changed according to item importance and operational policy.

[0114] Additionally, the quality inspection unit (140) may be configured to determine whether to re-inspect by performing a threshold comparison on the calculated Q. The discrepancy analyzer transmits a flag to the re-inspection decision unit if D exceeds a preset allowable range, and the re-inspection decision unit may be configured to register the re-inspection target in a queue. Re-inspection may be performed as either re-inspection by the same inspector or cross-inspection.

[0115] In one embodiment, the quality inspection unit (140) may be configured to record detailed grounds along with the quality inspection score in the history storage unit. The recorded items may include the total inspection score, a set of partial inspection scores, calibration parameters, inconsistency indicators, uncertainty indicators, a list of rule violations, inspector weighting factors, and version information of the coefficients used. The recorded history may be used for label adjustment in the AI ​​labeling adjustment unit (150) and for subsequent audit verification.

[0116] Additionally, the quality inspection unit (140) may be configured to support the simultaneous execution of full inspection and partial inspection, taking into account throughput and latency in the operating environment. When executing simultaneously, resource allocation policies and priorities may be applied, and parallel processing units may be divided by item group. The score calibrator is applied to both batch processing and real-time processing, and the parameter synchronization cycle may be set according to the processing mode.

[0117]

[0118] The AI ​​labeling adjustment unit (150) may be configured to combine reliability-related indicators input from the step-by-step reliability measurement unit (130) and the quality inspection unit (140) to maintain, change, or delete labels of the training dataset, and to record the change history so as to correct the initial reliability value during reprocessing. Here, the labels may include all the map information used for training, such as account categories, whether VAT is deductible, VAT type, tax rate code, item code, and voucher type.

[0119] More specifically, the AI ​​labeling adjustment unit (150) may be composed of an indicator aggregator, a comprehensive reliability calculator, a judgment engine, a label editor, a metadata manager, a dataset writer, a rollback manager, and an export interface. The indicator aggregator may be configured to receive as input a step reliability history, current interval reliability, quality inspection score, inconsistency indicator, rule violation list, and collection quality metadata.

[0120] In addition, according to one embodiment, the comprehensive reliability calculator can generate a single comprehensive reliability value by combining input indicators according to a predefined combination rule or correction function. When combining, it may be configured to ensure reproducibility by recording the latest version of each indicator and the version of the correction parameter used together. The calculated comprehensive reliability can be provided as a judgment criterion used in the recommended reliability flowchart of FIG. 7.

[0121] Additionally, the judgment engine may be configured to determine one of label retention, label change, or label deletion by comparing the overall reliability with a preset allowable value. The allowable value may be set differently for each label item group, such as account category, value-added tax deduction status, and tax rate code, and may be set to be applied sequentially according to a priority table in the event of a conflict in priority between item groups. If the judgment result is determined to be pending, it may be routed to an inspection queue and linked to the re-inspection procedure of the quality inspection department (140).

[0122] In one embodiment, the label editor may be configured to directly modify or delete labels based on the results of the decision engine. When changing, the existing label, the changed label, the code for the reason for the change, the versions of the decision rule and correction function used, and the referenced quality inspection basis may be recorded together. When deleting, a soft deletion method may be applied that leaves the code for the reason for deletion and the reference history, and marks the labels to be excluded only from the training targets in the dataset.

[0123] Additionally, the metadata manager may be configured to persist the label change history as metadata for the dataset. The metadata may include a record identifier, processing time, overall reliability, judgment result, threshold used, quality inspection score, stage reliability summary indicator, operator intervention status, transaction identifier, pipeline stage number, and schema version. When the same data is reprocessed, the metadata manager may query the cumulative change history and the latest quality inspection result to send a request to the stage reliability measurement unit (130) to correct the initial reliability value.

[0124] Additionally, the dataset writer can be configured to record the results of labeling to the training dataset storage. The recording method may support one or more of snapshot versions and incremental commits, and each commit may include the dataset version, coverage, number of changes, number of failures, and number of warnings. During recording, the source feature data, label data, and metadata can be stored separately to enable selective retraining and analysis.

[0125] According to one embodiment, the rollback manager may support rollbacks at the change commit level. It may be configured to restore to a labeled state prior to a specified version upon a rollback request, and to record the reason for the restoration and the scope of impact in the change history. The rollback may be performed within the same pipeline or a separate audit interface and may be configured to be completed after data integrity verification.

[0126] Additionally, the AI ​​labeling adjustment unit (150) can be operated using a combination of batch processing and real-time processing. Real-time processing is applied to transactions requiring event-based changes, and batch processing can be applied to periodic bulk data updates. Parameter synchronization between the two processing modes can be managed by pipeline version control.

[0127] Additionally, in one embodiment, the AI ​​labeling adjustment unit (150) may include a semantic constraint checker for verifying the consistency of the labels. The semantic constraint checker may be configured to check the validity of combinations of account names and whether value-added tax is deductible, consistency between tax rate codes and tax types, and compatibility between item codes and industry codes, and to generate an automatic correction candidate or switch to a hold in case of violation.

[0128] Additionally, the AI ​​labeling adjustment unit (150) can perform label propagation for related records through relationship propagation rules. When defined connection rules are satisfied, such as identical transaction bundles, identical voucher row groups, and identical merchants within the same period, the changed label can be propagated within a limited range, and the change history and basis can be set to be recorded even during propagation.

[0129] In one embodiment, the AI ​​labeling adjustment unit (150) may include a conflict resolution policy. If a rule-based result and a historical statistics-based result conflict, a procedure to apply priority or hold as a neutral label may be performed according to the policy. Item groups that frequently conflict may be designated as targets with a high priority for re-inspection by the quality inspection unit (140).

[0130] Additionally, the AI ​​labeling coordination unit (150) may include access control and an audit trail to ensure compliance with data governance. Access control may be configured with role-based permissions, and the audit trail may be configured to record the subject requesting the label change, the approver, and the processing result in an immutable repository. Label items containing personally identifiable information may be pseudonymized or tokenized according to policy.

[0131] Additionally, according to one embodiment, the AI ​​labeling coordination unit (150) may include an export interface for linking with a training pipeline. The export interface may be configured to generate a training snapshot by specifying a specific dataset version and a label version, and to provide a schema, a feature dictionary, a label catalog, and a quality metric summary together. The export result may be configured to be linked with the model training history to enable reproducible training.

[0132] Additionally, the AI ​​labeling coordination unit (150) can parallelize the processing pipeline for performance and scalability. The judgment engine and the label editor can be divided into item groups or partition keys and executed in parallel, and the bundle units requiring consistency can be configured to be processed sequentially within transaction boundaries. In the event of a failure, partial failure isolation and retry policies can be applied.

[0133]

[0134] The data output unit (160) may be configured to receive outputs generated by the step-by-step reliability measurement unit (130), the quality inspection unit (140), and the AI ​​labeling adjustment unit (150), and to perform storage, visualization, and external provision. Here, the outputs may include generated data with assigned reliability, an adjusted AI training dataset, a summary of quality inspection results, processing history, and metadata.

[0135] More specifically, the data output unit (160) may be composed of an output orchestrator, a serializer, a repository manager, a transmission module, an access controller, a version manager, a visualization renderer, and an audit logger. The output orchestrator may be configured to determine the type and priority of outputs transmitted from the upper module and to determine the target repository, file format, and transmission path.

[0136] According to one embodiment, the serializer may be configured to support one or more formats among CSV, JSON Lines, Parquet, and Avro for structured data. Additionally, it may be configured to ensure write performance and query efficiency by applying compression and partitioning during large-scale data processing. One or more of the collection date, operator identifier, data type, and dataset version may be used as partitioning criteria.

[0137] Additionally, the repository manager can operate by separating the online repository for short-term operations from the archive repository for long-term preservation. The online repository provides high-speed access for immediate lookup and visualization, while the archive repository can be configured to preserve version snapshots and incremental commits. When archiving, hashes and manifests for integrity verification can be recorded together.

[0138] Additionally, the data output unit (160) may include a dataset version management system. The version manager may be configured to record the dataset identifier, schema version, pipeline version, and versions of the combined rules and correction parameters used together, so that reproducible restoration and export are possible in the same environment. If a rollback is required, it may be configured so that restoration to a specified version is performed atomically within the transaction boundary.

[0139] In one embodiment, the transport module may be configured to support one or more of HTTPS-based APIs, SFTP, object store uploads, message queue publishing, and webhook calls. Data provided externally may be accompanied by a metadata manifest, and the manifest may include a dataset range, number of records, field list, code scheme, confidence distribution summary, quality inspection summary, creation time, and time zone.

[0140] Additionally, the data output unit (160) may include a visualization renderer. The visualization renderer may be configured to display summary indicators, such as the distribution of account categories, the distribution of value-added tax types, the quality inspection discrepancy rate, the number of cases by overall reliability interval, and the number of cases by label change reason code, in a dashboard format. The visualization results may be provided in the form of an image or an interactive view and may be included together with a table when generating a report.

[0141] Additionally, access controllers can be configured to restrict data access and export requests by applying role-based permissions. Security in the transmission and storage sections can be configured to be maintained through encrypted channels and encrypted storage. For items containing personally identifiable information, pseudonymization, partial masking, and tokenization policies can be applied during the export phase, and recovery keys can be stored in a separate secure storage.

[0142] According to one embodiment, the data output unit (160) may include internationalization and time zone management functions. Numbers and currency notations, and date and time notations are serialized according to a specified locale and time zone, and file encoding may be configured to be unified into a standard character set to ensure interoperability.

[0143] Additionally, the data output unit (160) may include a commit procedure for quality assurance. After writing to a temporary path, visualization is performed through atomic movement, and the commit log may record the file path, byte size, hash, and creation time. In the event of a transmission failure, an exponential backoff retry is applied, and if the maximum number of retries is exceeded, it may be moved to a quarantine area and configured to send an operator notification.

[0144] Additionally, notifications can be issued upon the occurrence of critical conditions through the alert function. Alerts may be issued via defined channels if a decrease in overall reliability, an exceedance of the quality inspection inconsistency rate, an accumulation of transmission failures, or a dataset schema inconsistency is detected. The alert includes the scope of impact and recommended actions, which can be utilized for subsequent processing.

[0145] In one embodiment, the dataset export interface may be configured to support linkage with a training pipeline. Upon export, a training snapshot is generated, and the training, validation, and test splits, along with a label catalog, a feature dictionary, and a quality metric summary, may be provided. Re-exporting of the same version may be configured to be handled idempotently to maintain the same hash value.

[0146] Additionally, the data output unit (160) can perform audit logging and data lineage recording. The audit log records the requesting entity, approver, execution result, number of impact records, and reason for failure, and the lineage record may include the input source, applied rule, model and parameters used, and reference to the generated output. The records are sequentially stored in an immutable repository and can be utilized for compliance checks and post-analysis.

[0147] In addition, parallel writing and write buffers can be applied for performance and scalability. Output operations are parallelized at the partition level, and overload signals can be transmitted to upper modules through backpressure control. Operational metrics regarding throughput and latency are collected periodically and can be used for capacity planning.

[0148] According to one embodiment, the data output unit (160) can transmit the generated data with confidence and the adjusted AI training dataset to the next recipient.

[0149] For example, it may be provided to an AI learning pipeline server (300) in a batch or streaming manner. At this time, the data output unit (160) may transmit a manifest containing a dataset snapshot version, an incremental commit hash, a label catalog, and a summary of recommendation confidence and quality inspection scores, and may commit to a history repository upon receiving an acknowledgment response issued by the receiving side. The transmission protocol may be implemented as REST or gRPC, and for large-scale batch transmissions, a column-oriented format such as Parquet and a pre-agreed schema version may be applied. To prevent duplicate reflection, an idempotent key may be assigned to each transmission unit.

[0150] Additionally, in one embodiment, the data output unit (160) may transmit labels requiring operational reflection to an ERP or tax filing linkage server (400). At this time, the data output unit (160) may transmit only records confirmed to have a reliability above a threshold and a quality inspection pass status to the staging area for pre-operational reflection verification, and after performing a mapping table version match check, proceed with the actual reflection through a two-stage commit procedure. The reflection result may be returned as an item-specific mapping log along with a success code and recorded in the change history.

[0151] Additionally, in one embodiment, the data output unit (160) may be exported to a data lake and audit repository (500) for long-term storage. At this time, a ledger-style change history, policy version, and signature value may be stored together, and an immutable storage policy may be applied. If personal information or identification information is included, pseudonymization and encryption may be applied.

[0152] Additionally, in one embodiment, the data output unit (160) can maintain a routing policy for each recipient to simultaneously branch the same snapshot to an AI learning pipeline server (300), an ERP or tax filing linkage server (400), a data lake, and an audit repository (500), and in the event of transmission failure, exponential retries and checkpoint-based re-transmission may be performed. All communications include a hash value and a signature for verifying transmission integrity, and a version negotiation procedure with the receiving side schema registry may be performed beforehand.

[0153]

[0154] FIG. 3 is a schematic diagram illustrating the hardware configuration of an automatic tax accounting data step-by-step reliability measurement and AI labeling adjustment system (100) according to one embodiment of the present invention.

[0155] Referring to FIG. 3, the step-by-step reliability measurement and AI labeling adjustment system (100) for automatic tax accounting data according to embodiments is implemented in the form of a computing device including hardware (200) and may include memory (210), a processor (220), a communication module (230), and an input / output unit (240).

[0156] The memory (210) is a non-transient computer-readable recording medium and may include a permanent mass storage device such as RAM (random access memory), ROM (read only memory), disk drive, SSD (solid state drive), flash memory, etc. Here, the permanent mass storage device such as ROM, SSD, flash memory, disk drive, etc. may be included in the device or server described above as a separate permanent storage device distinct from the memory (210).

[0157] Additionally, the memory (210) may store an operating system and at least one program code (e.g., code for a security module or an application installed to provide a specific service). These software components may be loaded from a computer-readable recording medium separate from the memory (210). This separate computer-readable recording medium may include a computer-readable recording medium such as a floppy drive, disk, tape, DVD / CD-ROM drive, or memory card.

[0158] In another embodiment, software components may be loaded into memory (210) via a communication module (230) rather than a computer-readable recording medium. For example, at least one program may be loaded into memory (210) based on a computer program installed by files provided over a network by developers or a file distribution system (e.g., an application store service server) that distributes installation files for applications.

[0159] The processor (220) may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input / output operations. Instructions may be provided to the processor (220) by memory (210) or a communication module (230). For example, the processor (220) may be configured to execute instructions received according to program code stored in a recording device such as memory (210).

[0160] The communication module (230) can provide a function for communicating with a user terminal, etc., through a network. Additionally, the communication module (230) can provide a function for the screen providing system to communicate with one or more other devices through a wired or wireless network. That is, the communication module (230) is a part that realizes each function module described above with reference to FIG. 2 by controlling its function by a processor (220) that references memory (210).

[0161] The input / output unit (240) may be a means for interfacing with an external input / output device (not shown). For example, the external input device may include devices such as a keyboard, mouse, microphone, camera, etc., and the external output device may include devices such as a display, speaker, haptic feedback device, etc. As another example, the input / output unit (240) may be a means for interfacing with a device in which the functions for input and output are integrated into one, such as a touchscreen.

[0162] Additionally, in other embodiments, the automatic tax accounting data step-by-step reliability measurement and AI labeling adjustment system (100) may include more hardware components than those shown in FIG. 3 depending on the nature of the device to which it is applied. For example, it may be implemented to include at least some of the input / output devices described above, or it may include additional components such as a transceiver, a GPS (Global Positioning System) module, a camera, various sensors, a DB, etc. As a more specific example, it may be implemented to include various additional components such as an accelerometer or gyroscope sensor, a camera module, various physical buttons, buttons using a touch panel, input / output ports, and a vibrator for vibration.

[0163] However, the components and forms of the computing device described in this specification are merely exemplary, and the configuration of the computing device in which the step-by-step reliability measurement and AI labeling adjustment system (100) of automatic tax accounting data is implemented may differ from that described in this specification due to the adoption of other known technologies or future advancements in information and communication technology.

[0164]

[0165] Next, we will examine the step-by-step reliability measurement process according to one embodiment of the present invention.

[0166] FIG. 4 is a flowchart illustrating the step-by-step reliability measurement process according to one embodiment of the present invention.

[0167] Referring to Fig. 4, first, in step S401, a basic expected confidence level can be set based on the initial output of the automatic classification generation program. The basic expected confidence level is set to 95.0 percent, and the posterior confidence level is calculated when deductive and inductive instance verifications are passed. For example, an example formula can be derived as (95.0% × 99.9%) / ((95.0% × 99.9%) + (5% × 1%)) = 99.9%.

[0168] Next, in step S402, if the reliability of step S401 is less than 99.9 percent and the final account automatic classification and generation code have been generated, the update reliability can be recalculated. For example, the example formula can be derived as (95.0% × 99.9%) / ((95.0% × 99.9%) + (5% × 0.1%)) = 99.9%. Additionally, if the character composition of the raw data is not the native language, it can be corrected by reflecting the possibility of linking with the transaction amount. For example, the example formula can be derived as (95.0% × 50%) / ((95.0% × 50%) + (5% × 50.0% × (Transaction Amount - 300,000) / 300,000)) = ?%.

[0169] In addition, at step S403, the reliability can be updated by reflecting a specific code combination or a predefined text pattern. For example, if the account automatic classification code is 262 and the final lifestyle-based account change code is 338 and the reliability at step S402 is 99.9 percent, an example formula can be derived as, for example, (95.0% × 99.9%) / ((95.0% × 99.9%) + (5% × 0.1%)) = 99.9%.

[0170] Conversely, if a pattern similar to "DeepMind" or "Input($-B)" is detected in the A / C P'meter, the weights are adjusted. For example, the example formula can be derived as (95.0%×70.0%) / ((95.0%×70.0%)+(5%×30.0%×(Transaction Amount-300,000) / 300,000))=?%.

[0171] Additionally, in the S404 step, a conservative penalty may be applied if "other progress costs" are included in the A / C P'meter. For example, the example formula can be derived as follows: (95.0% × 60.0%) / ((95.0% × 60.0%) + (5% × 70.0%)) = 94.2%.

[0172] Additionally, in step S405, if a combination of strings related to meetings or preprocessed token patterns such as "deep" or "yeon" are included, the reliability can be calculated based on the transaction amount boundary. For example, the example formula can be derived as (95.0%×30.0%) / ((95.0%×30.0%)+(5%×80.0%×(transaction amount-100,000) / 100,000))=?%. Meanwhile, if it is a card transaction and the usage amount is 200,000 won or more, and there is a discrepancy between the slip-based pre-pattern detection and the actual account creation, for example, the example formula can be derived as (95.0%×80.0%) / ((95.0%×80.0%)+(5%×95.0%))=94.1%.

[0173] In addition, at the S406 stage, a correction penalty may be applied when the item combination fit is low under conditions where the S405 stage reliability is 99.9 percent or higher and the A / C P'meter is generated.

[0174] For example, the example formula can be derived as (95.0%×50.0%) / ((95.0%×50.0%)+(5%×90.0%))=91.3%. When applying different pattern matching under similar conditions, for example, the example formula can be derived as (95.0%×70.0%) / ((95.0%×70.0%)+(5%×80.0%))=94.3%.

[0175] Additionally, in step S407, if the reliability of step S406 is 99.9 percent or less and the amount used is less than 100,000 won, the adjustment range may be reduced. For example, the example formula can be derived as (95.0%×95.0%) / ((95.0%×95.0%)+(5%×5.0%))=99.7%. Conversely, if the amount used is 500,000 won or more, the penalty is increased according to the amount boundary function. For example, the example formula can be derived as (95.0%×60.0%) / ((95.0%×60.0%)+(5%×80.0%×(transaction amount-500,000) / 500,000))=?%.

[0176] Additionally, at step S408, if the reliability of step S407 is 99.9 percent or higher, a combination of an offline payment method and a specific payment gateway is verified, and the transaction amount exceeds 100,000 won, a conservative penalty may be applied. For example, an example formula can be derived as (95.0% × 50.0%) / ((95.0% × 50.0%) + (5% × 90.0%)) = 93.1%.

[0177] Additionally, in step S409, if a specific generated code and amount code or size condition is determined to be abnormal, the reliability can be significantly lowered. For example, if the generated code is 199 and the amount used is greater or less than 0, the example formula can be derived as (95.0%×1.0%) / ((95.0%×1.0%)+(5%×90.0%))=29.7%.

[0178] In addition, at step S410, different matching information, such as interest withholding tax items or duplicate data check results, can be reflected. For example, if the generation code is 901 and there is an interest withholding amount in the account source data, the example formula can be derived as (95.0%×5.0%) / ((95.0%×5.0%)+(5%×95.0%))=50.0%. In this case, even if the code is the same, if it is identified as a card transaction record, for example, the example formula can be derived as (95.0%×2.0%) / ((95.0%×2.0%)+(5%×80.0%))=32.2%.

[0179] Meanwhile, if the S409 step is 99.9 percent or higher and a conflict is detected only in the third check of the debit card versus account data duplication check, for example, the example formula can be derived as (95.0%×80.0%) / ((95.0%×80.0%)+(5%×90.0%))=94.4%.

[0180] At this time, for account data where the transaction amount is 3 million or more and less than 5 million, an amount boundary correction is added. For example, the example formula can be derived as (95.0%×90.0%) / ((95.0%×90.0%)+(5%×90.1%×(transaction amount-5,000,000) / 5,000,000))=?%. Additionally, if a specific data box parameter is generated and the generation code is 731, 831, 691, or 781, 0.8 percent may be subtracted from the S409 step value.

[0181] Additionally, in step S411, sub-engines and internet search or chatbot-based automatic extraction results can be combined as weighting factors. For example, if the confidence level in step S410 is 95.0 percent, a warning indicator such as "95.0%_[Caution: multi-BizField]" can be added. Reinforcement combination is applied if it corresponds to a specific code group. For example, an example formula can be derived as (95.0%×90.0%) / ((95.0%×90.0%)+(5%×10.0%))=99.4%.

[0182] Additionally, in step S412, if the transaction amount is 70,000 won or more and crawling or chatbot output characteristic characters and keywords related to children or pets in A / C_P'meter are simultaneously identified, it can be indicated that the account is extracted based on a recommendation and the reliability can be calculated. For example, the result may be displayed as "97.1%_[Caution: Recommendation A / C extraction by crawling search_multi-BizField]".

[0183] Additionally, in step S413, the aggregated results can be organized to determine the reference confidence level. For example, if the confidence level in step S410 is 95.0 percent, it is maintained; otherwise, it can be replaced with the confidence level in step S412. If the existing independent machine's automatically generated code passes validation, reinforced coupling is applied. For example, the example formula can be derived as (95.0% × 98.0%) / ((95.0% × 98.0%) + (5% × 0.1%)) = 99.9%.

[0184] Finally, in Step S414, if the reliability of Step S413 is below the baseline reliability, the account can be automatically converted to a safe account and VAT can be set to non-deductible. For example, an advance or withdrawal account can be designated, and an adjustment rule disallowing VAT deduction can be applied.

[0185]

[0186] Next, we will examine the quality inspection process according to one embodiment of the present invention step by step.

[0187] FIG. 5 is a flowchart illustrating the quality inspection process according to one embodiment of the present invention in steps.

[0188] Referring to FIG. 5, first, in step S501, the quality inspection unit (140) can receive generated data including step-by-step reliability and perform a preliminary verification. The input may include raw metadata provided by the data collection unit, classification results and code mapping from the automatic classification generation unit, interval reliability history from the step-by-step reliability measurement unit, and a voucher image reference.

[0189] In addition, the existence of required fields, data types and formats, consistency of currency and tax rates, matching of total amounts, supply prices, and tax amounts, alignment of time zones and timestamps, and referential integrity of merchant identifiers and business partner codes can be checked. In one embodiment, a verification session identifier is generated for each input record, and reproducibility can be ensured by recording the input hash, transmission path, and schema version together. In the event of a verification failure, the failure type code, field offset, and reprocessing policy are attached, and the process is branched to a quarantine queue; in the event of a success, the process proceeds to the inspection plan stage.

[0190] Next, in step S502, the quality inspection department (140) may establish an inspection plan. The plan may include target data bundles, the proportion of full inspection and partial inspection, the number of augmented samples for partial inspection, thresholds and warning criteria, and resource allocations.

[0191] Next, a risk score can be calculated using step-by-step reliability, input quality indicators, and past discrepancy patterns, and the partial inspection weight can be expanded for bundles with high risk scores. In one embodiment, the plan is finalized on a bundle basis by applying rules for identical transaction bundles, identical voucher row groups, and identical merchants and periods, and the augmentation seed and plan version are fixed to minimize variation in results during repeated execution.

[0192] Next, in step S503, the quality inspection department (140) may assign an inspector and look up the inspector profile. The profile may include a proficiency weighting factor based on inspection grade or historical accuracy, a field weighting factor based on cumulative accuracy by label item group, and a status weighting factor corresponding to a predefined status code.

[0193] According to one embodiment, reviewers detected to have repetitive bias or conflicts of interest within the same item group may be automatically excluded and configured to be transferred to a cross-review queue if available personnel are insufficient. The assignment result is recorded along with the session identifier and can be used as a weighting factor in a subsequent correction step.

[0194] Next, in step S504, the quality inspection department (140) can perform a condition-based overall inspection. The overall inspection can evaluate the amount matching, the valid combination of tax type and tax rate code, the mutual constraints between account items and whether VAT is deductible, the proof requirements for each transaction type, and the validity of cross-reference between code systems using a set of rules.

[0195] In one embodiment, for each rule, a pass or fail, a raw score and a list of violation items, and a referenced code table version are calculated, and the raw score and grounds at the record level can be stored. The violation items include field paths and expected value ranges so that they can be quickly reproduced upon re-examination.

[0196] Next, in step S505, the quality inspection unit (140) may generate augmentation samples for partial inspection. Augmentation may consist of one or more of text synonym substitution, keyword masking, generation of amount boundary adjacent cases, and variation in voucher image resolution. In one embodiment, amount boundary augmentation may generate bidirectional deviation samples centered on amounts close to deduction limits or tax rate conversion boundaries, and text augmentation may be set to preserve the presence and location of key keywords after stop word removal. Additionally, each augmentation sample may have a corresponding key with the original, a conversion history, and a fixed random number seed recorded to ensure comparability.

[0197] Next, in step S506, the quality inspection unit (140) may perform partial inspections for each augmented sample. Partial inspections consist of item-unit or segment-unit checks, and raw scores per sample, a list of triggered rules, and auxiliary indicators such as image recognition or text parser reliability may be collected together. In one embodiment, retries within a limit number of times may be performed on samples that failed to process, and in the event of repeated failures, the cause of failure may be marked with an exclusion mark and reflected in the denominator correction. When executing in parallel, the order may be guaranteed using a sample partition key.

[0198] Next, in step S507, the quality inspection department (140) may perform score calibration. Isotonic correction or temperature scaling may be applied to correct the scale difference between the total inspection score and the partial inspection score distribution.

[0199] According to one embodiment, calibration parameters are periodically retrained on a past confirmed dataset, and when applied, the parameter version and the training dataset version may be recorded together. In addition, distribution truncation and cumulative distribution function-based mapping may be performed in parallel to reduce outlier sensitivity.

[0200] Next, in step S508, the quality inspection unit (140) can perform an inconsistency analysis. The average difference between the total inspection score and the partial inspection score, the variance, the entropy, the inconsistency rate by item group, and the sensitivity by augmentation type can be calculated. According to one embodiment, if the inconsistency rate in a specific item group exceeds a threshold, a warning flag can be set, and detailed rules of the rule set or text parser dictionary can be presented as a candidate list to be marked as targets for future improvement. Additionally, frequent confusion pairs can be extracted and passed to the conflict resolution policy of the AI ​​labeling coordination unit.

[0201] Next, in step S509, the quality inspection unit (140) can calculate a quality inspection score by combining the inspector weighting factor and the interval reliability provided by the step reliability measurement unit (130). The calculation may be performed using the previously presented mathematical formula 2, and after calculation, clipping and rounding policies may be applied to normalize the result to an interval [0, 1] or a percentage scale.

[0202] In one embodiment, contribution decomposition is calculated to store the contribution rates of total inspection, partial inspection, discrepancy penalty, inspector weighting, and step-by-step reliability combination items together, and a re-evaluation trigger can be set when excessive variance of scores within the same group is detected.

[0203] Next, in step S510, the quality inspection unit (140) may perform a re-inspection decision. The criteria may be set as when the quality inspection score is below a threshold or when the discrepancy indicator exceeds an allowable range. According to one embodiment, a choice is made between re-inspection by the same inspector and cross-inspection, and in the case of cross-inspection, an inspector with a high weighting in the same item group's specialized field may be assigned first. Additionally, a limit on the number of re-inspections and a loop prevention rule may be applied to prevent infinite recirculation of the same case.

[0204] Next, in step S511, the quality inspection unit (140) records the results and grounds in the history storage unit and, if necessary, transmits them to the data output unit (160). The recorded items may include the total inspection score, the set of partial inspection scores and calibration parameters, the discrepancy indicator, the list of rule violations, the inspector weighting factor, the version of the rule and correction function used, the session identifier, and the input hash. In one embodiment, they may be sequentially stored in an immutable storage, and the referential integrity of the records, voucher images, and logs may be periodically verified.

[0205] Next, in step S512, the quality inspection unit (140) transmits the quality inspection score, warning flag, and item group inconsistency pattern summary to the AI ​​labeling adjustment unit (150), and may provide input quality indicator summary and rule sensitivity information as feedback to the step-by-step reliability measurement unit (130). According to one embodiment, the feedback may include a request for prior information correction in a subsequent section and a suggestion for gate function adjustment. Additionally, the state of highly sensitive rules may be managed so that they are changed only after supervisor approval.

[0206] Next, in step S513, the quality inspection unit (140) may perform an exception handling branch. If an exception is detected, such as input validation failure, absence of calibration parameters, processing timeout, or unavailable external reference table, it may be isolated in a quarantine queue, and the cause code, reprocessing policy, and retry result may be recorded. In one embodiment, a retry index backoff, an alternative path switching, and a rule prohibiting partial result reflection may be applied step by step. When the exception is resolved, it returns to the original step or branches to a re-inspection path.

[0207] Finally, in step S514, the quality inspection unit (140) can generate a quality inspection summary and transmit it to the output interface of the data output unit (160). The summary may include a score distribution, a discrepancy rate, a re-inspection rate, a top list of major violation items, rule and parameter versions, a dataset version, a processing time, and a time zone.

[0208] According to one embodiment, the summary is provided as both a dashboard and a file snapshot, and when provided externally, a metadata manifest and an integrity hash may be transmitted together. Additionally, warning items within the summary may include links to allow operators to immediately view the re-examination queue and the rule improvement request queue.

[0209]

[0210] Next, we will examine the procedure for calculating and applying recommendation reliability in the AI ​​labeling adjustment process according to one embodiment of the present invention step by step.

[0211] FIG. 6 is a flowchart illustrating the step-by-step process of calculating and applying recommendation reliability in an AI labeling adjustment process according to an embodiment of the present invention.

[0212] Referring to FIG. 6, first, in step S601, the AI ​​labeling adjustment unit (150) can perform recommendation reliability calculation.

[0213] The input may include the interval reliability history and current reliability of the step-by-step reliability measurement unit (130), the quality inspection score and inconsistency indicator and rule violation list of the quality inspection unit (140), and the collected quality metadata and classification basis summary provided by the data collection unit (110) and the automatic classification generation unit (120).

[0214] Additionally, threshold tables by item group, weighted combination policies, and correction parameters can be loaded together. Recommendation confidence can be calculated after adjusting weights by aligning the scales of input metrics and reflecting missing masks. Scale alignment can be applied based on a correction table periodically retrained from past confirmed datasets. If missing items exist, the blending ratio can be adjusted by utilizing previous confidence and quality inspection distribution summaries as gates.

[0215] In addition, conservative penalties may be applied upfront to high-impact item groups, such as account categories and VAT deduction eligibility. The versions of the correction parameters used, threshold set versions, and input summaries are recorded along with the calculated recommendation confidence to support subsequent reproduction and auditing.

[0216] Next, in step S602, the AI ​​labeling adjustment unit (150) can determine the application of recommendation reliability and labeling behavior.

[0217] One of the following is selected by referring to the threshold table by item group: maintain, change, or delete the label. If it is determined to maintain, propagation within the restricted range may be performed for relationship groups defined as the same transaction group or the same document row group.

[0218] In the case where it is determined to be a change, a candidate with a higher priority that combines rule suitability and past same business partner selection statistics can be proposed from among multiple candidates calculated by the automatic classification generation unit (120). After the validity of the combination of account item and value-added tax type is re-verified through a semantic constraint checker, the journal entry can be updated.

[0219] If it is determined to be deleted or held in abeyance, a safety account and non-deduction settings may be applied to ensure the safety of accounting processing. The case may be routed to the re-inspection queue of the Quality Inspection Department (140) for cross-inspection or supplementary verification.

[0220] According to one embodiment, a bundled transaction boundary may be set to prevent irreversible changes during propagation. The schema version and code system version of the records to be propagated are matched through a pre-check and then applied; if a conflict is detected, it is switched to a hold and can be returned to the re-check path. All decisions are provided with a policy version and the basis so that they can be managed to enable subsequent comparative review.

[0221] Finally, in step S603, the AI ​​labeling adjustment unit (150) can perform result persistence and feedback delivery. At this time, the changed labels and the maintained labels are distinguished and committed to the dataset writer, and a snapshot version and an incremental commit hash can be generated. The commit metadata may include the recommendation confidence RRR, applied threshold, branch path, correction parameters and rule versions used, conflict resolution results, relationship propagation range, and whether operator intervention occurred.

[0222] Additionally, the label change history is persisted in the metadata area, and a reference pointer can be generated so that the step-by-step reliability measurement unit (130) can correct the initial reliability value when the same data is reprocessed. Item groups with high inconsistencies or rule combinations with frequent conflicts are summarized and transmitted to the quality inspection unit (140) to be used as a basis for weighted adjustment of the next inspection plan.

[0223] Additionally, an export request is issued to the data output unit (160) for external learning pipeline linkage, and a specified dataset version, label catalog, and quality metric summary may be delivered together. Exceptionally, if a schema mismatch or storage lock is detected, the rollback manager restores to the previous stable snapshot, and the case is marked as pending and can be registered in the reprocessing queue.

[0224]

[0225] Next, with reference to FIGS. 7 to 9, we will examine the quality inspection UI flow of the present invention based on example screens.

[0226] FIG. 7 is a drawing illustrating an example screen of an inspector's level and status evaluation according to an embodiment of the present invention, FIG. 8 is a drawing illustrating an example screen of an inspection result display by quality inspection item according to an embodiment of the present invention, and FIG. 9 is a drawing illustrating an example screen of statistics and comparison of quality inspection results according to an embodiment of the present invention.

[0227] First, referring to Fig. 7, the quality control inspector level is displayed on the left panel along with inspector identification information, and the status evaluation for the same inspector is displayed separately on the right panel.

[0228] The inspector's name is placed at the top of the screen, and the area to the left below contains a level evaluation table labeled P'm_101. The table may consist of a classification, level, and check column. The classification item displays, for example, top_tier, expert level, and staff, with level values ​​assigned to each item such as 1, 2, and 3, and the selected item can be activated via the check column. Additionally, this level value can be used as an integer code referenced by the quality inspection department when calculating correction weights.

[0229] The right area includes a condition evaluation table labeled P'm_102. The table displays physical and mental conditions separately in distinct column groups, and a level and check value can be entered for each condition. In one embodiment, three levels of good, average, and poor are provided, and a selected level may be indicated as 1 in the check column, while an unselected level may be indicated as 0. Additionally, the condition evaluation is stored along with a time stamp within the same session and can be utilized as a condition weighting factor in the score calibration and quality inspection score correction steps described in FIG. 5.

[0230] The level code and status code entered on this screen are persisted as part of the inspector profile and can also be used as a basis for priority adjustment when determining cross-inspection assignment logic and re-inspection. According to one embodiment, the combination of level and status may be configured such that the allowed range is predefined based on the policy version, and a warning notification is displayed if a combination outside the range is entered.

[0231] Referring to FIG. 8, in one embodiment, three inspection tracks are arranged in parallel at the top of the screen, and the results of the aggregation step are displayed as they move toward the right. The areas of Epoch Quality Inspection_3 (data augmentation-based), Epoch Quality Inspection_2 (one-point manual work), and Total Quality Inspection_1 (condition algorithm-based) are arranged in order from left to right.

[0232] The table in each area includes accounting item columns such as customer name, supply price, tax amount, service charge, total amount, VAT deduction status, VAT type, account code, and item code, and the inspection results of individual vouchers can be displayed on a row-by-row basis. In one embodiment, cells with finalized calculation results are displayed as having an unmodifiable attribute, while cells where an inspector can verify supporting evidence or suggest modifications are displayed as being in an input-enabled state. For example, if a code such as 812 is displayed in the account code column of a specific row, this means that the result mapped by the automatic classification generation unit has been reflected identically for each quality inspection track.

[0233] Additionally, in one embodiment, a pass and a violation flag are stored internally in the row of each area, and the flag combination is utilized in the discrepancy analysis step of FIG. 5. In the augmentation-based track, text synonym substitution or sensitivity in samples close to monetary boundaries may be recorded together on a row basis, and in the one-point manual track, reviewer comments and links to evidence may be attached. The total track on the right aggregates the results of two preceding tracks to display the final judgment for each item, and the aggregation logic and parameter version may be stored together as metadata for each row.

[0234] In addition, in one embodiment, summary indicators such as the total of the current page, the number of violations, and the distribution of deduction status may be displayed on the status bar at the bottom of the screen. Furthermore, a batch processing function for the same transaction group or the same voucher row group may be provided, so as to enable the consistent application of maintenance, modification, and hold.

[0235] Referring to FIG. 9, in one embodiment, the Epoch quality inspection results and the total quality inspection results are displayed side by side on the left side of the screen, normalized to the same coordinate system, and a summary of the unit pipeline one-point review of the dataset may be placed in the center. In the quality inspection management output area on the right, the final confirmed results by item may be organized and provided in a table format.

[0236] Additionally, the statistics panel provides, for example, violation rates by item group, discrepancies, re-inspection rates, and inspector weighted factor distributions, and can be configured to enable partial aggregation through period filters and vendor filters. In one embodiment, a pivot table is automatically generated with axes for deduction status, VAT type, account code, and item code, and the number of cases and the rate can be displayed together in each intersecting cell. At the bottom of the screen, policy-important indicators, such as the total non-deductions, can be highlighted and displayed, and the total value is persisted along with the dataset version and processing time.

[0237] Additionally, in one embodiment, the comparison panel may be configured to visually represent the difference between the Epoch track and the Total track, thereby enabling the identification of sections where performance degradation occurs in item groups with high rule sensitivity or in augmented samples. If the difference value exceeds a threshold range, a warning indicator is generated, and the corresponding section may be automatically registered as a re-examination plan. The central one-point summary panel displays whether the manual review results and the automatic aggregation results match, and if they fail to match, a conflict resolution policy may be proposed.

[0238] Additionally, in one embodiment, the right output area functions as a preview of the final output to be delivered to the data output unit. The column configuration is arranged, for example, by customer name, supply price, tax amount, service charge, total amount, VAT deduction status, VAT type, account code, and item code, and recommendation confidence and quality inspection scores may be added as metadata to each row. Additionally, upon an export request, a snapshot version and an integrity hash are generated, and whether to deliver to an external learning pipeline can be selected.

[0239]

[0240] Next, we will examine the entire procedure of the step-by-step reliability measurement and AI labeling adjustment method for automated tax accounting data according to one embodiment of the present invention in order.

[0241] FIG. 10 is a flowchart illustrating the overall procedure of a step-by-step reliability measurement and AI labeling adjustment method for automatic tax accounting data according to an embodiment of the present invention.

[0242] Referring to FIG. 10, first, in step S1001, data subject to accounting processing can be collected and standardized through the data collection unit (110). The data subject to collection may include card approval results, bank deposits and withdrawals, electronic tax invoices, payment slips, salary deduction data, etc.

[0243] Additionally, in one embodiment, schema mapping, currency and tax rate unit normalization, time zone unification, deduplication, and required field validation may be performed. Auxiliary metadata such as business partner codes, industry codes, and voucher image references may be combined, and record-specific collection paths, source hashes, and code scheme versions may be recorded as metadata.

[0244] Next, in step S1002, the summary, merchant name, item, and amount structure are parsed through the automatic classification generation unit (120), and the account category and whether value-added tax is deductible can be automatically classified by combining rule matching and embedding similarity matching. The basis for classification, candidate labels and scores, and a list of conflict candidates can be stored, and an accounting processing code can be generated according to the classification result. The mapping table version and the code conversion policy are listed together with the label to ensure compatibility between versions.

[0245] Next, in step S1003, the classification results from the previous step and auxiliary metadata are input through the step reliability measurement unit (130) to calculate the step reliability. The calculation may reflect the amount boundary sensitivity, VAT type consistency, account change history, transaction type and keyword appearance pattern, and external reference match degree. If there are missing items, a correction value with applied replacement conditions may be used, and the reliability history by time point is calculated together and provided as input for the subsequent quality inspection and label adjustment steps.

[0246] Next, in step S1004, a multi-stage quality inspection can be performed through the quality inspection unit (140) to produce a quality inspection result. First, a full inspection based on a set of rules is executed to evaluate the amount matching, tax rate code validity, combination constraints of account items and deduction status, and supporting documentation requirements.

[0247] In addition, in one embodiment, partial inspection may be performed on augmented samples generated by text synonym substitution, keyword masking, amount boundary adjacent cases, etc. The level code and status code of the inspector may be identified and corresponding weights applied so that the score may be corrected, and the total score, partial score distribution, discrepancy indicator, and list of rule violations may be recorded.

[0248] Next, in step S1005, the quality inspection result and the step reliability can be combined through the AI ​​labeling adjustment unit (150) to calculate the recommendation reliability, and one of maintaining, changing, or deleting the label can be determined according to the threshold policy for each item group.

[0249] If a change is required, among the multiple candidates generated by the automatic classification generation unit, the candidate with the highest rule fit and vendor selection statistics is selected, and only the result that passes the semantic constraint validator can be reflected in the journal entry. If the decision is to maintain, limited propagation may be performed within the same transaction group or the same journal entry row group; if the decision is to hold or delete, it may be routed to the re-examination queue. For all decisions, the policy version and rationale may be assigned as metadata.

[0250] Finally, in step S1006, the generated data with confidence and the adjusted AI training dataset can be stored or provided externally through the data output unit (160). A commit-unit snapshot version and an incremental commit hash can be generated, and the dataset version, label catalog, and quality metric summary can be managed together.

[0251] Additionally, in one embodiment, when exporting to an external learning pipeline, change history, recommendation confidence, and quality inspection score may be included as metadata, and in the event of an exception, the previous stable snapshot may be restored according to a rollback procedure and then placed in a reprocessing queue.

[0252]

[0253] As described above, the system and method for step-by-step reliability measurement and AI labeling adjustment of automatic tax accounting data according to the present invention can be configured to automatically perform the collection, classification, reliability calculation, quality inspection, label adjustment, and output of data subject to accounting processing by integrating a data collection unit, an automatic classification generation unit, a step-by-step reliability measurement unit, a quality inspection unit, an AI labeling adjustment unit, and a data output unit into a pipeline.

[0254] Through this, the present invention has the advantage of enabling quantitative data quality evaluation that reflects stepwise differences compared to existing single classification and verification methods, allowing for stable and adaptive data classification and verification even in the event of changes in data characteristics or exceptional situations, and objectively and precisely improving inspection quality by adjusting the quality inspection score by reflecting the inspector's proficiency, weights by specialized field, and predefined physical and mental conditions during the quality inspection process.

[0255]

[0256] Meanwhile, since the description of the technology disclosed in this specification is merely an example for structural or functional explanation, the scope of the disclosed technology should not be interpreted as being limited by the examples described in the text. That is, since the examples are subject to various modifications and may take various forms, the scope of the disclosed technology should be understood to include equivalents capable of realizing the technical concept. Furthermore, since the purposes or effects presented in the disclosed technology do not imply that a specific example must include all of them or only such effects, the scope of the disclosed technology should not be understood as being limited by them.

[0257] Furthermore, when it is stated that one component is “connected” to another component, it should be understood that while it may be directly connected to that other component, there may also be other components in between. Conversely, when it is stated that one component is “directly connected” to another component, it should be understood that there are no other components in between. Meanwhile, other expressions describing the relationships between components, such as “between” and “between” or “adjacent to” and “directly adjacent to,” should be interpreted in the same way.

[0258] A singular expression should be understood to include a plural expression unless the context clearly indicates otherwise, and terms such as “include” or “have” are intended to specify the existence of the set-up features, numbers, steps, actions, components, parts, or combinations thereof, and should be understood not to preclude the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0259] The present invention is not limited to the specific preferred embodiments described above, and any person skilled in the art to which the invention pertains can make various modifications without departing from the essence of the invention claimed in the claims, and any modifications are within the scope of the claims as long as they relate to the technical concept forming such modifications.

Claims

A data collection unit (110) that collects data subject to accounting processing; An automatic classification generation unit (120) that automatically classifies account items and whether value-added tax is deductible based on the collected accounting processing target data and generates an accounting processing code; A step-by-step reliability measurement unit (130) that calculates step-by-step reliability by applying a calculation formula reflecting multiple probability elements and weights to each processing step of the automatic classification generation unit (120) based on Bayesian probability theory; A quality inspection unit (140) that performs a multi-stage quality inspection and evaluates the level and status of the inspector on generated data including the measured step-by-step reliability to produce a quality inspection result; AI labeling adjustment unit (150) that automatically adjusts or updates the labels of the AI ​​deep learning training dataset by reflecting the above quality inspection results and step-by-step reliability; and A data output unit (160) including a data output unit that stores or provides externally a generated data with confidence and an adjusted AI training dataset. Automated tax accounting data step-by-step reliability measurement and AI labeling adjustment system. In paragraph 1, The above step-by-step reliability measurement unit (130) is, Calculate the update reliability by combining a previously calculated reliability value with a pre-set weight and reflecting at least one of the transaction amount, VAT type, whether the account item has changed, the transaction type, or the appearance of a pre-defined keyword. When some items in the data are missing or incomplete, applying replacement conditions and reflecting the correction coefficient calculated by estimating uncertainty due to randomness in the reliability calculation, Automated tax accounting data step-by-step reliability measurement and AI labeling adjustment system. In paragraph 1, The above quality inspection department (140) is, Calculating the inspector's proficiency based on inspection grades or past inspection accuracy, and adjusting the quality inspection score by applying weights set for each of the inspector's areas of expertise together with state weights based on predefined classifications of physical and mental states, Automated tax accounting data step-by-step reliability measurement and AI labeling adjustment system. In paragraph 1, The above AI labeling adjustment unit (150) is, Calculating overall reliability by combining step-by-step reliability and quality inspection scores, changing or deleting the labels of data where the overall reliability is below a pre-set tolerance value, and maintaining the labels of data where the overall reliability is above the tolerance value and reflecting them in the training dataset. Automated tax accounting data step-by-step reliability measurement and AI labeling adjustment system. A step in which a data collection unit collects data subject to accounting processing; The automatic classification generation unit automatically classifies account items and whether value-added tax is deductible based on the collected accounting processing target data, and generates an accounting processing code; A step of a step reliability measurement unit that calculates step reliability by reflecting multiple probability factors and weights; A step in which the quality inspection department evaluates the level and status of the inspector regarding generated data including measured step-by-step reliability, performs multi-stage quality inspection, and produces a quality inspection result; A step in which an AI labeling adjustment unit automatically adjusts or updates the labels of an AI deep learning training dataset by reflecting the quality inspection results and step-by-step reliability; and A data output unit comprising the step of storing or providing externally the generated data with assigned confidence and the adjusted AI training dataset, Step-by-step reliability measurement and AI labeling adjustment method for automated tax accounting data.

Images