Fair competition review method based on feature decoupling joint distillation and reinforcement learning

Abstract

The application relates to a fair competition review method based on feature decoupling combined distillation and reinforcement learning, and belongs to the technical field of natural language processing and fair competition review. The method comprises the following steps: establishing rule violation category standards according to the latest regulations, collecting policy texts and historical cases, and jointly constructing a rule violation judgment dataset and a rule violation subdivision classification dataset through expert labeling; a teacher model is constructed, the rule violation judgment dataset is used for training, and the optimal weight is saved; a light student model is constructed, the decoupling factor and the probability distribution of the teacher model are inherited through knowledge distillation, and the student model is jointly trained in combination with real labels; a rule violation subdivision classification network is trained based on Lawformer; the student model and the subdivision network are integrated to form a two-stage review process: whether a rule violation exists is judged by the student model, and then the specific rule violation clause category of the rule violation text is judged by the subdivision network, so that step-by-step analysis from coarse granularity to fine granularity is realized. The application can improve the sensitivity and classification accuracy of low-frequency categories.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and fair competition review technology, specifically relating to a fair competition review method based on feature decoupling joint distillation and reinforcement learning. Background Technology

[0002] With the continuous evolution of Natural Language Processing (NLP) technology, deep learning-based text understanding methods have been widely applied in various scenarios such as government affairs, judiciary, and finance. Among them, the semantic recognition and automatic judgment capabilities of policy documents are becoming one of the core technologies in the construction of digital government. The fair competition review system requires administrative agencies at all levels to review whether the content of policy measures involving the economic activities of market entities excludes or restricts competition before formulating such measures, in order to ensure fair market competition. However, in reality, policy texts are generally lengthy, complex in structure, and rigorously abstract in expression. Manual review, line by line, is not only costly and inefficient, but also susceptible to subjective bias, making it difficult to support the current needs for high-frequency policy updates and real-time supervision.

[0003] Although pre-trained language models based on Transformer (such as BERT, RoBERTa, etc.) have made significant progress in text classification and semantic modeling, direct application of large-scale models still faces multiple challenges: (1) low inference efficiency and high deployment cost, making it difficult to run efficiently in actual review systems; (2) ambiguous semantic boundaries, with some policy statements being compliant in form but illegal in substance, and the model's ability to extract such deep semantic relationships is limited.

[0004] To address the aforementioned issues, this invention proposes a fair competition review method based on feature decoupling, joint distillation, and reinforcement learning. This method fully leverages the advantages of large-scale pre-trained models in semantic modeling. By constructing a teacher-student model architecture, the class probability distribution output by the teacher model is used as a "soft label" to guide a lightweight student model in learning its deep semantic discriminative features. This significantly reduces the model parameter size and inference time while preserving the expressive power of key features. This strategy greatly improves review efficiency with minimal loss of accuracy. In this invention, the distillation-optimized lightweight model is used for the first stage of violation identification. In the second stage, a Lawformer pre-trained model from the legal domain is introduced as a violation sub-classification model to perform fine-grained classification of texts already identified as violations, forming a two-layer structure of "initial screening + sub-classification," achieving an efficient and highly accurate automated review process. The overall framework balances performance and efficiency, possessing strong practicality and promotional value. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a fair competition review method based on feature decoupling joint distillation and reinforcement learning.

[0006] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a fair competition review method based on feature decoupling joint distillation and reinforcement learning, comprising the following steps: S1. Establish standards for the categories of violations involved in fair competition review in accordance with the latest laws and regulations, and form a complete system of categories of violations in fair competition review; S2. Obtain policy document texts and historical fair competition review violation cases, combine them with domain experts for fine annotation, and construct a fair competition review classification task dataset, which includes a violation judgment dataset and a violation sub-classification dataset; S3. Construct a teacher model for judging violations in fair competition review; the teacher model adopts a two-stream feature decoupling network of the general cross-domain pre-trained language model BERT and the legal domain Lawformer model; train the teacher model using data from the violation judgment dataset, and dynamically save the model weights with the best performance as the teacher model in the subsequent distillation stage; S4. Construct a student model for judging violations of fair competition review. The student model adopts a shallow bidirectional gated recurrent neural network BiGRU and a state space model Mamba. Distillation learning is used to distill the knowledge of the teacher model in the previous step into the student model. The training input of the student model is consistent with that of the teacher model. During the distillation learning process, the core decoupling factor and output probability distribution of the teacher model are used as multiple labels to pass to the student model and are trained together with the real category labels. S5. Construct a violation classification network based on the Lawformer model in the legal domain; train the violation classification network using concentrated data from the violation classification dataset based on the reinforcement learning DQN algorithm to obtain the trained violation classification network; S6. Integrate the knowledge-distilled student model with a trained violation classification network based on the Lawformer model in the legal domain to construct a complete two-stage fair competition review classification method. After completing the initial judgment on whether the policy text "contains violations", the violation classification network performs fine-grained classification on the sentences judged to "contain violations", and outputs the specific fair competition review clause category corresponding to the violation text; realize the step-by-step review process from "violation judgment" to "violation classification".

[0007] Furthermore, the categories of violations in the fair competition review in step S1 include market access, qualification restrictions, interference with business conduct, inappropriate government subsidies, and market exclusivity policies.

[0008] Furthermore, in step S2, the collected policy document texts and historical fair competition review violation cases are manually labeled to clarify the specific clause number violated by each policy statement, and to ensure that the labeling follows the principle of one statement corresponding to one type of violation standard. After labeling, the statement content is segmented, deduplicated, and standardized, and data cleaning is performed using common noise filtering methods. Finally, the processed statements and their corresponding violation category numbers are uniformly organized into a structured table format to obtain the fair competition review classification task dataset.

[0009] Furthermore, the data processing procedure in the dual-stream feature decoupling network in step S3 is as follows: S31. Input Text The first hidden representation is output after the first branch. The first branch uses the general cross-domain pre-trained language model BERT; the input text... The second hidden representation is obtained through the second branch. The second branch uses the Lawformer, a pre-trained model specifically designed for the legal field. S32. The first hidden representation Second hidden representation The inputs are fed into a shared bidirectional long short-term memory (BiLSTM) network to obtain the first new hidden representation. Second new hidden representation ; S33. The first new hidden representation Second new hidden representation After processing by an adaptive gating fusion mechanism, the fused features are obtained. The formula is expressed as follows: , , in, Indicates the gating weight; This represents the dot product operation; This represents the Sigmoid activation function; This indicates a splicing operation; Gate represents a nonlinear gated network based on a multilayer perceptron; the nonlinear gated network based on a multilayer perceptron includes a first fully connected layer, a second fully connected layer, and a nonlinear activation layer; S34. A factor decoder module is introduced, which consists of two independent linear layers. The first linear layer of the factor decoder performs a linear transformation on the fused features to extract the core factors related to the task objective. The second linear layer extracts the interference factor. The formula is expressed as follows: , , in, and These represent the weight matrices of the first and second linear layers, respectively. and These represent the bias terms of the first linear layer and the second linear layer, respectively; the core factors extracted by the first linear layer of the factor decoder are supervised by the subsequent main task branch, while the interference factors extracted by the second linear layer are supervised by the subsequent adversarial branch. S35. Introduce a gradient inversion layer (GRL) to the interference factor during the backpropagation phase of the adversarial branch. The features are then processed to obtain the gradient-reversed features. The features after gradient reversal After performing a pseudo-classification task using a multilayer perceptron with adversarial branching, the pseudo-classification probability distribution is obtained. The formula is expressed as follows: , , in, This indicates the gradient inversion layer operation; This represents an adversarial branching multilayer perceptron; This represents the pseudo-classification probability distribution of the output of the adversarial branching multilayer perceptron; S36. Core factors in the main branch The input is fed into the main branch multilayer perceptron for classification prediction, yielding the class prediction probability distribution of the main branch. The formula is expressed as follows: , in, This represents a main branch multilayer perceptron; This represents the predicted probability value output by the main branch multilayer perceptron; S37. Decouple the core factor and the interference factor by minimizing the loss function through mutual information to constrain the decoupling process. The formula is as follows: , , in, Mutual information minimization loss; MSE represents mean squared error loss. This represents the Pearson correlation coefficient; This represents the covariance between the core factor and the confounding factor. and These represent the standard deviations of the core factor and the interference factor, respectively. S38. After the dual-stream feature decoupling network has been trained and converged, its key structural parameters are extracted and fixed to obtain the trained fair competition review violation judgment teacher model.

[0010] Furthermore, in step S3, the adversarial branch is trained under supervision using an adversarial loss function: , in, Indicates adversarial loss; C represents the number of categories; This is a real label; The first digit in the probability distribution of the output of an adversarial branch multilayer perceptron represents the digit of the probability distribution. Class probability values; the main branch uses the standard cross-entropy loss function for supervised training, and the formula for the loss function is as follows: , in, Indicates the loss of the main branch; The first branch of the multilayer perceptron output represents the first... Class prediction probability value.

[0011] Furthermore, in step S3, the total loss of the two-stream feature decoupling network... The formula is expressed as: ,in, , , These represent the weight coefficients corresponding to the main branch loss, adversarial loss, and mutual information minimization loss, respectively.

[0012] Furthermore, step S4 specifically includes: S41. The input text is processed using a pre-trained word vector embedding layer consistent with BERT, and then fed into a shallow bidirectional gated recurrent neural network (BiGRU). The output features of the shallow bidirectional gated recurrent neural network (BiGRU) are then fed into the state space model Mamba to obtain the output of the state space model Mamba. The output of the state space model Mamba is processed through an MLP to obtain the class probability distribution, and the violation judgment result is output through an argmax layer. For the input text T, the above formula is expressed as follows: , , , in, The hidden feature representation extracted by the student model is represented by Mamba, the state-space model is represented by BiGRU, and the bidirectional gated recurrent unit is represented by BiGRU. This indicates the class number predicted by the student model. This represents the class probability prediction distribution output by the student model; S42. Under the condition of inputting the same text sample, predict the probability distribution of the categories of the main branch extracted in step S3. With core factors This serves as a supervisory signal during the distillation phase of the student model; multiple supervisory signals are introduced during the student model training phase through the joint distillation loss function; the formula for the joint distillation loss function is as follows: , in, Represents the loss function of combined distillation; Indicates soft label loss; Indicates characteristic distillation loss; Indicates supervised classification loss; , , These represent the weights of the soft label loss, feature distillation loss, and supervised classification loss, respectively. Soft tag loss The distillation loss of the teacher model and the student model on the prediction distribution is expressed by the following formula: , in, Indicates temperature parameter; Indicates KL divergence; Characteristic distillation loss The hidden feature representation extracted for the student model and the core factors decoupled from the teacher model The feature alignment loss between them is expressed by the following formula: , in, represents the mean squared error; n represents the feature dimension. Represents the L2 norm; Supervision and classification of losses Similarly, using the binary cross-entropy loss, the formula is as follows: , in, The first output of the student model represents the... Class prediction probability value; S43. During the training of the student model, the parameters of the teacher model are frozen until the student model converges, and the weights are saved.

[0013] Furthermore, the violation subdivision classification network described in step S5 optimizes the classifier's decision-making ability based on a reinforcement learning strategy: S51. Construct a Q-value network, which is a linear layer; extract features from the text data in the violation subdivision classification dataset using a pre-trained Lawformer model to obtain the text feature vector; input the text feature vector into the Q-value network to estimate the Q-value for each violation category; S52. The Q-learning algorithm is used to train the Q-value network. Within the framework of the Q-learning algorithm, each violation category of the text is defined as an "action," and the feature vector of the text after being encoded by the Lawformer pre-trained model in the legal domain is the "state." The model selects an action in each state and calculates the reward based on the actual label. The formula for calculating the target Q-value is as follows: , in, Indicates the target Q value; Indicates the discount factor; Indicates the current state; Indicates an action; This represents the maximum Q value for all actions in the next state; Represents the set of all possible actions for the next state; Indicates an immediate reward; when When the model predicts correctly, it means the prediction is correct; when... When this happens, it indicates that the model prediction is incorrect; S53. In the Q-learning algorithm, action selection adopts... ; S54. The mean squared error (MSE) loss function is used to measure the difference between the model's predicted Q value and the target Q value. The formula is as follows: , in, This represents the Q-value loss; Indicates the state Take action below Expected rewards afterward; S55. Train and test the violation subdivision classification network, and save the optimal weights.

[0014] The advantages of this invention are: 1) This invention addresses the challenges of complex legal terminology and highly abstract semantic representation in policy texts by proposing a feature decoupling framework based on a two-stream network. This framework handles feature extraction and decoupling from the input text through three independent branches: Core Factor Classification Task: First, starting from the main classification task, the model uses core factors for classification; Mutual Information Minimization Loss: The model minimizes the mutual information between core factors and interfering factors, reducing their correlation and ensuring their independence in the feature space, thus achieving effective decoupling between core factors and interfering factors; Interfering Factor Branch and GRL Adversarial Training: The model introduces a gradient inversion layer (GRL) in the interfering factor branch to perform adversarial training on the interfering factors, thereby suppressing their influence on the core factors and ensuring their independence during optimization. Through this three-branch mechanism, this invention can explicitly optimize the decoupling of core factors and interfering factors, ensuring that core factors accurately capture class features while minimizing the interference of interfering factors on the classification task. Ultimately, the model can better distinguish and understand core factors in the classification task, thereby improving classification accuracy and enhancing the model's generalization ability.

[0015] 2) To address the stringent requirements of reasoning efficiency and resource consumption in practical applications of legal compliance review models, this invention introduces a knowledge distillation mechanism. By constructing a teacher-student model architecture, it significantly reduces inference costs while maintaining high recognition performance. Specifically, in the dual-stream semantic feature extraction framework, the teacher model guides the lightweight student model's training by transferring deep semantic knowledge. During distillation, the student model not only learns real label information but also acquires richer inter-class semantic boundary cognition by aligning with the core factor features extracted by the teacher model and mimicking the teacher model's prediction distribution, thereby effectively compressing parameter scale and improving inference speed. This strategy is particularly suitable for deployment environments with limited computing power, enabling rapid response and efficient operation of policy text compliance review systems without sacrificing detection capabilities.

[0016] 3) This invention further introduces the Q-learning algorithm from reinforcement learning to optimize the model's classification decision-making ability, especially in situations where the datasets for subdivided violation classification are extremely imbalanced. Specifically, this paper uses a reinforcement learning framework to treat each text's violation category as an "action" and defines a reward mechanism to guide the model to continuously optimize its classification strategy. In each state, the model obtains an immediate reward by selecting an appropriate violation category and comparing it with the actual label. To optimize classification performance, the model uses a Q-value network for decision-making, dynamically adjusting its ability to discriminate low-frequency violation categories. In this way, reinforcement learning effectively solves the data imbalance problem and improves the sensitivity to low-frequency categories and classification accuracy. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a schematic diagram illustrating the fair competition review standards for this invention. Figure 3 This is a structural diagram of the dual-stream feature decoupling network model of the present invention; Figure 4 This is a structural diagram of the student violation judgment model of the present invention; Figure 5 This is a flowchart of the complete two-stage fair competition review method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 In this embodiment, as Figure 1 As shown, this invention provides a fair competition review method based on feature decoupling joint distillation and reinforcement learning, the specific steps of which include: S1. Establish standards for the categories of violations involved in fair competition review in accordance with the latest laws and regulations, and form a complete system of violation categories for fair competition review, which will serve as the basic labeling system for subsequent violation judgment and sub-classification; Specifically, based on the latest promulgated and currently effective laws, regulations, and policy documents, a systematic review of legal provisions, administrative regulations, and policy guidelines related to fair competition review is conducted to establish complete and detailed standards for violation categories, serving as the foundational classification system for subsequent classification model training and violation identification. The fair competition review violation categories comprise 54 categories, covering multiple dimensions such as market access, qualification restrictions, interference in business operations, inappropriate government subsidies, and market exclusion policies, comprehensively reflecting potential exclusionary or restrictive competition behaviors that may exist in policy documents.

[0021] S2. By using web crawlers, data cleaning and manual review, and AI generation, policy document texts and historical fair competition review violation cases are obtained. Combined with domain experts, fine annotation is performed to construct a fair competition review classification task dataset. This dataset includes a violation judgment dataset and a violation sub-classification dataset. Examples of the datasets are shown in Table 1.

[0022] Table 1 Dataset Examples Specifically, for the 54 established categories of fair competition review violations, a team of domain experts with relevant policy review and antitrust knowledge conducted meticulous manual annotation of a large number of collected policy texts and historical review cases. This involved clearly identifying the specific clause number violated by each policy statement and ensuring that the annotation followed the principle of one statement corresponding to one violation standard. After annotation, the statements were segmented, deduplicated, and standardized, and data was cleaned using common noise filtering methods. Finally, the processed statements and their corresponding violation category numbers were organized into structured tabular data, constructing a high-quality annotated dataset specifically for fair competition review tasks. This dataset includes two subsets: a binary classification dataset for violation judgment, labeled 0 (no violation) or 1 (violation); and a multi-class dataset for fine-grained classification, labeled sequentially from 0 to 53 according to the 54 violation standards. These datasets serve as the foundation for model training and evaluation. Examples of some of the 54 established categories of fair competition review violations are shown below. Figure 2 As shown.

[0023] S3. Construct a teacher model for judging violations in fair competition review; the teacher model adopts a two-stream feature decoupling network of the general cross-domain pre-trained language model BERT and the legal domain Lawformer model; train the teacher model using data from the violation judgment dataset, and dynamically save the model weights with the best performance as the teacher model in the subsequent distillation stage.

[0024] Specifically, such as Figure 3 As shown, the data processing process in the two-stream feature decoupling network is as follows: S31. Input Text The first hidden representation is output after the first branch. The first branch uses the general cross-domain pre-trained language model BERT; the input text... The second hidden representation is obtained through the second branch. The second branch uses the Lawformer pre-trained model, which is specifically designed for the legal field; the formula is as follows: , , in, This represents a general cross-domain pre-trained language model. This refers to Lawformer, a pre-trained model specifically designed for the legal field. S32. The first hidden representation Second hidden representation The inputs are fed into a shared bidirectional long short-term memory (BiLSTM) network to obtain the first new hidden representation. Second new hidden representation The formula is expressed as follows: , , in, This represents a bidirectional long short-term memory network; This represents the first new hidden representation; This represents the second new hidden representation; S33. The first new hidden representation Second new hidden representation After processing by an adaptive gating fusion mechanism, the fused features are obtained. The formula is expressed as follows: , , in, Indicates the gating weight; This represents the dot product operation; This represents the Sigmoid activation function; This indicates a splicing operation; Gate represents a nonlinear gated network based on a multilayer perceptron; the nonlinear gated network based on a multilayer perceptron includes a first fully connected layer, a second fully connected layer, and a nonlinear activation layer, and the data processing procedure of this network is as follows: , in, , These represent the weights of the first fully connected layer and the second fully connected layer, respectively. , These represent the bias parameters of the first fully connected layer and the second fully connected layer, respectively. Represents the ReLU activation function; This represents the input to a nonlinear gated network based on a multilayer perceptron; This represents the output of a nonlinear gated network based on a multilayer perceptron; S34. A factor decoder module is introduced. This module consists of two independent linear layers. Its purpose is to decompose the features fused by the shared encoder into core factors and interference factors. Specifically, the first linear layer of the factor decoder performs a linear transformation on the fused features to extract the core factors relevant to the task objective. The second linear layer extracts the interference factor. This factor represents redundant information or noise irrelevant to the target task. Although the two independent linear layer structures are identical, their parameters are not shared; the formula is as follows: , , in, and These represent the weight matrices of the first and second linear layers, respectively. and The first linear layer is the bias term for the second linear layer; the core factors extracted by the first linear layer of the factor decoder are supervised by the subsequent main task branch, while the interference factors extracted by the second linear layer are supervised by the subsequent adversarial branch. S35. Introduce a gradient inversion layer (GRL) to the interference factor during the backpropagation phase of the adversarial branch. The features are then processed to obtain the gradient-reversed features. The features after gradient reversal After performing a pseudo-classification task using a multilayer perceptron with adversarial branching, the pseudo-classification probability distribution is obtained. The formula is expressed as follows: , , in, This indicates the gradient inversion layer operation; This represents an adversarial branching multilayer perceptron; The pseudo-classification probability distribution of the adversarial branch multilayer perceptron output is represented; optimization is performed using the adversarial loss function. , in, Indicates adversarial loss; C represents the number of categories; This is a real label; The first digit in the probability distribution of the output of an adversarial branch multilayer perceptron represents the digit of the probability distribution. Class probability value; S36. Core factors in the main branch The input is fed into the main branch multilayer perceptron for classification prediction, yielding the class prediction probability distribution of the main branch. The formula is expressed as follows: , in, This represents a main branch multilayer perceptron; This represents the predicted probability value output by the main branch multilayer perceptron; the process uses the standard cross-entropy loss function for supervised training, and the formula for the loss function is as follows: , in, Indicates the loss of the main branch; The first branch of the multilayer perceptron output represents the first... Class prediction probability value; S37. Decouple the core factor and the interference factor by minimizing the loss function through mutual information to constrain the decoupling process. The formula is as follows: , , in, Mutual information minimization loss; MSE represents mean squared error loss. This represents the Pearson correlation coefficient; This represents the covariance between the core factor and the confounding factor. and These represent the standard deviations of the core factor and the interference factor, respectively. The total loss of the dual-stream feature decoupling network The formula is expressed as: ,in, , , These represent the weight coefficients corresponding to the main branch loss, adversarial loss, and mutual information minimization loss, respectively. + + =1, =0.4, =0.3, =0.3; S38. After the dual-stream feature decoupling network has been trained and converged, its key structural parameters are extracted and fixed to obtain the trained fair competition review violation judgment teacher model.

[0025] S4. For example Figure 4 As shown, a student model for judging violations in fair competition review is constructed. The student model adopts a shallow bidirectional gated recurrent neural network BiGRU and a state-space model Mamba. Distillation learning is used to distill the knowledge of the teacher model in the previous step into the student model. The training input of the student model is consistent with that of the teacher model. In the process of distillation learning, the core decoupling factor and output probability distribution of the teacher model are used as multiple labels to pass to the student model and are trained together with the real category labels.

[0026] Specifically, S41. The input text is processed using a pre-trained word vector embedding layer consistent with BERT, and then fed into a shallow bidirectional gated recurrent neural network (BiGRU). The output features of the shallow bidirectional gated recurrent neural network (BiGRU) are input into the state space model Mamba to obtain the output of the state space model Mamba. The output of the state space model Mamba is processed by an MLP to obtain the output category probability distribution, and the violation judgment result is output through an argmax layer. For the input text T, the above process is represented as follows: , , , in, The hidden feature representation extracted by the student model is represented by Mamba, the state-space model is represented by BiGRU, and the bidirectional gated recurrent unit is represented by BiGRU. This indicates the class number predicted by the student model. This represents the class probability prediction distribution output by the student model; S42. Under the condition of inputting the same text sample, predict the probability distribution of the categories of the main branch extracted in step S3. (i.e., "soft labels") and core factors As a supervisory signal during the distillation stage of the student model, multiple supervisory signals are introduced during the student model training phase through the joint distillation loss function; the formula for the joint distillation loss function is as follows: , in, Represents the loss function of combined distillation; Indicates soft label loss; Indicates characteristic distillation loss; Indicates supervised classification loss; , , represent the weights of the soft label loss, feature distillation loss, and supervised classification loss, respectively. In this invention = =0.3, ; Soft tag loss The distillation loss of the teacher model and the student model on the prediction distribution is expressed by the following formula: , in, Indicates temperature parameter, =2; Indicates KL divergence; Characteristic distillation loss Hidden feature representation extracted for student model Core factors decoupled from the teacher model The feature alignment loss between them is expressed by the following formula: , in, represents the mean squared error; n represents the feature dimension. Represents the L2 norm; Supervision and classification of losses Similarly, using the binary cross-entropy loss, the formula is as follows: , in, The first output of the student model represents the... Class prediction probability value; S43. During the training of the student model, the parameters of the teacher model are frozen until the student model converges, and the weights are saved.

[0027] S5. Construct a violation subdivision classification network based on the Lawformer model in the legal domain; train the violation subdivision classification network using centralized data from the violation subdivision classification dataset to obtain a trained violation subdivision classification network; the violation subdivision classification network optimizes the decision-making ability of the classifier based on a reinforcement learning strategy.

[0028] S51. Construct a Q-value network, wherein the Q-value network is a linear layer; input text The pre-trained Lawformer model extracts features to obtain the input text. Feature vectors; input text The feature vectors are input into the Q-value network to estimate the Q-value of each violation category, which is the expected reward after selecting the violation category; S52. The Q-learning algorithm is used to train the Q-value network. Within the framework of the Q-learning algorithm, each violation category of the text (the sentence to be reviewed) is defined as an "action," and the feature vector of the text after being encoded by the Lawformer pre-trained model in the legal domain is defined as a "state." The model selects an action (i.e., selects a violation category) in each state and calculates the reward based on the actual label. The formula for calculating the target Q-value is as follows: , in, Indicates the target Q value; Indicates the discount factor; Indicates the current state, which is the input text. Semantic representation; This indicates the action, which corresponds to the violation category in the text. This represents the maximum Q value for all actions in the next state; Represents the set of all possible actions for the next state; Indicates an immediate reward; when When the model predicts correctly, it means the prediction is correct; when... When this happens, it indicates that the model prediction is incorrect; S53. In the Q-learning algorithm, action selection adopts... During training, the model will have a certain probability Choosing a random action (i.e., randomly selecting a violation category), in most cases the model... Given the probability, select the action with the largest current Q value and execute a greedy strategy; specifically, assume the current model outputs a Q value of The model will be based on Choose an action. That is: pass The model can find a balance between exploration (exploring new categories) and utilization (selecting the category currently considered optimal), and gradually learn more effective classification strategies; S54. The mean squared error (MSE) loss function is used to measure the difference between the model's predicted Q value and the target Q value. The formula is as follows: , in, This represents the Q-value loss; Indicates the state Next, take action Expected rewards afterward; S55. Train and test the violation subdivision classification network, and save the optimal weights.

[0029] S6. Integrate the knowledge-distilled student model with a trained violation classification network based on the Lawformer model in the legal domain to construct a complete two-stage fair competition review classification method. After completing the initial determination of whether a policy text "contains violations," the violation classification network performs fine-grained classification of sentences determined to "contain violations," outputting the specific fair competition review clause category corresponding to the violation text; thus realizing a step-by-step review process from "violation judgment" to "violation classification." The flowchart of the two-stage fair competition review method is as follows: Figure 5 As shown.

[0030] Example 2 This invention uses a self-built fair competition review dataset. The specific details of the violation judgment dataset are as follows: the training dataset contains 7278 sentences, with 4266 violations and 3012 non-violations; the test dataset contains 1456 sentences, with 443 violations and 1022 non-violations. Experimental results show that the proposed two-stream gating network model has better performance. This experiment selected several advanced pre-trained feature extraction networks for comparative experiments, and the test results are shown in Tables 2 and 3.

[0031] Table 2. Performance of violation detection classification models with different network structures on the test dataset. All the above models are based on the Transformer encoder structure, but each has different optimization directions: the Nezha model focuses on improving the processing ability of Chinese text through innovative pre-training methods; the Lawformer model combines the professional needs of the legal field and has made specific optimizations for legal texts, enhancing its performance in legal scenarios; the Ernie model introduces a knowledge-enhanced pre-training method, making it more adaptable and generalizable in cross-domain tasks; the BERT-base-chinese model is a standard Chinese BERT model, widely used in various Chinese NLP tasks, and has strong versatility. Dual-net is a two-stream network, ours represents the violation judgment teacher model in this invention, and ours(student) represents the violation judgment student model in this invention.

[0032] Table 3. Performance of violation subdivision classification models with different network structures on the test dataset. As can be seen from the experimental results in Table 2, the dual-stream network constructed in this invention, combined with the feature decoupling mechanism, outperforms each individual model and the backbone dual-stream network model of the dual-stream feature decoupling network. This fully demonstrates the effectiveness of the feature decoupling mechanism described in this invention. Furthermore, the student model after distillation also achieves performance close to that of a single pre-trained model, even though its parameter size is only 27M, which is only 1 / 40th of the parameter size of the teacher model and 1 / 20th of the parameter size of the standard BERT. This fully demonstrates the superiority of this invention. Similarly, this invention trains and tests the violation subdivision classification model based on reinforcement learning on a violation subdivision classification dataset. The test results are shown in Table 3. As can be seen from the experimental results, this invention achieves the best results in accuracy and F1 score by utilizing the reinforcement learning training strategy, and the F1 score is significantly improved. This also shows that this strategy not only optimizes the model's ability to judge violation subdivision categories, but also has better classification ability for those categories with few samples.

[0033] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fair competition review method based on feature decoupling joint distillation and reinforcement learning, characterized in that, Includes the following steps: S1. Establish standards for the categories of violations involved in fair competition review in accordance with the latest laws and regulations, and form a complete system of categories of violations in fair competition review; S2. Obtain policy document texts and historical fair competition review violation cases, combine them with domain experts for fine annotation, and construct a fair competition review classification task dataset, which includes a violation judgment dataset and a violation sub-classification dataset; S3. Construct a teacher model for judging violations in fair competition review; the teacher model adopts a two-stream feature decoupling network of the general cross-domain pre-trained language model BERT and the legal domain Lawformer model; train the teacher model using data from the violation judgment dataset, and dynamically save the model weights with the best performance as the teacher model in the subsequent distillation stage; The data processing procedure in the two-stream feature decoupling network is as follows: S31. Input Text The first hidden representation is output after the first branch. The first branch uses the general cross-domain pre-trained language model BERT; the input text... The second hidden representation is obtained through the second branch. The second branch uses the Lawformer, a pre-trained model specifically designed for the legal field. S32. The first hidden representation Second hidden representation The inputs are fed into a shared bidirectional long short-term memory (BiLSTM) network to obtain the first new hidden representation. Second new hidden representation ; S33. The first new hidden representation Second new hidden representation After processing by an adaptive gating fusion mechanism, the fused features are obtained. The formula is expressed as follows: ， ， in, Indicates the gating weight; This represents the dot product operation; This represents the Sigmoid activation function; This indicates a splicing operation; Gate represents a nonlinear gated network based on a multilayer perceptron; the nonlinear gated network based on a multilayer perceptron includes a first fully connected layer, a second fully connected layer, and a nonlinear activation layer; S34. A factor decoder module is introduced, which consists of two independent linear layers. The first linear layer of the factor decoder performs a linear transformation on the fused features to extract the core factors related to the task objective. The second linear layer extracts the interference factor. The formula is expressed as follows: ， ， in, and These represent the weight matrices of the first and second linear layers, respectively. and These represent the bias terms of the first linear layer and the second linear layer, respectively; the core factors extracted by the first linear layer of the factor decoder are supervised by the subsequent main task branch, while the interference factors extracted by the second linear layer are supervised by the subsequent adversarial branch. S35. Introduce a gradient inversion layer (GRL) to counteract interference factors during the backpropagation phase of the adversarial branch. The features are then processed to obtain the gradient-reversed features. The features after gradient inversion After performing a pseudo-classification task using a multilayer perceptron with adversarial branching, the pseudo-classification probability distribution is obtained. The formula is expressed as follows: ， ， in, This indicates the gradient inversion layer operation; This represents an adversarial branching multilayer perceptron; This represents the pseudo-classification probability distribution of the output of the adversarial branching multilayer perceptron; S36. Core factors in the main branch The input is fed into the main branch multilayer perceptron for classification prediction, yielding the class prediction probability distribution of the main branch. The formula is expressed as follows: ， in, This represents a main branch multilayer perceptron; This represents the predicted probability value output by the main branch multilayer perceptron; S37. Decouple the core factor and the interference factor by minimizing the loss function through mutual information to constrain the decoupling process. The formula is as follows: ， ， in, Mutual information minimization loss; MSE represents mean squared error loss. This represents the Pearson correlation coefficient; This represents the covariance between the core factor and the confounding factor. and These represent the standard deviations of the core factor and the interference factor, respectively. S38. After the dual-stream feature decoupling network has been trained and converged, its key structural parameters are extracted and fixed to obtain the trained fair competition review violation judgment teacher model. S4. Construct a student model for judging violations of fair competition review. The student model adopts a shallow bidirectional gated recurrent neural network BiGRU and a state-space model Mamba. Distillation learning is used to distill the knowledge of the teacher model in the previous step into the student model. The training input of the student model is consistent with that of the teacher model. During the distillation learning process, the core decoupling factor and output probability distribution of the teacher model are used as multiple labels to pass to the student model and are trained together with the real category labels. S5. Construct a violation classification network based on the Lawformer model in the legal domain; train the violation classification network using concentrated data from the violation classification dataset based on the reinforcement learning DQN algorithm to obtain the trained violation classification network; S6. Integrate the knowledge-distilled student model with a trained violation classification network based on the Lawformer model in the legal domain to construct a complete two-stage fair competition review classification method. After completing the initial judgment on whether the policy text "contains violations", the violation classification network performs fine-grained classification on the sentences judged to "contain violations", and outputs the specific fair competition review clause category corresponding to the violation text; realize the step-by-step review process from "violation judgment" to "violation classification".

2. The fair competition review method based on feature decoupling joint distillation and reinforcement learning according to claim 1, characterized in that, The categories of violations in the fair competition review in step S1 include market access, qualification restrictions, interference with business practices, inappropriate government subsidies, and market exclusivity policies.

3. The fair competition review method based on feature decoupling joint distillation and reinforcement learning according to claim 2, characterized in that, In step S2, the collected policy document texts and historical fair competition review violation cases are manually labeled to clarify the specific clause number violated by each policy statement, and to ensure that the labeling follows the principle of one statement corresponding to one type of violation standard. After labeling, the statement content is segmented, deduplicated, and standardized, and data cleaning is performed using common noise filtering methods. Finally, the processed statements and their corresponding violation category numbers are uniformly organized into a structured table format to obtain the fair competition review classification task dataset.

4. The fair competition review method based on feature decoupling joint distillation and reinforcement learning according to claim 3, characterized in that, In step S3, the adversarial branch is trained under supervision using an adversarial loss function: ， in, Indicates adversarial loss; C represents the number of categories; This is a real label; The first digit in the probability distribution of the output of an adversarial branch multilayer perceptron represents the digit of the probability distribution. Class probability values; the main branch uses the standard cross-entropy loss function for supervised training, and the formula for the loss function is as follows: ， in, Indicates the loss of the main branch; The first branch of the multilayer perceptron output represents the first... Class prediction probability value.

5. The fair competition review method based on feature decoupling joint distillation and reinforcement learning according to claim 4, characterized in that, In step S3, the total loss of the two-stream feature decoupling network is... The formula is expressed as: ,in, , , These represent the weight coefficients corresponding to the main branch loss, adversarial loss, and mutual information minimization loss, respectively.

6. The fair competition review method based on feature decoupling joint distillation and reinforcement learning according to claim 5, characterized in that, Step S4 specifically includes: S41. The input text is processed using a pre-trained word vector embedding layer consistent with BERT, and then fed into a shallow bidirectional gated recurrent neural network (BiGRU). The output features of the shallow bidirectional gated recurrent neural network (BiGRU) are then fed into the state space model Mamba to obtain the output of the state space model Mamba. The output of the state space model Mamba is processed through an MLP to obtain the class probability distribution, and the violation judgment result is output through an argmax layer. For the input text T, the above formula is expressed as follows: ， ， ， in, The hidden feature representation extracted by the student model is represented by Mamba, the state-space model is represented by BiGRU, and the bidirectional gated recurrent unit is represented by BiGRU. This indicates the class number predicted by the student model. This represents the class probability prediction distribution output by the student model; S42. Under the condition of inputting the same text sample, predict the probability distribution of the categories of the main branch extracted in step S3. With core factors This serves as a supervisory signal during the distillation phase of the student model; multiple supervisory signals are introduced during the student model training phase through the joint distillation loss function; the formula for the joint distillation loss function is as follows: ， in, Represents the loss function of combined distillation; Indicates soft label loss; Indicates characteristic distillation loss; Indicates supervised classification loss; , , These represent the weights of the soft label loss, feature distillation loss, and supervised classification loss, respectively. Soft tag loss The distillation loss of the teacher model and the student model on the prediction distribution is expressed by the following formula: ， in, Indicates temperature parameter; Indicates KL divergence; Characteristic distillation loss The hidden feature representation extracted for the student model and the core factors decoupled from the teacher model The feature alignment loss between them is expressed by the following formula: ， in, represents the mean squared error; n represents the feature dimension. Represents the L2 norm; Supervision and classification of losses Similarly, using the binary cross-entropy loss, the formula is as follows: ， in, The first output of the student model represents the... Class prediction probability value; S43. During the training of the student model, the parameters of the teacher model are frozen until the student model converges, and the weights are saved.

7. The fair competition review method based on feature decoupling joint distillation and reinforcement learning according to claim 6, characterized in that, The violation segmentation classification network described in step S5 optimizes the classifier's decision-making ability based on a reinforcement learning strategy: S51. Construct a Q-value network, which is a linear layer; extract features from the text data in the violation subdivision classification dataset using a pre-trained Lawformer model to obtain the text feature vector; input the text feature vector into the Q-value network to estimate the Q-value for each violation category; S52. The Q-learning algorithm is used to train the Q-value network. Within the framework of the Q-learning algorithm, each violation category of the text is defined as an "action," and the feature vector of the text after being encoded by the Lawformer pre-trained model in the legal domain is the "state." The model selects an action in each state and calculates the reward based on the actual label. The formula for calculating the target Q-value is as follows: ， in, Indicates the target Q value; Indicates the discount factor; Indicates the current state; Indicates an action; This represents the maximum Q value for all actions in the next state; Represents the set of all possible actions for the next state; Indicates an immediate reward; when When the model predicts correctly, it means the prediction is correct; when... When this happens, it indicates that the model prediction is incorrect; S53. In the Q-learning algorithm, action selection adopts... ; S54. The mean squared error (MSE) loss function is used to measure the difference between the model's predicted Q value and the target Q value. The formula is as follows: ， in, Indicates Q-value loss; Indicates the state Take action below Expected rewards afterward; S55. Train and test the violation subdivision classification network, and save the optimal weights.

Images