A large model-based conversational knowledge tutoring method and device

Abstract

The application belongs to the technical field of data processing, and particularly relates to a dialogue type knowledge tutoring method and device based on a large model. The method comprises the following steps: S1, receiving an error exercise and an error answer submitted by a user; S2, determining the type of the error exercise; S3, according to the type of the error exercise, calling corresponding model parameter sets from a server and loading the model parameter sets to a large language model; and S4, based on the large language model, generating guiding dialogue content for the error answer, wherein the guiding dialogue content comprises at least one guiding question, and dynamically adjusting a dialogue strategy according to the answer content of the user to the guiding question until the user obtains a correct answer or reaches a preset upper limit of the number of interactions. The application realizes accurate control of a dialogue strategy of a large model, ensures strong correlation between error exercises and tutoring, and significantly improves tutoring efficiency and individualization.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, and specifically relates to a dialogic knowledge tutoring method and device based on a large model. Background Technology

[0002] In recent years, with the development of artificial intelligence technology, online education and tutoring systems have gone through three main stages of development. The first stage is static question bank systems, such as various online question bank websites, which feature question retrieval based on keyword matching and fixed parsing templates. The second stage is rule engine systems, which combine OCR technology for question recognition and expert-based parsing generation for learning and tutoring. However, the quality of their question parsing is low, and they mainly rely on manual annotation, making it unable to handle open-ended questions. The third stage is the AI-assisted systems that have been publicly released in recent years. These systems achieve natural language interaction and have a certain ability to understand context, but these online tutoring systems still have the following problems.

[0003] First, existing AI-assisted systems rely on static feedback mechanisms, providing only correct answers or fixed explanation templates, lacking personalized guidance tailored to the root causes of errors. Second, they suffer from poor adaptability to different question types; most systems cannot distinguish the differences in tutoring strategies between mathematical calculation questions and conceptual understanding questions. Finally, the application of large-scale models is crude; when directly calling general dialogue models, the generated content lacks structure, easily including irrelevant information or skipping key derivation steps. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a dialogic knowledge tutoring method and apparatus based on a large model. By quantifying cognitive gaps, it enables tiered guidance of knowledge points, thereby improving tutoring efficiency.

[0005] The first aspect of this application is a dialogic knowledge tutoring method based on a large model, which mainly includes:

[0006] Step S1: Receive the user's submitted incorrect exercises and their incorrect answers;

[0007] Step S2: Determine the type of the incorrect exercise;

[0008] Step S3: Based on the type of the error exercise, retrieve the corresponding model parameter set from the server and load it into the large language model, including:

[0009] Step S31: Obtain one or more specific parameters and one or more general parameters that match the type of the error exercise, and jointly construct a model parameter set. The specific parameters are characterized by Boolean values ​​to indicate whether the control strategy corresponding to the specific parameter can participate in the dialogue generation process of the large language model. The general parameters are characterized by their values ​​to indicate the degree to which the control strategy corresponding to the general parameter participates in the large language model.

[0010] Step S32: Inject the general parameters and special parameters into the large language model using the SDK or structured prompt words;

[0011] Step S4: Based on the large language model, generate guided dialogue content for the incorrect answer. The guided dialogue content includes at least one guiding question, and dynamically adjust the dialogue strategy according to the user's answer to the guiding question until the user arrives at the correct answer or reaches the preset interaction limit.

[0012] Preferably, step S2 further includes:

[0013] The probabilities of each type of incorrect exercise are obtained by using keyword matching, grammatical structure analysis, and pre-trained model classification.

[0014] The final type of the incorrect exercise is determined by weighted calculation.

[0015] Preferably, step S4 further includes:

[0016] Step S41: Calculate the logical difference between the incorrect answer and the correct answer;

[0017] Step S42: Based on the logical gap, calculate the initial level of the guided dialogue content and determine the dialogue guidance path from the initial level to the highest level. Each guiding question on the dialogue guidance path aims to narrow the gap between the user's cognition and the correct answer.

[0018] Step S43: Based on real-time user feedback, dynamically adjust the difficulty or direction of subsequent guidance questions.

[0019] Preferably, in step S41, for numerical error questions, the logical gap is represented by the relative error between the correct and incorrect answers; for non-numerical error questions, the logical gap is calculated by the semantic similarity between the correct and incorrect answers.

[0020] Preferably, step S43 further includes:

[0021] Step S431: Calculate the correction progress based on user feedback. :

[0022] ;

[0023] in, This refers to the logical gap between the user's current answer and the correct answer. This refers to the logical discrepancy between the answer provided by the user in the previous feedback and the correct answer.

[0024] Step S432, when the correction progress If the problem level exceeds the first set value, the problem level is raised; otherwise, the current problem level is retained. The correction progress is calculated after k consecutive iterations. If none of the values ​​exceed the first set value, then a new dialogue guidance path will be used.

[0025] Preferably, the method further includes:

[0026] Step S5: Record the complete dialogue process between the user and the large language model;

[0027] Step S6: Based on the complete dialogue process, generate an analysis report on the user's knowledge gaps.

[0028] The second aspect of this application provides a dialogic knowledge tutoring device based on a large model, mainly comprising:

[0029] The error problem acquisition module is used to receive user-submitted error problems and their incorrect answers;

[0030] The exercise type acquisition module is used to determine the type of the incorrect exercise;

[0031] A model parameter loading module is used to retrieve the corresponding model parameter set from the server and load it into the large language model based on the type of the error exercise; the model parameter loading module includes:

[0032] The parameter determination unit is used to acquire one or more specific parameters and one or more general parameters that match the type of the error exercise, and to jointly construct a model parameter set. The specific parameters are characterized by Boolean values ​​to indicate whether the control strategy corresponding to the specific parameter can participate in the dialogue generation process of the large language model. The general parameters are characterized by their values ​​to indicate the degree to which the control strategy corresponding to the general parameter participates in the large language model.

[0033] The parameter injection unit is used to inject the general parameters and special parameters into the large language model through SDK or structured prompt words;

[0034] The dialogue control module is used to generate guided dialogue content for the incorrect answers based on the large language model. The guided dialogue content includes at least one guiding question, and the dialogue strategy is dynamically adjusted according to the user's answer to the guiding question until the user arrives at the correct answer or the preset number of interactions is reached.

[0035] Preferably, the exercise type acquisition module includes:

[0036] The exercise type probability acquisition unit is used to obtain the probability of each type of the incorrect exercise based on keyword matching, grammatical structure analysis and pre-trained model classification.

[0037] The exercise type determination unit is used to determine the final type of the incorrect exercise through weighted calculation.

[0038] Preferably, the dialogue control module includes:

[0039] A logical gap calculation unit is used to calculate the logical gap between the incorrect answer and the correct answer;

[0040] The guiding question initialization unit is used to calculate the initial level of the guided dialogue content based on the logical gap, and determine the dialogue guidance path from the initial level to the highest level. Each guiding question on the dialogue guidance path aims to narrow the gap between the user's cognition and the correct answer.

[0041] The dialogue adjustment unit is used to dynamically adjust the difficulty or direction of subsequent guiding questions based on real-time user feedback.

[0042] Preferably, in the logical gap calculation unit, for numerical error questions, the logical gap is represented by the relative error between the correct and incorrect answers; for non-numerical error questions, the logical gap is calculated by the semantic similarity between the correct and incorrect answers.

[0043] Preferably, the dialogue adjustment unit includes:

[0044] The correction progress calculation subunit is used to calculate the correction progress based on the user feedback. :

[0045] ;

[0046] in, This refers to the logical gap between the user's current answer and the correct answer. This refers to the logical discrepancy between the answer provided by the user in the previous feedback and the correct answer.

[0047] Problem-level control subunit, used to control the progress of corrections. If the problem level exceeds the first set value, the problem level is raised; otherwise, the current problem level is retained. The correction progress is calculated after k consecutive iterations. If none of the values ​​exceed the first set value, then a new dialogue guidance path will be used.

[0048] Preferably, the device further includes:

[0049] The dialogue process recording module is used to record the complete dialogue process between the user and the large language model;

[0050] The analysis report generation module is used to generate an analysis report on the user's knowledge gaps based on the complete dialogue process.

[0051] This application achieves precise control of the large-scale model dialogue strategy, ensuring a strong correlation between incorrect exercises and tutoring, and significantly improving tutoring efficiency and personalization. Attached Figure Description

[0052] Figure 1 This is a flowchart of one implementation of the dialogic knowledge tutoring method based on a large model, as described in this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are only some, not all, of the embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0054] According to the first aspect of this application, a dialogic knowledge tutoring method based on a large model, such as... Figure 1 As shown, it mainly includes:

[0055] Step S1: Receive the user's submitted incorrect exercises and their incorrect answers;

[0056] Step S2: Determine the type of the incorrect exercise;

[0057] Step S3: Based on the type of the error exercise, retrieve the corresponding model parameter set from the server and load it into the large language model, including:

[0058] Step S31: Obtain one or more specific parameters and one or more general parameters that match the type of the error exercise, and jointly construct a model parameter set. The specific parameters are characterized by Boolean values ​​to indicate whether the control strategy corresponding to the specific parameter can participate in the dialogue generation process of the large language model. The general parameters are characterized by their values ​​to indicate the degree to which the control strategy corresponding to the general parameter participates in the large language model.

[0059] Step S32: Inject the general parameters and special parameters into the large language model using the SDK or structured prompt words;

[0060] Step S4: Based on the large language model, generate guided dialogue content for the incorrect answer. The guided dialogue content includes at least one guiding question, and dynamically adjust the dialogue strategy according to the user's answer to the guiding question until the user arrives at the correct answer or reaches the preset interaction limit.

[0061] This application presents an intelligent tutoring method based on a large model. By integrating keyword matching, grammatical analysis, and pre-trained model classification, it achieves high-precision question type identification. It employs collaborative control of dedicated and general parameters to ensure adaptability to question types across different subjects. In step S4, a cognitive gap quantification mechanism is introduced, which can dynamically generate hierarchical guidance questions strongly correlated with incorrect answers and adjust the dialogue strategy based on real-time feedback, thus maintaining the advantages of natural interaction while ensuring the rigor of teaching.

[0062] First, in step S1, this application reads the questions that the user answered incorrectly and their wrong answers by scanning the test paper or by having the user upload the test paper. It is understood that the system can also obtain the correct answers at the same time.

[0063] Next, in step S2, the types of these exercises are determined so that different large language models can be used for knowledge point instillation. In this embodiment, the types of exercises include, but are not limited to, mathematical calculation problems, logical reasoning problems, concept comprehension problems, or language application problems.

[0064] In some alternative implementations, step S2 further includes:

[0065] The probabilities of each type of incorrect exercise are obtained by using keyword matching, grammatical structure analysis, and pre-trained model classification.

[0066] The final type of the incorrect exercise is determined by weighted calculation.

[0067] This embodiment can determine the subject or question type category of the incorrect exercises submitted by the user, so that the server can call the appropriate model parameter set. Specifically, in step S21, this application identifies the type of incorrect exercises through multiple methods, and then in step S22, performs a comprehensive weighted calculation. This combined approach improves the accuracy of exercise type identification.

[0068] Keyword matching is primarily suited for mathematical calculation problems. For example, detecting keywords like "solve equations" or "calculate" increases the probability of classifying it as a mathematical calculation problem. Detecting keywords like "what," "why," or "explain" increases the probability of classifying it as a concept comprehension problem. Detecting keywords like "if...then" increases the probability of classifying it as a logic reasoning problem. Grammatical structure analysis is mainly suited for language-related questions, such as fill-in-the-blank questions. By analyzing missing sentence components, it categorizes the questions. Pre-trained model classification is mainly for complex question types, such as open-ended questions. Using BERT / CNN to perform multi-label classification on the question text can improve the accuracy of question type recognition. For example, a pre-labeled question type pair could be {Newton's Second Law Formula, Physics Formula Question}.

[0069] Different question types require different large language models with different parameters for guidance, which can accurately match the guidance strategy. For example, it can avoid using mathematical reasoning to handle Chinese questions, while also reducing the load on the large model, filtering out irrelevant knowledge by question type, and improving response speed. Different model parameters are pre-stored in the database or server and are dynamically loaded after the question type is identified.

[0070] Subsequently, in step S3, model parameters are selected, and the selected set of model parameters is loaded into the large language model.

[0071] First, it should be noted that the dynamically loaded model parameters in this application differ from the weight parameters of a large language model. The weight parameters of a large language model are internal trainable parameters within the model's neural network, requiring training and updates to permanently alter the model's capabilities; for example, the GPT-3 model contains 175 billion parameters. The dynamically loaded model parameter set in step S3 of this application consists of externally adjustable variables that control the model's generation behavior. These variables are adjusted in real-time via APIs or prompts and are only effective during a single inference process. These parameters do not directly modify the weights of the large model. Instead, they achieve output control in different directions on the same model through external control switches and prompts. For example, logical deduction dialogues for math problems, metaphorical analogy dialogues for concept problems, or grammatical analysis dialogues for language problems, using implicit input prompts to make the large language model output the desired result.

[0072] Secondly, it should be noted that the parameters in the model parameter set of this application are mainly divided into two types: one is dedicated parameters that match the type of the error exercise, and the other is general parameters. Dedicated parameters address the presence or absence of errors, while general parameters address the participation depth issue.

[0073] For example, the general parameter `tem` controls the randomness of the generation. For math problems, its value is set to 0.1-0.3 to characterize the rigor of the output; for conceptual questions, its value is set to 0.5-0.7 to characterize the flexibility of the output. Another example is the general parameter `top`, which limits the range of candidate words. For math problems, its value is set to 0.7 to indicate that the output needs to focus on formulas; for conceptual questions, its value is set to 0.9 to indicate that the output is open-ended and associative. Finally, the general parameter `maxlen` controls the maximum length of the generated text. For math problems, its value is set to 100 for a concise representation of the answer; for conceptual questions, its value is set to 200 to allow for a detailed explanation of the answer.

[0074] The dedicated parameter `ana` indicates whether analogical explanations are allowed; a value of `False` disables them, and `True` enables them. The dedicated parameter `check` indicates whether the formula validation plugin is enabled; a value of `True` enables automatic calculation validation, and a value of `False` disables it. The dedicated parameter `logy` is an analogy parameter that can be enabled when the error question type is a concept question. For example, if a user asks what inflation is, with the dedicated parameter `logy` enabled, the output of the large language model could be:

[0075] Inflation is like inflating a balloon. The money (balloon) increases, but the goods (air inside the balloon) do not increase, causing prices (balloon volume) to expand. Similarly, deflation is the process of the balloon leaking air.

[0076] As mentioned above, the exercise type determines the values ​​of general and specific parameters. Subsequently, in step S32, the aforementioned model parameter set is injected into the large language model. This can be achieved through methods such as prompt-based injection, API injection, and SDK injection. Prompt-based injection implicitly controls model behavior through structured prompts. API-level injection is suitable for closed-source models, primarily using the JSON body of HTTP requests. SDK-level injection is mainly applied to open-source models.

[0077] Once the large language model is determined, knowledge tutoring can be conducted using the guided dialogue of this application in step S4.

[0078] In some alternative implementations, step S4 further includes:

[0079] Step S41: Calculate the logical difference between the incorrect answer and the correct answer;

[0080] Step S42: Based on the logical gap, calculate the initial level of the guided dialogue content and determine the dialogue guidance path from the initial level to the highest level. Each guiding question on the dialogue guidance path aims to narrow the gap between the user's cognition and the correct answer.

[0081] Step S43: Based on real-time user feedback, dynamically adjust the difficulty or direction of subsequent guidance questions.

[0082] In some optional implementations, in step S41, for numerical error questions, the logical gap is represented by the relative error between the correct and incorrect answers; for non-numerical error questions, the logical gap is calculated by the semantic similarity between the correct and incorrect answers.

[0083] In this embodiment, the correct answer is assumed to be Aco and the incorrect answer to be Aer. The logical gap is defined as a measure of the difference between the two and can be calculated using relative error or semantic similarity.

[0084] (1) For numerical problems, such as mathematical calculations, the difference is measured using relative error:

[0085] CG=|Aco-Aer| / Aco.

[0086] If CG ≤ preset threshold, such as 0.1, the error is considered minor and only a simple prompt is needed; otherwise, step-by-step guidance is required.

[0087] (2) For conceptual questions, such as multiple-choice questions and short-answer questions, semantic similarity is used for calculation. For example, cosine similarity is used for calculation, and the formula is:

[0088] CG=1-sim(E(Aco),E(Aer)).

[0089] Here, E represents the embedding representation of the large model, such as BERT encoding, and sim is the cosine similarity calculation function. The lower the similarity, the greater the calculated logical difference (CG), and the lower the level of the initial guiding question should be, meaning a more basic initial dialogue is needed for guidance.

[0090] Based on the logical gap (CG), guided questions can generally be divided into three levels. Level 1 has the highest CG value, for example, CG ≥ 0.6, indicating that the user may not understand the corresponding exercise at all. At this time, fundamental misunderstandings should be corrected and knowledge gaps should be traced, such as formula errors or conceptual confusion. Level 2 has a CG value between 0.3 and 0.6, indicating that the user can understand the exercise to some extent, but the knowledge points are not well grasped. Guided dialogue can focus on prompting key steps or omissions. Level 3 has the lowest CG value, for example, CG < 0.3. Guided dialogue is used to refine and correct, locate detailed errors, such as calculation details or expression optimization.

[0091] For example, a level 1 dialogue might be: You used formula X, but do the conditions in the problem satisfy the prerequisites for applying that formula? A level 2 dialogue might be: In step 2, why did you choose addition instead of subtraction? A level 3 dialogue might be: Please recalculate a certain step.

[0092] In step S42, the initial dialogue level is determined. Each level typically contains multiple dialogues, thus generating multiple guidance paths from that level to the top level. For example, level 1 contains 10 optional dialogues (a1-a10), level 2 contains 20 optional dialogues (b1-b20), and level 3 contains 30 optional dialogues (c1-c30). When the initial level is determined to be level 2, (b1,b2,b3)-(c1,c2,c3) is a dialogue guidance path, (b4,b5,b6)-(c4,c5) is a dialogue guidance path, and (b8,b10)-(c7,c8) is also a dialogue guidance path.

[0093] Understandably, each guided dialogue path aims to guide users to the correct answers to incorrect exercises through hierarchical dialogue, and different guided dialogue paths are adapted to different users.

[0094] In step S43, it is necessary to elevate the level or adjust the dialogue guidance path based on user feedback. For example, in some optional embodiments, step S43 further includes:

[0095] Step S431: Calculate the correction progress based on user feedback. :

[0096] ;

[0097] in, This refers to the logical gap between the user's current answer and the correct answer. This refers to the logical discrepancy between the answer provided by the user in the previous feedback and the correct answer.

[0098] Step S432, when the correction progress If the problem level exceeds the first set value, the problem level is raised; otherwise, the current problem level is retained. The correction progress is calculated after k consecutive iterations. If none of the values ​​exceed the first set value, then a new dialogue guidance path will be used.

[0099] This embodiment can accurately locate the root cause of errors by quantifying logical gaps and dynamically adjusting strategies, avoiding ineffective guidance, and can adaptively generate questions to match the user's cognitive level.

[0100] For example, the first setting is 0.5, k is 2, and for a math calculation problem, the correct answer is 100, while the user's initial incorrect answer is 50. The initial logical gap CG can be calculated to be 0.5. The CG value is between 0.3 and 0.6, so the initial level is determined to be level 2. The initial dialogue guidance path (b1,b2,b3)-(c1,c2,c3) is selected. The system first provides dialogue b1, such as a basic concept dialogue. Based on this dialogue, the user corrects their answer to 80, so the logical gap is updated to 0.2. At this time, the calculation correction progress is 0.6. Since it is greater than the first setting value of 0.5, the level is promoted to level 3, and the guidance dialogue changes from the basic concept dialogue to the detailed calculation check dialogue c1, until the user obtains the correct answer.

[0101] If, after the first dialogue, the user corrects their answer to 60, the update logic gap is 0.4. The calculated correction progress is 0.2, which is less than the first set value of 0.5. Therefore, the current question level is retained, and dialogue b2 is provided. If, based on comparison b2, the correction progress calculated based on the user's corrected answer is still less than the first set value, the correction progress requirement for two consecutive calculations is met. If none of them exceed the first set value, a new dialogue guidance path needs to be changed, for example (b4,b5,b6)-(c4,c5).

[0102] In some alternative implementations, the method further includes:

[0103] Step S5: Record the complete dialogue process between the user and the large language model;

[0104] Step S6: Based on the complete dialogue process, generate an analysis report on the user's knowledge gaps.

[0105] This embodiment, based on the complete dialogue process, can identify the user's knowledge gaps. For example, too many dialogues on basic concepts indicate that the user has a poor grasp of basic concepts, while too many dialogues on detailed calculations indicate that the user has a careless tendency.

[0106] This application represents a leap from general question answering to professional tutoring. Compared to existing solutions, it significantly improves error diagnosis accuracy and guidance effectiveness, supporting multi-scenario applications from K-12 to vocational education. Through adaptive parameter control of question types and quantification of cognitive gaps, it addresses content reliability issues while maintaining the flexibility of the large language model, providing a new reference model for the development of intelligent education.

[0107] The second aspect of this application provides a large-model-based dialogic knowledge tutoring device corresponding to the above method, mainly comprising:

[0108] The error problem acquisition module is used to receive user-submitted error problems and their incorrect answers;

[0109] The exercise type acquisition module is used to determine the type of the incorrect exercise;

[0110] A model parameter loading module is used to retrieve the corresponding model parameter set from the server and load it into the large language model based on the type of the error exercise; the model parameter loading module includes:

[0111] The parameter determination unit is used to acquire one or more specific parameters and one or more general parameters that match the type of the error exercise, and to jointly construct a model parameter set. The specific parameters are characterized by Boolean values ​​to indicate whether the control strategy corresponding to the specific parameter can participate in the dialogue generation process of the large language model. The general parameters are characterized by their values ​​to indicate the degree to which the control strategy corresponding to the general parameter participates in the large language model.

[0112] The parameter injection unit is used to inject the general parameters and special parameters into the large language model through SDK or structured prompt words;

[0113] The dialogue control module is used to generate guided dialogue content for the incorrect answers based on the large language model. The guided dialogue content includes at least one guiding question, and the dialogue strategy is dynamically adjusted according to the user's answer to the guiding question until the user arrives at the correct answer or the preset number of interactions is reached.

[0114] In some optional implementations, the exercise type acquisition module includes:

[0115] The exercise type probability acquisition unit is used to obtain the probability of each type of the incorrect exercise based on keyword matching, grammatical structure analysis and pre-trained model classification.

[0116] The exercise type determination unit is used to determine the final type of the incorrect exercise through weighted calculation.

[0117] In some alternative implementations, the dialogue control module includes:

[0118] A logical gap calculation unit is used to calculate the logical gap between the incorrect answer and the correct answer;

[0119] The guiding question initialization unit is used to calculate the initial level of the guided dialogue content based on the logical gap, and determine the dialogue guidance path from the initial level to the highest level. Each guiding question on the dialogue guidance path aims to narrow the gap between the user's cognition and the correct answer.

[0120] The dialogue adjustment unit is used to dynamically adjust the difficulty or direction of subsequent guiding questions based on real-time user feedback.

[0121] In some optional implementations, in the logical gap calculation unit, for numerical error questions, the logical gap is represented by the relative error between the correct and incorrect answers; for non-numerical error questions, the logical gap is calculated by the semantic similarity between the correct and incorrect answers.

[0122] In some alternative implementations, the dialogue adjustment unit includes:

[0123] The correction progress calculation subunit is used to calculate the correction progress based on the user feedback. :

[0124] ;

[0125] in, This refers to the logical gap between the user's current answer and the correct answer. This refers to the logical discrepancy between the answer provided by the user in the previous feedback and the correct answer.

[0126] Problem-level control subunit, used to control the progress of corrections. If the problem level exceeds the first set value, the problem level is raised; otherwise, the current problem level is retained. The correction progress is calculated after k consecutive iterations. If none of the values ​​exceed the first set value, then a new dialogue guidance path will be used.

[0127] In some alternative embodiments, the apparatus further includes:

[0128] The dialogue process recording module is used to record the complete dialogue process between the user and the large language model;

[0129] The analysis report generation module is used to generate an analysis report on the user's knowledge gaps based on the complete dialogue process.

[0130] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A large model-based conversational knowledge tutoring method, characterized by, include: Step S1: Receive the user's submitted incorrect exercises and their incorrect answers; Step S2: Determine the type of the incorrect exercise; Step S3: Based on the type of the error exercise, retrieve the corresponding model parameter set from the server and load it into the large language model, including: Step S31: Obtain one or more specific parameters and one or more general parameters that match the type of the error exercise, and jointly construct a model parameter set. The specific parameters are characterized by Boolean values ​​to indicate whether the control strategy corresponding to the specific parameter can participate in the dialogue generation process of the large language model. The general parameters are characterized by their values ​​to indicate the degree to which the control strategy corresponding to the general parameter participates in the large language model. Step S32: Inject the general parameters and special parameters into the large language model using the SDK or structured prompt words; Step S4: Based on the large language model, generate guided dialogue content for the incorrect answer. The guided dialogue content includes at least one guiding question, and dynamically adjust the dialogue strategy according to the user's answer to the guiding question until the user arrives at the correct answer or reaches the preset interaction limit. Step S4 further includes: Step S41: Calculate the logical difference between the incorrect answer and the correct answer; Step S42: Based on the logical gap, calculate the initial level of the guided dialogue content and determine the dialogue guidance path from the initial level to the highest level. Each guiding question on the dialogue guidance path aims to narrow the gap between the user's cognition and the correct answer. Step S43: Based on real-time user feedback, dynamically adjust the difficulty or direction of subsequent guidance questions; Step S43 further includes: Step S431: Calculate the correction progress based on user feedback. : ； in, This refers to the logical gap between the user's current answer and the correct answer. This refers to the logical discrepancy between the answer provided by the user in the previous feedback and the correct answer. Step S432, when the correction progress If the problem level exceeds the first set value, the problem level is raised; otherwise, the current problem level is retained. The correction progress is calculated after k consecutive iterations. If none of the values ​​exceed the first set value, then a new dialogue guidance path will be used.

2. The dialogic knowledge tutoring method based on a large model as described in claim 1, characterized in that, Step S2 further includes: The probabilities of each type of incorrect exercise are obtained by using keyword matching, grammatical structure analysis, and pre-trained model classification. The final type of the incorrect exercise is determined by weighted calculation.

3. The dialogic knowledge tutoring method based on a large model as described in claim 1, characterized in that, In step S41, for numerical error questions, the logical gap is represented by the relative error between the correct and incorrect answers; for non-numerical error questions, the logical gap is calculated by the semantic similarity between the correct and incorrect answers.

4. The dialogic knowledge tutoring method based on a large model as described in claim 1, characterized in that, The method further includes: Step S5: Record the complete dialogue process between the user and the large language model; Step S6: Based on the complete dialogue process, generate an analysis report on the user's knowledge gaps.

5. A dialogic knowledge tutoring device based on a large model, characterized in that, include: The error problem acquisition module is used to receive user-submitted error problems and their incorrect answers; The exercise type acquisition module is used to determine the type of the incorrect exercise; A model parameter loading module is used to retrieve the corresponding model parameter set from the server and load it into the large language model based on the type of the error exercise; the model parameter loading module includes: The parameter determination unit is used to acquire one or more specific parameters and one or more general parameters that match the type of the error exercise, and to jointly construct a model parameter set. The specific parameters are characterized by Boolean values ​​to indicate whether the control strategy corresponding to the specific parameter can participate in the dialogue generation process of the large language model. The general parameters are characterized by their values ​​to indicate the degree to which the control strategy corresponding to the general parameter participates in the large language model. The parameter injection unit is used to inject the general parameters and special parameters into the large language model through SDK or structured prompt words; The dialogue control module is used to generate guided dialogue content for the incorrect answer based on the large language model. The guided dialogue content includes at least one guiding question, and the dialogue strategy is dynamically adjusted according to the user's answer to the guiding question until the user arrives at the correct answer or the preset number of interactions is reached. The dialogue control module includes: A logical gap calculation unit is used to calculate the logical gap between the incorrect answer and the correct answer; The guiding question initialization unit is used to calculate the initial level of the guided dialogue content based on the logical gap, and determine the dialogue guidance path from the initial level to the highest level. Each guiding question on the dialogue guidance path aims to narrow the gap between the user's cognition and the correct answer. The dialogue adjustment unit is used to dynamically adjust the difficulty or direction of subsequent guiding questions based on real-time user feedback. The dialogue adjustment unit includes: The correction progress calculation subunit is used to calculate the correction progress based on the user feedback. : ； in, This refers to the logical gap between the user's current answer and the correct answer. This refers to the logical discrepancy between the answer provided by the user in the previous feedback and the correct answer. Problem-level control subunit, used to control the progress of corrections. If the problem level exceeds the first set value, the problem level is raised; otherwise, the current problem level is retained. The correction progress is calculated after k consecutive iterations. If none of the values ​​exceed the first set value, then a new dialogue guidance path will be used.

6. The dialogic knowledge tutoring device based on a large model as described in claim 5, characterized in that, The exercise type acquisition module includes: The exercise type probability acquisition unit is used to obtain the probability of each type of the incorrect exercise based on keyword matching, grammatical structure analysis and pre-trained model classification. The exercise type determination unit is used to determine the final type of the incorrect exercise through weighted calculation.

7. The dialogic knowledge tutoring device based on a large model as described in claim 5, characterized in that, In the logical gap calculation unit, for numerical error questions, the logical gap is represented by the relative error between the correct and incorrect answers; for non-numerical error questions, the logical gap is calculated by the semantic similarity between the correct and incorrect answers.

8. The dialogic knowledge tutoring device based on a large model as described in claim 5, characterized in that, The device further includes: The dialogue process recording module is used to record the complete dialogue process between the user and the large language model; The analysis report generation module is used to generate an analysis report on the user's knowledge gaps based on the complete dialogue process.

Images