A self-adaptive tutoring method and system based on support feasible region constraint

Abstract

The application belongs to the technical field of intelligent education, and specifically discloses a self-adaptive learning guiding method and system based on support feasible domain constraint, which comprises the following steps: obtaining problem analysis data based on the natural language question answered by a learner; obtaining the learner state based on the problem analysis data and the answer of the learner to the natural language question, wherein the learner state comprises a concept mastery state, a process promotion state and an expression state which are independent of each other; generating a learning guiding decision according to a preset control rule based on the learner state, wherein the learning guiding decision comprises a candidate support feasible domain, and the candidate support feasible domain is a set composed of at least two numbered continuous support levels; inputting the learning guiding decision, the answer of the learner to the natural language question and the learner state into a large language model which has been trained to obtain a selected support level in the form of structured data output by the large language model and a learning guiding speech, wherein the selected support level is within the candidate support feasible domain, and the learning guiding speech is output to the learner.

Description

Technical Field

[0001] This application belongs to the field of intelligent education technology, and more specifically, relates to an adaptive guided learning method and system based on scaffold feasible region constraints. Background Technology

[0002] Scaffolded instruction, a guided teaching method derived from Vygotsky's zone of proximal development theory, centers on the idea that teachers provide adjustable cognitive support based on learners' current cognitive levels, rather than directly providing answers. With the deep integration of generative artificial intelligence technology and intelligent education, implementing this teaching method as a computable algorithm, enabling intelligent learning guidance systems to adaptively adjust the intensity of guidance according to learners' current states, is an important research direction in the field of intelligent education.

[0003] Existing related technologies suffer from the following types of solutions and their shortcomings, which limit the controllability, interpretability, and strategy flexibility of adaptive learning: The first type is a learning service solution based on ontology description language and rule reasoning. This type of solution recommends resources to learners through ontology modeling and rule matching. Although it has strong interpretability, the ontology maintenance cost is high, and the ability to conduct open dialogue and multi-turn interaction in natural language is weak, making it difficult to adapt to problem-solving scenarios that require fine-grained multi-turn interaction.

[0004] The second type is a retrieval-based guided learning scheme based on multi-level keyword matching. This type of scheme only completes static content retrieval and does not have the ability to dynamically perceive and guide the learner's current cognitive state. It is difficult to meet the core requirement of scaffolded instruction to dynamically adjust the intensity of cognitive support.

[0005] The third type is a learning approach based on hierarchical language models. This approach uses a higher-level language model to generate a single deterministic control instruction based on a fixed multi-stage state machine, containing fields such as target stage, action type, skill focus, emotional strategy, and constraints. A lower-level language model then generates a natural language response based on this instruction. This type of approach also typically introduces an emotional vector composed of several emotional dimensions and a single scalar skill mastery score weighted by immediate performance score and historical average score. However, this type of approach has the following technical problems: First, the upper-level model outputs a unique and definite control instruction in each round, and the teaching strategy space is compressed into a hard switch between several fixed levels (such as scaffolding, cue prompts, and direct challenge). The lower-level model can only passively follow the instructions to generate discourse and cannot make fine adjustments between several adjacent scaffolding intensities according to the specific context. This single-point deterministic output does not match the core idea of ​​the zone of proximal development theory that "scaffolding intensity should be dynamically adjusted within the adjacent interval".

[0006] Secondly, this type of solution compresses the learner's state into a single scalar of skill mastery score and affective vector. When learners exhibit typical heterogeneous states such as "having mastered the concept but having difficulty expressing it" or "expressing it fluently but having confusion about the concept," the system is unable to make differentiated responses.

[0007] Third, the lower-level model only undertakes the single task of text generation and cannot learn the strategic skill of "choosing the appropriate scaffold strength under different learner states," which is directly related to teaching effectiveness. The strategy selection is entirely determined by the upper-level model through threshold switching, reducing the lower-level model to a mere text expresser and failing to achieve synergistic optimization of strategy and expression.

[0008] Fourth, offline training samples are usually directly mapped from the original questions to single-round "instruction-heuristic probing" pairs, which means that the lower-level model only learns the single-round questioning paradigm, lacks experience in strategy selection in a complete dialogue context, and is difficult to generalize to multi-round real interactions. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this application aims to provide an adaptive learning method and system based on scaffold feasible region constraints. This aims to solve the problems of low flexibility in existing adaptive learning methods, which are caused by the teaching strategy space being compressed into hard switching between several fixed levels, the learner state being compressed into a single scalar, and the lower-level model only undertaking the single task of text generation.

[0010] To achieve the above objectives, in a first aspect, this application provides an adaptive guided learning method based on scaffold feasible region constraints, comprising: Based on learners' responses to natural language questions, obtain question analysis data; Based on the learner's answer to the natural language question, the learner's state is obtained, which includes mutually independent concept mastery state, process progress state, and expression state; Based on the learner's state, a guidance decision is generated according to a preset control rule. The guidance decision includes a candidate scaffold feasible region, which is a set consisting of at least two consecutively numbered scaffold levels. The learning guidance decision, the learner's answer to the natural language question, and the learner's state are input into the trained large language model to obtain the selected scaffold level and learning guidance discourse in the form of structured data output by the large language model. The selected scaffold level is within the feasible domain of the candidate scaffolds, and the learning guidance discourse is output to the learner.

[0011] This application models learner states as independent concept mastery, process progression, and expression states, enabling the system to accurately perceive learners' heterogeneous cognitive states along three independent cognitive dimensions. This avoids the state confusion problem caused by compressing multidimensional cognitive states into a single scalar, and allows for differentiated responses to typical heterogeneous states such as "concept mastered but difficult to express" or "fluent expression but conceptual confusion." By generating a candidate scaffolding feasible domain consisting of at least two consecutively numbered scaffolding levels based on learner states, the teaching strategy output is expanded from the single-scaffolding level hard switching in existing solutions to a continuous strategy feasible domain, enabling the guided learning process to... The scaffolding strength during the process can be dynamically adjusted within the adjacent interval, achieving the teaching objective of "dynamically adjusting the scaffolding strength within the zone of proximal development," significantly improving the flexibility of teaching strategies. By having the trained large language model simultaneously output the selected scaffolding level and the guiding discourse in the same inference process, and with the selected scaffolding level constrained within the feasible domain of the candidate scaffolding, the large language model has the ability to select strategies while generating guiding discourse. This avoids the architectural defects of existing solutions where the discourse generation module only undertakes the task of text expression and the strategy selection is completely determined by external rigidity. It achieves the synergistic optimization of strategy selection and discourse generation, improving the flexibility of adaptive guiding.

[0012] According to the adaptive guided learning method based on the feasible region constraint of the scaffold provided in this application, the set of elements of the feasible region of the candidate scaffold is taken from a six-level scaffold level set. The six-level scaffold level set includes, in order, open questioning level, directed questioning level, evidence element prompting level, principle prompting level, local correction level, and explicit explanation plus restatement level. The information increment carried by adjacent levels satisfies the condition that the higher the level number, the greater the information increment.

[0013] This application provides a fine-grained and semantically clear scaffolding intensity grading system for scaffolded instruction by taking the elements of the feasible domain of candidate scaffolds from a six-level scaffolding hierarchy that includes open-ended probing, directed probing, evidence element prompts, principle prompts, local correction, and explicit explanation plus repetition. Furthermore, the information increment carried by adjacent levels satisfies the monotonicity that the higher the level number, the greater the information increment. This grading system covers a complete scaffolding intensity spectrum from the lightest guidance (open-ended probing) to the most intense support (explicit explanation plus repetition), enabling the guidance system to accurately match the appropriate intensity of cognitive support within the learner's zone of proximal development. The monotonicity of information increment between levels ensures that the level selection within the feasible domain has interpretable pedagogical semantics, facilitating the review and understanding of guidance strategies by teachers and curriculum researchers.

[0014] According to the adaptive learning method based on scaffold feasible region constraints provided in this application, the preset control rules include state mapping rules, misunderstanding priority correction rules, expression lower limit protection rules, and continuity constraint rules.

[0015] This application provides a complete and auditable hybrid control mechanism for generating the feasible domain of candidate stents by specifically defining the preset control rules as a four-level cascaded rule: state mapping rule, misunderstanding priority correction rule, expression lower limit protection rule, and continuity constraint rule.

[0016] The state mapping rule determines the initial interval of the feasible region based on the learner's conceptual mastery state and process progression state, achieving a reasonable mapping between scaffolding strength and the learner's cognitive level. The misunderstanding-priority correction rule covers the feasible region with a set of high-scaffolding levels when a specific misunderstanding is detected, ensuring that the system can prioritize providing high-intensity corrective support when misunderstandings are discovered, promptly correcting learners' cognitive biases. The expression lower limit protection rule removes low-numbered scaffolding levels when learners have expression difficulties, avoiding providing learners with insufficient open-ended guidance that would increase their cognitive burden. The continuity constraint rule ensures that the feasible region is always a continuous interval rather than a discrete set of points, guaranteeing the smoothness of scaffolding level selection within the feasible region and the gradualness of teaching strategy adjustments. The cascading application mechanism of the four-level rules provides clear rule basis for each round of guidance decisions, making the entire guidance decision chain an interpretable and auditable teaching process log.

[0017] According to the adaptive guided learning method based on scaffold feasible region constraints provided in this application, the training process of the large language model includes: A training dataset is constructed at the granularity of the learning unit. Each learning unit contains at least one original question, target concept, core knowledge point set, standard reasoning step chain, candidate misunderstanding type set, and scaffolding prompt chain. The scaffolding prompt chain consists of reference scripts corresponding to multiple scaffolding levels. Perform a complete multi-turn dialogue simulation for each learning unit and generate a dialogue log containing the context of the multi-turn dialogue. The dialogue logs are flattened and extracted according to the dialogue rounds to obtain training samples. The input fields of each training sample include the candidate scaffold feasible domain, target reasoning steps, teaching action type, reference dialogue, learner's current round of speech, learner status and historical dialogue summary. The output field is structured data that contains both the selected scaffold level and the guiding speech as keys. Based on the training samples, a low-rank parameter fitting method is used to jointly train the base large language model.

[0018] This application constructs a complete training dataset at the granularity of the learning unit, including the original question, target concept, core knowledge point set, standard reasoning step chain, candidate misunderstanding type set, and scaffolding prompt chain. This provides a structured and systematic training data foundation for the training of the large language model, enabling the training data to naturally carry prior knowledge in the teaching field. By simulating complete multi-turn dialogues for each learning unit and extracting the data in a flattened manner according to the dialogue turn, each training sample carries the contextual information of complete multi-turn dialogues. This avoids the problem of missing multi-turn context caused by training samples being only single-turn "instruction-heuristic probing" pairs in existing solutions, effectively improving the generalization ability of the large language model to multi-turn real interaction scenarios. By using the selected scaffolding level and the learning discourse as joint output fields, and employing a low-rank parameter adaptation method to jointly train the base large language model, the model learns both the task of "choosing which scaffolding level within the feasible domain" and "how to generate the corresponding discourse" simultaneously. The supervision signals for both tasks are provided in the same forward propagation, and collaborative optimization is achieved through shared underlying semantic representations. This enables the trained large language model to have the dual capabilities of policy selection and discourse generation, rather than just undertaking the single task of text generation.

[0019] According to the adaptive guided learning method based on scaffold feasible region constraints provided in this application, the method further includes: If the selected scaffold level output by the large language model is not within the feasible region of the candidate scaffolds, a rewriting mechanism is triggered and the output is overwritten with the closest scaffold level within the feasible region of the candidate scaffolds.

[0020] This application provides a post-verification guarantee during the inference phase for the hard constraints of the feasible region by triggering a rewriting mechanism when the selected scaffold level output by the large language model is not within the feasible region of the candidate scaffolds, and then overwriting the output with the closest scaffold level within the feasible region of the candidate scaffolds. This rewriting mechanism ensures that regardless of the generation process of the large language model, the final output scaffold level always falls within the policy feasible region determined by the hybrid control rules, eliminating the risk of scaffold strength exceeding the limit from an engineering perspective, and ensuring the controllability and reliability of the learning guidance decision. This mechanism, together with the feasible region consistency constraint during the training phase, forms a double insurance, jointly ensuring the strict satisfaction of the hard constraints of the feasible region.

[0021] According to the adaptive tutoring method based on scaffold feasible region constraints provided in this application, after outputting the tutoring discourse to the learner, the method further includes: Receive the learner's response, update the learner's status based on the response, and output a process jump flag.

[0022] This application constructs a complete closed-loop control process of "diagnosis-decision-generation-evaluation-update" by receiving learners' responses after the guidance discourse is output to them, updating the learners' status, and outputting process jump markers. This closed-loop mechanism enables the system to dynamically update its cognitive state perception based on the learners' actual responses to the guidance discourse, achieving round-by-round adaptive adjustment of scaffolding strength, rather than a one-time static output. The introduction of process jump markers allows the system to judge the guidance progress based on the learners' real-time cognitive state, deciding whether to continue guidance on the current problem, enter the summary stage, enter the transfer problem, or end the guidance, achieving autonomous process management of the guidance process and improving the intelligence and adaptability of the guidance system.

[0023] According to the adaptive guided learning method based on scaffold feasible region constraints provided in this application, the process jump indicator is determined comprehensively based on the current mastery level, concept mastery status, process progress status, number of consecutive rounds without progress, and cumulative number of rounds.

[0024] This application provides a multi-dimensional and refined set of rules for decision-making in the learning process by specifically defining the criteria for process jump indicators as a comprehensive assessment of current mastery level, concept mastery status, process progress status, number of consecutive rounds without progress, and cumulative rounds. The joint assessment of current mastery level and cognitive status ensures that summarization or transfer is triggered only when learners have truly achieved a high level of mastery, avoiding erroneous jumps caused by relying on a single dimension. The introduction of consecutive rounds without progress enables automatic detection of learners' learning bottlenecks. When learners fail to make progress for several consecutive rounds, the system can automatically jump to the summary branch and provide fallback support, avoiding the waste of learners' time and cognitive resources in ineffective interactions. The upper limit on the cumulative rounds provides a hard termination boundary for the learning process, ensuring that the learning process does not continue indefinitely and guaranteeing the system's practicality and user experience.

[0025] Secondly, this application provides an adaptive guided learning system based on scaffold feasible region constraints, comprising: The first acquisition module is used to obtain question analysis data based on the learner's answers to natural language questions; The second acquisition module is used to acquire learner status based on the question analysis data and learner's answers to the natural language question. The learner status includes mutually independent concept mastery status, process progress status and expression status. The generation module is used to generate a guidance decision based on the learner's state and according to preset control rules. The guidance decision includes a candidate scaffold feasible region, which is a set consisting of at least two consecutively numbered scaffold levels. The guidance module is used to input the guidance decision, the learner's answer to the natural language question, and the learner's state into the trained large language model to obtain the selected scaffold level and guidance discourse in the structured data form of the large language model output. The selected scaffold level is within the feasible domain of the candidate scaffolds, and the guidance discourse is output to the learner.

[0026] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the adaptive guided learning method based on scaffold feasible region constraints as described in the first aspect or any possible implementation thereof.

[0027] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the adaptive guided learning method based on scaffold feasible region constraints as described in the first aspect or any possible implementation of the first aspect.

[0028] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to execute the adaptive guided learning method based on scaffold feasible region constraints described in the first aspect or any possible implementation of the first aspect.

[0029] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0030] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: (1) By modeling learner states as independent concept mastery states, process progress states, and expression states, the system can accurately perceive learners’ heterogeneous cognitive states along three independent cognitive dimensions, avoiding the state confusion problem caused by compressing multidimensional cognitive states into a single scalar, and can make differentiated responses to typical heterogeneous states such as “concept mastered but difficult to express” or “fluent expression but confused concepts”.

[0031] (2) By generating a candidate scaffolding feasible domain consisting of at least two consecutively numbered scaffolding levels based on learner status, the teaching strategy output is extended from the single scaffolding level hard switching in the existing scheme to a continuous strategy feasible domain, so that the scaffolding strength in the learning process can be dynamically adjusted in the adjacent interval, achieving the teaching goal of "dynamically adjusting the scaffolding strength in the zone of proximal development", and significantly improving the flexibility of the teaching strategy.

[0032] (3) By having the large language model trained to output the selected scaffold level and the guidance discourse in the same reasoning process, and the selected scaffold level is constrained within the feasible range of the candidate scaffold, the large language model has the ability to select strategies while generating guidance discourse. This avoids the architectural defects of the existing scheme where the discourse generation module only undertakes the task of text expression and the strategy selection is completely determined by external hardness, and realizes the synergistic optimization of strategy selection and discourse generation.

[0033] (4) By constructing a complete training dataset at the granularity of the learning unit, including the original question, target concept, core knowledge point set, standard reasoning step chain, candidate misunderstanding type set and scaffolding prompt chain, a structured and systematic training data foundation is provided for the training of the large language model, so that the training data naturally carries the prior knowledge of the teaching field; by performing a complete multi-round dialogue simulation for each learning unit and extracting it in a flattened manner according to the dialogue round, each training sample carries the context information of the complete multi-round dialogue, avoiding the problem of missing multi-round context caused by the training sample being only a single round of "instruction-heuristic follow-up question" in the existing scheme, and effectively improving the generalization ability of the large language model to multi-round real interaction scenarios.

[0034] By combining the above-mentioned beneficial effects, the problem of low flexibility in existing adaptive learning methods is solved, which is caused by the compression of the teaching strategy space into hard switching between several fixed levels, the compression of learner state into a single scalar, and the lower-level model only undertaking the single task of text generation. The technical effect of "improving the flexibility of adaptive learning" is achieved. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating the adaptive guided learning method based on scaffold feasible region constraints provided in an embodiment of this application. Figure 2 This is a schematic diagram of the cascading of hybrid control rules provided in an embodiment of this application; Figure 3 This is a flowchart of the dual-task joint training provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the adaptive learning device based on the feasible domain constraint of the support provided in the embodiments of this application; Figure 5This is a schematic diagram of the structure of the adaptive learning system based on scaffold feasible region constraints provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0038] In this article, the term "and / or" describes 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, or B existing alone. The symbol " / " in this article indicates that the related objects are in an "or" relationship; for example, A / B means A or B.

[0039] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0040] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0041] Next, combined Figures 1-4 The adaptive guided learning method based on scaffold feasible region constraints provided in the embodiments of this application is introduced.

[0042] Figure 1 This is a flowchart illustrating the adaptive guided learning method based on scaffold feasible region constraints provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps: Step S1: Obtain question analysis data based on the learner's answers to natural language questions; Step S2: Based on the question analysis data and learners' answers to natural language questions, obtain the learner status, which includes independent concept mastery status, process progress status, and expression status. Step S3: Based on the learner's state, generate a guidance decision according to preset control rules. The guidance decision includes the feasible region of candidate scaffolds, which is a set consisting of at least two consecutively numbered scaffold levels. Step S4 involves inputting the guidance decision, learner's answers to natural language questions, and learner status into the trained large language model to obtain the selected scaffold level and guidance discourse in the structured data form output by the large language model. The selected scaffold level is within the feasible range of candidate scaffolds, and the guidance discourse is output to the learner.

[0043] Optionally, the problem analysis data may include the problem focus, task format, target concept, reasoning step chain consisting of standard reasoning step identifiers, key cognitive bottlenecks, and a set of candidate misunderstanding types. The problem focus is a brief summary of the current problem-solving objective. The task format is a six-element set consisting of concept understanding, relational reasoning, process explanation, phenomenon prediction, experimental design, and evidence evaluation. The target concept is the core scientific concept or knowledge point examined by the problem. The standard reasoning step chain is an ordered sequence consisting of reasoning step identifiers, which are an eight-element set consisting of identifying the problem objective, recalling relevant concepts, establishing conceptual relationships, analyzing question conditions, eliminating distracting options, constructing a reasoning chain, verifying the rationality of the answer, and reflecting on the reasoning process. The first seven items constitute the reasoning step chain of the main problem stage and are arranged in strictly ascending order of number; the eighth item is only used when the process jump identifier is "transfer". Key cognitive bottlenecks are the cognitive stages where learners of the corresponding age group are most likely to get stuck. The set of candidate misunderstanding types is a priori list of learner misunderstanding types that may occur within the problem domain.

[0044] Optionally, learners' answers to natural language questions, along with pre-defined structured prompts, can be input into a large language model to obtain question analysis data in a structured data format (such as JSON) output by the large language model.

[0045] Optionally, obtain the learner's discourse in round t regarding the natural language problem. Using a large language model, based on problem analysis data, current discourse and historical dialogue context, learner state vectors are output along three independent cognitive dimensions. and independently output misunderstanding indication information. .

[0046] The state values ​​for the three cognitive dimensions are as follows: Concept mastery state Taken from the set of concept mastery states ,in This indicates that a concept is missing or ambiguous. This indicates a grasp of the conceptual aspects. This indicates a basic grasp of the concepts; and the progress of the process. Taken from the process advancement state set ,in This indicates that the current reasoning step is blocked. This indicates that the current reasoning step has partially progressed. Indicates the completion of the current reasoning step; expresses the state. Taken from the set of expressed states ,in Indicates difficulty in expressing oneself. Indicates basic expression, This indicates a complete expression.

[0047] The states of the three cognitive dimensions are independent and uncoupled, and the output is three independent discrete state values, rather than converting the three dimensions into a single scalar. This cognitive state diagnosis method is fundamentally different from the non-independent state representation method in the existing scheme, which is composed of an emotion vector superimposed with a single scalar skill mastery score.

[0048] Optionally, an indication of whether a specific misunderstanding was detected in this round can be output independently. When a specific misunderstanding matching the candidate misunderstanding type set is detected in the learner's utterance, Set to 1, and attach a misunderstanding description; otherwise Set it to 0.

[0049] Optionally, learner status and preset prompts can be input into the large language model, enabling the large language model to generate learning guidance decisions based on preset control rules.

[0050] Optionally, the guided learning decision may include the goal reasoning steps, teaching action types, reference dialogue, process jump indicators, and candidate scaffolding feasibility domains. The core field of the guided learning decision is the candidate scaffolding feasibility domain. The candidate stent feasible region is a set consisting of 2 to 3 stent level numbers that are consecutive.

[0051] Optionally, the type of teaching action can be freely selected. Problem Focus Conditions or evidence extraction Concept activation, Relationship or explanation construction, Misunderstanding correction The six-element set consisting of transfer or summarization is used to forcibly set the teaching action type when a specific misunderstanding is detected. .

[0052] Optionally, the target reasoning step is taken from an element in the reasoning step chain.

[0053] Optionally, the reference wording is a reference expression text corresponding to the selected support level range.

[0054] Optionally, the process jump identifier is taken from a quadruple set consisting of continue, summary, transfer, and done.

[0055] Alternatively, the feasible region of the candidate stent can be... Target reasoning steps Teaching actions Reference script and the learner's current round of discourse Learner status The historical dialogue summary is concatenated into a structured input and provided to the trained large language model. This large language model is trained through a dual-task joint training process, and it simultaneously outputs the selected scaffold level in the form of structured data during the same inference process. With teaching discourse The core constraint is the hard constraint of the feasible region. .

[0056] Optionally, the scaffold selection tendency of a large language model is implicitly induced by the training data: when the learner is in a high-level state (e.g., , , When learners are at a low level (e.g., when they tend to choose the lowest level in the feasible region); when learners are at a low level (e.g., when they are in a low-level state), they tend to choose the lowest level in the feasible region. , , When a misunderstanding is detected, the system tends to select the highest level within the feasible region; when a misunderstanding is detected, it tends to select a higher level within the feasible region. This strategy selection capability is fundamentally different from the architecture of existing schemes where the lower-level model only undertakes the single task of text generation, and the strategy selection is entirely determined by the threshold switching of the upper-level model.

[0057] The adaptive guided learning method based on scaffold feasible region constraints provided in this application models learner states as independent concept mastery states, process progression states, and expression states. This enables the system to accurately perceive learners' heterogeneous cognitive states along three independent cognitive dimensions, avoiding the state confusion problem caused by compressing multidimensional cognitive states into a single scalar. It can provide differentiated responses to typical heterogeneous states such as "concept mastered but difficult to express" or "fluent expression but conceptual confusion." By generating candidate scaffold feasible regions consisting of at least two consecutively numbered scaffold levels based on learner states, the teaching strategy output is extended from the single-scaffold level hard switching in existing solutions to a continuous scaffold level. The feasible domain of the strategy enables the scaffolding strength in the learning process to be dynamically adjusted within the nearest interval, achieving the teaching objective of "dynamically adjusting the scaffolding strength within the zone of proximal development" and significantly improving the flexibility of the teaching strategy. By having the large language model trained to simultaneously output the selected scaffolding level and the learning discourse in the same inference process, and with the selected scaffolding level constrained within the feasible domain of the candidate scaffolding, the large language model has the ability to select strategies while generating learning discourse. This avoids the architectural defects of existing solutions where the discourse generation module only undertakes the task of text expression and the strategy selection is completely determined by external rigidity. It achieves the synergistic optimization of strategy selection and discourse generation, and improves the flexibility of adaptive learning.

[0058] In some embodiments, the set of elements of the candidate stent feasible domain is taken from a set of six stent levels. The set of six stent levels includes, in order, open questioning level, directed questioning level, evidence element prompting level, principle prompting level, local correction level, and explicit explanation and restatement level. The information increment carried by adjacent levels satisfies the condition that the higher the level number, the greater the information increment.

[0059] The set of elements of the feasible region of candidate stents is taken from the set of six stent levels. The six levels of the support system represent open questioning, targeted questioning, evidence element prompting, principle prompting, local correction, and explicit explanation plus restatement, respectively. The information increment carried by adjacent levels satisfies the monotonicity that the higher the level number, the greater the information increment.

[0060] In some embodiments, the preset control rules include state mapping rules, misunderstanding priority correction rules, expression lower limit protection rules, and continuity constraint rules.

[0061] Figure 2 This is a schematic diagram of the cascading of hybrid control rules provided in the embodiments of this application, such as... Figure 2 As shown, the hybrid control rules are applied through a four-level constraint cascade: Level 1, state mapping rules. Based on the binary of conceptual understanding state and process progression state Determine the initial feasible region In this embodiment, specifically: when hour ;when hour ;when hour ;when hour This mapping satisfies the monotonicity that the lower the learner's state level, the higher the average scaffold number covered by the feasible region. Those skilled in the art can make adaptive adjustments to the mapping table according to actual application scenarios.

[0062] Level Two: Misunderstanding-First Correction Rule ,when When the feasible region is covered, And the types of teaching movements in this round Forced to be set (Misunderstanding correction); When The feasible region remains unchanged.

[0063] Level 3, Expression Lower Bound Protection Rules ,when When this happens, stent levels with numbers lower than a preset lower limit are removed from the feasible region. In this embodiment, the preset lower limit is set to 3, meaning that stents with numbers lower than a preset lower limit are removed. and ,when The feasible region remains unchanged.

[0064] Level 4, Continuity and Cardinality Constraint Rules The feasible region output by the aforementioned three-level rules is made continuous to ensure the final... Satisfying cardinality constraint With continuity constraint (the difference in the numbers of any adjacent elements arranged in ascending order is equal to 1).

[0065] Ultimately, the feasible domain of candidate stents The generation process expression is:

[0066] in , , These represent the current round's concept mastery state, process progress state, and expression state, respectively.

[0067] Optionally, when generating guided learning decisions, the following fields are also output: target reasoning steps. Teaching actions Reference script Process jump indicator .

[0068] In some embodiments, the training process of a large language model includes: The training dataset is constructed at the granularity of the learning unit. Each learning unit contains at least one original question, target concept, core knowledge point set, standard reasoning step chain, candidate misunderstanding type set, and scaffolding prompt chain. The scaffolding prompt chain consists of reference scripts corresponding to multiple scaffolding levels. Perform a complete multi-turn dialogue simulation for each learning unit and generate a dialogue log containing the context of the multi-turn dialogue. The dialogue logs are flattened and extracted according to the dialogue rounds to obtain training samples. The input fields of each training sample include the candidate scaffold feasible domain, target reasoning steps, teaching action type, reference dialogue, learner's current round of speech, learner status and historical dialogue summary. The output field is structured data that contains two keys: selected scaffold level and guidance speech. Based on the training samples, a low-rank parameter fitting method is used to jointly train the base large language model.

[0069] Figure 3 This is a flowchart of the dual-task joint training process provided in an embodiment of this application, such as... Figure 3 As shown, the specific training steps for the large language model are as follows: Step 1: Construction of the learning unit and organization of the training dataset.

[0070] The training dataset is constructed using guided learning units as the granularity for organizing offline training data. Each guided learning unit It must contain at least the following fields: Original question Target concept Core knowledge points collection Standard reasoning chain Candidate Misunderstanding Type Set support chain ,in For stent level The corresponding reference dialogue. The scaffolding cue chain is pre-generated by teachers in this field or auxiliary tools with language comprehension capabilities, serving as the discourse anchor points in the training phase of the guided discourse generation module.

[0071] Step two: Complete multi-turn dialogue simulation.

[0072] Language models with multi-turn dialogue capabilities play the roles of both teacher and learner in each guided learning unit. Conduct a complete multi-round dialogue simulation. The simulation process strictly follows the protocol from steps S1 to S4. Each round generates a dialogue log containing five types of records: problem analysis record, learner diagnosis record, guidance decision record, teacher's discourse record, and learner response evaluation record. The simulation continues until the process jump flag is set to summary, transfer, or done, resulting in a complete dialogue log. .

[0073] Step 3: Flatten the training samples.

[0074] All dialogue logs are flattened and extracted to obtain a training sample set. The input fields for each training sample include: target concept, target reasoning steps, instructional action type, candidate scaffold feasible region, reference dialogue, learner's current turn of speech, learner status, and a summary of historical dialogues. The output fields are structured data, containing two keys: selected scaffold level and guiding dialogue. This complete dialogue simulation and flattened extraction mechanism, with guiding units as the granularity, ensures that each training sample carries complete multi-turn dialogue context information.

[0075] Step 4: Low-rank parameter adaptation joint training.

[0076] Using a pre-trained large language model as the base model, a low-rank parameter fitting method is employed for joint training on the training sample set. The training objective is to minimize the composite loss function.

[0077] in To determine the classification loss for the selected stent level, cross-entropy is used. To assess the generation loss of guided discourse, a token-by-token negative log-likelihood approach is used. The feasible region consistency penalty term is defined as the sum of probabilities assigned by the model to stent grades outside the feasible region of the candidate stent, satisfying the condition that the selected stent grade predicted by the model belongs to the feasible region of the candidate stent. ,otherwise >0. , , The loss weighting coefficients are used; those skilled in the art can make adaptive adjustments to the aforementioned parameters according to the actual application scenario.

[0078] The dual-task joint training mechanism enables the guided discourse generation module to simultaneously learn two tasks: "choosing which scaffold level within the feasible domain" (classification task) and "how to generate the corresponding discourse" (generation task). The supervision signals for both tasks are provided in the same forward propagation, allowing them to achieve collaborative optimization through shared underlying semantic representations. This dual-task joint training mechanism is fundamentally different from existing solutions where the lower-level model only undertakes the single task of text generation.

[0079] After training, the adapted parameters are saved to form the final guided discourse generation module. During the reasoning phase, the guided discourse generation module outputs the selected scaffold level and guided discourse in a structured data format, and ensures that the hard constraints of the feasible region are met through structured output parsing and post-verification.

[0080] In some embodiments, the method further includes: If the selected scaffold level output by the large language model is not within the feasible region of the candidate scaffolds, the rewriting mechanism is triggered and the output is overwritten with the closest scaffold level within the feasible region of the candidate scaffolds.

[0081] Optionally, the selected scaffold level output by the large language model needs to be constrained within the feasible region of the candidate scaffolds. This constraint is implemented during the training phase through a feasible region consistency penalty term in the composite loss function. Guided by this approach, the inference stage is ensured through structured output parsing and post-validation: when the large language model outputs... When a constraint is violated, a rewrite mechanism is triggered, and the output is overridden with the nearest scaffold level within the feasible domain.

[0082] In some embodiments, after outputting the guiding remarks to the learners, the method further includes: Receive the learner's response, update the learner's status based on the response, and output a process jump indicator.

[0083] guiding discourse Output to learners and receive their responses The learner response assessment module outputs updated state values ​​along three independent cognitive dimensions: concept mastery, process progression, and expression, thus obtaining the new state. .

[0084] The core feature of this step is that the state is updated along three independent cognitive dimensions. The response assessment does not output a single scalar score after weighted conversion of the three dimensions, but maintains the independence of the three dimensions. This assessment method is fundamentally different from the non-orthogonal state representation method in the existing scheme, which is composed of an emotional vector superimposed with a single scalar skill mastery score.

[0085] Optionally, output the learner's current mastery level. (Taken from the ternary set of low, medium, and high) Progress Indicators (Boolean value) and process jump indicator .

[0086] In each round of decision-making, the feasible domain of candidate scaffolds and the selected scaffold level are explicit discrete variables. Combined with the three-dimensional independent output of learner status and the explicit values ​​of process jump indicators, the entire dialogue chain becomes an auditable, replayable, and optimizable teaching process log, which is convenient for teachers and teaching researchers to review and improve.

[0087] In some embodiments, the process jump indicator is determined based on a combination of the current level of understanding, concept mastery status, process progress status, number of consecutive rounds without progress, and cumulative number of rounds.

[0088] Optionally, when =high、 and hour, = summary; when =high、 And when migration issues exist, = transfer; when no progress is made after reaching a preset threshold number of rounds. = summary; when hour, = done; In other cases = continue. When The method returns to step S2 and proceeds to the next round; when Each time, they enter the corresponding branch.

[0089] in Based on the current level of understanding, This is the preset maximum number of migration trigger rounds. The maximum allowed number of rounds is preset; the parameter is set by those skilled in the art according to the actual application scenario.

[0090] In one embodiment of this application, the upper limit of the number of migration trigger rounds. Set to 10, maximum allowed number of rounds The threshold is set to 12, and the threshold for no continuous progress is set to 3 rounds; those skilled in the art can make adaptive adjustments to the aforementioned parameters according to the actual application scenario.

[0091] In one embodiment of this application, taking a primary school science knowledge point guidance scenario as an example, after a learner submits a preliminary answer about a scientific phenomenon (such as desert plant roots, buoyancy, combustion conditions, etc.), the method of this invention executes sequentially: the problem structure analysis module parses the target concept, reasoning step chain, and candidate misunderstanding types of the problem; the learner cognitive state diagnosis module assesses the learner's current state along three independent cognitive dimensions—concept, reasoning process, and language expression—and independently determines whether there is a specific misunderstanding; the scaffold feasible domain decision module outputs candidate scaffold feasible domains containing 2 to 3 consecutive levels according to mixed control rules; the guidance discourse generation module selects a suitable scaffold level for the current state within the feasible domain and generates corresponding guidance discourse; after the learner answers according to the guidance discourse, the learner response evaluation module updates the learner's state along the three independent cognitive dimensions and determines the direction of process jump. When the learner's three-dimensional cognitive state reaches the highest level and the current mastery level is judged as high, the method jumps to the summary or transfer branch; when no progress is made in multiple consecutive rounds, the method completes the bottom-line guidance with the highest intensity scaffold and then jumps to the summary branch. Each round of decision-making in the entire mentoring process generates an explicit candidate scaffolding feasible domain, a selected scaffolding level, and a learner status update record, making the entire dialogue chain auditable and replayable.

[0092] Figure 4 This is a schematic diagram of the structure of the adaptive learning guidance device based on the feasible region constraint of the support provided in the embodiments of this application, as shown below. Figure 4 As shown, in one embodiment of this application, an adaptive guidance device based on scaffold feasible region constraints is provided to implement the above-described method embodiments. The device in this embodiment includes a processor and a memory, with the processor connected to the memory. The processor is used to call and execute a program stored in the memory, and the memory is used to store the program, which is at least used to execute the adaptive guidance method based on scaffold feasible region constraints in the above embodiments.

[0093] The processor can be a heterogeneous architecture of CPU, GPU, or NPU, capable of efficiently executing tasks such as model inference, rule matching, and state updates. The memory can be RAM, ROM, or Flash, used to store program code, guided learning unit data, learner historical conversation data, model parameters, and other related data. All modules communicate via a system bus. The device also includes an input / output interface module responsible for receiving natural language questions input by learners, outputting guided learning utterances and learner state update results, and enabling interaction with learners.

[0094] The adaptive learning system based on scaffold feasible region constraints provided in this application is described below. The adaptive learning system based on scaffold feasible region constraints described below can be referred to in correspondence with the adaptive learning method based on scaffold feasible region constraints described above.

[0095] Figure 5 This application provides a schematic diagram of the structure of an adaptive guided learning system based on scaffold feasible region constraints, as shown in the embodiments below. Figure 5 As shown, the system 500 includes: The first acquisition module 510 is used to obtain question analysis data based on the learner's answers to natural language questions. The second acquisition module 520 is used to acquire learner status based on question analysis data and learners’ answers to natural language questions. Learner status includes independent concept mastery status, process progress status and expression status. The generation module 530 is used to generate a learning guidance decision based on the learner's state and according to preset control rules. The learning guidance decision includes a candidate scaffold feasible region, which is a set consisting of at least two consecutively numbered scaffold levels. The learning guidance module 540 is used to input the learning guidance decision, the learner's answer to the natural language question and the learner's status into the trained large language model, obtain the selected scaffold level and learning guidance discourse in the form of structured data output by the large language model, select the scaffold level within the feasible range of the candidate scaffolds, and output the learning guidance discourse to the learner.

[0096] It should be understood that the above system is used to execute the methods in the above embodiments. The corresponding program modules in the system are similar in implementation principle and technical effect to those described in the above methods. The working process of the system can be referred to the corresponding process in the above methods, and will not be repeated here.

[0097] Based on the methods in the above embodiments, Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown in the illustration, this application provides an electronic device that may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions from the memory 630 to execute the adaptive guided learning method based on scaffold feasible region constraints described in the above embodiment.

[0098] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the adaptive guided learning method based on scaffold feasible region constraints described in the various embodiments of this application.

[0099] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the adaptive guided learning method based on scaffold feasible region constraints in the above embodiments.

[0100] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the adaptive guided learning method based on scaffold feasible region constraints in the above embodiments.

[0101] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0102] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0103] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0104] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0105] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An adaptive guided learning method based on scaffold feasible region constraints, characterized in that, include: Based on learners' responses to natural language questions, obtain question analysis data; Based on the problem analysis data and the learner's answers to the natural language question, the learner's state is obtained, which includes mutually independent concept mastery state, process progress state, and expression state. Based on the learner's state, a guidance decision is generated according to a preset control rule. The guidance decision includes a candidate scaffold feasible region, which is a set consisting of at least two consecutively numbered scaffold levels. The learning guidance decision, the learner's answer to the natural language question, and the learner's state are input into the trained large language model to obtain the selected scaffold level and learning guidance discourse in the form of structured data output by the large language model. The selected scaffold level is within the feasible domain of the candidate scaffolds, and the learning guidance discourse is output to the learner.

2. The adaptive guided learning method based on scaffold feasible region constraints according to claim 1, characterized in that, The set of feasible domains of the candidate stents is taken from a set of six stent levels. The set of six stent levels includes, in order, open questioning level, targeted questioning level, evidence element prompting level, principle prompting level, local correction level, and explicit explanation plus restatement level. The information increment carried by adjacent levels satisfies the condition that the higher the level number, the greater the information increment.

3. The adaptive guided learning method based on scaffold feasible region constraints according to claim 2, characterized in that, The preset control rules include state mapping rules, misunderstanding priority correction rules, expression lower limit protection rules, and continuity constraint rules.

4. The adaptive guided learning method based on scaffold feasible region constraints according to claim 1, characterized in that, The training process of the large language model includes: A training dataset is constructed at the granularity of the learning unit. Each learning unit contains at least one original question, target concept, core knowledge point set, standard reasoning step chain, candidate misunderstanding type set, and scaffolding prompt chain. The scaffolding prompt chain consists of reference scripts corresponding to multiple scaffolding levels. Perform a complete multi-turn dialogue simulation for each learning unit and generate a dialogue log containing the context of the multi-turn dialogue. The dialogue logs are flattened and extracted according to the dialogue rounds to obtain training samples. The input fields of each training sample include the candidate scaffold feasible domain, target reasoning steps, teaching action type, reference dialogue, learner's current round of speech, learner status and historical dialogue summary. The output field is structured data that contains two keys: selected scaffold level and guidance speech. Based on the training samples, a low-rank parameter fitting method is used to jointly train the base large language model.

5. The adaptive guided learning method based on scaffold feasible region constraints according to claim 1, characterized in that, The method further includes: If the selected scaffold level output by the large language model is not within the feasible region of the candidate scaffolds, a rewriting mechanism is triggered and the output is overwritten with the closest scaffold level within the feasible region of the candidate scaffolds.

6. The adaptive guided learning method based on scaffold feasible region constraints according to claim 1, characterized in that, After outputting the guiding discourse to the learners, the method further includes: Receive the learner's response, update the learner's status based on the response, and output a process jump flag.

7. The adaptive guided learning method based on scaffold feasible region constraints according to claim 1, characterized in that, The process jump indicator is determined based on a comprehensive assessment of the current level of understanding, concept mastery, process progress, number of consecutive rounds without progress, and cumulative number of rounds.

8. An adaptive guided learning system based on scaffold feasible region constraints, characterized in that, include: The first acquisition module is used to obtain question analysis data based on the learner's answers to natural language questions; The second acquisition module is used to acquire learner status based on the question analysis data and learner's answers to the natural language question. The learner status includes mutually independent concept mastery status, process progress status and expression status. The generation module is used to generate a guidance decision based on the learner's state and according to preset control rules. The guidance decision includes a candidate scaffold feasible region, which is a set consisting of at least two consecutively numbered scaffold levels. The guidance module is used to input the guidance decision, the learner's answer to the natural language question, and the learner's state into the trained large language model to obtain the selected scaffold level and guidance discourse in the structured data form of the large language model output. The selected scaffold level is within the feasible domain of the candidate scaffolds, and the guidance discourse is output to the learner.

9. An electronic device, characterized in that, include: At least one memory for storing computer programs; At least one processor is configured to execute a program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the adaptive guided learning method based on scaffold feasible region constraints as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is run on the processor, the processor performs the adaptive guided learning method based on scaffold feasible region constraints as described in any one of claims 1-7.

Images