Multi-agent collaborative and dynamic knowledge graph-based multi-modal adaptive learning method, system, device, computer readable storage medium and computer program product

Abstract

The application provides a multi-modal adaptive learning method and system based on multi-agent collaboration and dynamic knowledge graph, a device, a computer readable storage medium and a computer program product. The method comprises: collecting learning task input information; performing multi-modal preprocessing and perception on the structured learning task input package to obtain a multi-modal feature package; performing candidate pruning and task scheduling based on the multi-modal feature package to output a candidate sample package; performing reasoning diagnosis, data set prior analysis, dynamic graph reading and student state reading in parallel according to the candidate sample package to obtain a parallel execution result; generating a personalized learning path and learning suggestion according to the multi-modal feature package and the parallel execution result package in combination with a fusion strategy and a dynamic knowledge graph; and updating the student state and performing abnormal processing according to the interaction result of the current learning task. The application improves the error mode recognition stability, learning path rationality and system continuous service capability under complex input conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent education and adaptive learning technology, specifically to a multimodal adaptive learning method, system, device, computer-readable storage medium, and computer program product based on multi-agent collaboration and dynamic knowledge graph. Background Technology

[0002] Existing intelligent learning systems typically analyze question text or student answer text, mainly relying on a single question-answering model, static question bank, or single-path semantic retrieval method to output knowledge point recognition results and learning suggestions. While such solutions can provide some assistance when the input is complete and the task structure is simple, in real-world teaching scenarios, student-submitted learning tasks often simultaneously include question text, student answers, standard answers, speech-to-text transcription, screenshots, photos of whiteboard writing, function graphs, and historical learning records, making the input format much more complex.

[0003] Meanwhile, existing solutions have at least the following shortcomings: First, a single reasoning path is highly sensitive to the completeness of the question stem and the quality of the input. When the question stem is missing, the text is blurry, the image is occluded, or the speech is disjointed, it is prone to unstable attribution errors. Second, existing recommendation logic usually lacks explicit constraints on the prerequisite dependencies, bridging relationships, and difficulty gradients of knowledge points, and the generated learning path is prone to skipping necessary foundations. Third, student profile updates and knowledge graph updates are often performed separately, making it difficult to form a closed-loop mechanism where a single diagnostic result can be continuously fed back to subsequent tasks. Fourth, when the external reasoning engine times out, connection is abnormal, or structured output fails, the system is prone to long waiting times or even direct service interruption, making it difficult to meet the low latency requirements of online teaching.

[0004] Therefore, a new adaptive learning scheme is urgently needed to at least solve some of the above-mentioned technical problems. Summary of the Invention

[0005] In view of this, this application proposes a multimodal adaptive learning method, system, device, computer-readable storage medium, and computer program product based on multi-agent collaboration and dynamic knowledge graphs, to improve the stability of error pattern recognition, the rationality of the learning path structure, and the continuous service capability of the system while ensuring processing efficiency. Specifically, this application is implemented through the following technical solutions: According to a first aspect of the embodiments of this specification, a multimodal adaptive learning method based on multi-agent collaboration and dynamic knowledge graph is provided, the method comprising the following steps: Step S1: Collect the input information of this learning task and encapsulate it into a structured learning task input package. The input information of this learning task includes required field information and optional field information. The required field information includes the question text, student answer and standard answer. The optional field information includes at least one of the following: speech transcription result, image data, image file name, learning style and historical interaction record. Step S2: Perform multimodal preprocessing and perception on the input package of the structured learning task to obtain a multimodal feature package; Step S3: Based on the multimodal feature package, perform candidate sample pruning and task scheduling, and output candidate sample package; Step S4: Based on the candidate sample package, perform inference diagnosis, dataset prior analysis, dynamic graph reading and student status reading in parallel to obtain the parallel execution result package; Step S5: Based on the multimodal feature package and the parallel execution result package, a personalized learning path and learning suggestions are generated using a preset fusion strategy and a path legitimization strategy based on a dynamic knowledge graph. Step S6: Update student status and handle anomalies based on the interaction results of this learning task, and output the final adaptive learning result.

[0006] According to a second aspect of the embodiments of this specification, a multimodal adaptive learning system based on multi-agent collaboration and dynamic knowledge graph is provided, the system comprising: The input acquisition unit is used to acquire the input information of this learning task and encapsulate it into a structured learning task input package. The input information of this learning task includes required field information and optional field information. The required field information includes the question text, student answer and standard answer. The optional field information includes at least one of the following: speech transcription result, image data, image file name, learning style and historical interaction record. A multimodal preprocessing and sensing unit is used to perform multimodal preprocessing and sensing on the input package of the structured learning task to obtain a multimodal feature package; The candidate pruning and task scheduling unit is used to perform candidate sample pruning and task scheduling based on the multimodal feature package and output the candidate sample package. The parallel execution unit is used to perform inference diagnosis, dataset prior analysis, dynamic graph reading and student status reading in parallel based on the candidate sample package, and obtain a parallel execution result package. The learning path generation unit is used to generate personalized learning paths and learning suggestions based on the multimodal feature package and the parallel execution result package, using a preset fusion strategy and a path legitimization strategy based on a dynamic knowledge graph. The result output unit is used to update the student status and handle anomalies based on the interaction results of the current learning task, and output the final adaptive learning result.

[0007] According to a third aspect of the embodiments of this specification, an apparatus is provided, comprising: A processor; and a computer-readable storage medium storing computer program instructions that, when executed by the processor, cause the processor to perform any of the methods described above.

[0008] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided, characterized in that a computer program is stored on the computer-readable storage medium, the computer program being executed by a processor using the method described in any of the preceding claims.

[0009] According to a fifth aspect of the embodiments of this specification, a computer program product is provided, characterized in that the computer program product includes a computer program / instructions, which, when executed by a processor, implement the method described in any of the preceding claims.

[0010] The embodiments of this application have at least the following technical effects: This application unifies heterogeneous inputs such as text, speech, and images into structured features through multimodal preprocessing. Then, a confidence fusion mechanism combining multi-agent parallel diagnosis and prior dataset analysis effectively overcomes the sensitivity of a single inference path to input quality, significantly improving the stability of error pattern recognition under noisy conditions. Simultaneously, through path legitimization processing of dynamic knowledge graphs and closed-loop updates of student states, the generated learning paths simultaneously satisfy knowledge point prerequisites, bridging relationships, and difficulty gradient constraints, possessing ample personalized features. Furthermore, by leveraging lazy loading, cache reuse, and runtime circuit breaker fallback mechanisms, the system can continue to output structured results even when the external inference engine malfunctions, ensuring low latency and high availability in online teaching scenarios. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a multimodal adaptive learning method based on multi-agent collaboration and dynamic knowledge graph in an embodiment of this application. Figure 2 This is a schematic diagram of the overall process of multimodal adaptive learning based on multi-agent collaboration and dynamic knowledge graph in an embodiment of this application; Figure 3This is a schematic diagram of a simulation result in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a multimodal adaptive learning system based on multi-agent collaboration and dynamic knowledge graph in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a device according to an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0014] This application provides a multimodal adaptive learning method based on multi-agent collaboration and dynamic knowledge graphs, such as... Figure 1 The diagram illustrates a flowchart of a multimodal adaptive learning method based on multi-agent collaboration and dynamic knowledge graphs, as described in this application. The method includes the following steps S1-S6: Step S1: Collect the input information of this learning task and encapsulate it into a structured learning task input package. The input information of this learning task includes required fields and optional fields. The required fields include the question text, student answers, and standard answers. The optional fields include at least one of the following: speech transcription results, image data, image file names, learning styles, and historical interaction records.

[0015] Combination Figure 2 This document provides a schematic diagram of the overall process of multimodal adaptive learning based on multi-agent collaboration and dynamic knowledge graphs, as described in an embodiment of this application. Specifically, the input information for this learning task in this step includes the question text, student answers, and standard answers, as well as optional fields such as speech-to-text results, image data, image file names, learning styles, and historical interaction records. The question text, student answers, and standard answers are mandatory fields; the speech-to-text results, image data, image file names, learning styles, and historical interaction records are optional fields. First, integrity checks are performed on each input field. Then, the validated input information is uniformly packaged into a learning task input package, serving as the basis for subsequent multimodal analysis and multi-agent scheduling. The output of this step is a structured learning task input package.

[0016] Step S2: Perform multimodal preprocessing and perception on the input package of the structured learning task to obtain a multimodal feature package.

[0017] The structured learning task input package collected in the above steps undergoes multimodal preprocessing and perception. This preprocessing may include text normalization, image enhancement, image quality and layout analysis, scene judgment, and prompt generation. Step S2 transforms the original multimodal input into structured feature results that can be directly used for candidate pruning and inference diagnosis. The output of this step is a multimodal feature package.

[0018] Step S3: Based on the multimodal feature package, perform candidate sample pruning and task scheduling, and output candidate sample package.

[0019] After obtaining the multimodal feature package, candidate samples are selected from historical samples and pruned. After the candidate pruning is completed, lazy loading initialization, background warm-up, cache reading and parallel task distribution are performed according to the current request type, so as to prepare for subsequent parallel inference diagnosis and prior analysis. The output of this step is the candidate sample package.

[0020] Step S4: Based on the candidate sample package, perform inference diagnosis, dataset prior analysis, dynamic graph reading and student status reading in parallel to obtain the parallel execution result package.

[0021] After obtaining the candidate sample package, inference diagnosis, dataset prior analysis, dynamic graph reading, and student status reading are performed in parallel, outputting structured inference diagnosis results, dataset prior analysis results, graph snapshots, and student status information, respectively. Structured inference diagnosis results include information such as similarity, candidate knowledge points, error attribution, thought process, and diagnostic confidence. Dataset prior analysis results include information such as similar cases, prior learning paths, candidate knowledge points, and error patterns. Student status information includes information such as the student's mastery of relevant knowledge points, distribution of weaknesses, and historical interaction status.

[0022] Step S4 completes the parallel diagnosis and prior analysis of the current learning task. The output of this step is the parallel execution result package.

[0023] Step S5: Based on the multimodal feature package and the parallel execution result package, a personalized learning path and learning suggestions are generated using a preset fusion strategy and a path legitimization strategy based on a dynamic knowledge graph.

[0024] Combining the multimodal feature package and parallel execution result package output from the above steps, the results are fused and paths are generated. Furthermore, a path legalization strategy based on dynamic knowledge graphs is used for path legalization processing, and finally, personalized learning paths and learning suggestions are output.

[0025] Step S6: Update student status and handle anomalies based on the interaction results of this learning task, and output the final adaptive learning result.

[0026] The core of this step is to achieve closed-loop updates of student profiles, knowledge graphs, and system status, and to provide service guarantees in case of anomalies. Finally, the adaptive learning results of this learning task include diagnostic scores, knowledge point identification results, error patterns, personalized learning paths, learning suggestions, student profiles, and knowledge graph visualization results.

[0027] This application unifies heterogeneous inputs such as text, speech, and images into structured features through multimodal preprocessing. Then, a confidence fusion mechanism combining multi-agent parallel diagnosis and prior dataset analysis effectively overcomes the sensitivity of a single inference path to input quality, significantly improving the stability of error pattern recognition under noisy conditions. Simultaneously, through path legitimization processing of dynamic knowledge graphs and closed-loop updates of student states, the generated learning paths simultaneously satisfy knowledge point prerequisites, bridging relationships, and difficulty gradient constraints, possessing ample personalized features. Furthermore, by leveraging lazy loading, cache reuse, and runtime circuit breaker fallback mechanisms, the system can continue to output structured results even when the external inference engine malfunctions, ensuring low latency and high availability in online teaching scenarios.

[0028] In some embodiments of this application, step S2 includes: step S21, performing text normalization and keyword extraction on the question text, the student answer, the standard answer, and the speech-to-text result to obtain a set of text keywords; step S22, performing image enhancement and binarization processing on the image data to obtain an enhanced image and a binarized image; step S23, calculating image quality indicators and layout structure indicators based on the enhanced image and the binarized image; and step S24, outputting scene labels, collection suggestions, and intelligent agent assistance prompts according to the image quality indicators and the layout structure indicators.

[0029] Step S2 is the fundamental feature extraction step in the entire adaptive learning method. Its core objective is to uniformly transform the raw inputs from different modalities into structured, quantifiable, and directly callable feature representations, providing a reliable input basis for subsequent candidate pruning and inference diagnosis. Step S2 includes steps S21-S24, and the specific implementation details of each step are as follows: Step S21: This step performs unified standardization processing and keyword extraction on the four types of text inputs: question text, student answers, standard answers, and speech-to-text results.

[0030] (1) Whitespace removal: First, remove the whitespace characters at the beginning and end of each text field, replace multiple consecutive spaces in the text with a single space, and remove tabs and newlines to make the text content compact and consistent.

[0031] (2) Case normalization: English characters in the text are uniformly converted to lowercase to eliminate matching failures caused by case differences; for mathematical symbols, variable names, formula fragments and proper nouns, their original writing forms are retained to avoid changes in the meaning of formulas or terms due to normalization.

[0032] (3) Symbol standardization: Convert full-width punctuation marks in the text to half-width punctuation marks, unify quotation marks and parentheses with different encodings to standard ASCII characters, and unify mathematical symbols with different writing styles (such as "×" and "..."). The symbols “”, “÷”, and “ / ” are unified into a standard representation to ensure consistency in subsequent semantic analysis.

[0033] (4) Keyword extraction: Extract text keywords based on preset word element rules, domain keyword list and frequency of occurrence, while retaining formula fragments and key phrases as independent word elements.

[0034] After the above four processing steps, a set of text keywords is output, which serves as the basic input for subsequent candidate sample pruning and similarity calculation.

[0035] Step S22: Image enhancement and binarization This step performs six image processing operations sequentially on the image data (including screenshots of the question, photos of the whiteboard, function graphs, etc.).

[0036] (1) Orientation correction: First, read the EXIF ​​orientation information of the image and perform automatic orientation correction on the image; if the EXIF ​​orientation information is unavailable, rotation correction can be performed based on the page boundary, text orientation or image subject orientation estimation results to ensure that the image content orientation is correct.

[0037] (2) Grayscale conversion: Convert the color image to a grayscale image to eliminate the interference of color channels on subsequent quality analysis and reduce the amount of computation.

[0038] (3) Automatic contrast enhancement: Automatically enhances the contrast of grayscale images, making the strokes of text, the outlines of graphics, the details in the dark and the details in the bright areas of the image clearer. It is especially suitable for shooting dim images caused by insufficient light.

[0039] (4) Median filtering: The median filtering algorithm is used to remove salt and pepper noise in the image to avoid noise interference with edge detection and layout analysis.

[0040] (5) Sharpening: The Laplacian operator is used to sharpen the image, highlighting the edge information of the strokes of the text and the outline of the graphics, which is convenient for subsequent layout structure analysis.

[0041] (6) Binarization: The threshold is determined based on the grayscale histogram and the grayscale image is converted into a binary image, that is, the pixel value only takes two states, black or white. This binary image will be used for the calculation of subsequent layout structure indicators.

[0042] After the above six processing steps, the enhanced image (used for continuous value calculation of quality indicators) and the binarized image (used for discrete value calculation of layout structure indicators) are output respectively.

[0043] Step S23: Calculation of Image Quality Indicators and Page Layout Indicators This step calculates multiple quantization metrics based on the enhanced and binarized images to accurately characterize the physical features and layout structure of the images. These metrics include the following: (1) Brightness: The arithmetic mean of the gray values ​​of all pixels in the enhanced image, which represents the overall brightness of the image. Low brightness usually means that the shooting environment is not bright enough.

[0044] (2) Contrast: Enhances the standard deviation of the gray values ​​of image pixels, characterizing the degree of dispersion of gray distribution. A higher contrast means that the difference between light and dark is obvious.

[0045] (3) Edge density: After performing edge intensity statistics on the enhanced image, the proportion of edge pixels or edge response to the total pixels represents the richness of the text and graphic outlines. High edge density indicates rich content details.

[0046] (4) Foreground ratio: The proportion of foreground pixels representing text, handwriting or graphic outlines in a binarized image to the total number of pixels. It represents the proportion of the content in the entire image. A low foreground ratio indicates that the image contains a large amount of invalid background.

[0047] (5) Text line estimate: Based on the horizontal projection of the enhanced image, the number of projection peaks is counted to represent the number of text lines in the document. A higher text line estimate indicates that the content is mainly multi-line text.

[0048] (6) Content region estimation: Based on the connected component analysis of the binarized image, the number and distribution of independent content blocks are estimated to characterize the degree of image segmentation. A higher content region estimation value indicates that the image contains multiple independent content regions.

[0049] (7) Center density: The pixel density in the central area of ​​the image (such as the central 1 / 4 area), which helps to determine whether the image is a screenshot. Screenshots usually have content concentrated in the central area.

[0050] (8) Occupied area ratio: The ratio of the effective content area to the total image area. It helps to comprehensively judge the image quality. A low occupied area ratio indicates that there is too little effective content.

[0051] Step S24: Scene Judgment and Prompt Generation This step uses the image quality index and layout structure index calculated in step S23, combined with the image file name and user notes, to make a comprehensive scene judgment and generate corresponding tags and prompts.

[0052] Scene Label Output: Based on the comprehensive judgment results, select one or more labels from the following scene types for output: Question Screenshot, Document Page, Whiteboard Photo, Screenshot, Function / Coordinate Graph, General Learning Image. The criteria for determining the scene label include: if the estimated value of the text line is high and the foreground proportion is moderate with high center density, it is determined to be a question screenshot or document page; if the brightness is low, the contrast is high, and the edge density is low, it is determined to be a whiteboard photo; if it contains obvious coordinate axes and curves, it is determined to be a function / coordinate graph.

[0053] Image acquisition suggestion output: When the image quality index is lower than the preset threshold (such as brightness below 50 or foreground ratio below 20%), an acquisition suggestion is output to prompt the user to retake the photo or add text description.

[0054] Intelligent agent-assisted prompts output: Scene labels and quality indicators jointly determine subsequent processing strategies. When the scene label is a screenshot of the question and the quality indicators are qualified, the image can be used as a valid contextual input for image understanding or reasoning diagnosis; when local OCR dependency is available and there is recognizable text in the image, the OCR text recognition result is further output; when the scene label is a function / coordinate graph, the label prioritizes calling the image intelligent agent for graph understanding.

[0055] Through the collaborative processing of steps S21 to S24, the original multimodal input is completely converted into a multimodal feature package. This feature package includes a set of text keywords, a set of image quality indicators, a set of layout structure indicators, scene labels, acquisition suggestions, and intelligent agent assistance prompts, which can be directly called by the subsequent candidate pruning and task scheduling module and the inference diagnosis module.

[0056] This application's embodiments eliminate the interference of text format differences on semantic matching by performing normalization processing on text input; effectively improve the analyzability of images under adverse conditions such as low light and reflection by performing enhancement processing on image input; and generate scene labels and intelligent agent assistance prompts by calculating multiple quantitative indicators and combining scene judgment, enabling the system to adaptively select subsequent processing strategies according to the actual physical characteristics and content type of the image, thereby uniformly converting heterogeneous original multimodal inputs into structured, quantifiable, and directly callable feature representations, providing a high-quality feature foundation for subsequent candidate sample pruning and multi-agent parallel inference diagnosis, and significantly improving the processing stability and accuracy of the entire adaptive learning system under complex input conditions.

[0057] In some embodiments of this application, step S3 includes: step S31, constructing a lexical set and a topic candidate set based on the question text, the student answer, and the standard answer; step S32, calculating a coarse ranking score for candidate samples based on the lexical set, the topic candidate set, and historical sample cache data, and determining a coarsely ranked candidate sample set based on the coarse ranking score; step S33, performing a fine ranking on the coarsely ranked candidate sample set using a preset similarity algorithm to obtain the candidate sample package, wherein the candidate sample package includes similar cases, candidate knowledge points, candidate error patterns, and candidate learning paths.

[0058] The core objective of step S3 is to quickly select the candidate samples most relevant to the current learning task from a massive amount of historical samples, and to perform differentiated task scheduling based on the request type, providing high-quality candidate inputs and efficient allocation of computing resources for subsequent parallel inference diagnosis and prior analysis stages. Step S3 includes steps S31-S34, and the specific implementation details of each step are as follows: Step S31: Topic Prediction and Lexical Construction The purpose of this step is to extract structured lexical units from the text input of the current learning task that can be used for fast matching, and to determine the knowledge domain to which the task belongs, so as to provide a basic index for subsequent coarse ranking.

[0059] Lexical segmentation: The three required text fields—question text, student answers, and standard answers—are concatenated sequentially into a unified query text. Lexical segmentation is then performed on this query text. Lexical segmentation is not simply character splitting; rather, it extracts the smallest matching semantic units based on a domain dictionary and formula recognition rules. Specifically, this includes: ordinary text keywords, formula fragments, and key phrases. After segmentation, all lexical units are deduplicated, and a lexical set is constructed. This set serves as the basic unit for subsequent inverted index matching and similarity calculation.

[0060] Topic prediction: Input the same query text into the topic judgment module based on domain dictionary, keyword statistics, or preset topic rules to determine which knowledge domain (mathematics, science, technology, or engineering) the current task is more likely to belong to, and construct a topic candidate set Θ accordingly. For example, if the query text contains words such as "quadratic function" and "discriminant", the topic prediction result is "mathematics", and Θ will contain the topic tag "mathematics".

[0061] The output of this step is a lexical set and a topic candidate set Θ.

[0062] Step S32: Coarse ranking of candidate samples This step performs a fast coarse sort on all historical samples based on a three-way index: the lexical inverted index, the topic index, and the historical cache, to filter out an initial candidate set with low computational cost. The lexical inverted index is a pre-built inverted table that records which historical samples each lexical appears in, supporting fast retrieval based on the number of shared lexical terms. The topic index is a sample index organized by topic classification, supporting fast filtering based on topic matching. The historical cache is a high-speed storage area that stores summaries of recently hit similar cases, topic tags, keyword sets, similarity scores, and summaries of previous learning paths, supporting O(1) time cache hit judgment.

[0063] Coarse ranking score calculation: Calculate the coarse ranking score for the i-th historical sample using the following formula: H_i = n_i + 2·I(topic_i ∈ Θ) + I(cache_i = 1) Where n_i represents the number of shared terms between the i-th historical sample and the current query text, providing a basic relevance score; topic_i represents the topic category of the i-th historical sample; Θ represents the topic candidate set for the current task; I is an indicator function, which takes a value of 1 when the condition in parentheses is met, and 0 otherwise; cache_i indicates whether the i-th historical sample hits the cache, taking a value of 1 if it does, and 0 otherwise.

[0064] The number of shared terms, n_i, reflects the direct relevance at the text level; topic matching is given a double weighting, reflecting the significant impact of knowledge domain consistency on candidate sample selection; cache hits are given an additional bonus, reflecting the timeliness and reference value of the most recently interacted samples. After the scores of the three items are combined, they are sorted from high to low according to H_i, and the top K samples (K is a preset threshold, such as 50) are taken as the candidate sample set after coarse ranking.

[0065] Step S33: Candidate Sample Refinement This step performs a more refined multi-dimensional similarity calculation on the historical samples that have entered the candidate set after the coarse ranking, and finally outputs a high-quality candidate sample package.

[0066] (1) Question text similarity: Based on semantic vectors, keyword overlap or a combination of both, it measures the semantic similarity of the question descriptions.

[0067] (2) Student answer similarity: Based on at least one of the following calculations, namely, answer text overlap, formula fragment matching and semantic similarity, the similarity between students' answering methods and content is measured.

[0068] (3) Similarity of standard answer: Based on the degree of overlap between standard answer text, formula fragments and knowledge point keywords, it measures the similarity of knowledge points involved in the correct answer.

[0069] (4) Word set similarity: Based on the word overlap rate, it measures the degree of matching between keywords and formula fragments.

[0070] Fine-grained ranking score fusion: The four types of similarity are first normalized to the [0,1] interval, and then weighted and fused according to preset weights to form a fine-grained ranking score. The sample with the highest fine-grained ranking score is taken as the similar case output. At the same time, candidate knowledge points (a list of knowledge points involved in the sample), candidate error patterns (typical error types recorded in the sample), and candidate learning paths (historical recommendation learning paths corresponding to the sample) are extracted from the sample and the candidate set to form a candidate sample package.

[0071] Step S34: Runtime Scheduling After candidate pruning is completed, this step performs differentiated task scheduling based on the current request type to ensure efficient utilization of computing resources and orderly startup of each module.

[0072] The request types in this application embodiment include, for example, plain text requests, voice requests, image requests, and multimodal requests, which respectively enable lexical / topic analysis, speech-to-text understanding, image quality and layout analysis, or combinations thereof. If the request type is a comprehensive recommendation request, adaptive, pattern-driven, clustering-driven, and comprehensive recommendation strategies are further enabled. The output of this step is a candidate sample package.

[0073] Four scheduling strategies: (1) Lazy loading initialization: only load the agents and models necessary for the current request type to avoid loading all modules and wasting resources; (2) Background warm-up: Frequently called modules (such as inverted word index and BERT vector model) are preloaded into memory to reduce the latency of the first call; (3) Cache reading: Prioritize reading intermediate results from the historical cache (such as summaries of recent similar cases) to avoid duplicate calculations; (4) Parallel task distribution: The four tasks of reasoning diagnosis, prior analysis, graph reading and student status reading are distributed to different computing units for parallel execution, making full use of multi-core or distributed computing resources.

[0074] The output of this step is a candidate sample package, which, together with the multimodal feature package, serves as the input for the parallel analysis step S4.

[0075] This application's embodiments construct a structured index foundation through lexical segmentation and topic prediction. A coarse-ranking scoring formula, incorporating shared lexical quantity, double bonus for topic matching, and cache hit bonus, is then used to rapidly screen historical samples. Subsequently, a refined ranking based on the normalized fusion of similarity across four dimensions—question text, student answers, standard answers, and lexical sets—further enhances the relevance and accuracy of candidate samples. Finally, by combining request-type-based lazy loading, background warm-up, cache reuse, and parallel task distribution scheduling strategies, the computational overhead of candidate sample retrieval is significantly reduced while ensuring the input quality and execution efficiency of subsequent multi-agent parallel analysis. This enables the system to quickly locate high-value reference samples in massive historical data, providing efficient and accurate data support for personalized learning path generation.

[0076] In some embodiments of this application, the parallel execution results include structured reasoning diagnostic results, dataset prior analysis results, graph snapshots, and student status information. Step S4 includes: Step S41, using a reasoning diagnostic agent to output the structured reasoning diagnostic results based on the question text, the student's answer, the standard answer, the learning style, and the candidate sample package. The structured reasoning diagnostic results include similarity, knowledge point candidates, error attribution, thought process, and diagnostic confidence. Step S42, using a dataset prior analysis agent to output the dataset prior analysis results based on historical sample cache data. The dataset prior analysis results include similar cases, prior learning paths, and error pattern statistics. Step S43, using a dynamic knowledge graph reading strategy to read the pre-dependencies, relevance, and difficulty information of the key knowledge points to form the graph snapshot. Step S44, using an adaptive student model to read the student's mastery of the corresponding knowledge points, distribution of weaknesses, and historical interaction status to obtain the student status information.

[0077] Step S4 is the core diagnostic step of the entire adaptive learning method. Its core objective is to comprehensively diagnose the current learning task from four dimensions—the reasoning agent, historical statistics, knowledge graph, and student profile—through a four-way parallel analysis mechanism, ensuring that the diagnostic results achieve optimal accuracy, robustness, and personalization. Step S4 includes steps S41-S44, and the specific implementation details of each step are as follows: Step S41: Reasoning and Diagnosis This step involves the reasoning diagnostic agent performing deep semantic analysis on the current learning task and outputting structured diagnostic results.

[0078] The reasoning diagnostic agent receives multiple types of input information. The first type consists of the question text, student answers, and standard answers, forming the core textual basis for diagnosis. The second type is learning style, used to assist in adjusting the expression and suggested direction of the diagnostic output. The third type is vector hints, which are not the original vectors themselves, but hint text formed by concatenating three parts: similar case recall results, candidate knowledge point keywords, and multimodal preprocessed feature summaries. This hint text is used to stabilize the output of the reasoning diagnostic agent, enabling it to generate structured results even when the input quality fluctuates. The fourth type is candidate sample packages, providing historical reference for reasoning diagnosis.

[0079] The reasoning diagnostic agent outputs structured diagnostic results. These include: similarity, which characterizes the overall closeness between the current task and historical samples; knowledge point candidates, a list of knowledge points identified by the agent that are relevant to the current task; error attribution, a classification of the reasons for student errors, such as conceptual confusion, calculation errors, or misreading the question; a thought process, a step-by-step reasoning chain generated by the agent, demonstrating the complete derivation path from the question to the conclusion; and diagnostic confidence, a self-assessment score of the agent for each of the above results, ranging from 0 to 1, with higher values ​​indicating greater confidence in the outcome.

[0080] Step S42: Prior Analysis of Dataset This step uses prior analysis of the dataset to analyze the statistical patterns of the agent based on historical data and outputs prior analysis results as a supplement and verification for reasoning and diagnosis.

[0081] The dataset prior analysis agent performs statistical analysis based on multiple types of inputs. The first type is historical samples, which include both existing samples in the training library and newly added samples continuously fed back through the interaction logs, ensuring that the analysis is based on the latest learning behavior data. The second type is keywords for the current task, used to locate relevant records in the historical samples. The third type is the hit cache, used to quickly locate similar cases that have appeared frequently recently. The fourth type is seed knowledge points, which are candidate knowledge points output from the reasoning and diagnosis stage, serving as anchor points for statistical analysis.

[0082] The dataset prior analysis agent outputs the dataset prior analysis results. These include: similar cases, which are historical cases matching the current task from statistical perspectives, cache hit rates, or similarity calculations; prior learning paths, which are high-frequency learning paths derived from historical data statistics, reflecting the effective learning order of most students in similar situations; candidate knowledge points, which are high-frequency knowledge points selected from statistical frequency perspectives, forming cross-validation with the knowledge point candidates output by inference diagnosis; and error pattern statistics, which show the frequency of each error pattern in historical data, used to help determine the typicality and severity of the current error.

[0083] Step S43: Dynamic graph reading This step uses dynamic graph planning to extract structured knowledge relationships related to key knowledge points from a dynamic knowledge graph, forming a graph snapshot.

[0084] Based on the key knowledge points output in step S41, an immediate query request is initiated to the knowledge graph management module. The knowledge graph management module then retrieves all associated structured knowledge relationships based on the location of each key knowledge point within the graph.

[0085] The reading results form a snapshot of the learning graph. Among these, prerequisite dependencies are a list of knowledge points that must be mastered before learning the focus knowledge point, ensuring the learning path doesn't skip necessary foundations; correlations are other knowledge points related to the focus knowledge point, used to expand the breadth of the learning path; difficulty information is the difficulty level of the focus knowledge point, used for subsequent priority calculations; bridging nodes are intermediate knowledge points connecting different knowledge domains, providing transitional paths when there are knowledge gaps between the focus knowledge point and the knowledge points already mastered by the student; edge weight information is the co-occurrence strength between knowledge points, reflecting the closeness of their association.

[0086] Step S44: Read student status This step uses an adaptive student model to read the student's personalized status information on relevant knowledge points, providing student-dimensional constraints for subsequent path generation.

[0087] The adaptive student model retrieves multiple types of information from the student profile database based on student identifiers. The first type is knowledge point mastery, which is calculated statistically based on the student's scores in historical interaction records on relevant knowledge points, reflecting the student's familiarity with each knowledge point. The second type is the distribution of weaknesses, which is statistically analyzed based on the set of knowledge points where the student made the most errors in history, reflecting the knowledge areas where the student most needs to strengthen. The third type is historical interaction status, which is analyzed based on the student's recent learning behavior sequence, reflecting the student's recent learning activity and learning trends.

[0088] The above three types of information come from the system's existing student profiles, recent interaction logs, and knowledge point update records. It should be noted that this information is not read-only. After this task is completed, the system will write back and update the student profile based on the interaction results, including updating the mastery level, adjusting the distribution of weak points, and supplementing historical interaction status, thereby forming a continuously iterative closed loop of student status.

[0089] Through four parallel executions in steps S41-S44, a comprehensive diagnosis of the current learning task is completed from four dimensions: deep semantic reasoning, historical statistical patterns, knowledge structure constraints, and student personalized status. The results of the four analyses complement and cross-validate each other: reasoning diagnosis provides deep semantic understanding, prior analysis provides statistical pattern support, graph reading provides knowledge structure constraints, and student status reading provides personalized basis. The output of this step is a diagnostic result package, which includes structured reasoning diagnostic results, dataset prior analysis results, graph snapshots, and student status information, serving as the unified input for the result fusion and path generation stage in step S5.

[0090] This application embodiment receives multidimensional input including vector hints through a reasoning diagnostic agent and outputs structured diagnostic results including thought processes and confidence levels. A dataset prior optimization module outputs statistical similar cases and error pattern frequencies based on continuously flowing historical samples. A dynamic graph planning module instantly reads the pre-dependencies, relevance, bridging nodes, and difficulty information of key knowledge points to form a graph snapshot. An adaptive student model reads and writes back to update students' knowledge point mastery and weakness distribution. These four analyses are executed in parallel and complement each other, enabling the diagnostic results to simultaneously possess deep semantic understanding capabilities, historical statistical robustness, knowledge structure integrity, and student-specific adaptability. This significantly improves the accuracy of error attribution and the rationality of learning path generation in complex learning scenarios.

[0091] In some embodiments of this application, the multimodal feature package includes multimodal perception results, and step S5 includes: step S51, fusing the structured reasoning diagnostic results, the dataset prior analysis results, and the multimodal perception results using a preset fusion strategy to obtain a fusion result; step S52, extracting the target knowledge point set, target error pattern, and focus knowledge point set from the fusion result; step S53, combining the prior learning path, target knowledge point set, student mastery, and graph snapshot to form an original learning path; step S54, performing path legalization processing on the original learning path to obtain the personalized learning path, wherein the path legalization processing includes inserting prerequisite knowledge points, deduplicating duplicate knowledge points, retaining bridging nodes, and limiting path length; step S55, calculating the knowledge point priorities in the personalized learning path based on the graph snapshot and the student status information, and generating the learning suggestions based on the knowledge point priorities.

[0092] Step S5 is the decision output stage of the entire adaptive learning method. Its core objective is to perform confidence fusion on the results of the four parallel analyses and generate learning paths and suggestions that meet the requirements of knowledge point pre-dependencies, difficulty gradients, and personalized needs under the constraints of a dynamic knowledge graph. Step S5 includes steps S51-S55, and the specific implementation details of each step are as follows: Step S51: Confidence Fusion This step performs a weighted fusion of the inference diagnosis results, the dataset prior results, and the multimodal perception results, unifying the heterogeneous diagnostic information into comparable confidence scores.

[0093] The multimodal feature package output in step S2 and the structured diagnostic package output in step S4 are not directly involved in the fusion in this step. Instead, they are first converted into comparable confidence scores. Specifically, C_reason is the confidence score of the inference diagnostic result (from S41, with a value of 0 to 1), C_dataset is the confidence score of the dataset prior result (from S42, with a value of 0 to 1), and C_multi is the confidence score of the multimodal perception result (from S2, with a value of 0 to 1). All three are scalars rather than feature vectors.

[0094] Then, the confidence scores obtained after the above transformation are weighted and fused according to the following formula: C_fuse = α·C_reason + β·C_dataset + γ·C_multi, and α + β + γ = 1 In this context, α, β, and γ are configurable weights, with recommended default values ​​of α=0.5, β=0.3, and γ=0.2. This weight allocation reflects the dominant role of the inference and diagnostic agent in the fusion process, while preserving the statistical support from prior data sets and the supplementary role of multimodal perception features. The fusion result C_fuse is also a confidence score between 0 and 1, with higher values ​​indicating greater confidence in the diagnostic conclusion.

[0095] Step S52: Extracting Key Knowledge Points This step extracts the core knowledge points and error mode information required for path construction based on the fusion result C_fuse.

[0096] Target knowledge point set extraction: The target knowledge point set is determined jointly from multiple sources. The first source is highly relevant knowledge points in the diagnostic results, i.e., the several items with the highest confidence among the knowledge point candidates output by the reasoning diagnostic agent; the second source is high-frequency knowledge points in the prior analysis results, i.e., the knowledge points that appear most frequently in similar situations, as statistically analyzed by the dataset prior analysis agent; the third source is the preceding or related nodes in the graph snapshot, i.e., nodes in the dynamic knowledge graph that have preceding dependencies or correlations with the focus knowledge points. After the knowledge points from multiple sources are merged, they are sorted in descending order of fusion confidence to form the target knowledge point set, which serves as the starting point for subsequent path construction.

[0097] Target error pattern extraction: Select the error attribution with the highest confidence from the fusion results as the target error pattern, which will be clearly marked in the learning suggestions.

[0098] Determining the set of key knowledge points: Based on the set of target knowledge points, further filter out the knowledge points that need to be focused on in this study, namely the knowledge points with the lowest mastery and the highest error frequency, to form the set of key knowledge points.

[0099] Step S53: Generation of the original learning path This step combines multi-source information to construct an unverified original learning path: the first type is the prior learning path, which is the high-frequency learning order derived from historical data statistics, serving as a reference skeleton for path construction; the second type is the target knowledge point set, which determines the core knowledge points that the path needs to cover; the third type is the student mastery level, used to determine which knowledge points need to be learned first and which can be skipped; the fourth type is the graph snapshot, which provides the prerequisite dependencies and related relationship constraints between knowledge points.

[0100] Following the order of the prior learning path, the target knowledge points are arranged sequentially, and relevant nodes from the graph snapshot are inserted between adjacent knowledge points to form an original learning path containing multiple knowledge point nodes. This original path has not yet undergone validity verification and may have issues such as missing prerequisite knowledge points, duplicate nodes, or excessive length.

[0101] Step S54: Path Validation Processing This step performs a validation operation on the original learning path to ensure that the generated path meets the knowledge structure constraints.

[0102] Prerequisite knowledge point insertion: Each target knowledge point in the path is checked one by one. If the knowledge point has a prerequisite knowledge point in the graph, and the student's mastery of the prerequisite knowledge point is lower than the preset threshold of 0.6, then the prerequisite knowledge point is inserted before the target knowledge point. For example, if the target knowledge point is "finding the root of a quadratic equation", its prerequisite knowledge point is "factorization", and the student's mastery of "factorization" is only 0.45, then the system will insert "factorization" before "finding the root of a quadratic equation" in the path.

[0103] Deduplication of knowledge points: Detect whether the same knowledge point is listed multiple times in the path. If so, keep the first occurrence of the node and remove subsequent duplicate nodes to avoid students wasting time on the same knowledge point.

[0104] Bridge node retention: When a bridge node exists between two knowledge points in the graph and is marked as a necessary transition node in the graph, it is retained. The purpose of bridging nodes is to provide a smooth transition when the knowledge points have a large span, avoiding gaps in the learning path.

[0105] Path length limit: Check the total number of nodes in the path. If it exceeds the preset limit (e.g., 15 nodes), truncate to the limit and prioritize the node with the highest priority value to ensure that the path length is moderate and the key points are highlighted.

[0106] After the above four legalization processes, a personalized learning path that meets the knowledge constraints is output.

[0107] Step S55: Priority Calculation and Recommendation Generation This step calculates the learning priority for each knowledge point in the personalized learning path and generates detailed learning suggestions accordingly.

[0108] Priority Calculation: The learning priority of the k-th knowledge point in the path can be calculated using the following formula: P_k=η1·(1-M_k)+η2·D_k+η3·E_k+η4·G_k Wherein, P_k is the priority value of knowledge point k, with a larger value indicating that it should be learned first; M_k is the current mastery level of knowledge point k (between 0 and 1, with higher values ​​indicating better mastery), and (1-M_k) reflects the degree of weakness, with lower mastery levels contributing more to this; D_k is the difficulty of knowledge point k (between 0 and 1, with higher values ​​indicating greater difficulty); E_k is the error frequency of this knowledge point (normalized value, with higher values ​​indicating more errors); G_k is the dependency strength of this knowledge point (between 0 and 1, with higher values ​​indicating a greater degree of dependence by other knowledge points); η1, η2, η3, and η4 are configurable weights. This formula design ensures that priority is simultaneously influenced by four dimensions: weakness, difficulty, error frequency, and dependency.

[0109] Learning suggestion generation: Based on the descending order of priority P_k, a learning suggestion is generated for each knowledge point in the path. Each suggestion may include the following information: knowledge point name, clearly indicating the learning target; action description, such as review, practice, or preview, indicating the learning method; current mastery level, reflecting the student's starting level; priority value, reflecting the importance of the knowledge point in the path; and suggested learning time, a recommended learning period calculated based on priority and the student's historical interaction status.

[0110] The output of this step is a personalized learning path and learning suggestions.

[0111] This application's embodiments unify heterogeneous diagnostic information into comparable confidence scores by performing weighted confidence fusion on inference diagnosis, dataset priors, and multimodal perception results, effectively improving the reliability of diagnostic conclusions. By extracting target knowledge points from multiple sources in the fusion results and constructing the original path by combining prior paths and graph snapshots, and then performing four-fold legalization processing—insertion of prerequisite knowledge points, deduplication, bridging retention, and length restrictions—the generated learning path strictly meets the prerequisite dependencies of knowledge points and difficulty gradient constraints. By comprehensively calculating the priority of knowledge points from multiple dimensions—weakness, difficulty, error frequency, and dependency strength—and generating detailed learning suggestions including action descriptions and suggested times, the output personalized learning path possesses both knowledge structure integrity and student-specific adaptability, significantly improving the rationality and executability of the learning path.

[0112] In some embodiments of this application, step S6 includes: step S61, writing the interaction data of the current learning task into the interaction log, wherein the interaction data of the current learning task includes the task score, topic, key knowledge points, and error patterns; step S62, updating the student profile and student status based on the task score, knowledge point relevance, and error labels; step S63, updating the knowledge point co-occurrence edge weights of the dynamic knowledge graph based on the knowledge point co-occurrence frequency; and step S64, triggering a runtime circuit breaker mechanism when the inference diagnostic agent experiences an abnormal state, bypassing the inference diagnostic agent within the circuit breaker window and switching to the dataset prior optimization mode to continue outputting diagnostic results.

[0113] Step S6 is the closing stage of the entire adaptive learning method. Its core objective is to persistently write the results of this interaction into the student profile, knowledge graph, and interaction log, enabling continuous iteration of the system state. It also provides automatic rollback protection for runtime anomalies, ensuring the continuity of online teaching services. Step S6 includes steps S61-S64, and the specific implementation details of each step are as follows: Step S61: Interaction Log Recording This step writes key information from this learning task into the interaction log, providing data support for subsequent analysis and model optimization.

[0114] The core information of this task will be written into the interaction log. This includes: the task score, reflecting the overall quality of the student's response; the topic category, recording the knowledge domain to which this task belongs; the key knowledge points, marking the core knowledge points identified in this diagnostic; and the error pattern, recording the student's error attributions.

[0115] While writing to the log, two statistics are performed. The first is the knowledge point activity statistics, which calculates the frequency of each knowledge point in all recent interactions, reflecting which knowledge points are the current learning hotspots; the second is the knowledge point co-occurrence statistics, which records which knowledge points appear simultaneously in the same interaction, providing a basis for co-occurrence counting for subsequent dynamic graph edge weight updates.

[0116] Step S62: Student Status Update This step updates the student profile in multiple dimensions based on the results of this interaction, so that the student profile can continuously reflect the student's latest learning status.

[0117] Mastery Update: Based on the score of this task, the relevance of each knowledge point to this task, and the error tags, the system adjusts the student's mastery level for the corresponding knowledge points. For knowledge points answered correctly, the mastery level is increased, with the increase being positively correlated with the score and the difficulty of the knowledge point; that is, answering correctly on more difficult knowledge points yields a greater increase. For knowledge points answered incorrectly, the system decreases the mastery level, with the decrease also being positively correlated with the difficulty of the knowledge point.

[0118] Stability Update: Tracks consecutive answers to the same knowledge point. If a student answers the same knowledge point correctly multiple times in a row, the stability of that knowledge point is improved, indicating that the student has a relatively solid grasp of the knowledge point; if the student answers incorrectly multiple times in a row, the stability is reduced, indicating that there are still significant problems with that knowledge point.

[0119] Error label count update: Increase the count for each error pattern identified this time, which will be used for subsequent error pattern statistical analysis and identification of high-frequency error patterns.

[0120] Step S63: Dynamic graph edge weight update This step updates the edge weights between knowledge points in the dynamic knowledge graph based on the co-occurrence of knowledge points in this interaction, so that the graph can continuously reflect the latest knowledge point relationships.

[0121] Specifically, edge weight updates can be performed using the following formula: W_ij(t+1)=λ·W_ij(t)+(1-λ)·N_ij / Max(N_i,1) Where W_ij(t) represents the co-occurrence edge weight between knowledge point i and knowledge point j at time t, reflecting the correlation strength between the two in historical interactions; N_ij represents the number of times knowledge point i and knowledge point j co-occur in the current interaction, i.e. the newly added correlation evidence; N_i represents the total number of times knowledge point i appears, used for normalization processing; λ is a smoothing coefficient used to control the fusion ratio of historical edge weights and new evidence; Max(N_i,1) is used to avoid division by zero error, and is set to 1 when N_i is 0.

[0122] The new edge weights are obtained by weighting historical edge weights and current co-occurrence evidence. This design ensures the overall stability of the graph, preventing drastic fluctuations due to a single interaction, while also promptly reflecting newly emerging knowledge point relationships, allowing the graph to continuously evolve throughout the teaching process.

[0123] Step S64: Runtime Circuit Breaker and Rollback This step provides the system with an automatic detection and service guarantee mechanism for runtime exceptions, ensuring that the system can continue to output structured results when the external inference engine is unavailable.

[0124] The system continuously monitors the operational status of the inference diagnostic agent and triggers a circuit breaker when any of the following anomalies occur: connection timeout (the connection time with the inference diagnostic agent exceeds a preset threshold); read timeout (the time to read the result from the inference diagnostic agent exceeds a preset threshold); or structured output anomaly (the result returned by the inference diagnostic agent does not conform to the preset format definition).

[0125] After the circuit breaker is triggered, three operations are performed. First, runtime failure information is recorded, including the failure type, the timestamp of occurrence, and the request identifier, to facilitate subsequent troubleshooting. Second, within the preset circuit breaker window, subsequent requests directly bypass the inference diagnostic agent and no longer wait for its response. Third, the system switches to the dataset prior optimization mode, where the dataset prior optimization module continues to output knowledge point predictions, error patterns, learning paths, and learning suggestions.

[0126] This mechanism ensures that even if the external inference engine becomes unavailable due to network fluctuations, service overload, or program anomalies, the system can still continuously return structured diagnostic results and learning suggestions based on historical statistical patterns, meeting the stringent requirements for low latency and high availability in online teaching scenarios.

[0127] After completing steps S61-S64, output the diagnostic score, knowledge point identification results, error patterns, personalized learning paths, learning suggestions, updated student profiles, and graph visualization results as the complete output of this adaptive learning task.

[0128] This application's embodiments accumulate structured data for subsequent analysis through interactive log recording and statistics on knowledge point activity and co-occurrence relationships. Multi-dimensional updates of mastery, stability, and error labels ensure that student profiles continuously reflect students' latest learning status. An edge weight update formula with a smoothing coefficient allows the dynamic knowledge graph to maintain stability while promptly incorporating new knowledge point association evidence. A triple detection mechanism for connection timeouts, read timeouts, and structured output anomalies in the reasoning diagnostic agent, along with a fallback mechanism that automatically switches to the dataset prior optimization mode within a circuit breaker window, achieves comprehensive closed-loop updates and anomaly self-healing for student status, knowledge graph, and system services. This enables the system to continuously improve diagnostic accuracy and path rationality through iterative development, while simultaneously ensuring service continuity and low latency requirements in online teaching scenarios.

[0129] To further verify the actual effect of this application, the following experimental data were used for verification: Experiment 1: Stability Analysis of Error Pattern Recognition under Noisy Conditions To verify the diagnostic stability of this application under complex input conditions, simulation experiments were conducted using the project's built-in training / test separated synthetic dataset. The training set contained 209 samples, and the test set contained 91 samples. In noisy scenarios, a random degradation with a character retention rate of 0.65 was applied to the test set question stems, and a random degradation with a character retention rate of 0.72 was applied to the student answers to simulate situations such as missing question stems, noisy input, and incomplete expression.

[0130] like Figure 3 As shown, a simulation result diagram of an embodiment of this application is provided. Under noise-free conditions, the error pattern recognition accuracy of the method of this application is 83.52%; under noisy conditions, the error pattern recognition accuracy of the method of this application is 81.32%, which is higher than the 43.96% of the single semantic method, indicating that this application has a more stable error recognition capability under multimodal input and multi-agent collaborative conditions.

[0131] Experiment 2: Path Matching Rate and Processing Latency Analysis Furthermore, under the same test conditions, the path matching rate and average processing latency of the proposed method and the single semantic method were compared. The proposed method achieved a path matching rate of 97.53% under noisy conditions, which is higher than the 90.48% of the single semantic method. The average processing latency of the proposed method under noisy conditions was 5.729ms, which, although higher than the 3.742ms of the single semantic method, still maintained a low millisecond-level return, while achieving higher path stability and diagnostic completeness.

[0132] This application also provides a multimodal adaptive learning system 400 based on multi-agent collaboration and dynamic knowledge graph, such as... Figure 4 As shown, a schematic diagram of a multimodal adaptive learning system based on multi-agent collaboration and dynamic knowledge graph is provided in an embodiment of this application. The system 400 includes: The input acquisition unit 410 is used to acquire the input information of the current learning task and encapsulate it into a structured learning task input package. The input information of the current learning task includes required field information and optional field information. The required field information includes the question text, student answer and standard answer. The optional field information includes at least one of the following: speech transcription result, image data, image file name, learning style and historical interaction record. The multimodal preprocessing and perception unit 420 is used to perform multimodal preprocessing and perception on the input package of the structured learning task to obtain a multimodal feature package; The candidate pruning and task scheduling unit 430 is used to perform candidate sample pruning and task scheduling based on the multimodal feature package and output the candidate sample package. Parallel execution unit 440 is used to perform inference diagnosis, dataset prior analysis, dynamic graph reading and student status reading in parallel according to the candidate sample package, and obtain parallel execution result package; The learning path generation unit 450 is used to generate personalized learning paths and learning suggestions based on the multimodal feature package and the parallel execution result package, using a preset fusion strategy and a path legitimization strategy based on a dynamic knowledge graph. The result output unit 460 is used to update the student status and handle anomalies based on the interaction results of the current learning task, and output the final adaptive learning result.

[0133] It is understood that the multimodal adaptive learning system based on multi-agent collaboration and dynamic knowledge graphs described above can realize all the steps of the multimodal adaptive learning method based on multi-agent collaboration and dynamic knowledge graphs provided in the foregoing embodiments. The relevant explanations of the multimodal adaptive learning method based on multi-agent collaboration and dynamic knowledge graphs are applicable to the multimodal adaptive learning system based on multi-agent collaboration and dynamic knowledge graphs, and will not be repeated here.

[0134] Figure 5 This is a schematic diagram of the structure of a device according to an embodiment of this application. For example... Figure 5 As shown, the device includes one or more processors (or processing units), and may also include one or more memories coupled to the processors, and may also include a communication module coupled to the processors.

[0135] A communication module can be used to communicate with other devices or apparatuses, such as sending or receiving data and / or signals. A communication module may have at least one communication module for communication. A communication module may include any interface necessary for communicating with other devices. Exemplarily, a communication module may be a transceiver, circuit, bus, module, or other type of communication module.

[0136] The processor may include, but is not limited to, one or more of the following: a general-purpose computer, a special-purpose computer, a microcontroller, a digital signal processor (DSP), or a controller-based multi-core controller architecture. The device may have multiple processors, such as application-specific integrated circuit (ASIC) chips, which are time-dependent on a clock synchronized with the main processor.

[0137] The memory may include one or more non-volatile memories and one or more volatile memories. Examples of non-volatile memories include, but are not limited to, at least one of the following: read-only memory (ROM), electrically programmable read-only memory (EPROM), flash memory, hard disk, compact disc (CD), digital video disc (DVD), or other magnetic and / or optical storage. Examples of volatile memories include, but are not limited to, at least one of the following: random access memory (RAM), or other volatile memories that do not persist during the duration of a power outage.

[0138] A computer program consists of computer-executable instructions that are executed by an associated processor. Programs can be stored in ROM. A processor can perform any appropriate action and processing by loading the program into RAM.

[0139] Possible implementations of this application can be achieved through a program, enabling the communication device to execute any of the processes discussed in the foregoing embodiments. Possible implementations of this application can also be achieved through hardware or a combination of software and hardware.

[0140] In some implementations, the program may be tangibly contained in a computer-readable storage medium, which may include in a device (such as in memory) or other storage device accessible by the device. The program may be loaded from the computer-readable storage medium into RAM for execution. The computer-readable storage medium may include any type of tangible non-volatile memory, such as ROM, EPROM, flash memory, hard disk, CD, DVD, etc.

[0141] This application also provides a computer-readable storage medium storing computer instructions or program code thereon, which, when executed by a processor, causes the processor to perform the methods and functions involved in any of the above embodiments. A computer-readable medium can be any tangible medium that contains or stores a program for or relating to an instruction execution system, apparatus, or device. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. More detailed examples of computer-readable storage media include electrical connections with one or more wires, magnetic media (e.g., disks, floppy disks, hard disks, magnetic tapes, magnetic storage devices), optical media (e.g., optical storage devices, DVDs), semiconductor media (e.g., solid-state drives), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), or any suitable combination thereof.

[0142] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. Embodiments of this application also provide at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. This computer program product includes one or more computer-executable instructions, such as instructions included in a program module, which execute in a device on a target real or virtual processor to perform the processes, methods, and functions involved in any of the above embodiments. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0143] This application also proposes a computer program product, including a computer program or instructions that, when run on a computer, cause the computer to perform the processes, methods, and functions described in the above embodiments. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided as needed. The machine-executable instructions for the program modules can be executed locally or in a distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0144] Generally, the various embodiments of this application can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software, which can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are shown and described as block diagrams, flowcharts, or represented using some other illustration, it should be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as, as non-limiting examples, in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0145] It should be noted that although embodiments of this application have been described above with reference to the accompanying drawings, these embodiments are not independent of each other, and they can be combined to obtain other embodiments. The methods, situations, categories, and classifications of embodiments in this application are only for the convenience of description and should not constitute a special limitation. Various methods, categories, situations, and features in embodiments can be combined with each other if logically consistent. The various embodiments of this application can be arbitrarily combined to achieve different technical effects. The embodiments of this application will not list various combinations.

[0146] Furthermore, although the operation of the methods of this disclosure is described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Rather, the steps depicted in the flowcharts may be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps. It should also be noted that the features and functions of two or more devices according to this disclosure may be embodied in one device. Conversely, the features and functions of one device described above may be further divided and embodied by multiple devices.

[0147] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0148] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A multimodal adaptive learning method based on multi-agent collaboration and dynamic knowledge graph, characterized in that, The method includes the following steps: Step S1: Collect the input information of this learning task and encapsulate it into a structured learning task input package. The input information of this learning task includes required field information and optional field information. The required field information includes the question text, student answer and standard answer. The optional field information includes at least one of the following: speech transcription result, image data, image file name, learning style and historical interaction record. Step S2: Perform multimodal preprocessing and perception on the input package of the structured learning task to obtain a multimodal feature package; Step S3: Based on the multimodal feature package, perform candidate sample pruning and task scheduling, and output candidate sample package; Step S4: Based on the candidate sample package, perform inference diagnosis, dataset prior analysis, dynamic graph reading and student status reading in parallel to obtain parallel execution results; Step S5: Based on the multimodal feature package and the parallel execution result package, a personalized learning path and learning suggestions are generated using a preset fusion strategy and a path legitimization strategy based on a dynamic knowledge graph. Step S6: Update student status and handle anomalies based on the interaction results of this learning task, and output the final adaptive learning result.

2. The method according to claim 1, characterized in that, Step S2 includes: Step S21: Perform text normalization and keyword extraction on the question text, the student answers, the standard answers, and the speech-to-text results to obtain a set of text keywords; Step S22: Perform image enhancement and binarization processing on the image data to obtain an enhanced image and a binarized image; Step S23: Calculate image quality indicators and layout structure indicators based on the enhanced image and the binarized image; Step S24: Output scene labels, acquisition suggestions, and intelligent agent assistance prompts based on the image quality index and the layout structure index.

3. The method according to claim 1, characterized in that, Step S3 includes: Step S31: Construct a word set and a topic candidate set based on the question text, the student answers, and the standard answers; Step S32: Calculate the coarse ranking score of the candidate samples based on the word set, the topic candidate set, and the historical sample cache data, and determine the coarse ranking candidate sample set based on the coarse ranking score. Step S33: Use a preset similarity algorithm to perform fine ranking on the coarsely ranked candidate sample set to obtain the candidate sample package. The candidate sample package includes similar cases, candidate knowledge points, candidate error patterns, and candidate learning paths.

4. The method according to claim 1, characterized in that, The parallel execution results include structured reasoning diagnostic results, dataset prior analysis results, graph snapshots, and student status information. Step S4 includes: Step S41: Based on the question text, the student's answer, the standard answer, the learning style, and the candidate sample package, the reasoning diagnostic agent outputs the structured reasoning diagnostic result, which includes similarity, knowledge point candidates, error attribution, thought process, and diagnostic confidence. Step S42: Based on the historical sample cache data, the dataset prior analysis agent outputs the dataset prior analysis results, which include similar cases, prior learning paths, and error pattern statistics. Step S43: Use the dynamic knowledge graph reading strategy to read the prerequisite dependencies, related relationships and difficulty information of the key knowledge points to form the graph snapshot; Step S44: Use the adaptive student model to read the student's mastery of the corresponding knowledge point, the distribution of weak points, and the historical interaction status to obtain the student status information.

5. The method according to claim 4, characterized in that, The multimodal feature package includes multimodal perception results, and step S5 includes: Step S51: The structured reasoning diagnostic results, the dataset prior analysis results, and the multimodal perception results are fused using a preset fusion strategy to obtain a fusion result; Step S52: Extract the target knowledge point set, target error pattern, and focus knowledge point set from the fusion result; Step S53: Combine the prior learning path, the set of target knowledge points, the students' mastery level, and the map snapshot to form the original learning path; Step S54: Perform path legalization processing on the original learning path to obtain the personalized learning path. The path legalization processing includes inserting prerequisite knowledge points, removing duplicate knowledge points, retaining bridging nodes, and limiting the path length. Step S55: Calculate the priority of knowledge points in the personalized learning path based on the map snapshot and the student status information, and generate the learning suggestions based on the knowledge point priorities.

6. The method according to claim 1, characterized in that, Step S6 includes: Step S61: Write the interaction data of this learning task into the interaction log. The interaction data of this learning task includes the score, topic, key knowledge points and error patterns of this task. Step S62: Update the student profile and student status based on the score of this task, the relevance of the knowledge points, and the error tags; Step S63: Update the co-occurrence edge weights of knowledge points in the dynamic knowledge graph based on the number of times each knowledge point co-occurs. The update of the co-occurrence edge weights satisfies the following formula: W_ij(t+1)=λ·W_ij(t)+(1-λ)·N_ij / Max(N_i,1); Where W_ij(t) represents the co-occurrence edge weight between knowledge point i and knowledge point j at time t; N_ij represents the number of times knowledge point i and knowledge point j co-occur in this interaction; N_i represents the total number of times knowledge point i appears; λ is the smoothing coefficient; Max(N_i,1) is used to avoid division by zero; Step S64: If the inference diagnostic agent is in an abnormal state, the runtime circuit breaker mechanism is triggered. Within the circuit breaker window, the inference diagnostic agent is bypassed and the system switches to the dataset prior optimization mode to continue outputting diagnostic results.

7. A multimodal adaptive learning system based on multi-agent collaboration and dynamic knowledge graph, characterized in that, The system includes: The input acquisition unit is used to acquire the input information of this learning task and encapsulate it into a structured learning task input package. The input information of this learning task includes required field information and optional field information. The required field information includes the question text, student answer and standard answer. The optional field information includes at least one of the following: speech transcription result, image data, image file name, learning style and historical interaction record. A multimodal preprocessing and sensing unit is used to perform multimodal preprocessing and sensing on the input package of the structured learning task to obtain a multimodal feature package; The candidate pruning and task scheduling unit is used to perform candidate sample pruning and task scheduling based on the multimodal feature package and output the candidate sample package. The parallel execution unit is used to perform inference diagnosis, dataset prior analysis, dynamic graph reading and student status reading in parallel based on the candidate sample package, and obtain parallel execution results; The learning path generation unit is used to generate personalized learning paths and learning suggestions based on the multimodal feature package and the parallel execution result package, using a preset fusion strategy and a path legitimization strategy based on a dynamic knowledge graph. The result output unit is used to update the student status and handle anomalies based on the interaction results of the current learning task, and output the final adaptive learning result.

8. A device, characterized in that, include: processor; And a computer-readable storage medium storing computer program instructions that, when executed by the processor, cause the processor to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes computer program instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 6.

Images