Core thinking chain multi-scene precise service method for party and century learning and education scene

Abstract

The invention is suitable for the field of discipline inspection and supervision, and provides a core thinking chain multi-scene precise service method for party and century learning and education scenes. A core thinking chain including'problem extraction-problem optimization-intention judgment-problem response-result optimization-result output 'is constructed, requirements can be split in multiple dimensions according to personnel identities, time nodes and behavior types, and through technologies such as keyword professional conversion, regulation version accurate matching and user identity dynamic adaptation, the requirements can be met. Knowledge base contents such as in-party laws and regulations, discipline inspection and supervision terminologies, typical cases and the like are accurately called; according to the method, the professional supporting capacity is enhanced, the key information extraction accuracy is improved, dual accurate adaptation is achieved, retrieval and output logic is optimized, the illusion rate is strictly controlled, and authority is guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of discipline inspection and supervision, and in particular relates to a multi-scenario precise service method for the core thinking chain in the context of Party discipline learning and education. Background Technology

[0002] Currently, in Party discipline learning and education scenarios, general-purpose large language models have been gradually applied to service scenarios such as discipline and law Q&A, competition quizzes, and case analysis. Its core working principle is: based on publicly available internet data and a basic general knowledge base, it uses natural language processing technology to semantically understand user input questions and directly generate response content. Some applications supplement this with simple keyword search functionality, attempting to match relevant content from publicly available Party regulations texts to assist in generating answers, thus meeting the Party discipline learning needs of Party members and ordinary Party members. This type of technology does not require building customized processes for specific fields, relies on the basic semantic understanding capabilities of general-purpose models, has low deployment costs, and is suitable for general knowledge Q&A scenarios.

[0003] However, existing technologies have the following problems: Lack of professional terminology and business logic: Existing general-purpose language models have not been specifically trained for the field of discipline inspection and supervision, lack a precise understanding of discipline inspection and supervision terminology, and have not incorporated the unique business logic of discipline inspection and supervision work, resulting in insufficient professionalism in the output content and failing to meet the professional needs of discipline education; Poor targeting of key information extraction: Existing technology can only perform general keyword extraction and has not established an extraction mechanism for the core dimensions of discipline inspection and supervision scenarios, making it impossible to accurately capture the key elements in the problem, thus creating hidden dangers for subsequent accurate matching; Lack of adaptation between regulatory versions and user identities: Existing technology does not establish a mechanism to link case time with regulatory versions, which can easily lead to logical errors in judging past behavior using current regulations; at the same time, it cannot dynamically identify user identities, outputting the same content to users with different identities, which lacks targeting. The search strategy is too simplistic: Existing technologies mostly use single keyword search or general semantic matching methods, without designing differentiated search strategies based on the different intentions of discipline inspection and supervision scenarios. This results in low matching accuracy of content such as laws, cases, and standards, and makes it difficult to quickly locate the core evidence. The illusion rate remains high: the general language model has a high illusion rate in professional fields and lacks a specific verification mechanism for Party discipline learning scenarios. The output is prone to erroneous information that does not conform to Party regulations and typical cases, which affects the seriousness and accuracy of discipline education. Therefore, a multi-scenario precise service method based on the core thinking chain for Party discipline learning and education is needed to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The core thinking chain for Party discipline learning and education scenarios provides a multi-scenario precision service method, which includes building a core thinking chain for discipline education based on large language models, local knowledge bases, retrieval-enhanced generation (RAG) and knowledge graph technology, and achieving precision service through multi-dimensional demand decomposition and a six-step professional process. The six-step process is as follows: Receive questions related to Party discipline learning and education; Extract key information from the problem, including the identity of the person involved, the time frame, the behavior, the situation of the problem, and the industry sector. Optimize problem information according to the characteristics of discipline inspection and supervision scenarios, including the transformation of non-standard keywords and the rewriting of problems; Analyze the intent of the problem and categorize it; Differentiated knowledge retrieval, logical processing, and response generation are performed based on intent; Optimize and output response content according to the principle of "prioritizing standardized reminders"; The local knowledge base includes Party regulations, laws and regulations, rules and regulations, professional terminology of discipline inspection and supervision, business logic and typical cases. By accurately matching multi-dimensional needs with the content of the knowledge base, the illusion rate of large models is reduced.

[0006] A further technical solution involves calculating the weights of key information extracted from the problem using the following formula: ; in, For the first The extraction weight of key information categories (personnel identity / time point / behavior / problem situation / industry sector), This is an information importance coefficient (ranging from 0.6 to 0.8), used to reflect the priority of core dimensions in the discipline inspection and supervision scenario. For the first The importance score of information type in the disciplinary education scenario (0-10 points). The information correlation coefficient (with a value of 0.2-0.4, and...) ), For the first The relevance score between the information and the core demands of the problem (0-10 points, 10 points for complete relevance). We can prioritize extracting key information that has core value for subsequent processes, avoid blind information extraction, ensure that no core elements are omitted, and provide a high-quality data foundation for subsequent problem optimization and intent judgment.

[0007] Further technical solutions, including keyword conversion for problem optimization, employ the following formula: ; in, For input questions containing non-standard keywords and colloquial expressions (such as "pig slaughtering feast", "image project", "political achievement project", etc.). This is a dictionary of professional terminology for discipline inspection and supervision. It includes commonly used expressions in Party regulations, standardized terms used in discipline inspection and supervision business scenarios, and a table showing the correspondence between non-standard expressions and professional terms. This is a dictionary-based keyword mapping function that outputs standardized keywords by querying a pre-defined mapping relationship in the dictionary. (For example, transforming "pig-slaughtering feast" into "banquet" and associating "image project" with "illegal construction activities that are divorced from reality and waste manpower and resources"); Ensure semantic consistency between the question description and the knowledge base content, solve the problem of understanding non-standard expressions, and improve the accuracy of subsequent retrieval and matching.

[0008] A further technical solution is intended to determine the use of a classification formula with a confidence threshold: other ; in, This is the optimized question text vector, generated through a semantic encoding model, which includes both semantic features and business dimension features of the question. This is an intent classification model based on BERT pre-training and fine-tuned using inspection and monitoring scenario data, outputting intent categories. (Including checking discipline, boundaries, regulations, standards, policies, and cases) For classification confidence, The confidence threshold is set (ranging from 0.75 to 0.85). If the confidence level reaches the threshold, the corresponding intent will be applied. If the confidence level does not reach the threshold, the search will be classified as "other" and a generalized search will be initiated. Accurately identify users' core needs, avoid deviations in response direction, provide clear guidance for differentiated searches, and improve overall service response efficiency.

[0009] A further technical solution involves matching case timelines with regulatory versions using a formula: ; in, To best match the regulatory version, This is a collection of valid versions of Party regulations (including the effective and expiration dates and scope of application for each version). This refers to the time of the incident or the duration of the behavior extracted from the question. For version Effective date For version Expiration time (if not expired, take the current time) ), For indicator functions (satisfying) exist The value is 1 if the interval is specified, otherwise it is 0. For version Relevance to the problem industry sector (value range: 0-1); It automatically filters out the relevant and applicable versions of laws and regulations at the time the case occurred, avoiding errors in the application of laws and regulations and providing accurate timeliness basis for qualitative and quantitative disciplinary actions.

[0010] Further technical solutions, such as user identity matching with regulatory content, adopt the following formula: ; in, A set of regulations to be adapted. This is a collection of all legal content in the local knowledge base (including clause texts, applicable objects, and quantitative standards). To extract user identities from the questions (such as county-level leading cadres, ordinary party members, and discipline inspection and supervision staff). For the content of regulations Clearly defined scope of applicable identities For determining identity affiliation, when belong When a subset of the law is included, the content of the law is... Included in the adaptation set; Provide targeted legal content for users with different identities to avoid a "one-size-fits-all" approach, improve content adaptability and user experience, and meet the differentiated needs of discipline inspection and supervision cadres and ordinary party members.

[0011] A further technical solution, the formula for the differentiated retrieval strategy for problem response, is as follows: ; in, For the final search results, To retrieve weighting coefficients, when searching for regulations or standards. (Focusing on precise text matching), when searching for cases or boundary-based intents. (Focusing on related searches) To enhance the generated text matching results for retrieval (based on keyword and semantic similarity matching of legal clauses and standard provisions). The entity association retrieval results for the knowledge graph (including the association relationships of legal clauses, cases, and terms); Achieve intent-driven, precise retrieval, balancing the accuracy of text matching with the comprehensiveness of related retrieval, and improve the accuracy and coverage of knowledge matching.

[0012] A further technical solution, the optimized content sorting formula is as follows: ; in, To optimize the output sequence, In response to the generation of a content list, (Standard reminder weight) (Weight of punitive measures) (Based on the weight of legal provisions) To standardize the relevance score of reminders, To score the relevance of the punitive measures, The score is based on the relevance of the legal provisions (0-10 points). It is a descending sorting function; This forms a clear logic of "standard reminder - legal basis - punishment measures," highlighting the warning orientation of disciplinary education and improving the logic and educational effect of the response content.

[0013] A further technical solution involves using a cross-validation formula to control the hallucination rate: ; in, For the final illusion rate, For local knowledge base With knowledge graph The number of errors in the dual-validation judgment ( , To validate the number of errors in the knowledge base, (Number of validation errors for knowledge graph) This represents the total number of output contents. Strictly control the error rate of output content, ensure that the response content is consistent with Party regulations and typical cases, and maintain the authority and seriousness of discipline education.

[0014] Further technical solutions and multi-scenario adaptation formulas are as follows: ; in, For scenario-based service processes, Based on the six-step process vector, To adapt operators to the process, Scenario Adaptation Matrix (Question and Answer Consultation Scenario) Competition question-and-answer scenario Case analysis scenario The matrix elements represent the resource allocation weights for the corresponding process steps; It enables flexible expansion of the basic process to multiple scenarios, meeting diverse needs such as Q&A consultation, competition answering, and case analysis, and improving the versatility and practicality of the method.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention enhances professional support capabilities: through a keyword mapping and conversion mechanism, an intent classification mechanism, and a differentiated retrieval strategy, it constructs a dictionary of professional terms for discipline inspection and supervision and a dedicated business logic mapping mechanism, accurately understands professional terms and business scenarios, solves the problem of missing professional terms and business logic in existing technologies, and ensures that the response content fully complies with the norms of discipline inspection and supervision work. This invention improves the accuracy of key information extraction: by using a multi-dimensional decomposition framework and a key information extraction weighting mechanism, it can specifically extract core dimension information such as personnel identity and time nodes, solving the problem of poor targeting of key information extraction in existing technologies and providing reliable data support for subsequent accurate matching. This invention achieves dual precise adaptation: through a precise matching mechanism of legal version and a user identity adaptation mechanism, a dual adaptation module of "time-law" and "identity-content" is constructed, which automatically matches the legal version at the time of the case and the corresponding adaptation content of the user identity, solving the problem of the lack of adaptation in existing technologies and providing precise basis for qualitative and quantitative disciplinary and legal clarification. This invention optimizes the retrieval and output logic: through differentiated retrieval strategies and result optimization sorting mechanisms, it achieves intent-driven precise retrieval and "warning priority" output sorting, solving the problems of single retrieval and chaotic logic in existing technologies, and significantly improving the accuracy of the legal basis and the hit rate of core information in the response content; This invention strictly controls the illusion rate to ensure authority: by supporting a local knowledge base, employing a dual retrieval strategy, and establishing a dual verification control mechanism for the illusion rate, a multi-level content verification system is constructed. This effectively reduces the illusion rate, solves the problem of high illusion rates in existing technologies, ensures that the response content is completely consistent with Party regulations and typical cases, and maintains the authority and seriousness of disciplinary education.

[0016] To more clearly illustrate the structural features and effects of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0020] like Figure 1 As shown, this embodiment of the invention provides a multi-scenario precise service method for the core thinking chain in the context of Party discipline learning and education. This method includes constructing a core thinking chain for discipline education based on large language models, local knowledge bases, retrieval-enhanced generation (RAG), and knowledge graph technology. Precise services are achieved through multi-dimensional demand decomposition and a six-step professional process. The six-step process is as follows: Receive questions related to Party discipline learning and education; Extract key information from the problem, including the identity of the person involved, the time frame, the behavior, the situation of the problem, and the industry sector. Optimize problem information according to the characteristics of discipline inspection and supervision scenarios, including the transformation of non-standard keywords and the rewriting of problems; Analyze the intent of the problem and categorize it; Differentiated knowledge retrieval, logical processing, and response generation are performed based on intent; The response content was optimized and output according to the principle of "prioritizing standardized reminders".

[0021] In this embodiment, the local knowledge base includes Party regulations, laws and regulations, rules and regulations, professional terms of discipline inspection and supervision, business logic and typical cases. By accurately matching multi-dimensional needs with the content of the knowledge base, the illusion rate of large models is reduced. The core thinking chain construction and six-step process design, through multi-dimensional demand decomposition and full-process professional management, solve the problem of poor targeting of key information extraction in existing technologies, provide framework support for precise processing in each link, and ensure that services focus on the core needs of discipline inspection and supervision scenarios.

[0022] Specifically, the formula for calculating the weight of key information extracted from the problem is as follows: ; in, For the first The extraction weight of key information categories (personnel identity / time point / behavior / problem situation / industry sector), This is an information importance coefficient (ranging from 0.6 to 0.8), used to reflect the priority of core dimensions in the discipline inspection and supervision scenario. For the first The importance score of information type in the disciplinary education scenario (0-10 points). The information correlation coefficient (with a value of 0.2-0.4, and...) ), For the first The score for the relevance of the information to the core demands of the problem (0-10 points, 10 points for complete relevance).

[0023] In this embodiment, the key information extraction weighting mechanism quantifies the importance and relevance of information, clarifies the priority of core dimensions in the discipline inspection and supervision scenario, avoids the blindness of information extraction in existing technologies, ensures that no key elements are omitted, and lays a high-quality data foundation for subsequent problem optimization and intent judgment.

[0024] Specifically, the keyword conversion for question optimization uses the following formula: ; in, For input questions containing non-standard keywords and colloquial expressions (such as "pig slaughtering feast", "image project", "political achievement project", etc.). This is a dictionary of professional terminology for discipline inspection and supervision. It includes commonly used expressions in Party regulations, standardized terms used in discipline inspection and supervision business scenarios, and a table showing the correspondence between non-standard expressions and professional terms. This is a dictionary-based keyword mapping function that outputs standardized keywords by querying a pre-defined mapping relationship in the dictionary. (For example, transforming "pig-killing feast" into "banquet" and associating "image project" with "illegal construction behavior that is divorced from reality and wastes manpower and resources").

[0025] In this embodiment, the keyword mapping and transformation mechanism constructs a mapping function based on a dictionary of professional terms for discipline inspection and supervision, which realizes the accurate transformation of non-standard keywords into professional terms, solves the problem that the general model has difficulty understanding non-standard expressions of discipline inspection and supervision, and ensures effective matching between questions and knowledge base content.

[0026] Specifically, the intent determination uses a classification formula with a confidence threshold: other ; in, This is the optimized question text vector, generated through a semantic encoding model, which includes both semantic features and business dimension features of the question. This is an intent classification model based on BERT pre-training and fine-tuned using inspection and monitoring scenario data, outputting intent categories. (Including checking discipline, boundaries, regulations, standards, policies, and cases) For classification confidence, The confidence threshold is set (ranging from 0.75 to 0.85). If the confidence level reaches the threshold, the corresponding intent will be applied. If the confidence level does not reach the threshold, the category will be classified as "other" and a generalized search will be initiated.

[0027] In this embodiment, the intent classification mechanism with confidence threshold filters valid intent categories through confidence threshold, avoiding the problem of ambiguous intent judgment in existing technologies. It can accurately identify core intents such as checking regulations and laws, providing clear guidance for differentiated retrieval and improving response efficiency.

[0028] Specifically, the formula for matching case dates with regulatory versions is as follows: ; in, To best match the regulatory version, This is a collection of valid versions of Party regulations (including the effective and expiration dates and scope of application for each version). This refers to the time of the incident or the duration of the behavior extracted from the question. For version Effective date For version Expiration time (if not expired, take the current time) ), For indicator functions (satisfying) exist The value is 1 if the interval is specified, otherwise it is 0. For version Relevance to the problem industry sector (value range 0-1).

[0029] In this embodiment, the precise matching mechanism for regulatory versions establishes a rigid correlation between the case time and the effective and expiration times of regulations by determining the time interval and weighting the relevance to the field. This completely solves the problem of inaccurate adaptation of existing technical regulatory versions and ensures the timeliness and legality of the response basis.

[0030] Specifically, the formula for matching user identity with legal content is as follows: ; in, A set of regulations to be adapted. This is a collection of all legal content in the local knowledge base (including clause texts, applicable objects, and quantitative standards). To extract user identities from the questions (such as county-level leading cadres, ordinary party members, and discipline inspection and supervision staff). For the content of regulations Clearly defined scope of applicable identities For determining identity affiliation, when belong When a subset of the law is included, the content of the law is... It was included in the adaptation set.

[0031] In this embodiment, the user identity and regulatory content matching mechanism determines the affiliation of users and achieves accurate matching between users with different identities and the corresponding regulatory content. This solves the problem of poor user identity matching in the existing technology and makes the response content more targeted and practical.

[0032] Specifically, the formula for the differentiated retrieval strategy for question response is: ; in, For the final search results, To retrieve weighting coefficients, when searching for regulations or standards. (Focusing on precise text matching), when searching for cases or boundary-based intents. (Focusing on related searches) To enhance the generated text matching results for retrieval (based on keyword and semantic similarity matching of legal clauses and standard provisions). This refers to the entity association retrieval results of the knowledge graph (including the association relationships of legal clauses, cases, and terms).

[0033] In this embodiment, the differentiated retrieval strategy, combining the advantages of retrieval enhancement generation and knowledge graph-related retrieval, dynamically adjusts the retrieval weight according to the intent type, thereby improving the comprehensiveness and accuracy of knowledge retrieval. This solves the problem of incomplete coverage by the single retrieval method in the prior art and provides rich content support for high-quality responses.

[0034] Specifically, the formula for ranking the optimized content is as follows: ; in, To optimize the output sequence, In response to the generation of a content list, (Standard reminder weight) (Weight of punitive measures) (Based on the weight of legal provisions) To standardize the relevance score of reminders, To score the relevance of the punitive measures, The score is based on the relevance of the legal provisions (0-10 points). This is a descending sorting function.

[0035] In this embodiment, the result optimization sorting mechanism optimizes the output sequence by reasonably allocating the weights of normative reminders, punitive measures, and legal basis, following the logic of "warning first, basis second, and punishment supplement." This meets the core requirements of discipline education, solves the problems of chaotic logic and lack of focus in existing technologies, and enhances the educational and guiding effect.

[0036] Specifically, the hallucination rate control uses a cross-validation formula: ; in, For the final illusion rate, For local knowledge base With knowledge graph The number of errors in the dual-validation judgment ( , To validate the number of errors in the knowledge base, (Number of validation errors for knowledge graph) This represents the total number of output contents.

[0037] In this embodiment, the illusion rate dual verification control mechanism strictly controls the error rate of the output content through dual cross-verification of the local knowledge base and the knowledge graph, which solves the problem of high illusion rate in the prior art and ensures the accuracy and seriousness of the response content.

[0038] Specifically, the multi-scenario adaptation formula is as follows: ; in, For scenario-based service processes, Based on the six-step process vector, To adapt operators to the process, Scenario Adaptation Matrix (Question and Answer Consultation Scenario) Competition question-and-answer scenario Case analysis scenario The matrix elements represent the resource allocation weights for the corresponding process steps.

[0039] In this embodiment, the multi-scenario adaptation mechanism adjusts the resource allocation weight of the basic process through the scenario adaptation matrix, enabling the basic service process to be flexibly extended to multiple scenarios such as Q&A consultation, competition answering, and case analysis. This solves the problem of poor scenario adaptability in existing technologies and can meet the diverse needs of Party discipline learning and education.

[0040] Working principle and usage process of this invention: This method is supported by four core technologies: large language model, local knowledge base, retrieval augmentation generation (RAG), and knowledge graph. It constructs a core thinking chain specifically for discipline education, and achieves precise services in multiple scenarios through a closed-loop logic of "input - six-step processing - output - continuous service". Each step of the process forms a deep integration of technology and business. The specific working principle is as follows: (a) Problem Input: Receiving requirements from multiple channels and identifying scenarios; Users can input various needs through platforms related to Party discipline learning and education (including mini-programs, apps, and web pages), with input formats including text questions, case descriptions, and competition question entries. Upon receiving input, the system automatically identifies and labels the scenario type, which includes Q&A consultations (such as inquiries about Party discipline and laws by individual Party members), competition quizzes (such as answering questions in Party discipline knowledge competitions), case analyses (such as qualitative and quantitative disciplinary analyses of cases handled by discipline inspection and supervision cadres), and legal learning (such as interpretations and queries of specific legal provisions). Simultaneously, the system performs preliminary format validation on the input content to ensure there are no garbled characters or invalid characters, providing a qualified foundation for subsequent processing. (II) Problem Extraction: Precise Filtering of Core Dimension Information; A pre-trained large language model is launched and loaded with special prompts for the discipline inspection and supervision scenario. These prompts include definitions of core business dimensions and key information identification rules, guiding the model to focus on extracting core information. Based on a key information extraction weight calculation mechanism, the model calculates weights and prioritizes five core dimensions of information in the input content: personnel identity (e.g., leadership rank, ordinary party member, discipline inspection and supervision staff), time nodes (e.g., case occurrence time, duration of behavior), behavior description (e.g., decision-making behavior, banquet behavior, fund usage behavior), problem situation (e.g., degree of loss, scope of impact), and industry field (e.g., government work, grassroots work, project construction). Key information with higher weight values ​​is extracted first. During extraction, the model automatically excludes redundant information unrelated to discipline inspection and supervision business (e.g., irrelevant personal information, irrelevant details in the scenario description), and stores the extracted key information in a structured manner, forming structured data of "key information - attributes - relationships," providing accurate data support for subsequent steps. (III) Problem Optimization: Professional Transformation and Logical Consolidation; Keyword Professional Conversion: The system utilizes a built-in dictionary of professional terms for discipline inspection and supervision, covering commonly used expressions in Party regulations, standardized terminology for discipline inspection and supervision business scenarios, and a table of correspondences between non-standard expressions. Through a keyword mapping and conversion mechanism, non-standard keywords and colloquial expressions in the input questions are accurately converted. For example, "pig-killing feast" is transformed into "banquet," and "image project" is directly associated with professional discipline inspection and supervision expressions such as "unrealistic and wasteful construction activities," ensuring semantic consistency between the question description and the knowledge base content. Problem logic rewriting: Combining the business logic of discipline inspection and supervision (such as the core logic of "characterization-quantification-accountability"), the transformed problems are restructured and logically supplemented. For problems with ambiguous expressions, necessary business scenario premises are added (e.g., rewriting "How to handle decision-making errors by leading cadres" as "How should we characterize and handle cases where leading cadres are irresponsible or negligent in their work, leading to decision-making errors and adverse effects"). For problems with multiple intertwined logics, logical decomposition and sorting are performed (e.g., breaking down "How should we handle cases where a cadre both violates regulations by hosting banquets and making decisions" into two sub-problems: "Characteristic handling of a cadre's violation of regulations by hosting banquets" and "Characteristic handling of a cadre's violation of regulations by making decisions," with the relationships marked respectively), making the problems more consistent with the business processing logic of discipline inspection and supervision, and reducing the difficulty of subsequent intent judgment and retrieval. (iv) Intent Judgment: Accurately Identify Core Needs; The optimized question text is transformed into a fixed-dimensional text vector using a semantic encoding model. This vector contains core information such as the semantic features and business dimension features of the question. The text vector is then input into an intent classification model pre-trained with BERT and fine-tuned using data from disciplinary inspection and supervision scenarios. The model classifies the core demands of the question, with the results covering seven categories: disciplinary action (qualitative assessment of the behavior and the level of punishment), boundary assessment (whether a behavior constitutes a violation of discipline or law), legal assessment (relevant Party regulations or legal provisions), standard assessment (specific standards for disciplinary action and applicable conditions), policy assessment (relevant disciplinary education policy requirements), case assessment (results in the handling of similar typical cases), and other categories (cannot be classified into the above categories). During the judgment process, the model calculates the classification confidence score for each intent category. If the highest confidence score reaches a set threshold, it is determined to be an intent of that category, and the corresponding processing logic is initiated; if the threshold is not reached, it is classified as "other," and a generalized retrieval mode is automatically initiated to expand the search scope to cover potential demands. Simultaneously, the system associates and stores the intent judgment results with structured key information to clarify the core direction of subsequent responses. (v) Problem Response: Differentiated Search and Precise Matching; Based on the intent assessment results, a differentiated retrieval strategy is initiated, combining the advantages of Retrieval Enhancement Generation (RAG) and knowledge graph-based retrieval to achieve accurate multi-dimensional content matching: Search strategy adaptation: The search weight is dynamically adjusted according to different intent types. For intents to search for laws and standards, the focus is on retrieving enhanced text matching results, prioritizing the accurate location of corresponding legal clauses and standard provisions. For intents to search for cases and boundaries, the focus is on entity association search results from the knowledge graph, retrieving similar cases and professional interpretations related to boundary definition through the association between entities (such as behavior type, identity type, and degree of loss). Multidimensional precise matching: Time-Regulations Matching: The system uses a precise matching mechanism for regulations versions to compare the extracted case occurrence time and duration of the behavior with the effective and expiration times of Party regulations in the local knowledge base. It also considers the correlation between the industry sector of the problem and the applicable scope of the regulations to select valid and suitable regulations within that time period, thus avoiding errors in the application of regulations. Identity-Content Matching: Through a mechanism that matches user identity with legal content, the system filters out applicable legal content and business processing logic corresponding to the extracted user identity information. For example, it provides content containing details of qualitative and quantitative disciplinary measures and key points of investigation and evidence collection for discipline inspection and supervision cadres, and provides easy-to-understand interpretations of laws and regulations and reminders of behavioral norms for ordinary party members. Deep content association: Retrieve Party regulations, laws and regulations, explanations of discipline inspection and supervision terminology, business logic descriptions, typical cases, etc. that match key information and intent types from the local knowledge base. At the same time, use knowledge graphs to mine the relationships between search results (such as the correspondence between regulations and typical cases, and the relationship between professional terms and the explanation of regulations), forming a multi-dimensional and highly relevant preliminary response content set. (vi) Result optimization: logical sorting and content purification; Content weighting and sorting: Based on the optimized sorting mechanism, the weights of each item in the initial response set are calculated. The weights consist of three parts: relevance of regulatory reminders, relevance of punitive measures, and relevance of legal basis. Regulatory reminders have the highest weight, ensuring the core requirement of disciplinary education is "warning first." The content is then sorted in descending order based on the weight calculation results, forming a core logical sequence of "regulatory reminders - legal basis - punitive measures," supplemented with auxiliary content such as business logic explanations and related case references. Content refinement and optimization: The sorted content undergoes data cleaning, removing duplicate information and irrelevant content, and correcting non-standard and illogical statements; legal citations are standardized, clearly specifying the law's name, clause number, and specific content; penalties are hierarchically organized, clearly presented according to severity level and applicable conditions; and regulatory reminders are simplified to ensure clear understanding for users of different identities. The optimized final response framework is logically clear, highlights key points, and is accurately expressed. (vii) Results Output and Continuous Service: Scenario-based Presentation and Closed-Loop Response; Scenario-based output format: Based on the scenario type identified at each question input stage, a multi-scenario adaptation mechanism is invoked to adjust the presentation format of the response content. For Q&A scenarios, a concise format of "standard reminder + core answer + legal basis" is used, facilitating quick access to key information. For competition scenarios, the legal provision number and core knowledge points are highlighted, along with supplementary answer analysis points. For case analysis scenarios, a detailed format of "qualitative conclusion + logical breakdown + supporting evidence + handling suggestions" is adopted to meet the professional needs of discipline inspection and supervision cadres. For legal learning scenarios, a structure of "clause content + professional interpretation + applicable scenarios + precautions" is used to assist users in deeply understanding the legal content. The optimized response content is then organized according to the corresponding format and returned through the user input platform, ensuring that the presentation format adapts to the scenario requirements. Continuous service response: The system supports users in asking follow-up questions based on the returned results. The follow-up questions are automatically linked with historical key information, intent judgment results, and response content, repeating the above process of "question extraction - question optimization - intent judgment - question response - result optimization - result output" to achieve a closed-loop service of "one question - multiple interactions - accurate answer", ensuring that users' deeper needs are fully met. At the same time, data such as user question content, key information, intent type, and response content are recorded in the system log, providing data support for subsequent knowledge base optimization and model iteration.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios, characterized in that: Based on large language models, local knowledge bases, retrieval-enhanced generation (RAG) and knowledge graph technologies, we construct a core thinking chain for discipline education and achieve precise services through multi-dimensional demand decomposition and a six-step professional process. The six-step process is as follows: Receive questions related to Party discipline learning and education; Extract key information from the problem, including the identity of the person involved, the time frame, the behavior, the situation of the problem, and the industry sector. Optimize problem information according to the characteristics of discipline inspection and supervision scenarios, including the transformation of non-standard keywords and the rewriting of problems; Analyze the intent of the problem and categorize it; Differentiated knowledge retrieval, logical processing, and response generation are performed based on intent; The response content was optimized and output according to the principle of "prioritizing standardized reminders".

2. The core thinking chain multi-scenario precise service method for Party discipline learning and education scenarios as described in claim 1, characterized in that, The formula for calculating the weight of key information extracted from the problem is as follows: ； in, For the first The extraction weight of key information categories (personnel identity / time point / behavior / problem situation / industry sector), This is an information importance coefficient (ranging from 0.6 to 0.8), used to reflect the priority of core dimensions in the discipline inspection and supervision scenario. For the first The importance score of information type in the disciplinary education scenario (0-10 points). The information correlation coefficient (with a value of 0.2-0.4, and...) ), For the first The score for the relevance of the information to the core demands of the problem (0-10 points, 10 points for complete relevance).

3. The multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios as described in claim 1, characterized in that, The keyword conversion for problem optimization uses the following formula: ； in, For input questions containing non-standard keywords and colloquial expressions (such as "pig-killing feast", "image project", "political achievement project", etc.), This is a dictionary of professional terminology for discipline inspection and supervision. It includes commonly used expressions in Party regulations, standardized terms used in discipline inspection and supervision business scenarios, and a table showing the correspondence between non-standard expressions and professional terms. This is a dictionary-based keyword mapping function that outputs standardized keywords by querying a pre-defined mapping relationship in the dictionary. (For example, transforming "pig-slaughtering feast" into "banquet" and associating "image project" with "illegal construction activities that are divorced from reality and waste manpower and resources").

4. The multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios as described in claim 1, characterized in that, Intent determination uses a classification formula with confidence thresholds: other ; in, This is the optimized question text vector, generated through a semantic encoding model, which includes both semantic features and business dimension features of the question. This is an intent classification model based on BERT pre-training and fine-tuned using inspection and monitoring scenario data, outputting intent categories. (Including checking discipline, boundaries, regulations, standards, policies, and cases) For classification confidence, The confidence threshold is set (ranging from 0.75 to 0.85). If the confidence level reaches the threshold, the corresponding intent will be applied. If the confidence level does not reach the threshold, the category will be classified as "other" and a generalized search will be initiated.

5. The multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios according to claim 1, characterized in that, The formula for matching case dates with regulatory versions is as follows: ； in, To best match the regulatory version, This is a collection of valid versions of Party regulations (including the effective and expiration dates and scope of application for each version). This refers to the time of the incident or the duration of the behavior extracted from the question. For version Effective date For version Expiration time (if not expired, take the current time) ), For indicator functions (satisfying) exist (Take 1 if the interval is specified, otherwise take 0). For version Relevance to the industry sector in question (value range: 0-1).

6. The multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios according to claim 1, characterized in that, The formula for matching user identity with legal content is as follows: ； in, A set of regulations to be adapted. This is a collection of all legal content in the local knowledge base (including clause texts, applicable objects, and quantitative standards). To extract user identities from the questions (such as county-level leading cadres, ordinary party members, and discipline inspection and supervision staff). For the content of regulations Clearly defined scope of applicable identities For determining identity affiliation, when belong When a subset of the law is included, the content of the law is... It was included in the adaptation set.

7. The multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios according to claim 1, characterized in that, The formula for the differentiated retrieval strategy for question response is: ； in, For the final search results, To retrieve weighting coefficients, when searching for regulations or standards. (Focusing on precise text matching), when searching for cases or boundary-based intents. (Focusing on related searches) To enhance the generated text matching results for retrieval (based on keyword and semantic similarity matching of legal clauses and standard provisions). This refers to the entity association retrieval results of the knowledge graph (including the association relationships of legal clauses, cases, and terms).

8. The multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios according to claim 1, characterized in that, The optimized content sorting formula is as follows: ； in, To optimize the output sequence, In response to the generation of a content list, (Standard reminder weight) (Weight of punitive measures) (Based on the weight of legal provisions) To standardize the relevance score of reminders, To score the relevance of the punitive measures, The score is based on the relevance of the legal provisions (0-10 points). This is a descending sorting function.

9. The multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios according to claim 1, characterized in that, The hallucination rate control uses a cross-validation formula: ； in, For the final illusion rate, For local knowledge base With knowledge graph The number of errors in the dual-validation judgment ( , To validate the number of errors in the knowledge base, (Number of validation errors for knowledge graph) This represents the total number of output contents.

10. The multi-scenario precise service method for the core thinking chain in Party discipline learning and education scenarios according to claim 1, characterized in that, The formula for multi-scenario adaptation is: ； in, For scenario-based service processes, Based on the six-step process vector, To adapt operators to the process, Scenario Adaptation Matrix (Question and Answer Consultation Scenario) Competition question-and-answer scenario Case analysis scenario The matrix elements represent the resource allocation weights for the corresponding process steps.

Images