Knowledge Mastery-Based Learning Interaction Methods and Devices

Abstract

This application discloses a learning interaction method and apparatus based on knowledge mastery. The method includes: receiving user-inputted interaction data related to a target concept; extracting evidence features from the interaction data and calculating a contradiction penalty score; updating the current mastery score corresponding to the target concept based on the evidence features and the contradiction penalty score to obtain a target mastery score; determining a retrieval strategy and unlocking status based on the target mastery score, evidence features, and a list of currently active misunderstandings maintained by a misunderstanding tracker module; generating original response data based on the retrieval strategy, unlocking status, and a preset knowledge base; and outputting corresponding final response data after the original response data passes leakage protection detection. This method can comprehensively optimize intelligent tutoring logic, thereby steadily improving the quality of adaptive teaching and the effectiveness of actual learning.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a learning interaction method and apparatus based on knowledge mastery. Background Technology

[0002] In AI-driven educational applications such as intelligent tutoring systems, adaptive learning platforms, and massively open online courses (MOOCs), while large language models can answer questions and explain concepts, they often output complete answers directly without verifying whether learners truly understand the knowledge points. This bypasses the necessary thinking process in learning, reducing the effectiveness of knowledge acquisition. Existing solutions typically estimate learners' mastery of knowledge points probabilistically based on behavioral data such as correct or incorrect answers. However, this method infers mastery solely from correct or incorrect answers, resulting in low accuracy. Furthermore, it cannot dynamically adjust the tutoring system's response strategy based on the estimated mastery level; regardless of the learner's level of mastery, the system outputs the same response, making it difficult to provide differentiated tutoring content based on the learner's actual comprehension. Summary of the Invention

[0003] In view of the above problems, this application provides a learning interaction method and apparatus based on knowledge mastery, which can solve the problems of inaccurate knowledge mastery estimation and inability to dynamically adjust response strategies according to knowledge mastery in existing methods.

[0004] Firstly, this application provides a learning interaction method based on knowledge mastery, including: Receive interactive data from the user related to the target concept; Evidence features are extracted from the interaction data, and a contradiction penalty calculation is performed on the interaction data to obtain a contradiction penalty score; The current mastery score corresponding to the target concept is updated based on the evidence features and the contradiction penalty score to obtain the target mastery score; Based on the target mastery score, the evidence features, and the current active misunderstanding list maintained by the misunderstanding tracker module, the retrieval strategy and unlocking status are determined; Based on the retrieval strategy, the unlock status, and the preset knowledge base, generate the original response data; After the original response data passes the leakage protection detection, the corresponding final response data is output based on the original response data.

[0005] In the above technical solution, the method can accurately quantify the learner's knowledge mastery by combining the evidence features of interactive data and the contradiction penalty score. At the same time, it dynamically adjusts the retrieval and content unlocking strategies by combining active misunderstanding content, and generates differentiated tutoring responses that are adapted to the learner's comprehension level. This effectively avoids the problem of missing thinking caused by direct and complete answers, thereby improving the accuracy of knowledge mastery assessment, realizing personalized adaptive learning interaction, and strengthening the learning effectiveness of knowledge points.

[0006] In some implementations, after receiving user-inputted interactive data related to the target concept, the method further includes: The interaction data is subjected to intent detection to obtain intent detection results; Determine whether the intent detection result includes an evaluation intent; If the evaluation intent is not included, the current global session state is updated based on the interaction data to obtain the target global session state; Determine whether the target global session state is in a talkback state; If it is a review session, the current mastery score corresponding to the target concept is calculated based on time decay to obtain the decayed current mastery score.

[0007] In the above technical solution, the method can accurately identify the assessment intent to activate the assessment blocking mechanism, thereby avoiding the risk of direct feedback of results. At the same time, through session state control, it ensures that the mastery time decay is only triggered in the repetition state, which conforms to the memory forgetting law and ensures the timeliness and rationality of mastery update.

[0008] In some implementations, extracting evidence features based on the interaction data includes: Structured features are extracted from the user feedback content in the interaction data; wherein, the structured features include coverage, accuracy, and coherence. Obtain the list of currently active misunderstandings maintained by the misunderstanding tracker module and the total number of times prompts have been used for the target concept; The initial evaluation score is calculated based on the structured features, and the initial evaluation score is discretized based on a threshold to obtain the review score. The misunderstanding clarity is calculated based on the sum of the severity of each misunderstanding in the current active misunderstanding list and the preset maximum severity budget. Calculate the micro-test score based on the micro-test data in the interactive data; The prompt efficiency is calculated based on the total number of prompts used and the preset maximum number of prompts. The evidentiary features include the repetition score, the misunderstanding clarity, the microtest score, and the prompting efficiency.

[0009] In the above technical solution, the method can comprehensively collect features related to learning behavior and expression quality, and quantify them into diversified evidence indicators, thereby achieving refined and standardized calculation and fusion of learning evidence, and thus comprehensively reflecting the learner's true level of understanding and error cognition.

[0010] In some implementations, the step of calculating the contradiction penalty on the interaction data to obtain a contradiction penalty score includes: Obtain the correct statement data corresponding to the target concept; A consistency check is performed on the interaction data based on the correct statement data to obtain a set of contradictory results. When the set of contradictory results includes contradictory results, a contradiction penalty score is calculated based on the set of contradictory results. When the set of contradictory results only includes results that do not contain contradictions, the contradiction penalty score is determined to be 0; The consistency checks include equation consistency checks, unit consistency checks, and symbol mapping consistency checks. The set of contradictory results includes results that are not contradictory and / or results that are contradictory; The contradictory results include at least one contradiction type and a severity weight corresponding to each contradiction type; The contradiction types include one or more of the following: equation consistency type, unity consistency type, and symbolic mapping consistency type.

[0011] In the above technical solution, the method can comprehensively identify various cognitive contradictions in the user's expression through multi-dimensional consistency checks, and reasonably calculate the corresponding penalty score according to the actual type and severity of the contradiction, thereby achieving objective correction of mastery based on the user's actual cognitive defects.

[0012] In some implementations, updating the current mastery score corresponding to the target concept based on the evidence features and the contradiction penalty score to obtain the target mastery score includes: The evidence features are weighted and fused to obtain the total evidence score; The target mastery score is calculated based on the current mastery score, the total evidence score, the contradiction penalty score, the preset learning rate, and the preset penalty rate corresponding to the target concept.

[0013] In the above technical solution, the method can comprehensively measure the user's understanding by weighted fusion of various learning evidences, and at the same time introduce contradictory penalty constraints to constrain the negative impact of cognitive errors, thereby making the updated mastery results more reasonable.

[0014] In some implementations, determining the retrieval strategy and unlocking status based on the target mastery score, the evidence features, and the currently active misunderstanding list maintained by the misunderstanding tracker module includes: Based on the preset misunderstanding removal conditions, the misunderstanding status of the currently active misunderstanding list maintained by the misunderstanding tracker module is updated to obtain the target active misunderstanding list; The unlocking status is determined based on the target mastery score, the retelling score in the evidence features, the target active misunderstanding list, and the preset composite unlocking conditions; Based on the target mastery score and the unlock status, determine the retrieval strategy; The retrieval strategy is one of the following: full access strategy, scaffold enhancement strategy, and scaffold only strategy.

[0015] In the above technical solution, the method can combine real-time misunderstanding status, knowledge mastery level and user comprehension performance to jointly determine content unlocking permissions, and adaptively match differentiated knowledge base retrieval strategies based on comprehensive evaluation results. This allows for control over the scope of teaching content push from the source, avoiding premature output of complete answers, and thus adapting to the learning needs of different cognitive levels.

[0016] In some implementations, generating raw response data based on the retrieval strategy, the unlock status, and a preset knowledge base includes: The search is performed in a preset knowledge base according to the search strategy to obtain the search data; Generate a response generation instruction based on the retrieval strategy and the unlock status constraint; Based on the retrieved data and the response generation instruction, the large language model is invoked to generate the original response data.

[0017] In the above technical solution, the method can accurately filter knowledge base content according to a predetermined retrieval strategy, and at the same time reasonably constrain the model generation logic by combining the unlock status, so that the large language model outputs original response content that is adapted to the user's current cognitive level.

[0018] In some implementations, after generating the original response data based on the retrieval strategy, the unlock status, and a preset knowledge base, the method further includes: The original response data is detected according to preset leakage protection rules to obtain leakage protection detection results; wherein, the leakage protection detection results are used to indicate whether a leakage type is detected, and the leakage type includes at least one of step count exceeding the limit, final answer marking, numerical equation matching, and answer keyword matching; When the leakage protection detection result indicates that no leakage type was detected, the original response data is determined to have passed the leakage protection detection, and the step of outputting the corresponding final response data based on the original response data is executed.

[0019] In the above technical solution, the method can comprehensively screen various answer leakage elements in the response content according to the sound leakage protection rules, and strictly block non-compliant complete answer information, thereby ensuring the implementation effect of tiered tutoring and progressive learning.

[0020] In some embodiments, the method further includes: When the leakage protection detection result indicates that at least one type of leakage has been detected, obtain the current unlock status; When the current unlock state is not a fully unlocked state, the original response data is rewritten to obtain security response data; Output the corresponding final response data based on the security response data; The current unlock state can be one of the following: fully unlocked, locked, or forced locked.

[0021] In the above technical solution, the method can flexibly adapt to the actual unlocking status after identifying the problem of leakage of answer content, and rewrite and optimize the response content in a timely manner for scenarios where the answer is not fully unlocked, thereby effectively avoiding the problem of premature leakage of key answers and ensuring the stable operation of the tiered tutoring mechanism.

[0022] Secondly, this application provides a learning interaction device based on knowledge mastery, the learning interaction device based on knowledge mastery comprising: The receiving unit is used to receive interactive data related to the target concept input by the user; Extraction unit, used to extract evidence features based on the interaction data; The calculation unit is used to calculate the contradiction penalty on the interactive data to obtain the contradiction penalty score; The mastery update unit is used to update the current mastery score corresponding to the target concept according to the evidence features and the contradiction penalty score, so as to obtain the target mastery score; The determining unit is used to determine the retrieval strategy and unlocking status based on the target mastery score, the evidence features, and the currently active misunderstanding list maintained by the misunderstanding tracker module. The generation unit is used to generate raw response data based on the retrieval strategy, the unlocking status, and a preset knowledge base; The output unit is used to output the corresponding final response data based on the original response data after the original response data passes the leakage protection detection.

[0023] In the above technical solution, the device can accurately quantify the learner's knowledge mastery by combining the evidence features of interactive data and the contradiction penalty score. At the same time, it can dynamically adjust the retrieval and content unlocking strategies by combining active misunderstanding content, and generate tutoring responses that are adapted to the learner's comprehension level. This effectively avoids the problem of missing thinking caused by direct and complete answers, thereby improving the accuracy of knowledge mastery assessment, realizing personalized adaptive learning interaction, and strengthening the learning effectiveness of knowledge points.

[0024] Thirdly, this application provides an electronic device, the electronic device including a memory and a processor, the memory for storing a computer program, the processor running the computer program to cause the electronic device to perform the knowledge mastery-based learning interaction method as described in any one of the first aspects.

[0025] Fourthly, this application provides a readable storage medium storing a computer program, which, when executed by a processor, performs the knowledge mastery-based learning interaction method described in any one of the first aspects.

[0026] Fifthly, this application provides a computer program product comprising a computer program that, when executed by a processor, performs the knowledge mastery-based learning interaction method described in any one of the first aspects.

[0027] The beneficial effects of this application are as follows: First, it can objectively measure learners' knowledge mastery based on multi-source learning evidence, correcting the shortcomings of traditional subjective assessments; second, it can identify and lock down users' misconceptions in real time, and reasonably restrict the scope of open teaching content based on misunderstanding prevention mechanisms; third, through full-process leakage protection and content gating control, it standardizes the content generation boundaries of large models from a technical perspective, preventing the arbitrary disclosure of complete answers; it can also accurately identify unauthorized use intentions, ensuring the academic integrity of daily learning scenarios; furthermore, it can dynamically adjust knowledge base retrieval and content access permissions based on mastery status, adaptively matching suitable learning materials; and finally, based on this, it achieves reasonable control over the pace of knowledge output and compels users to actively think and explore. Therefore, overall, this application can comprehensively optimize intelligent tutoring logic, thereby steadily improving the quality of adaptive teaching and the effectiveness of actual learning. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating a knowledge mastery-based learning interaction method in some embodiments of this application. Figure 2 This is a schematic diagram of the structure of a knowledge-based learning interaction device in some embodiments of this application; Figure 3 This is a schematic diagram of the structure of an electronic device in some embodiments of this application. Detailed Implementation

[0030] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0032] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more (including two), similarly, "multiple sets" refers to two or more sets (including two sets), and "multiple pieces" refers to two or more pieces (including two pieces) unless otherwise explicitly defined.

[0033] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0034] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0035] To address the shortcomings of existing intelligent tutoring assessment methods, such as subjective bias, lack of content output control, and insufficient resource adaptation, this application proposes a knowledge-mastery-based learning interaction method. This method focuses on learners' knowledge mastery, integrating multiple learning evidences to achieve objective quantitative assessment. It incorporates a conflict-punishment mechanism for cognitive errors to dynamically update mastery levels. Furthermore, it sets composite unlocking conditions based on misunderstanding status and learning performance, employs a tiered retrieval strategy to differentiate knowledge base access, and introduces assessment intent recognition and content leakage prevention mechanisms. These multi-dimensional measures ensure compliant control of answer content, reasonably manage the pace of knowledge output, and guide users to think independently. Based on this, this application ultimately constructs a standardized and adaptive intelligent tutoring model, effectively improving the overall effectiveness of intelligent teaching.

[0036] like Figure 1 As shown, some embodiments of this application provide a learning interaction method based on knowledge mastery, which includes: S100: Receive interactive data related to the target concept input by the user.

[0037] In this embodiment, the target concept refers to the course knowledge content that learners need to understand and master, such as subject knowledge points, theoretical definitions, formulas and theorems, and professional principles.

[0038] In this embodiment, interactive data refers to multimodal conversation data such as user-inputted response text, answer content, quiz answers, question statements, feedback information, and operational behaviors during the learning interaction process.

[0039] As an optional implementation, before step S100, the method may pre-compile the global initialization configuration, wherein the step of pre-compiling the global initialization configuration may include: The initialization of the mastery vector, loading of all fixed parameters and preferred parameters, loading of the preset knowledge base, and initialization of the dialogue state manager are completed in advance.

[0040] For example, when a user is a new learner, the user has no interaction history. In this case, the mastery vector is initialized as M0=[0.5,0.5,...,0.5]. This method uses a neutral prior of 0.5 to represent the state of unknown knowledge (not a state of no knowledge) and supports subsequent bidirectional adjustment of scores based on positive and negative evidence.

[0041] In this embodiment, the method also preloads all fixed parameter ranges, preferred parameters, and detailed rules, wherein the complete parameter configuration can be found in Table 1.

[0042] Table 1. Complete Initialization Parameter Configuration Table

[0043] In this embodiment, the preset knowledge base stores educational content with content block type tags. These content blocks are divided into six categories: definition, example, hint, partial solution, complete solution, and answer.

[0044] In this embodiment, the method also needs to initialize the nine-state rotation mechanism. The default initial state is set to the greeting state (S0), which provides basic support for the subsequent interaction flow.

[0045] In this embodiment, the nine-state rotation mechanism refers to a full-process conversation state transition control mechanism that covers greetings, guided questions, concept review, micro-quizzes, Q&A interaction, error correction feedback, prompt guidance, evaluation and detection, and conversation end.

[0046] S200. Extract evidence features based on interactive data.

[0047] In this embodiment, the method can completely extract four types of evidence data from all user interactions and text responses in this round: Explanation Score (EB), Clarity of Misunderstanding (MC), Micro-quiz Score (MQ), and Cueing Efficiency (HU). These four types of evidence are the core input features for subsequent mastery calculation.

[0048] S300. Calculate the contradiction penalty on the interactive data to obtain the contradiction penalty score.

[0049] In this embodiment, the method employs a dual-mode detection approach combining rules and a Large Language Model (LLM) to perform multi-dimensional consistency checks on user interaction data (explained in detail later). It also combines the severity weights corresponding to different contradiction types to quantitatively calculate the corresponding contradiction penalty score for this round. The contradiction penalty score is a penalty calculated based on the logical inconsistency between the detected user input and correct knowledge.

[0050] S400. Update the current mastery score corresponding to the target concept based on the evidence characteristics and the contradiction penalty score to obtain the target mastery score.

[0051] In this embodiment, the method uses the evidence-weighted fusion result of the evidence features and the quantified contradiction penalty score, combined with fixed parameters such as preset learning rate and preset penalty rate, to iteratively update the current mastery score of the target concept according to a unified calculation rule, and outputs the target mastery score within the constraint interval, thereby achieving precise iterative updates of the mastery score of a single target concept. The mastery score represents a quantitative value of the learner's current understanding level of a specific concept, and its value can range from 0 to 1.

[0052] S500 determines the retrieval strategy and unlock status based on the target mastery score, evidence features, and the current active misunderstanding list maintained by the misunderstanding tracker module.

[0053] In this embodiment, after updating the misunderstanding status of the active misunderstanding list maintained in real time, the method comprehensively determines the unlocking status by combining the target mastery score, the retelling score in the evidence features, and the preset composite unlocking conditions, and then combines the mastery score with the unlocking status to match the corresponding retrieval strategy.

[0054] The Misunderstanding Tracker module is a functional component used to record, track, and manage the misunderstandings and their severity status exhibited by learners. The Current Active Misunderstandings List is a list of misunderstandings and their attribute information that have not yet been clarified or corrected by learners and are still in an active state.

[0055] In this embodiment, the retrieval strategy is a set of rules determined based on the learner's mastery level, used to determine what type of teaching content to retrieve from the knowledge base.

[0056] S600 generates raw response data based on the retrieval strategy, unlock status, and preset knowledge base.

[0057] In this embodiment, the method performs content retrieval in a preset knowledge base according to a predetermined retrieval strategy. Simultaneously, it generates a response generation instruction with constraints based on the unlock status, and invokes a large language model based on the retrieved data and the response generation instruction to generate original response data that meets the constraints. The unlock status is a mode flag indicating whether the composite unlock conditions are currently met, allowing the system to output a complete solution.

[0058] For example, this method can filter compliant content blocks within the knowledge base according to the selected retrieval strategy, generate corresponding constraint instructions based on the current unlock status, and call the large language model to generate original response data that meets the control requirements.

[0059] S700: After the original response data passes the leakage protection detection, the corresponding final response data is output based on the original response data.

[0060] In this embodiment, the method utilizes a bilingual leakage protection detection mechanism throughout the entire process. The original response data must first undergo a standardized leakage protection detection process. Only after passing the detection and approval can the final response data be directly output based on the original response data.

[0061] To ensure the effectiveness of differentiated instruction and progressive learning, step S700 may also include: S711. Detect the original response data according to the preset leakage protection rules, obtain the leakage protection detection result, and execute step S712 or S713 according to the leakage protection detection result.

[0062] In this embodiment, the leakage protection detection result is used to indicate whether a leakage type has been detected. Leakage types include at least one of the following: step count exceeding limit, final answer marking, numerical equation matching, and solution keyword matching. Step count exceeding limit is a violation type triggered because the number of steps in the generated response exceeds a preset budget. Final answer marking is a violation type triggered because the generated response contains keywords indicating a conclusion (such as "therefore," "the answer is"). Numerical equation matching is a violation type triggered because the generated response contains equation patterns of the form where variables equal numerical values. Solution keyword matching is a violation type triggered because the generated response contains specific solution-final phrases (such as "in conclusion," "solved").

[0063] In this embodiment, the method can also achieve accurate identification of various answer leakage risks through leakage prevention rules that cover both Chinese and English.

[0064] In this embodiment, the method can use a joint detection mechanism of Chinese and English bilingual regular expression matching and keyword weighting to accurately identify various risks of answer leakage. The specific detection rules are shown in Table 2.

[0065] Table 2. Bilingual Content Leakage Prevention Detection Rules

[0066] S712. When the leakage protection detection result indicates that no leakage type was detected, determine that the original response data passed the leakage protection detection (i.e., the original response data passed the leakage protection detection), and execute the step in S700 to output the corresponding final response data based on the original response data.

[0067] In this embodiment, if no leaked content is identified in the answer, the original response data can be determined to be compliant, and the process can directly proceed to the output of the final response data.

[0068] S713. When the leakage protection detection result indicates that at least one type of leakage has been detected, obtain the current unlock status.

[0069] In this embodiment, the current unlocking state is one of three: fully unlocked, locked, or forced locked. A fully unlocked state indicates that all composite unlocking conditions have been met, and the system can output a complete solution. A locked state indicates that the composite unlocking conditions have not been met, and the system is forcibly restricted to outputting only a scaffolded response. A forced locked state is an unoverridable locked state triggered by evaluation intent detection or special conditions, and output is forcibly restricted regardless of the mastery score.

[0070] S714. When the current unlock status is not fully unlocked, the original response data is rewritten to obtain the security response data.

[0071] In this embodiment, the method rewrites and optimizes the original response data for issues involving leaked answers, performs content masking and step trimming, thereby removing illegal and leaked content and generating compliant and secure response data.

[0072] S715. Output the corresponding final response data based on the security response data.

[0073] In this embodiment, the method can use the processed security response data as compliance content and use it to complete the final response data output delivery. Based on this, the method can use the rewritten security response data as the final output to achieve progressive scaffolding teaching guidance.

[0074] In this embodiment, progressive scaffolding instruction refers to a tiered tutoring approach that avoids directly providing complete solutions. Instead, it uses progressive aids such as thought process hints, partial analysis, and error-prone guidance to build cognitive scaffolds, guiding learners to think and explore independently, and gradually unlocking learning content as needed.

[0075] In the above embodiments, the method can accurately quantify the learner's knowledge mastery by combining the evidence features of interactive data and the contradiction penalty score, optimize the timeliness of mastery assessment based on time decay, and dynamically adjust the retrieval and content unlocking strategies in combination with active misunderstanding content. It can also generate differentiated tutoring responses that are adapted to the learner's comprehension level, thereby effectively avoiding the problem of missing thinking caused by direct and complete answers, thus improving the accuracy of knowledge mastery assessment, realizing personalized adaptive learning interaction, and strengthening the learning effectiveness of knowledge points.

[0076] In some embodiments, after step S100, the method may further include: S800: Perform intent detection on the interaction data and obtain the intent detection result.

[0077] In this embodiment, the method can adopt a two-stage detection approach using both Chinese and English languages, thereby distinguishing between normal learning interactions and evaluation queries based on a category-weighted keyword database, thus achieving accurate intent detection results.

[0078] In this embodiment, the specific details of bilingual intent detection are shown in Table 3.

[0079] Table 3. Configuration Table for Bilingual Assessment Intent Detection

[0080] S900. Determine whether the intent detection result includes the evaluation intent. If yes, end this process; otherwise, proceed to step S1000.

[0081] In this embodiment, the evaluation intent refers to the user's interactive intent to actively request standard answers, complete problem-solving steps, direct scoring, batch assessment, and final exam answers, with the aim of evaluating results and directly obtaining complete solutions.

[0082] In this embodiment, once an evaluation intent is detected, the method will automatically trigger an evaluation blocking mechanism. This evaluation blocking mechanism is used to restrict the output of complete answer content and terminate the current interaction process.

[0083] In this embodiment, the session state update process proceeds normally if there is no evaluation intention.

[0084] For example, as soon as the method recognizes that the user intends to ask for the answer directly, the complete solution to the problem, or simply to get a score on the exam, the method will immediately activate a strict anti-cheating / anti-direct answer mode.

[0085] In this mode, the method will directly lock content permissions, not providing the complete content; only allow hints of the thought process, not the full solution; automatically cover the final answer; provide a maximum of one solution step; and never release the complete answer or solution process, no matter how well you have mastered the knowledge points.

[0086] S1000: Update the current global session state based on the interaction data to obtain the target global session state.

[0087] In this embodiment, the method can be based on a nine-state rotation mechanism, combined with the user interaction data of the current round to update the session context and running state in real time, and generate an updated target global session state.

[0088] In this embodiment, the nine-state rotation mechanism is shown in Table 4.

[0089] Table 4. Complete Details of the Nine-State Rotation Mechanism

[0090] Here, "hints" refers to the number of times a user receives a hint. This is the data continuously recorded by the method, used to manage how many times a user receives a hint. The word "mask" refers to covering, shielding, or hiding, specifically the forced concealment of the final answer, final result, or final value.

[0091] S1100. Determine whether the target global session state is in the replay state. If yes, proceed to step S1200. If no, do not adjust the current mastery score corresponding to the target concept (i.e., the current mastery score updated in step S400).

[0092] In this embodiment, the method limits the triggering conditions for mastery decay through this step and initiates forgetting modeling operations only in the retelling state. For other conversation states, the method does not perform mastery decay adjustment, thereby ensuring rigorous and consistent evaluation logic. The retelling state is a state node in the dialogue state machine specifically used to require learners to restate the learned content in their own words to verify their understanding.

[0093] S1200: Perform time-based decay calculation on the current mastery score corresponding to the target concept to obtain the decayed current mastery score (i.e., the current mastery score updated in step S400).

[0094] In this embodiment, the method can update mastery by simulating the forgetting pattern of memory based on the time interval between two adjacent user interactions and the exponential decay formula.

[0095] In this embodiment, the mastery estimator reads the mastery vector M of the current knowledge point. t =[m1,m2,...,m K Based on the time interval Δt (hours) since the last interaction, an exponential forgetting model is introduced to complete the mastery decay. The decay calculation formula is as follows: m k =e ( λΔt) m k t ; Among them, M t Let t be the array representing the overall mastery of knowledge points. m1 m K Individual points are awarded for mastering the 1st, 2nd... Kth independent knowledge points; K represents the total number of knowledge points; m k t This represents the original mastery score of the k-th knowledge point at the end of the last interaction. k: The kth target knowledge point / concept; t: The previous interaction time; Δt is the time interval between two dialogues / interactions, in hours; λ represents the rate of memory forgetting, and its preferred decay coefficient is 0.01 / hour; where, the larger the value, the faster the memory is forgotten; the smaller the value, the longer the memory is retained. e is a natural constant with a fixed mathematical value of approximately 2.718; m k ′ represents the latest mastery score of the k-th knowledge point after time-based forgetting decay.

[0096] In this embodiment, the method completes the prior decay update of mastery based on the preferred parameter λ=0.01 / hour, and the decay details for different time intervals are shown in Table 5.

[0097] Table 5 shows the attenuation details for different time intervals.

[0098] In the above embodiments, the method can accurately identify the assessment intent to activate the assessment blocking mechanism, thereby avoiding the risk of direct feedback of results. At the same time, through session state management, it ensures that the mastery time decay is only triggered in the repetition state, which conforms to the memory forgetting law and ensures the timeliness and rationality of mastery updates.

[0099] In some embodiments, step S200 may include: S210. Extract structured features from the user feedback content in the interaction data; whereby the structured features include coverage, accuracy and coherence.

[0100] In this embodiment, the method can quantify the quality of students' self-recited content from three dimensions: content coverage, correctness of knowledge points, and logical coherence of statements. Coverage is an indicator used to quantify the completeness of key knowledge elements mentioned by learners when evaluating their recitation. Correctness is an indicator used to quantify the accuracy of the content presented by learners when evaluating their recitation. Coherence is an indicator used to quantify the fluency and coherence of logical expression when evaluating the content presented by learners.

[0101] S220. Obtain the list of currently active misunderstandings maintained by the misunderstanding tracker module and the total number of times the prompts have been used for the target concept.

[0102] In this embodiment, the method can retrieve the set of identified erroneous concepts and the historical frequency of prompt calls in real time, providing raw data for misunderstanding assessment and prompt efficiency calculation. The total number of prompts used represents the cumulative number of scaffolded prompts provided by the system to the learner during the interaction with a specific concept.

[0103] S230. Calculate the initial evaluation score based on the structured features, and perform threshold-based discretization on the initial evaluation score to obtain the playback score.

[0104] In this embodiment, the method can combine multi-dimensional structured features to weighted calculate the original review score, and the calculation formula is as follows: EBscore_raw=0.4×coverage+0.4×correctness+0.2×coherence; Where coverage is the degree of coverage, with a value range of [0,1], and is used to measure the completeness of the mention of key knowledge points; Correctness is the degree of accuracy, with a value range of [0,1], and is used to measure the accuracy of the content statement; Coherence is a measure of the fluency of expressed logic, with values ​​ranging from [0,1] to [0.0,1.0].

[0105] In this embodiment, the method can classify scores into three levels based on a threshold: A score below 0.33 is recorded as 0 points, 0.33 to 0.66 is recorded as 0.5 points, and a score above 0.66 is recorded as 1 point, resulting in the standardized recall score EB.

[0106] S240. Calculate the misunderstanding clarity based on the sum of the severity of each misunderstanding in the current active misunderstanding list and the preset maximum severity budget.

[0107] In this embodiment, the misunderstanding clarity MC is used to measure the overall severity of students' cognitive errors. The higher the value, the fewer the errors and the clearer the cognition.

[0108] Misunderstanding clarity is used to quantify the severity of students' cognitive errors. The specific calculation formula is as follows: MC=1-(Σseverity_i) / max_severity_budget; Among them, severity i The severity of the misunderstanding; Σseverity i This represents the sum of the severity of all currently active misunderstandings. max_severity_budget is the pre-set maximum severity budget (upper limit).

[0109] In this embodiment, the misunderstanding clarity (MC) is automatically constrained within the numerical range of [0,1]. Misunderstanding clarity is a score reflecting the learner's clarity of conceptual understanding. The maximum severity budget is a preset upper limit threshold used in the misunderstanding clarity calculation to normalize the severity of misunderstandings exhibited by the learner.

[0110] S250. Calculate the microtest score based on the microtest data in the interactive data.

[0111] In this embodiment, the micro-quiz score MQ is in discrete form, with a value set of {0, 0.5, 1.0}, corresponding to three states: incorrect answer, no participation in the micro-quiz, and correct answer, respectively. The micro-quiz data consists of structured test questions and their answers used to quickly assess learners' mastery of specific knowledge points. The micro-quiz score is a discrete numerical value quantified based on the learner's correct answers on the micro-quiz.

[0112] S260. Calculate the prompt efficiency based on the total number of prompts used and the preset maximum number of prompts.

[0113] In this embodiment, prompting efficiency is used to reflect students' ability to solve problems independently; the fewer prompts used, the higher the prompting efficiency. The formula for calculating prompting efficiency HU is as follows: HU=max(0,1-hints_used / max_hints); Wherein, HU represents prompt efficiency; hints_used is the number of hints that have been used; max_hints is the maximum number of hints allowed by default; max(0,…) is used to ensure that the result is not less than 0, so that the final output is in the range of 0 to 1 (that is, the output range is limited to [0,1] by truncating the maximum value).

[0114] In this embodiment, the evidentiary features include repetition score, misunderstanding clarity, microtest score, and prompting efficiency. That is, the aforementioned repetition score, misunderstanding clarity, microtest score, and prompting efficiency constitute the evidentiary features in this application.

[0115] In the above embodiments, the method can comprehensively collect features related to learning behavior and expression quality, and quantify them into diversified evidence indicators, thereby achieving refined and standardized calculation and fusion of learning evidence, and thus comprehensively reflecting the learner's true level of understanding and error cognition.

[0116] In some embodiments, step S300 may include: S310. Obtain the correct statement data corresponding to the target concept.

[0117] In this embodiment, this step uses standard concept definitions, formula rules, unit specifications, and variable mapping relationships from the knowledge base as the consistency verification benchmark data.

[0118] S320. Perform a consistency check on the interactive data based on the correctly stated data to obtain a set of contradictory results.

[0119] In this embodiment, consistency checks may include one or more of the following: equation consistency checks, unit consistency checks, or symbol mapping consistency checks. Specifically, the equation consistency type is used to identify inconsistencies in the mathematical expressions between learner statements and correct knowledge. The unit consistency type is used to identify inconsistencies in the use of physical units or dimensions in learner statements. The symbol mapping consistency type is used to identify inconsistencies in the meanings that learners assign to variable symbols that are inconsistent with the correct definitions.

[0120] S330. When the set of contradictory results includes results that are contradictory, calculate the contradiction penalty score based on the set of contradictory results; when the set of contradictory results only includes results that are not contradictory, determine the contradiction penalty score to be 0.

[0121] In this embodiment, the set of contradictory results includes results that are not contradictory and / or results that are contradictory. That is, the set of contradictory results can contain multiple results, which can all be non-contradictory, all be contradictory, or be partially contradictory and partially non-contradictory.

[0122] In this embodiment, contradictory results correspond to at least one contradiction type. Each contradiction type has a corresponding severity weight.

[0123] In this embodiment, the contradiction type includes one or more of the following: equation consistency type (i.e., whether the mathematical expressions are consistent), unit consistency type (i.e., whether the physical units are consistent), and symbol mapping consistency type (i.e., whether the variable symbols are mapped to the same quantity).

[0124] For example, the formula for calculating the penalty for a single contradiction can be as follows: g(C)=contradiction_detected×severity_weight; The formula for calculating multiple contradictions can be as follows: g(C)=min(1,Σseverity_weight_i / severity_budget); Where g(C) is the contradiction penalty score; contradiction_detected indicates whether a contradiction was detected (1 for yes, 0 for no). severity_weight represents the severity weight of a single conflict; ∑severity_weight_i is the sum of the severity weights of multiple conflicts; severity_budget is the preset total budget (maximum value) for severity. min(1,…) means limiting the final result to the range of 0 to 1.

[0125] In this embodiment, the severity and weight configurations corresponding to different conflict types are shown in Table 6.

[0126] Table 6. Severity and Weight Allocation for Different Types of Conflicts

[0127] In this embodiment, the method is also configured with a misunderstanding removal rule: when two consecutive rounds of error-free mode are met and the relevant concept is correct in at least one micro-test, the misunderstanding is removed; wherein, active misunderstandings that are not removed continue to block the unlocking of the complete solution.

[0128] In the above embodiments, the method can comprehensively identify various cognitive contradictions in user statements through multi-dimensional consistency checks, and reasonably calculate corresponding penalty scores based on the actual type and severity of the contradictions, thereby achieving objective correction of mastery by combining the user's actual cognitive deficiencies. In some embodiments, step S400 may include: S410. Weighted fusion of evidence features yields the total evidence score.

[0129] In this embodiment, the method can integrate four types of learning evidence using a multi-evidence weighted fusion function, and the fusion formula is as follows: f(E)=w1 EB+w2 MC+w3 MQ+w4 HU; Where f(E) is the total evidence score; EB stands for Standardized Recall Score; MC stands for misunderstanding clarity score; MQ is the penalty score for contradictions; HU represents the efficiency score. w1, w2, w3, and w4 correspond to the weighting coefficients of the four types of evidence, respectively.

[0130] In this embodiment, the system global parameters are preset with preferred values: w1=0.40, w2=0.25, w3=0.25, w4=0.10, and the sum is 1.0.

[0131] S420. Calculate the target mastery score based on the current mastery score, total evidence score, contradiction penalty score, preset learning rate, and preset penalty rate corresponding to the target concept.

[0132] In this embodiment, the method can comprehensively consider historical mastery decay values, positive learning evidence, and contradictory error penalties to iteratively update the mastery.

[0133] In this embodiment, the preset learning rate is a hyperparameter that controls the weight of positive evidence in the mastery score update. The preset penalty rate is a hyperparameter that controls the weight of contradictory penalty scores in the mastery score update.

[0134] The complete update formula is as follows: m_k^(t+1)=clip(m'_k+α·f(E)-β·g(C),0,1); Where m_k^(t+1) is the new mastery score of the target concept k after the (t+1)th round update; m'_k is the current mastery score after attenuation processing; α is the learning rate, which controls the extent to which positive evidence increases mastery. f(E) is the total evidence score calculated earlier; β is the penalty rate, which controls the extent to which contradictory errors reduce mastery. g(C) is the contradiction penalty score; clip(x,0,1) is a cutoff function that limits the result to between 0 and 1 to avoid exceeding a reasonable range.

[0135] In this embodiment, the learning rate α ranges from [0.10, 0.50], and is preferably 0.30; The penalty rate β ranges from [0.05, 0.40], with a preferred value of 0.20; Using the clip function, the final mastery is forced to be constrained within the normalized interval [0,1].

[0136] In the above embodiments, the method can comprehensively measure the user's understanding by weighted fusion of various learning evidences, and at the same time introduce contradictory penalty constraints to constrain the negative impact of cognitive errors, thereby making the updated mastery results more reasonable.

[0137] In some embodiments, step S500 may include: S510. Based on the preset misunderstanding removal conditions, update the misunderstanding status of the current active misunderstanding list maintained by the misunderstanding tracker module to obtain the target active misunderstanding list.

[0138] In this embodiment, the misunderstanding lifecycle includes four states: DETECTED, ACTIVE, ADDRESSED, and CLEARED. Different states correspond to different content unlocking and blocking rules and misunderstanding severity superposition rules.

[0139] S520. Determine the unlock status based on the target mastery score, the retelling score in the evidence characteristics, the target active misunderstanding list, and the preset composite unlock conditions.

[0140] In this embodiment, the composite unlocking condition is a gating logic composed of multiple sub-conditions, such as mastery level, quality of feedback, and no active misunderstandings.

[0141] In this embodiment, the method employs a joint determination rule that simultaneously satisfies multiple conditions. The formula for the composite unlocking condition is as follows: UNLOCK=(m_k≥τ)∧(EB≥0.5)∧(ActiveMisconceptions= ); UNLOCK indicates the unlocked state (true if the condition is met, false if the condition is not met); m_k is the current mastery score of the target concept k; τ is the preset mastery threshold; EB stands for Standardized Recall Score; ∧ represents the logical "AND", indicating that all conditions must be met simultaneously; ActiveMisconceptions is a list of active misunderstandings corresponding to the target concept; An empty set indicates that there are no active misunderstandings.

[0142] In this embodiment, the complete teaching content access can only be unlocked if all three conditions mentioned above are met simultaneously.

[0143] S530. Determine the retrieval strategy based on the target mastery score and unlock status; the retrieval strategy is one of the following: full access strategy, scaffold enhancement strategy, or scaffold only strategy.

[0144] In this embodiment, a full access strategy can be a retrieval strategy that allows the retrieval of all types of content from the knowledge base, including complete solutions and final answers. A scaffolding enhancement strategy can be a retrieval strategy that allows the retrieval of definitions, examples, hints, and partial solutions, but not complete solutions and final answers. A scaffolding-only strategy can be a retrieval strategy that only allows the retrieval of basic scaffolding content such as definitions, examples, and hints from the knowledge base.

[0145] In this embodiment, the method adopts the A+ coupling retrieval strategy rule, and the specific determination logic is as follows: If the unlock status is UNLOCKED, then a full access policy is adopted; If the mastery level m_k≥0.5 but the device is not unlocked, a scaffold enhancement strategy is adopted. If the mastery level m_k < 0.5 or the patient is in the S8 assessment blockage state, then a stent-only strategy is adopted.

[0146] Wherein, A+ is the coupling retrieval strategy identifier defined in this embodiment; UNLOCKED indicates that the device is unlocked. m_k is the current mastery score of the target concept k; FULL_ACCESS is the full access policy; SCAFFOLD_PLUS is a scaffold enhancement strategy; SCAFFOLD_ONLY is a bracket-only strategy; S8 is the preset assessment and blocking phase.

[0147] In the above embodiments, the method can combine real-time misunderstanding status, knowledge mastery level and user comprehension performance to jointly determine content unlocking permissions, and adaptively match differentiated knowledge base retrieval strategies based on comprehensive evaluation results, thereby controlling the scope of teaching content push from the source, avoiding premature output of complete answer content, and thus adapting to the learning needs of different cognitive foundations.

[0148] In some embodiments, step S600 may include: S610. Search the preset knowledge base according to the search strategy to obtain the search data.

[0149] In this embodiment, the knowledge base is pre-divided into six categories of tagged content blocks. During retrieval, inaccessible content types are filtered according to the retrieval strategy. The specific classification of content blocks is shown in Table 7 below.

[0150] Table 7. Detailed Classification of Knowledge Base Content Blocks

[0151] S620. Generate a response generation instruction based on the retrieval strategy and unlock status constraints.

[0152] In this embodiment, the generation instruction will embed constraint parameters such as step budget, answer masking, and content block whitelist to limit the generation range and output format of the large model.

[0153] S630. Based on the retrieved data and the response generation instructions, call the large language model to generate the original response data.

[0154] In this embodiment, the method can generate original response content that matches the student's cognitive level based on the filtered compliant knowledge base content and the current dialogue state.

[0155] In the above embodiments, the method can accurately filter knowledge base content according to a predetermined retrieval strategy, and at the same time reasonably constrain the model generation logic by combining the unlock status, so that the large language model outputs original response content that is adapted to the user's current cognitive level.

[0156] In some embodiments, this method can be deployed as a bilingual STEM tutoring assistant for learners of calculus and linear algebra, adapting to the learning scenarios of university students.

[0157] This method relies on a domain-specific concept library, misunderstanding templates, and micro-test rules to support various engineering problem scenarios, including electrical engineering, mechanical engineering, and computer engineering. Furthermore, the core mastery gating logic remains consistent across disciplines.

[0158] For example, the customized operating parameter configuration in this embodiment is shown in Table 8.

[0159] Table 8 Customized Operating Parameter Configuration Table

[0160] Specifically, this embodiment takes the learning and interaction process of the knowledge point of integration by parts as an example to illustrate the complete quantitative calculation process for successfully unlocking the content. The key parameters and specific calculation process are shown in Table 9.

[0161] Table 9. Examples of Specific Parameters and Calculation Processes

[0162] Meanwhile, this embodiment takes a learning interaction scenario with persistent cognitive misunderstandings as an example to specifically demonstrate the system's unlocking, interception, and control logic, and intuitively explain the constraint and regulation effect of the cognitive misunderstanding blocking mechanism on the opening of learning content. Specific experiments and effect examples are shown in Table 10.

[0163] Table 10 Cases of Content Unlocking Rejection under the Misunderstanding Blocking Mechanism

[0164] In some embodiments, the method can be configured as a high-constraint standard exam preparation mode.

[0165] This method can improve the knowledge mastery requirements and increase the punishment for errors by adjusting the core control parameters, so as to adapt to the pre-exam intensive and standardized learning scenario.

[0166] In this embodiment, the parameter comparison between the standard mode and the exam preparation mode is shown in Table 11.

[0167] Table 11 Comparison of parameters between the standard model and the exam preparation model

[0168] It is evident that, under this model, the method will raise the content unlocking threshold, reduce the gain effect of positive learning evidence, enhance the punishment for cognitive contradictions and conceptual errors, and tighten the misunderstanding severity budget, thereby regulating learning logic and answer rigor from multiple dimensions to meet the training requirements of high-intensity test preparation.

[0169] In some embodiments, the method can be configured with a relaxed beginner mode for learners with weak foundations.

[0170] This method can lower the evaluation threshold, weaken the punishment for mistakes, and strengthen positive incentives to match the cognitive patterns of learners in the introductory stage, thereby achieving smooth and gradual personalized teaching.

[0171] In this embodiment, the parameter comparison between standard mode and beginner mode is shown in Table 12.

[0172] Table 12 Comparison of parameters between Standard Mode and Beginner Mode

[0173] It is evident that this model appropriately relaxes the conditions for unlocking content, strengthens the positive incentives for effective learning behaviors, reduces the punishment for minor cognitive biases and expression flaws, and reasonably relaxes the control of misunderstandings, thereby effectively protecting the learning enthusiasm of beginners and making it more suitable for teaching scenarios for beginners with zero experience.

[0174] In some embodiments, the method can integrate a subject adaptation module to achieve cross-scenario deployment and adaptation for multiple engineering disciplines. The system's core algorithms, mastery assessment mechanisms, and content unlocking logic are universal, requiring only the replacement of domain-specific resources to quickly adapt to the teaching needs of different engineering majors.

[0175] The specific adaptation content for each engineering discipline is shown in Table 13.

[0176] Table 13 Specific Adaptation Contents for Each Engineering Discipline

[0177] Each subject adapter comes with four types of exclusive compatible content: First, a domain-specific concept classification system corresponding to the mastery degree vector; Second, a database of typical misunderstandings regarding severity weighting; Third, mini-quiz assessment templates that are tailored to the subject knowledge points; Fourth, content block retrieval and filtering rules that meet the requirements of subject teaching.

[0178] Based on this, the method can be rapidly deployed and applied across multiple disciplines by relying on modular adaptive design, thus making the method applicable to a wider range of scenarios.

[0179] like Figure 2 As shown, some embodiments of this application provide a structural schematic diagram of a learning interaction device based on knowledge mastery. It should be understood that this device is related to... Figure 1 The method executed in the middle corresponds to the steps involved in the aforementioned method. The specific functions and effects of the device can be found in the description above. To avoid repetition, detailed descriptions are omitted here.

[0180] The knowledge-based interactive learning device includes: The receiving unit 1310 is used to receive interactive data related to the target concept input by the user; Extraction unit 1320 is used to extract evidence features based on interactive data; The calculation unit 1330 is used to calculate the contradiction penalty on the interactive data and obtain the contradiction penalty score; The mastery update unit 1340 is used to update the current mastery score corresponding to the target concept based on the evidence features and the contradiction penalty score, so as to obtain the target mastery score; The determination unit 1350 is used to determine the retrieval strategy and unlocking status based on the target mastery score, evidence features, and the current active misunderstanding list maintained by the misunderstanding tracker module. The generation unit 1360 is used to generate raw response data based on the retrieval strategy, unlock status, and preset knowledge base; The output unit 1370 is used to output the corresponding final response data based on the original response data after the original response data passes the leakage protection detection.

[0181] In some embodiments, after the receiving unit 1310 receives interactive data related to the target concept input by the user, the knowledge mastery-based learning interaction device further includes: The intent detection unit 1380 is used to perform intent detection on the interaction data and obtain the intent detection result. The judgment unit 1390 is used to determine whether the intent detection result includes the evaluation intent; The update unit 1400 is used to update the current global session state based on the interaction data to obtain the target global session state when the judgment unit 1390 determines that the intent detection result does not include the evaluation intent. The judgment unit 1390 is also used to determine whether the target global session state is a talkback state; The attenuation calculation unit 1410 is used to perform time-based attenuation calculation on the current mastery score corresponding to the target concept when the judgment unit 1390 determines that the target global session state is a talk state, so as to obtain the attenuated current mastery score.

[0182] In some embodiments, the extraction unit 1320 includes: Extraction subunit 1321 is used to extract structured features based on user feedback content in the interaction data; wherein, the structured features include coverage, accuracy and coherence; The first acquisition subunit 1322 is used to acquire the list of currently active misunderstandings maintained by the misunderstanding tracker module and the total number of times the prompts have been used for the target concept; The first calculation subunit 1323 is used to calculate the initial evaluation score based on the structured features, and to perform threshold-based discretization processing on the initial evaluation score to obtain the review score. The first calculation subunit 1323 is also used to calculate the misunderstanding clarity based on the sum of the severity of each misunderstanding in the current active misunderstanding list and the preset maximum severity budget; The first calculation subunit 1323 is also used to calculate the micro-test score based on the micro-test data in the interactive data; The first calculation subunit 1323 is also used to calculate the prompt efficiency based on the total number of prompts used and the preset maximum number of prompts; Among the evidentiary features are replay score, misunderstanding clarity, microtest score, and prompting efficiency.

[0183] In some embodiments, the computing unit 1330 includes: The second acquisition subunit 1331 is used to acquire the correct statement data corresponding to the target concept; Check subunit 1332 is used to perform consistency checks on the interactive data based on the correct statement data to obtain a set of contradictory results; The second calculation subunit 1333 is used to calculate the contradiction penalty score based on the contradiction result set when the contradiction result set includes contradictory results; and to determine the contradiction penalty score as 0 when the contradiction result set only includes non-contradictory results. Consistency checks include equation consistency checks, unit consistency checks, and symbol mapping consistency checks. The set of contradictory results includes both consistent results and / or contradictory results; The contradictory results include at least one contradiction type and a severity weight corresponding to each contradiction type; Contradictory types include one or more of the following: equation consistency type, unity consistency type, and symbolic mapping consistency type.

[0184] In some embodiments, the mastery update unit 1340 includes: The fusion subunit 1341 is used to weight and fuse the evidence features to obtain the total evidence score; The third calculation subunit 1342 is used to calculate the target mastery score based on the current mastery score, total evidence score, contradiction penalty score, preset learning rate, and preset penalty rate corresponding to the target concept.

[0185] In some embodiments, the determining unit 1350 includes: The update subunit 1351 is used to update the misunderstanding status of the current active misunderstanding list maintained by the misunderstanding tracker module according to the preset misunderstanding clearing conditions, so as to obtain the target active misunderstanding list. Subunit 1352 is used to determine the unlocking status based on the target mastery score, the replay score in the evidence characteristics, the target active misunderstanding list, and the preset composite unlocking conditions. Subunit 1352 is also used to determine the retrieval strategy based on the target mastery score and unlock status; The retrieval strategy is one of the following: full access strategy, scaffold enhancement strategy, or scaffold only strategy.

[0186] In some embodiments, the generation unit 1360 includes: The retrieval subunit 1361 is used to perform a retrieval in a preset knowledge base according to a retrieval strategy to obtain retrieval data; The generation subunit 1362 is used to generate a response generation instruction based on the retrieval strategy and unlocking state constraints; Subunit 1363 is invoked to generate raw response data by calling the large language model based on the retrieved data and response generation instructions.

[0187] In some embodiments, after the generation unit 1360 generates the original response data according to the retrieval strategy, unlock status, and preset knowledge base, the knowledge mastery-based learning interaction device further includes: The leakage detection unit 1420 is used to detect the original response data according to the preset leakage protection rules and obtain the leakage protection detection result; wherein, the leakage protection detection result is used to indicate whether a leakage type is detected, and the leakage type includes at least one of step count exceeding the limit, final answer marking, numerical equation matching and answer keyword matching; The output unit 1370 is also configured to determine that the original response data has passed the leakage protection detection when the leakage detection unit 1420 detects that no leakage type has been detected, and to perform the step of outputting the corresponding final response data based on the original response data.

[0188] In some embodiments, the method further includes: The acquisition unit 1430 is used to acquire the current unlocking status when the leakage protection detection result indicates that at least one type of leakage has been detected; The rewrite unit 1440 is used to rewrite the original response data to obtain security response data when the current unlock state obtained by the acquisition unit 1430 is not a fully unlocked state. The output unit 1370 is also used to output the corresponding final response data based on the security response data; The current unlock status is one of the following: fully unlocked, locked, or forced locked.

[0189] like Figure 3 As shown, this application provides an electronic device 1500, which includes a processor 1501 and a memory 1502. The processor 1501 and the memory 1502 are interconnected and communicate with each other through a communication bus 1503 and / or other forms of connection mechanism (not shown). The memory 1502 stores a computer program that can be executed by the processor 1501. When the computing device is running, the processor 1501 executes the computer program to perform the method in any of the aforementioned optional implementations.

[0190] This application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the method in any of the aforementioned optional implementations.

[0191] The computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0192] This application provides a computer program product, which includes a computer program that, when run by a processor, executes the method in any of the aforementioned optional implementations.

[0193] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application. In particular, as long as there is no conflict, the various technical features mentioned in the embodiments can be combined in any way. This application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A learning interaction method based on knowledge mastery, characterized in that, include: Receive interactive data from the user related to the target concept; Evidence features are extracted from the interaction data, and a contradiction penalty calculation is performed on the interaction data to obtain a contradiction penalty score; The current mastery score corresponding to the target concept is updated based on the evidence features and the contradiction penalty score to obtain the target mastery score; Based on the target mastery score, the evidence features, and the current active misunderstanding list maintained by the misunderstanding tracker module, the retrieval strategy and unlocking status are determined; Based on the retrieval strategy, the unlock status, and the preset knowledge base, generate the original response data; After the original response data passes the leakage protection detection, the corresponding final response data is output based on the original response data.

2. The learning interaction method based on knowledge mastery according to claim 1, characterized in that, After receiving the user-input interactive data related to the target concept, the method further includes: The interaction data is subjected to intent detection to obtain intent detection results; Determine whether the intent detection result includes an evaluation intent; If the evaluation intent is not included, the current global session state is updated based on the interaction data to obtain the target global session state; Determine whether the target global session state is in a talkback state; If it is a review session, the current mastery score corresponding to the target concept is calculated based on time decay to obtain the decayed current mastery score.

3. The learning interaction method based on knowledge mastery according to claim 1, characterized in that, The step of extracting evidence features based on the interaction data includes: Structured features are extracted from the user feedback content in the interaction data; wherein, the structured features include coverage, accuracy, and coherence. Obtain the list of currently active misunderstandings maintained by the misunderstanding tracker module and the total number of times prompts have been used for the target concept; The initial evaluation score is calculated based on the structured features, and the initial evaluation score is discretized based on a threshold to obtain the review score. The misunderstanding clarity is calculated based on the sum of the severity of each misunderstanding in the current active misunderstanding list and the preset maximum severity budget. Calculate the micro-test score based on the micro-test data in the interactive data; The prompt efficiency is calculated based on the total number of prompts used and the preset maximum number of prompts. The evidentiary features include the repetition score, the misunderstanding clarity, the microtest score, and the prompting efficiency.

4. The learning interaction method based on knowledge mastery according to claim 1, characterized in that, The step of calculating the contradiction penalty on the interactive data to obtain a contradiction penalty score includes: Obtain the correct statement data corresponding to the target concept; A consistency check is performed on the interaction data based on the correct statement data to obtain a set of contradictory results. When the set of contradictory results includes contradictory results, a contradiction penalty score is calculated based on the set of contradictory results. When the set of contradictory results only includes results that do not contain contradictions, the contradiction penalty score is determined to be 0; The consistency checks include equation consistency checks, unit consistency checks, and symbol mapping consistency checks. The set of contradictory results includes results that are not contradictory and / or results that are contradictory; The contradictory results include at least one contradiction type and a severity weight corresponding to each contradiction type; The contradiction types include one or more of the following: equation consistency type, unity consistency type, and symbolic mapping consistency type.

5. The learning interaction method based on knowledge mastery according to claim 1, characterized in that, The step of updating the current mastery score corresponding to the target concept based on the evidence features and the contradiction penalty score to obtain the target mastery score includes: The evidence features are weighted and fused to obtain the total evidence score; The target mastery score is calculated based on the current mastery score, the total evidence score, the contradiction penalty score, the preset learning rate, and the preset penalty rate corresponding to the target concept.

6. The learning interaction method based on knowledge mastery according to claim 1, characterized in that, The step of determining the retrieval strategy and unlocking status based on the target mastery score, the evidence features, and the currently active misunderstanding list maintained by the misunderstanding tracker module includes: Based on the preset misunderstanding removal conditions, the misunderstanding status of the currently active misunderstanding list maintained by the misunderstanding tracker module is updated to obtain the target active misunderstanding list; The unlocking status is determined based on the target mastery score, the retelling score in the evidence features, the target active misunderstanding list, and the preset composite unlocking conditions; Based on the target mastery score and the unlock status, determine the retrieval strategy; The retrieval strategy is one of the following: full access strategy, scaffold enhancement strategy, and scaffold only strategy.

7. The learning interaction method based on knowledge mastery according to claim 1, characterized in that, The step of generating raw response data based on the retrieval strategy, the unlock status, and the preset knowledge base includes: The search is performed in a preset knowledge base according to the search strategy to obtain the search data; Generate a response generation instruction based on the retrieval strategy and the unlock status constraint; Based on the retrieved data and the response generation instruction, the large language model is invoked to generate the original response data.

8. The learning interaction method based on knowledge mastery according to claim 1, characterized in that, After generating the original response data based on the retrieval strategy, the unlock status, and the preset knowledge base, the method further includes: The original response data is detected according to preset leakage protection rules to obtain leakage protection detection results; wherein, the leakage protection detection results are used to indicate whether a leakage type is detected, and the leakage type includes at least one of step count exceeding the limit, final answer marking, numerical equation matching, and answer keyword matching; When the leakage protection detection result indicates that no leakage type was detected, the original response data is determined to have passed the leakage protection detection, and the step of outputting the corresponding final response data based on the original response data is executed.

9. The learning interaction method based on knowledge mastery according to claim 8, characterized in that, The method further includes: When the leakage protection detection result indicates that at least one type of leakage has been detected, obtain the current unlock status; When the current unlock state is not a fully unlocked state, the original response data is rewritten to obtain security response data; Output the corresponding final response data based on the security response data; The current unlock state can be one of the following: fully unlocked, locked, or forced locked.

10. A learning interactive device based on knowledge mastery, characterized in that, The knowledge-based learning interaction device includes: The receiving unit is used to receive interactive data related to the target concept input by the user; Extraction unit, used to extract evidence features based on the interaction data; The calculation unit is used to calculate the contradiction penalty on the interactive data to obtain the contradiction penalty score; The mastery update unit is used to update the current mastery score corresponding to the target concept according to the evidence features and the contradiction penalty score, so as to obtain the target mastery score; The determining unit is used to determine the retrieval strategy and unlocking status based on the target mastery score, the evidence features, and the currently active misunderstanding list maintained by the misunderstanding tracker module. The generation unit is used to generate raw response data based on the retrieval strategy, the unlocking status, and a preset knowledge base; The output unit is used to output the corresponding final response data based on the original response data after the original response data passes the leakage protection detection.

Images