Course dynamic optimization method and system, electronic equipment and storage medium

Abstract

The invention provides a course dynamic optimization method and system, an electronic device and a storage medium, and relates to the technical field of online education, and the method achieves the quantitative association of a course unit-knowledge fragment through a soft mapping matrix, enables the perplexity to be bidirectionally transmitted between the course unit and the knowledge fragment, and achieves the quantitative traceability. The comprehensive confusion degree is obtained by fusing the confusion degrees of the course unit layer and the knowledge fragment layer, so that the high-confusion course unit can be positioned more accurately and explainably, and fine optimization of the course can be realized; through a course optimization decision with cost constraint, automatic generation of a course optimization scheme with controllable cost is realized; besides, personalized learning path optimization of a course structure layer is realized, and a course optimization strategy continuously acts on subsequent learners, so that dynamic evolution of course contents and personalized learning paths is realized.

Description

Technical Field

[0001] This invention relates to the field of online education technology, and in particular to a method, system, electronic device, and storage medium for dynamic course optimization. Background Technology

[0002] With the development of online education and Large Language Model (LLM), more and more platforms can achieve the following functions: automatically convert resources such as documents, web pages, and videos into online courses; automatically generate quizzes and assignments and grade them; and provide AI tutor dialogues based on course content.

[0003] However, according to the inventor's analysis, the existing technology still has the following shortcomings: 1. Lack of end-to-end traceability: It is difficult to trace the source of a course knowledge point, test question, or AI Tutor reply back to the document page number, web page paragraph, or video timestamp; teachers and regulators find it difficult to verify the accuracy and compliance of the content.

[0004] 2. Disconnect between learning behavior analysis and course optimization: Existing learning management systems (LMS) mostly remain at the statistical level (completion rate, accuracy rate, duration, etc.), and mainly rely on manual reading of reports and experience to revise courses; the system itself rarely diagnoses problems or generates optimization solutions automatically based on behavioral data, and it is even more difficult to form a continuous "generation-feedback-regeneration" closed loop.

[0005] 3. Personalized learning paths often remain at the question level: Adaptive learning systems mostly adjust the question bank based on question difficulty and answer accuracy; they rarely perform unified modeling of "course unit - source content - tutor dialogue - behavioral data" and then optimize personalized learning paths at the course structure level. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, electronic device, and storage medium for dynamic course optimization, so as to at least solve one of the above-mentioned problems.

[0007] In a first aspect, the present invention provides a method for dynamic optimization of courses, comprising: At least one source learning resource is parsed and segmented to obtain multiple knowledge fragments and their identification data. The knowledge fragment identification data includes fragment identifier, source type, source location, text content, and vector representation. An initial course consisting of multiple course units is constructed based on a pre-defined course outline, and a soft mapping matrix is ​​generated, which consists of soft mapping weights between each course unit and each knowledge segment. The soft mapping weights are obtained based on the semantic similarity between the corresponding course unit and the corresponding knowledge segment. During the process of learners learning the current course, learner learning behavior data is acquired and standardized to obtain an event set consisting of learning behavior events; where the current course includes the initial course or the optimized new version course, and the learning behavior events include learner identifier, course unit identifier, event type, timestamp and event parameters; Based on the event set and soft mapping matrix, the overall confusion level of each relevant course unit is determined. The overall confusion level integrates the first confusion level of the course unit level and the second confusion level of the knowledge segment level. For high-confusion course units determined based on comprehensive confusion level, under preset cost constraints, a combination of target optimization actions is selected from multiple candidate optimization actions to generate a course optimization scheme; the current course is optimized based on the course optimization scheme, and the optimized new version of the course is released. Based on the comprehensive confusion level and cognitive load estimate of each candidate course unit, as well as the learner's current mastery level and selection frequency of each candidate course unit, the next course unit to be learned is planned for the learner; among them, the candidate course units are course units for which all prerequisites have been met but not completed.

[0008] In an optional implementation, the soft mapping matrix is ​​determined through the following process: The title, learning objectives, and keywords of each course unit are encoded to obtain a course unit vector, and the text content of each knowledge segment is encoded to obtain a knowledge segment vector. Calculate the cosine similarity between each pair of course unit vectors and knowledge segment vectors; Based on the cosine similarity of all knowledge segments corresponding to the same course unit, the cosine similarity of each knowledge segment corresponding to the course unit is normalized to obtain the soft mapping weight between the course unit and the corresponding knowledge segment.

[0009] In an optional implementation, the overall confusion level of each relevant course unit is determined based on the event set and the soft mapping matrix, including: Based on the event set, determine the first level of confusion for each relevant course unit at the course unit level; Using a soft mapping matrix, the first perplexity is propagated to the knowledge fragment layer to obtain the second perplexity of the corresponding course unit at the knowledge fragment layer; The overall confusion level of each course unit is obtained by combining the first and second confusion levels.

[0010] In an optional implementation, based on the event set, the first level of confusion for each relevant course unit at the course unit level is determined, including: For each course unit, at least two of the following metrics are collected from its relevant learning behavior events: error rate, repeat visit intensity, help-seeking density, average dwell time, and bounce rate. The statistically obtained indicator values ​​are normalized. The first perplexity of the course unit level is obtained by weighted summation of the normalized index values.

[0011] In an optional implementation, the first and second perplexity of each course unit are combined to obtain the comprehensive perplexity of the course unit, including: For each course unit, obtain all relevant knowledge fragments associated with it through a soft mapping matrix; Determine the maximum value of the second perplexity of all relevant knowledge fragments in the knowledge fragment layer; The overall perplexity of a course unit is obtained by weighted summing of the first perplexity and the maximum perplexity of the course unit.

[0012] In an optional implementation, under a preset cost constraint, a combination of target optimization actions is selected from multiple candidate optimization actions to generate a course optimization scheme, including: For each highly confusing course unit, obtain a set of candidate optimization actions, which includes at least one of the following: supplementing explanations, adding exercises, splitting the unit, and adjusting the order. Estimate the expected confusion reduction benefit and implementation cost for each candidate optimization action in the candidate optimization action set; With the goal of maximizing total revenue under the constraints of total cost budget and the number of optimized actions applied in a single course unit, a greedy approximation algorithm is used to select the target optimized action combination from the candidate optimized action set.

[0013] In an optional implementation, based on the overall confusion and cognitive load estimates of each candidate course unit, as well as the learner's current mastery level and selection frequency for each candidate course unit, the next course unit to be learned is planned for the learner, including: The current course structure is abstracted into a directed acyclic graph, and all candidate course units for the learner are identified. For each candidate course unit, the score is calculated using the following formula: ; in, S L ( u j ) as candidate course units u j The rating,G L ( u j (Based on candidate course units) u j The expected knowledge gain is obtained by calculating the corresponding current mastery level and overall confusion level. CL L ( u j ) as candidate course units u j Cognitive load estimation, Explore L ( u j (Based on candidate course units) u j The corresponding number of selections and the exploration items determined using the multi-armed gambling machine concept. η and ξ All are preset weighting coefficients; The candidate course unit with the highest score is selected as the next course unit.

[0014] Secondly, the present invention provides a dynamic course optimization system, comprising: The resource parsing and segmentation module is used to parse and segment at least one source learning resource to obtain multiple knowledge fragments and their knowledge fragment identification data. The knowledge fragment identification data includes fragment identifier, source type, source location, text content, and vector representation. The traceable course construction module is used to construct an initial course consisting of multiple course units based on a preset course outline, and generate a soft mapping matrix composed of soft mapping weights between each course unit and knowledge segment; wherein, the soft mapping weights are obtained based on the semantic similarity between the corresponding course unit and the corresponding knowledge segment. The learning behavior acquisition module is used to acquire and standardize learner learning behavior data during the process of learners learning the current course, and obtain an event set consisting of learning behavior events; wherein, the current course includes the initial course or the optimized new version course, and the learning behavior event includes learner identifier, course unit identifier, event type, timestamp and event parameters; The course unit confusion determination module is used to determine the comprehensive confusion of each relevant course unit based on the event set and the soft mapping matrix. The comprehensive confusion combines the first confusion of the course unit layer and the second confusion of the knowledge fragment layer. The course optimization and release module is used to select a combination of target optimization actions from multiple candidate optimization actions to generate a course optimization plan for high-confusion course units determined based on comprehensive confusion level, under preset cost constraints; optimize the current course based on the course optimization plan, and release the optimized new version of the course. The learning path planning module is used to plan the next course unit for learners based on the comprehensive confusion and cognitive load estimates of each candidate course unit, as well as the learner's current mastery of each candidate course unit and the number of times they have selected it. The candidate course units are those for which all prerequisites have been met but not yet completed.

[0015] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the course dynamic optimization method of any of the foregoing embodiments.

[0016] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the course dynamic optimization method of any of the foregoing embodiments.

[0017] The present invention provides a method, system, electronic device, and storage medium for dynamic course optimization. The method includes: parsing and segmenting at least one source learning resource to obtain multiple knowledge fragments and their identification data, wherein the identification data includes fragment identifier, source type, source location, text content, and vector representation; constructing an initial course consisting of multiple course units based on a preset course syllabus, and generating a soft mapping matrix composed of soft mapping weights between each course unit and each knowledge fragment; wherein the soft mapping weights are obtained based on the semantic similarity between the corresponding course unit and the corresponding knowledge fragment; and acquiring and standardizing learner learning behavior data during the learner's current course learning process to obtain an event set composed of learning behavior events; wherein the current course includes the initial course or an optimized new version course, and the learning behavior... Events include learner identifiers, course unit identifiers, event types, timestamps, and event parameters. Based on the event set and soft mapping matrix, the overall confusion level of each relevant course unit is determined, which integrates the first confusion level at the course unit level and the second confusion level at the knowledge fragment level. For high-confusion course units determined based on the overall confusion level, under preset cost constraints, a combination of target optimization actions is selected from multiple candidate optimization actions to generate a course optimization plan. The current course is optimized based on the course optimization plan, and a new version of the optimized course is released. Based on the overall confusion level and cognitive load estimation of each candidate course unit, as well as the learner's current mastery level and selection frequency for each candidate course unit, the next course unit to be learned is planned for the learner. Candidate course units are those for which all prerequisites have been met but not yet completed. This approach, using a soft mapping matrix to quantify the association between course units and knowledge fragments, allows confusion levels to propagate bidirectionally between them, thus enabling quantifiable traceability. By integrating the confusion levels from both the course unit and knowledge fragment layers to obtain a comprehensive confusion level, high-confusion course units can be located more accurately and interpretably, facilitating refined course optimization. Cost-constrained course optimization decisions enable the automatic generation of cost-controlled course optimization solutions. Furthermore, it achieves personalized learning path optimization at the course structure layer and continuously applies course optimization strategies to subsequent learners, realizing the dynamic evolution of course content and personalized learning paths. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1A flowchart illustrating a dynamic course optimization method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a course dynamic optimization system provided in an embodiment of the present invention; Figure 3 A schematic diagram of the overall system architecture provided in an embodiment of the present invention; Figure 4 A flowchart illustrating another course dynamic optimization method provided in an embodiment of the present invention; Figure 5 A schematic diagram of a traceable data model provided in an embodiment of the present invention; Figure 6 A flowchart illustrating a learning behavior analysis and two-layer perplexity estimation method provided in an embodiment of the present invention; Figure 7 A flowchart illustrating a personalized learning path loop and version iteration closed loop provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

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

[0021] Analysis of existing technologies reveals a need for a new technical solution: one that can establish a fine-grained, traceable link from source learning resources to courses, assessments, and AI tutoring; one that can also feed back behavioral data to drive dynamic optimization of course content and structure; and one that enables individualized path adaptation. Based on this, embodiments of the present invention provide a method, system, electronic device, and storage medium for dynamic course optimization based on a traceable knowledge link and a closed loop of learning behavior, which can achieve at least one of the following technical effects: 1. Establish a fine-grained traceability relationship from source learning resources to course structure, assessment content, and AI tutoring dialogues; 2. Standardize the collection of learners' multidimensional behavioral data in the course (watching, answering questions, engaging in dialogue, exiting, etc.) and associate it with specific course units and their source materials; 3. Based on learning behavior data, automatically identify "high-confusion points" and "high-value resources" in the course and generate quantifiable indicators; 4. Based on the recognition results, automatically generate course optimization plans, including supplementary explanations, changing the order, splitting / merging lesson sections, adjusting the density of exercises, etc., and generate personalized learning paths; 5. Continuously apply course optimization strategies to subsequent learners to achieve dynamic evolution of course content and personalized learning paths.

[0022] To facilitate understanding of this embodiment, a method for dynamic optimization of courses disclosed in this embodiment of the invention will first be described in detail.

[0023] This invention provides a method for dynamic course optimization. This method can be executed by electronic devices with specific data processing capabilities and is applicable to scenarios such as AI course generation platforms, learning management systems, and intelligent teaching aid platforms. See also... Figure 1 The diagram shows a flowchart of a dynamic course optimization method, which mainly includes the following steps S110 to S160: Step S110: parse and segment at least one source learning resource to obtain multiple knowledge fragments and their identifier data.

[0024] The knowledge fragment identification data includes fragment identifiers, source types, source locations, text content, and vector representations. In practice, different types of source learning resources, such as documents, web pages, and videos, are first parsed (e.g., text extraction, chapter structure recognition, speech-to-text transcription), and then segmented into semantically relatively complete knowledge fragments. Subsequently, a set of standardized identifier data is generated for each knowledge fragment. The fragment identifier ensures its uniqueness, the source type and source location record its original source for accurate traceability, the text content carries core knowledge information, and the vector representation is converted into computer-processable mathematical vectors through semantic encoding, laying the foundation for subsequent semantic similarity-based course construction, content retrieval, and intelligent association.

[0025] Step S120: Construct an initial course consisting of multiple course units based on a preset course outline, and generate a soft mapping matrix composed of soft mapping weights between each course unit and knowledge segment; wherein, the soft mapping weights are obtained based on the semantic similarity between the corresponding course unit and the corresponding knowledge segment.

[0026] In this embodiment, the teaching content is divided into multiple course units according to the preset course syllabus. A soft mapping matrix is ​​generated by calculating the semantic similarity between the descriptive text (such as title and objective) of each course unit and the text of each knowledge segment. The weight values ​​(soft mapping weights) in the soft mapping matrix quantitatively reflect the contribution of each knowledge segment to the construction of the knowledge content of a specific course unit. This establishes a flexible and quantifiable traceability link while maintaining the logical structure of the course, providing a data foundation for subsequent accurate problem diagnosis and content optimization based on the knowledge source.

[0027] In some possible embodiments, the soft mapping matrix described above can be determined through the following process: encoding the title, learning objectives, and keywords of each course unit to obtain a course unit vector, and encoding the text content of each knowledge segment to obtain a knowledge segment vector; calculating the cosine similarity between each pair of course unit vectors and knowledge segment vectors; and normalizing the cosine similarity of each knowledge segment corresponding to the same course unit based on the cosine similarity of all knowledge segments corresponding to the same course unit to obtain the soft mapping weight between the course unit and the corresponding knowledge segment.

[0028] Step S130: During the process of the learner learning the current course, the learner's learning behavior data is acquired and standardized to obtain an event set consisting of learning behavior events; wherein, the current course includes the initial course or the optimized new version course, and the learning behavior events include learner identifier, course unit identifier, event type, timestamp and event parameters.

[0029] In this embodiment, raw behavioral data generated by learners during their interaction with the course (such as watching videos, answering questions, and conversing with AI teaching assistants) can be continuously collected through data tracking or interfaces. This data is then standardized into structured learning behavior events according to a unified standard. Each learning behavior event includes key information such as learner identifier (who is being identified), course unit identifier (where is being identified), event type (what was being done), timestamp (when was being identified), and event parameters (recording specific details such as correct or incorrect answers, dialogue content, duration, and score), thus forming a complete and traceable event set. This standardization process ensures the comparability and aggregability of multi-source heterogeneous behavioral data, providing a high-quality data foundation for subsequent accurate quantitative analysis and confusion calculation.

[0030] Step S140: Based on the event set and the soft mapping matrix, determine the comprehensive perplexity of each relevant course unit. The comprehensive perplexity integrates the first perplexity of the course unit layer and the second perplexity of the knowledge fragment layer.

[0031] In this embodiment, at the course unit level, a first level of confusion is calculated by aggregating all learning behavior events related to that course unit (such as the error rate of answering questions, the number of times the lesson is repeated, the frequency of seeking help, etc.). This first level directly reflects the degree of difficulty encountered by the learner in that course unit. Then, using the quantitative correlation defined by the soft mapping matrix, the confusion weight of the course unit level is allocated and propagated to each of its associated knowledge segments to calculate a second level of confusion. This second level of confusion is used to identify which underlying knowledge points or teaching content itself has problems. Finally, the confusion levels of the course unit level and the knowledge segment level are merged to obtain a comprehensive level of confusion. This allows us to discover "which course unit" is difficult from a macro perspective and diagnose "which specific knowledge points or content in the course unit" cause the difficulty from a micro perspective, achieving precise problem localization from phenomenon to root cause.

[0032] In some possible embodiments, step S140 above may include: determining the first perplexity of each relevant course unit at the course unit layer based on the event set; propagating the first perplexity to the knowledge fragment layer using a soft mapping matrix to obtain the second perplexity of the corresponding course unit at the knowledge fragment layer; and fusing the first and second perplexities of each course unit to obtain the comprehensive perplexity of the course unit.

[0033] In one possible implementation, determining the first level of confusion for each relevant course unit at the course unit level based on the event set may include: for each course unit, statistically analyzing at least two of the following indicators from its relevant learning behavior events: error rate, repeated access intensity, help-seeking density, average dwell time, and bounce rate; normalizing the statistically obtained indicator values; and weighting and summing the normalized indicator values ​​to obtain the first level of confusion at the course unit level.

[0034] In one possible implementation, the first and second perplexity of each course unit are combined to obtain the comprehensive perplexity of the course unit. This may include: for each course unit, obtaining all relevant knowledge fragments associated with it through a soft mapping matrix; determining the maximum value of the second perplexity of all relevant knowledge fragments in the knowledge fragment layer; and weighted summing the first perplexity of the course unit with the maximum value to obtain the comprehensive perplexity of the course unit.

[0035] Step S150: For high-confusion course units determined based on comprehensive confusion level, under preset cost constraints, select a combination of target optimization actions from multiple candidate optimization actions to generate a course optimization scheme; optimize the current course based on the course optimization scheme, and release the optimized new version of the course.

[0036] In this embodiment, high-confusion course units requiring intervention can be identified first based on overall confusion level. These units are then treated as a resource allocation problem: a series of candidate optimization actions (such as supplementing case studies, adding exercises, splitting units, etc.) are evaluated for each high-confusion unit, and the expected effects (such as the benefit of reduced confusion level) and implementation costs (such as development time and computing resources) of each action are estimated. Then, under preset cost constraints (such as total budget and maximum number of actions per course unit), a target optimization action combination is selected from all candidate optimization actions using an optimization algorithm (such as a greedy strategy). This combination aims to achieve the greatest overall improvement in teaching effectiveness with limited resource input. Finally, the course optimization plan can be executed automatically (such as calling a content generation model to create supplementary materials and adjusting the course structure), and a new version of the course can be released, thus completing the closed loop from "problem diagnosis" to "solution generation and deployment," ensuring that course content can continuously iterate based on real learning data.

[0037] In some possible embodiments, the above-mentioned target optimization action combination can be determined as follows: for each high-confusion course unit, obtain a candidate optimization action set, which includes at least one of supplementary explanation, adding exercises, splitting the unit, and adjusting the order; estimate the expected confusion reduction benefit and implementation cost for each candidate optimization action in the candidate optimization action set; with the goal of maximizing the total benefit under the constraints of the total cost budget and the number of optimization actions applied per course unit, select the target optimization action combination from the candidate optimization action set using a greedy approximation algorithm.

[0038] Step S160: Based on the comprehensive confusion level and cognitive load estimate of each candidate course unit, as well as the learner's current mastery level and selection frequency of each candidate course unit, plan the next course unit to be learned for the learner; wherein, the candidate course unit is a course unit for which all prerequisites have been met but not completed.

[0039] In this embodiment, at each path decision, all available candidate course units for the learner (i.e., course units for which prerequisites have been met and which the learner has not yet studied) are first identified. Then, a score is calculated for each candidate course unit. This score weighs three key factors: first, the expected knowledge gain, determined by the unit's overall perplexity (representing the general difficulty value of the content) and the learner's current mastery estimate (representing the personal value to the learner), prioritizing high-value units that are "difficult for the general public but not for me personally"; second, cognitive load estimation, avoiding recommending overly long or complex units that could lead to learner burnout; and third, the exploration factor, proactively guiding learners to explore units with fewer selections by recording their historical selection frequency, preventing them from getting trapped in information cocoons or missing potentially important content. Ultimately, the candidate course unit with the highest score is selected as the next course unit, thus achieving a personalized adaptive learning path that efficiently improves knowledge level while also considering learning experience and breadth of knowledge.

[0040] In some possible embodiments, step S160 above may include: abstracting the course structure of the current course into a directed acyclic graph (DAG), and determining all candidate course units for the learner; for each candidate course unit, calculating the score using the following formula: ;in, S L ( u j ) as candidate course units u j The rating, G L ( u j (Based on candidate course units) u j The expected knowledge gain is obtained by calculating the corresponding current mastery level and overall confusion level. CL L ( u j ) as candidate course units u j Cognitive load estimation, Explore L ( u j (Based on candidate course units) u j The corresponding number of selections and the exploration items determined using the multi-armed gambling machine concept. η and ξ All are preset weighting coefficients; the candidate course unit with the highest score is selected as the next course unit.

[0041] The course dynamic optimization method provided in this invention achieves quantitative correlation between course units and knowledge fragments through a soft mapping matrix, enabling bidirectional propagation of confusion levels between course units and knowledge fragments, thus achieving quantifiable traceability. By integrating the confusion levels from both the course unit and knowledge fragment layers to obtain a comprehensive confusion level, high-confusion course units can be located more accurately and interpretably, facilitating refined course optimization. Through cost-constrained course optimization decisions, cost-controllable course optimization schemes are automatically generated. Furthermore, personalized learning path optimization at the course structure layer is achieved, and the course optimization strategy is continuously applied to subsequent learners, realizing the dynamic evolution of course content and personalized learning paths.

[0042] Corresponding to the above-described dynamic course optimization method, this embodiment of the invention also provides a dynamic course optimization system. See [link to related document]. Figure 2 The diagram shown illustrates the structure of a dynamic course optimization system. The system includes: The resource parsing and segmentation module 201 is used to parse and segment at least one source learning resource to obtain multiple knowledge fragments and their knowledge fragment identification data. The knowledge fragment identification data includes fragment identifier, source type, source location, text content, and vector representation. The traceable course construction module 202 is used to construct an initial course consisting of multiple course units based on a preset course outline, and generate a soft mapping matrix composed of soft mapping weights between each course unit and knowledge segment; wherein, the soft mapping weights are obtained based on the semantic similarity between the corresponding course unit and the corresponding knowledge segment. The learning behavior acquisition module 203 is used to acquire and standardize the learner's learning behavior data during the process of the learner learning the current course, and obtain an event set consisting of learning behavior events; wherein, the current course includes the initial course or the optimized new version course, and the learning behavior event includes learner identifier, course unit identifier, event type, timestamp and event parameters; The course unit confusion determination module 204 is used to determine the comprehensive confusion of each relevant course unit based on the event set and the soft mapping matrix. The comprehensive confusion integrates the first confusion of the course unit layer and the second confusion of the knowledge fragment layer. The course optimization and release module 205 is used to select a combination of target optimization actions from multiple candidate optimization actions to generate a course optimization plan for high-confusion course units determined based on comprehensive confusion level, under preset cost constraints; optimize the current course based on the course optimization plan, and release the optimized new version of the course. The learning path planning module 206 is used to plan the next course unit for learners based on the comprehensive confusion and cognitive load estimation of each candidate course unit, as well as the learner's current mastery and selection frequency of each candidate course unit; wherein, the candidate course unit is a course unit for which all prerequisites have been met but not completed.

[0043] The course dynamic optimization system provided in this invention achieves quantitative correlation between course units and knowledge fragments through a soft mapping matrix, enabling bidirectional propagation of confusion levels between course units and knowledge fragments, thus achieving quantifiable traceability. By integrating the confusion levels from both the course unit and knowledge fragment layers to obtain a comprehensive confusion level, it can more accurately and interpretably locate high-confusion course units, thereby facilitating refined course optimization. Through cost-constrained course optimization decisions, it achieves the automatic generation of cost-controllable course optimization schemes. In addition, it also realizes personalized learning path optimization at the course structure layer and continuously applies course optimization strategies to subsequent learners, achieving dynamic evolution of course content and personalized learning paths.

[0044] Furthermore, the aforementioned traceable course construction module 202 is specifically used for: encoding the title, learning objectives, and keywords of each course unit to obtain a course unit vector; encoding the text content of each knowledge segment to obtain a knowledge segment vector; calculating the cosine similarity between each pair of course unit vectors and knowledge segment vectors; and normalizing the cosine similarity of each knowledge segment corresponding to the same course unit based on the cosine similarity of all knowledge segments corresponding to the same course unit to obtain the soft mapping weight between the course unit and the corresponding knowledge segment.

[0045] Furthermore, the aforementioned course unit perplexity determination module 204 is specifically used to: determine the first perplexity of each relevant course unit at the course unit level based on the event set; propagate the first perplexity to the knowledge fragment level using a soft mapping matrix to obtain the second perplexity of the corresponding course unit at the knowledge fragment level; and merge the first and second perplexities of each course unit to obtain the comprehensive perplexity of the course unit.

[0046] Furthermore, the course unit confusion determination module 204 is also used to: for each course unit, statistically analyze at least two of the following indicators from its related learning behavior events: error rate, repeated access intensity, help-seeking density, average dwell time, and bounce rate; normalize the statistically obtained indicator values; and perform weighted summation on the normalized indicator values ​​to obtain the first confusion level of the course unit.

[0047] Furthermore, the course unit perplexity determination module 204 is also used to: for each course unit, obtain all relevant knowledge fragments associated with it through the soft mapping matrix; determine the maximum value of the second perplexity of all relevant knowledge fragments in the knowledge fragment layer; and perform a weighted summation of the first perplexity of the course unit and the maximum value to obtain the comprehensive perplexity of the course unit.

[0048] Furthermore, the course optimization and release module 205 is specifically used for: obtaining a set of candidate optimization actions for each high-confusion course unit, the set of candidate optimization actions including at least one of supplementary explanation, adding exercises, splitting the unit, and adjusting the order; estimating the expected confusion reduction benefit and implementation cost for each candidate optimization action in the set of candidate optimization actions; and selecting the target optimization action combination from the set of candidate optimization actions with the goal of maximizing the total benefit under the constraints of the total cost budget and the number of optimization actions applied in a single course unit, using a greedy approximation algorithm.

[0049] Furthermore, the learning path planning module 206 described above is specifically used to: abstract the course structure of the current course into a directed acyclic graph, and determine all candidate course units for the learner; for each candidate course unit, calculate the score using the following formula: ; in, S L ( u j ) as candidate course units u j The rating, G L ( u j (Based on candidate course units) u j The expected knowledge gain is obtained by calculating the corresponding current mastery level and overall confusion level. CL L ( u j ) as candidate course units u j Cognitive load estimation, Explore L ( u j (Based on candidate course units) u j The corresponding number of selections and the exploration items determined using the multi-armed gambling machine concept. η and ξ All are preset weighting coefficients; Select the candidate course unit with the highest score as the next course unit.

[0050] The course dynamic optimization system provided in this embodiment has the same implementation principle and technical effects as the aforementioned course dynamic optimization method embodiment. For the sake of brevity, any parts not mentioned in the course dynamic optimization system embodiment can be referred to the corresponding content in the aforementioned course dynamic optimization method embodiment.

[0051] For ease of understanding, the course dynamic optimization method and system provided in the embodiments of the present invention will be described in detail below.

[0052] The embodiments of this invention mainly propose: 1. Traceable Knowledge Link Construction Method: Parse source learning resources such as documents, web pages, and videos into knowledge fragments and encode them into vectors; construct a soft mapping of "course unit - knowledge fragment" through vector similarity + softmax weight to achieve quantifiable traceability.

[0053] 2. Two-layer perplexity estimation method: Standardize learning behavior events (viewing, answering questions, asking for help from a tutor, rewatching, etc.); calculate the perplexity of course units based on unit-level behavior indicators, and then propagate the perplexity to the segment level through mapping weights to form a "unit-segment" two-layer perplexity model.

[0054] 3. Decision-making method for course optimization strategy with cost constraints: Model optimization actions such as supplementary explanation, adding exercises, splitting units, and adjusting the order as a "benefit-cost" decision problem; automatically select the optimal action combination under resource constraints through multi-objective optimization (or greedy approximation).

[0055] 4. Combining Bandit's personalized learning path planning method: Abstracting the course into a directed acyclic graph; for candidate course units, dynamically planning paths for individual learners based on expected knowledge gain, cognitive load estimation, and multi-armed gambler (using a UCB (Upper Confidence Bound) type strategy) exploration item scoring.

[0056] like Figure 3 As shown, the overall architecture (i.e., the overall system architecture) of the course dynamic optimization system provided in this embodiment of the invention includes: 1. Resource Acquisition and Parsing Module 301: This tool is used to collect data from various learning resources, including text documents, web page content, and audio / video resources. It performs operations such as format parsing, text extraction, audio transcription, and chapter recognition, dividing the source resources into multiple knowledge fragments and generating a unique fragment identifier and metadata for each fragment. It can also be used to generate initial course outlines.

[0057] 2. Traceable Course Construction Module 302: This is used to construct course units based on the course syllabus and establish a soft mapping relationship between course units and knowledge fragments; the mapping relationship is stored in the form of a weight matrix for subsequent confusion propagation and source tracing query.

[0058] 3. Learning Behavior Collection Module 303: This is used to collect learners' browsing, answering, homework submission, AI tutoring dialogues, media playback, and other behaviors when using the course platform; and to standardize these behaviors into learning behavior events in a unified format and write them into the event storage.

[0059] 4. Learning Behavior Analysis and Confusion Node Identification Module 304: It is used to aggregate and analyze learning behavior events, calculate multi-dimensional indicators at the course unit level and knowledge segment level; and automatically identify high-confusion units and high-confusion segments in the course based on a two-layer confusion estimation algorithm.

[0060] 5. Course Optimization Decision Module 305: Based on the results of confusion analysis and the action candidate set, this algorithm uses a cost-constrained decision-making algorithm to determine the optimal action combination and generate a course optimization plan that includes supplementary explanations, structural adjustments, and practice reinforcement.

[0061] 6. Course Publishing and Path Arrangement Module 306: It is used to manage course versions, publish updated courses, and create personalized learning paths for different learners; during the learning process, it calls path planning algorithms to dynamically select the next learning unit for learners.

[0062] 7. Data Storage and Management Module 307: It is used to store source learning resources, knowledge fragments, course units, mapping relationships, learning behavior events, indicator data, and course version information; and provides a unified data access interface to the outside world. For example... Figure 3 As shown, the data storage and management module 307 may include an indicator and confusion result library, a course unit and version library, a knowledge fragment library, a learning behavior event library, and a source learning resource library.

[0063] The modules mentioned above can be deployed on the same server (i.e., the server side), or they can be deployed in a distributed cluster or microservice in a cloud platform environment and communicate with each other over the network.

[0064] like Figure 3As shown, the teacher's end is connected to the course publishing and path arrangement module 306. The teacher's end displays a teacher interface, which can be used for course editing and version management. Teachers can view / manage optimization plans (corresponding to the course optimization decision module 305), view analysis results / confirm optimizations (corresponding to the learning behavior analysis and confusion node identification module 304), and upload resources / edit courses (corresponding to the resource acquisition and parsing module 301) through the teacher's end. The learning terminal is also connected to the course publishing and path arrangement module 306. The learning terminal displays a learner interface, which can be used for course playback / quizzes / assignments / AI Tutors. Learners can browse courses / answer questions / assignments / AI dialogues through the learning terminal.

[0065] like Figure 4 As shown, the course dynamic optimization method provided in this embodiment of the invention includes steps S401 to S406.

[0066] S401: Source learning resource collection and knowledge fragment construction.

[0067] Content is collected from at least one source learning resource, which includes at least one of text documents, web page content, and audio / video resources. The source learning resource is parsed and segmented to obtain multiple knowledge fragments, and knowledge fragment identification data is generated for each knowledge fragment. The knowledge fragment identification data includes: fragment identifier (fragment_id); source type (e.g., document / web / video); source location (e.g., document page number and paragraph, web page DOM path, video timestamp range, etc.); fragment text content or audio transcription content; vector representation embedding; and other metadata.

[0068] Specifically, the system can acquire learning resources from multiple heterogeneous data sources, including but not limited to: Document resources: such as PDF, Word, PPT, etc.; Web resources: such as course introduction pages, standard entry pages, blog posts, etc.; Multimedia resources: such as instructional videos, audio explanations, etc.

[0069] For document and web page resources, the system extracts plain text content and structural information (chapter titles, paragraphs, lists, etc.) through a parsing engine, and then segments them according to preset rules: using chapter / section titles as the main delimiters; controlling the number of words or tokens within a single segment to not exceed a limit, such as 512-1024 tokens; and maintaining semantic integrity as much as possible. For multimedia resources, the system calls a speech recognition model to transcribe the audio portion while retaining timestamp information; then segments the data according to time windows or semantic boundaries.

[0070] For each knowledge fragment, the following data structure is generated: fragment_id: A globally unique identifier; source_type: such as "document" "web" "video"; source_loc: such as "Document A, Page 3, Paragraph 2" or "Video B, 00:10:05–00:10:50 interval"; content: Plain text content or transcription result; Embedding: The vector representation obtained through the text encoding model Enc(·); metadata includes the chapter title, topic tags, and difficulty estimate.

[0071] Knowledge fragments are stored in a knowledge fragment library, which is used for subsequent course construction and retrieval.

[0072] S402: Construction of traceable curriculum units and generation of soft mapping relationships.

[0073] Based on the pre-generated course syllabus, the course is divided into multiple course units. (This includes lesson units, test units, and homework units.)

[0074] Let the set of knowledge fragments be... For each course unit u j The system concatenates its title, learning objectives, keywords, and other information into text. t j .

[0075] 1. Using the encoding model Enc(·) to calculate the vector representation, we obtain: , ; in, t j For unit u j The title, target, and keywords are combined into a single text.

[0076] 2. Calculate the similarity between course units and knowledge segments: .

[0077] 3. To obtain measurable coverage weights, softmax normalization is introduced, i.e., for a fixed... j Perform softmax: ; Where λ>0 is a temperature parameter used to control the smoothness of the weight distribution. When the threshold τ is configurable, the fragment is considered... f i For unit u j Make a substantial contribution and record the mapping relationship: ( u j → f i , w j,i ), thus forming a soft mapping matrix of "course unit – knowledge fragment" ( W =( w j,i This can not only be used for tracing the source, but also provides a weighted basis for the subsequent propagation of perplexity from the unit layer to the fragment layer.

[0078] S403: Collection and standardization of learning behavior events.

[0079] During the learner's learning process, data is collected and standardized into an event set E. Each event includes: learner_id, unit_id, a list of optional fragment_ids; event_type (view_unit / answer_question / submit_assignment / ai_chat / video_play, etc.); timestamp, event_params (duration, score, question content, etc.).

[0080] For example: view_unit: Records entry / exit and calculates dwell time; answer_question: Records the question_id, whether it is correct, and the time taken; ai_chat: Records the question text, response text, and a list of unit_id and fragment_id used by the AI ​​during retrieval.

[0081] Specifically, after the course is released, learners participate in the learning process through the front-end interface. The system records learning behaviors through an embedded SDK or server-side mechanism and converts them into standardized events. Typical event types include, but are not limited to: 1. Course access event view_unit.

[0082] Examples of fields: learner_id, unit_id, enter_timestamp, leave_timestamp, duration, scroll_depth, etc.

[0083] 2. The quiz response event `answer_question`.

[0084] Examples of fields: learner_id, unit_id, question_id, is_correct, score, response_time, etc.

[0085] 3. Assignment submission / grading events: submit_assignment / grade_assignment.

[0086] It includes assignment label, score, and summary comments.

[0087] 4. AI tutoring dialogue event ai_chat.

[0088] Examples of fields: learner_id, unit_id, question_text, answer_text, retrieved_units (a list of course units retrieved by the AI), retrieved_fragments (a list of retrieved knowledge fragments), timestamp, etc.

[0089] 5. Media playback events: video_play / video_seek / video_replay.

[0090] Record the playback, pause, drag, and replay intervals of video resources, which can be mapped back to a specific fragment_id.

[0091] All events are written to the event store in a structured format. Each event must include at least: event_id, learner_id, unit_id, event_type, timestamp, and event_params, and optionally record a list of associated fragment_ids.

[0092] like Figure 5 As shown, the traceable data model can: parse / segment source learning resources (documents / web pages / videos) to obtain knowledge fragments; and calculate weights using softmax. w j,i The soft mapping matrix is ​​obtained. W A traceable link is formed between course units and knowledge fragments through a soft mapping matrix; learner interactions for course units are converted into learning behavior events, which can be used to query related knowledge fragments and course units in reverse.

[0093] S404: Learning Behavior Analysis and Two-Level Distress Estimation.

[0094] This step performs aggregate analysis on learning behaviors, calculates the confusion level at the course unit level and the knowledge fragment level, and performs two-layer fusion.

[0095] 1) Course unit level indicators and confusion level.

[0096] For each course unit u j Collect the following data from the event collection: M j :and u j The total number of related response events; M j err :and u j The number of related incorrect answers; V j : u j Number of visits; U j Access Unit u j The number of unique learners; Q j Quote u j The number of AI tutoring questions asked; T j Learners in this unit u j The total duration of stay; B j In this unit u j The number of times you immediately quit the course.

[0097] Calculate basic indicators: Error rate: ; Repeat access intensity: ; Request density: ; Average length of stay: ; Bounce rate: ; in, This is a preset smoothing term used to ensure that the calculation program can run normally under any extreme conditions (such as no access or data being 0).

[0098] These metrics are then normalized (e.g., min-max or Z-score) to obtain: , , , , , all ∈ [0,1].

[0099] Define the course unit perplexity (i.e., the first perplexity): ; Among them, weight , k =1, 2, 3, 4, 5. These weights can be set empirically or obtained by fitting historical data.

[0100] 2) Knowledge fragment layer perplexity (i.e., second perplexity).

[0101] Using soft mapping weights w j,i This propagates the level of confusion from course units to the knowledge fragment layer: .

[0102] In this way, the confusion level of each knowledge segment is accumulated not only by how many course units it is cited, but also by the confusion level of the course unit itself, in the form of weights.

[0103] 3) Dual-layer fusion confusion level.

[0104] Course Units u j Let the set of related knowledge fragments be First, calculate the maximum perplexity of the related knowledge fragments: ; Then calculate the overall perplexity: Among them, the weighting coefficient β ∈[0,1], used to control the relative weights of the "direct action signal" and the "source segment signal".

[0105] According to C j The set of highly confusing course units is selected by sorting by size or comparing with a preset threshold; simultaneously, it can be based on... C i frag Identify highly confusing fragments to provide a more granular reference for subsequent optimizations.

[0106] like Figure 6As shown, the process of learning behavior analysis and two-layer confusion estimation mainly includes: aggregating and statistically analyzing the Event data in the learning behavior event database (aggregating by course unit identifier); then calculating indicators, including error rate, repeated access intensity, help density, average dwell time, and bounce rate; then normalizing the indicators; then calculating the confusion of course units, propagating confusion to knowledge fragments, and merging confusion in two layers; finally identifying high-confusion nodes, selecting high-confusion course units and high-confusion knowledge fragments, and obtaining a set of high-confusion course units and a set of high-confusion knowledge fragments.

[0107] This embodiment achieves accurate problem localization from behavioral data to source fragments through a soft mapping matrix and two-layer perplexity propagation.

[0108] S405: Generation of dynamic optimization schemes for courses based on multi-objective decision-making.

[0109] Once the set of highly confused units and the set of highly confused segments are determined, an optimization scheme is generated based on a predefined set of actions.

[0110] 1. Motion modeling: For each course unit u j Define the set of candidate optimization actions ,For example: a (1): Supplementary explanations (generating additional explanations, examples, etc. based on related fragments); a (2): Intensive practice (increase the number of questions or targeted practice); a (3): Break down units (break down units with large content spans into multiple smaller units); a (4): Adjust the unit order or insert a "Preliminary Knowledge" unit; Other actions include adding illustrations, case studies, and project assignments.

[0111] For action (j,k), estimate its expected perplexity reduction gain. Implementation costs Cost j ( k The benefits can be estimated based on historical optimization results or heuristic rules.

[0112] 2. Global optimization goal: Given the constraints of limited overall development resources B and a maximum of M actions per unit, the objective is: ; The constraints are: ; ;in, Indicates for the unit j Adopt an action k , λ is the cost weight coefficient.

[0113] 3. Greedy approximation implementation: Considering that the solution complexity of 0-1 integer programming is relatively high, this embodiment adopts a greedy approximation algorithm: Step 1: Calculate the unit cost-benefit ratio for each (j,k): ; where is a preset smoothing term.

[0114] Step 2: Sort all candidate actions in descending order according to R j ( k ); Step 3: Traverse the sorted list in sequence. For each (j,k), if the number of actions already selected for the current unit < M and the total cost does not exceed the budget B, then set , and update the total cost; Step 4: Until the budget is exhausted or all candidate actions have been traversed.

[0115] Finally obtained x j (k) Determine the specific course optimization plan. Based on this, the system: invokes the content generation model to automatically generate supplementary explanations and examples with the corresponding knowledge fragments as the context; updates the course structure configuration, that is, re-divides, splits or adjusts course units (such as splitting, re-ordering, inserting prerequisite knowledge units, etc.); updates the exercise question bank and the organizational structure of questions (such as updating the question set, the number of exercises and the difficulty structure). The generated optimization plan is written into the course configuration and a new version identifier version_id is generated.

[0116] S406: Course release, personalized path planning and closed-loop iteration.

[0117] 1. Course version release: Release the course content and structure after applying the optimization plan as a new version. New learners entering the course default to using the latest version; for learners who are already studying, they can choose whether to migrate to the new version or only apply partial optimizations (such as adding supplementary content without changing the structure of the completed units) according to the platform policy.

[0118] 2. Personalized learning path planning: The system abstracts the course structure into a directed acyclic graph , where represents the prerequisite relationship between course units.

[0119] For a learner L, at any learning step, the set of candidate course units is the set of all units for which prerequisites have been met but not yet completed. .

[0120] For each candidate course unit u j Define the scoring function: ; Among them, knowledge gain estimation: . For learners L Course Units j The level of mastery can be estimated (through answer results and historical performance). C j The overall level of confusion for this course unit; Cognitive load estimation Calculated using empirical or simple regression models based on factors such as the complexity and length of the unit content and learner profiles; Exploration Item (Multi-armed Gambling Machine Idea): Let the learner... L Past choices u j The number of times The total number of selections is N L ,but .

[0121] At each path decision, the system selects... As the learner's next learning unit, and updated upon completion based on their performance (such as answer results, duration of stay, etc.). and This enables online adaptive path planning; among which, For the above .

[0122] 3. Closed-loop iteration: With the optimization of the course version, graph structures G and C j The course will be updated, and the path planning will also evolve accordingly, thus achieving a two-layer closed loop of "course version iteration + online adaptive path".

[0123] Under the new version of the course and personalized paths, the system continues to execute steps S403-S405, continuously collecting learning behaviors, updating confusion levels, and generating further optimization plans. This achieves a complete closed-loop iteration: source resources → course construction → learning behaviors → confusion level analysis → course and path optimization → new version of the course → a new round of learning behaviors. After each round of optimization, the system can compare the overall changes in confusion levels between different versions, as well as changes in indicators such as learner completion rate and average score, to further verify the effectiveness of the optimization strategy and adjust the algorithm parameters.

[0124] like Figure 7 As shown, the individual learner path loop within a single version includes: obtaining the set of completed units in the learner's current state, and filtering candidate course units based on prerequisite relationships; calculating a score for each candidate course unit, selecting the candidate course unit with the highest score, and pushing the content of that unit as supplementary resources; updating the relevant parameters in the scoring algorithm after completing the learning, and recording the learning behavior event; and then, the step of filtering candidate course units based on prerequisite relationships can be re-executed. The individual learner path loop can flow back the learning behavior event to the course version iteration closed loop, which includes: collecting learning behavior under the new version, recalculating the perplexity, generating a new round of optimization schemes, generating a new version course graph G(V) based on the optimization schemes, and releasing the new version course; and then, the step of collecting learning behavior under the new version can be re-executed, and the new version course can provide the latest G(V) and course units for filtering candidate course units.

[0125] In other embodiments, the present invention may further: adjust the weights of each indicator in the perplexity calculation; use more complex optimization algorithms to replace greedy approximations (such as heuristic search or reinforcement learning strategies); replace the exploration items in personalized path planning with other bandit strategies such as Thompson Sampling; and encode learning behavior events as event data conforming to the xAPI or Caliper standards for integration with third-party Learning Record Store (LRS) or Learning Management System (LMS) systems. All the above variations are within the framework of "traceable knowledge link + two-layer perplexity + optimization decision + path bandit" constructed by the present invention, and are considered conventional modifications by those skilled in the art, and should not depart from the scope of protection of the present invention.

[0126] In summary, the embodiments of the present invention have the following beneficial effects: Computable and propagable traceability links: Through soft mapping matrices, quantitative associations of "course units – knowledge fragments – source resources" are achieved, enabling confusion to propagate bidirectionally between units and fragments.

[0127] Two-layer perplexity model: Compared with statistics only at the unit level, this embodiment of the invention achieves problem localization at the source fragment level by defining and propagating three types of perplexity, providing a basis for fine optimization.

[0128] Course optimization decision with budget constraints: The optimization action is explicitly modeled as a combination optimization problem with benefits and costs. With the help of a greedy approximation algorithm, the "most cost-effective" optimization combination can be automatically selected in scenarios with limited development resources. This is far superior to simple threshold control and if-else rules.

[0129] Bandit-based path planning: UCB-like strategies are used at the course structure layer to integrate knowledge gain, cognitive load, and exploration items to provide a dynamically evolving learning path for individual learners; compared with traditional adaptive systems that only adjust the difficulty of questions, this embodiment of the invention achieves higher-dimensional adaptation at the course structure layer.

[0130] End-to-end closed loop: From source resources → knowledge fragments → course units → learning behavior → confusion level → optimization decision → new version course → personalized path, a complete closed loop is formed, and each link has clear algorithm support, which is feasible and scalable.

[0131] like Figure 8 As shown, an electronic device 800 provided in this embodiment of the invention includes: a processor 801, a memory 802 and a bus. The memory 802 stores a computer program that can run on the processor 801. When the electronic device 800 is running, the processor 801 and the memory 802 communicate through the bus, and the processor 801 executes the computer program to implement the above-mentioned course dynamic optimization method.

[0132] Specifically, the memory 802 and processor 801 mentioned above can be general-purpose memory and processor, without any specific limitations.

[0133] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the course dynamic optimization method described in the preceding method embodiments. The computer-readable storage medium includes various media capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), RAM, magnetic disk, or optical disk.

[0134] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0135] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic optimization of a course, characterized in that, include: At least one source learning resource is parsed and segmented to obtain multiple knowledge fragments and their knowledge fragment identification data. The knowledge fragment identification data includes fragment identifier, source type, source location, text content, and vector representation. An initial course consisting of multiple course units is constructed based on a pre-defined course outline, and a soft mapping matrix is ​​generated, which consists of soft mapping weights between each course unit and each knowledge segment; wherein, the soft mapping weights are obtained based on the semantic similarity between the corresponding course unit and the corresponding knowledge segment. During the process of learners learning the current course, the learners' learning behavior data is acquired and standardized to obtain an event set consisting of learning behavior events; wherein, the current course includes the initial course or the optimized new version course, and the learning behavior event includes learner identifier, course unit identifier, event type, timestamp and event parameters; Based on the event set and the soft mapping matrix, the overall confusion level of each relevant course unit is determined, wherein the overall confusion level integrates the first confusion level of the course unit layer and the second confusion level of the knowledge fragment layer; For high-confusion course units determined based on the comprehensive confusion level, under a preset cost constraint, a combination of target optimization actions is selected from multiple candidate optimization actions to generate a course optimization scheme; the current course is optimized based on the course optimization scheme, and a new version of the optimized course is released. Based on the comprehensive confusion level and cognitive load estimate of each candidate course unit, as well as the learner's current mastery level and selection frequency of each candidate course unit, the next course unit to be learned is planned for the learner; wherein, the candidate course unit is a course unit for which all prerequisites have been met but not completed.

2. The course dynamic optimization method according to claim 1, characterized in that, The soft mapping matrix is ​​determined through the following process: The title, learning objectives, and keywords of each course unit are encoded to obtain a course unit vector, and the text content of each knowledge segment is encoded to obtain a knowledge segment vector. Calculate the cosine similarity between each pair of the course unit vector and the knowledge segment vector; Based on the cosine similarity of all knowledge segments corresponding to the same course unit, the cosine similarity of each knowledge segment corresponding to the course unit is normalized to obtain the soft mapping weight between the course unit and the corresponding knowledge segment.

3. The course dynamic optimization method according to claim 1, characterized in that, The step of determining the overall confusion level of each relevant course unit based on the event set and the soft mapping matrix includes: Based on the set of events, determine the first perplexity of each relevant course unit at the course unit level; Using the soft mapping matrix, the first perplexity is propagated to the knowledge fragment layer to obtain the second perplexity of the corresponding course unit at the knowledge fragment layer; The first and second confusion levels of each course unit are combined to obtain the overall confusion level of the course unit.

4. The course dynamic optimization method according to claim 3, characterized in that, The step of determining the first confusion level of each relevant course unit at the course unit level based on the event set includes: For each course unit, at least two of the following metrics are statistically analyzed from its related learning behavior events: error rate, repeat visit intensity, help-seeking density, average dwell time, and bounce rate. The statistically obtained indicator values ​​are normalized. The first perplexity of the course unit level is obtained by weighted summation of the normalized index values.

5. The course dynamic optimization method according to claim 3, characterized in that, The process of integrating the first and second confusion levels of each course unit to obtain the overall confusion level of the course unit includes: For each course unit, obtain all relevant knowledge fragments associated with it through the soft mapping matrix; Determine the maximum value of the second perplexity of all the relevant knowledge fragments in the knowledge fragment layer; The comprehensive confusion level of the course unit is obtained by weighted summing the first confusion level of the course unit with the maximum value.

6. The course dynamic optimization method according to claim 1, characterized in that, The step of selecting a combination of target optimization actions from multiple candidate optimization actions to generate a course optimization scheme under a preset cost constraint includes: For each of the highly confusing course units, a set of candidate optimization actions is obtained, which includes at least one of the following: supplementing explanation, adding exercises, splitting the unit, and adjusting the order. Estimate the expected confusion reduction benefit and implementation cost for each candidate optimization action in the set of candidate optimization actions; With the goal of maximizing total revenue under the constraints of total cost budget and the number of optimized actions applied in a single course unit, a greedy approximation algorithm is used to select the target optimized action combination from the candidate optimized action set.

7. The course dynamic optimization method according to claim 1, characterized in that, The method of planning the next course unit for the learner based on the comprehensive confusion and cognitive load estimation of each candidate course unit, as well as the learner's current mastery level and selection frequency for each candidate course unit, includes: The course structure of the current course is abstracted into a directed acyclic graph, and all candidate course units of the learner are determined. For each candidate course unit, the score is calculated using the following formula: ； in, S L ( u j ) as candidate course units u j The rating, G L ( u j (Based on candidate course units) u j The expected knowledge gain is obtained by calculating the corresponding current mastery level and overall confusion level. CL L ( u j ) as candidate course units u j Cognitive load estimation, Explore L ( u j (Based on candidate course units) u j The corresponding number of selections and the exploration items determined using the multi-armed gambling machine concept. η and ξ All are preset weighting coefficients; The candidate course unit with the highest score is selected as the next course unit.

8. A dynamic course optimization system, characterized in that, include: The resource parsing and segmentation module is used to parse and segment at least one source learning resource to obtain multiple knowledge fragments and their knowledge fragment identification data. The knowledge fragment identification data includes fragment identifier, source type, source location, text content, and vector representation. The traceable course construction module is used to construct an initial course consisting of multiple course units based on a preset course outline, and generate a soft mapping matrix composed of soft mapping weights between each course unit and each knowledge segment; wherein, the soft mapping weights are obtained based on the semantic similarity between the corresponding course unit and the corresponding knowledge segment. The learning behavior acquisition module is used to acquire and standardize the learner's learning behavior data during the process of the learner learning the current course, and obtain an event set consisting of learning behavior events; wherein, the current course includes the initial course or the optimized new version course, and the learning behavior event includes learner identifier, course unit identifier, event type, timestamp and event parameters; The course unit confusion determination module is used to determine the comprehensive confusion of each relevant course unit based on the event set and the soft mapping matrix. The comprehensive confusion integrates the first confusion of the course unit layer and the second confusion of the knowledge fragment layer. The course optimization and release module is used to select a combination of target optimization actions from multiple candidate optimization actions to generate a course optimization scheme for high-confusion course units determined based on the comprehensive confusion level, under a preset cost constraint; optimize the current course based on the course optimization scheme, and release the optimized new version of the course. The learning path planning module is used to plan the next course unit to be learned for the learner based on the comprehensive confusion and cognitive load estimation of each candidate course unit, as well as the learner's current mastery level and selection number of each candidate course unit; wherein, the candidate course unit is a course unit for which all prerequisites have been met but not completed.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the course dynamic optimization method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the course dynamic optimization method according to any one of claims 1-7.

Images