Consistency calibration and standardized output method and system of AI model identification result

Abstract

The invention provides a consistency calibration and standardized output method and system for an AI model recognition result. The method comprises the steps of obtaining yellow page number data in a standardized format; generating an input instruction of a structured AI model by using a stored cue word-based template construction algorithm, performing yellow page number data identification and verification processing through a fine-tuned general AI large model, and converting a processing result into a preliminary standardized AI identification result according with a preset specification according to a fixed field output format rule; according to the preliminarily standardized AI recognition result, the recognition result is intelligently corrected by adopting common error mapping table matching and error case knowledge base verification, the accuracy index of the current batch is calculated, and a calibration recognition result with a confidence score is output; and generating a final consistency calibrated and standardized number identification result by using standardized output format conversion. According to the invention, the accuracy and consistency of yellow page number identification are effectively improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method and system for calibrating and standardizing the output of AI model recognition results. Background Art

[0002] Yellow page number recognition technology is an important part of realizing the digitization of enterprise and merchant information. This technical field mainly involves multiple technical links such as data collection, recognition, verification and standardized output. With the development of Internet technology and the application of artificial intelligence, yellow page number recognition has gradually developed from traditional manual entry to automatic recognition, but data format diversity and standardized output are still important technical challenges facing this field.

[0003] The common yellow page number recognition technology on the market currently mainly uses a combination of OCR image recognition and rule matching to extract information such as phone numbers and addresses from documents of various formats through preset templates. For example, some systems use regular expression-based matching algorithms to identify phone numbers in different formats, while others use entity extraction models based on deep learning to identify and classify various fields in merchant information.

[0004] The most relevant technology currently is the number recognition and verification method based on the AI ​​large model, which uses artificial intelligence technology to batch process and identify yellow page data. Its main technical principle is to use the collected documents of various formats as input, perform natural language understanding and information extraction through the AI ​​model, and output structured merchant information data. This method can handle a certain degree of format changes and data differences by training the model to learn various number formats and contextual relationships.

[0005] However, the existing technology has significant technical defects: first, processing the yellow pages number format problem requires a lot of code, and it is impossible to predict and deal with all possible error formats; second, data operation and maintenance personnel need to spend a lot of time to organize the data input into the large model, and developers also need to write a lot of code to process the model output data; in addition, there is a lack of a pre-processing mechanism for data garbled during the use of the model, which makes it difficult to ensure the consistency and accuracy of the model output results. Summary of the invention

[0006] In view of this, the present application provides a method and system for consistency calibration and standardized output of AI model recognition results, which solves the problems in the prior art that processing the yellow pages number format requires a large amount of code, data collation is time-consuming, and there is a lack of data garbled preprocessing mechanism.

[0007] The embodiment of the present application provides a method for consistency calibration and standardized output of AI model recognition results, including: receiving yellow page number data in a preset format set, and using a format recognition algorithm to identify the source type and format characteristics of the yellow page number data, and using an input format unification algorithm to normalize key fields to obtain yellow page number data in a standardized format; using a stored prompt word template-based construction algorithm to generate a structured AI model input instruction, performing yellow page number data recognition and verification processing through a fine-tuned general AI large model, and converting the processing result into a preliminary standardized AI recognition result that meets the preset specification according to a fixed field output format rule; based on the preliminary standardized AI recognition result, using common error mapping table matching and error case knowledge base verification to intelligently correct the recognition result, calculate the accuracy index of the current batch, and output the calibration recognition result with a confidence score, and use standardized output format conversion to generate a final consistency calibration and standardized number recognition result; based on the final consistency calibration and standardized number recognition result, using a performance indicator calculation model to calculate the overall recognition accuracy and the scores of each dimension, identify high-frequency error patterns to generate an error pattern classification report, and update error correction rules and general AI large model optimization strategies.

[0008] Optionally, the method of utilizing an input format unification algorithm to normalize key fields and obtain yellow pages number data in a standardized format includes: based on the source type and format characteristics of the yellow pages number data, utilizing a number type standard library to match and verify the numbers in the yellow pages number data to obtain structured data with a type identifier; and utilizing an automated format input template generation tool and a preset format template architecture to convert and generate yellow pages number data in a standardized format based on the structured data with a type identifier.

[0009] Optionally, yellow page number data recognition and verification processing is performed through a fine-tuned general AI big model, and the processing result is converted into a preliminary standardized AI recognition result that meets preset specifications according to the fixed field output format rule, including: using the general AI big model to process the yellow page number data in the standardized format based on the structured AI model input instructions to obtain the original AI recognition result; based on the original AI recognition result, using the preset fixed field output format rule to perform format standardization processing on the recognition results one by one, and output the preliminary standardized AI recognition result.

[0010] Optionally, the recognition results are intelligently corrected by matching common error mapping tables and verifying error case knowledge bases, calculating the accuracy index of the current batch, and outputting calibrated recognition results with confidence scores, including: using a preset common error mapping table to match the preliminary standardized AI recognition results, identifying and correcting character confusion and format errors, and obtaining a first preliminary calibrated recognition result; based on the recognition result of the first preliminary calibrated, using the error case knowledge base for in-depth verification and correction, updating the error case knowledge base based on the verification and correction results, and outputting a second preliminary calibrated recognition result; based on the recognition result of the second preliminary calibrated, using a dynamic feedback algorithm to calculate the accuracy index of the current batch, and obtaining the calibrated recognition result with a confidence score.

[0011] Optionally, the use of a dynamic feedback algorithm to calculate the accuracy index of the current batch to obtain the calibrated recognition result with a confidence score includes: based on the recognition result of the second preliminary calibration, using a logarithmic function to process the quantitative ratio of the input yellow pages number data to the output recognition result to obtain a preliminary performance evaluation score, wherein the preliminary performance evaluation score reflects the data processing effect of the recognition result of the second preliminary calibration; based on the preliminary performance evaluation score, using a dynamic weight adjustment algorithm to calculate the evaluation index of each dimension to generate a multi-dimensional performance score, wherein the multi-dimensional performance score includes the number format accuracy, the address format accuracy and the merchant information accuracy; based on the multi-dimensional performance score, using a preset confidence calculation model to comprehensively evaluate the reliability of the recognition result, when the score of any dimension in the multi-dimensional performance is lower than the first preset threshold, the recognition result of the corresponding dimension is recalibrated to output the calibrated recognition result with a confidence score.

[0012] Optionally, the performance indicator calculation model is used to calculate the overall recognition accuracy and the scores of each dimension, identify high-frequency error patterns to generate an error pattern classification report, and update error correction rules and model optimization strategies, including: based on the final consistency calibrated and standardized number recognition results, the performance indicator calculation model is used to calculate the overall recognition accuracy and the scores of each dimension to obtain a model performance evaluation report; using an error pattern analysis algorithm to identify high-frequency error patterns and abnormal conditions in the model performance evaluation report, and outputting the error pattern classification report; based on the error pattern classification report, a rule base update mechanism is used to update error correction rules and model optimization strategies.

[0013] Optionally, before performing yellow page number data identification and verification processing through the fine-tuned general AI large model, it also includes: constructing a linear-nonlinear hybrid machine learning framework, including: according to the yellow page number data in the standardized format, constructing a hybrid machine learning framework including a linear model and a nonlinear model to generate a hybrid processing framework; wherein the linear model is used to process data items that conform to the first preset structural rules, and the nonlinear model is used to process data items that require context understanding; according to the hybrid processing framework, a time attention mechanism is introduced, a time feature extractor is used to extract time tags and timing patterns, and an attention calculation unit is used to assign weights to data in different time periods to obtain yellow page number data features with time weights; according to the yellow page number data features with time weights, the UpperConfidenceBound algorithm is used to dynamically adjust the balance between exploring new strategies and using known strategies, and output an optimized processing strategy for yellow page number data identification.

[0014] Optionally, before using the input format unification algorithm to normalize key fields, it also includes: constructing an equivariant network preprocessing framework for Markov data, including: receiving the yellow pages number data in the preset format set, constructing it as a Markov process, and using the equivariant network construction algorithm to generate an equivariant network structure including a Markov transition modeler, an equivariant feature extractor and a structural consistency verifier; based on the equivariant network structure, analyzing the generalization boundary of the equivariant network structure when processing yellow pages number data, and generating a targeted data enhancement strategy using theoretical generalization error upper bound calculations and empirical test results; based on the data enhancement strategy, analyzing the preset complexity index and dependency structure of the yellow pages number data, dynamically adjusting the order and network structure parameters of the Markov process according to the analysis results, and outputting an optimized equivariant network preprocessing model.

[0015] Optionally, before using common error mapping table matching and error case knowledge base verification to perform intelligent correction on the recognition results, it also includes: establishing an unbounded Gaussian difference private sampling mechanism, including: receiving the preliminary standardized AI recognition results, using an unbounded Gaussian mechanism and a differential privacy protection mechanism, adding Gaussian distributed noise to the yellow pages number data, and obtaining data with differential privacy protection; according to the data with differential privacy protection, dynamically adjusting the parameters of the differential private sampling using a gradient descent method or reinforcement learning technology, optimizing the sampling strategy, and outputting the optimized unbounded Gaussian difference private sampling parameters; based on the optimized unbounded Gaussian difference private sampling parameters, tracking the privacy consumption in each processing step through a privacy budget management mechanism, and automatically adjusting the processing strategy of the unbounded Gaussian difference private sampling mechanism when it is detected that the privacy budget is lower than a preset threshold, to generate a final calibration processing strategy, and the calibration processing strategy is used to guide the subsequent intelligent correction process.

[0016] The embodiment of the present application also provides a system for consistency calibration and standardized output of AI model recognition results, including a data preprocessing module, an AI model enhancement processing module, a post-processing calibration module and a dynamic optimization module connected in sequence, wherein: the data preprocessing module is used to receive yellow page number data in a preset format set, and adopt a format recognition algorithm to identify the source type and format characteristics of the yellow page number data, and use an input format unification algorithm to normalize key fields, and output the yellow page number data in a standardized format to the AI ​​model enhancement processing module; the AI ​​model enhancement processing module is used to generate input instructions for a structured AI model using a stored algorithm based on a prompt word template, perform yellow page number data recognition and verification processing through a fine-tuned general AI large model, and convert the processing results into preliminary standardized data that conforms to preset specifications according to fixed field output format rules. AI recognition results, and the preliminary standardized AI recognition results are transmitted to the post-processing calibration module; the post-processing calibration module is used to intelligently correct the recognition results according to the preliminary standardized AI recognition results, using common error mapping table matching and error case knowledge base verification, calculate the accuracy index of the current batch, and obtain the calibration recognition results with confidence scores, and use standardized output format conversion to generate the final consistency calibration and standardized number recognition results, and send the final consistency calibration and standardized number recognition results to the dynamic optimization module; the dynamic optimization module is used to calculate the overall recognition accuracy and the scores of each dimension based on the final consistency calibration and standardized number recognition results, using the performance indicator calculation model, identify high-frequency error patterns to generate error pattern classification reports, and update error correction rules and general AI large model optimization strategies.

[0017] An embodiment of the present application also provides a computer system, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method for consistency calibration and standardized output of the above-mentioned AI model recognition results.

[0018] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the method for consistency calibration and standardized output of the above-mentioned AI model recognition results.

[0019] An embodiment of the present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the method for consistency calibration and standardized output of the above-mentioned AI model recognition results.

[0020] This application has the following technical effects: by constructing a data preprocessing module, using an input format unification algorithm and an automated JSON input template generation tool, the multi-source heterogeneous yellow page number data is converted into a standardized AI model input, which greatly reduces the impact of data input diversity on model output; by designing an AI model enhancement layer, using constrained output template design and fixed field output format rules, the fine-tuned AI large model can output structured and consistent recognition results, thereby improving the standardization of model output; by implementing a post-processing calibration algorithm, establishing a common error mapping table and an error case knowledge base, and combining a dynamic feedback algorithm, the AI ​​model recognition results are accurately corrected and the confidence is evaluated, which significantly improves the accuracy of the recognition results; by establishing a dynamic optimization mechanism, using a performance indicator calculation model and an error pattern analysis algorithm, a closed-loop feedback system from recognition results to rule updates is constructed, which achieves continuous self-optimization of the system and ensures long-term recognition accuracy; by innovatively designing a full-process standardized processing framework for yellow page number recognition, preprocessing, model processing, and post-processing are organically combined, which solves the technical bottleneck of traditional methods in processing diversified data and greatly reduces the workload of data operation and maintenance and developers. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following is a brief introduction to the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can also be obtained based on these drawings without creative work.

[0022] Figure 1 It is a flow chart of a method for consistency calibration and standardized output of AI model recognition results provided in an embodiment of the present application; Figure 2 is a flowchart for implementing data preprocessing provided in an embodiment of the present application; Figure 3 It is a flowchart for implementing data identification and verification processing provided by an embodiment of the present application; Figure 4 is a flowchart for implementing the generation of number recognition results provided by an embodiment of the present application; Figure 5 It is a system structure diagram for consistency calibration and standardized output of AI model recognition results provided in an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the embodiments of the present disclosure clearer, the technical scheme in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the present disclosure for protection, but merely represents the selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present disclosure.

[0024] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0025] The term "and / or" herein only describes an association relationship, indicating that three relationships may exist. For example, A and / or B may represent the following three situations: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C may represent including any one or more elements selected from the set consisting of A, B, and C.

[0026] The specific implementation of the present application is described in detail below in conjunction with the accompanying drawings. It should be understood that the following description is only exemplary and is not intended to limit the scope of protection of the present application.

[0027] like Figure 1 As shown, the embodiment of the present application provides a method for consistency calibration and standardized output of AI model recognition results, including: S1: Receive yellow page number data in a preset format set, and use a format recognition algorithm to identify the source type and format characteristics of the yellow page number data, and use an input format unification algorithm to normalize key fields to obtain yellow page number data in a standardized format.

[0028] First, in the data preprocessing stage implemented in this application, diverse yellow page number data from different sources are received. These data may exist in various forms, including but not limited to txt text files, Excel tables, CSV files and other formats. These data have significant differences in format and structure, which brings challenges to the subsequent AI model processing.

[0029] Specifically, a special format recognition algorithm was designed, which can automatically analyze the characteristics of input data, including file extensions, content format characteristics, data separator types, and other information in multiple dimensions. Through this intelligent analysis, the source type and format characteristics of yellow page number data can be accurately identified, providing a basis for subsequent standardization processing. For example, for spreadsheet format, its header structure, data column type, and relationship between columns will be identified; for text format, its separators and data organization methods will be analyzed.

[0030] Subsequently, the input format unification algorithm is applied to normalize the identified data. The core function of this algorithm is to accurately extract key fields from raw data in various formats, such as phone numbers, area codes, addresses, business names, etc., and perform preliminary normalization on these fields. For example, for phone numbers, non-standard separators and spaces are removed to unify the format; for address information, the address components are decomposed and structured. This normalization process ensures that data from different sources can be converted into a unified standard format, creating conditions for subsequent AI model processing.

[0031] like Figure 2 As shown, S1 includes: S1.1: Based on the source type and format characteristics of the yellow page number data, the number type standard library is used to match and verify the numbers in the yellow page number data to obtain structured data with type identification.

[0032] After completing the preliminary data source identification and format feature extraction, it is necessary to further accurately identify the type and verify the validity of the key element in the yellow page number data - the telephone number. The implementation of this step depends on the pre-established number type standard library, which contains the standard format patterns of various types of numbers.

[0033] In the specific implementation, possible phone number fields are first extracted from the preprocessed data. Then, the pattern rules defined in the number type standard library are used to perform pattern matching and type recognition on these numbers. For example, the number will be identified as a mobile phone number, landline number, or special service number based on its features such as the number of digits and prefix. During the recognition process, specific verification rules are also applied, such as the matching check between the area code and the number of digits, and the validity verification of the number prefix, to ensure the accuracy of the recognition results.

[0034] For numbers with questionable format identification, such as the fixed-line format differences between "023xxxxxxx" and "23xxxxxxx", they will be distinguished through a specific area code verification algorithm. For example, it will check whether "023" is a valid area code and combine it with the subsequent number length to determine the correct format. Similarly, for the distinction between the "0955" area code and the "95" service number, the preset area code database and service number rules will be referenced for identification.

[0035] Through this process, structured data with clear type identification is finally output. Each phone number is given a clear type label (such as mobile number, landline number, service number, etc.) and validity identification, providing an important basis for subsequent processing.

[0036] S1.2: Based on the structured data with type identification, an automated format input template generation tool and a preset format template architecture are used to convert and generate yellow page number data in a standardized format.

[0037] After obtaining structured data with type identification, it is necessary to convert this data into a standardized format so that the AI ​​model can process it efficiently. This step is achieved through the automated format input template generation tool and the preset format template architecture.

[0038] First, the structured data is processed using an automated format input template generation tool. The tool has preset regular expression templates for more than 30 common data formats, covering various phone number formats, area code formats, address formats, etc. Through these templates, various fields in the data can be identified and extracted, even if these fields have various forms of expression in the original data.

[0039] For example, when processing table data containing the number "010XXX456", the merchant name "test merchant", the address "test address", and the longitude and latitude "test longitude, test dimension", the following JSON data structure is automatically generated: {'phone':['XXX456','XXX457'], 'areaCode': '010', 'address': 'test address', 'lng': 123.123, 'lat': 96.233} This standardized JSON format contains clearly defined fields and values, which can be directly used to build prompt word templates for input into AI models. In this way, the data diversity of the input model is effectively reduced and the accuracy of the model output format is improved.

[0040] During the conversion process, preset format template schemas are also applied to ensure the consistency and integrity of the conversion results. These schemas define the fields that must be included in the standard output, the relationship between the fields, and the format requirements of the values.

[0041] For example, for phone number data, the schema requires that area codes and local numbers be stored separately and in a specific format; for address data, the schema defines the components of the address and how they are organized.

[0042] Through the above processing, the yellow pages number data in a standardized format is finally output. This data is not only unified in structure, but also highly consistent in field definition and value representation, providing an ideal input for subsequent AI model processing.

[0043] Before using the input format unification algorithm to normalize the key fields, it also includes: building an equivariant network preprocessing framework for Markov data, including: A1: Receive the yellow page number data in the preset format set, construct it as a Markov process, and use an equivariant network construction algorithm to generate an equivariant network structure including a Markov transition modeler, an equivariant feature extractor and a structure consistency verifier.

[0044] Constructing an equivariant network preprocessing framework for Markov data is an important enhancement to the data preprocessing module.

[0045] In this step, we first receive various types of yellow page number data in a preset format set, which may come from various sources and in different formats. We innovatively treat these data as a Markov process, assuming that each element in the data (such as the digits in the number, the components in the address) has a conditional dependency relationship with the previous and next elements. For example, in a telephone number, the area code part and the local number part usually follow different distribution rules, and the choice of the area code will affect the legitimacy and distribution characteristics of the subsequent numbers.

[0046] Based on this Markov property, an equivariant network construction algorithm is used to generate a complex network structure, which consists of three core components: Markov transition modeler, equivariant feature extractor and structural consistency verifier. The Markov transition modeler is mainly responsible for capturing the conditional dependencies between data elements. It can learn and represent the statistical laws between area codes and subsequent numbers, for example; the equivariant feature extractor is specifically used to extract format-independent data features, making it robust to slight changes in the data (such as changes in separators, order adjustments, etc.); the structural consistency verifier ensures that the processed data maintains structural integrity and avoids the data structure being destroyed during the processing process.

[0047] A key feature of the equivariant network is that it remains invariant or has a predictable transformation relationship to specific transformations of the input data (such as translation, rotation, permutation, etc.). This makes it possible to effectively cope with various format changes when processing yellow page number data, such as the use of different separators, adjustment of the order of numbers, etc., thereby improving the robustness and accuracy of preprocessing.

[0048] A2: Based on the equivariant network structure, the generalization boundary of the equivariant network structure when processing yellow page number data is analyzed, and a targeted data enhancement strategy is generated using theoretical generalization error upper bound calculations and empirical test results.

[0049] After constructing the equivariant network structure, it is necessary to evaluate its generalization ability boundary when processing yellow page number data, that is, to determine the range of data changes that the network can reliably process. This step is achieved by combining theoretical analysis with empirical testing.

[0050] At the theoretical analysis level, the generalization error upper bound calculation method is used to evaluate the theoretical performance upper limit of the equivariant network when dealing with different types of data changes. This calculation takes into account factors such as the structural complexity of the equivariant network, the size of the training data, the Markov characteristics of the data, and the equivariant constraints, and mathematically determines the theoretical limit of the network's generalization ability.

[0051] At the same time, a large number of empirical tests were conducted to evaluate the network's processing performance under these changes by inputting various variations of yellow page number data into the equivariant network, such as adding different degrees of noise, changing the data format, adjusting the data structure, etc. The test results were combined with theoretical analysis to accurately identify the strong and weak areas of the network's generalization ability.

[0052] Based on the above analysis results, a targeted data enhancement strategy is generated. For data types with weak generalization ability of equivariant networks, such as non-standard number representations in a specific format or complex address structure changes, more variant samples of such data will be generated for enhanced training; while for data types where the network already has good generalization ability, unnecessary enhancements will be reduced to optimize the efficiency of computing resources.

[0053] This data enhancement strategy not only improves the ability to process various non-standard data formats, but also enables it to adapt to new data forms that may appear in the future, thereby enhancing its long-term availability and stability.

[0054] A3: According to the data enhancement strategy, the preset complexity index and dependency structure of the yellow page number data are analyzed, and the order and network structure parameters of the Markov process are dynamically adjusted according to the analysis results, and the optimized equivariant network preprocessing model is output.

[0055] After obtaining the targeted data enhancement strategy, the preset complexity index and dependency structure of the yellow page number data are further analyzed to achieve dynamic structural optimization of the Markov model. This enables the preprocessing module to adaptively process yellow page data of different complexities.

[0056] First, the complexity of the Yellow Pages data is assessed, taking into account factors such as the structural complexity of the data, the strength of dependencies between elements, and the diversity of formats. For example, telephone number data in a standard format has a low complexity score, while address data containing non-standard formats, multiple nested structures, or special characters has a high complexity score.

[0057] Based on the complexity evaluation results, the order of the Markov process is dynamically adjusted. For data with simple structures and clear dependencies (such as standard format telephone numbers), low-order Markov models, such as first-order or second-order models, may be used, which can reduce the computational complexity while ensuring processing quality; while for data with complex structures and rich dependencies (such as multi-part address expressions), high-order Markov models may be used, or even variant models such as hidden Markov models, to better capture long-distance dependencies in the data.

[0058] At the same time, the structural parameters of the equivariant network will be adjusted according to the data characteristics, such as the number of network layers, number of nodes, connection weights, etc. For data types that require stronger expressiveness, the network complexity will increase; while for simple data types, the network structure may be simplified to improve processing efficiency.

[0059] Through this dynamic optimization, the final output is an equivariant network preprocessing model that adapts to specific data characteristics. This model can optimize the use efficiency of computing resources while ensuring processing accuracy. This method not only improves the adaptability of the model, but also reduces unnecessary computing overhead, providing significant performance advantages for practical applications.

[0060] S2: Generate input instructions for a structured AI model using the stored algorithm based on the prompt word template, perform yellow page number data recognition and verification processing through a fine-tuned general AI large model, and convert the processing results into preliminary standardized AI recognition results that meet preset specifications based on fixed field output format rules.

[0061] In the AI ​​model enhancement processing stage of the present invention, fine-tuning the general AI model is a key step to ensure recognition accuracy. The general AI model used in the present invention is based on the Transformer architecture, including a 12-layer attention mechanism and a 768-dimensional hidden layer, using a ReLU activation function. More than 500,000 manually annotated yellow page number data were used as a training set during the fine-tuning process, covering various formats of phone numbers, addresses, and business information.

[0062] The following hyperparameter settings are used in the fine-tuning process: the initial value of the learning rate is set to 3e -5 , using the cosine annealing scheduling strategy; the batch size is 32; the number of training rounds is 5; the cross entropy loss function is used; the AdamW optimizer is used, and the weight decay is 0.01. To prevent overfitting, a 10% random dropout is also introduced.

[0063] The model evaluation uses accuracy, precision, recall and F1 score as the main indicators, and the overall accuracy on the validation set reaches 95.2%. In particular, the recognition accuracy for standard format numbers is as high as 98.7%, and the recognition accuracy for abnormal format numbers is also 93.1%, which is significantly better than traditional methods.

[0064] After completing data preprocessing, the AI ​​model enhancement layer processing stage is entered. In this stage, the stored algorithm based on the prompt word template is first used to generate structured AI model input instructions. The algorithm receives the standardized JSON format yellow page number data output from the data preprocessing module as the basic input. These data have been processed in a unified format and contain key information such as number, area code, address, longitude and latitude, and have a consistent data structure.

[0065] The algorithm is built using preset prompt word templates to analyze standardized JSON data. The algorithm selects the corresponding prompt word template based on the task type (such as number identification, number verification, number appeal, or number error correction). Each template contains key instructions and format requirements to guide the AI ​​model to correctly understand and process data.

[0066] Specifically, the constructed model input instructions usually include the following key parts: task description, input data format description, expected output format definition, processing rule description, and example demonstration. The task description clearly tells the model the specific task to be completed; the input data format description helps the model understand the meaning of each field in the JSON structure; the expected output format definition clearly specifies the data structure and field requirements that the model should output; the processing rule description contains the processing principles for special cases; and the example demonstration provides input-output comparison to help the model better understand the task requirements.

[0067] During the construction process, the prompt word template algorithm places special emphasis on the constraints of the output format. For example, it clearly specifies that landline numbers must contain area codes and seven digits, mobile phone numbers must be eleven digits, and each number type needs to output key fields such as original data, cleaned data, confidence, and anomaly identification.

[0068] After the structured input instructions are generated, they are input into a specially fine-tuned general AI model for processing. The fine-tuning process here is key. The general AI model has been trained with a large amount of standard yellow page number data so that it can accurately understand the various formats and characteristics of yellow page number data. The fine-tuning process mainly focuses on allowing the model to learn standard data patterns, master number format rules, and accurately extract and verify key information from complex texts.

[0069] like Figure 3 As shown, S2 includes: S2.1: Using the general AI large model, the yellow pages number data in the standardized format is processed based on the structured AI model input instructions to obtain the original AI recognition result.

[0070] Once the structured AI model input instructions and the standardized format of the yellow pages number data are prepared, they are fed into the fine-tuned general AI model. This processing is a key link in the core function and directly affects the quality of the final recognition result.

[0071] During the processing, the AI ​​model first parses the input instructions and data. The instruction part tells the model the type of task to be performed and the processing method, while the data part provides the yellow pages number information to be analyzed. Due to the early data preprocessing and structured instruction generation, the model can clearly understand the task requirements and data structure, greatly improving the processing efficiency and accuracy.

[0072] Subsequently, the model initiates a series of processing steps: first, it conducts an in-depth analysis of the input data to understand the various fields and information contained therein; then, it processes the data specifically according to the task requirements, such as verifying the legitimacy of the number format, identifying the number type, and correcting obvious format errors; finally, it generates processing results based on the preset output requirements.

[0073] Throughout the entire processing process, the model fully utilizes the domain knowledge it has acquired through fine-tuning to accurately analyze the Yellow Pages number data. For example, the model can recognize that "010-123XXXXX" and "010123XXXXX" are actually the same number, but with different formats; it can determine that "400-8XX-88XX" is a service hotline rather than an ordinary landline number; it can accurately extract valid phone number information from mixed text, etc.

[0074] At the same time, the model will also generate confidence scores during the processing process, reflecting its degree of confidence in the processing results. These scores provide an important reference for subsequent calibration and optimization. For example, for numbers with clear formats and in compliance with standards, the model may give a higher confidence level; while for numbers with non-standard formats or containing ambiguous information, it may give a lower confidence level.

[0075] Through this process, raw AI recognition results are output, which contain the model's preliminary processing and understanding of the input data, but may not yet fully meet all the requirements of the standardized output format and require further format standardization in the next step.

[0076] S2.2: Based on the original AI recognition result, a preset fixed field output format rule is used to perform format standardization processing on the recognition results one by one, and the preliminary standardized AI recognition result is output.

[0077] After obtaining the original AI recognition results, these results need to be formatted and standardized to ensure the consistency and standardization of the output results. This step is achieved by applying preset fixed field output format rules.

[0078] Although the original AI recognition results contain the model's processing results for the Yellow Pages number data, they may be inconsistent or incomplete in format and structure. For example, some fields may be missing, the format may not be uniform, and the field value representation may not be consistent. These inconsistencies will cause difficulties in subsequent processing and use, so they need to be standardized.

[0079] Apply the preset fixed field output format rules to process these raw results one by one. These rules include format requirements and conversion rules for different fields, especially strict specifications for phone numbers. For example, landline numbers are required to be output in the format of "area code + 7 digits", such as "010-123XXXXX"; mobile phone numbers must strictly follow the 11-digit format standard, such as "139123XXXXX", and no additional separators are allowed. There are also corresponding format specifications for fields such as address and business name.

[0080] During the specific implementation process, the corresponding format rules will be applied to each field output by the AI ​​model one by one. For the number field, it will be checked whether it meets the preset format requirements. If not, the corresponding format conversion will be performed; for the address field, structured standardization will be performed to ensure the order and format of the address components are unified; for other fields such as latitude and longitude, merchant name, etc., there are also corresponding format standards.

[0081] In addition, it will ensure that the output contains all necessary standard fields, such as raw data, cleaned data, confidence, and anomaly identification, to ensure the integrity and consistency of the output. For necessary fields that the AI ​​model fails to generate, default values ​​will be added or marked as missing according to the rules.

[0082] Through this process, preliminary standardized AI recognition results that meet the preset specifications are output. These results are highly unified in format and structure, providing a good foundation for subsequent calibration processing. This standardization not only improves the readability and processability of the results, but also creates conditions for error detection and correction in the next step.

[0083] Before the yellow pages number data recognition and verification processing is performed through the fine-tuned general AI large model, it also includes: building a linear-nonlinear hybrid slot machine learning framework, including: B1: Based on the yellow page number data in the standardized format, a hybrid slot machine learning framework including a linear model and a nonlinear model is constructed to generate a hybrid processing framework; wherein the linear model is used to process data items that conform to the first preset structural rules, and the nonlinear model is used to process data items that require context understanding.

[0084] Before entering the AI ​​model processing stage, by building a linear-nonlinear hybrid learning framework, the model's processing capability and adaptability to yellow page number data can be significantly improved. Based on the yellow page number data in a standardized format, this framework creates a hybrid structure that integrates linear and nonlinear models, enabling more flexible response to various data processing scenarios.

[0085] The core feature of the hybrid processing framework is the intelligent allocation of processing tasks based on data characteristics. Specifically, the linear model in is mainly responsible for processing highly structured and well-defined data items, such as telephone numbers in standard formats, standardized area codes, etc. This type of data usually conforms to the first preset structural rules, has clear format definitions and verification standards, and is suitable for processing using linear models with high computational efficiency and concise logic. For example, to determine whether an 11-digit number is a valid mobile phone number, a linear model can be used to quickly verify whether the first three digits are a legitimate operator code.

[0086] At the same time, the nonlinear model (mainly the large language model LLM) is responsible for processing highly complex data items that require contextual understanding, such as irregular address descriptions and mixed information merchant descriptions. This type of data usually requires understanding of semantics, context, and implicit relationships, which is difficult to be effectively processed by simple rule matching. Therefore, it is necessary to utilize the powerful semantic understanding and contextual analysis capabilities of nonlinear models. For example, extracting structured address information from a text such as "the first floor of Lenovo Building, located at the southwest corner of the intersection of North Third Ring Road West and Zhongguancun South Street" requires the deep semantic understanding capabilities of nonlinear models.

[0087] Based on the characteristics of the input data and the historical processing results, it will dynamically decide which model to assign specific data to. This intelligent allocation mechanism not only improves processing efficiency, but also gives full play to the expertise of various models and optimizes the overall processing effect. At the same time, for some complex data, it is also possible to call linear and nonlinear models at the same time, and then combine the processing results of the two to further improve accuracy.

[0088] This hybrid architecture makes full use of the efficiency of linear models and the complex expression capabilities of nonlinear models, making it more flexible to cope with various yellow page data recognition scenarios, thereby improving the overall processing performance.

[0089] B2: Based on the hybrid processing framework, a temporal attention mechanism is introduced, and a temporal feature extractor is used to extract time tags and timing patterns. An attention calculation unit is used to assign weights to data in different time periods to obtain yellow pages number data features with time weights.

[0090] Based on the hybrid slot machine learning framework, an innovative temporal attention mechanism is introduced to enable the model to focus on the temporal dimension of the data, thereby more effectively capturing the characteristics of the yellow page number data evolving over time. This mechanism is particularly suitable for processing long-running yellow page data, as the data format and characteristics often change over time.

[0091] The temporal attention mechanism first extracts time tags and time series patterns from the data through a temporal feature extractor. It records and analyzes the changing trends of the features of the yellow pages data processed in different time periods, including new merchant types, regional format changes, or new number formats. For example, it may be observed that new phone number prefixes appear in a certain time period, or that the address representation method in a specific area changes. These time-related patterns are recorded and form a temporal feature library.

[0092] The attention calculation unit then assigns weights to data from different time periods. This process gives higher weights to recent data and data from similar historical scenarios based on the time stamp and historical performance of the data. For example, if new number format patterns are found in the past month, these new patterns will be given higher attention when processing the current data; at the same time, if special data patterns have appeared in specific seasons (such as during holidays) in history, these patterns will also be given more attention in similar time periods.

[0093] Through this weight distribution, the yellow pages number data features with time weights are obtained, which enables the model to prioritize the most relevant time patterns. This feature representation with time dimension enables the model to not only understand the current structure of the data, but also perceive the time evolution trend of the data, thereby improving the ability to process emerging data patterns.

[0094] The temporal attention mechanism is implemented through three key components: a temporal feature extractor (extracting time tags and temporal patterns from data), an attention calculation unit (assigning weights to data in different time periods), and an adaptive fusion layer (integrating the outputs of different models according to the attention weights). This mechanism enables automatic adaptation to format changes and emerging patterns in yellow page data that evolve over time, such as new number formats, changes in address representation, etc., greatly improving long-term usability and adaptability.

[0095] B3: Based on the yellow page number data characteristics with time weight, the UpperConfidenceBound algorithm is used to dynamically adjust the balance between exploring new strategies and using known strategies, and output an optimized processing strategy for yellow page number data recognition.

[0096] After obtaining the yellow page number data features with time weights, the core mechanism of hybrid slot machine learning is realized: dynamic balance exploration and utilization. This step achieves an intelligent balance between "exploring new strategies" and "utilizing known effective strategies" through the Upper Confidence Bound (UCB) algorithm, ensuring that stable and efficient performance can be maintained while continuously improving and adapting to new situations.

[0097] Specifically, the ratio between exploration and utilization will be adjusted dynamically based on the feedback from the processing results. When encountering new data patterns or when the effectiveness of existing strategies decreases, the exploration ratio will be increased and new processing methods will be tried; and when a strategy is proven to be efficient and reliable, the utilization of the strategy will be increased to improve processing efficiency.

[0098] The UCB algorithm takes into account both the historical performance and uncertainty of each processing strategy when achieving this balance. The algorithm calculates a score for each possible processing strategy, which consists of two parts: the average historical performance of the strategy (i.e., the utilization term) and the confidence upper bound representing uncertainty (i.e., the exploration term). The strategy with the highest score will be selected for trial, thus naturally balancing exploration and utilization.

[0099] For example, when dealing with area code recognition tasks, multiple recognition strategies may be maintained at the same time: rule-based matching, context-based inference, regional information-based verification, etc. The UCB algorithm dynamically decides which strategy to use based on the historical performance and current uncertainty of these strategies. If a certain strategy (such as rule-based matching) performs well in dealing with a specific type of area code, it will be used more frequently; but if a new area code format appears, other strategies will be given enough opportunities to try to find a more suitable processing method.

[0100] Through this dynamic balance mechanism, we can maintain a high recognition accuracy rate while continuously exploring better processing methods to achieve continuous self-optimization. This is especially important for long-term yellow page number recognition, because data features and patterns may change over time, and we need to be able to adapt to these changes and continuously optimize its performance.

[0101] Finally, the optimized yellow pages number data recognition and processing strategies are output. These strategies combine historical experience and innovative attempts, can effectively process various yellow pages data, and continue to optimize as the data characteristics change.

[0102] S3: Based on the preliminary standardized AI recognition results, the recognition results are intelligently corrected by using common error mapping table matching and error case knowledge base verification, the accuracy index of the current batch is calculated, and the calibration recognition results with confidence scores are output, and the final consistency calibration and standardized number recognition results are generated by using standardized output format conversion.

[0103] like Figure 4 As shown, S3 includes: S3.1: Using a preset common error mapping table, matching the preliminary standardized AI recognition results, identifying and correcting character confusion and format errors, and obtaining a first preliminary calibrated recognition result; In the first step of the post-processing calibration phase, the initial standardized AI recognition results are matched and corrected using a preset common error mapping table. This step targets common character confusion and format errors that may occur during AI model processing, and accurately corrects them through pattern matching and replacement.

[0104] The common error mapping table is an important part of the system. It contains more than 2,000 high-frequency error cases, covering a variety of common character confusion and format error types. For example, the confusion between the number "0" and the letter "O" (such as "0XX0" should be "OXXO"), the confusion between the vertical bar "|" and the number "1" and the letter "I" (such as "10XX5|10XX6" should be "10XX5, 10XX6"), and the incorrect handling of special characters (such as "95XXX" should be "95XXX"), etc. These error patterns and their correct corresponding forms are pre-entered and continuously updated.

[0105] The calibration process uses pattern matching and replacement to traverse each field in the recognition result and compare it with the pattern in the error mapping table. When a match is found, the incorrect pattern is replaced with the correct form according to the mapping relationship. This replacement is not limited to simple character replacement, but also includes more complex pattern conversion, such as character combination replacement in a specific context.

[0106] For example, when a character sequence such as "13o" is detected in the recognition result, it will be recognized that the "o" is actually a misidentification of the number "0" and corrected to "130"; similarly, for a sequence such as "I0XX6", it will be recognized that the "I" should be the number "1" and corrected accordingly. This correction is based on the analysis and summary of a large number of actual error cases and can effectively handle most common character confusion situations.

[0107] In addition, the details of each correction are recorded, including the original text, the type of error identified, the corrected text, etc. These records will be used for subsequent error analysis and mapping table updates. In this way, the mapping table can be continuously improved and expanded to improve the coverage of new error types.

[0108] Through this process, the first preliminary calibration recognition results are output, which have corrected most common character confusions and format errors, laying the foundation for subsequent deep calibration. Through this precise error correction, the accuracy of the recognition results is significantly improved, and the manual correction workload of data operation and maintenance personnel is reduced.

[0109] S3.2: Based on the recognition result of the first preliminary calibration, use the error case knowledge base to perform in-depth verification and correction, update the error case knowledge base according to the verification and correction results, and output the recognition result of the second preliminary calibration.

[0110] After the initial correction based on the common error mapping table, the more complex deep calibration phase begins. This step uses the error case knowledge base to perform a deeper level of verification and correction on the recognition results of the first preliminary calibration, dealing with complex error patterns that are difficult to resolve with simple character mapping.

[0111] The error case knowledge base is another key component. Unlike the common error mapping table, it contains more complex and context-related error patterns, which are usually difficult to solve with simple character mapping and require more semantic and contextual information. The error case knowledge base is automatically updated every week, which can continuously absorb new error types and patterns to maintain its validity and applicability.

[0112] During the implementation process, we first analyze the various fields and overall structure of the recognition results to identify possible complex error patterns. For example, we may detect that the address information contains inconsistent regional information, such as the province and city do not match; or we may identify that the merchant name contains possible industry abbreviations or misuse of professional terms. Then, we match these potential error patterns with the cases in the knowledge base to find similar error types and solutions.

[0113] When a match is found, the corresponding correction strategy is applied to correct the recognition result. These correction strategies usually consider not only local characters or words, but also broader context and semantic relationships. For example, for the correction of address information, reference will be made to the geographic database and address format specifications; for the correction of business information, industry background and common expressions will be considered.

[0114] In addition, the error case knowledge base will be updated based on the correction results. When an error is successfully corrected, the error pattern and its correction strategy will be recorded to enhance the coverage of the knowledge base; if a new type of error that cannot be corrected is encountered, it will also be recorded as a reference for updating the knowledge base. This dynamic update mechanism ensures that the knowledge base can be continuously expanded and improved to adapt to new error types.

[0115] An important feature of the error case knowledge base is its automatic update mechanism. New error cases are regularly analyzed, patterns and rules are extracted, and added to the knowledge base, so that the knowledge base can be continuously expanded and improved to adapt to new error types. This dynamic update mechanism ensures that various error situations can be handled continuously and effectively.

[0116] Finally, the second preliminary calibration recognition results are output, which have corrected most of the complex error patterns and greatly improved the recognition accuracy. By combining the common error mapping table and the error case knowledge base, comprehensive coverage and accurate correction of various error types are achieved.

[0117] S3.3: Based on the recognition result of the second preliminary calibration, the accuracy index of the current batch is calculated using a dynamic feedback algorithm to obtain the calibration recognition result with a confidence score.

[0118] After completing two rounds of calibration, the quality of the calibration results needs to be quantitatively evaluated for subsequent processing and optimization. This step is achieved through a dynamic feedback algorithm that calculates the accuracy index of the current batch and generates a confidence score for each recognition result.

[0119] The dynamic feedback algorithm is an adaptive evaluation mechanism that evaluates performance by analyzing the relationship between input and output. The algorithm first collects the recognition results of the second preliminary calibration, then compares and analyzes them with the original input data, and calculates the accuracy and reliability indicators of the processing.

[0120] S3.3.1: Based on the recognition results of the second preliminary calibration, a logarithmic function is used to process the quantitative ratio of the input yellow pages number data to the output recognition results to obtain a preliminary performance evaluation score, wherein the preliminary performance evaluation score reflects the data processing effect of the recognition results of the second preliminary calibration.

[0121] In this step, the recognition result of the second preliminary calibration is firstly quantitatively evaluated. Specifically, a logarithmic function is used to process the quantitative ratio of the input yellow page number data and the output recognition result. This method can effectively smooth the influence of the data volume difference on the evaluation result.

[0122] The following formula is used to calculate the preliminary performance evaluation score:

[0123] I is the amount of input data; O is the amount of data finally returned by the model; Pm is the accuracy rate returned by the model (expressed as a percentage, such as 0.8 represents 80%); Ph is the accuracy of manual statistics (expressed as a percentage, such as 0.8 represents 80%); W1 is the weight of the input data volume; W2 is the weight of the output data volume; W3 is the weight of the model accuracy; W4 is the weight of manual accuracy; C is other parameters (which can be set according to actual conditions, such as model running time, data processing rate, etc.); Use logarithmic functions to process input and output to avoid excessive impact of large numbers on the results, while retaining the trend of data volume changes; if F is high, it means that the model performance is good, and the current model can be maintained or slightly optimized; if F is low, it is necessary to count various parameters to accurately find out the cause of model errors, so as to specifically analyze and optimize the model.

[0124] Based on the overall score, determine whether the results need to be optimized, and based on the scores of each dimension, determine in which aspects the output format needs to be optimized and modified.

[0125] In actual applications, the system will calculate the overall accuracy F value, and also calculate the scores of various dimensions (such as number format, address format, merchant name, etc.). If the F value is high, it means that the overall performance of the model is good; if the F value is low, the system will further analyze the scores of various dimensions to find out the links with poor performance and provide direction for subsequent optimization.

[0126] In addition, the system will calculate a confidence score for each recognition result, reflecting the system's confidence in the correctness of the result. The confidence calculation takes into account multiple factors, including the output confidence of the AI ​​model and the matching degree during the correction process.

[0127] Finally, this step outputs calibration recognition results with confidence scores. These results not only contain the calibrated data, but also come with confidence information and performance indicators, providing an important reference for subsequent standardized output and system optimization.

[0128] This calculation method takes into account two key factors: processing completeness (i.e. what proportion of the input data can be processed) and processing accuracy (i.e. how much of the processing results are correct). These two factors together determine the overall performance.

[0129] The preliminary performance evaluation score directly reflects the data processing effect of the recognition result of the second preliminary calibration. If the score is high, it means that most of the input data can be processed with high processing accuracy; if the score is low, it means that there may be problems with incomplete data processing or insufficient accuracy, and the specific reasons need to be further analyzed.

[0130] In this way, the overall quality of the calibration results can be preliminarily evaluated, providing a basis for subsequent multi-dimensional evaluation. At the same time, this evaluation method can also adapt to processing scenarios with different data volumes and types, providing consistent and comparable performance indicators.

[0131] S3.3.2: Based on the preliminary performance evaluation score, a dynamic weight adjustment algorithm is used to calculate the evaluation indicators of each dimension and generate a multi-dimensional performance score, wherein the multi-dimensional performance score includes the accuracy of number format, address format and merchant information.

[0132] After obtaining the preliminary performance evaluation scores, it is necessary to further analyze the performance of each dimension in order to accurately identify possible problem areas. This step is achieved through a dynamic weight adjustment algorithm that can dynamically calculate the evaluation indicators of each dimension based on the importance and historical performance of different dimensions.

[0133] First, several key evaluation dimensions are defined, including number format accuracy, address format accuracy, and merchant information accuracy. These dimensions reflect the performance when processing different types of data. For example, number format accuracy measures the ability to identify and calibrate phone number formats; address format accuracy evaluates the effect of processing complex address information; merchant information accuracy reflects the accuracy of processing information such as merchant names and descriptions.

[0134] The dynamic weight adjustment algorithm assigns appropriate weights to each dimension based on its importance in the current task and historical performance. For example, if it is found that incorrect number format is the most common problem in historical processing, then a higher weight may be given to number format accuracy; or if the current task pays special attention to the accuracy of address information, the weight of address format accuracy may be increased.

[0135] Based on these weights, the evaluation index pi of each dimension is calculated to form a multi-dimensional performance score. This multi-dimensional score not only provides a comprehensive view of the overall performance, but also accurately locates possible problem areas. For example, it may be found that the overall performance is good, but the address format accuracy is relatively low, which indicates the direction that needs to be improved first.

[0136] Through this multi-dimensional evaluation, we can understand its performance characteristics more precisely and provide a clear direction and basis for subsequent optimization. At the same time, this evaluation method can also adapt to different types of task requirements and provide task-related performance indicators.

[0137] S3.3.3: Based on the multi-dimensional performance score, the reliability of the recognition result is comprehensively evaluated using a preset confidence calculation model. When the score of any dimension in the multi-dimensional performance is lower than a first preset threshold, the recognition result of the corresponding dimension is recalibrated and the calibrated recognition result with a confidence score is output.

[0138] After obtaining the multi-dimensional performance scores, the final evaluation and calibration phase begins. This step uses a preset confidence calculation model to comprehensively evaluate the reliability of the recognition results and perform additional calibration if necessary.

[0139] The confidence calculation model is a complex assessment that takes into account multiple factors to evaluate the reliability of each recognition result. These factors include the confidence of the original output of the AI ​​model, the degree of matching during the calibration process, and the performance scores of each dimension. The model calculates a confidence score for each recognition result, usually expressed as a floating point number between 0 and 1, reflecting the degree of confidence in the correctness of the result.

[0140] We will focus on the scores of each dimension in the multi-dimensional performance. When the score of any dimension is found to be lower than the preset threshold, it will be considered that there may be problems with the recognition results of that dimension and additional calibration is required. For example, if the address format accuracy is found to be lower than the threshold, all address information will be recalibrated; or if the number format accuracy is found to be insufficient, the phone number information will be additionally verified and corrected.

[0141] The recalibration process will adopt corresponding strategies according to the specific problem areas. For example, for number format issues, stricter format validation rules may be applied; for address information issues, reference checks against geographic databases may be added. This targeted calibration can effectively solve accuracy issues in specific areas and improve overall recognition quality.

[0142] Finally, the calibration recognition results with confidence scores are output. These results not only contain data that have undergone multiple rounds of calibration, but also come with an evaluation score for the reliability of each result. This output format with confidence enables downstream applications to flexibly handle different results based on the reliability level. For example, high-confidence results can be used directly, while low-confidence results may require manual review or additional verification.

[0143] Through this comprehensive evaluation and targeted calibration mechanism, the accuracy and reliability of recognition results can be significantly improved, while also providing detailed performance indicators and problem location for subsequent optimization.

[0144] Before the common error mapping table matching and error case knowledge base verification are used to intelligently correct the recognition results, it also includes: establishing an unbounded Gaussian difference private sampling mechanism, including: C1: Receive the preliminary standardized AI recognition results, use the unbounded Gaussian mechanism and the differential privacy protection mechanism to add Gaussian distributed noise to the yellow page number data, and obtain data with differential privacy protection.

[0145] Before performing intelligent correction, the unbounded Gaussian difference private sampling mechanism is first implemented, which is an advanced technology designed to protect data privacy while improving calibration effects. This step first receives the preliminary standardized AI recognition results output from the AI ​​model enhancement layer, and then processes the data through a carefully designed privacy protection mechanism.

[0146] Differential privacy is a widely recognized data protection technology. Its core idea is to add carefully designed noise to data or query results to ensure that even if the attacker has all the information except the target data, he cannot determine whether the target data is included in the data set. This protection is particularly important when processing yellow pages number data that may contain sensitive information, which can effectively prevent the leakage and abuse of personal information.

[0147] The unbounded Gaussian mechanism is used as a specific implementation of differential privacy. Compared with the traditional bounded differential privacy mechanism, the unbounded Gaussian mechanism can process data with unbounded value ranges, which makes it particularly suitable for processing various numerical and text information in yellow page number data. By adding noise that follows a Gaussian distribution to the data, the mechanism can theoretically provide the best privacy protection performance while maintaining data availability.

[0148] In specific operations, the privacy budget (i.e., the strength of privacy protection) will be dynamically adjusted according to the sensitivity of the data and the purpose of use. For example, for data fields containing sensitive information such as personal phone numbers, a higher privacy protection strength will be applied, adding relatively large noise; while for public commercial information, a relatively low protection strength may be used to obtain better data utility.

[0149] Through this mechanism, data with differential privacy protection is output, which not only protects privacy but also retains enough information for subsequent processing. This balance is particularly important for yellow page number processing, because it needs to find the best balance between protecting personal privacy and providing useful services.

[0150] C2: Based on the data with differential privacy protection, the parameters of differential private sampling are dynamically adjusted using the gradient descent method or reinforcement learning technology to optimize the sampling strategy and output the optimized unbounded Gaussian differential private sampling parameters.

[0151] After obtaining data with differential privacy protection, it is necessary to further optimize the parameters of differential private sampling to achieve the best balance between privacy protection and data utility. This step achieves dynamic optimization of sampling parameters through gradient descent or reinforcement learning technology.

[0152] An adaptive algorithm is designed to dynamically adjust differential private sampling parameters according to data characteristics and usage purposes. The algorithm considers multiple factors, such as data sensitivity, required accuracy, privacy risk, etc., so that it can find the best balance between privacy protection and data utility in different situations.

[0153] When using the gradient descent method for parameter optimization, we first define an objective function that comprehensively measures the privacy protection level and data utility. Then, by iteratively adjusting the sampling parameters (such as noise amplitude, distribution parameters, etc.), we gradually find the parameter combination that makes the objective function optimal. This gradient-based optimization method can efficiently find the optimal solution in a large parameter space and is suitable for dealing with complex optimization problems.

[0154] Alternatively, reinforcement learning techniques may be used for parameter optimization. In this case, parameter adjustment is considered a decision-making process, and the optimal parameter adjustment strategy is learned through interaction with the environment (i.e., data processing). By experimenting with different parameter settings, observing their effects, and adjusting the strategy based on the effects, a strategy that maximizes long-term returns (i.e., a comprehensive indicator of privacy protection and data utility) is gradually learned.

[0155] Through these optimization methods, the sampling strategy can be continuously adjusted and improved so that the added noise can effectively protect privacy without excessively affecting the availability of data. This optimization is an ongoing process, and the strategy will be continuously adjusted based on the feedback from the calibration results to adapt to different types of data and usage scenarios.

[0156] Finally, the optimized unbounded Gaussian difference private sampling parameters are output, which can guide the addition of the most appropriate amount of noise when processing yellow page number data to achieve the best balance between privacy protection and data utility. This carefully adjusted parameter setting is the key foundation for providing high-quality and secure services.

[0157] C3: Based on the optimized unbounded Gaussian difference private sampling parameters, the privacy consumption in each processing step is tracked through the privacy budget management mechanism. When it is detected that the privacy budget is lower than the preset threshold, the processing strategy of the unbounded Gaussian difference private sampling mechanism is automatically adjusted to generate the final calibration processing strategy, which is used to guide the subsequent intelligent correction process.

[0158] After obtaining the optimized unbounded Gaussian difference private sampling parameters, a privacy budget management mechanism is established, which is a key control to ensure that the expected level of privacy protection is maintained during long-term operation. The privacy budget is a limited resource that will be gradually consumed as data queries and processing proceed, so it needs to be managed carefully.

[0159] A mathematical model was established to theoretically analyze the privacy protection performance of the entire calibration process, including the quantitative calculation of privacy loss, the application of the combination theorem of privacy protection, and the impact of different operations on the privacy budget. These theoretical analyses provide a scientific basis for the management of the privacy budget.

[0160] Based on these theoretical analyses, a privacy budget management mechanism is designed, which can track and control the privacy consumption in each processing step. The amount of noise used in each processing operation and the corresponding privacy loss are recorded, the total privacy consumption is calculated cumulatively, and compared with the preset privacy budget upper limit.

[0161] When it is detected that the privacy budget is close to or below a preset threshold, an automatic adjustment mechanism is triggered. Such adjustments may include increasing the amount of noise to strengthen privacy protection, limiting the frequency of sensitive operations, adjusting data access permissions, etc. For example, more noise protection may be added to particularly sensitive fields (such as personal phone numbers), or the frequency of access to these fields may be reduced.

[0162] This dynamic adjustment ensures that the privacy protection level does not fall below the preset standard while maintaining basic functions and performance. When the privacy budget is about to be exhausted, it may prompt the user to reset the privacy budget or switch to a more conservative processing mode to prevent privacy leakage.

[0163] Finally, the final calibration processing strategy is generated, which takes into account the privacy protection requirements and data processing effects, and provides clear guidelines for the subsequent intelligent correction process. This strategy ensures that during the calibration process, data quality can be effectively improved while maintaining a sufficient level of privacy protection, achieving the dual goals of security and utility.

[0164] This privacy budget management mechanism and dynamic adjustment strategy are important guarantees for the secure processing of sensitive yellow pages data. It enables the provision of provable privacy protection guarantees and enhances compliance and security when processing sensitive data.

[0165] S4: Based on the final consistency calibration and standardized number recognition results, the performance indicator calculation model is used to calculate the overall recognition accuracy and the scores of each dimension, identify high-frequency error patterns to generate an error pattern classification report, and update error correction rules and general AI large model optimization strategies.

[0166] S4 includes: S4.1: Based on the final consistency calibration and standardized number recognition results, the performance indicator calculation model is used to calculate the overall recognition accuracy and the scores of each dimension to obtain a model performance evaluation report.

[0167] After completing the processing and calibration of the yellow page number data, we entered the first step of the dynamic optimization phase: performance evaluation. This step comprehensively evaluated the processing performance of the model through quantitative index calculation and analysis, providing data support for subsequent optimization.

[0168] First, the final consistency calibration and standardized number recognition results are received as the basic data for evaluation. These results have been fully processed and calibrated and represent the current best output quality. These results are analyzed using a preset performance indicator calculation model, which includes a series of indicators and algorithms for quantitative performance evaluation.

[0169] In terms of overall performance evaluation, the overall recognition accuracy was calculated, which is the core indicator for measuring the overall processing quality. The overall recognition accuracy is calculated by comparing the processing results with the standard answers (or the correct results confirmed by other verification methods) and calculating the proportion of correct recognitions. This indicator directly reflects the overall accuracy of processing the yellow page number data.

[0170] In addition to the overall accuracy, the scores of various key dimensions are also calculated. These dimensions include but are not limited to number format accuracy, address information accuracy, merchant data completeness, etc. Each dimension is quantitatively analyzed through a specially designed evaluation algorithm to generate a corresponding performance score. This multi-dimensional evaluation enables the precise identification of performance strengths and weaknesses, providing direction for targeted optimization.

[0171] In addition, some auxiliary performance indicators are calculated, such as processing speed, resource consumption, stability, etc. These indicators reflect the operating status and efficiency from different angles.

[0172] By integrating the above indicators and analysis results, a comprehensive model performance evaluation report is generated. This report not only contains the specific values ​​of each indicator, but also includes in-depth content such as historical trend analysis, similar comparison, performance fluctuation analysis, etc. This comprehensive evaluation report provides managers and developers with a clear performance view, helping them understand the current status and potential problems.

[0173] Most importantly, this performance evaluation report provides a solid data foundation for subsequent error analysis and optimization strategy formulation, ensuring that optimization work can be based on objective data rather than subjective judgment, thereby improving the accuracy and effectiveness of optimization.

[0174] S4.2: Using an error pattern analysis algorithm, identify high-frequency error patterns and abnormal conditions in the model performance evaluation report, and output the error pattern classification report.

[0175] After obtaining the model performance evaluation report, we entered the deeper error analysis stage. This step uses a special error pattern analysis algorithm to analyze and classify the problems found in the performance evaluation to help understand the nature and laws of the problems.

[0176] First, error cases and anomalies are extracted from the performance evaluation report. These cases represent processing failures or suboptimal situations and are the focus of optimization. Then, the error pattern analysis algorithm is applied to conduct in-depth analysis of these cases.

[0177] The error pattern analysis algorithm uses a variety of advanced pattern recognition and clustering techniques to identify underlying common patterns and regularities among seemingly disparate error cases. For example, the algorithm may find that multiple error cases involve a specific type of area code format processing, or appear in a specific context of address information processing. This pattern recognition capability enables the abstraction of more general error types from specific error cases.

[0178] In the process of identifying patterns, the algorithm pays special attention to high-frequency error patterns, that is, those types of errors that appear many times repeatedly. These high-frequency error patterns usually represent more serious problems, and solving these problems may lead to more significant performance improvements. For example, if number formats with special characters are frequently processed incorrectly, then optimizing this specific processing logic may greatly improve the overall accuracy.

[0179] In addition to high-frequency errors, the algorithm also identifies anomalies, which are types of errors that are less common but may have a significant impact. These anomalies may represent potential risks or emerging challenges under specific conditions and require special attention.

[0180] Based on the results of pattern recognition and analysis, a detailed error pattern classification report is generated. This report classifies and sorts the identified error patterns by type, frequency, severity, and other dimensions, and provides representative cases and analysis of each type of error. The report also includes the time distribution of error patterns, correlation analysis, and possible root cause speculation, providing a comprehensive reference for subsequent rule updates and model optimization.

[0181] Through this kind of error analysis, we can extract key issues and optimization directions from a large number of specific errors, making the optimization work more targeted and improving resource utilization efficiency and optimization effects.

[0182] S4.3: Based on the error pattern classification report, a rule base update mechanism is adopted to update the error correction rules and model optimization strategy.

[0183] After completing the analysis and classification of error patterns, we entered the last step of the dynamic optimization phase: rule updating and optimization strategy formulation. This step is the key link of closed-loop feedback, which converts the conclusions obtained from the analysis into specific optimization actions to achieve continuous progress.

[0184] First, we classify the error patterns and report to determine the rules and policies that need to be updated first. This prioritization takes into account multiple factors, such as the frequency of the error, the impact, the difficulty of solving it, and resource constraints. Through this comprehensive consideration, we can select the most valuable optimization direction under limited resource conditions.

[0185] In terms of updating error correction rules, a rule base update mechanism is used. This mechanism can automatically generate or modify corresponding error correction rules based on the identified error patterns. For example, if the analysis finds frequent errors when processing area codes in a specific format, the relevant area code recognition and verification rules will be updated; or if it is found that a certain type of special characters often causes confusion, the character replacement rules will be updated.

[0186] For more complex error patterns, you may need to add entirely new rules or modify the combination logic of existing rules. For example, if you find that the performance is poor when processing multi-nested address information, you may need to design new address parsing rules or adjust the priority and application order of existing rules.

[0187] These rule updates are not limited to the post-processing calibration stage, but may also involve relevant rules in the pre-processing and AI model processing stages, forming a collaborative optimization of the entire process.

[0188] In terms of model optimization strategy, we will formulate targeted model improvement plans based on the analysis results. These plans may include model retraining, parameter adjustment, structural modification, and other methods. For example, if we find that the model performs poorly on a certain type of data, we may need to increase the training samples of this type of data; or if we find that certain features have a significant impact on performance, we may need to adjust the way the model handles these features.

[0189] Detailed optimization suggestions will be generated, including the specific content of the optimization, expected results, implementation plan, etc., for reference by managers and developers. After review and confirmation, these suggestions will be applied to the next iteration.

[0190] Through this closed-loop feedback mechanism, it is able to continuously learn and improve from actual operations, and continuously improve its ability and efficiency in processing yellow pages number data. This continuous optimization mechanism is the key guarantee for maintaining high performance and adaptability in the long term, enabling it to cope with ever-changing data characteristics and business needs.

[0191] like Figure 5 As shown, the present application also provides a system for consistency calibration and standardized output of AI model recognition results, including a data preprocessing module, an AI model enhancement processing module, a post-processing calibration module and a dynamic optimization module connected in sequence, wherein: The data preprocessing module is used to receive the yellow page number data in a preset format set, and use a format recognition algorithm to identify the source type and format characteristics of the yellow page number data, and use an input format unification algorithm to normalize key fields, and output the yellow page number data in a standardized format to the AI ​​model enhancement processing module; The AI ​​model enhancement processing module is used to generate input instructions of a structured AI model using a stored algorithm based on a prompt word template, perform yellow page number data recognition and verification processing through a fine-tuned general AI large model, and convert the processing result into a preliminary standardized AI recognition result that meets the preset specification according to the fixed field output format rule, and transmit the preliminary standardized AI recognition result to the post-processing calibration module; The post-processing calibration module is used to intelligently correct the recognition results according to the preliminary standardized AI recognition results, using common error mapping table matching and error case knowledge base verification, calculate the accuracy index of the current batch, and obtain the calibration recognition results with confidence scores, and generate the final consistency calibration and standardized number recognition results by using standardized output format conversion, and send the final consistency calibration and standardized number recognition results to the dynamic optimization module; The dynamic optimization module is used to calculate the overall recognition accuracy and the scores of each dimension based on the final consistency calibration and standardized number recognition results using the performance indicator calculation model, identify high-frequency error patterns to generate error pattern classification reports, and update error correction rules and general AI large model optimization strategies. The disclosed embodiment also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the steps of the method for consistency calibration and standardized output of AI model recognition results described in the above method embodiment are executed. Among them, the storage medium can be a volatile or non-volatile computer-readable storage medium.

[0192] In addition, an embodiment of the present disclosure also provides a computer program product, which stores a computer program. When the computer program is executed by a processor, the steps of the method for consistency calibration and standardized output of AI model recognition results provided in any of the above embodiments of the present disclosure are executed. For details, please refer to the above method embodiments, which will not be repeated here.

[0193] The computer program product may be implemented in hardware, software or a combination thereof. In an optional embodiment, the computer program product is embodied as a computer storage medium, which may be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (SDK) and the like.

[0194] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the above-described devices and systems can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed devices, systems and methods can be implemented in other ways. The system embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the system or unit can be electrical, mechanical or other forms.

[0195] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0196] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0197] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0198] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The protection scope of the present disclosure is not limited thereto. Although the present disclosure is described in detail with reference to the aforementioned embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the aforementioned embodiments within the technical scope disclosed in the present disclosure, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure shall be based on the protection scope of the claims.

Claims

1. A method for consistency calibration and standardized output of AI model recognition results, characterized in that: include: Receiving yellow page number data in a preset format set, and using a format recognition algorithm to identify the source type and format characteristics of the yellow page number data, and using an input format unification algorithm to normalize key fields to obtain yellow page number data in a standardized format; The stored algorithm based on the prompt word template is used to generate the input instructions of the structured AI model, and the yellow page number data recognition and verification processing is performed through the fine-tuned general AI large model, and the processing results are converted into preliminary standardized AI recognition results that meet the preset specifications according to the fixed field output format rules; Based on the preliminary standardized AI recognition results, the recognition results are intelligently corrected by using common error mapping table matching and error case knowledge base verification, the accuracy index of the current batch is calculated, and the calibration recognition results with confidence scores are output, and the final consistency calibration and standardized number recognition results are generated by using standardized output format conversion; According to the final consistency calibration and standardized number recognition results, the performance indicator calculation model is used to calculate the overall recognition accuracy and the scores of each dimension, high-frequency error patterns are identified to generate error pattern classification reports, and error correction rules and general AI large model optimization strategies are updated.

2. The method according to claim 1, characterized in that The method of utilizing the input format unification algorithm to normalize key fields and obtain yellow page number data in a standardized format includes: Based on the source type and format characteristics of the yellow page number data, the numbers in the yellow page number data are matched and verified using a number type standard library to obtain structured data with a type identifier; According to the structured data with type identification, an automated format input template generation tool and a preset format template architecture are used to convert and generate yellow page number data in a standardized format.

3. The method according to claim 1, characterized in that: The yellow pages number data is identified and verified through a fine-tuned general AI model, and the processing results are converted into preliminary standardized AI recognition results that meet the preset specifications according to the fixed field output format rules, including: Using the general AI large model to process the yellow pages number data in the standardized format based on the structured AI model input instructions to obtain an original AI recognition result; According to the original AI recognition result, the preset fixed field output format rule is adopted to perform format standardization processing on the recognition results one by one, and the preliminary standardized AI recognition result is output.

4. The method according to claim 1, characterized in that Intelligently correct the recognition results using common error mapping table matching and error case knowledge base verification, calculate the accuracy index of the current batch, and output the calibrated recognition results with confidence scores, including: Using a preset common error mapping table, matching the preliminary standardized AI recognition results, identifying and correcting character confusion and format errors, and obtaining a first preliminary calibrated recognition result; According to the recognition result of the first preliminary calibration, the error case knowledge base is used to perform in-depth verification and correction, the error case knowledge base is updated according to the verification and correction results, and the recognition result of the second preliminary calibration is output; According to the recognition result of the second preliminary calibration, the accuracy index of the current batch is calculated using a dynamic feedback algorithm to obtain the calibration recognition result with a confidence score.

5. The method according to claim 4, characterized in that The method of calculating the accuracy index of the current batch by using a dynamic feedback algorithm to obtain the calibration recognition result with a confidence score includes: According to the recognition result of the second preliminary calibration, the ratio of the number of input yellow page number data to the output recognition result is processed by using a logarithmic function to obtain a preliminary performance evaluation score, wherein the preliminary performance evaluation score reflects the data processing effect of the recognition result of the second preliminary calibration; Based on the preliminary performance evaluation score, a dynamic weight adjustment algorithm is used to calculate the evaluation index of each dimension to generate a multi-dimensional performance score, wherein the multi-dimensional performance score includes the number format accuracy, the address format accuracy and the merchant information accuracy; According to the multi-dimensional performance score, the reliability of the recognition result is comprehensively evaluated using a preset confidence calculation model. When the score of any dimension in the multi-dimensional performance is lower than a first preset threshold, the recognition result of the corresponding dimension is recalibrated and the calibrated recognition result with a confidence score is output.

6. The method according to claim 1, characterized in that The performance indicator calculation model is used to calculate the overall recognition accuracy and the scores of each dimension, identify high-frequency error patterns to generate error pattern classification reports, and update error correction rules and model optimization strategies, including: According to the final consistency calibration and standardized number recognition results, the performance indicator calculation model is used to calculate the overall recognition accuracy and the scores of each dimension to obtain a model performance evaluation report; Using an error pattern analysis algorithm, identifying high-frequency error patterns and abnormal conditions in the model performance evaluation report, and outputting the error pattern classification report; According to the error pattern classification report, a rule base update mechanism is adopted to update the error correction rules and model optimization strategy.

7. The method according to claim 1, characterized in that Before the yellow pages number data recognition and verification process is performed through the fine-tuned general AI large model, it also includes: Build a linear-nonlinear hybrid slot machine learning framework, including: According to the yellow page number data in the standardized format, a hybrid slot machine learning framework including a linear model and a nonlinear model is constructed to generate a hybrid processing framework; wherein the linear model is used to process data items that conform to a first preset structural rule, and the nonlinear model is used to process data items that require context understanding; According to the hybrid processing framework, a temporal attention mechanism is introduced, a temporal feature extractor is used to extract time tags and time series patterns, and an attention calculation unit is used to assign weights to data in different time periods to obtain yellow page number data features with time weights; According to the yellow page number data characteristics with time weight, the UpperConfidenceBound algorithm is used to dynamically adjust the balance between exploring new strategies and using known strategies, and output an optimized processing strategy for yellow page number data identification.

8. The method according to claim 1, characterized in that Before normalizing the key fields using the input format normalization algorithm, it also includes: Construct an equivariant network preprocessing framework for Markov data, including: Receiving the yellow page number data in the preset format set, constructing it as a Markov process, and using an equivariant network construction algorithm to generate an equivariant network structure including a Markov transition modeler, an equivariant feature extractor, and a structure consistency verifier; According to the equivariant network structure, the generalization boundary of the equivariant network structure when processing yellow page number data is analyzed, and a targeted data enhancement strategy is generated by using theoretical generalization error upper bound calculation and empirical test results; According to the data enhancement strategy, the preset complexity index and dependency structure of the yellow page number data are analyzed, the order and network structure parameters of the Markov process are dynamically adjusted according to the analysis results, and the optimized equivariant network preprocessing model is output.

9. The method according to claim 1, characterized in that: Before intelligently correcting the recognition results using common error mapping table matching and error case knowledge base verification, it also includes: Establish an unbounded Gaussian difference private sampling mechanism, including: Receiving the preliminary standardized AI recognition result, using an unbounded Gaussian mechanism and a differential privacy protection mechanism to add Gaussian distributed noise to the yellow page number data, and obtaining data with differential privacy protection; According to the data with differential privacy protection, dynamically adjust the parameters of differential private sampling by using gradient descent method or reinforcement learning technology, optimize the sampling strategy, and output the optimized unbounded Gaussian differential private sampling parameters; According to the optimized unbounded Gaussian difference private sampling parameters, the privacy consumption in each processing step is tracked through the privacy budget management mechanism. When it is detected that the privacy budget is lower than the preset threshold, the processing strategy of the unbounded Gaussian difference private sampling mechanism is automatically adjusted to generate a final calibration processing strategy, which is used to guide the subsequent intelligent correction process.

10. A system for consistency calibration and standardized output of AI model recognition results, characterized in that: It includes a data preprocessing module, an AI model enhancement processing module, a post-processing calibration module and a dynamic optimization module which are connected in sequence, wherein: The data preprocessing module is used to receive the yellow page number data in a preset format set, and use a format recognition algorithm to identify the source type and format characteristics of the yellow page number data, and use an input format unification algorithm to normalize key fields, and output the yellow page number data in a standardized format to the AI ​​model enhancement processing module; The AI ​​model enhancement processing module is used to generate input instructions of a structured AI model using a stored algorithm based on a prompt word template, perform yellow page number data recognition and verification processing through a fine-tuned general AI large model, and convert the processing result into a preliminary standardized AI recognition result that meets the preset specification according to the fixed field output format rule, and transmit the preliminary standardized AI recognition result to the post-processing calibration module; The post-processing calibration module is used to intelligently correct the recognition results according to the preliminary standardized AI recognition results, using common error mapping table matching and error case knowledge base verification, calculate the accuracy index of the current batch, and obtain the calibration recognition results with confidence scores, and generate the final consistency calibration and standardized number recognition results by using standardized output format conversion, and send the final consistency calibration and standardized number recognition results to the dynamic optimization module; The dynamic optimization module is used to calculate the overall recognition accuracy and the scores of each dimension based on the final consistency calibration and standardized number recognition results using a performance indicator calculation model, identify high-frequency error patterns to generate error pattern classification reports, and update error correction rules and general AI large model optimization strategies.

Images