Imbiguous term detection method and system fusing symbol rule and machine learning

Abstract

The invention discloses an discrimination clause detection method and system fusing symbol rules and machine learning, and relates to the technical field of data processing, and the method comprises the following steps: policy text semantic analysis and clause extraction; performing symbol reasoning matching on terms based on a predefined rule base; performing compliance classification and property category labeling on terms by adopting a multi-task learning model; fusing a rule judgment result and model prediction to generate a final review conclusion; and outputting an examination report containing the property and basis of the illegal terms, and updating the rule base and the model parameters based on manual feedback. According to the discrimination term detection method and system, accurate identification of hidden discrimination terms in policy texts, fine-grained property classification and explainable output of judgment basis under the small sample condition are realized through a method of fusing symbol rule reasoning and a differentiable machine learning model, and meanwhile, a continuous self-evolution capability is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a discrimination clause detection method and system fusing symbolic rules and machine learning. BACKGROUND

[0002] In recent years, with the continuous improvement of the fair competition review system in China, the identification and handling of possible discriminatory clauses in documents (such as local protection, differential treatment, market access restrictions, etc.) have become an important part of document review. The current approach mainly relies on manual review, i.e., the staff of the policy-making or document-issuing unit manually reviews the documents against the Fair Competition Review System Implementation Rules, the Anti-Monopoly Law, and other laws and regulations to determine whether there are contents that hinder the equal competition of market entities. Manual review has the advantages of clear institutional basis and clear legal responsibility, but also has significant shortcomings: first, the review efficiency is limited by human resources, and it takes a long time to respond to a large number of documents; second, manual judgment relies on the legal knowledge and business experience of the reviewer, and different people may have different interpretations of the same clause, which is highly subjective and results in a lack of consistency; finally, the manual method is difficult to update and adapt to the rapidly changing policy language and expression methods, and it is easy to miss new types of hidden discriminatory clauses.

[0003] To improve efficiency and accuracy, some institutions and research teams have tried to introduce computer-aided review techniques. The first type of method is rule-based detection, which usually relies on a keyword library or pattern matching rules summarized by experts to compare policy texts with known violations, such as detecting sensitive expressions such as "limited to local enterprises" and "designated suppliers". The advantage of this method is simplicity and strong interpretability, but the disadvantage is that it lacks adaptability to expression diversity and implicit semantics, and it is easy to miss cases if the wording of the document is different from the rule setting (e.g., using synonyms, implicit expressions).

[0004] The second type of method is text classification based on machine learning or deep learning, which trains a model by labeling historical case data to automatically determine whether a clause is in violation. For example, using support vector machines (SVM), convolutional neural networks (CNN), long short-term memory networks (LSTM), or pre-trained language models based on Transformers (such as BERT) to classify policy clauses. This method is superior to simple rule matching in capturing semantic features and can identify some implicit violations. However, its shortcomings are also obvious: the performance of the model is highly dependent on large-scale labeled data, and policy document data often involves legal and administrative fields, which has high labeling costs and data scarcity; in addition, such models are "black boxes" and it is difficult to explain their decision-making basis, making them difficult to use directly as decision-making evidence in legal supervision scenarios.

[0005] Some studies explore the strategy of combining rules with machine learning, that is, adding rule matching results as additional input in text features, or weighting and fusing the high confidence results of rule matching with model prediction. This method takes into account the explainability of rules and the generalization ability of models to some extent, but still lacks a sustainable evolving rule system, that is, once the rules are set, they need to be maintained manually and cannot be automatically updated through model training. In addition, existing methods rarely support multi-task learning, usually only output simple binary classification results (violation / non-violation), and cannot simultaneously label the nature category of violation, which limits the ability to analyze clauses in a fine-grained manner.

[0006] The existing technology has not realized the following capabilities: first, the organic integration of symbolic logic rules and machine learning algorithms to combine the accurate reasoning of expert knowledge base with the generalization ability of data-driven models; second, accurately detecting hidden discriminatory clauses and classifying their nature (such as differential treatment, restriction of competition, etc.) under limited training data conditions; third, establishing a differentiable rule learning mechanism that can continuously optimize rule parameters based on new cases and feedback to realize adaptive updating of the rule base; fourth, ensuring the transparency of the review process and the explainability of the results, so that the review conclusion is reasonable and well-founded, to improve the recognition of AI review results by regulatory authorities. Due to the lack of the above key capabilities, there is currently no mature solution that can balance the detection accuracy under small sample conditions and the explainability of model decision-making, which is a technical bottleneck that needs to be solved in the field of intelligent review of discriminatory clauses.

[0007] In summary, although the existing technology has improved the automation level of discriminatory clause detection to some extent, it still has deficiencies in accuracy, explainability, and scalability, especially under small sample conditions. How to fully utilize the advantages of rule knowledge and machine learning to achieve accurate detection and nature classification of implicit violation clauses remains a problem to be solved. SUMMARY

[0008] The purpose of the present application is to provide a discriminatory clause detection method and system that combines symbolic rules and machine learning, to solve the defects in current review technology and achieve accurate identification, fine-grained nature classification, and explainable output of judgment basis for implicit discriminatory clauses in policy texts under small sample conditions.

[0009] Embodiments of the present application are implemented as follows: A discriminatory clause detection method that combines symbolic rules and machine learning, comprising: S1. Preprocessing the input policy text, parsing and dividing it into independent clause units, and generating semantic vector representation for each clause unit; S2. match the clause unit with a pre-defined symbol rule base containing violation patterns based on fair competition regulations, and output a preliminary determination result based on rule matching and a rule matching feature vector; S3. input the semantic vector representation of the clause unit and the rule matching feature vector into a multi-task learning model, which simultaneously performs a compliance binary classification task and a violation property multi-classification task, and outputs a compliance probability and a violation property probability distribution based on model learning; S4. fuse the preliminary determination result based on rule matching and the probability output based on model learning, and generate a final clause review conclusion according to a pre-set fusion strategy; S5. output the review result, which includes the determination conclusion of the violation clause, the violation property category and the determination basis, and perform parameter optimization on the symbol rule base and the multi-task learning model based on artificial feedback.

[0010] Further, in other preferred embodiments of the present application, the step of preprocessing the policy text comprises: segmenting the policy text into clauses with independent semantics using syntax analysis and dependency tree parsing; converting each clause text into a high-dimensional semantic vector using a pre-trained language model; extracting key entities in the clause as auxiliary features through named entity recognition technology, and concatenating them with the semantic vector to form an enhanced clause vector representation.

[0011] Further, in other preferred embodiments of the present application, the rules in the symbol rule base are parameterizable logical triples, which include trigger conditions, violation types and rule weights; the rule matching step uses fuzzy logic to calculate the matching degree of the clause and the rule trigger condition, and generates a rule matching vector containing the rule matching degree information.

[0012] Further, in other preferred embodiments of the present application, the multi-task learning model has a shared text encoding layer and branches into two independent output layers: The first output layer is used for compliance binary classification, and outputs the probability of clause violation through a Sigmoid function; The second output layer is used for violation property multi-classification, and outputs the probability distribution of the clause belonging to each violation property category through a Softmax function.

[0013] Further, in other preferred embodiments of the present application, the training of the multi-task learning model uses a joint loss function, which is the weighted sum of the cross-entropy loss of the compliance binary classification task and the cross-entropy loss of the violation property multi-classification task, and the model parameters of the two tasks are optimized simultaneously through the joint loss function.

[0014] Further, in other preferred embodiments of the present application, the strategy adopted in the fusion step is: If the confidence of rule matching is higher than the preset threshold, the rule determination result is adopted as the final conclusion as a priority; If the confidence of rule matching is lower than the preset threshold, the rule matching confidence and the model prediction probability are integrated, the final violation probability is calculated by weighted average, and the determination conclusion is generated according to the probability.

[0015] Further, in other preferred embodiments of the present application, the step of parameter optimization based on artificial feedback is a differentiable process, including: Compare the correct label of artificial review with the model prediction result, and calculate the loss; Update the parameters of the multi-task learning model and the rule weight and trigger condition parameters in the rule library by the gradient descent algorithm, and realize the adaptive co-evolution of the rule library and the model.

[0016] A discrimination clause detection system fusing symbolic rules and machine learning, for implementing the above-mentioned discrimination clause detection method, comprising: A text preprocessing module for receiving policy text, cleaning, structured parsing and clause division, and generating vectorized representation of clauses; A symbolic rule engine module embedded with a symbolic rule library for rule matching reasoning on clauses and outputting rule matching results; A multi-task machine learning module containing a trained multi-task learning model for simultaneous classification and prediction of compliance and violation properties of clauses; A decision fusion module for receiving and fusing the outputs of the rule engine module and the machine learning module, and generating the final review conclusion according to the preset strategy; A result output and feedback optimization module for structured output of review reports and receiving artificial feedback information to drive adaptive update of system parameters.

[0017] Further, in other preferred embodiments of the present application, the symbolic rule engine module and the multi-task machine learning module do not run in isolation, but interact through a shared feature bus; the rule matching result is input to the multi-task learning model in the form of a vector, and part of the intermediate layer output of the multi-task learning model can also be used to assist the fuzzy matching calculation of the rule engine.

[0018] Further, in other preferred embodiments of the present application, the closed-loop optimization mechanism implemented by the result output and feedback optimization module is specifically as follows: the system presents the clauses determined to be in violation and their determination basis to the user and provides a user correction interface; the correction results submitted by the user are recorded by the system as labeled training samples, and the incremental training of the multi-task machine learning model and the parameter adjustment of the symbolic rule base are triggered periodically or in real time, forming a closed loop of "detection-feedback-learning-optimization".

[0019] The beneficial effects of the embodiments of the present application are: The embodiments of the present application provide a discriminatory clause detection method and system fusing symbolic rules and machine learning, which comprises the following steps: policy text semantic analysis and clause extraction; symbolic reasoning matching of clauses based on a pre-defined rule base; compliance classification and property category labeling of clauses using a multi-task learning model; generation of a final review conclusion by fusing rule determination results and model prediction; output of a review report containing the properties and basis of the clauses in violation, and updating of the rule base and model parameters based on artificial feedback. The discriminatory clause detection method and system fuses symbolic rule reasoning and differentiable machine learning models, realizes accurate identification of implicit discriminatory clauses in policy texts, fine-grained property classification, and interpretable output of determination basis under small sample conditions, and has a continuous self-evolution capability. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are some of the embodiments of the present application but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application. Therefore, the following detailed description of the embodiments of the present application is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0021] The abbreviations and key terms involved in the following embodiments are defined as follows: EMBODIMENT

[0022] The embodiments provide a discriminatory clause detection method fusing symbolic rules and machine learning, which comprises: S1. Preprocessing of the input policy text, analyzing and dividing it into independent clause units, and generating semantic vector representations of each clause unit.

[0023] Further, the step of preprocessing the policy text comprises: The policy text is segmented into clauses with independent semantics using syntactic analysis and dependency tree parsing. Each clause text is converted into a high-dimensional semantic vector using a pre-trained language model (such as BERT): , wherein, s i represents the i th clause text, h i is the corresponding semantic vector representation, d is the vector dimension.

[0024] The key entities in the clause are extracted as auxiliary features through named entity recognition technology, and are concatenated with the semantic vector to form an enhanced clause vector representation.

[0025] Optionally, named entity recognition (NER) is performed to extract elements such as "region", "subject type", and "constraint verb" as auxiliary features to form an enhanced input: , wherein, e i is the entity feature vector, and the ";" symbol represents vector concatenation.

[0026] This step combines the BERT semantic representation with the entity features, allowing the model to capture both semantic and structural information of the clause, providing high-quality input for subsequent reasoning.

[0027] Further, the discrimination clause detection method provided by the embodiment fuses symbolic rules and machine learning, and further comprises: S2. Match the clause unit with a pre-defined symbolic rule library, which contains violation patterns based on fair competition regulations, and output the preliminary judgment result based on rule matching and rule matching feature vector.

[0028] Further, for the preprocessed clause text, a pre-constructed symbolic rule library is used for reasoning matching. The symbolic rule library is constructed as follows: , Each rule is represented as a logical triple, which includes a trigger condition, a violation type, and a rule weight: , wherein, C j is the trigger condition set (keyword / dependency pattern); T j is the corresponding violation type;w j ∈[0,1] is a rule weight.

[0029] The rule base is composed of a set of discrimination criteria summarized by industry experts according to regulations such as the Fair Competition Review Implementation Rules. Each rule defines the characteristic pattern of a certain type of discriminatory clause, such as "contains local priority label", "limits the qualification of a specific subject", etc. The rule engine traverses the rule base and determines whether each clause meets the triggering conditions of a certain rule: if the clause text contains the key word combination or structural pattern set by the rule, the rule engine outputs the preliminary judgment result and confidence score corresponding to the rule.

[0030] The rule matching step uses fuzzy logic to calculate the matching degree of the clause and the rule triggering condition: , A rule matching vector containing the matching degree information of each rule is generated: .

[0031] The rule matching vector output in this step will be one of the input features of the subsequent machine learning model to integrate symbolic knowledge into the model judgment. Through the fuzzy matching mechanism, it can identify semantically similar but differently expressed violation clauses, improving the robustness and coverage of the rule system.

[0032] Further, the discrimination clause detection method provided by the embodiment of the application further comprises: S3. The semantic vector representation of the clause unit and the rule matching feature vector are jointly input into a multi-task learning model, and the multi-task learning model simultaneously performs the compliance binary classification task and the violation property multi-classification task, and outputs the compliance probability and violation property probability distribution based on model learning.

[0033] Specifically, while obtaining the rule matching result, the semantic vector of the clause is input into the deep learning model for compliance discrimination and property label prediction. The embodiment uses a multi-task learning framework to build a neural network model: a shared text encoding layer extracts high-dimensional semantic feature representation of the clause, and then branches into two output layers: The first output layer is used for compliance binary classification (compliant / non-compliant) and outputs the probability of clause violation through the Sigmoid function: , wherein, represents the probability that the i-th clause is judged as "violation" (the value range is 0~1). If i >0.5, the clause is judged as non-compliant; σ ​(·) is a Sigmoid activation function, which is used to map the linear transformation result to a probability value, and the formula is: σ ( x )=1 / (1+ e −x ); W 1 is the weight matrix of the compliance classification branch, with a dimension of , where d z is the dimension of the input vector; b 1 is the bias term of the classification branch, which is a scalar used to adjust the classification boundary.

[0034] The second output layer is used for multi-classification of violation properties (such as regional discrimination, differential treatment, and restriction of competition), and outputs the probability distribution of each violation property category to which the clause belongs through the Softmax function: , where, represents the probability distribution vector of the i clause in each violation category, such as regional discrimination, differential treatment, and restriction of competition. Its dimension is K , where K is the total number of violation types; is used for multi-classification output, which normalizes the score of each category to a probability. The Softmax function converts the linear score of each category to a probability distribution, so that the sum of all elements of is 1.

[0035] Further, the training of the multi-task learning model uses a joint loss function, which is the weighted sum of the cross-entropy loss of the compliance binary classification task and the cross-entropy loss of the violation property multi-classification task. Through this joint loss function, the model parameters of the two tasks are optimized simultaneously.

[0036] The joint loss function is defined as: , where, is the binary cross-entropy loss; is the multi-class cross-entropy loss; α is the task weight hyperparameter.

[0037] The training goal of the model is to minimize the loss function of both the compliance discrimination and property classification tasks simultaneously, thereby improving the learning efficiency on small sample data in parameter sharing. The multi-task joint loss function is .

[0038] To incorporate the symbolic rule knowledge into the model, a rule matching vector is added as an additional feature at the encoding layer or the output layer of the model, or an additional loss weight penalty is given to the samples that violate the rules. This differentiable rule fusion mechanism enables the model to adjust the degree of dependence on rule features during the training process: when a certain rule frequently misjudges, the model will reduce its weight; when the rule is consistent with the training data, the model will strengthen its role. Through this strategy, the model gradually learns to make a comprehensive judgment on rules and data evidence, improving overall performance. Ultimately, the multi-task model obtained through iterative training can output the violation probability of each clause and the possible violation type.

[0039] Further, the discrimination clause detection method provided by the embodiment fuses symbolic rules and machine learning, and further comprises: S4. Fusing the preliminary judgment result based on rule matching and the probability output based on model learning, and generating a final clause review conclusion according to a preset fusion strategy.

[0040] The output probability of the symbolic rule engine and the output probability of the machine learning model to generate the final review judgment result.

[0041] , wherein λ ∈ [0, 1] is a fusion weight, which is dynamically adjusted to balance the influence of rules and models.

[0042] Specifically, for each clause, the judgment result of the rule engine and the model prediction result are fused: if a clause is judged by the rule engine as a discriminatory clause with high confidence, it can be directly marked as non-compliant and the corresponding property label is given, that is, when higher than the threshold value, directly outputting “non-compliant”; otherwise, the violation probability output by the model is considered, and the and confidence are determined.

[0043] When the model and the rule opinion are inconsistent, the influence of the two can be balanced through a threshold setting or a weighting strategy. For example, when the rule is not triggered but the model predicts a high risk, the clause is still judged as a suspected violation; on the contrary, when the rule is triggered but the model has low confidence, it can be further reviewed by artificial.

[0044] This fusion strategy ensures the robustness of the system, significantly reduces the misjudgment rate while ensuring “zero missed judgment”, neither misses the situations that are clearly prohibited by rules, nor identifies new types of unfairness that evade artificial rules through models, while balancing the rule interpretability and model flexibility.

[0045] Further, the discrimination clause detection method provided by the embodiment fuses symbolic rules and machine learning, and further comprises the following steps: S5. Output the review result, which comprises the determination conclusion of the non-compliant clause, the non-compliant nature category, and the determination basis, and perform parameter optimization on the symbolic rule library and the multi-task learning model based on artificial feedback.

[0046] Specifically, after detecting all the clauses in the entire document, the system generates a structured review result output For each clause determined to be non-compliant, the original content, the determination result (for example, "non-compliant-discrimination clause exists"), the specific non-compliant nature, and the determination basis are output. For example, if a clause is marked as "discrimination", the system will list the triggered rule item (for example, "local enterprise preferential procurement" violates article X of the fair competition review regulations) or the key expression focused by the model for reference of the reviewer. The compliant clause can also be marked as "no problem found".

[0047] Further, the parameter optimization based on artificial feedback is a differentiable process, comprising the following steps: Compare the correct label of artificial review with the model prediction result, and calculate the loss; Simultaneously update the parameters of the multi-task learning model and the rule weight and trigger condition parameters in the symbolic rule library through the gradient descent algorithm, to realize the adaptive co-evolution of the rule library and the model.

[0048] Specifically, the system is designed with an expert feedback interface; for misjudged samples, the system introduces a feedback learning mechanism: , wherein, η is the learning rate, y i is the artificial review result.

[0049] Through the above steps, a closed-loop discrimination clause detection framework is formed, which can continuously learn and improve, and realize the synchronous update of the rule library and the model parameters, to perform efficient and interpretable compliance review on policy texts.

[0050] The closed-loop self-optimization strategy realizes the closed-loop optimization of "determination-feedback-update", so that the detection system can continuously learn and adapt to new policy language expressions, and significantly improves the long-term accuracy and stability.

[0051] Further, the embodiment further provides a discrimination clause detection system fusing symbolic rules and machine learning, for realizing the discrimination clause detection method, which comprises the following: A text preprocessing module is configured to receive a policy text, perform cleaning, structured analysis, and clause division, and generate a vectorized representation of the clause. a symbolic rule engine module embedded with a symbolic rule library for rule matching inference on the clauses and outputting rule matching results; a multi-task machine learning module containing a trained multi-task learning model for simultaneous classification and prediction of compliance and non-compliance properties of the clauses; a decision fusion module for receiving and fusing the outputs of the rule engine module and the machine learning module and generating a final review conclusion according to a preset strategy; a result output and feedback optimization module for structuring a review report and receiving artificial feedback information to drive adaptive updating of system parameters.

[0052] Further, the symbolic rule engine module and the multi-task machine learning module do not run in isolation, but interact through a shared feature bus; the rule matching results are input to the multi-task learning model in the form of a vector, and part of the intermediate layer output of the multi-task learning model can also be used to assist the fuzzy matching calculation of the rule engine.

[0053] Further, the closed-loop optimization mechanism implemented by the result output and feedback optimization module is as follows: the system presents the clauses determined as non-compliant and their determination basis to the user and provides a user correction interface; the correction results submitted by the user are recorded by the system as labeled training samples, and the incremental training of the multi-task machine learning model and the parameter adjustment of the symbolic rule library are triggered periodically or in real time, forming a closed loop of "detection-feedback-learning-optimization".

[0054] In summary, the embodiment of the present application provides a discriminatory clause detection method and system that fuses symbolic rules and machine learning, which includes the following steps: policy text semantic analysis and clause extraction; symbolic inference matching of the clauses based on a pre-defined rule library; compliance classification and property category labeling of the clauses using a multi-task learning model; fusion of rule determination results and model prediction to generate a final review conclusion; output of a review report containing non-compliant clause properties and basis, and updating of the rule library and model parameters based on artificial feedback. The discriminatory clause detection method and system achieve precise identification of implicit discriminatory clauses in policy texts, fine-grained property classification, and interpretable output of determination basis under small sample conditions by fusing symbolic rule inference and differentiable machine learning models, and have continuous self-evolution capability.

[0055] The above only describes preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting discriminatory clauses that integrates symbolic rules and machine learning, characterized in that, include: S1. Preprocess the input policy text, parse it and divide it into independent clause units, and generate a semantic vector representation for each clause unit; S2. Match the clause unit with a predefined symbol rule base, which contains violation patterns based on fair competition regulations, and output a preliminary judgment result based on rule matching and a rule matching feature vector; S3. Input the semantic vector representation of the clause unit and the rule matching feature vector into a multi-task learning model. The multi-task learning model simultaneously performs a compliance binary classification task and a violation multi-classification task, and outputs the compliance probability and violation probability distribution based on model learning. S4. Integrate the preliminary judgment result based on rule matching with the probability output based on model learning, and generate the final clause review conclusion according to the preset fusion strategy; S5. Output the review results, which include the judgment conclusion of the violation clause, the nature and category of the violation and the judgment basis, and optimize the parameters of the symbol rule base and the multi-task learning model based on human feedback.

2. The method for detecting discriminatory clauses according to claim 1, characterized in that, The steps for preprocessing the policy text include: Using syntactic analysis and dependency tree parsing, policy texts are segmented into clauses with independent semantics; Each clause text is converted into a high-dimensional semantic vector using a pre-trained language model; Key entities in the clauses are extracted using named entity recognition technology as auxiliary features, and then concatenated with the semantic vector to form an enhanced clause vector representation.

3. The method for detecting discriminatory clauses according to claim 1, characterized in that, The rules in the symbol rule base are represented in the form of parameterizable logical triples, which include triggering conditions, violation types, and rule weights. The rule matching step uses fuzzy logic to calculate the matching degree between the clauses and the rule triggering conditions, and generates a rule matching vector containing matching degree information for each rule.

4. The method for detecting discriminatory clauses according to claim 1, characterized in that, The multi-task learning model has a shared text encoding layer and branches into two independent output layers: The first output layer is used for compliance binary classification, and outputs the probability of clause violation through the Sigmoid function; The second output layer is used for multiple classification of violation types. It outputs the probability distribution of each clause belonging to each violation type category through the Softmax function.

5. The method for detecting discriminatory clauses according to claim 4, characterized in that, The multi-task learning model is trained using a joint loss function, which is a weighted sum of the cross-entropy loss of the compliance binary classification task and the cross-entropy loss of the violation multi-class classification task. This joint loss function optimizes the model parameters for both tasks simultaneously.

6. The method for detecting discriminatory clauses according to claim 1, characterized in that, The strategy adopted in the fusion step is as follows: If the confidence level of the rule matching is higher than the preset threshold, the rule judgment result will be adopted as the final conclusion. If the confidence level of rule matching is lower than a preset threshold, the final violation probability is calculated by weighted averaging of the rule matching confidence level and the model prediction probability, and a judgment conclusion is generated based on this probability.

7. The method for detecting discriminatory clauses according to claim 1, characterized in that, The parameter optimization step based on human feedback is a differentiable process, including: The correct labels verified by humans are compared with the model prediction results, and the loss is calculated. By using the gradient descent algorithm, the parameters of the multi-task learning model and the rule weights and trigger condition parameters in the symbol rule base are updated simultaneously, thereby achieving adaptive co-evolution between the rule base and the model.

8. A discriminatory clause detection system integrating symbolic rules and machine learning, used to implement the discriminatory clause detection method according to any one of claims 1-7, characterized in that, include: The text preprocessing module is used to receive policy texts, perform cleaning, structured parsing and clause segmentation, and generate vectorized representations of the clauses; The symbol rule engine module, which has an embedded symbol rule library, is used to perform rule matching reasoning on the terms and output the rule matching results. The multi-task machine learning module contains a trained multi-task learning model for simultaneously classifying and predicting the compliance and non-compliance nature of the terms. The decision fusion module is used to receive and fuse the outputs of the rule engine module and the machine learning module, and generate a final review conclusion according to a preset strategy. The results output and feedback optimization module is used to output structured review reports and receive human feedback to drive adaptive updates of system parameters.

9. The discriminatory clause detection system according to claim 8, characterized in that, The symbol rule engine module and the multi-task machine learning module do not run in isolation, but interact with each other through a shared feature bus; the rule matching result is input to the multi-task learning model as a vector, and some intermediate layer outputs of the multi-task learning model can also be used to assist the fuzzy matching calculation of the rule engine.

10. The discriminatory clause detection system according to claim 8, characterized in that, The closed-loop optimization mechanism implemented by the result output and feedback optimization module is as follows: the system presents the clauses that are judged to be in violation and the basis for the judgment to the user, and provides a user correction interface; the correction results submitted by the user are recorded by the system as labeled training samples, and the incremental training of the multi-task machine learning model and the parameter adjustment of the symbol rule base are triggered periodically or in real time, forming a closed loop of "detection-feedback-learning-optimization".