A "value alignment" dialogue guiding method for ideological and political education and a credible evaluation system thereof

Abstract

This invention relates to the fields of intelligent teaching and trustworthy artificial intelligence, providing a "value alignment" dialogue guidance method and its trustworthy evaluation system for ideological and political education. The system collects multiple rounds of teacher-student dialogue and encapsulates them into data packets containing round identifiers, unified timestamps, and conversation context indexes; it structures course themes, teaching objectives, and normative constraints into a computable constraint set; it identifies dialogue states and generates control contexts, driving the selection of guidance actions and the generation of candidate responses; it performs value alignment and compliance checks on candidate responses, rewriting, downgrading, redirecting, or rejecting responses based on the results, and retains a handling log; it evaluates the alignment degree, guidance effectiveness, risk compliance, stability, and interpretability of the entire process, outputting scores, cause identification, and evidence chains, and writing them into audit storage, achieving controllable dialogue guidance, quantifiable evaluation, and traceable process.

Description

Technical Field

[0001] This invention relates to the field of ideological and political education technology, and in particular to a "value alignment" dialogue guidance method and its credible evaluation system for ideological and political education. Background Technology

[0002] The teaching organization of ideological and political education is gradually extending from classroom lectures and offline tutoring to a blended learning model supported by learning platforms, mobile terminals, and classroom interactive terminals. To enhance learner engagement and personalized guidance, existing teaching platforms generally incorporate intelligent question-and-answer, conversational tutoring, and learning analytics components: one approach uses a rule-based and knowledge-base retrieval-based question-and-answer system, providing fixed questions and answers around course chapters, knowledge points, and question banks; another approach uses natural language processing and human-computer dialogue generation models to generate explanations, examples, and guiding questions after learners input their questions; some solutions also incorporate security measures such as content filtering, sensitive word blocking, blacklists and whitelists, and manual spot checks to reduce the risk of inappropriate content output. Some products further provide learning profiles, answer performance statistics, and satisfaction evaluations to assist teachers in process management.

[0003] However, dialogic teaching for ideological and political education has clear value orientation and normative constraints. Existing general dialogue systems or general intelligent tutoring systems still have shortcomings in practical application: First, value goals are mostly in the form of teaching objective texts or course outlines, making it difficult to transform them into calculable and verifiable constraints. The dialogue generation process lacks executable goal constraints, causing the guidance direction to easily drift with the context, resulting in deviations, off-topic situations, or inconsistent stances. Second, existing security and compliance controls mostly remain at the level of word list filtering or single-round judgment, failing to combine multi-round contexts, cognitive stages, and risk levels for graded handling. When encountering controversial topics, they often can only simply block or give generalized responses, affecting the continuity of teaching and making it difficult to achieve "corrective guidance." Third... Existing evaluation mechanisms primarily rely on outcome satisfaction, single-dialogue scores, or human experience, lacking process-level quantitative assessments that cover alignment, guidance effectiveness, risk compliance, stability, and interpretability. This makes it difficult to pinpoint which round a problem occurred in, which criterion was triggered, and what actions were taken. Fourth, most systems lack audit evidence retention and evidence chain management capabilities. Dialogue segments, triggering rules, actions, and log indexes are not traceable, leading to reliance on manual review for teacher verification, responsibility allocation, and quality governance, which is costly and inconsistent. Fifth, existing solutions typically disperse dialogue guidance, content verification, and evaluation auditing across different stages or systems, lacking online closed-loop linkage and making it difficult to achieve synchronous constraints, real-time correction, and verifiable traceability before and after generation.

[0004] Therefore, there is an urgent need for a dialogue guidance technology for ideological and political education scenarios, which can structure value goals into computable constraints and integrate them with multi-round dialogue status recognition, guidance action selection, alignment and compliance verification, process-level credible assessment and audit evidence retention, so as to achieve controllable guidance, quantifiable assessment and traceable governance. Summary of the Invention

[0005] One objective of this invention is to propose a "value alignment" dialogue guidance method and its credible evaluation system for ideological and political education. This invention utilizes natural language processing and dialogue guidance technology to structure ideological and political value goals into computable constraints and realize an online verification, evaluation, and audit closed loop, which has the advantages of controllable guidance, quantifiable evaluation, and traceable process.

[0006] A "value alignment" dialogue guidance method and its credibility assessment system for ideological and political education according to an embodiment of the present invention include the following steps:

[0007] Step S1: Obtain dialogue input and teaching constraint information to form a dialogue data package;

[0008] Step S2: Structure the teaching constraint information into a verifiable constraint set, and configure trigger conditions and action identifiers for each constraint item in the constraint set;

[0009] Step S3: Perform dialogue state recognition and modeling on the dialogue data packet to obtain the dialogue state object, and associate the dialogue state object with the constraint set to generate a control context;

[0010] Step S4: Determine the target guidance action and its guidance parameters based on the control context, and call the dialogue generation model to generate candidate guidance outputs;

[0011] Step S5: Perform alignment and compliance checks on the candidate guidance outputs, determine the check results based on the constraint set, perform alignment control processing with the check results, output the guidance outputs after alignment control processing, and record the triggered rule identifiers and handling action identifiers;

[0012] Step S6: Perform a credibility assessment and audit evidence retention process on the dialogue process containing the guidance output, generate assessment results and evidence chain information according to the preset assessment index system, and write the assessment results and evidence chain information into the audit storage.

[0013] Optionally, the dialogue input and instructional constraint information are acquired and a dialogue data package is formed, including:

[0014] Collect dialogue content and establish a session identifier for the session;

[0015] Perform text normalization on the dialogue content to obtain normalized text, and establish a corresponding index between the original text and the normalized text;

[0016] Generate a unified time stamp for the dialogue content, and generate fragment time stamps corresponding to the normalized text;

[0017] Receive course topic information and generate topic identifiers, then write the topic identifiers into the session metadata corresponding to the session identifier;

[0018] Receive teaching objective information and generate objective identifiers, and establish the association between objective identifiers and theme identifiers;

[0019] Receive specification constraint information and generate constraint identifiers, and establish the association relationship between constraint identifiers and subject identifiers and target identifiers;

[0020] Based on the above processing results, construct the dialogue data packet and write the dialogue data packet to the session storage.

[0021] 3. The dialogue guidance method according to claim 1, characterized in that, the teaching constraint information is structured into a verifiable constraint set and trigger conditions and action identifiers are configured, including:

[0022] Read the teaching constraint information from the dialogue data packet, generate structured constraint configuration data, and set version identifier and scope identifier for the structured constraint configuration data;

[0023] Perform element parsing on the teaching constraint information to obtain an element set, and set identifiers and parameters for the elements in the element set;

[0024] A set of rules is generated based on a set of elements. The rules in the set of rules correspond to verifiable trigger conditions and action identifiers, respectively.

[0025] Merge the rule set to generate a constraint set, and configure verifiable trigger conditions, action identifiers, priority parameters and conflict resolution parameters for the constraint items in the constraint set.

[0026] Optionally, perform dialogue state recognition and modeling on the dialogue data packets and generate control context, including,

[0027] Read the dialogue data packet from step S1 and generate state analysis input;

[0028] Extract the necessary information for state analysis from the dialogue data packet to form the state analysis input;

[0029] Topic identification is performed on the state analysis input, and the identification results are matched with the set of topic elements to obtain topic deviation indicators;

[0030] Map state analysis inputs to position results and label input segments that support position categories;

[0031] The cognitive stage category is obtained through stage reasoning, and the stage is based on the index to record the position of the input fragment that supports the cognitive stage category;

[0032] In the risk identification process, the triggering fragment is first located and a hit index is generated, and then the risk level category is determined by the hit index;

[0033] The coverage strength of the value dimensions is calculated to form the value dimension distribution, and the evidence index is written synchronously with the coverage calculation;

[0034] Input the position results, cognitive stage categories, risk level categories and value dimension distribution into the consistency verification process, and output alignment deviation indicators. The alignment deviation indicators represent the deviation type, deviation intensity and trigger index.

[0035] The topic tags, stance results, cognitive stage categories, risk level categories, value dimension distribution and alignment deviation indicators are aggregated to form a dialogue state object, which is associated with the conversation identifier and interaction index.

[0036] A control context is generated from the dialogue state object and the constraint set, and the control context is written to the state store.

[0037] Optionally, the target guidance action and its guidance parameters are determined based on the control context, and candidate guidance outputs are generated, including:

[0038] Read the control context from step S4 and generate the guidance decision input;

[0039] Determine the guidance objective based on the guidance decision input, and set the objective parameters for the guidance objective;

[0040] Based on the dialogue stage information and risk level information in the dialogue state object, action gating is performed to obtain the set of allowed actions and the set of prohibited actions, and action parameters are configured for the actions.

[0041] The target guidance action is determined based on the guidance objective and the set of permitted actions, and guidance parameters corresponding to the target guidance action are generated.

[0042] Perform pruning on the historical dialogues pointed to by the session context index to obtain a set of context fragments, and configure fragment parameters for the context fragments;

[0043] The target guidance action, guidance parameters, context fragment set, and format constraints are assembled into the generated request data;

[0044] The dialogue generation model is invoked to perform generation processing on the generation request data, output candidate guidance outputs, and write the candidate guidance outputs to the output.

[0045] Optionally, alignment and compliance checks are performed on the candidate boot outputs, and the boot outputs are output after alignment control processing, including:

[0046] The dialogue guidance method according to claim 1 is characterized in that, performing alignment and compliance checks on candidate guidance outputs and outputting guidance outputs after alignment control processing includes:

[0047] Retrieve candidate guidance outputs and load constraint sets to form validation inputs. Make judgments on the trigger conditions of the disabling rules and record the disabling hit information when triggered.

[0048] Perform risk identification, input the identification results into risk rules to complete the judgment, output risk hit information, perform alignment quantification processing on candidate guidance outputs, and produce alignment results;

[0049] Establish a correspondence between candidate guidance outputs and key rules, determine the coverage gaps based on the correspondence, and form key verification information;

[0050] Extract the format features of the candidate guide output, compare them with the format rules, and return the comparison results as format verification information;

[0051] The system integrates prohibited hit information, risk hit information, alignment results, key point verification information, and format verification information to form a verification result object.

[0052] The decision to handle the situation is derived from the verification result object, and the action to handle the situation is determined based on the priority parameter and the conflict resolution parameter.

[0053] The execution of the action generates a guidance output and a action log. The guidance output and action log are then written to the audit storage for subsequent steps.

[0054] Optionally, the process of performing a credibility assessment and audit documentation on the dialogue process is described below:

[0055] Read data from the session storage, align and aggregate it according to the session identifier and interaction index, and generate evaluation records;

[0056] Perform structured extraction on the evaluation records to generate evaluation samples associated with the interaction index;

[0057] The evaluation sample is evaluated by calling a preset scoring function to generate a scoring result. The scoring function performs a weighted calculation on the indicator parameters and outputs a comprehensive conclusion identifier.

[0058] Based on the evaluation sample and the scoring result, generate cause location information and evidence chain information, and write the scoring result, cause location information and evidence chain information into the audit storage;

[0059] A retrieval index associated with the session identifier and the interaction index is generated in the audit storage for subsequent dialogue guidance calls and tracing.

[0060] Optionally, a credible assessment system for guiding "value alignment" dialogues in ideological and political education includes:

[0061] The data acquisition and encapsulation module acquires dialogue input and teaching constraint information, and generates dialogue data packages.

[0062] The constraint construction module structures teaching constraint information to generate a constraint set, and configures trigger conditions and action identifiers for the constraint items in the constraint set;

[0063] The state modeling and context generation module performs dialogue state recognition and modeling on the dialogue data packet to obtain a dialogue state object, and associates the dialogue state object with the constraint set to generate a control context.

[0064] The guidance generation module determines the target guidance action and guidance parameters based on the control context, and calls the dialogue generation model to output candidate guidance outputs.

[0065] The verification and handling module performs alignment and compliance verification on the candidate guidance output, generates verification results and determines the handling action, executes the handling action to output the guidance output, and records the handling log.

[0066] The assessment and audit module generates assessment results and evidence chain information for the dialogue process including guided outputs, and writes the assessment results and evidence chain information into the audit storage.

[0067] The beneficial effects of this invention are:

[0068] (1) By collecting multi-round dialogues through learner terminals, teacher terminals and classroom interaction terminals and encapsulating them into dialogue data packets with round identifiers, unified timestamps, role identifiers and conversation context indexes, and with the integrated access of course themes, teaching objectives and normative constraints, the unified organization and traceable management of ideological and political dialogue teaching inputs, scenarios and constraints are realized.

[0069] (2) By converting the value goals of ideological and political education into a set of computable constraints, and configuring verifiable trigger conditions, action identifiers, priority parameters and conflict resolution parameters for each constraint, the value goals are realized from textual expression to executable constraints in a structured manner, providing a deterministic basis for subsequent dialogue guidance and verification control.

[0070] (3) By identifying and modeling the distribution of topics, viewpoints, cognitive stages, risk levels and value dimensions in multi-round dialogues, the current dialogue state object is generated and bound to the set of computable constraints to form a control context. This realizes the online closed-loop connection of guiding decisions with dialogue state and value constraints as input, reducing the risk of deviation and target drift in multi-round dialogues. Attached Figure Description

[0071] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0072] Figure 1 This is a flowchart of a "value alignment" dialogue guidance method for ideological and political education proposed in this invention; Figure 2 This is a schematic diagram of the dialogue acquisition and data encapsulation structure in this invention, showing the process of learner terminal, teacher terminal and classroom interaction terminal acquiring multiple rounds of dialogue and forming dialogue data packets;

[0073] Figure 3 This is a schematic diagram of the selection of guiding actions and the generation of candidate guiding statements in this invention, demonstrating the process of selecting questions for clarification, presenting key value points, correcting viewpoints, comparing cases, summarizing and generalizing, and shifting to a refusal to answer, and generating candidate guiding statements under the control context. Detailed Implementation

[0074] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0075] refer to Figures 1 to 3 A dialogue-guided method for "value alignment" in ideological and political education and its credible evaluation system include the following steps:

[0076] In this embodiment, obtaining dialogue input and teaching constraint information to form dialogue data includes the following steps: The dialogue input between teachers and students is collected through learner terminals, teacher terminals, and classroom interaction terminals. Learner terminals can be any of mobile communication terminals, tablet computing terminals, or personal computing terminals. Teacher terminals can be personal computing terminals or tablet computing terminals. Classroom interaction terminals are interactive display large screen terminals. The speaker role identifier and turn number are obtained simultaneously during the collection. When the dialogue input is voice, audio data is collected by a voice acquisition device and audio segment identifiers are generated. The voice acquisition device is a built-in microphone in the terminal or a classroom microphone array. The audio data is converted into dialogue text by a speech transcription unit and an index is established to correspond to the audio segment identifiers and the dialogue text. When the dialogue input is text, the dialogue text is collected through the terminal text input device, which can be a keyboard input component or a touch input component, and text fragment identifiers are generated for the dialogue text. A unified timestamp is generated for each round of dialogue collected above using a unified time synchronization system clock, and a sentence timestamp sequence is generated for the sentences within the same round of dialogue. The course topic information is obtained through the course management interface on the teacher's terminal or the course resource library on the teaching platform server, and a topic identifier is generated. The course topic information includes the course name, chapter topic, knowledge point list and teaching progress identifier. The teaching objective information is obtained through the teaching objective configuration interface on the teacher's terminal or the teaching plan library of the teaching platform server, and objective identifiers are generated. The teaching objective information includes value objective items, ability objective items, knowledge objective items and assessment requirement items, and the association between objective identifiers and theme identifiers is established. The teacher terminal obtains normative constraint information and generates constraint identifiers through the normative constraint configuration interface or the compliance rule base of the teaching platform server. The normative constraint information includes teaching scenario type, applicable population scope, content compliance boundary, list of prohibited expressions and risk level judgment, and establishes the association relationship between constraint identifiers and topic identifiers and target identifiers. The obtained data is encapsulated into a conversation data packet. The conversation data packet contains a conversation number, round number, speaker role identifier, conversation text or audio segment identifier, unified timestamp, sentence timestamp sequence, conversation context index, topic identifier, target identifier and constraint identifier, and the conversation data packet is written to the conversation storage.

[0077] In this embodiment, the teaching constraint information is structured into a verifiable constraint set, and trigger conditions and action identifiers are configured for each constraint item in the constraint set, including: Read the course topic information, teaching objective information and normative constraint information in the dialogue data packet formed in step S1 of claim 2, generate structured constraint configuration data, and set version identifier and scope identifier for the structured constraint configuration data; Perform topic element analysis on the course topic information to obtain chapter topics, knowledge point lists and teaching progress indicators, form a topic element set, and assign topic identifiers and weight parameters to each topic element in the topic element set; The teaching objective information is analyzed to obtain value objective items, ability objective items, knowledge objective items and assessment requirement items, forming a set of objective elements. The value objective items are then converted into a set of value dimension labels. The set of value dimension labels includes at least national identity, rule of law awareness, social responsibility, collectivism, integrity and friendliness, professional ethics and scientific spirit. Target intensity parameters and achievement threshold parameters are set for each value dimension label. Perform position mapping processing on the target element set and the topic element set to generate target position labels. The target position labels include at least support labels, opposition labels and neutral labels. Set an applicable condition set for each target position label. The applicable condition set includes dialogue topic matching conditions, learner viewpoint and position conditions and risk level conditions. The prohibited items are parsed for the normative and restrictive information to form a set of prohibited content rules. The set of prohibited content rules includes at least the prohibited expression list rules, scenario incompatibility rules, identity incompatibility rules, information authenticity risk rules, and deviation from value dimension rules. Each prohibited content rule is set with a rule identifier, triggering conditions, and handling action identifier. Triggering conditions include keyword triggering, semantic matching triggering, stance conflict triggering, and risk level triggering. The teaching objective information is extracted and processed to form a set of rules for essential points. The set of rules for essential points includes at least the rules for core concept coverage, the rules for value point coverage, the rules for the integrity of the argument chain, and the rules for summarization. Each rule for essential points is set with a rule identifier, triggering conditions, and action identifier. Triggering conditions include round triggering, knowledge point missing triggering, insufficient coverage of value dimensions triggering, and cognitive stage triggering. Standardize the analysis of risk judgment criteria for information enforcement, and form a set of risk judgment rules. The set of risk judgment rules should include at least the following: erroneous fact judgment criteria, biased stance judgment criteria, discrimination and hatred judgment criteria, illegal and irregular guidance judgment criteria, privacy leakage judgment criteria, and incitement to confrontation judgment criteria. For each risk judgment rule, a judgment identifier, triggering condition, and action identifier are set. The triggering conditions include semantic classification triggering, similar case triggering, and sensitive entity triggering. The course topic information and teaching scenario type are processed to generate format constraints, forming a set of output format rules. The set of output format rules includes at least language style rules, address and politeness rules, citation basis rules, step-by-step guidance structure rules and length limit rules. Each output format rule is set with a rule identifier, trigger condition and action identifier. The acquired sets are merged into a computable constraint set, and each constraint in the computable constraint set is assigned a verifiable trigger condition, action identifier, priority parameter, and conflict resolution parameter. The action identifier includes at least rewrite action, downgrade action, redirection action, and rejection action.

[0078] In this embodiment, performing dialogue state recognition and modeling on the dialogue data packet to obtain a dialogue state object, and associating the dialogue state object with a constraint set to generate a control context includes: Read the dialogue data packet generated in step S1, extract the session number, round number, speaker role, standardized dialogue text, session context index, course topic information, teaching objective information and normative constraint information, and generate the input for this round's state analysis; Perform topic identification processing on the input of this round of state analysis to obtain the topic tags and topic strength parameters of this round, and match the topic tags of this round with the topic element set generated in step S2 to output the topic matching results and the off-topic identifier; Perform opinion and stance identification processing on the normalized dialogue text to obtain opinion and stance categories. The opinion and stance categories include at least support, opposition and neutrality. Output the stance basis fragment index and stance confidence parameter. The dialogue text and conversation context index are used to perform cognitive stage identification processing to obtain cognitive stage categories. The cognitive stage categories include at least the concept understanding stage, opinion formation stage, value judgment stage, and action intention stage, and output the stage-based fragment index and stage confidence parameters. The dialogue text performs risk element identification processing, extracts sensitive entities, sensitive expressions, controversial topics and instructional guidance expressions, calls the risk judgment rule set generated in step S2 to perform judgment matching on risk elements, and outputs risk hit details and risk level categories. The risk level categories include at least low risk, medium risk and high risk. The dialogue text is processed to determine the value dimension. The value dimension label set generated in step S2 is used to calculate the value dimension distribution. The value dimension distribution includes the coverage strength parameter and coverage missing identifier corresponding to each value dimension label, and outputs the evidence fragment index corresponding to each value dimension label. Perform consistency verification on the distribution of opinion and stance categories, cognitive stage categories, risk level categories and value dimensions, and generate alignment deviation markers. The alignment deviation markers include deviation type markers, deviation intensity parameters and trigger fragment indexes. Construct the current dialogue state object, which includes the conversation number, round number, topic tag for this round, opinion / position category, cognitive stage category, risk level category, value dimension distribution, alignment deviation indicator, basis fragment index set, and confidence parameter set; The constructed current dialogue state object is bound to the computable constraint set generated in step S2 to generate the current round control context. The current round control context includes at least the session number, round number, current dialogue state object, rule identifier set of computable constraint set, rule priority set and conflict resolution parameter set, and the current round control context is written to the session state storage.

[0079] In this embodiment, determining the target guidance action and its guidance parameters based on the control context, and then calling the dialogue generation model to generate candidate guidance outputs includes: Read the current round control context generated in step S3, extract the current dialogue state object, the rule identifier set of the computable constraint set, the rule priority set and the conflict resolution parameter set, and generate the guidance decision input; The guidance decision input is processed to determine the guidance objectives, resulting in the current guidance objective set. The current guidance objective set includes at least the objective value dimension label, objective position label, essential points to be achieved, issues to be clarified, and risks to be avoided. Target strength parameters and achievement threshold parameters are set for each objective in the current guidance objective set. Action gating is performed on the cognitive stage category and risk level category of the current dialogue state object to obtain the set of allowed actions and the set of prohibited actions. The set of allowed actions and the set of prohibited actions cover question clarification, presentation of key points, correction of viewpoints, case comparison, induction and summary and refusal to answer. Action priority parameters and triggering conditions are set for each action. Based on the current set of guiding objectives and the set of allowed actions, action selection processing is performed to obtain the target guiding action and generate the guiding parameters corresponding to the target guiding action. The guiding parameters include at least a list of target value dimension labels, target position labels, a list of essential points to be achieved, dialogue style constraints, output structure constraints, length limit parameters, and a list of prohibited expressions. Perform context trimming on the historical rounds of dialogue pointed to by the conversation context index in step S1 to obtain a set of context fragments. The set of context fragments includes fragments related to the topic tags of the current round, fragments related to the basis of viewpoints and positions, fragments related to the evidence of value dimensions, and fragments related to the risk hit details. Set fragment position index and fragment importance parameters for each fragment. The generated target guidance action, guidance parameters, context fragment set, and output format rule set are assembled into generation constraint input, and the generation constraint input is written into the generation request data of this round. The dialogue generation model is invoked to perform dialogue generation processing on the data of the current generation request, and candidate guiding dialogues are output. The candidate guiding dialogues include guiding content text, guiding action identifiers, essential point coverage markers, reference basis placeholders, and structured output markers.

[0080] In this embodiment, alignment and compliance checks are performed on the candidate guidance outputs. The check results are determined based on the constraint set, and alignment control processing is performed with the check results. The guidance output after alignment control processing is output, and the triggered rule identifier and handling action identifier are recorded, including: Read the candidate guidance scripts output in step S4 and read the computable constraint set output in step S2 to form the verification input for this round. The verification input for this round includes candidate guidance script text, guidance action identifier, mandatory point coverage marker, output structure marker, conversation number and round number. Perform prohibited content rule verification on the candidate guiding text, match the trigger conditions one by one according to the prohibited content rule set generated in step S2, and output the prohibited hit list. The prohibited hit list includes at least the hit rule identifier, the hit segment index, the hit type identifier, and the hit severity parameter. Perform risk criterion rule verification on the candidate guidance text, match the trigger conditions one by one according to the risk criterion rule set generated in step S2, and output the risk hit list. The risk hit list includes at least the hit criterion identifier, the hit segment index, the hit risk category and the hit severity parameter. Perform value dimension alignment verification on the candidate guiding text, and calculate the alignment result according to the value dimension label set generated in step S2 and the target position label. The alignment result includes at least the value dimension coverage parameter, position consistency parameter, deviation type identifier and deviation segment index. Perform mandatory key point rule verification on the candidate guiding scripts, match the trigger conditions one by one according to the mandatory key point rule set generated in step S2, and output the key point missing list. The key point missing list shall at least include the missing key point identifier, the missing position index and the missing severity parameter. Perform output format rule validation on the candidate guidance scripts, match the trigger conditions one by one according to the output format rule set generated in step S2, and output a format violation list. The format violation list includes at least the violation rule identifier, the violation location index and the violation type identifier. Summarize the hit list, risk hit list, step alignment results, key point missing list and format violation list, and generate a verification result object. The verification result object shall include at least the compliance conclusion identifier, alignment conclusion identifier, risk level identifier and disposal priority parameter. Based on the verification result object, the alignment control process is triggered and the current round of disposal decision is generated. The alignment control process selects at least one disposal action from rewriting, downgrading, redirection and rejection. The disposal action selection follows the rule priority parameters and conflict resolution parameters configured in step S2, and outputs a set of disposal action identifiers and a set of disposal reason identifiers. The alignment control process is performed to obtain the output of this round of guidance. The rewriting corresponds to rewriting the content of the candidate guidance text while retaining the essential points coverage mark. The downgrading corresponds to adjusting the candidate guidance text to a low-risk expression and reducing the scope of controversial content. The turning corresponds to generating alternative guidance texts that are consistent with the course theme and teaching objectives. The refusal corresponds to generating a refusal text and generating subsequent guidance suggestion items. Record the set of action identifiers for the output, the triggered rule identifiers and criterion identifiers, the hit fragment index, the pre-processing text summary index, the post-processing text summary index, the session number, the round number, and the unified timestamp to form the current round's verification and processing log. Write the current round's guidance output and the current round's verification and processing log into the audit storage.

[0081] In this embodiment, a credibility assessment and audit evidence retention process is performed on the dialogue process including the guided output. Assessment results and evidence chain information are generated according to a preset assessment indicator system, and the assessment results and evidence chain information are written into the audit storage, including: Read the dialogue data packet generated in step S1, the candidate guidance script generated in step S4, the current round guidance output and the current round verification and processing log output in step S5, and align and aggregate them according to the session number and round sequence number to form a round-level evaluation record set. The round-level evaluation record set is structured and extracted to generate a process-level evaluation sample set. Each evaluation sample in the process-level evaluation sample set includes learner input fragment index, guidance action identifier, guidance parameter summary index, guidance output fragment index, trigger rule identifier, criterion identifier, disposal action identifier, risk level identifier, and uniform timestamp. For each evaluation sample, alignment calculation is performed. The value dimension label set, target stance label, and essential point rule set generated in step S2 are read. The value dimension coverage strength parameter, stance consistency parameter, and essential point coverage parameter are calculated for the guided output segment, and an alignment score is generated. The alignment score is calculated using the following formula, and the formula is recorded in LaTeX code:

[0082]

[0083] in: For the value dimension weight parameters, For the first Each value dimension covers the intensity parameter. For the position consistency parameter, For key point coverage parameters, For the quantity of value dimensions, and For each evaluation sample, a guidance effectiveness calculation is performed. The cognitive stage category and opinion / stance category obtained in step S3 are read, and the learner input segment indexes from adjacent rounds are read. The cognitive stage transfer parameter, clarification completion parameter, and opinion convergence parameter are calculated to generate a guidance effectiveness score. The guidance effectiveness score is calculated using the following formula, and the formula is recorded in LaTeX code:

[0084]

[0085] in For cognitive stage transfer parameters, To clarify the completion parameters, For the convergence parameters of the viewpoint, , and For each assessment sample, risk compliance calculation is performed. The prohibited hit list, risk hit list and format violation list generated in step S5 are read. The violation count parameter, maximum severity parameter and adequacy parameter are calculated. A risk compliance score is generated, and a risk compliance conclusion identifier and trigger item detail index are generated. For consecutive evaluation samples under the same session number, perform stability calculation processing, read the guidance action identifier, disposal action identifier and alignment score sequence of each round, calculate and output fluctuation parameters, action jump parameters and score fluctuation parameters, generate stability score, and generate abnormal fluctuation location information and corresponding round sequence number set; For each evaluation sample, interpretability generation processing is performed to construct cause location information. The cause location information includes at least the correspondence between hit rule identifiers and criterion identifiers, hit fragment index, action identifier, alignment deviation type identifier, and evidence fragment index, and forms evidence chain information. The evidence chain information includes at least the dialogue fragment index, trigger rule identifier, criterion identifier, log index, and timestamp. The obtained scores are summarized to generate an evaluation result object. The evaluation result object includes at least the alignment score, guidance effectiveness score, risk compliance score, stability score, set of interpretability items, comprehensive score and level label, and records the calculation parameter index and threshold label of each score. The evaluation results and evidence chain information are written into the audit storage. The audit storage record includes at least the session number, round number, unified timestamp, scoring result index, cause location information index, evidence chain information index, and review status identifier. An audit retrieval index and a traceability index are generated for subsequent rounds of dialogue to guide the use of the audit retrieval index and traceability index.

[0086] In this embodiment, a credible assessment system for guiding dialogue on "value alignment" in ideological and political education is characterized in that the system includes: The dialogue acquisition and data encapsulation module is used to acquire learners' multi-turn dialogue input and receive course topic information, teaching objective information and normative constraint information, forming a dialogue data package containing turn identifiers, unified timestamps, role identifiers and conversation context indexes; The value goal structuring and constraint generation module is configured to structure the value goals of ideological and political education into a computable constraint set based on the course theme information, teaching goal information and normative constraint information output by the dialogue collection and data encapsulation module, and to configure verifiable trigger conditions and action identifiers for each constraint in the computable constraint set. The dialogue state recognition and control context construction module is configured to perform dialogue state recognition and modeling on the dialogue data packets output by the dialogue acquisition and data encapsulation module, to obtain the current dialogue state object containing viewpoint category, cognitive stage category, risk level category, value dimension distribution and alignment deviation identifier, and is configured to bind the current dialogue state object with the computable constraint set output by the value target structuring and constraint generation module to form the control context of this round. The guidance action selection and candidate dialogue generation module is configured to select the target guidance action from the preset guidance action set and generate guidance parameters based on the current round control context output by the dialogue state recognition and control context construction module. It is also configured to call the dialogue generation model to generate candidate guidance dialogues based on the target guidance action, guidance parameters and the conversation context provided by the dialogue acquisition and data encapsulation module. The alignment verification and handling control module is configured to perform value alignment verification and compliance verification on the candidate guidance scripts generated by the guidance action selection and candidate script generation module. It is also configured to determine the verification result according to the computable constraint set output by the value target structuring and constraint generation module and trigger the alignment control processing corresponding to the verification result. The alignment control processing includes rewriting, downgrading, redirection and rejection. It outputs the current round of guidance output after alignment control processing and records the triggered rule identifier and handling action identifier. The Trustworthy Assessment and Audit Evidence Retention Module is configured to perform trustworthy assessment and audit evidence retention processing on the dialogue process including the output of the alignment verification and handling control module in this round of guidance. It generates assessment results based on the preset assessment index system, and outputs assessment scores, cause location information and evidence chain information. The evidence chain information includes dialogue segment index, trigger rule identifier, criterion identifier, log index and timestamp. The assessment results and evidence chain information are written to the audit storage for subsequent rounds of dialogue guidance.

[0087] Example 1:

[0088] This embodiment selects "classroom discussion + after-class intelligent tutoring" in ideological and political courses in colleges and universities as the implementation scenario, and focuses on solving three common problems of general dialogue systems in ideological and political teaching: discussions are prone to going off-topic and deviations, and value orientation is difficult to be explicitly constrained; when encountering controversial viewpoints, there is a lack of graded handling, which can easily lead to non-compliance risks or simple blocking that can cause teaching interruptions; and teachers' after-class review lacks evidence chains and triggering basis, making it difficult to locate the problem rounds and handling processes, resulting in high governance costs and low review efficiency.

[0089] The implementation location was set in the smart classroom and on-campus teaching platform testing environment of a university in Pudong New Area, Shanghai (hereinafter referred to as "University A" for privacy protection). Offline classes were held in "Smart Classroom 3-308" on the third floor of the main teaching building of University A. The classroom was equipped with an interactive display screen, a ceiling-mounted microphone array, wireless network, and a classroom interaction system. Online and after-class tutoring were conducted on the on-campus teaching platform, "Ideological and Political Learning Space." The system was deployed in the computer room of University A's Information Center (2nd floor of the Information Building, north of the main teaching building), on the same intranet segment as the teaching platform server. Teachers and students accessed the system through the campus network. The trial operation period was from November 3, 2025 to December 12, 2025, covering a complete 6-week teaching cycle. Centralized teacher review and data collection were completed from December 15, 2025 to December 21, 2025.

[0090] Table 1 Summary of Trial Operation Locations, Times, Scale of Participation, Deployment Methods, and Data Collection Equipment

[0091] project content Trial operation location A university in Pudong New Area, Shanghai; Offline classrooms: Smart Classroom 3-308, 3rd Floor, Main Teaching Building; Servers: Computer Room, Information Center, 2nd Floor, Information Building Trial run time November 3, 2025 - December 12, 2025 (6 weeks); Centralized review for teachers: December 18, 2025 - December 20, 2025 (14:00-17:30) Courses and Classes The ideological and political education module of the general basic courses consists of 2 courses; 6 classes; and 4 instructors. Scale of participation 236 students; 1,247 total dialogue sessions; 14,986 total dialogue rounds. Deployment method The system is deployed on the same intranet segment as the school's teaching platform; terminals access it through the campus network; and classrooms use the smart classroom wireless network. Data acquisition equipment Learner terminals (mobile communication terminals / tablet computing terminals / personal computing terminals); teacher terminals (personal computing terminals / tablet computing terminals); classroom interactive terminals (interactive display screens); voice acquisition equipment (terminal-built-in microphones or classroom ceiling-mounted microphone arrays). Input method percentage Voice input accounted for 37%; text input accounted for 63%. Terminal type proportion Mobile communication terminals 58%; tablet computing terminals 12%; personal computing terminals 30%.

[0092] The participants consisted of 236 students and 4 instructors from 6 classes across two general education courses. Class discussions were held every Tuesday and Thursday from 9:50 AM to 11:25 AM, while after-class tutoring was available daily from 6:00 PM to 11:00 PM. Student-used devices were categorized based on actual usage: mobile communication terminals accounted for 58%, tablets for 12%, and personal computing terminals for 30%. Input methods were distributed as follows: voice input 37%, text input 63%, with voice input primarily occurring during class discussions and text input mainly during after-class tutoring. Classroom audio was captured by a microphone array, while personal device audio was captured by the terminal's built-in microphone. Voice was transcribed into text in real-time and then integrated into the same communication channel.

[0093] In classroom applications, before starting a lesson, teachers select the day's chapter topic and learning objectives on the teaching platform. The system retrieves the chapter topic, knowledge point list, and teaching schedule from the course resource library; value objectives, ability objectives, knowledge objectives, and assessment requirements from the teaching plan library; and scenario type, applicable audience, content boundaries, prohibited expressions, and risk level criteria from the compliance rule library. When students initiate a dialogue in the classroom or on the platform, the system automatically encapsulates each round of input into a dialogue data packet. The data packet includes a session number, round identifier, unified timestamp, role identifier, and session context index, and also writes the topic identifier, objective identifier, and constraint identifier into the session metadata. This ensures that when the same student repeatedly asks follow-up questions in class and after class, the system can maintain consistency in the discussion topic, objectives, and boundaries, avoiding the break in context when switching devices.

[0094] During the discussion, students often raise questions that express their stance or emotions. Instead of directly outputting the generated content as is, the system first structures the value objectives into a set of computable constraints, and unifies the rules for prohibited content, mandatory points, risk criteria, and output format into a single verifiable configuration. The system then models the dialogue's state, outputting the current dialogue state object, including viewpoint, cognitive stage, risk level, value dimension distribution, and alignment deviation indicator, and binds this state to the set of computable constraints to form a control context. The control context drives the selection of guiding actions: when the concept is identified as unclear and the risk is low, the system tends to ask clarifying questions and provide key value points. When a deviation from the established stance is identified and the controversy escalates, the system shifts to case comparison and viewpoint correction. When a high-risk situation is identified and triggers a criterion, the system executes a redirection or refusal to answer, and provides a compliant direction for continued discussion. After candidate responses are generated, they enter the alignment and compliance verification chain. The system outputs the verification results according to the rules and performs rewriting, downgrading, redirection, or refusal to answer actions before finally presenting the guidance output for this round to the students. Every rule trigger, every action, and the text summary index and timestamp before and after each action are written to the audit log, forming an evidence chain entry. Teachers can replay the specific round and triggering basis.

[0095] Table 2 System Response Delay Statistics and Network Congestion Records

[0096] Scene Median response time (seconds) 95th percentile response time (seconds) Remark Classroom network conditions 1.48 2.91 Location: Smart Classroom 3-308 Evening rush hour after school (6:00 PM - 11:00 PM) 1.62 3.27 Platform peak access period Network Congestion Day (Classroom) 1.48 3.98 Brief periods of congestion occurred on November 20, 2025, and December 4, 2025, causing the 95th percentile to rise.

[0097] To better reflect the pace of real teaching, system logs were logged daily during the trial run. A total of 1,247 sessions were generated over six weeks, including 412 classroom discussion sessions and 835 after-school tutoring sessions; a total of 14,986 rounds; 5,471 rounds of voice input and 9,515 rounds of text input; an average of 12.0 rounds per session. Relevant statistics are shown in Table 2. The median response latency recorded by the system was 1.48 seconds under classroom network conditions and 1.62 seconds during the evening peak after class. The 95th percentile response latency was 2.91 seconds and 3.27 seconds, respectively. Relevant latency statistics are shown in Table 3. The wireless network at the classroom location "Smart Classroom 3-308" experienced two short-term congestions on November 20, 2025, and December 4, 2025, causing the 95th percentile latency in the classroom scenario to rise to 3.98 seconds. The system was still able to complete verification, processing, and evidence retention, and the log link remained continuous.

[0098] Table 3. Comparison of key indicators between control and experimental settings.

[0099] index Comparison mode Experimental mode illustrate High-risk rounds (per thousand rounds) 5.2 3.1 Normalization of high-risk rounds requiring teacher intervention Total number of high-risk rounds (rounds) 78 53 Control period 2 weeks / Experiment period 4 weeks, for traceability and verification The proportion of repeated follow-up questions in high-risk rounds 44% 16% The percentage of repeated follow-up questions during high-risk treatment The proportion of off-topic conversations (lasting ≥3 rounds) 20.3% 7.9% Deviating from the current chapter's theme and continuing for three or more rounds

[0100] To verify the differences from the general solution, University A configured a control mode on the same platform. This mode only performed basic sensitive word filtering and manual sampling, without introducing value-based structured constraints, dialogue-driven guidance, alignment verification, or process-level evaluation and evidence retention. The control mode ran from November 3rd to November 16th, 2025 (2 weeks), while the experimental mode ran from November 17th to December 12th, 2025 (4 weeks). Statistics were normalized based on the number of events per thousand rounds to eliminate differences in cycle length.

[0101] Regarding risk compliance, the control model recorded 78 high-risk rounds requiring teacher intervention within two weeks, translating to 5.2 rounds per thousand rounds; the experimental model recorded 53 high-risk rounds within four weeks, translating to 3.1 rounds per thousand rounds. See Table 4 for the comparison. Regarding the handling results, in the control model, the proportion of students repeatedly asking follow-up questions due to "direct blocking or vague prompts" in high-risk rounds was 44%; in the experimental model, after redirecting or refusing to answer and providing compliant learning directions, the proportion of repeated follow-up questions decreased to 16%. See Table 4 for the comparison. Regarding off-topic content, in the control model, the proportion of off-topic conversations "deviating from the current chapter topic and lasting for more than three rounds" was 20.3%, while in the experimental model it was 7.9%. See Table 4 for the comparison. The determination of off-topic conversations was based on the consistency judgment rules from the teaching platform and the teacher's review conclusions. The judgment records were exported and archived on December 16, 2025.

[0102] Regarding value alignment and the effectiveness of guidance, University A organized a centralized review by teachers. The review took place in Room 5-512 on the 5th floor of the main teaching building, from 2:00 PM to 5:30 PM on December 18, 19, and 20, 2025. The review sample consisted of 400 stratified samples from classroom and after-class conversations, totaling 4,600 rounds, ensuring coverage of different classes, topics, and input methods. Teachers used a standardized scale for scoring, which included five items: coverage of key values, consistency of stance, coherence of guidance, appropriateness of compliant handling, and interpretability and verifiability. Each item was worth 20 points, for a total of 100 points. The average total score for the control model was 72.1, with a standard deviation of 9.4; the average total score for the experimental model was 87.6, with a standard deviation of 6.8. Looking only at the coverage of key value points, the control model averaged 14.2 / 20, while the experimental model averaged 17.8 / 20. Regarding the appropriateness of compliant handling, the control model averaged 13.6 / 20, while the experimental model averaged 18.1 / 20. Relevant review data is shown in Table 5. A typical difference clearly recorded in the teacher reviews was that the control model often resulted in a "one-size-fits-all" approach when controversial statements were raised, leading to the interruption of the discussion; the experimental model, after triggering the rules, was able to shift to factual clarification and presentation of key value points within the course framework, allowing the discussion to continue.

[0103] Regarding process traceability and teaching governance costs, the platform statistically analyzed the time required for teachers to locate disputed rounds. In the control mode, the median time from receiving student feedback to locating the specific issue round was 3 minutes and 52 seconds; in the experimental mode, teachers could directly click on evidence chain entries to view dialogue segment indexes, trigger rule identifiers, criterion identifiers, and action identifiers, reducing the median time to 1 minute and 24 seconds. See Table 5 for the relevant comparison. Regarding record coverage, the control mode achieved a round coverage rate of 35.7% with "input-output replayable records"; the experimental mode achieved a round coverage rate of 98.9% with complete "trigger-action-evidence chain-log index" records. See Table 5 for the relevant comparison. The remaining missing rounds mainly occurred during the classroom network congestion period from 10:23 to 10:27 on December 4, 2025. The system recorded the session number and round, but audio transcription failure resulted in some missing segments, which were subsequently supplemented by text from the students after class.

[0104] Based on the above time, location, and data from Tables 1 to 3, it can be seen that in real classroom and after-school tutoring environments, by structuring value objectives into calculable constraints and integrating them with multi-round dialogue state identification, guided action selection, alignment and compliance verification, process-level credible assessment, and audit evidence retention, the system can reduce the proportion of off-topic deviations and high-risk rounds without changing the teacher's teaching process, improve the coverage of key value points and the appropriateness score of compliance handling, and significantly reduce the cost of teacher review and positioning, forming a traceable, auditable, and governable dialogic ideological and political education closed loop.

[0105] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dialogue-guided method for "value alignment" in ideological and political education, characterized in that: Includes the following steps: Step S1: Obtain dialogue input and teaching constraint information to form a dialogue data package; Step S2: Structure the teaching constraint information into a verifiable constraint set, and configure trigger conditions and action identifiers for each constraint item in the constraint set; Step S3: Perform dialogue state recognition and modeling on the dialogue data packet to obtain the dialogue state object, and associate the dialogue state object with the constraint set to generate a control context; Step S4: Determine the target guidance action and its guidance parameters based on the control context, and call the dialogue generation model to generate candidate guidance outputs; Step S5: Perform alignment and compliance checks on the candidate guidance outputs, determine the check results based on the constraint set, perform alignment control processing with the check results, output the guidance outputs after alignment control processing, and record the triggered rule identifiers and handling action identifiers; Step S6: Perform a credibility assessment and audit evidence retention process on the dialogue process containing the guidance output, generate assessment results and evidence chain information according to the preset assessment index system, and write the assessment results and evidence chain information into the audit storage.

2. The dialogue guidance method according to claim 1, characterized in that, Acquire dialogue input and instructional constraint information and form a dialogue data package, including: Collect dialogue content and establish a session identifier for the session; Perform text normalization on the dialogue content to obtain normalized text, and establish a corresponding index between the original text and the normalized text; Generate a unified time stamp for the dialogue content, and generate fragment time stamps corresponding to the normalized text; Receive course topic information and generate topic identifiers, then write the topic identifiers into the session metadata corresponding to the session identifier; Receive teaching objective information and generate objective identifiers, and establish the association between objective identifiers and theme identifiers; Receive specification constraint information and generate constraint identifiers, and establish the association relationship between constraint identifiers and subject identifiers and target identifiers; Based on the above processing results, construct the dialogue data packet and write the dialogue data packet to the session storage.

3. The dialogue guidance method according to claim 1, characterized in that, The teaching constraint information is structured into a verifiable set of constraints and configured with trigger conditions and action identifiers, including: Read the teaching constraint information from the dialogue data packet, generate structured constraint configuration data, and set version identifier and scope identifier for the structured constraint configuration data; Perform element parsing on the teaching constraint information to obtain an element set, and set identifiers and parameters for the elements in the element set; A set of rules is generated based on a set of elements. The rules in the set of rules correspond to verifiable trigger conditions and action identifiers, respectively. Merge the rule set to generate a constraint set, and configure verifiable trigger conditions, action identifiers, priority parameters and conflict resolution parameters for the constraint items in the constraint set.

4. The dialogue guidance method according to claim 1, characterized in that, Perform dialogue state recognition and modeling on dialogue data packets and generate control context. include, Read the dialogue data packet from step S1 and generate state analysis input; Extract the necessary information for state analysis from the dialogue data packet to form the state analysis input; Topic identification is performed on the state analysis input, and the identification results are matched with the set of topic elements to obtain topic deviation indicators; Map state analysis inputs to position results and label input segments that support position categories; The cognitive stage category is obtained through stage reasoning, and the stage is based on the index to record the position of the input fragment that supports the cognitive stage category; In the risk identification process, the triggering fragment is first located and a hit index is generated, and then the risk level category is determined by the hit index; The coverage strength of the value dimensions is calculated to form the value dimension distribution, and the evidence index is written synchronously with the coverage calculation; Input the position results, cognitive stage categories, risk level categories and value dimension distribution into the consistency verification process, and output alignment deviation indicators. The alignment deviation indicators represent the deviation type, deviation intensity and trigger index. The topic tags, stance results, cognitive stage categories, risk level categories, value dimension distribution and alignment deviation indicators are aggregated to form a dialogue state object, which is associated with the conversation identifier and interaction index. A control context is generated from the dialogue state object and the constraint set, and the control context is written to the state store.

5. The dialogue guidance method according to claim 1, characterized in that, Based on the control context, the target guidance action and its guidance parameters are determined, and candidate guidance outputs are generated, including: Read the control context from step S4 and generate the guidance decision input; Determine the guidance objective based on the guidance decision input, and set the objective parameters for the guidance objective; Based on the dialogue stage information and risk level information in the dialogue state object, action gating is performed to obtain the set of allowed actions and the set of prohibited actions, and action parameters are configured for the actions. The target guidance action is determined based on the guidance objective and the set of permitted actions, and guidance parameters corresponding to the target guidance action are generated. Perform pruning on the historical dialogues pointed to by the session context index to obtain a set of context fragments, and configure fragment parameters for the context fragments; The target guidance action, guidance parameters, context fragment set, and format constraints are assembled into the generated request data; The dialogue generation model is invoked to perform generation processing on the generation request data, output candidate guidance outputs, and write the candidate guidance outputs to the output.

6. The dialogue guidance method according to claim 1, characterized in that, Perform alignment and compliance checks on the candidate boot outputs and output the boot outputs after alignment control processing. include, The dialogue guidance method according to claim 1 is characterized in that, performing alignment and compliance checks on candidate guidance outputs and outputting guidance outputs after alignment control processing includes: Retrieve candidate guidance outputs and load constraint sets to form validation inputs. Make judgments on the trigger conditions of the disabling rules and record the disabling hit information when triggered. Perform risk identification, input the identification results into risk rules to complete the judgment, output risk hit information, perform alignment quantification processing on candidate guidance outputs, and produce alignment results; Establish a correspondence between candidate guidance outputs and key rules, determine the coverage gaps based on the correspondence, and form key verification information; Extract the format features of the candidate guide output, compare them with the format rules, and return the comparison results as format verification information; The system integrates prohibited hit information, risk hit information, alignment results, key point verification information, and format verification information to form a verification result object. The decision to handle the situation is derived from the verification result object, and the action to handle the situation is determined based on the priority parameter and the conflict resolution parameter. The execution of the action generates a guidance output and a action log. The guidance output and action log are then written to the audit storage for subsequent steps.

7. The dialogue guidance method according to claim 1, characterized in that, The aforementioned process involves performing credibility assessments and audit evidence retention on the dialogue process: Read data from the session storage, align and aggregate it according to the session identifier and interaction index, and generate evaluation records; Perform structured extraction on the evaluation records to generate evaluation samples associated with the interaction index; The evaluation sample is evaluated by calling a preset scoring function to generate a scoring result. The scoring function performs a weighted calculation on the indicator parameters and outputs a comprehensive conclusion identifier. Based on the evaluation sample and the scoring result, generate cause location information and evidence chain information, and write the scoring result, cause location information and evidence chain information into the audit storage; A retrieval index associated with the session identifier and the interaction index is generated in the audit storage for subsequent dialogue guidance calls and tracing.

8. A credible evaluation system for guiding dialogue on "value alignment" in ideological and political education, characterized in that: The system includes: The data acquisition and encapsulation module acquires dialogue input and teaching constraint information, and generates dialogue data packages. The constraint construction module structures teaching constraint information to generate a constraint set, and configures trigger conditions and action identifiers for the constraint items in the constraint set; The state modeling and context generation module performs dialogue state recognition and modeling on the dialogue data packet to obtain a dialogue state object, and associates the dialogue state object with the constraint set to generate a control context. The guidance generation module determines the target guidance action and guidance parameters based on the control context, and calls the dialogue generation model to output candidate guidance outputs. The verification and handling module performs alignment and compliance verification on the candidate guidance output, generates verification results and determines the handling action, executes the handling action to output the guidance output, and records the handling log. The evaluation and audit module generates evaluation results and evidence chain information for the dialogue process including the guidance output, and writes the evaluation results and evidence chain information into the audit storage for subsequent dialogue guidance calls and traceability.

Images