Intelligent learning plan generation method and device based on large model

Abstract

The invention provides an intelligent learning plan generation method and device based on a large model, and belongs to the technical field of educational informatization. Identifying entities and key terms; constructing a course knowledge graph; carrying out matching calculation with courses through mixed retrieval on the basis of a course knowledge graph and key information features, and carrying out dual-channel collaborative multi-channel retrieval in combination with vector similarity and keyword matching; the multi-target course arrangement information is optimized based on a genetic algorithm, course arrangement is carried out in combination with a time planner of a constraint satisfaction problem model, and a course time arrangement plan is generated; and related lecturers and students are informed of confirmed results. According to the method, the curriculums and student requirements are matched, the comprehensiveness and accuracy of curriculum retrieval are improved through weighted sorting, the curriculums meeting the learning plan requirements can be found more accurately, and the curriculum matching efficiency is improved. A better course arrangement scheme is found through global search, the resource utilization rate is improved, and a scientific and reasonable course time arrangement plan is generated.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of educational informatization, and specifically relates to a large model-based intelligent learning plan generation method and device based on generative artificial intelligence and natural language processing technology. BACKGROUND

[0002] With the rapid development of information technology and the continuous acceleration of knowledge update speed, the formulation of personalized learning plans has become increasingly important for improving learning efficiency and meeting the needs of different learners.

[0003] In terms of learning data processing, simple word segmentation, word frequency statistics, and bag-of-words models are mainly relied on, making it difficult to deeply understand the semantic information and contextual associations of text. When dealing with complex texts such as policy documents and course descriptions, it is difficult to accurately identify entities, key provisions, and core content, and to generate high-quality semantic vector representations. This makes subsequent course matching and learning plan generation lack precise data support, and there is a large deviation between recommended courses and actual needs of learners.

[0004] In terms of course resource organization and management, existing course recommendation systems are mostly based on label matching or collaborative filtering algorithms, lacking deep mining of knowledge structures and association relationships of course content. The prerequisite and subsequent relationship between courses, domain dependency relationship, and the association between courses and related policies and regulations are often ignored, resulting in a lack of systematicness and logical coherence in recommended courses.

[0005] Existing course scheduling needs to consider multiple factors such as course duration, classroom resources, instructor time, and student time. Traditional course scheduling methods usually use single objective optimization, such as only considering classroom utilization or course compactness, which can easily lead to imbalance in other objectives, such as unreasonable course time arrangement, causing student learning fatigue, or instructor teaching time conflicts. SUMMARY

[0006] The present application provides a large model-based intelligent learning plan generation method, which collects learning-related information. The BERT+BiLSTM model is used for processing to identify entities and key provisions, complete abstract generation and topic modeling, and solve the problems of one-sided data collection and low efficiency.

[0007] The method comprises: Step S101: Extracting learning data in the learning plan; Step S102: Processing the learning data using the BERT+BiLSTM model to identify entities and key provisions, and simultaneously performing abstract generation and topic modeling to obtain key information features of the learning data; Step S103: constructing a course knowledge graph, and labeling the course knowledge graph, the labeling content being a course field, an applicable object, and a policy association; Step S104: based on the course knowledge graph and key information features, performing hybrid retrieval and matching calculation with courses, combining vector similarity and keyword matching double-channel collaborative multi-path retrieval, and performing weighted sorting through BM25 and FAISS similarity scores; Step S105: based on a genetic algorithm, optimizing multi-objective course scheduling information, and combining a time planner of a constraint satisfaction problem model to schedule courses, and generating a course time arrangement plan; Step S106: using a visual course schedule editing interface supporting drag-and-drop adjustment to provide a user with course plan auditing and adjustment, and notifying relevant teachers and students of the confirmed result.

[0008] It should be further explained that step S101 specifically includes: The learning data in the learning plan includes course completion, learning duration, and examination score information; The learning data is monitored in real time, and a data update threshold and a time interval are set. When the learning data changes or reaches the update condition, the data crawling task is triggered to make the learning plan optimized based on the latest data; Sensitive information is encrypted when learning data is crawled, and user privacy data is encrypted.

[0009] It should be further explained that step S102 specifically includes: Step S1021: performing multi-modal alignment and noise filtering on the crawled learning data, and through timestamp association and content relevance matching, unifying a structured course schedule, an unstructured policy text, and a semi-structured student interaction record to the same space-time context, and filtering historical data irrelevant to the current learning plan goal, repeated records, and format error content; Step S1022: based on the input characteristics of BERT, adding a position identification vector to the learning data, which is used to label the association order of policy provisions and course goals, and a domain marking vector; Step S1023: performing four sub-tasks of entity recognition, key clause extraction, abstract generation, and theme modeling on the learning data in parallel, and sharing the underlying feature representation of BERT+BiLSTM; wherein the entity recognition and key clause extraction take the clause content of the learning plan as the context constraint of the entity, and the abstract generation and theme modeling realize content focusing through an attention mechanism; Step S1024: Confidence calibration and result screening are performed on the multi-task output results. Thresholds are set based on the category confidence of entity recognition, the semantic similarity between the summary and the original text, low-confidence results are filtered, and secondary verification is performed to retain valid information features that meet the requirements.

[0010] It should be further explained that step S1023 specifically includes: Step S10231: Semantic hierarchical coding is performed on the learning data. Based on the deep semantic representation capability of the BERT pre-training model, the original text is input into BERT for context coding to obtain a word vector sequence containing global semantic information. At the same time, the sequence is modeled by a BiLSTM layer to capture the cross-sentence association relationship between policy provisions and course objectives, and a hidden state sequence that fuses context and sequential information is output. Step S10232: Context constraints are applied to the entity recognition and key clause extraction tasks. The explicit clause content in the learning plan is input into the gating unit of the BiLSTM layer as a bias term when calculating the hidden state, and the core objectives of the current learning plan are associated when extracting entities and clauses. Step S10233: A multi-scale attention mechanism is designed for the summary generation and topic modeling tasks. The summary generation uses content-based self-attention to calculate the similarity between word vectors and global semantic centers to obtain key information. Step S10234: The domain keywords extracted by the entity recognition task are fed back to the summary generation task in real time through the bottom-layer hidden state sequence of BERT+BiLSTM. The high-frequency topics identified by the topic modeling task are passed to the key clause extraction task as a constraint condition to limit the scope of clause extraction.

[0011] It should be further explained that step S1024 specifically includes: Step S10241: Confidence evaluation is performed on the multi-task output results. Context consistency confidence and semantic coherence confidence are defined to form a comprehensive confidence vector. Step S10242: Based on the learning plan target type and data distribution characteristics, the confidence thresholds of each dimension are adjusted. Step S10243: Secondary verification is performed on the filtered learning plan results. The verification includes: entities appearing in policy provisions and course outlines at the same time, summaries covering at least 3 learning plan topics, and learning topic modeling results intersecting with the policy keyword library. Step S10244: The results that pass the secondary verification are integrated. The entities and corresponding policy provisions are mapped, the summaries and learning topic labels are bound, and an information feature set is formed.

[0012] It needs to be further explained that step S103 specifically includes: Extract the course name, duration, lecturer from the course management system, and obtain the courseware, video, and exercise data from the teaching resource library. Also, collect the completion rate and evaluation score data from the student behavior database, and the policy provisions and compliance requirements data associated from the policy file library to form the course knowledge base. Based on course field annotation, use professional skills, Python development, and data analysis learning plan structure, and define student tags combined with student portraits. Policy association annotation is achieved through keyword matching and policy provision mapping, and the association degree is annotated. Build a course knowledge graph, define matching relationships including course-policy coverage, course-teaching resources, and course-student tags, and enhance the semantic expression dimension of the graph. Check the logical consistency of the course field and the applicable students, the matching of the policy association degree and the course content, use graph neural networks for link prediction of the knowledge graph, fill in the relationship between unannotated courses and students, and finally output the course knowledge graph.

[0013] It needs to be further explained that step S104 specifically includes: Step S1041: Map the text nodes and multimedia nodes in the course knowledge graph and the student completion rate and interaction comments in the learning data to a unified semantic space through a cross-modal embedding model to generate feature vectors containing visual, textual, and numerical features. Step S1042: Define the priority strategy of dual-channel retrieval based on the target type and context information of the learning plan. Step S1043: When performing dual-channel collaborative retrieval, use a gated attention mechanism to fuse vector similarity and keyword matching scores. By defining a gating value, adjust the weighting coefficients of BM25 and FAISS based on the domain relevance and policy association strength of the current retrieval results. Step S1044: Perform feasibility verification on the weighted sorted candidate courses, combine course duration, lecturer available time, and student regional restriction conditions, filter courses that do not meet actual conditions, and sort the remaining results to output the final matching course list.

[0014] It needs to be further explained that step S1043 specifically includes: Fuse the feature vectors into continuous gating values between 0 and 1. According to the learning plan phase, course type, and student group characteristics, set weight adjustment rules to trigger FAISS weight priority processing and trigger BM25 weight emphasis on skill keywords. The BM25 and the FAISS score are nonlinearly transformed by using a piecewise function, and a weight proportion is allocated according to a gating value; when the gating value is greater than 0.7, the FAISS weight = gating value * 0.8 + 0.2, and the BM25 weight = 1-FAISS weight; The candidate courses ranked by the weight are matched with a historical success case library, if the matching degree is lower than a threshold, manual review is triggered, and the gating value and the weight parameter of each adjustment are recorded.

[0015] It should be further explained that step S105 specifically includes: According to the actual needs of the learning plan, the optimization target is expanded to three dimensions of non-overlapping course time periods, lecturer, classroom and equipment load balancing at the same period, student learning time length on a single day not exceeding a threshold, and course type alternation arrangement; Each target sets a quantifiable mathematical expression: time conflict number = Σ (course i and course j overlapping period), resource utilization rate = Σ (lecturer i daily teaching time length / total available time length); Constraint conditions are extracted from the course knowledge graph and key information features, including: course duration is fixed, lecturer unavailable time period, student regional restriction, and course field and student adaptation degree ≥80%, policy related courses are preferentially arranged, and the constraints are converted into adjustable penalty coefficients; An improved genetic algorithm optimization process is defined, a real number coding method is used to represent the course scheduling, domain knowledge guidance is introduced in the crossover operation, constraint checking is combined with the CSP model in the mutation operation, and the weighted sum of the three optimization targets is calculated at the same time when each generation population is evaluated; A time planner is constructed, the initial course scheduling plan output by the genetic algorithm is detected for time conflicts by using a backtracking algorithm, and alternative resources in the course knowledge graph are called to make local adjustments, the optimization target score is recalculated after the adjustment, and until the final course scheduling plan without conflicts and meeting all constraints is generated.

[0016] According to another embodiment of the present application, an electronic device is provided, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the steps of the large model-based intelligent learning plan generation method when executing the program.

[0017] From the above technical solutions, the present application has the following advantages: The intelligent learning plan generation method based on a large model provided by the application extracts learning data, processes the learning data by using a BERT+BiLSTM model, utilizes the powerful context semantic understanding capability of BERT and the bidirectional feature extraction advantage of BiLSTM, and accurately identifies entities and key clauses. A course knowledge graph is constructed, and the course and its related relationships are presented in a structured form. Based on the mixed retrieval of the course knowledge graph and key information features, combined with the vector similarity and keyword matching dual-channel retrieval, the comprehensiveness and accuracy of course retrieval are improved, and courses that meet the learning plan requirements can be more accurately found, and the course matching efficiency is improved. The genetic algorithm is used for multi-objective course scheduling optimization, and the constraint satisfaction problem model is combined, so that multiple objectives and complex constraint conditions such as course time distribution, resource conflict, and student preference can be considered at the same time. A better course scheduling scheme is found through global search, resource utilization is improved, a scientific and reasonable course time arrangement plan is generated, and teaching management efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the present application, the drawings required to be used in the description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0019] Figure 1 The flowchart of the intelligent learning plan generation method based on a large model is shown. Figure 2 The flowchart of the intelligent learning plan generation method based on a large model is shown. Figure 3 The flowchart of another embodiment of the intelligent learning plan generation method based on a large model is shown. Figure 4 The schematic diagram of an electronic device is shown. DETAILED DESCRIPTION

[0020] The intelligent learning plan generation method based on a large model provided by the application realizes intelligent matching and arrangement of training courses by constructing a structured instructor course knowledge base and using RAG (Retrieval-Augmented Generation) technology to integrate multi-modal document information such as policy files and leadership speeches. The system implementing the method can include a course vectorization module, a document semantic analysis module, a time planning algorithm module, and a notification module, and can complete the whole process from demand analysis to course table generation, improving the efficiency of education and training management.

[0021] The intelligent learning plan generation method based on a large model related to the present application will be described in detail below. For the purpose of illustration but not for the purpose of limitation, specific details such as specific system structures, techniques, and the like are presented in order to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details.

[0022] It should be understood that when used in the specification, the term "comprises" indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. The terms "comprise", "contain", "have" and their variants mean "including but not limited to", unless otherwise specifically emphasized.

[0023] The phrase "one embodiment" or "some embodiments" appearing in the specification means that a specific feature, structure, or characteristic described in the embodiment is included in one or more embodiments of the present application. Thus, the phrases "in one embodiment", "in some embodiments", "in other some embodiments", "in yet some embodiments" appearing in different places in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized.

[0024] In the embodiments of the present application, computer program code for carrying out operations of the present disclosure can be written in one or more programming languages or combinations thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" language or similar programming languages. Program code can be executed entirely on a user computer, partially on a user computer, as a separate software package, partially on a user computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, through the Internet using an Internet service provider).

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts fall within the scope of protection of the present application.

[0026] Referring to Figure 1 FIG. 1 shows a flowchart of a method for generating an intelligent learning plan based on a large model according to an embodiment. The method comprises the following steps: Step S101: Extract learning data in the learning plan.

[0027] In some embodiments, structured data, semi-structured data, and unstructured data are extracted from learning management systems, online course platforms, user terminal devices, and other channels by developing interfaces that adapt to multiple data sources. At the same time, the data is preliminarily formatted and encoded to ensure that it can be recognized and processed by subsequent modules.

[0028] The structured data of this embodiment can be student basic information, course completion progress, and test scores. The semi-structured data can be course syllabus and chapter directory. The unstructured data can be policy documents, student feedback comments, and learning logs. In this way, the learning plan generation process can fully consider various influencing factors to improve the accuracy and practicality of the learning plan.

[0029] In some specific embodiments, multi-source data fusion technology is used to extract learning data in the learning plan from management systems, online learning platforms, and mobile learning applications, including course completion, learning duration, test scores, and other multi-dimensional information, to enrich data sources and improve data comprehensiveness and accuracy.

[0030] The extracted learning data is monitored in real time. When the learning data changes or meets the update conditions, the data extraction task is triggered by setting the data update threshold and time interval, ensuring the timeliness and freshness of the learning data, and enabling the learning plan generation to be optimized based on the latest data.

[0031] Optionally, sensitive information is encrypted when extracting learning data to ensure the security of data transmission and storage, while complying with privacy policies and relevant regulations to strictly protect user privacy data and enhance user trust in data usage.

[0032] Step S102: Process the learning data using the BERT+BiLSTM model to identify entities and key clauses, and perform abstract generation and topic modeling to obtain key information features of the learning data.

[0033] In this embodiment, the preprocessed learning data is input into the BERT+BiLSTM model. The BERT model first understands the context of the data and captures the semantic representation of the words in the text. The BiLSTM model further processes the features output by BERT in both directions to identify entities such as course names, policy and regulation names, student names, and key clauses such as course requirements and policy provisions.

[0034] In the aspect of abstract generation, important sentences and paragraphs in the text are selected based on the attention mechanism to generate the abstract; in the process of topic modeling, the potential topics of the text are mined by using algorithms such as LDA, and finally the key information feature vectors of the learning data are integrated.

[0035] It can be seen that the BERT model is based on the Transformer architecture and learns the context semantics of the text through the self-attention mechanism; the BiLSTM model can capture the semantic dependency of the text from the front and back directions through the bidirectional recurrent structure, thereby improving the accuracy of entity and key clause recognition. The abstract generation and topic modeling respectively use the attention mechanism to focus on important content and the probability model to mine the potential semantic structure.

[0036] Step S103: Constructing a course knowledge graph, and labeling the course knowledge graph, the labeling content being a course field, an applicable object, and a policy association.

[0037] The present embodiment takes a course as a core node, and constructs a course knowledge graph by taking the relationships between courses, such as prerequisite relationships, parallel relationships, and associations of courses with lecturers, students, and policies, as edges.

[0038] Each course node is labeled in detail to clearly indicate its belonging course field, such as a programming language or financial accounting. The applicable object is, for example, a beginner, an advanced student, or a specific professional group, and the association with a related policy and regulation is, for example, which policies require the course content. At the same time, the knowledge graph completion algorithm is used to mine and fill in potential node relationships, and to perfect the knowledge graph structure.

[0039] The present embodiment uses a graph database technology to store course-related entities and relationships, and converts course information into a structured knowledge graph through a knowledge representation learning method. The labeling process is based on a combination of manual labeling and automatic labeling. The automatic labeling extracts key information in the text to perform preliminary labeling, and the manual labeling is used for auditing and correction to ensure the accuracy of the labeling.

[0040] As an embodiment of step S103, step S103 specifically includes: The course name, duration, and lecturer are extracted from the course management system, the courseware, video, and exercise data are obtained from the teaching resource library, the completion rate and test score data are collected from the student behavior database, and the policy provisions and compliance requirements data are obtained from the policy file library, thereby forming a course knowledge base.

[0041] The course field annotation adopts a hierarchical structure of professional skills, Python development, and data analysis. The applicable object annotation is defined in combination with the student profile, such as age, position, and existing skills, and further defines tags for in-school students, in-service personnel, and personnel transferring positions. The policy correlation annotation is matched through keywords and policy provisions, and the annotation is directly related, such as the course content covering the original text of the provisions, and indirectly related to the course objectives meeting the provisions.

[0042] The course knowledge graph is constructed, and in addition to traditional course nodes, policy text segment nodes, teaching resource nodes, and student feedback nodes are added. The relationship types are defined, including course-policy coverage, which is the coverage relationship between courses and policy provisions; course-resource association, which is the support relationship between courses and teaching resources; and course-student adaptation, which is the matching relationship between courses and student tags, so as to enhance the semantic expression dimension of the course knowledge graph.

[0043] The execution rule verification checks the logical consistency of the course field and the applicable object, such as the Python development field not annotating the applicable object of early childhood education. The policy correlation degree and the matching of the course content, such as the directly related course containing at least 3 policy provision original texts. The graph neural network (GNN) is used to link the knowledge graph for prediction, to evaluate the potential rationality of the unannotated relationship, and to trigger manual review for low confidence relationships. Finally, the verified high-quality course knowledge graph is output.

[0044] The present embodiment improves the timeliness and professionalism of annotation through rules. It is considered that policy documents will be updated over time, such as the release of the new version of the vocational skills training subsidy policy. Therefore, based on the policy timeliness, such as the effective date of the provision and the adjustment of the annotation standard based on the field professionalism; its implementation is based on the Drools rule engine, which regularly synchronizes the latest policy documents and updates the annotation template to ensure that the annotation results are consistent with the current policy environment.

[0045] The present embodiment guarantees the quality of the graph through verification. Rule verification solves the mismatch between the field and the object, but may miss complex and unreasonable relationships. Verification through graph neural network mining potential association makes up for the deficiency of manual annotation; in the implementation, rule verification excludes early childhood education by writing logical constraints such as field = professional skill applicable object, and model verification trains GNN model to predict unannotated relationships, such as the correlation probability of course A-policy provision B. The edges with a probability lower than the threshold trigger manual review, and finally a high-confidence knowledge graph is output.

[0046] Step S104: Based on the course knowledge graph and key information features, mixed retrieval and course matching calculation are performed through a double-channel collaborative multi-path retrieval combining vector similarity and keyword matching, and weighted sorting is performed through BM25 and FAISS similarity scores.

[0047] The learning data key information features obtained in step S102 and the course knowledge graph constructed in step S103 are taken as inputs, and vector similarity retrieval is performed. Optionally, the FAISS index can be used, the IVF PQ algorithm is adopted, keyword matching retrieval can use the ElasticSearch, and the IK tokenizer is used for retrieval.

[0048] For vector similarity retrieval, the cosine similarity of the learning data features and the course vectors is calculated in a high-dimensional space. For keyword matching retrieval, after the text is segmented by the tokenizer, the BM25 algorithm is used to calculate the matching score of the keywords and the course text. Finally, according to the characteristics and requirements of the learning plan, the weighting coefficients are set, the scores of the two retrieval methods are weighted and fused, the retrieved courses are sorted, and the courses with high matching degree are selected. In this way, both semantic fuzzy queries and implicit related courses can be processed, and courses containing specific keywords can be quickly located, improving the comprehensiveness and accuracy of the retrieval.

[0049] Step S105: Based on the genetic algorithm, the multi-objective course scheduling information is optimized, and the time planner of the constraint satisfaction problem model is combined to schedule the courses, and the course time arrangement plan is generated.

[0050] In this embodiment, the course matching results, course duration, classroom resources, lecturer time, student time, and other course scheduling related information are taken as inputs to construct a multi-objective optimization problem, and the objectives include course time distribution balance, avoidance of classroom and lecturer conflicts, and meeting student learning time preferences.

[0051] Optionally, the genetic algorithm is used to iteratively optimize the course scheduling scheme, and a new course scheduling scheme is generated through selection, crossover, mutation, and other operations. Meanwhile, the constraint satisfaction problem model is combined to check the constraint conditions of each scheme, such as classroom capacity limits and lecturer non-conflict periods. In this embodiment, the course scheduling scheme that meets the constraint conditions and has the optimal objective function can be selected, and finally a detailed course time arrangement plan is generated.

[0052] The constraint satisfaction problem model verifies the feasibility of the generated scheme to ensure that the scheme meets the actual resource and time limit conditions. This improves the rationality of the course scheduling and the resource utilization rate, generates a scientific and efficient course time arrangement plan, and meets the multi-faceted needs of students and teaching resources.

[0053] As step S105 of this embodiment, the following specific steps are also involved.

[0054] In combination with the actual needs of the learning plan, the optimization goals are expanded to three dimensions: no overlap of course time periods, load balancing of lecturers, classrooms, and equipment at the same time period, and single-day learning time of students not exceeding the threshold and alternating arrangement of course types. A quantifiable mathematical expression is set for each goal: Time conflict =∑(overlapping time periods of course i and course j), Resource utilization =∑(lecturer i's teaching time length on the same day / total available time length).

[0055] The constraints are extracted from the course knowledge graph and key information features, including: fixed course duration, lecturer unavailable time period, student regional restriction, and course field and student adaptation degree ≥80%, policy-related course priority scheduling. The constraints are converted into adjustable penalty coefficients.

[0056] An improved genetic algorithm optimization process is defined, using real number coding to represent course scheduling, introducing domain knowledge guidance in the crossover operation, combining CSP model constraint checking in the mutation operation, and calculating the weighted sum of the three optimization goals at each generation population evaluation. A time planner is constructed to detect time conflicts through backtracking algorithm and call alternative resources in the course knowledge graph for local adjustment on the initial scheduling plan output by the genetic algorithm. After adjustment, the optimization goal score is recalculated until the final scheduling plan without conflicts and meeting all constraints is generated.

[0057] In this embodiment, the single-day learning time standard deviation of each student experience is measured, and the initial weights of each goal are set by expert experience, such as time conflict weight 0.4, resource utilization 0.3, and student experience 0.3. The constraints are extracted from the course knowledge graph and key information features, including: fixed course duration, lecturer unavailable time period, student regional restriction, and course field and student adaptation degree ≥80%, policy-related course priority scheduling. The constraints are converted into adjustable penalty coefficients. For example, the applicable object of course A is in-service personnel, excluding school students during the period, and the penalty coefficient of soft constraint is stored in the configuration table when the adaptation degree is less than 70%, and the penalty coefficient is +0.2. CSP constraint checking is combined to ensure that the population individuals are always within the feasible domain; the implementation method needs to design an adaptability function, use simulated binary crossover (SBX) operator in crossover, and use polynomial mutation operator in mutation, and reserve high adaptability individuals to enter the next generation.

[0058] The resource pool in the course knowledge graph is called, such as the list of backup lecturers, the list of available classrooms, and the target score is recalculated for each adjusted plan to ensure that the adjusted plan still meets the multi-objective optimization requirements.

[0059] In this way, the core direction of multi-objective optimization is clearly defined, and quantifiable targets are defined from three key dimensions of time, resources, and student experience, providing clear measurement standards for subsequent course scheduling optimization. These mathematical expressions convert abstract optimization requirements into specific and calculable indicators, facilitating evaluation and comparison in algorithms. From the course knowledge graph and key information features, various constraint conditions are extracted, distinguishing between hard and soft constraints, and converting soft constraints into penalty coefficients through a rule engine. This allows not only to meet hard, non-negotiable conditions during course scheduling, but also to consider soft, adjustable conditions through a penalty mechanism, enhancing the adaptability of the course scheduling plan to complex real-world situations.

[0060] Based on the above objectives and constraints, the genetic algorithm is improved specifically. Real number coding is adopted to fit the actual elements of course scheduling, and domain knowledge guidance and constraint checking mechanisms are introduced in the crossover and mutation operations to make the algorithm more directional and reasonable when searching for course scheduling solutions, avoiding the generation of invalid or unreasonable solutions. The multi-objective weighted sum is calculated during the evaluation of each generation population to ensure the algorithm evolves towards the overall optimal direction. For the initial course scheduling plan output by the genetic algorithm, the conflict resolution mechanism of the time planner is used to detect time conflicts through a backtracking algorithm, and the alternative resources in the course knowledge graph are used for local adjustment. After adjustment, the optimization target score is re-evaluated, and the process is iterated until a final course scheduling plan without conflicts and meeting all constraints is obtained. The entire process forms a closed loop, realizing a complete course scheduling optimization process from target definition, constraint handling, algorithm optimization to conflict resolution. The initial course scheduling plan is checked and adjusted to ensure that the final generated course scheduling plan has no time conflicts and meets all the set constraints. By continuously iterating the optimization target score, the quality of the course scheduling plan is further improved, providing reliable protection for the smooth conduct of teaching activities.

[0061] Step S106: Use a visual course schedule editing interface that supports drag-and-drop adjustments to allow users to review and adjust the course plan. After confirmation, the result is notified to the relevant instructors and students.

[0062] This embodiment can display the course time scheduling plan in an intuitive calendar view, and users can adjust the course time, classroom, and other information through mouse drag-and-drop operations. The interface provides a review function, allowing users to view course details, policy requirements associated with the course, and student adaptation, etc. When the user confirms that the course plan is correct, the system notifies the relevant instructors and students of the course plan information through various channels such as internal messages, emails, and SMS, and updates the course scheduling status in the management system. This improves user engagement and satisfaction. The notification function ensures that course plan information is accurately conveyed to instructors and students in a timely manner, avoiding information transmission omissions or delays, and ensuring the smooth conduct of teaching activities.

[0063] In an embodiment of the present application, based on step S102, a possible embodiment will be given below to specifically illustrate the non-limiting embodiment. As shown in Figure 2 Step S102 specifically includes: Step S1021: Multi-modal alignment and noise filtering of the captured learning data, unified to the same space-time context through timestamp association, content correlation matching of structured curriculum, unstructured policy text, and semi-structured staff interaction records, filtering historical data irrelevant to the current learning plan target, repeated records, and format error content; Step S1022: Based on the input characteristics of BERT, adding a position identification vector to the learning data, used to mark the association order of policy provisions and curriculum objectives, and a domain marking vector.

[0064] Step S1023: Parallel execution of entity recognition, key clause extraction, abstract generation, and topic modeling four sub-tasks of learning data, sharing the underlying feature representation of BERT+BiLSTM; among them, entity recognition and key clause extraction take the clause content of the learning plan as the context constraint, and abstract generation and topic modeling realize content focusing through attention mechanism.

[0065] In this embodiment, step S1023 specifically includes: Step S10231: Semantic hierarchical coding of learning data, based on the deep semantic representation ability of the BERT pre-training model, inputting the original text into BERT for context coding to obtain a word vector sequence containing global semantic information; at the same time, through the BiLSTM layer, the sequence is modeled for bidirectional long-range dependence, capturing the cross-sentence association relationship between policy provisions and curriculum objectives, and outputting a hidden state sequence that integrates context and sequence information.

[0066] Step S10232: Apply context constraints to entity recognition and key clause extraction tasks, take the explicit clause content in the learning plan as a conditional input to the gating unit of the BiLSTM layer, embed the vector as a bias term when calculating the hidden state, and associate the core target of the current learning plan when extracting entities and clauses.

[0067] Step S10233: Design a multi-scale attention mechanism for abstract generation and topic modeling tasks, among them, abstract generation adopts content-based self-attention, calculates the similarity between word vectors and global semantic centers to obtain key information.

[0068] Step S10234: Through the underlying hidden state sequence of BERT+BiLSTM, the domain keywords extracted by the entity recognition task are fed back to the abstract generation task in real time; the high-frequency topics identified by the topic modeling task are passed to the key clause extraction task as a constraint condition to limit the scope of clause extraction.

[0069] As an example of the present application, assume that key information features need to be generated for an enterprise's artificial intelligence engineer skill improvement learning plan. The input data includes: enterprise internal policy documents such as the digital talent training subsidy details, course outlines containing chapters such as "Python programming" and "machine learning practice", student historical learning records, 80% of student feedback that deep learning chapters are difficult, and an AI talent skill demand white paper. The following is the specific processing flow of step S1023: All policy provisions, course outlines, student comments, and industry reports are input into the pre-trained Chinese BERT model for context encoding. The Transformer layer of BERT will make the vector based on the word "Python programming" not only contain the basic semantics of "Python", but also integrate the information of related words such as "practical projects" and "data analysis" in the context. Subsequently, these word vectors are input into the BiLSTM layer for bidirectional long-range dependency modeling. The forward unit of BiLSTM captures the sequential echo of "training must include practical operation" in policy provisions and "experimental courses account for 30%" in course outlines, and the backward unit captures the reverse association of "hope to increase practical operation" in student comments and "improve hands-on ability" in course objectives. Finally, BiLSTM outputs a hidden state sequence, and each time step vector integrates global semantics and long-range dependencies.

[0070] In this example, the core goal of the learning plan is to "cultivate AI engineers that meet the needs of the enterprise's digital transformation", so ensure that the entities "Python programming", "machine learning", and "training duration > 100 hours" and "need to pass certification exams" are strongly associated with this goal.

[0071] In specific operations, the goal description of the learning plan is converted into a word embedding vector and embedded as a bias term in the BiLSTM gate unit calculation. For example, when BiLSTM processes the policy provision "training content must cover Python programming and machine learning", the calculation formula of the gate unit changes from the traditional ht=σ(Wihxt+Whhht−1+b) to ht=σ(Wihxt+Whhht−1+b+c), where c is the word embedding vector of the "digital transformation" goal. This operation forces the model to prioritize information related to "digital transformation" and "AI engineers" when extracting entities and provisions.

[0072] For the abstract generation task, a content-based self-attention mechanism is adopted. For example, the course outline text "This course includes Python programming, which will be input into BERT+BiLSTM to obtain a hidden state sequence. The self-attention mechanism calculates the similarity between each word vector and the global semantic center (obtained by averaging all word vectors), for example, the hidden state of "Python programming" has a similarity of 0.85 (high correlation) with the global center, the similarity of "supervised learning" is 0.72 (moderate correlation), and the similarity of "experimental course proportion" is 0.65 (low correlation). Finally, the abstract generation task will preferentially select high-similarity keywords such as "Python programming" and "machine learning" to generate the abstract "This course focuses on Python programming and machine learning, two core directions."

[0073] For the topic modeling task, a cross-attention mechanism based on documents is adopted. For example, the course outline, student comments, and industry reports are considered as different documents, and the "Python programming" paragraph vector in the "technology module" dimension scores 0.9, and the "career development" dimension scores 0.3. The cross-attention mechanism compares the topic distribution of different documents and extracts the cross-document "Python programming" that is frequently mentioned in the course outline, student comments, and industry reports, and finally determines the core topic as "Python programming practical ability training". Through the above process, the key information features extracted from the multi-source text in step S1023 not only cover the global semantics, but also focus on the "digital transformation" and "practical ability" of the learning plan, while avoiding information omission through task coordination.

[0074] Step S1024: Confidence calibration and result screening of multi-task output results, based on the class confidence of entity recognition, the semantic similarity between the abstract and the original text, and the threshold setting, filtering low-confidence results, and performing secondary verification to retain valid information features that meet the requirements.

[0075] It should be noted that step S1024 specifically includes: Step S10241: Confidence evaluation of multi-task output results, defining context consistency confidence and semantic coherence confidence to form a comprehensive confidence vector.

[0076] Step S10242: Adjust the confidence threshold of each dimension based on the learning plan target type and data distribution characteristics.

[0077] Step S10243: Secondary verification of the filtered learning plan results, including: entities appearing in both policy provisions and course outlines, abstracts covering at least three learning plan topics, and learning topic modeling results intersecting with the policy keyword library.

[0078] Step S10244: The results of the secondary verification are integrated to map the entities to the corresponding policy clauses, and the summaries are bound to the learning topic labels to form the information feature set.

[0079] The multi-dimensional confidence evaluation and secondary verification mechanism of the embodiment strictly reviews the multi-task output results from multiple aspects such as context logic and semantic coherence, effectively filters out low-quality information with logical contradictions and ambiguous semantics, and improves the reliability of the final information feature set. Based on the learning plan target type and the data distribution characteristics, the confidence threshold is adjusted to make the screening process flexible to adapt to different application scenarios and data conditions. By establishing the association between entities and policy clauses and between summaries and learning topic labels, the results of the verification are structurally integrated, and scattered information is converted into an information feature set with clear logical relationships. This structured processing facilitates subsequent quick retrieval and utilization, improves the value of the data, and provides strong support for course matching and accurate learning plan generation.

[0080] As can be seen from the above steps, the addition of the location identification vector and the domain label vector, combined with the learning plan clauses as context constraints and attention mechanism focusing on content, enables the model to more accurately understand the semantic and logical relationships of the learning data, thereby improving the accuracy of entity recognition, key clause extraction, summary generation, and topic modeling, and extracting more accurate key information features. The confidence calibration and result screening link strictly controls the quality of the multi-task output results by setting thresholds and secondary verification, filters low-confidence information, ensures that the final retained effective information features are real and reliable, effectively deals with complex data scenarios, meets the diverse needs of data processing in different learning plan development processes, and enhances the universality and applicability of the method.

[0081] In an embodiment of the present application, based on step S104, a possible embodiment will be given below to illustrate the specific implementation thereof. As shown in Figure 3 Step S104 specifically includes: Step S1041: The text nodes and multimedia nodes in the course knowledge graph are mapped to a unified semantic space through a cross-modal embedding model together with the student completion rate and interaction comments in the learning data to generate feature vectors containing visual, textual, and numerical information.

[0082] Step S1042: Based on the target type of the learning plan and the context information, a priority strategy for double-channel retrieval is defined.

[0083] Step S1043: When performing double-channel collaborative retrieval, a gating attention mechanism is used to fuse the vector similarity and the keyword matching score, and by calculating the gating value, the weighting coefficients of BM25 and FAISS are adjusted based on the field relevance and policy association strength of the current retrieval results. Step S1043 specifically includes: fusing the feature vectors into a continuous gating value between 0 and 1; setting a weight adjustment rule according to the learning plan stage, course type and student group characteristics, triggering FAISS weight priority processing and triggering BM25 weight focus on skill keywords; using a piecewise function to perform nonlinear transformation on the BM25 and FAISS scores, and then assigning weight proportions according to the gating value, when the gating value > 0.7, FAISS weight = gating value x 0.8 + 0.2, BM25 weight = 1-FAISS weight; matching the weighted ranked candidate courses with the historical success case library, if the matching degree is lower than the threshold, triggering manual review, and recording the gating value and weight parameters adjusted each time.

[0084] Step S1044: performing feasibility check on the weighted ranked candidate courses, combining course duration, lecturer available time, student geographical restriction constraints, filtering courses that do not meet the actual conditions, and ranking the remaining results, and outputting the final course matching list.

[0085] In some specific embodiments, the cross-modal data mapping and dual-channel retrieval are combined, which can fully utilize the multi-dimensional information of courses and learning data, and greatly improve the matching accuracy of courses and learning plan requirements. According to the learning plan target and context, the retrieval priority and weight are adjusted, so that the retrieval strategy can flexibly adapt to different scene requirements. No matter what type of learning plan and student group, appropriate retrieval scheme can be developed to improve system applicability. Through the feasibility checking link, the actual course offering conditions are taken into account to avoid recommending courses that cannot be implemented due to time conflicts, geographical restrictions, etc., and to ensure that the finally recommended courses can be implemented.

[0086] As an embodiment of the above, assume that an online education platform wants to develop a "Python data analysis skill certification" learning plan for student Li San.

[0087] In step S1041, the platform maps the text introduction, teaching video of the Python data analysis course in the course knowledge graph, and the display of Li San's past course completion rate of 70% in programming courses, interactive comments in the learning community, and expression of interest in data visualization to a unified semantic space through cross-modal embedding model to generate video picture features, course introduction keywords, and completion rate feature vectors.

[0088] Step S1042 determines the dual-channel retrieval priority strategy as: prioritizing keyword matching retrieval of courses related to the Python data analysis skill certification examination outline, while also considering vector similarity retrieval of similar high-quality courses based on the learning plan target of "skill certification".

[0089] In step S1043, dual-channel collaborative retrieval is performed. The feature vector is first processed into a gating value 0.6. Since Li San is in the skill authentication sprint phase, and the course type is a technical operation type, the system sets the trigger BM25 weight according to the weight adjustment rule, focusing on skill keywords such as "Python data cleaning" and "Pandas library application". After nonlinear transformation of the BM25 and FAISS scores by the segmentation function, the weight is calculated according to the gating value, FAISS weight = 0.6 x 0.8 + 0.2 = 0.68, BM25 weight = 1 - 0.68 = 0.32, and the retrieval results are weighted and sorted. Then the candidate courses are compared with the case library of historical successful Python data analysis skill authentication, and it is found that the matching degree of a certain course is only 60%, which is lower than the threshold 70%, triggering manual review, and recording the gating value 0.6 and the weight parameter.

[0090] Finally, in step S1044, the candidate courses after weighted sorting are subjected to feasibility verification. It is found that one of the courses has a high matching degree, but the course duration is 60 hours, far exceeding the 30 hours of learning time that Li San can invest, and is filtered out; the lecturer of another course cannot teach online in the area where Li San is located, and is also excluded. The remaining courses that meet the conditions are sorted again, and a final adaptive course list containing 3 courses is output to Li San.

[0091] As an example of the present application, the above-mentioned intelligent learning plan generation method based on a large model can be implemented based on an intelligent learning plan generation system. The multi-modal document analysis module in the system uses a BERT+BiLSTM model to perform entity recognition and key clause extraction on policy documents. The course knowledge graph construction module associates the course domain, target object, and policy of the teacher course library. The hybrid search engine module can combine vector similarity (FAISS) and keyword matching (ElasticSearch) dual-channel retrieval, where the FAISS index uses IVF_PQ (inverted index + product quantization), and the ElasticSearch uses IK tokenizer + ik_max_word (index) + ik_smart (query). The course scheduling algorithm module is based on a time planner of the constraint satisfaction problem (CSP) model. The human-computer interaction module supports a visual course table editing interface for drag-and-drop adjustment.

[0092] It should be understood that the size of the serial number of each step in the above-mentioned embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0093] As Figure 4As shown, the present application also provides an electronic device comprising a display module 103, a memory 102, a processor 101, and a computer program stored in the memory and capable of running on the processor 101, wherein the processor 101 implements the steps of the large model-based intelligent learning plan generation method when executing the program.

[0094] In embodiments of the present application, the electronic device includes, but is not limited to, a laptop computer, a desktop computer, a workstation, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections, and relationships, and their functions, are shown as examples only and are not meant to limit implementations of the present application described and / or claimed herein.

[0095] In embodiments of the present application, the processor 101 can be implemented by using at least one of an application-specific integrated circuit, a programmable logic device, a field programmable gate array, a processor, a controller, a microcontroller, a microprocessor, an electronic unit designed to perform the functions described herein, and in some cases, such implementation can be implemented in a controller. For software implementation, the implementation of such as processes or functions can be implemented with separate software modules allowing at least one function or operation to be performed, and the software code can be implemented by a software application (or program) written in any appropriate programming language and stored in a memory and executed by a controller.

[0096] The display module 103 is used to display information input by a user or information provided to a user. The display module 103 can include a display panel, which can be configured in the form of a liquid crystal display, an organic light-emitting diode, etc.

[0097] The memory 102 can be used to store software programs and various data. The memory 102 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device.

[0098] The electronic device implementing the intelligent learning plan generation method is a unit and algorithm steps of each example described in combination with the embodiments disclosed herein, which can be realized by electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described generally in the above description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0099] Through the above description of the embodiments, those skilled in the art can easily understand that the electronic device implementing the intelligent learning plan generation method described herein can be realized by software or by a combination of software and necessary hardware. Therefore, the technical solution of the disclosed embodiment according to the intelligent learning plan generation method can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.) or a network, and includes a number of instructions to make a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) execute the indexing method according to the disclosed embodiment.

[0100] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the examples shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating intelligent learning plans based on large models, characterized in that, The methods include: S101: Capture learning data from the learning plan; S102: The BERT+BiLSTM model is used to process the learning data, identify entities and key terms, and perform summary generation and topic modeling to obtain key information features of the learning data. S103: Construct a course knowledge graph and label the course knowledge graph, including the course domain, target audience, and policy relevance; S104: Based on the course knowledge graph and key information features, the course is matched with the course through hybrid retrieval. The dual-channel collaborative multi-way retrieval of vector similarity and keyword matching is combined, and the weighted ranking is performed by BM25 and FAISS similarity scores. S105: Optimize multi-objective scheduling information based on genetic algorithm, and combine it with a time planner based on constraint satisfaction problem model to generate a course schedule plan; S106: A visual timetable editing interface that supports drag-and-drop adjustments allows users to review and adjust course plans, and the results are then communicated to relevant instructors and students.

2. The intelligent learning plan generation method based on a large model according to claim 1, characterized in that, Step S101 specifically includes: The learning data captured from the learning plan includes course completion rate, learning duration, and assessment results. The captured learning data is monitored in real time. By setting data update thresholds and time intervals, when the learning data changes or meets the update conditions, the data capture task is triggered, so that the learning plan is optimized based on the latest data. Sensitive information is encrypted when collecting learning data, and user privacy data is encrypted.

3. The intelligent learning plan generation method based on a large model according to claim 1, characterized in that, Step S102 specifically includes: S1021: Perform multimodal alignment and noise filtering on the captured learning data. By linking timestamps and matching content relevance, structured course schedules, unstructured policy texts, and semi-structured student interaction records are unified into the same spatiotemporal context. Historical data, duplicate records, and content with format errors that are irrelevant to the current learning plan objectives are filtered out. S1022: Based on the input characteristics of BERT, add positional identifier vectors to the learning data to mark the association order between policy clauses and course objectives, as well as domain label vectors; S1023: Parallel execution of four sub-tasks of learning data: entity recognition, key clause extraction, summary generation, and topic modeling, sharing the underlying feature representation of BERT+BiLSTM; among them, entity recognition and key clause extraction use the clause content of the learning plan as the contextual constraint of the entity, and summary generation and topic modeling achieve content focusing through an attention mechanism. S1024: Perform confidence calibration and result filtering on the output results of multiple tasks. Set thresholds based on the category confidence of entity recognition and the semantic similarity between the summary and the original text to filter low-confidence results and perform secondary verification to retain valid information features that meet the requirements.

4. The intelligent learning plan generation method based on a large model according to claim 3, characterized in that, Step S1023 specifically includes: S10231: Semantic hierarchical encoding is performed on the learning data. Based on the deep semantic representation capability of the BERT pre-trained model, the original text is input into BERT for context encoding to obtain a word vector sequence containing global semantic information. At the same time, the sequence is modeled with bidirectional long-range dependencies through BiLSTM layers to capture the cross-sentence association between policy clauses and course objectives, and output a hidden state sequence that integrates context and sequence information. S10232: Apply contextual constraints to the entity recognition and key clause extraction tasks, take the explicit clause content in the learning plan as the condition input to the gate unit of the BiLSTM layer, embed the vector as the bias term when calculating the hidden state, and associate the core objective of the current learning plan when extracting entities and clauses. S10233: Design a multi-scale attention mechanism for the tasks of summarization and topic modeling. The summarization adopts content-based self-attention, and obtains key information by calculating the similarity between word vectors and global semantic centers. S10234: By using the underlying hidden state sequence of BERT+BiLSTM, the domain keywords extracted by the entity recognition task are fed back to the summary generation task in real time; the high-frequency topics identified by the topic modeling task are passed as constraints to the key clause extraction task to limit the scope of clause extraction.

5. The intelligent learning plan generation method based on a large model according to claim 3, characterized in that, Step S1024 specifically includes: S10241: Evaluate the confidence of the multi-task output results, define the context consistency confidence and semantic coherence confidence, and form a comprehensive confidence vector; S10242: Adjust the confidence thresholds for each dimension based on the learning plan target type and data distribution characteristics; S10243: Perform secondary verification on the selected learning plan results. Verification includes: the entity appears in both the policy clauses and the course syllabus; the summary must cover at least 3 learning plan topics; and there is an intersection between the learning topic modeling results and the policy keyword database. S10244: The results of secondary verification will be integrated, entities will be mapped to corresponding policy provisions, and summaries will be bound to learning topic tags to form a set of information features.

6. The intelligent learning plan generation method based on a large model according to claim 1, characterized in that, Step S103 specifically includes: The course knowledge base is formed by extracting course names, durations, and instructors from the course management system, as well as courseware, videos, and exercise data from the teaching resource database, and collecting completion rate and assessment scores from the student behavior database, and policy clauses and compliance requirements data associated with the policy document database. Based on course domain labeling, a learning plan structure of professional skills, Python development, and data analysis is adopted, and student tags are defined in combination with student profiles. Policy association labeling is carried out through keyword matching and policy clause mapping to mark the degree of relevance. Construct a course knowledge graph and define matching relationships including course-policy coverage, course-teaching resources, and course-student tags to enhance the semantic expression dimension of the graph; Examine the logical consistency between the course domain and the applicable students, the policy relevance and the matching of the course content, use graph neural networks to predict links in the knowledge graph, fill in the relationships between unlabeled courses and students, and finally output the course knowledge graph.

7. The intelligent learning plan generation method based on a large model according to claim 1, characterized in that, Step S104 specifically includes: S1041: Map the text nodes and multimedia nodes in the course knowledge graph with the student completion rate and interactive comments in the learning data to a unified semantic space through a cross-modal embedding model to generate feature vectors containing visual, text, and numerical data. S1042: Define a priority strategy for dual-channel retrieval based on the target type and context information of the learning plan; S1043: When performing dual-channel collaborative retrieval, a gated attention mechanism is used to fuse vector similarity and keyword matching scores. By defining a gate value, the weighting coefficients of BM25 and FAISS are adjusted based on the domain relevance and policy relevance of the current retrieval results. S1044: Perform feasibility verification on the weighted candidate courses. Based on the course duration, available time of the instructor, and geographical restrictions of the students, filter out courses that do not meet the actual conditions, sort the remaining results, and output the final list of matched courses.

8. The intelligent learning plan generation method based on a large model according to claim 7, characterized in that, Step S1043 specifically includes: The feature vectors are fused into continuous gating values ​​between 0 and 1; Based on the learning plan stage, course type, and student group characteristics, set weight adjustment rules to prioritize FAISS weights and emphasize skill keywords when triggering BM25 weights. A piecewise function is used to perform a nonlinear transformation on the BM25 and FAISS scores, and then the weight ratio is allocated according to the gate value. When the gate value is > 0.7, the FAISS weight = gate value × 0.8 + 0.2, and the BM25 weight = 1 - FAISS weight. The weighted candidate courses are matched with a database of historical success cases. If the matching degree is lower than the threshold, manual review is triggered, and the gate value and weight parameters are recorded each time.

9. The intelligent learning plan generation method based on a large model according to claim 1, characterized in that, Step S105 specifically includes: Based on the actual needs of the learning plan, the optimization goals were expanded to three dimensions: no overlap in course time slots, balanced load of lecturers, classrooms, and equipment during the same time slots, and daily learning time for students not exceeding the threshold and alternating course types. Each objective can be set with a quantifiable mathematical expression as follows: Number of time conflicts = Σ (overlapping time periods between course i and course j), Resource utilization rate = Σ (teaching hours of lecturer i on the day / total available time)). Constraints are extracted from the course knowledge graph and key information features, including: fixed course duration, unavailable time periods for instructors, geographical restrictions for students, and course domain and student suitability ≥ 80%, with policy-related courses given priority in scheduling. These constraints are then transformed into adjustable penalty coefficients. An improved genetic algorithm optimization process is defined, using real number encoding to represent course scheduling, introducing domain knowledge guidance in the crossover operation, combining constraint checks of the CSP model in the mutation operation, and simultaneously calculating the weighted sum of the three optimization objectives during each generation population evaluation. A time planner is constructed. The initial course schedule output by the genetic algorithm is processed by a backtracking algorithm to detect time conflicts and make local adjustments by calling alternative resources in the course knowledge graph. After adjustment, the optimization target score is recalculated until a final course schedule that is conflict-free and meets all constraints is generated.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent learning plan generation method based on a large model as described in any one of claims 1 to 9.

Images