Heuristic question answering method, system and equipment based on multi-agent interaction and medium

Abstract

The invention discloses a heuristic question answering method, a heuristic question answering system, heuristic question answering equipment and a heuristic question answering medium based on multi-agent interaction, and the heuristic question answering method, the heuristic question answering system, the heuristic question answering equipment and the medium are corresponding schemes. Meanwhile, a complex teaching function is divided into a plurality of agents, and each agent performs own functions, so that the overall interpretability and schedulability are greatly improved, and later rule adjustment and strategy optimization are facilitated; besides, the multi-agent framework and a specific big language model training mode are relatively decoupled, an existing open source or commercial big language model can be directly used, and rapid deployment is achieved. When a more powerful large model is replaced in the future, the scheme can be upgraded only by re-adapting prompts and interfaces of the agents without changing the whole multi-agent structure.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and personalized teaching technology, and in particular to a heuristic question-answering method, system, device and medium based on multi-agent interaction. Background Technology

[0002] With the rapid development of online education and the widespread adoption of smart learning devices, personalized teaching technologies based on big data and artificial intelligence (AI) have become an important component of school teaching and individual learning. These technologies can automatically analyze teaching resources, diagnose students' cognitive states during the learning process, plan learning paths, and provide students with targeted teaching content and assessment methods.

[0003] In smart education scenarios, personalized Q&A is one of the most core and important applications. In traditional classrooms, teachers face numerous questions from students, and due to limited time and energy, it is difficult to provide detailed, one-on-one guidance to each student. With the development of natural language processing and deep learning technologies, personalized Q&A systems can automatically understand student questions and provide students with immediate, high-quality answers and explanations, thereby improving learning efficiency and enhancing the learning experience.

[0004] Existing personalized Q&A technologies mainly include the following two categories: (1) A solution based on a pre-set Q&A database or rule engine.

[0005] This type of solution pre-builds a large number of question-and-answer pairs or rules, matching student questions with questions in a database to return corresponding answers. This approach is simple to implement, has a fast response time, and provides relatively high accuracy within its coverage area. However, it heavily relies on a manually constructed question-and-answer database and rule system, making it difficult to cover all question scenarios. When student questions exceed the preset range, the system struggles to provide effective responses. Furthermore, the answers received by all students are essentially fixed, making it impossible to personalize the response to the diverse cognitive states of different students.

[0006] (2) A scheme based on large language model prompting engineering and fine-tuning.

[0007] With the development of Large Language Models (LLMs), a new type of technology has emerged that uses pre-trained language models and employs prompt engineering and fine-tuning to achieve personalized Q&A. This approach leverages the powerful natural language understanding and generation capabilities of LLMs to provide students with high-quality, comprehensive answers across a wide range of data by designing appropriate prompts or fine-tuning specific subject data. However, existing methods generally follow a "question-answer" paradigm: when a student asks a question, the system directly returns a complete answer or a complete solution process. While this approach can solve the immediate problem, it often neglects the identification and tracking of the student's cognitive state, failing to pinpoint exactly which step or knowledge point the student is struggling with, thus hindering truly personalized tutoring. Furthermore, LLMs often implicitly determine the expression style, content length, and information depth simultaneously during a single generation process, lacking an explicit selection process for teaching strategies. These strategies are often implicitly embedded in the weights within the large model, making them difficult to control and interpret.

[0008] Therefore, a new technical solution is needed to explicitly separate key teaching functions such as "cognitive diagnosis, teaching strategy selection, teaching content generation, and recommended after-class exercises" to construct a controllable, explainable, and heuristic personalized Q&A solution. Summary of the Invention

[0009] The purpose of this invention is to provide a heuristic question-answering method, system, device, and medium based on multi-agent interaction, which can improve learning effectiveness and efficiency, and enhance the interpretability and schedulability of the solution.

[0010] The objective of this invention is achieved through the following technical solution: A heuristic question-answering method based on multi-agent interaction is proposed, which configures a multi-agent framework, with each step executed by a separate agent; the steps include: Step 1: Conduct cognitive diagnosis using students' historical learning records and output students' cognitive status; the historical learning records include students' answers to after-class exercises. Step 2: Based on students' cognitive state and their historical dialogue data, and according to the pre-set teaching strategy library and strategy decision rules, output teaching strategies. Step 3: Under the constraints of teaching strategies, combined with students' cognitive state and the students' previous dialogue data, generate teaching content that follows the Socratic teaching style, realize heuristic Q&A, and update students' historical dialogue data. Step 4: After the heuristic Q&A session, combine the updated student historical dialogue data, student cognitive status, and student historical learning records to output after-class exercises suitable for students.

[0011] A heuristic question-answering system based on multi-agent interaction is provided to implement the aforementioned method. This system is configured with a multi-agent framework, including: a cognitive diagnosis agent, a teaching strategy selection agent, a teaching content generation agent, and a homework recommendation agent; wherein: A cognitive diagnostic agent is used to perform cognitive diagnosis using students' historical learning records and output the students' cognitive status; the historical learning records include students' answers to after-class exercises. The teaching strategy selection agent is used to combine students' cognitive state and students' historical dialogue data, and output teaching strategies according to the preset teaching strategy library and strategy decision rules. The teaching content generating agent is used to generate teaching content that follows the Socratic teaching style under the constraints of teaching strategies, combined with students' cognitive state and students' previous dialogue data, to achieve heuristic Q&A and update students' historical dialogue data. The intelligent agent that recommends homework assignments is used to output homework assignments suitable for students after the heuristic Q&A session, by combining updated historical dialogue data, students' cognitive status, and students' historical learning records.

[0012] A processing device includes: one or more processors; and a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0013] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0014] As can be seen from the technical solutions provided by the present invention above, (1) the whole solution forms a closed loop structure of "explanation-practice-feedback", which improves the learning effect and learning efficiency; (2) the complex teaching function is divided into multiple agents, each agent performs its own duties, which greatly improves the overall interpretability and schedulability, and facilitates the later rule adjustment and strategy optimization; (3) the multi-agent framework of the present invention is relatively decoupled from the specific large language model training method, and can directly use existing open source or commercial large language models to achieve rapid deployment; when replacing a more powerful large model in the future, there is no need to change the overall multi-agent structure, and only the prompts and interfaces of the agents need to be re-adapted to achieve the solution upgrade. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart of a heuristic question-answering method based on multi-agent interaction provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a multi-agent framework provided in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of a heuristic question-answering system based on multi-agent interaction, provided as an embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0021] First, the following explanations are provided for the terms that may be used in this article: The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0022] The following provides a detailed description of a heuristic question-answering method, system, device, and medium based on multi-agent interaction provided by this invention. Contents not described in detail in the embodiments of this invention belong to prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of this invention, they are performed according to conventional conditions in the art or conditions recommended by the manufacturer. Where the manufacturers of the instruments used in the embodiments of this invention are not specified, they are all conventional products that can be purchased commercially.

[0023] Example 1 This invention provides a heuristic question-answering method based on multi-agent interaction. In this method, a multi-agent framework is configured, and each step is executed by a separate agent, such as... Figure 1 As shown, the steps of the entire method are as follows: Step 1: Use students' historical learning records to conduct cognitive diagnosis and output the students' cognitive status.

[0024] In this embodiment of the invention, step 1 can be executed by a cognitive diagnostic agent, as follows: Where f is the cognitive diagnostic agent, and M is the student's historical learning record. This represents the student's cognitive state; where the student's historical learning record M is represented as: ;in, This is the information for the i-th homework question. This represents the student's answer to the i-th homework question. To grade the answers to the i-th homework question, The representative answered correctly. The number of incorrect answers represents the number of homework questions the student has already completed.

[0025] In this embodiment of the invention, the historical learning record includes students' answers to after-class exercises, so as to achieve a closed loop for the entire scheme.

[0026] Step 2: Based on students' cognitive state and their historical dialogue data, and according to the pre-set teaching strategy library and strategy decision rules, output teaching strategies.

[0027] In this embodiment of the invention, step 2 can be performed by selecting an agent through a teaching strategy, as follows: The historical dialogue data of students is mapped into semantic vectors by a text encoder to obtain the semantic representation c of the historical dialogue data; where the historical dialogue data contains the dialogue data of each round of students, and the dialogue data of each round contains the teaching content and the dialogue content output by students. The semantic representation c of historical dialogue data is combined with the students' cognitive state. Concatenate them into a unified feature vector x; The pre-set teaching strategy library is denoted as The teaching is output through the following strategic decision-making rules: ; ; Where softmax is the normalized exponential function, W is the weight parameter, b is the bias parameter, and s is the policy score vector, where the th... Item represents the first One strategy The score; P is a probability vector, where the first... Item represents the first One strategy The probability of being selected; L represents the number of strategies; Choose the strategy with the highest probability from the probability vector P as the teaching strategy.

[0028] The teaching strategy library contains corresponding strategy categories, including: supplementing prior knowledge, diagnostic questioning, heuristic questioning, focused explanation, and testing and consolidation; or, it may also contain control parameters, including: the granularity and quantity of prompts, the number of supplemented prior knowledge items, and whether to arrange one or more of the exercises.

[0029] Step 3: Under the constraints of the teaching strategy, combined with the students' cognitive state and the interaction records of the students in the previous dialogue round (i.e., the previous dialogue data), generate teaching content that follows the Socratic teaching style, realize heuristic Q&A, and update the students' historical dialogue data.

[0030] In this embodiment of the invention, step 3 can be executed by a teaching content generation agent. This teaching content generation agent follows a Socratic structure of first asking questions, then providing partial explanations, and then asking questions again to generate personalized teaching content for students. It constructs a progressive dialogue path through continuous heuristic questioning, guided follow-up questions, and partial prompts to achieve heuristic Q&A. The teaching content generation agent integrates a reward model, which uses a multi-dimensional reward evaluation mechanism to automatically score the quality of personalized teaching content, obtain a final reward score, and optimize the teaching content generation agent based on the final reward score.

[0031] In this embodiment of the invention, a progressive dialogue path is constructed through continuous heuristic questioning, guided follow-up questions, and partial prompts to achieve heuristic Q&A. Specifically, when interacting with students, heuristic questions guide students to think independently, and follow-up questions are asked in depth in subsequent interactions. The whole process is a progressive interaction process. Of course, even if students cannot give the correct answer in subsequent interactions, the teaching content generation intelligent experience will provide some prompts rather than directly giving the complete answer.

[0032] During the interaction, one execution process constitutes one round (dialogue round). Taking the teaching of the knowledge point "quadratic function" as an example: In the first round, the teaching strategy selection agent chooses a strategy, such as "diagnostic questioning." The teaching content generation agent then generates teaching content, such as the question, "Do you know what a quadratic function is?" to test the student's basic ability. The real student will give a response; if the student's ability is weak, they may reply, "I don't know what a function is." Then, in the next round, based on the dialogue data from the previous round (in this example, the dialogue data from the first round), the teaching strategy selection agent may choose a new strategy, "precedence knowledge supplementation." The teaching content generation agent then generates teaching content such as "Let's first look at the definition of a function; a function is xxxx" for the student.

[0033] In this embodiment of the invention, the multi-dimensional rewards include: necessary rewards and non-necessary rewards, and each category contains several reward items; the necessary rewards are used to measure the effectiveness and logical correctness of personalized teaching content; the non-necessary rewards are used to evaluate the quality of the personalized teaching content in terms of expression.

[0034] Calculate the score for each reward item in the personalized learning content, and then calculate the final reward score using the following formula: ; Where R is the reward score. The reward score for the j-th necessary reward item. The number of reward items in the necessary reward category. The reward score for the nth non-essential reward item. This refers to the number of non-essential reward items.

[0035] During the training phase, given a student's cognitive state and teaching strategies, a candidate set of teaching content is generated by the teaching content generating agent, denoted as: ,in, For the k-th candidate teaching content, Group size; The final reward score for each candidate teaching content group is obtained using a reward model. And calculate the relative advantage value of the candidate teaching content within the group: ; in, The final reward score for the k-th candidate teaching content is... Let be the group relative advantage value of the k-th candidate teaching content. and These are functions for calculating the mean and standard deviation, respectively. Based on the relative advantage value of candidate teaching content within the group, the parameters of the teaching content generating agent are optimized by combining an algorithm that optimizes agent behavior strategy based on reward function.

[0036] In this embodiment of the invention, algorithms such as Proximal Policy Optimization (PPO), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO) can be used to optimize the agent generating the teaching content by combining the relative advantage values ​​of the candidate teaching content within the group. The relevant optimization process can be implemented with reference to conventional techniques, and will not be elaborated upon in this invention.

[0037] Step 4: After the heuristic Q&A session, combine the updated student historical dialogue data, student cognitive status, and student historical learning records to output after-class exercises suitable for students.

[0038] In this embodiment of the invention, step 4 can be performed by an agent recommended through after-class exercises, as follows: (1) Extract the set of knowledge points F from the updated student historical dialogue data; (2) Based on the knowledge point set F, multiple candidate questions are selected from the question bank by matching. (3) Calculate the similarity between each candidate question and each question in the student's historical learning record M. If the similarity between a candidate question and a question in M ​​is higher than the threshold, the corresponding candidate question is removed. (4) Based on the students' cognitive state, select from the candidate questions eliminated in the previous step and output the students' homework questions.

[0039] The above-mentioned solution provided by the embodiments of the present invention has the following advantages: (1) It forms a closed-loop structure of "explanation-practice-feedback", which improves the learning effect and learning efficiency; (2) It divides the complex teaching function into multiple agents, each agent performs its own duties, which greatly improves the overall interpretability and schedulability, and facilitates the later rule adjustment and strategy optimization; (3) The multi-agent framework is relatively decoupled from the specific large language model training method, and existing open source or commercial large language models can be used directly to achieve rapid deployment; when replacing a more powerful large model in the future, there is no need to change the overall multi-agent structure, and the solution can be upgraded by simply re-adapting the prompts and interfaces of the agents.

[0040] To more clearly demonstrate the technical solution and its effects provided by the present invention, the method provided by the embodiments of the present invention will be described in detail below with reference to specific examples.

[0041] I. Overall Overview of the Plan

[0042] In this embodiment of the invention, heuristic question answering is achieved through a multi-agent framework. It forms a closed-loop structure of explanation-practice-feedback based on four core agents. In addition to the four core agents, a corresponding conversation management module and a front-end interface module are also provided. The conversation management module is used to obtain the student's current teaching dialogue records and historical learning records, manage them uniformly, and provide them to the relevant agents. The front-end interface module is responsible for interacting with the student's end, inputting the teaching content generated by the teaching content generator into the student's end, and receiving mathematical input from the student's end for the teaching content generator.

[0043] Meanwhile, the multi-agent framework also includes a multi-agent interaction and control mechanism to define the calling order and information flow of each agent. For example, at the beginning of each round of dialogue, the cognitive diagnosis agent is called first; based on the cognitive diagnosis results, the teaching strategy selection agent is called; then, the teaching content generation agent is called to generate the teaching content for students in this round; after the teaching is completed, the homework recommendation agent is called to provide homework for students; the multi-round interaction process is recorded for subsequent rounds of cognitive diagnosis and strategy selection. The above mechanism can be implemented by configuring a multi-agent collaborative scheduling module, which uses a configurable strategy graph or finite state machine to achieve adaptive flow control for different teaching scenarios (new knowledge learning, error correction, pre-exam review, etc.).

[0044] In practical applications, each intelligent agent can share one or more large language models, or they can be deployed on different large language models. The remaining modules can also be implemented based on servers or terminal devices with processors and memory.

[0045] like Figure 2 As shown, an example of the overall multi-agent framework for implementing the above method is presented.

[0046] II. Detailed introduction of each intelligent agent.

[0047] 1. Cognitive diagnostic intelligent agent.

[0048] In this embodiment of the invention, a set M is used to represent the student's historical learning records, denoted as: ;in, Information for the i-th homework question may include the question number, the knowledge point to which the question belongs, and the question stem; The answer to the i-th homework question can include the answer and the solution process. To grade the answers to the i-th homework question, The representative answered correctly. The number of incorrect answers represents the number of homework questions the student has already completed.

[0049] In implementation, M can include records of questions that students have answered in the Q&A method provided by this invention, and can also incorporate data from external homework systems and examination systems.

[0050] The output of the cognitive diagnostic agent is a vector. This represents the student's cognitive state, indicating the student's ability or mastery level in several dimensions, such as different knowledge points or different ability dimensions (understanding, application, reasoning, etc.). For the set of real numbers, The state dimension can be designed and determined according to the actual situation, such as by the number of knowledge points or the number of ability dimensions. In actual use, it is not required to... Each dimension is explicitly named, as long as it ensures that subsequent agents (such as agents that select teaching strategies) can use the vector to make decisions.

[0051] In this embodiment, the cognitive diagnostic agent f is used to map M to Formalization The mapping process of the cognitive diagnostic agent can be achieved through any model or algorithm that meets the requirements, including but not limited to: cognitive diagnostic models based on item response theory (IRT); simple scoring models based on knowledge point mastery statistics; vector coding models based on machine learning or large models; and computational models based on rules or expert experience.

[0052] In this embodiment of the invention, since the intelligent agent recommending after-class exercises continuously provides new questions for students to learn, M is continuously updated as students complete new questions or provide new answers. The cognitive diagnostic agent also recalculates new information before the next round of dialogue. This enables dynamic tracking of students' abilities and status.

[0053] 2. Selecting intelligent agents as teaching strategies.

[0054] In educational theory, teaching strategies determine a teacher's teaching methods, the depth of content, and the sequence of the teaching process. Different students, different knowledge points, and different learning stages require different teaching methods. For example: (1) When students have a weak grasp of a certain knowledge point, presenting complex solutions directly often has little effect. In this case, teaching should prioritize supplementing the previous knowledge points before returning to the current problem to explain it.

[0055] (2) When teachers find it difficult to determine whether students’ specific difficulties lie in understanding the meaning of the question or in the calculation steps, they need to design a series of diagnostic questions to accurately locate the learning obstacles through the students’ answers.

[0056] (3) When students have a certain foundation and only hesitate at a certain step, they can be guided to deduce the key steps on their own through heuristic questioning, rather than being given the answer directly.

[0057] Therefore, in the multi-agent question-answering method of this invention, a dedicated teaching strategy selection agent needs to be set up. Specifically, the teaching strategy selection agent is used to dynamically select the corresponding teaching strategy in each round of question-answering based on the student's cognitive state and the student's interaction record in the previous dialogue round, and provide the selected strategy to the teaching content generation agent, so as to reflect the principles of "teaching according to aptitude" and "heuristic teaching" emphasized in pedagogy in actual dialogue.

[0058] During the t-th dialogue, the teaching strategy selection agent receives the following input: (1) The student’s cognitive state is output by the cognitive diagnostic agent before the start of this teaching dialogue (heuristic Q&A).

[0059] (2) Dialogue context (i.e., historical dialogue data): The historical dialogue data at the beginning of this round (including the most recent rounds of interaction with students) is recorded as a text set. ,in It is the output of the student in the u-th round (i.e., the dialogue content output by the student). This refers to the teaching content generated by the intelligent agent in the u-th round of teaching (i.e., the output of this invention). This is achieved through a text encoder. Will The text is mapped to semantic vectors, and then the semantic representation c of the complete dialogue history is obtained. Large language models or small models such as BERT (a bidirectional encoder representation model based on transform neural networks) can be used to process historical dialogue data using methods such as pooling and summation.

[0060] In this embodiment of the invention, the teaching content output by the teaching content generating agent is a piece of text. For example, in the example provided above, "Do you know what a quadratic function is?" is the teaching content. The student's output is the feedback after receiving the teaching content, which is a piece of text. For example, in the example provided above, "I don't know what a function is" is the student's output.

[0061] In this embodiment of the invention, there is no limitation on the processing method of historical dialogue data. It is only necessary to ensure that the obtained semantic representation c can represent the semantic information of the historical dialogue data. As an example, the following method can be used: ; Where m is the dimension of the semantic representation c.

[0062] semantic representation and Concatenate them into a unified feature vector: Suppose there are L candidate teaching strategy categories (labeled by experts), such as: supplementing prior knowledge, diagnostic questioning, heuristic questioning, focused explanation, assessment and reinforcement, etc., and the strategy set (teaching strategy library) is denoted as . Calculate each policy The policy score is then used to determine the probability of selection using the softmax function (normalized exponential function). For the entire teaching strategy base, the strategy score sequence and probability sequence are calculated using the following formula: ; ; Where W is the weight parameter, b is the bias parameter, and s is the policy score vector, where the th... Item for , representing the One strategy The score; P is a probability vector, where the first... Item for , representing the One strategy The probability of being selected; L represents the number of strategies.

[0063] The strategy with the highest probability is selected as the teaching strategy input and used to generate the agent in subsequent teaching content.

[0064] In this embodiment of the invention, the strategy categories include: supplementing prior knowledge, diagnostic questioning, heuristic questioning, focused explanation, and testing and consolidation; or, it may also include control parameters, which include: the granularity and quantity of prompts, the number of supplemented prior knowledge items, and whether to arrange one or any combination of exercises.

[0065] In this embodiment of the invention, the weight parameter W and the bias parameter b are learnable parameters. Manually annotated historical dialogue data can be utilized. ,in They are the first Each data point contains student cognitive states, historical dialogue data, and expert-annotated optimal strategy categories, where U is the total number of data points. The loss function used is cross-entropy. ; in, Indicates a given and Below, the selected teaching strategy is The probability of.

[0066] Furthermore, regression or classification losses with control parameters can be incorporated to form multi-task learning objectives. This invention does not limit the training method; users can choose the method according to their specific circumstances.

[0067] 3. The teaching content generates an intelligent agent.

[0068] (1) The teaching content generates the definition of the intelligent agent.

[0069] The instructional content generation agent is used in each round of Q&A to select teaching strategies provided by the cognitive diagnostic agent based on the student's cognitive state and the teaching strategies provided by the cognitive diagnostic agent, thereby generating individualized instructional content for each student. The teaching content generation agent uses a large language model to generate teaching content that follows the Socratic teaching style to ensure that the Q&A process is inspiring and explainable. It constructs a progressive dialogue path through continuous heuristic questioning, guided follow-up questions, and local prompts.

[0070] As an example, the corresponding teaching content can include: heuristic questions for students; targeted feedback and guidance for students' answers; necessary explanations of concepts, breakdowns of steps, and analogies; and corrections and explanations of students' incorrect thinking.

[0071] (2) Construction of a multidimensional reward model.

[0072] To ensure that the teaching behaviors generated by the intelligent agent align with educational objectives and possess high-quality inspiration, this invention integrates a multi-dimensional reward and evaluation mechanism into the intelligent agent. This mechanism automatically scores the quality of the teaching content generated by the intelligent agent in each round, including both necessary and unnecessary rewards.

[0073] (2.1) Core Rewards: Used to measure the effectiveness and logical correctness of teaching. Mainly includes: Correct / incorrect identification: Can it accurately identify the correct and incorrect points in a student's answer?

[0074] Socratic conformity: Whether it follows the heuristic dialogue structure dominated by "guidance, questioning, and reflection".

[0075] Formatting compliance: Whether the output format conforms to the preset teaching standards.

[0076] (2.2) Unnecessary rewards (Aux): Used to assess the quality of expression, including fluency, clarity and naturalness of expression.

[0077] The aforementioned rewards can be generated by training a reward model using supervised learning. This model can be either a classifier or a regressor. For example, in the "correct / incorrect identification" dimension, real teacher-student dialogue data can be collected. Experts can then label each round of teacher feedback to determine whether the teacher accurately identified the student's error type and key knowledge points, and based on this, construct a binary classification label of "correct identification / incorrect identification" to train a reward classifier. Alternatively, this can be further refined into continuous scores within the [0,1] interval for training a regression-based reward model. For "quality at the expression level," multiple versions of intelligent agent teaching content can be collected, and experts can score each response on three dimensions: "fluency," "clarity," and "naturalness" (e.g., 1–5 points), thereby training a regression-based reward model. Alternatively, a three-category label of "excellent / average / poor" can be constructed for training a categorized reward model. Considering that the methods involved here are all conventional techniques, they will not be elaborated further.

[0078] The final reward score R is calculated using the following formula: necessary rewards are multiplied, and non-necessary rewards are added, to ensure that the model can only achieve a high score when all core teaching competencies are met simultaneously: ; Where R is the reward score. The reward score for the j-th necessary reward item. The number of reward items in the necessary reward category. The reward score for the nth non-essential reward item. This refers to the number of non-essential reward items.

[0079] Based on this reward function, algorithms such as Proximal Policy Optimization (PPO), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO) can be used for training.

[0080] (3) Training method based on group relative policy optimization (GRPO).

[0081] To enable the teaching content generating agent to continuously optimize its strategy, this invention employs the Group Relative Policy Optimization (GRPO) algorithm for training.

[0082] Candidate Trajectory Generation: During the training phase, given a student's cognitive state and teaching strategies, the intelligent learning system generates a set of candidate teaching content. ,in, For the k-th candidate teaching content, Group size.

[0083] Relative advantage calculation: Input all candidate teaching content into the above reward model to obtain the final reward score. Calculate the relative advantage value within the group to eliminate the effect of variance in the reward distribution: ; in, The final reward score for the k-th candidate teaching content is... Let be the group relative advantage value of the k-th candidate teaching content. and These are functions for calculating the mean and standard deviation, respectively.

[0084] Subsequently, based on the relative advantage of candidate teaching content within the group, the parameters of the teaching content generating agent can be optimized using an algorithm that optimizes agent behavior strategies based on reward functions.

[0085] 4. Recommended exercises for after class: intelligent agents.

[0086] In this embodiment of the invention, the after-class exercise recommendation agent is used to provide students with personalized reinforcement exercises after the current teaching dialogue is completed, realizing a closed-loop teaching system of classroom simulation, explanation, practice, and feedback. Its core tasks include: extracting knowledge points from students' historical dialogue data, and searching the question bank for reinforcement questions that are different from the questions that have already appeared and of suitable difficulty to provide to students.

[0087] (1) Input and output definitions.

[0088] The inputs to the intelligent agent that recommends after-class exercises include: the student's cognitive state, output by the cognitive diagnosis intelligent agent; the student's historical dialogue data D, obtained through continuous interaction between the student and the teaching content generating intelligent agent; and the student's historical learning records M, used to avoid generating duplicate questions and dynamically adjust the difficulty and type of questions.

[0089] The output is a set of questions. Each question includes the question stem, question type, answer, reference analysis, and corresponding knowledge points.

[0090] (2) Knowledge point extraction.

[0091] The recommended intelligent agent first focuses on the teaching dialogue. The analysis was performed, and the set of knowledge points F involved was extracted: ; in, The knowledge point extraction process can be implemented in the following ways: using large language models or small classification / named entity recognition models to map student questions, teaching content, and teacher explanations to knowledge point tags; or, using keyword matching, semantic embedding vector similarity, or graph query methods to associate text content with a predefined set of knowledge points.

[0092] (3) Title search and deduplication.

[0093] When searching for questions, the question recommendation agent needs to ensure that the question set Q does not overlap with the questions in the student's historical learning records M, and that it covers the knowledge points in F. The generation process includes: Candidate question search: Based on the knowledge point set F, multiple candidate questions are selected from the question bank through matching.

[0094] Deduplication filtering: Calculate the similarity between candidate questions and each question in the student's historical learning record M (such as the cosine similarity of the question stem text vector), and exclude duplicate or similar questions.

[0095] Difficulty Adjustment: Establishing a Predictive Model for Student Answers .Model For a question Text information Encode the question to obtain its semantic vector. Combined with students' cognitive state Predicting the student's state regarding the question Correct answer rate The calculation process is as follows: ; in, These are the prediction weight parameters and the prediction bias parameters. It's a vector concatenation operation. It uses the sigmoid function. This model can be pre-trained using real student answer data (i.e., historical learning records M) to obtain question representations. The model parameters are fixed. When the agent recommends homework questions, for each candidate question, this model can predict the probability of a student answering correctly under their current cognitive state. Questions with excessively high correct answers (students have already mastered the material and there's no need for repeated practice) or excessively low correct answers (the questions are too difficult and exceed the student's ability) are filtered out. The levels of excessively high and low probabilities can be controlled by setting maximum and minimum threshold values. For example, [0.2, 0.5] can be used as an interval for advanced questions, and [0.5, 0.8] as an interval for basic reinforcement questions. In this example, 0.2 is the minimum threshold value; a correct answer probability less than the minimum threshold value indicates an excessively low correct answer probability. 0.8 is the maximum threshold value; a correct answer probability greater than the maximum threshold value indicates an excessively high correct answer probability.

[0096] Final Problem Recommendation: Output the final problem set Q for students to practice and consolidate their knowledge.

[0097] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0098] Example 2 This invention also provides a heuristic question-answering system based on multi-agent interaction, which is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 3 As shown, the system is configured with a multi-agent framework, including: a cognitive diagnosis agent, an instructional strategy selection agent, an instructional content generation agent, and a homework recommendation agent; among which: A cognitive diagnostic agent is used to perform cognitive diagnosis using students' historical learning records and output the students' cognitive status; the historical learning records include students' answers to after-class exercises. The teaching strategy selection agent is used to combine students' cognitive state and students' historical dialogue data, and output teaching strategies according to the preset teaching strategy library and strategy decision rules. The teaching content generating agent is used to generate teaching content that follows the Socratic teaching style under the constraints of teaching strategies, combined with students' cognitive state and students' previous dialogue data, to achieve heuristic Q&A and update students' historical dialogue data. The intelligent agent that recommends homework assignments is used to output homework assignments suitable for students after the heuristic Q&A session, by combining updated historical dialogue data, students' cognitive status, and students' historical learning records.

[0099] Since the main technical details involved in each intelligent agent have been described in detail in the previous embodiments, they will not be repeated here.

[0100] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0101] Example 3 The present invention also provides a processing device, such as Figure 4 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0102] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0103] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example: Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc. The output device can be a display terminal; The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0104] Example 4 The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0105] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0106] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A heuristic question-answering method based on multi-agent interaction, characterized in that, Configure a multi-agent framework, where each step is executed by a separate agent; the steps include: Step 1: Conduct cognitive diagnosis using students' historical learning records and output students' cognitive status; the historical learning records include students' answers to after-class exercises. Step 2: Based on students' cognitive state and their historical dialogue data, and according to the pre-set teaching strategy library and strategy decision rules, output teaching strategies. Step 3: Under the constraints of teaching strategies, combined with students' cognitive state and the students' previous dialogue data, generate teaching content that follows the Socratic teaching style, realize heuristic Q&A, and update students' historical dialogue data. Step 4: After the heuristic Q&A session, combine the updated student historical dialogue data, student cognitive status, and student historical learning records to output after-class exercises suitable for students.

2. The heuristic question-answering method based on multi-agent interaction according to claim 1, characterized in that, Step 1 is executed by a cognitive diagnostic agent, and is represented as follows: Where f is the cognitive diagnostic agent, and M is the student's historical learning record. For students' cognitive state; The student's history learning record M is represented as: ;in, This is the information for the i-th homework question. This represents the student's answer to the i-th homework question. To grade the answers to the i-th homework question, The representative answered correctly. The number of incorrect answers represents the number of homework questions the student has already completed.

3. The heuristic question-answering method based on multi-agent interaction according to claim 1, characterized in that, Step 2 involves selecting an agent to execute a teaching strategy, and the process is as follows: The text encoder maps students’ historical dialogue data into semantic vectors to obtain the semantic representation c of the historical dialogue data. The semantic representation c of historical dialogue data is combined with the students' cognitive state. Concatenate them into a unified feature vector x; The pre-set teaching strategy library is denoted as The teaching is output through the following strategic decision-making rules: ； ； Where W is the weight parameter, b is the bias parameter, and s is the policy score vector, where the th... Item represents the first One strategy The score; P is a probability vector, where the first... Item represents the first One strategy The probability of being selected; L represents the number of strategies; Select the strategy with the highest probability from the probability vector P as the teaching strategy; The teaching strategy library contains corresponding strategy categories, including: supplementing prior knowledge, diagnostic questioning, heuristic questioning, focused explanation, and testing and consolidation; or, it may also contain control parameters, including: the granularity and quantity of prompts, the number of supplemented prior knowledge items, and whether to arrange one or more of the exercises.

4. The heuristic question-answering method based on multi-agent interaction according to claim 1, characterized in that, Step 3 is executed by a teaching content generating agent. This teaching content generating agent follows a Socratic structure of asking questions first, then providing partial explanations, and then asking questions again. It generates personalized teaching content for students and constructs a progressive dialogue path through continuous heuristic questioning, guided follow-up questions, and partial prompts to achieve heuristic answering. The teaching content generating agent integrates a reward model, which uses a multi-dimensional reward evaluation mechanism to automatically score the quality of personalized teaching content, obtain a final reward score, and optimize the teaching content generating agent based on the final reward score.

5. The heuristic question-answering method based on multi-agent interaction according to claim 4, characterized in that, The aforementioned multi-dimensional reward evaluation mechanism automatically scores the quality of personalized teaching content, and the final reward score includes: The multi-dimensional rewards include: necessary rewards and non-necessary rewards, with each category containing several reward items; the necessary rewards are used to measure the effectiveness and logical correctness of personalized teaching content; the non-necessary rewards are used to evaluate the quality of the personalized teaching content in terms of expression. Calculate the score for each reward item in the personalized learning content, and then calculate the final reward score using the following formula: ； Where R is the reward score. The reward score for the j-th necessary reward item. The number of reward items in the necessary reward category. The reward score for the nth non-essential reward item. This refers to the number of non-essential reward items.

6. The heuristic question-answering method based on multi-agent interaction according to claim 4, characterized in that, The optimization of the teaching content generation agent based on the final reward score includes: During the training phase, given a student's cognitive state and teaching strategies, a candidate set of teaching content is generated by the teaching content generating agent, denoted as: ,in, For the k-th candidate teaching content, Group size; The final reward score for each candidate teaching content group is obtained using a reward model. And calculate the within-group relative advantage value of the candidate teaching content: ； in, The final reward score for the k-th candidate teaching content is... Let be the group relative advantage value of the k-th candidate teaching content. and These are functions for calculating the mean and standard deviation, respectively. Subsequently, based on the relative advantage value of candidate teaching content within the group, the parameters of the teaching content generating agent are optimized using an algorithm that optimizes agent behavior strategy based on reward function.

7. The heuristic question-answering method based on multi-agent interaction according to claim 1, characterized in that, Step 4 involves recommending the agent to perform tasks based on the homework assignments. The process is as follows: Extract the set of knowledge points F involved from the updated student historical dialogue data; Based on the knowledge point set F, multiple candidate questions are selected from the question bank through matching. Calculate the similarity between each candidate question and each question in the student's history learning record. If the similarity between a candidate question and a candidate question in the history learning record is higher than a threshold, the corresponding candidate question is removed. Based on the students' cognitive level, the candidate questions eliminated in the previous step are selected and the students' homework questions are output.

8. A heuristic question-answering system based on multi-agent interaction, characterized in that, For implementing the method according to any one of claims 1 to 7, the system is configured with a multi-agent framework, including: a cognitive diagnosis agent, a teaching strategy selection agent, a teaching content generation agent, and a homework recommendation agent; wherein: A cognitive diagnostic agent is used to perform cognitive diagnosis using students' historical learning records and output the students' cognitive status; the historical learning records include students' answers to after-class exercises. The teaching strategy selection agent is used to combine students' cognitive state and students' historical dialogue data, and output teaching strategies according to the preset teaching strategy library and strategy decision rules. The teaching content generating agent is used to generate teaching content that follows the Socratic teaching style under the constraints of teaching strategies, combined with students' cognitive state and students' previous dialogue data, to achieve heuristic Q&A and update students' historical dialogue data. The intelligent agent that recommends homework assignments is used to output homework assignments suitable for students after the heuristic Q&A session, by combining updated historical dialogue data, students' cognitive status, and students' historical learning records.

9. A processing device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images