AI intelligent education tutoring system based on thinking chain and retrieval enhancement generation technology

Abstract

The invention relates to the technical field of artificial intelligence education, and discloses an AI intelligent education tutoring system based on a thinking chain and a retrieval enhancement generation technology. Transparent generation and dynamic verification of a problem solving logic chain are realized through a CoT-RAG fusion architecture, and the interdisciplinary confusion rate is reduced in combination with a knowledge graph technology of subject boundary isolation; the non-question-bank-dependent generation engine creates questions in real time based on semantic association, and the coverage rate is increased; federated learning and differential privacy technologies are integrated, and the safety of user data is ensured while the generalization ability of the model is improved; the learning behavior analysis unit dynamically optimizes a knowledge recommendation path through time sequence modeling, and cooperates with the interactive logic correction module to form a self-evolution closed loop. Finally, an intelligent education system which is high in interpretability, high in professional precision and compliant in privacy is constructed, and the learning efficiency and the interdisciplinary problem solving capability are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence education technology, and specifically to an AI intelligent education tutoring system based on thought chain and retrieval enhancement generation technology. Background Art

[0002] The AI tutoring system is an educational tool that uses artificial intelligence technologies (such as machine learning, natural language processing, and computer vision) to provide personalized learning support. It analyzes students' learning data and behaviors, dynamically adjusting teaching content and methods to help students master knowledge more efficiently.

[0003] However, current AI-powered tutoring systems have significant flaws. Traditional large language models (such as ChatGPT) are unable to demonstrate problem-solving thought chains, resulting in a logical black box. The cross-disciplinary answer confusion rate exceeds 40%, resulting in insufficient professional precision. Furthermore, they rely heavily on question banks, making it difficult to dynamically generate adaptive questions. The technical root cause lies in the lack of dynamic integration between the Chain of Thought (CoT) problem-solving module and Retrieval-Augmented Generation (RAG), resulting in unexplainable reasoning processes and redundant and inefficient knowledge retrieval. Therefore, we propose an AI-powered tutoring system based on Chain of Thought and Retrieval-Augmented Generation technologies. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides an AI intelligent education tutoring system based on thought chain and retrieval enhancement generation technology to solve the background technology problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an AI intelligent education and tutoring system based on thought chaining and retrieval enhancement generation technology, comprising the following collaborative modules:

[0006] The user interaction module is used to receive homework, test questions, and natural language questions input by users, pass the questions to downstream modules for processing, and finally output feedback information containing an explainable reasoning chain;

[0007] The CoT-RAG fusion architecture module is directly connected to the user interaction module. It uses thought chain technology to break down received questions into multi-level logical steps. It also drives the Retrieval Enhanced Generation (RAG) mechanism to extract relevant information from multi-source heterogeneous data and generate a transparent logical chain for solving problems.

[0008] The knowledge graph module, which isolates subject boundaries, builds an interdisciplinary knowledge graph and receives the decomposition results from the CoT-RAG fusion architecture module. Through the built-in semantic reinforcement model and subject classifier, it dynamically identifies and isolates interdisciplinary interference items, providing knowledge-related data with professional accuracy for downstream generation modules.

[0009] The non-question bank-dependent generation engine generates adaptive exercises and analysis content in real time based on the semantic association data output by the knowledge graph module and the user's historical learning level. The generation process is independent of the pre-existing question bank and the generated content is fed back to the user interaction module for display;

[0010] The interactive logic correction module is bidirectionally connected to the user interaction module and the CoT-RAG fusion architecture module, allowing users to modify the reasoning path by clicking on the logic node, triggering the system to regenerate the corrected answer and synchronize the corrected data to the knowledge graph module and learning behavior analysis unit;

[0011] The learning behavior analysis unit receives behavioral data from the user interaction module and the logic correction module in real time, and dynamically adjusts the retrieval priority of the knowledge graph module and the disassembly granularity of the CoT-RAG fusion architecture module through time series modeling to form a closed-loop optimization link.

[0012] Preferably, the CoT-RAG fusion architecture module includes:

[0013] The symbolic logic reasoning unit abstracts the questions passed in by the user interaction module into formal logical expressions and generates a preliminary reasoning framework;

[0014] The neural network completion unit fills in the logical details based on the Transformer architecture, annotates the confidence level of each reasoning step, and passes the results to the knowledge graph module isolated by subject boundaries for verification;

[0015] The dynamic verification interface calls the semantic relationship data of the knowledge graph module to verify the factual basis of the logical node in real time.

[0016] Preferably, the knowledge graph module with subject boundary isolation achieves cross-disciplinary interference suppression in the following ways:

[0017] Define discipline-specific semantic relationship constraints to limit the direct association of cross-disciplinary knowledge points;

[0018] Use graph embedding technology to map different subject knowledge points into independent vector spaces to reduce the risk of semantic confusion;

[0019] Set up dynamic filters to automatically cut off paths with interdisciplinary correlations below a threshold during the retrieval process, ensuring that the data input into the non-question bank-dependent generation engine meets the subject professional requirements.

[0020] Preferably, the non-question bank dependent generation engine includes:

[0021] The semantic relevance calculation unit generates the question trunk based on the node relevance strength output by the knowledge graph module;

[0022] The difficulty control unit dynamically adjusts the complexity of question parameters based on the user's historical answer accuracy rate provided by the learning behavior analysis unit;

[0023] The parsing generator converts the reasoning steps of the CoT-RAG fusion architecture module into natural language parsing and embeds hyperlinks to related knowledge points provided by the knowledge graph module.

[0024] Preferably, the interactive logic correction module supports:

[0025] Perform logical addition, deletion, and modification operations on any node in the thinking chain generated by the CoT-RAG fusion architecture module;

[0026] Automatically generate a logical comparison view before and after the correction, highlight the difference nodes, and feed the correction results back to the knowledge graph module to update the semantic association;

[0027] A logical error pattern library is constructed based on the correction records to optimize the subsequent reasoning path generation.

[0028] Preferably, the learning behavior analysis unit realizes dynamic adjustment in the following manner:

[0029] Establish a three-dimensional time series matrix of users, knowledge points, and logical errors, and associate the behavioral data of the user interaction module with the feedback information of the logic correction module;

[0030] Use the LSTM model to predict user knowledge weaknesses and adjust the retrieval priority of the knowledge graph module;

[0031] Automatically simplify or refine the inference chain display granularity of the CoT-RAG fusion architecture module based on the user logic modification frequency.

[0032] Preferably, the system further includes a privacy protection mechanism:

[0033] Differential privacy technology is used to inject noise into the behavioral data uploaded by the user interaction module;

[0034] The federated learning framework is used to support the collaborative training of local models by multiple terminal devices, uploading only the desensitized model parameters to the central server and updating them in collaboration with the knowledge graph module.

[0035] Preferably, the knowledge graph module establishes a dynamic connection channel with the standardized education question database, including:

[0036] Question bank change monitoring interface to capture new knowledge points and question type changes in real time;

[0037] An automated semantic alignment unit that maps new content to existing knowledge graph nodes;

[0038] The version control module maintains the knowledge graph branches of different textbook versions.

[0039] Preferably, when calling the knowledge graph module, the dynamic verification interface synchronously triggers the subject classifier to predict the subject attribution probability of the retrieval results, and dynamically adjusts the verification threshold of the semantic reinforcement model according to the prediction results.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] This invention uses a CoT-RAG fusion architecture to transparently generate and dynamically verify problem-solving logic chains. This technology, combined with knowledge graph technology that isolates subject boundaries, reduces interdisciplinary confusion. A non-question bank-dependent generation engine creates questions in real time based on semantic associations, improving coverage. The integration of federated learning and differential privacy technologies enhances model generalization while ensuring user data security. A learning behavior analysis unit dynamically optimizes knowledge recommendation paths through temporal modeling, collaborating with an interactive logic correction module to form a self-evolving closed loop. Ultimately, this system creates an intelligent education system with strong interpretability, high professional precision, and privacy compliance, significantly improving learning efficiency and interdisciplinary problem-solving capabilities.

[0042] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a block diagram of the AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology of the present invention;

[0044] Figure 2 FIG. 4 is a flow chart of the CoT-RAG fusion architecture of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this technical field without making any creative efforts shall fall within the scope of protection of the present invention.

[0046] See also Figure 1 and Figure 2 The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology in this invention is composed of six core modules. Through the close collaboration of data flow and control flow, a complete intelligent education and tutoring closed loop is formed. The interaction logic of each module is as follows:

[0047] The user interaction module receives multimodal input (text, image, voice), parses it and passes it to the CoT-RAG fusion architecture module;

[0048] The CoT-RAG module breaks down the problem into a multi-level logical chain and calls the knowledge graph module to verify the semantic association;

[0049] After filtering out interdisciplinary interference, the knowledge graph module provides professional data to the non-question bank-dependent generation engine;

[0050] The generation engine generates questions and analyses in real time, and feeds them back to the user interaction module for display;

[0051] Users modify the logic chain through the interactive logic correction module, and the correction data will synchronously update the knowledge graph and learning behavior analysis unit;

[0052] The learning behavior analysis unit dynamically adjusts system parameters to form a closed-loop optimization link;

[0053] Privacy protection mechanisms run through all data transmission processes to ensure data security.

[0054] 1. Module functions and implementation details

[0055] 1. User interaction module

[0056] Support multimodal input (text, voice, image) and emotion recognition;

[0057] The output contains feedback information of the interpretable reasoning chain;

[0058] Input processing:

[0059] a. Text input is parsed into structured questions through natural language understanding (NLU), and semantic features are extracted using the ERT model, which is expressed as:

[0060]

[0061] Among them, h [CLS] This is the sentence aggregation vector output by the BERT model, with a dimension of 768. The input text is encoded as a sequence of word vectors. After encoding through a multi-layer Transformer, the vector of the [CLS] tag is taken as the semantic representation of the sentence.

[0062] Example:

[0063] Biological problem: The user inputs "analyze the characteristics of the light reaction stage in photosynthesis", and BERT extracts key semantics ("biological process", "energy conversion").

[0064] Chemistry question: A user uploads a diagram of a carbon dioxide gas experimental setup, and after OCR recognition, the text "How to prepare carbon dioxide gas?" is generated.

[0065] b. Image input (such as a picture of a mathematical formula) is converted into text using OCR technology. OCR recognizes the mathematical formula picture as LaTeX code:

[0066]

[0067] Where I is the input image matrix; s is the candidate character sequence (such as the recognition result is "H2O→H2+O2").

[0068] c. Voice input is transcribed into text through ASR (automatic speech recognition).

[0069] d. Emotion Recognition:

[0070] Analyze user voice intonation or text sentiment words (such as "too difficult" and "I don't understand") and use the LSTM model to predict sentiment labels:

[0071] y emotion =softmax(W e ·h T +b e )

[0072] Among them, h T is the hidden state of the last time step of LSTM (dimension 256); is the sentiment classification weight matrix (3 categories: positive / neutral / negative); b e is the bias term.

[0073] The specific steps are:

[0074] The parsed structured data (such as text semantic vectors and formula LaTeX) is passed to the CoT-RAG module;

[0075] Emotional tags trigger the system to adjust feedback strategies (such as adding guidance steps when there is negative emotion).

[0076] 2. CoT-RAG Converged Architecture Module

[0077] Break down the problem into multi-level logical steps and drive dynamic retrieval to generate answers.

[0078] Submodule process:

[0079] a. Symbolic logic reasoning unit:

[0080] Example 1:

[0081] Input: User questions such as "Prove the Pythagorean theorem";

[0082] Processing: Abstract the problem into a formal logical expression, such as " a 2 +b 2 =c 2 ”;

[0083] Output: Generate a preliminary reasoning framework, "Define a right triangle → Derive area relationships → Apply algebraic identities";

[0084] Example 2:

[0085] Input: User questions (e.g., “Analyze the impact mechanism of climate change on agricultural yields”);

[0086] Processing: Convert the problem into a causal chain of events;

[0087] Output: Generate a preliminary reasoning framework: greenhouse gas increase → temperature rise → crop growth cycle changes → yield fluctuations.

[0088] b. Neural network completion unit:

[0089] Input: symbolic logic framework;

[0090] Processing: Fill in the details based on the Transformer model (such as adding "proving equal areas by puzzle method"); expressed as:

[0091]

[0092] Among them, Q, K, V are query, key, and value matrices (dimension 768×64); d k =64 is the scaling factor to prevent gradient explosion.

[0093] Output: Complete logic chain, with the confidence level of each step marked (e.g., “Step 3 confidence level 92%”).

[0094] c. Dynamic verification interface:

[0095] Input: logical chain nodes (e.g. "The Pythagorean theorem applies to all right triangles");

[0096] Processing: Call the knowledge graph module to verify the correctness of the facts (search for counterexamples); expressed as:

[0097] Score(e)=PageRank(e)×cos(v e , v q )

[0098] Among them, e is the entity in the knowledge graph; v e , v q are the embedding vectors of entities and queries.

[0099] Output: Marks a node as verified (green) or questionable (red).

[0100] The specific steps are:

[0101] The symbolic logic framework is passed to the knowledge graph module for semantic verification;

[0102] The verification results are returned to the neural network completion unit to correct the confidence (such as marking low-confidence nodes in red).

[0103] 3. Knowledge graph module with discipline boundary isolation

[0104] Build an interdisciplinary knowledge graph and dynamically filter interdisciplinary interference.

[0105] a. Graph construction:

[0106] Data sources: standardized education question banks, open course resources, and academic papers;

[0107] Node definition: knowledge points, question types;

[0108] Relationship definition: "belonging to the discipline", "prerequisite";

[0109] Use the Node2Vec algorithm to generate subject-specific embeddings:

[0110]

[0111] Where f(u) is the embedding vector of node u (dimension 128); N S (u) is the set of neighbor nodes sampled by random walk;

[0112] Biology knowledge point: "Photosynthesis" is strongly associated with "chloroplasts", but has no direct connection with "circuit diagrams".

[0113] b. Discipline isolation mechanism:

[0114] Semantic constraint rules: limit direct cross-disciplinary connections (e.g., "chemical equation" is not directly connected to "geometry proof");

[0115] Graph embedding technology: Mapping mathematical knowledge points to vector space A and physical knowledge points to space B, reducing the risk of confusion;

[0116] Dynamic filter: When searching, cut off interdisciplinary paths with a correlation degree of less than 0.7 (for example, blocking "Pythagorean Theorem → Chemical Reaction Rate"), expressed as:

[0117]

[0118] Where s is the path correlation score (0-1); τ filter=0.7 is the filtering threshold.

[0119] The specific steps are:

[0120] Receive the disassembly results of the CoT-RAG module (such as "photosynthesis") and retrieve related knowledge points;

[0121] The filtered data is passed to the non-question bank generation engine to ensure the professionalism of the questions.

[0122] 4. Non-question bank dependent generation engine

[0123] Generate questions and analysis adapted to the user's level in real time.

[0124] Submodule process:

[0125] a. Semantic relevance calculation unit:

[0126] Input: knowledge graph node;

[0127] Processing: Calculate the cosine similarity with the associated nodes, expressed as:

[0128]

[0129] Among them, v i ·v j Embed vectors for knowledge points;

[0130] Output: Generate the main body of the question;

[0131] Example:

[0132] Physical connection: Generate questions that combine "Newton's laws" and "friction".

[0133] b. Difficulty Control Unit:

[0134] Input: User historical accuracy rate;

[0135] Processing: Adjust the question parameters based on the IRT model, expressed as:

[0136]

[0137] Among them, θ is the user ability value (0-1); β is the difficulty of the question; α is the discrimination of the question;

[0138] Output: Generate question variants;

[0139] Example:

[0140] Scientific and technological literature topic: If the user's classical Chinese comprehension ability θ = 0.6, a medium-difficulty "Translation of Selections from Basic Concepts of Quantum Mechanics" will be generated.

[0141] Parsing generator:

[0142] Input: logical chain of CoT-RAG;

[0143] Processing: Convert symbolic logic into natural language, expressed as:

[0144] Parsing = GPT-3.5 (step 1, step 2, knowledge point link)

[0145] Among them, step 1 and step 2 are specific reasoning steps in the logic chain generated by the CoT-RAG module.

[0146] Output: embedded knowledge point hyperlink;

[0147] The specific steps are:

[0148] Receive semantic association data from the knowledge graph, generate questions, and then feed them back to the user interaction module;

[0149] The user's answer results are transmitted to the learning behavior analysis unit to update the ability value θ.

[0150] Technical implementation:

[0151] Semantic relevance is calculated using cosine similarity;

[0152] Difficulty control is based on the Item Response Theory (IRT) model;

[0153] Parsing and generation use the template engine + GPT-3.5 hybrid mode.

[0154] 5.Interactive logic correction module

[0155] Allows users to modify the logic chain and optimize the system reasoning path.

[0156] a. Node editing:

[0157] The user deletes the incorrect node (e.g., "Assume all metals conduct electricity") and inserts the correct logic (e.g., "Mercury conducts electricity in liquid state");

[0158] b. Comparison view generation:

[0159] The system highlights the difference nodes before and after the modification (for example, the newly added nodes are marked in blue). The difference analysis algorithm is expressed as follows:

[0160] Δ=Diff(Chain old ,Chain ncw )

[0161] Among them, Chain old Chain is the graph structure of the original logical chain; ncw is the corrected logic chain;

[0162] Automatically evaluate the confidence change of the corrected logic chain (from 85% to 93%);

[0163] Example:

[0164] Environmental Science Correction: User corrected "Global warming directly causes glacier disappearance" to "Global warming is the main cause of accelerated glacier melting."

[0165] c. Data Feedback:

[0166] Synchronize the revised data to the knowledge graph;

[0167] Update the knowledge graph weight, expressed as:

[0168]

[0169] Among them, w ij is the association weight between knowledge points i and j; learning rate λ = 0.1.

[0170] The specific steps are:

[0171] The corrected data is synchronized to the knowledge graph module (updating the association weights) and the learning behavior analysis unit (recording error patterns).

[0172] 6. Learning Behavior Analysis Unit

[0173] Analyze user learning behavior data in real time and dynamically optimize knowledge graph retrieval priority and logical chain display granularity.

[0174] a. Construction of three-dimensional time series matrix:

[0175] Dimension 1: user ID;

[0176] Dimension 2: knowledge points;

[0177] Dimension 3: Logical error type;

[0178] Construct the user-knowledge-error tensor:

[0179] τ[u,k,e] = the number of times user u makes error e on knowledge point k

[0180] Example:

[0181] Chemical error statistics: Users frequently make errors in Chemical Equilibrium Calculation, which are marked as high-frequency error types.

[0182] b.LSTM prediction model:

[0183] Input: the user's most recent N learning records (such as answer records, correction operations, and duration of stay);

[0184] Model structure:

[0185] h t =LSTM(x t ,h t-1 )

[0186] Among them, x t is the feature vector of time step t (such as knowledge point ID, error type, and answering time); h t is the hidden state (dimension 128)

[0187] Output: Predict the probability of error at knowledge point k in the future:

[0188]

[0189] in, is the classification weight matrix; b p is the bias term;

[0190] Action: Improve the retrieval weight of relevant knowledge points in the knowledge graph.

[0191] C. Dynamic adjustment strategy:

[0192] Knowledge graph retrieval priority:

[0193] If it is predicted that the user has a high probability of making mistakes in the knowledge point k in the future, then its retrieval weight in the knowledge graph is increased:

[0194]

[0195] in, is the original retrieval weight of the knowledge point; α is the learning rate; p crror (k) is the error probability prediction value of knowledge point k; Retrieve weights for updated knowledge points.

[0196] Logic chain display granularity:

[0197] Automatically adjust the level of detail of the logic chain based on the frequency of user corrections:

[0198]

[0199] If Granularity > 30%, refine the display (such as splitting steps);

[0200] If Granularity < 10%, simplify the presentation (such as merging steps).

[0201] The specific steps are:

[0202] Input: Receives the correction records from the interactive logic correction module and the answer data from the user interaction module;

[0203] Output:

[0204] Send retrieval weight adjustment instructions to the knowledge graph module;

[0205] Sends a logic chain granularity adjustment instruction to the CoT-RAG module.

[0206] 7. Privacy protection mechanism

[0207] Ensure the security and compliance of user data during transmission, storage, and processing.

[0208] a. Differential privacy processing:

[0209] Noise injection: Add Laplace noise to user behavior data (such as answer records and correction operations), expressed as:

[0210]

[0211] in, is the data after adding noise (masked data); x is the original user data (such as answer records, correction operations); Lap is the Laplace distribution used to generate noise; Δf is the sensitivity (the maximum impact of a single data point, set to 1); = 0.1: privacy budget (controls noise intensity);

[0212] The desensitized data is uploaded to the central server;

[0213] Example:

[0214] Original data: "User B made 2 mistakes in the knowledge point 'Chemical Equilibrium'" → desensitized to: "User ID=3B4F made 2 mistakes in Subject 3-Knowledge Point 12".

[0215] b. Federated Learning Framework:

[0216] Local training: Each terminal device trains the subject classifier model locally, protecting the original data from leaving the device, expressed as:

[0217]

[0218] in, is the local data set of terminal i; θ global is the global model parameter;

[0219] Parameter aggregation:

[0220] The central server aggregates the parameters of each terminal:

[0221]

[0222] Among them, n i is the data volume of terminal i; n = ∑ni is the total data volume; is the local model parameter of the i-th terminal device; N is the total number of terminal devices participating in federated learning.

[0223] c. Edge computing architecture:

[0224] Sensitive data (such as voice input and images) is processed on the local device, and only the desensitized metadata is uploaded to the server.

[0225] The specific steps are:

[0226] User interaction module: The original data is processed with differential privacy and then passed to other modules;

[0227] Knowledge graph module: The parameters updated by federated learning are synchronized to the semantic reinforcement model.

[0228] 8. Dynamic connection between knowledge graph and question bank

[0229] Synchronize changes to external question banks to the knowledge graph in real time to ensure content timeliness.

[0230] a. Change the monitoring interface:

[0231] Technical implementation: Monitor the change events of the question bank database (such as adding, deleting, and modifying questions) through Webhook;

[0232] Trigger condition: When a new question is detected in the question bank, structured data (such as question stem, knowledge point labels, and difficulty coefficient) is automatically captured.

[0233] b. Semantic alignment unit:

[0234] Mapping new knowledge points to existing graph nodes (e.g. linking "Probability Distribution" to the "Statistics" branch) is represented as:

[0235] Sim BERT (q,k)=cos(h q ,h k )

[0236] Among them, h q is the BERT encoding vector of the newly added question; h k is the embedding vector of knowledge graph node k;

[0237] Alignment rules:

[0238] If Sim BERT ≥0.8, then map to the existing node;

[0239] If Sim BERT <0.8, create a new node.

[0240] c. Version Control:

[0241] Branch management: maintain knowledge graph branches for different textbook versions (such as the People's Education Press edition and the Beijing Normal University edition);

[0242] Switching logic: Automatically load the corresponding version according to the user's region:

[0243]

[0244] The specific steps are:

[0245] Non-question bank generation engine: Added a new question trigger engine to generate adapted questions (such as "generate exercises based on knowledge points of the People's Education Press edition");

[0246] User interaction module: displays version switching options, and users can manually select the textbook version.

[0247] 9. Dynamic verification interface adjustment mechanism

[0248] Dynamically optimize the knowledge graph retrieval verification threshold based on subject classification results to ensure the professionalism of the answers.

[0249] a. Subject classifier prediction:

[0250] Model architecture: Lightweight classification model based on FastText:

[0251]

[0252] Among them, P(c|x) is the probability that the input text x belongs to subject c; c ′ are all possible subject categories; x is the bag-of-words model vector of the input text; w c is the classification weight of subject c;

[0253] Example:

[0254] Input "Calculate the focus of the parabola" and output the mathematical probability P(mathematics) = 0.95.

[0255] b. Dynamic adjustment of threshold:

[0256] The PID control algorithm is used to dynamically adjust the verification threshold τ according to the actual confusion rate val :

[0257]

[0258] Wherein, e(t) is the current error defined as e(t) = target confusion rate - actual confusion rate (for example, the target confusion rate is set to 10%); K p , K i , K d are proportional, integral and differential coefficients respectively; ∑e(t) is the cumulative sum of historical errors; is the error change rate.

[0259] Adjustment logic:

[0260] If the actual confusion rate is higher than the target value (e.g. 15%), the threshold is raised to filter more strictly;

[0261] If the actual confusion rate is lower than the target value (eg, 5%), the threshold is lowered to increase the search range.

[0262] The specific steps are:

[0263] Knowledge graph module: receives the adjusted threshold and dynamically filters low-correlation paths;

[0264] CoT-RAG module: The verification interface re-evaluates the logical node confidence level according to the new threshold.

[0265] 2. System application scenario examples

[0266] Example 1: Tutoring on mechanics problems combining physics and mathematics

[0267] Scenario description: High school students need to solve a mechanics problem involving the movement of an object on an inclined plane. The system demonstrates how to coordinate multiple modules to provide precise guidance.

[0268] 1. User input:

[0269] Method: Voice question "How to calculate the acceleration of the wooden block on the inclined plane?"

[0270] deal with:

[0271] The user interaction module transcribes it into text through ASR and triggers emotion recognition (detecting the user's anxious tone).

[0272] 2. Disassembly of CoT-RAG module:

[0273] Symbolic logic reasoning unit generation framework:

[0274] Step 1: Force Analysis → Step 2: Newton's Second Law → Step 3: Algebraic Calculation

[0275] Neural network completion unit filling details:

[0276] Step 1: Decompose the force of gravity into its components F parallel to and perpendicular to the inclined plane || =mgsinθ;

[0277] Step 2: Apply the formula

[0278] 3. Knowledge graph verification:

[0279] The subject classifier detects that the problem belongs to physics (with a probability of 0.96), filtering out the interference of the mathematical knowledge point "trigonometric function image";

[0280] The dynamic filter cuts off paths with correlation less than 0.7 (such as “slope angle → trigonometric function periodicity”).

[0281] 4. Non-question bank generation engine:

[0282] Semantic association calculation: associate "incline motion" with "friction calculation";

[0283] Difficulty control: Based on the user's historical accuracy rate (60% in physics and mechanics), a medium-difficulty question is generated: "If the friction coefficient μ = 0.2, calculate the acceleration";

[0284] Parsing generator output:

[0285] "Step 1: Decomposing the Gravity F || =5×9.8×sin30°=24.5N→Newton’s second law”.

[0286] 5. Interactive logic correction:

[0287] User Action: Insert the step "Verify that the bevel angle is 30°";

[0288] System response:

[0289] Generate a revised comparison view, highlighting newly added nodes;

[0290] Update the association weight w between "angle verification" and "inclined motion" in the knowledge graph ij Increased from 0.4 to 0.6.

[0291] 6. Learning behavior analysis:

[0292] LSTM prediction: Detects users frequently correcting the "force analysis" step and predicts a +20% error probability in the future;

[0293] Dynamic adjustment: Improve the search priority of mechanical knowledge points, and make it mandatory for subsequent questions to include force decomposition diagrams.

[0294] 7. Privacy Protection:

[0295] The user's voice data is processed with differential privacy (=0.1) and desensitized to "User ID=7F3A asks: Physics-Slant Problem".

[0296] 8. Dynamic verification interface adjustment:

[0297] The PID controller detects the actual confusion rate of 8% (lower than the target 10%) and reduces the verification threshold τ val To 0.6, allowing more associated physics formulas to pass.

[0298] Example 2: Analysis of Chinese Argumentative Essays and Their Relationship with Environmental Science

[0299] Scenario description: Students need to analyze the argumentative logic of an essay on the pros and cons of technological development and systematically demonstrate their ability to relate interdisciplinary knowledge.

[0300] 1. User input:

[0301] Method: Upload the text "Analysis of the validity of the core arguments in the argumentative essay "The Double-edged Sword of Technology"."

[0302] Processing: BERT extracts keywords such as “technological development,” “pros and cons analysis,” and “validity of arguments.”

[0303] 2. Disassembly of CoT-RAG module:

[0304] Symbolic logic reasoning unit generation framework:

[0305] Step 1: Identify the argument → Step 2: Evaluate the evidence → Step 3: Analyze the logic of the conclusion.

[0306] Neural network completion unit filling details:

[0307] Step 2: Connect the impact of the Industrial Revolution with contemporary environmental issues.

[0308] 3. Knowledge graph verification:

[0309] The subject classifier identifies them as Chinese (probability 0.8) and Environmental Science (probability 0.2), triggering the interdisciplinary isolation mechanism;

[0310] The semantic reinforcement model verifies the correlation between “Industrial Revolution” and “Science and Technology Ethics Discussion” (score 0.82), and it is allowed to be retained.

[0311] 4. Non-question bank generation engine:

[0312] Difficulty Control: Based on the user's argumentative essay analysis ability (θ = 0.65), the generated topic is: "Compare the core argumentation strategies of two argumentative essays with different perspectives (supportive / cautious) on scientific and technological development";

[0313] Parsing generator embedded hyperlinks: discussion of science and technology ethics, argumentative essay structure.

[0314] 5. Interactive logic correction:

[0315] User Action: Delete "social cost" in step 3 and replace it with "long-term ecological impact";

[0316] System response:

[0317] Update the out-degree weight of the "ecological impact" node in the knowledge graph;

[0318] Record this correction to the logic error pattern library (type: "Argumentation focus deviation").

[0319] 6. Learning behavior analysis:

[0320] Update of the 3D Time Series Matrix: User ID=5D2B added the error type "Insufficient Evidence Relevance" to the "Argumentative Essay Analysis" knowledge point;

[0321] Adjustment of display granularity: Due to users frequently revising the argument evaluation steps, subsequent analysis and refinement of the display of the logical chain of "scientific facts → argument support" are carried out.

[0322] 7. Privacy Protection:

[0323] The user text input is processed by the federated learning framework, the text analysis model is trained locally, and only the desensitization parameters θ are uploaded. local .

[0324] 8. Dynamic connection of knowledge graph:

[0325] The question bank monitoring interface detects the newly added question "Analyze how to cite data in argumentative essays to enhance persuasiveness" and automatically aligns it to the "Argumentation Method" node of the knowledge graph.

[0326] Example 3: Chemical Experiment Design and Safety Verification

[0327] Scenario description: Students design a "hydrogen production" experiment, systematically verify the safety of the steps, and generate an optimization plan.

[0328] 1. User input:

[0329] Method: Upload the image of the experimental setup (reaction of zinc particles with dilute sulfuric acid).

[0330] Processing: OCR recognizes the text "Zn+H2SO4→ZnSO4+H2↑".

[0331] 2. Disassembly of CoT-RAG module:

[0332] Symbolic logic reasoning unit generation framework:

[0333] Step 1: Reactive Writing → Step 2: Gas Collection Method → Step 3: Safety Precautions Dynamic Verification Interface Call Knowledge Graph:

[0334] Verify the safety of the "water drainage method for collecting hydrogen" (retrieve the "hydrogen is flammable" warning).

[0335] 3. Knowledge graph verification:

[0336] Isolation of subject boundaries: filtering the physics knowledge point "gas diffusion rate";

[0337] Graph embedding technology: Mapping "hydrogen production" to chemical vector space, which is strongly associated with "explosion limit".

[0338] 4. Non-question bank generation engine:

[0339] Question generation: "If hydrochloric acid is used instead of sulfuric acid, what experimental conditions need to be adjusted?";

[0340] The analysis generator prompts: "Excessive hydrochloric acid concentration may produce Cl2 (chlorine gas hazard)."

[0341] 5. Interactive logic correction:

[0342] User Action: Add "Ventilation Equipment Check" in step 3;

[0343] System response:

[0344] The association weight w between “ventilation equipment” and “experimental safety” in the knowledge graph ij From 0.3 to 0.5;

[0345] Push the safety operation video link to the user interaction module.

[0346] 6. Learning behavior analysis:

[0347] LSTM prediction: detects when a user ignores "gas purity verification" and predicts a future error probability of +15%;

[0348] Dynamic adjustment: Subsequent experimental questions must include purity detection steps.

[0349] 7. Dynamic verification interface adjustment:

[0350] PID control: actual confusion rate 12% (target 10%), calculate the new threshold τ val = 0.65, strictly filtering low-correlation steps.

[0351] 3. Summary

[0352] This system achieves transparent reasoning chain generation and real-time creation of non-question bank questions through multi-disciplinary coverage (physics, Chinese, chemistry, etc.) and deep module collaboration (user interaction, CoT-RAG logical decomposition, and dynamic verification of knowledge graphs). It combines the learning behavior analysis unit to dynamically optimize the knowledge recommendation priority, uses PID control algorithms and federated learning frameworks to balance retrieval accuracy and data security, and uses differential privacy technology to desensitize user behavior data. Ultimately, it constructs an intelligent education closed loop with strong interpretability, high professional accuracy, and privacy compliance, significantly improving learning efficiency and interdisciplinary problem-solving capabilities.

Claims

1. The AI intelligent education tutoring system based on thought chain and retrieval enhancement generation technology is characterized by: Includes the following modules that work together: The user interaction module is used to receive homework, test questions, and natural language questions input by users, pass the questions to downstream modules for processing, and finally output feedback information containing an explainable reasoning chain; The CoT-RAG fusion architecture module is directly connected to the user interaction module. It uses thought chain technology to break down received questions into multi-level logical steps. It also drives the Retrieval Enhanced Generation (RAG) mechanism to extract relevant information from multi-source heterogeneous data and generate a transparent logical chain for solving problems. The knowledge graph module, which isolates subject boundaries, builds an interdisciplinary knowledge graph and receives the decomposition results from the CoT-RAG fusion architecture module. Through the built-in semantic reinforcement model and subject classifier, it dynamically identifies and isolates interdisciplinary interference items, providing knowledge-related data with professional accuracy for downstream generation modules. The non-question bank-dependent generation engine generates adaptive exercises and analysis content in real time based on the semantic association data output by the knowledge graph module and the user's historical learning level. The generation process is independent of the pre-existing question bank and the generated content is fed back to the user interaction module for display; The interactive logic correction module is bidirectionally connected to the user interaction module and the CoT-RAG fusion architecture module, allowing users to modify the reasoning path by clicking on the logic node, triggering the system to regenerate the corrected answer and synchronize the corrected data to the knowledge graph module and learning behavior analysis unit; The learning behavior analysis unit receives behavioral data from the user interaction module and the logic correction module in real time, and dynamically adjusts the retrieval priority of the knowledge graph module and the disassembly granularity of the CoT-RAG fusion architecture module through time series modeling to form a closed-loop optimization link.

2. The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology according to claim 1 is characterized in that: The CoT-RAG fusion architecture module includes: The symbolic logic reasoning unit abstracts the questions passed in by the user interaction module into formal logical expressions and generates a preliminary reasoning framework; The neural network completion unit fills in the logical details based on the Transformer architecture, annotates the confidence level of each reasoning step, and passes the results to the knowledge graph module isolated by subject boundaries for verification; The dynamic verification interface calls the semantic relationship data of the knowledge graph module to verify the factual basis of the logical node in real time.

3. The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology according to claim 1 is characterized in that: The knowledge graph module for subject boundary isolation achieves cross-disciplinary interference suppression in the following ways: Define discipline-specific semantic relationship constraints to limit the direct association of cross-disciplinary knowledge points; Use graph embedding technology to map different subject knowledge points into independent vector spaces to reduce the risk of semantic confusion; Set up dynamic filters to automatically cut off paths with interdisciplinary correlations below a threshold during the retrieval process, ensuring that the data input into the non-question bank-dependent generation engine meets the subject professional requirements.

4. The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology according to claim 1 is characterized in that: The non-question bank dependent generation engine includes: The semantic relevance calculation unit generates the question trunk based on the node relevance strength output by the knowledge graph module; The difficulty control unit dynamically adjusts the complexity of question parameters based on the user's historical answer accuracy rate provided by the learning behavior analysis unit; The parsing generator converts the reasoning steps of the CoT-RAG fusion architecture module into natural language parsing and embeds hyperlinks to related knowledge points provided by the knowledge graph module.

5. The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology according to claim 1 is characterized in that: The interactive logic correction module supports: Perform logical addition, deletion, and modification operations on any node in the thinking chain generated by the CoT-RAG fusion architecture module; Automatically generate a logical comparison view before and after the correction, highlight the difference nodes, and feed the correction results back to the knowledge graph module to update the semantic association; A logical error pattern library is constructed based on the correction records to optimize the subsequent reasoning path generation.

6. The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology according to claim 1 is characterized in that: The learning behavior analysis unit achieves dynamic adjustment in the following ways: Establish a three-dimensional time series matrix of users, knowledge points, and logical errors, and associate the behavioral data of the user interaction module with the feedback information of the logic correction module; Use the LSTM model to predict user knowledge weaknesses and adjust the retrieval priority of the knowledge graph module; Automatically simplify or refine the inference chain display granularity of the CoT-RAG fusion architecture module based on the user logic modification frequency.

7. The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology according to claim 1 is characterized in that: The system also includes privacy protection mechanisms: Differential privacy technology is used to inject noise into the behavioral data uploaded by the user interaction module; The federated learning framework is used to support the collaborative training of local models by multiple terminal devices, uploading only the desensitized model parameters to the central server and updating them in collaboration with the knowledge graph module.

8. The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology according to claim 1 is characterized in that: The knowledge graph module establishes a dynamic connection channel with the standardized education question database, including: Question bank change monitoring interface to capture new knowledge points and question type changes in real time; An automated semantic alignment unit that maps new content to existing knowledge graph nodes; The version control module maintains the knowledge graph branches of different textbook versions.

9. The AI intelligent education and tutoring system based on thought chain and retrieval enhancement generation technology according to claim 1 is characterized in that: When calling the knowledge graph module, the dynamic verification interface synchronously triggers the subject classifier to predict the subject attribution probability of the retrieval results, and dynamically adjusts the verification threshold of the semantic reinforcement model according to the prediction results.

Images