A data governance method and system based on a hybrid expert model

Abstract

This invention discloses a data governance method based on a hybrid expert model. Step S1: The original samples are comprehensively scored, and a first-round fine-tuning training dataset is constructed based on the comprehensive score. The sub-modules constituting the multi-perspective collaborative reasoning module are determined according to the manually labeled data governance task type. The first-round fine-tuning training dataset is used to perform a first-round supervised fine-tuning of the hybrid expert MoE model, generating a multi-perspective collaborative reasoning module containing the aforementioned sub-modules. Using the first-round supervised fine-tuning multi-perspective collaborative reasoning module, a single call to the multi-governance perspective collaborative module is executed on a new batch of data to perform reasoning, obtaining the output results of each sub-module and the comprehensive adjudication conclusion. The output results of the first-round supervised fine-tuning multi-perspective collaborative reasoning module are scored with comprehensive adjudication confidence and classified. This invention performs vertical domain supervised fine-tuning of the MoE base model, enabling the model to complete independent structured reasoning and comprehensive adjudication of multiple tasks in a single call.

Description

Technical Field

[0001] This invention relates to a data governance technology based on artificial intelligence. Background Technology

[0002] Data governance is a method of managing data throughout its entire lifecycle (from acquisition and use to disposal).

[0003] With the rapid development of business systems in vertical sectors such as government affairs, finance, healthcare, and industrial supervision, the scale of business data in different sectors is constantly expanding and is often subject to multiple constraints from different industry standards, business rules, and laws and regulations. Before data enters business analysis, risk control, or regulatory decision-making, it is usually necessary to carry out multi-dimensional data governance, including field standardization, entity unification, logical consistency verification, and compliance checks.

[0004] Some existing AI-based data governance technologies use only a single AI (artificial intelligence) model, making it difficult to simultaneously address multiple governance responsibilities such as format specification, logical verification, and compliance checks.

[0005] Other existing AI-based data governance technologies employ multi-AI model collaboration solutions, which obtain individual results by calling each AI model independently multiple times. This approach suffers from efficiency bottlenecks and requires business systems to be aware of and manage these multiple calls, resulting in high integration complexity.

[0006] In addition, existing AI models for data governance generally lack high-quality training data, making it difficult to cover complex scenarios and long-tail anomalous data, resulting in poor adaptability. Summary of the Invention

[0007] The technical problem to be solved by this invention is: how to select high-quality training data from a large amount of training data for training AI models, and how to perform multiple data governance tasks simultaneously by calling a trained AI model in a single call.

[0008] To address the aforementioned technical problems, this invention proposes a data governance method based on a hybrid expert model, comprising the following steps: Step S1: Collect the original samples for the data governance task. Score the original samples using perplexity, rule coverage scarcity, long-tail data mining metrics, and historical error contribution. Normalize and weight the scores of these four metrics to obtain a comprehensive score. Original samples with higher comprehensive scores are assigned to the manual validation pool; original samples with moderate comprehensive scores are entered into the model fine-tuning pool. Step S2: Manually label the samples in the manual validation pool to obtain the first part of the training data. Perform zero-sample pre-labeling on the samples in the model fine-tuning pool using a large language model to generate pre-labeled results, obtaining the second part of the training data. Merge the first and second parts of the training data to form the first round of fine-tuning training dataset. Step S3: Based on the type of data governance task manually labeled in the first round of fine-tuning training dataset, determine which sub-modules constitute the multi-perspective collaborative reasoning module. Step S4: The hybrid expert MoE model is subjected to the first round of supervised fine-tuning using the initial fine-tuning training dataset, generating a multi-perspective collaborative reasoning module containing the multiple sub-modules; each sub-module corresponds to a perspective; the multi-perspective collaborative reasoning module fuses and comprehensively adjudicates the output results of each sub-module. Step S5: Using the multi-perspective collaborative reasoning module after the first round of supervised fine-tuning, a single call to the multi-governance perspective collaborative module is executed on the new batch of data to perform reasoning, obtaining the output results of each sub-module and the comprehensive adjudication conclusion. Step S6: The output of the multi-perspective collaborative reasoning module after the first round of supervised fine-tuning is scored using perplexity, rule coverage scarcity, long-tail data mining metrics, historical error contribution, and divergence. The scores of the above five metrics are normalized and weighted to obtain a comprehensive decision confidence score, which is then classified. Among them, samples with high divergence and high error, along with the correct labels verified by humans, are entered into the retraining pool as the first part of the continuous retraining dataset. Samples with high divergence but acceptable confidence are manually corrected and included as the second part of the continuous retraining dataset. The continuous retraining dataset is used to continuously retrain the multi-perspective collaborative reasoning module after the first round of supervised fine-tuning.

[0009] Further, in step S1, the perplexity is used as a data quality quantification indicator. A generalized large language model is used to calculate the perplexity for each original sample, setting an upper and lower threshold. Original samples with perplexity higher than the upper threshold are considered high PPL data; those with perplexity between the upper and lower thresholds are considered medium PPL data; and those with perplexity lower than the lower threshold are considered low PPL data. Samples with high model uncertainty but acceptable quality are selected using the perplexity indicator. The rule coverage scarcity uses a preset value. The long-tail data mining indicator involves embedding the original samples and clustering them using an unsupervised clustering algorithm. A first score is obtained based on the inverse distance between the original sample and the cluster center, and a second score is obtained by inversely normalizing the prediction confidence of the original sample based on a single-parameter generalized large language model. The two scores are weighted and fused to obtain the long-tail data mining indicator, with a weight coefficient sum of 1. The long-tail data mining indicator simultaneously samples cluster center data and cluster boundary data. The historical error contribution uses a preset value.

[0010] Furthermore, in step S1, the original samples with lower comprehensive scores are included in the archive pool and do not participate in the construction of the training set and manual annotation in the current round; when the scoring strategy is updated or the rule base is expanded, the comprehensive scores of the samples in the archive pool are recalculated.

[0011] Further, in step S3, the multi-perspective collaborative reasoning module includes three sub-modules. The first sub-module performs a standardization governance task, responsible for unifying single-field formats, standardizing symbols, and standardizing terminology; the input is the original value of the current field and its corresponding format rules; the output is the standardization result, processing reason, and confidence level. The second sub-module performs a logical consistency governance task, responsible for verifying cross-field logical constraint relationships; the input is the current field, related fields, and logical constraint rules; the output is the consistency judgment result, correction suggestions, and confidence level. The third sub-module performs a domain compliance governance task, responsible for judging field compliance based on industry standards and regulations; the input is the current field, industry rules, and legal format requirements; the output is the compliance judgment result, correction suggestions, and confidence level.

[0012] Furthermore, in step S4, each sub-module is implemented by making a single structured call to the same fine-tuned MoE base model in conjunction with different governance strategy templates, and the outputs of each sub-module are independent of each other in terms of governance intent; the multi-perspective collaborative reasoning module generates the outputs of each sub-module and the comprehensive decision in a single reasoning, which is equivalent to the collaborative governance effect of multiple sub-models.

[0013] Furthermore, in step S4, when the output results of multiple sub-modules conflict, the multi-perspective collaborative reasoning module performs a comprehensive decision according to the following priority order: Method 1: Weighted fusion of the output results of different sub-modules based on the historical reliability weights of each sub-module. Method 2: Selecting the output results of high-priority sub-modules according to the governance task priority rules. Method 3: When the confidence level of the comprehensive decision segment is lower than a preset threshold, and the output results of at least two sub-modules are inconsistent, it is determined as a high-risk conflict sample and submitted for manual verification.

[0014] Furthermore, in step S6, the rule covers scarcity. Where r represents a rule; x i R(x) represents the data sample after treatment. i ) represents sample x i The set of hit rules; freq(r) represents the number of samples where rule r is actually triggered and corrected; ε is a smoothing term to prevent division by zero; the rule coverage scarcity index is used to filter samples involving low-frequency triggering rules. The historical error contribution... Where r represents a rule; x i R(x) represents the data sample after treatment. i ) represents sample x i A set of rules for hitting; W r This is used to give higher weight to low-frequency rules; ErrorRate(r) is the triggering error rate of rule r in historical governance. The degree of divergence... ; where D result D represents the textual dissimilarity of the output results of each submodule. conf D represents the dispersion of the confidence level of the output results of each submodule. rule α represents the conflict ratio of the hit rules for each submodule; α, β, and γ are normalized weight coefficients, and the sum of the three is 1.

[0015] Further, in step S6, samples with a divergence degree exceeding a preset divergence threshold and a historical error contribution degree exceeding a preset error contribution threshold are considered high-divergence and high-error samples; these samples, along with the manually verified correct labels, enter the retraining pool to construct a continuous retraining dataset. Samples with a divergence degree exceeding a preset divergence threshold but a comprehensive decision segment confidence level not lower than a preset confidence threshold are considered high-divergence but reasonably confident samples; these samples enter the manual verification pool, where erroneous parts of the pre-labeled results are manually corrected before being used to construct a continuous retraining dataset.

[0016] Furthermore, in step S6, samples with low overall decision confidence are placed in the archive pool and do not participate in the construction of the current round's continuous retraining dataset and manual verification pool; when the scoring strategy or rule base is updated, the samples in the archive pool are recalculated for overall decision confidence.

[0017] This invention also proposes a data governance system based on a hybrid expert model, comprising a first scoring module, an annotation module, a multi-perspective determination module, a first-round fine-tuning module, an inference module, and a second scoring module. The first scoring module collects original samples for the data governance task and scores them using perplexity, rule coverage scarcity, long-tail data mining metrics, and historical error contribution. These four scores are normalized and weighted to obtain a comprehensive score. Original samples with higher comprehensive scores are assigned to a manual validation pool; those with moderate comprehensive scores enter the model fine-tuning pool. The annotation module manually annotates the data in the manual validation pool to obtain the first part of training data. It also performs zero-sample pre-annotation on the data in the model fine-tuning pool using a large language model to obtain the second part of training data. The first and second parts of training data are merged to form the first-round fine-tuning training dataset. The multi-perspective determination module determines which sub-modules constitute the multi-perspective collaborative inference module based on the data governance task type manually annotated in the first-round fine-tuning training dataset. The first-round fine-tuning module is used to perform the first round of supervised fine-tuning on the hybrid expert model using the first-round fine-tuning training dataset, generating a multi-view collaborative reasoning module containing the multiple sub-modules; each sub-module corresponds to a viewpoint; the multi-view collaborative reasoning module fuses and comprehensively adjudicates the output results of each sub-module. The reasoning module is used to perform reasoning on a new batch of data using the multi-view collaborative reasoning module after the first round of supervised fine-tuning, obtaining the output results of each sub-module and the comprehensive adjudication conclusion. The second scoring module is used to score the output results of the multi-view collaborative reasoning module after the first round of supervised fine-tuning using perplexity, rule coverage scarcity, long-tail data mining indicators, historical error contribution, and divergence; the scores of the above five indicators are normalized and weighted and fused to obtain the comprehensive adjudication confidence, and then classified; among them, samples with high divergence and high error, along with the correct labels verified by humans, enter the retraining pool as the first part of the continuous retraining dataset; samples with high divergence but acceptable confidence are manually corrected and become the second part of the continuous retraining dataset; the continuous retraining dataset is used to continuously retrain the multi-view collaborative reasoning module after the first round of supervised fine-tuning.

[0018] The technical effect achieved by this invention is that by fine-tuning the MoE base model with vertical domain supervision, the model can complete independent structured reasoning and comprehensive decision-making for multiple tasks in a single call, and continuously optimize the model's capabilities. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the data governance method based on a hybrid expert model proposed in this invention.

[0020] Figure 2 This is a schematic diagram of the data governance system based on a hybrid expert model proposed in this invention.

[0021] The attached diagrams are labeled as follows: First scoring module 1, annotation module 2, multi-perspective determination module 3, first-round fine-tuning module 4, reasoning module 5, and second scoring module 6. Detailed Implementation

[0022] Please see Figure 1 The data governance method based on a hybrid expert model proposed in this invention includes the following steps.

[0023] Step S1: Collect raw samples for data governance tasks, including multi-dimensional and multi-scenario governance task data. Apply Perplexity (PPL) and Rule Coverage Scarcity (S) to the raw samples. rule Four indicators—long-tail data mining indicator B, historical error contribution rate E, etc.—are scored separately. The scores of the above four indicators are normalized and then weighted and merged to obtain a comprehensive score.

[0024] The perplexity level (PPL) is used as a quantitative indicator of data quality. A generalized large language model is used to calculate the PPL for each original sample, with upper and lower thresholds set. Original samples with a PPL above the upper threshold are considered high-PPL data, potentially indicating poor quality, high model learning difficulty, or the presence of domain-specific knowledge. These samples require focused review to improve the model's ability to learn from challenging samples. Original samples with a PPL between the upper and lower thresholds are considered medium-PPL data, representing moderate uncertainty for the model and containing significant information; these are prioritized for training. Original samples with a PPL below the lower threshold are considered low-PPL data, as the model has already mastered these samples, resulting in lower training value. The perplexity index effectively filters out sample data with high model uncertainty but acceptable quality, maximizing the information gain of the training data.

[0025] The long-tail data mining metric B involves embedding the original samples and clustering them using an unsupervised clustering algorithm. Data near the cluster centers is highly representative and requires sampling to uncover problem types not covered by manual annotation, thus supplementing the training data with diversity. Cluster boundary data has high uncertainty and also requires sampling. Through semi-supervised active learning, samples with low model prediction confidence but high information content are selected. The first score is obtained based on the inverse distance between the original sample and the cluster center, and the second score is obtained by inversely normalizing the prediction confidence of the original sample based on a single-parameter general large language model. The two scores are weighted and fused to obtain the long-tail data mining metric B, with the sum of the weight coefficients being 1. This strategy simultaneously samples cluster center data (high first score) and cluster boundary data (high second score), ensuring data coverage of both main and long-tail scenarios and improving the model's generalization ability.

[0026] The rule covers scarcity S rule Preset values ​​are used. The historical error contribution rate E also uses a preset value.

[0027] Original samples with high overall scores (e.g., those in the top 25%) are assigned to a manual validation pool, which consists of high-confidence samples.

[0028] Original samples with moderate overall scores (e.g., ranked between the top 25% and the bottom 25%) are entered into the model fine-tuning pool. These samples are considered to be of medium confidence.

[0029] Original samples with low overall scores (e.g., ranking in the bottom 25%) are placed in an archive pool; these are low-confidence samples. Low-confidence samples do not participate in the current round's training set construction or manual annotation. When the scoring strategy is updated or the rule base is expanded, low-confidence samples are brought back for secondary evaluation to prevent the loss of information from early low-scoring samples.

[0030] Step S2: The high-confidence samples in the manual verification pool are manually labeled to obtain the first part of the training data.

[0031] For the medium-confidence samples in the model fine-tuning pool, a single-parameter general large language model is called to perform zero-sample pre-labeling according to the current data governance strategy template, generating pre-labeled results containing standard values, governance reasons (including rule references) and confidence levels, thus obtaining the second part of the training data.

[0032] The first part of the training data is combined with the second part of the training data to form the first round of fine-tuning training dataset.

[0033] Step S3: Based on information such as the type of data governance task manually labeled in the first round of fine-tuning training dataset, determine which sub-modules constitute the multi-perspective collaborative reasoning module.

[0034] For example, the multi-perspective collaborative reasoning module includes three sub-modules. The first sub-module performs standardization governance tasks, responsible for unifying single-field formats, standardizing symbols, and standardizing terminology; the input is the original value of the current field and its corresponding format rules; the output is the standardization result, processing reason, and confidence level. The second sub-module performs logical consistency governance tasks, responsible for verifying cross-field logical constraint relationships; the input is the current field, related fields, and logical constraint rules; the output is the consistency judgment result, correction suggestions, and confidence level. The third sub-module performs domain compliance governance tasks, responsible for judging field compliance based on industry standards and regulations; the input is the current field, industry rules, and legal format requirements; the output is the compliance judgment result, correction suggestions, and confidence level.

[0035] Step S4: The Mixture of Experts (MoE) model is subjected to the first round of supervised fine-tuning using the initial fine-tuning training dataset, generating a multi-perspective collaborative reasoning module containing multiple sub-modules (determined in step S3). Each sub-module corresponds to a perspective. The multi-perspective collaborative reasoning module fuses and comprehensively adjudicates the outputs of each sub-module; the input to the comprehensive adjudication is the output, confidence level, and rule hit rate of each sub-module; the output of the comprehensive adjudication is the final governance result and the basis for the adjudication.

[0036] The sub-modules are not multiple independently deployed AI model instances, but rather implemented through a single structured call to the same fine-tuned MoE base model combined with different governance strategy templates. Through the design of training data and supervised fine-tuning using structured output formats for multiple sub-modules, the outputs of each sub-module are independent in terms of governance intent. Using this training method, the multi-view collaborative reasoning module generates the outputs of each sub-module and a comprehensive decision in a single reasoning iteration, effectively achieving the collaborative governance effect of multiple sub-models without requiring multiple independent calls.

[0037] When the outputs of multiple submodules conflict, the multi-perspective collaborative reasoning module makes a comprehensive decision according to the following priority order: Method 1: Weighted fusion of the outputs of different submodules based on their historical reliability weights. Method 2: Selection of the outputs of high-priority submodules based on governance task priority rules (e.g., logical consistency governance tasks take precedence over normalization governance tasks). Method 3: When the confidence level of the comprehensive decision segment is lower than a preset threshold, and the outputs of at least two submodules are inconsistent, it is determined as a high-risk conflict sample, submitted for manual verification, and the basis for the decision and the reason for manual intervention are recorded in the decision chain.

[0038] Step S5: Using the multi-perspective collaborative reasoning module after the first round of supervision and fine-tuning, perform a single call to the multi-governance perspective collaborative module to reason on the new batch of data, and obtain the output results of each sub-module and the comprehensive adjudication conclusion (governed data and basis).

[0039] Step S6: The output of the multi-perspective collaborative reasoning module after the first round of supervision and fine-tuning (the data after governance, used as the sample for this step) is analyzed using perplexity PPL and rule coverage scarcity S. rule B. Long-tail data mining metrics; E. Historical error contribution; D. Disagreement. total Five indicators were scored separately; the scores of the five indicators were normalized and then weighted and merged to obtain the overall decision confidence level.

[0040] The rules cover scarcity. Where r represents the rule. i R(x) represents the data sample after treatment. i ) represents sample x i The set of rules that were hit. freq(r) represents the number of samples in which rule r was actually triggered for correction, i.e., the number of samples in which the governance results are inconsistent with the pre-labeling, and which are manually labeled to confirm that the rule should be triggered. ε is a smoothing term to prevent division by zero. This metric prioritizes samples involving low-frequency triggering rules.

[0041] The contribution of historical errors Where r represents the rule. i R(x) represents the data sample after treatment. i ) represents sample x i A set of rules for hitting. This gives higher weight to low-frequency rules. ErrorRate(r) is the trigger error rate of rule r in historical governance, defined as "the proportion of historical samples in which the governance result is inconsistent with the manual annotation among the samples that hit rule r and rule r was actually triggered for correction".

[0042] The degree of divergence Among them, D result D represents the textual dissimilarity of the output results of each submodule. conf D represents the dispersion of the confidence level of the output results of each submodule. rule This represents the conflict ratio of the rules hit by each submodule. α, β, and γ are normalized weight coefficients, the sum of which is 1, and can be adjusted according to the business scenario.

[0043] When the sample's dissimilarity D total Samples exceeding a preset divergence threshold and whose historical error contribution E also exceeds the preset error contribution threshold are classified as "high divergence and high error". These samples, along with the manually verified correct labels, are entered into the retraining pool to build a continuously retraining dataset.

[0044] When the sample's dissimilarity D totalSamples exceeding the preset divergence threshold but with a comprehensive decision segment confidence level not lower than the preset confidence threshold are classified as "high divergence but acceptable confidence". These samples are then entered into a manual verification pool where errors in the pre-labeled results are manually corrected. The corrected samples are then added to the continuous retraining dataset.

[0045] The above two parts of data constitute the continuous retraining dataset, which is used to perform incremental LoRA (Low-Rank Adaptation) supervised fine-tuning on the multi-view collaborative reasoning module after the first round of supervised fine-tuning—that is, continuous retraining.

[0046] Samples with a high degree of divergence in a certain type of rule trigger updates to the governance strategy template.

[0047] Samples with low overall decision confidence (e.g., ranking in the bottom 25%) are placed in an archive pool and do not participate in building the continuous retraining dataset and human verification pool for the current round. They can be pulled back for evaluation when the scoring strategy or rule base is updated.

[0048] Please see Figure 2 The data governance system based on a hybrid expert model proposed in this invention includes a first scoring module 1, a labeling module 2, a multi-perspective determination module 3, a first-round fine-tuning module 4, an inference module 5, and a second scoring module 6. Figure 2 The system shown corresponds to Figure 1 The method shown.

[0049] The first scoring module 1 is used to collect the original samples for the data governance task, and to apply perplexity PPL and rule coverage scarcity S to the original samples. rule Four indicators—long-tail data mining metric B, historical error contribution rate E, etc.—are scored separately. These scores are then normalized and weighted to obtain a comprehensive score. Original samples with high comprehensive scores are assigned to the manual validation pool. Original samples with medium comprehensive scores are entered into the model fine-tuning pool. Original samples with low comprehensive scores are assigned to the archive pool.

[0050] The annotation module 2 is used to manually annotate the data in the manual validation pool to obtain the first part of the training data; it is also used to perform zero-sample pre-annotation on the single-parameter general-purpose large language model in the model fine-tuning pool to obtain the second part of the training data. The first part of the training data and the second part of the training data are merged to form the first round of fine-tuning training dataset.

[0051] The multi-view determination module 3 is used to determine which sub-modules constitute the multi-view collaborative reasoning module based on information such as the type of data governance task manually labeled in the first round of fine-tuning training dataset.

[0052] The first-round fine-tuning module 4 is used to perform the first round of supervised fine-tuning on the hybrid expert model using the first-round fine-tuning training dataset, generating a multi-view collaborative reasoning module containing multiple sub-modules. Each sub-module corresponds to a viewpoint. The multi-view collaborative reasoning module fuses and comprehensively decides the output results of each sub-module.

[0053] The reasoning module 5 is used to perform reasoning on the new batch of data using the multi-perspective collaborative reasoning module after the first round of supervision and fine-tuning, and to obtain the output results of each sub-module and the comprehensive adjudication conclusion (data after governance and basis).

[0054] The second scoring module 6 is used to evaluate the output of the multi-perspective collaborative reasoning module after the first round of supervision (the data after governance, serving as the sample for this step) using perplexity PPL and rule coverage scarcity S. rule B. Long-tail data mining metrics; E. Historical error contribution; D. Disagreement. total Five indicators were scored separately; the scores of the five indicators were normalized and then weighted and merged to obtain the comprehensive decision confidence level, which was then classified and processed. Among them, the degree of divergence D total "High-discrepancy and high-error" samples, exceeding a preset discrepancy threshold and with a historical error contribution E exceeding a preset error contribution threshold, are used as the first part of the continuous retraining dataset. Discrepancy D total The "high-discrepancy but reasonably confident" samples, which exceed the preset discrepancy threshold but whose overall decision segment confidence is not lower than the preset confidence threshold, are manually corrected for errors in the pre-labeled results and used as the second part of the continuous retraining dataset. The continuous retraining dataset is used to perform incremental LoRA supervised fine-tuning—that is, continuous retraining—on the multi-view collaborative reasoning module after the first round of supervised fine-tuning.

[0055] Compared with existing technologies, the data governance method and system proposed in this invention have the following beneficial effects.

[0056] First, this invention improves the accuracy and stability of complex vertical domain data governance through multi-governance perspective collaborative reasoning and vertical domain supervision fine-tuning. After fine-tuning, the accuracy of core field governance is significantly improved compared with the general model.

[0057] Second, by calling the multi-governance perspective collaborative reasoning module in a single call, this invention eliminates the efficiency bottleneck of calling different AI models multiple times. The end-to-end latency is close to that of a single model, and the system can meet the large-scale data governance throughput requirements of the production environment.

[0058] Third, this invention employs a multi-dimensional value assessment mechanism for both the original and processed data, selecting medium- to high-quality samples at different stages to construct the fine-tuning training dataset for the next stage. At each stage, only the selected high-quality samples are manually labeled or validated, while medium-quality samples can be pre-labeled by the model. This allows for prioritizing high-value samples with limited human resources, significantly reducing the workload of manual labeling.

[0059] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data governance method based on a hybrid expert model, characterized in that, Includes the following steps; Step S1: Collect the original samples for the data governance task, and score the original samples using perplexity, rule coverage scarcity, long-tail data mining indicators, and historical error contribution. After normalizing the scores of the above four indicators, weighted and fused them to obtain a comprehensive score. The original samples with higher comprehensive scores are assigned to the manual verification pool. Original samples with moderate overall scores are entered into the model fine-tuning pool; Step S2: Manually label the samples in the manual validation pool to obtain the first part of the training data; call the large language model to perform zero-sample pre-labeling on the samples in the model fine-tuning pool to generate pre-labeling results and obtain the second part of the training data; merge the first part of the training data and the second part of the training data to form the first round of fine-tuning training dataset. Step S3: Based on the data governance task type of manually labeled data in the first round of fine-tuning training dataset, determine which sub-modules constitute the multi-view collaborative reasoning module; Step S4: The hybrid expert MoE model is subjected to the first round of supervised fine-tuning using the first round of fine-tuning training dataset to generate a multi-view collaborative reasoning module containing the multiple sub-modules; each sub-module corresponds to a viewpoint; the multi-view collaborative reasoning module fuses and comprehensively decides the output results of each sub-module; Step S5: Using the multi-perspective collaborative reasoning module after the first round of supervision and fine-tuning, perform a single call to the multi-governance perspective collaborative module to reason on the new batch of data, and obtain the output results of each sub-module and the comprehensive adjudication conclusion; Step S6: The output of the multi-perspective collaborative reasoning module after the first round of supervision and fine-tuning is scored using perplexity, rule coverage scarcity, long-tail data mining indicators, historical error contribution, and divergence. The scores of the above five indicators are normalized and weighted to obtain the comprehensive decision confidence, and then classified. Among them, samples with high divergence and high error, along with the correct labels after manual verification, are entered into the retraining pool as the first part of the continuous retraining dataset. The samples with high divergence but acceptable confidence were manually corrected and used as the second part of the continuous retraining dataset; the continuous retraining dataset was used to continuously retrain the multi-view collaborative reasoning module after the first round of supervised fine-tuning.

2. The data governance method based on a hybrid expert model according to claim 1, characterized in that, In step S1, the perplexity is used as a data quality quantification indicator. A general large language model is used to calculate the perplexity of each original sample, and an upper and lower threshold are set. Original samples with perplexity exceeding the upper limit of the threshold are considered high PPL data; Original samples with perplexity between the upper and lower thresholds are used as mid-PPL data; Original samples with perplexity below the lower limit of the threshold are considered low PPL data; samples with high model uncertainty but acceptable quality are selected by using the perplexity index. The rule coverage scarcity uses a preset value; The long-tail data mining metric involves embedding the original samples and using an unsupervised clustering algorithm to cluster the original samples. The first score is obtained by inversely calculating the distance between the original sample and the cluster center. The second score is obtained by inversely normalizing the prediction confidence of the original sample based on a single-parameter general large language model. The two scores are weighted and fused to obtain the long-tail data mining index, with the sum of the weight coefficients being 1. The long-tail data mining index is used to simultaneously sample cluster center data and cluster boundary data. The historical error contribution rate uses a preset value.

3. The data governance method based on a hybrid expert model according to claim 1, characterized in that, In step S1, the original samples with lower comprehensive scores are included in the archive pool and do not participate in the construction of the training set and manual annotation in the current round; when the scoring strategy is updated or the rule base is expanded, the comprehensive scores of the samples in the archive pool are recalculated.

4. The data governance method based on a hybrid expert model according to claim 1, characterized in that, In step S3, the multi-perspective collaborative reasoning module includes three sub-modules; The first submodule performs the standardization governance task, responsible for unifying the format of single fields, standardizing symbols, and standardizing terminology; the input is the original value of the current field and the corresponding format rules; the output is the standardization result, the reason for processing, and the confidence level. The second submodule performs logical consistency governance tasks and is responsible for verifying cross-field logical constraint relationships. The input consists of the current field, related fields, and logical constraint rules. The output includes the consistency assessment result, correction suggestions, and confidence level. The third submodule performs domain compliance governance tasks and is responsible for judging field compliance based on industry standards and regulations; The input is the current field, industry rules, and legal format requirements; The output includes compliance assessment results, correction suggestions, and confidence levels.

5. The data governance method based on a hybrid expert model according to claim 1, characterized in that, In step S4, each sub-module is implemented by making a single structured call to the same fine-tuned MoE base model in conjunction with different governance strategy templates. The outputs of each sub-module are independent of each other in terms of governance intent. The multi-perspective collaborative reasoning module generates the outputs of each sub-module and the comprehensive decision in a single reasoning, which is equivalent to the collaborative governance effect of multiple sub-models.

6. The data governance method based on a hybrid expert model according to claim 1, characterized in that, In step S4, when there is a conflict in the output results of multiple sub-modules, the multi-perspective collaborative reasoning module makes a comprehensive decision according to the following priority order; Method 1: Weighted fusion of the output results of different submodules based on the historical reliability weights of each submodule; Method 2: Select the output results of high-priority sub-modules based on the governance task priority rules; Method 3: When the confidence level of the comprehensive decision segment is lower than the preset threshold and the output results of at least two sub-modules are inconsistent, it is judged as a high-risk conflict sample and submitted for manual verification.

7. The data governance method based on a hybrid expert model according to claim 1, characterized in that, In step S6, the rule covers scarcity. Where r represents a rule; x i R(x) represents the data sample after treatment. i ) represents sample x i The set of rules that are hit; freq(r) represents the number of samples in which rule r is actually triggered and corrected; ε is a smoothing term to prevent division by zero; the rule coverage scarcity index is used to filter samples involving low-frequency triggering rules; The contribution of historical errors Where r represents a rule; x i R(x) represents the data sample after treatment. i ) represents sample x i A set of rules for hitting; W r This is used to give higher weight to low-frequency rules; ErrorRate(r) is the trigger error rate of rule r in historical governance; The degree of divergence ; where D result D represents the textual dissimilarity of the output results of each submodule. conf D represents the dispersion of the confidence level of the output results of each submodule. rule α represents the conflict ratio of the hit rules for each submodule; α, β, and γ are normalized weight coefficients, and the sum of the three is 1.

8. The data governance method based on a hybrid expert model according to claim 1, characterized in that, In step S6, samples with a divergence degree exceeding a preset divergence degree threshold and a historical error contribution degree exceeding a preset error contribution threshold are considered high divergence and high error samples; such samples, along with the correct labels verified by manual verification, are entered into the retraining pool to construct a continuous retraining dataset. Samples with a divergence exceeding the preset divergence threshold but a comprehensive decision segment confidence level not lower than the preset confidence threshold are considered high-divergence but reasonably confident samples. These samples are entered into the manual verification pool, where errors in the pre-labeled results are manually corrected and used to construct a continuously retraining dataset.

9. The data governance method based on a hybrid expert model according to claim 8, characterized in that, In step S6, samples with low overall decision confidence are placed in the archive pool and do not participate in the construction of the current round's continuous retraining dataset and manual verification pool; when the scoring strategy or rule base is updated, the samples in the archive pool are recalculated for overall decision confidence.

10. A data governance system based on a hybrid expert model, characterized in that, It includes a first scoring module, a labeling module, a multi-perspective determination module, a first-round fine-tuning module, a reasoning module, and a second scoring module; The first scoring module is used to collect the original samples of the data governance task. The original samples are scored by perplexity, rule coverage scarcity, long-tail data mining indicators and historical error contribution. The scores of the above four indicators are normalized and then weighted and merged to obtain a comprehensive score. The original samples with higher comprehensive scores are allocated to the manual verification pool. Original samples with moderate overall scores are entered into the model fine-tuning pool; The annotation module is used to manually annotate the manual verification pool to obtain the first part of training data; it is also used to perform zero-sample pre-annotation on the large language model in the model fine-tuning pool to obtain the second part of training data; the first part of training data and the second part of training data are merged to form the first round of fine-tuning training dataset. The multi-view determination module is used to determine which sub-modules constitute the multi-view collaborative reasoning module based on the data governance task type of manually labeled data in the first round of fine-tuning training dataset. The first-round fine-tuning module is used to perform the first round of supervised fine-tuning on the hybrid expert model using the first-round fine-tuning training dataset, generating a multi-view collaborative reasoning module containing the multiple sub-modules; each sub-module corresponds to a viewpoint. The multi-perspective collaborative reasoning module integrates and makes a comprehensive decision on the output results of each sub-module. The reasoning module is used to perform reasoning on the new batch of data using the multi-perspective collaborative reasoning module after the first round of supervision and fine-tuning, and to obtain the output results of each sub-module and the comprehensive adjudication conclusion. The second scoring module is used to score the output of the multi-perspective collaborative reasoning module after the first round of supervision and fine-tuning using perplexity, rule coverage scarcity, long-tail data mining indicators, historical error contribution, and divergence. The scores of the above five indicators are normalized and weighted to obtain the comprehensive decision confidence, which is then classified. Among them, samples with high divergence and high error, along with the correct labels verified by humans, are entered into the retraining pool as the first part of the continuous retraining dataset. The samples with high divergence but acceptable confidence were manually corrected and used as the second part of the continuous retraining dataset; the continuous retraining dataset was used to continuously retrain the multi-view collaborative reasoning module after the first round of supervised fine-tuning.

Images