A knowledge base-based teaching content online generation method and system
By using CNN-BiLSTM and Transformer models in conjunction with knowledge graphs, personalized teaching content is generated, which solves the problems of low efficiency and poor adaptability of traditional teaching content acquisition methods and achieves efficient and personalized teaching content generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ANHUI EDUCATION NETWORK PUBLISHING
- Filing Date
- 2025-07-21
- Publication Date
- 2026-06-09
Smart Images
Figure CN120931442B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence teaching technology, and more specifically, to a method and system for online generation of teaching content based on a knowledge base. Background Technology
[0002] In today's era of booming digital education, the ways in which educational resources are acquired and utilized are undergoing profound changes. Traditional acquisition of teaching content mainly relies on teachers' personal knowledge, limited textbooks, and some general educational website resources. This approach has many drawbacks.
[0003] From a teacher's perspective, lesson preparation requires a significant investment of time and energy to collect and organize teaching materials from various sources, including texts, images, and videos. These materials are often scattered across different platforms, making integration difficult. Furthermore, the varying knowledge and teaching experience among teachers leads to inconsistent quality of teaching content. For instance, in some remote schools, teachers lack access to quality resources and are forced to rely solely on textbooks, resulting in a limited range of teaching methods that fails to meet the diverse learning needs of students.
[0004] From the student's perspective, with the deepening of the concept of personalized education, each student has significant differences in learning style, knowledge base, interests, and hobbies. They need personalized learning content to improve their learning outcomes. However, the universality of traditional teaching content cannot accurately adapt to the individual differences of students. For example, students with strong learning abilities may find the teaching content too simple and lacking in challenge; while students with weak foundations may find it difficult to keep up with the teaching progress and develop a fear of learning. Summary of the Invention
[0005] The purpose of this application is to provide a knowledge base-based online method and system for generating teaching content, so as to at least solve the technical problems existing in the teaching content acquisition methods of related technologies, such as teachers spending a lot of time and effort to collect and integrate materials, inconsistent quality of teaching content, and general content that cannot be adapted to individual student differences.
[0006] To achieve the above objectives, the embodiments of this application provide the following technical solutions.
[0007] In a first aspect, according to one embodiment of this application, a method for online generation of teaching content based on a knowledge base is provided, comprising the following steps:
[0008] In response to the teaching content generation request triggered by the instructor, the teaching content generation request is converted into graph data;
[0009] The graph data is semantically parsed using a semantic parsing model based on CNN-BiLSTM. The graph data is transformed into a fusion matrix by concatenating the adjacency matrix and the feature matrix through channel concatenation. Local correlation features between nodes are extracted by multi-scale convolution kernels of CNN. Then, the local feature sequence is globally semantically fused by the BiLSTM model to output structured retrieval conditions containing knowledge attribute labels, difficulty thresholds, and presentation format identifiers.
[0010] Based on the structured retrieval conditions, a preset knowledge graph is invoked for associated retrieval; wherein, the knowledge graph contains knowledge point entity nodes and causal, progressive, and supplementary relationship edges between nodes. Core knowledge point nodes are located by entity recognition, and the subordinate child nodes of the core knowledge point nodes are traversed to obtain the corresponding associated knowledge units. Knowledge units that meet the characteristics of the student profile are selected according to the difficulty threshold to form a candidate knowledge set.
[0011] The knowledge graph relation edge weights are weighted and calculated based on their matching degree with the teaching objectives to generate a unique sorted sequence of knowledge units. The sorted knowledge units are then integrated using a Transformer model, and semantic conflict detection is performed on the generated content and the expressions in the knowledge base. Conflicting knowledge units are labeled with credibility scores to form the teaching content.
[0012] Preferably, the method further includes the following steps:
[0013] During the preview of teaching content by students, dynamic evaluation data is collected through the real-time perception module to construct a temporary cognitive profile of students. The evaluation data includes the time students spend previewing teaching content, the key knowledge points they mark, and the questions they ask in real time.
[0014] By integrating temporary profiles with students' historical learning data, outdated knowledge units are removed and the presentation format of the remaining units is adjusted to output personalized teaching content.
[0015] Preferably, the step of converting the teaching content generation request into graph data includes:
[0016] By using word segmentation tools, speech recognition models, and handwriting feature extraction tools in natural language processing, multimodal element extraction is performed on the teaching content generation request to obtain entity elements, attribute elements, teaching scenario feature elements, speech features, and handwritten annotation features. Among them, the teaching scenario feature elements are used to represent the type of teaching scenario; speech features include speech rate and intonation changes, which are used to represent the urgency of the request or the key points to be emphasized; handwritten annotation features include underlines and asterisks, which are used to represent core needs.
[0017] Entity elements are transformed into entity nodes of graph data, and attribute elements are transformed into attribute nodes of graph data. Node associations are established through relational edges to obtain graph data. Entity nodes have a unique code in the subject knowledge base, and attribute nodes are attached to at least one entity node. The weight values of relational edges are determined through a dynamic relational weight self-adjustment mechanism. This dynamic relational weight self-adjustment mechanism includes: constructing a multimodal feature vector from extracted text scene features, speech features, and handwritten annotation features; inputting this vector into a weight allocation model based on an attention mechanism; calculating the influence weight of each modality feature; and determining the dynamic weight values of relational edges based on the weighted fusion result.
[0018] Preferably, the step of extracting local correlation features between nodes using multi-scale convolutional kernels of CNN includes:
[0019] Configure at least two different sizes of convolution kernels, including square convolution kernels and asymmetric convolution kernels. Each convolution kernel corresponds to a preset stride parameter and activation function, which are used to match node association patterns in different ranges in graph data, including symmetric association of close nodes and asymmetric association of mid-distance nodes.
[0020] The semantic complexity analysis module calculates the association asymmetry index of graph data, which includes the ratio of the number of entity nodes to the number of attribute nodes and the difference in the length of indirect association paths.
[0021] If the correlation asymmetry index exceeds a preset threshold, the asymmetric convolution kernel generation mechanism is activated. The weight parameters of the asymmetric convolution kernel are optimized through a generative adversarial network (GAN), and the asymmetric convolution kernel is used to adapt complex asymmetric correlation regions in the graph data.
[0022] Convolutional kernels of different scales are applied to local regions of the fusion matrix, and node association features within the corresponding regions are extracted by a sliding window calculation to generate multiple sets of local feature mapping matrices.
[0023] The local feature mapping matrices output by convolutional kernels of different scales are channel-concatenated to form an aggregated feature matrix containing multi-scale correlation information. The aggregated feature matrix is used as the input of the BiLSTM model.
[0024] Preferably, a fusion matrix of dimension N×N is defined as X, and two convolutional kernels of different scales are configured; wherein, the square small-scale convolutional kernel is represented as... Step size is The weight matrix is Asymmetric large-scale convolution kernels are represented as follows: Step size is The weight matrix is Weight matrix It was obtained through training and optimization using a Generative Adversarial Network (GAN).
[0025] For symmetric associations of nearby nodes, the calculation is performed by sliding a small-scale square convolution kernel, and is expressed as follows:
[0026]
[0027] For asymmetric associations of mid-distance nodes, the calculation is performed by sliding an asymmetric large-scale convolution kernel, and is expressed as follows:
[0028]
[0029] In the formula, Represents a small-scale feature mapping matrix. This represents the activation function. , The term represents the bias term, i and j represent the indexes of the feature map; p represents the row index of the convolution kernel and the local region of the fusion matrix, and q represents the column index of the convolution kernel and the local region of the fusion matrix. This represents the side length of the square convolution kernel. , These represent the number of rows and columns of the asymmetric convolution kernel, respectively. , This indicates the stride of the corresponding convolutional kernel; , This represents the weight matrix corresponding to the convolution kernel;
[0030] The square kernel feature mapping matrix F1 and the asymmetric kernel feature mapping matrix F2 are concatenated along the channel dimension to form the aggregated feature F, which is represented as: ,in, This indicates a channel cascading operation.
[0031] Preferably, the step of performing global semantic fusion on local feature sequences using a BiLSTM model to output structured retrieval conditions containing knowledge attribute labels, difficulty thresholds, and presentation format identifiers includes:
[0032] The aggregated feature matrix output by the CNN is converted into a one-dimensional local feature sequence by rows or columns. Each sequence element corresponds to a local associated feature, and a positional code is attached to represent the position of the feature in the global sequence.
[0033] Three parallel attention branches are introduced into the hidden layer of BiLSTM to generate knowledge attribute labels, difficulty thresholds, and presentation format identifiers, respectively. Each attention branch calculates the attention of the feature sequence through a weight matrix and fuses the attention weights with the hidden state of BiLSTM to generate a goal-oriented intermediate feature vector.
[0034] After concatenating the intermediate feature vectors of the three branches, the concatenation is input into the fully connected layer, which initially outputs knowledge attribute labels, difficulty thresholds, and presentation format identifiers.
[0035] A retrieval condition verification module is introduced, which constructs a feature sequence-valid retrieval condition mapping library based on historical generated data, calibrates the preliminary output, and finally outputs calibrated structured retrieval conditions.
[0036] Preferably, in the step of calling a preset knowledge graph for association retrieval, the method for constructing the knowledge graph includes the following steps:
[0037] Based on the knowledge base, the subject knowledge system extracts core knowledge points and their subordinate sub-knowledge points as entity nodes. At the same time, common student error patterns are extracted from the student's historical answer data as error pattern nodes, forming a hierarchical node system. Among them, the core knowledge point node corresponds to the key concept in the subject, the subordinate sub-node contains the detailed content of the knowledge point, and the error pattern node is associated with the knowledge point node that is prone to the error. Each node is assigned a unique identifier.
[0038] To address the logical connections between nodes, causal relationship edges, progressive relationship edges, supplementary relationship edges, as well as analytical relationship edges and induced relationship edges between error pattern nodes and knowledge units are constructed. Among them, causal relationship edges are used to connect knowledge points with causal deductions, progressive relationship edges are used to connect knowledge points with learning order, supplementary relationship edges are used to connect knowledge points that complement each other, analytical relationship edges are used to connect knowledge units that can explain the error pattern, and induced relationship edges are used to connect knowledge units that are prone to causing the error pattern.
[0039] Each relation edge is labeled with a weight value. The weight of the parsed relation edge is set based on the historical effectiveness of the knowledge unit in correcting the error, and the weight of the induced relation edge is set based on the association frequency between the knowledge unit and the error. A mapping relationship is established between each knowledge point node, error pattern node and corresponding knowledge unit. Each knowledge unit is associated with at least one node and an attribute label is attached to realize the association binding of knowledge point node-error pattern node-knowledge unit-attribute feature.
[0040] A dynamic update module for the knowledge graph is introduced, which regularly updates the node system, relational edge weights, and knowledge unit mapping by collecting new knowledge points, relationships, knowledge units, and new student error patterns from teaching practice.
[0041] Preferably, the step of fusing the sorted knowledge units using the Transformer model includes:
[0042] Obtain a weighted sequence of knowledge units, which contains several knowledge units sorted by their relevance strength and their matching degree with the teaching objectives.
[0043] Extract teaching style feature vectors, which are obtained by training based on the instructor's historical teaching content;
[0044] Each knowledge unit is transformed into a standardized feature vector and fused with the teaching style feature vector. This is then input into the self-attention layer of the Transformer encoder to calculate the semantic association weights and style adaptation weights between units.
[0045] Multi-dimensional fusion optimization is achieved through parallel processing of multi-head attention, whereby the multi-dimensional aspects include terminology consistency, logical coherence, and style consistency.
[0046] Based on the fusion feature vector output by the encoder, the decoder generates basic text content that conforms to the teaching style. According to the presentation format identifier in the structured retrieval conditions, the multimodal conversion module is called to convert the basic text content into teaching content of the corresponding modality. The multimodality includes text, static illustrations, dynamic demonstration videos and audio narration.
[0047] The final generated teaching content is a multimodal structured content that integrates teaching styles.
[0048] Preferably, the step of extracting the teaching style feature vector includes:
[0049] Collect historical teaching materials from instructors and categorize and label them;
[0050] Text features of various materials are extracted using a pre-trained BERT model, and a style feature library is constructed by combining the speech rate and pause features of the speech-to-text text.
[0051] The K-means clustering algorithm is used to cluster the style feature library to generate style labels;
[0052] Based on the feature matching degree of the instructor's latest teaching materials, the corresponding teaching style feature vector is dynamically updated.
[0053] According to another embodiment of this application, an online teaching content generation system based on a knowledge base is provided, including the following modules:
[0054] The data conversion module is used to respond to the teaching content generation request triggered by the instructor and convert the teaching content generation request into graph data;
[0055] The retrieval condition construction module is used to perform semantic parsing on the graph data using a semantic parsing model based on CNN-BiLSTM. The graph data is transformed into a fusion matrix that is concatenated by concatenating the adjacency matrix and the feature matrix through a channel concatenation method. Local correlation features between nodes are extracted by multi-scale convolution kernels of CNN. Then, the local feature sequence is globally semantically fused by the BiLSTM model to output structured retrieval conditions containing knowledge attribute labels, difficulty thresholds, and presentation format identifiers.
[0056] The associated retrieval module is used to call a preset knowledge graph to perform associated retrieval based on the structured retrieval conditions. The knowledge graph includes knowledge point entity nodes and causal, progressive, and supplementary relationship edges between nodes. The core knowledge point nodes are located by entity recognition, and the subordinate child nodes of the core knowledge point nodes are traversed to obtain the corresponding associated knowledge units. The knowledge units that meet the characteristics of the student profile are selected according to the difficulty threshold to form a candidate knowledge set.
[0057] The content generation module is used to perform weighted calculations based on the matching degree between the edge weights of the knowledge graph and the teaching objectives, generating a unique sorted sequence of knowledge units. The sorted knowledge units are then fused using the Transformer model, and semantic conflict detection is performed on the generated content and the expressions in the knowledge base. Conflicting knowledge units are labeled with credibility scores to form the teaching content.
[0058] Compared with the prior art, the technical effects of the online generation method and system for knowledge base-based teaching content in this application include the following aspects:
[0059] First, this application's embodiments extract multimodal elements, transforming the request text into graph data containing entity nodes, attribute nodes, and dynamic relational edges. Entity nodes correspond to unique codes in a subject knowledge base, attribute nodes exist in relation to entity nodes, and the weights of relational edges are dynamically determined through an attention-based weight allocation model. This not only makes implicit connections in the text explicit but also quantifies the constraint strength between elements through a dynamic weight adjustment mechanism. This solves the semantic loss problems caused by element isolation and poor adaptability of static weights in traditional text parsing, providing a structured and precise input foundation for subsequent semantic parsing.
[0060] Second, this application achieves high-precision semantic parsing through the collaborative mechanism of CNN and BiLSTM: CNN is configured with two types of convolutional kernels, square and asymmetric. The semantic complexity analysis module calculates the association asymmetry index. When the index exceeds the threshold, the GAN-optimized asymmetric convolutional kernel is activated to accurately capture complex asymmetric associations. BiLSTM introduces three parallel attention branches, combined with the retrieval condition verification module, which not only solves the problem of insufficient adaptation of single-scale convolutional kernels to complex graph structures, but also makes up for the shortcomings of traditional LSTM in target-oriented semantic fusion, thus significantly improving the output accuracy of structured retrieval conditions.
[0061] Third, the knowledge graph of this application achieves logic-driven associative retrieval through a hierarchical system of core knowledge point nodes, subordinate child nodes, and error pattern nodes, combined with causal, progressive, and supplementary relationship edges, as well as analytical and induced relationship edges of error pattern nodes: after locating the core node through entity recognition, the child nodes are traversed to obtain associated knowledge units, and at the same time, analytical units corresponding to common student errors are preferentially matched based on error pattern nodes; combined with difficulty thresholds and student profiles, it ensures that knowledge units not only focus on the subdivided content of the target knowledge points to avoid being too broad, but also adapt to the student's current level and weak points, thus solving the problems of mismatch between knowledge units and student level and lack of error targeting in traditional retrieval;
[0062] Fourth, this application generates a unique knowledge unit ranking sequence by weighted calculation of the matching degree between the knowledge graph relation edge weights and the teaching objectives. When the Transformer model is integrated, the instructor's teaching style feature vector is extracted first. After the knowledge unit features and style features are integrated, the consistency of terminology, logical coherence, and style uniformity are optimized in parallel through multi-head attention. Then, the multimodal transformation module generates content in the form of text, static illustrations, dynamic videos, etc., which avoids the problem of the relevance and teaching objectives being disconnected in traditional ranking. Furthermore, through style transfer and multimodal output, the generated content not only retains the core information of the knowledge unit and forms a coherent logical chain, but also adapts to the instructor's teaching habits and the student's learning preferences. Attached Figure Description
[0063] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:
[0064] Figure 1 This is a flowchart illustrating the implementation of an online teaching content generation method based on a knowledge base, according to an embodiment of this application.
[0065] Figure 2 Sub-processes of the knowledge base-based online teaching content generation method provided in the embodiments of this application Figure 1 ;
[0066] Figure 3 Sub-processes of the knowledge base-based online teaching content generation method provided in the embodiments of this application Figure 2 ;
[0067] Figure 4 Sub-processes of the knowledge base-based online teaching content generation method provided in the embodiments of this application Figure 3 ;
[0068] Figure 5 Sub-processes of the knowledge base-based online teaching content generation method provided in the embodiments of this application Figure 4 ;
[0069] Figure 6 A structural block diagram of a knowledge base-based online teaching content generation system provided in another embodiment of this application. Detailed Implementation
[0070] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.
[0071] The following describes the related technologies of the embodiments of this disclosure. The following related technologies are optional solutions and can be combined with the technical solutions of the embodiments of this disclosure in any way, and they all fall within the protection scope of the embodiments of this disclosure.
[0072] According to one embodiment of this application, a method for generating online teaching content based on a knowledge base is provided, the method including at least a portion of the following:
[0073] like Figure 1 As shown, step S101 is included, in which, in response to the teaching content generation request triggered by the teacher, the teaching content generation request is converted into graph data.
[0074] Specifically, step S101 is the process of transforming the instructor's teaching content generation requirements into structured graph data, which is achieved through multimodal feature extraction and node association construction.
[0075] In one implementation of this application, please refer to Figure 2 The steps to convert teaching content generation requests into graph data include the following:
[0076] S201: Using word segmentation tools, speech recognition models, and handwriting feature extraction tools in natural language processing, multimodal element extraction is performed on the teaching content generation request to obtain entity elements, attribute elements, teaching scene feature elements, speech features, and handwritten annotation features;
[0077] The characteristic elements of a teaching scenario are used to characterize the type of teaching scenario;
[0078] Speech features include speech rate and intonation variations, which are used to characterize the urgency of a request or the emphasis of a particular content;
[0079] Handwritten annotation features include underlines and asterisks, used to represent core requirements;
[0080] In one embodiment, the word segmentation tool uses the word segmentation module of the BERT pre-trained model to segment the text words of the teaching content generation request; the speech recognition model is used to convert speech signals into text and extract acoustic features such as speech rate and intonation through the speech feature analysis module; the handwriting feature extraction tool can use a convolutional neural network to extract features from handwritten annotation images, identify the position and shape of marks such as underlines and asterisks, and convert them into feature data that represents the core requirements.
[0081] Furthermore, the step of converting the teaching content generation request into graph data also includes step S202, in which entity elements are converted into entity nodes of graph data, attribute elements are converted into attribute nodes of graph data, and node associations are established through relation edges to obtain graph data.
[0082] Specifically, in step S202, entity nodes have a unique code in the subject knowledge base, and attribute nodes are attached to at least one entity node. Entity nodes (such as subjects, core knowledge points, etc.) have a unique code in the subject knowledge base, and each entity node has a clear and unique identity that can be accurately bound to the corresponding content in the knowledge base. Attribute nodes (such as difficulty, number of questions, duration, etc.) cannot exist independently and must be associated with at least one entity node. The attachment relationship clarifies the object to which the attribute belongs, preventing attribute information from becoming isolated elements.
[0083] In addition, in this embodiment, the weight value of the relation edge is determined by a dynamic relation weight self-adjustment mechanism. This mechanism, through the dynamic weighting of multimodal features, allows the association weight between nodes in the graph data to be adaptively adjusted according to the details of the teacher's request, thus solving the problem that traditional fixed weights cannot accurately match complex teaching needs.
[0084] Specifically, the dynamic relation weight self-adjustment mechanism provided in this embodiment includes: constructing a multimodal feature vector from the extracted text scene feature elements, speech features and handwritten annotation features, inputting it into a weight allocation model based on an attention mechanism, calculating the influence weight of each modality feature, and determining the dynamic weight value of the relation edge based on the weighted fusion result.
[0085] For example, for teaching scenario feature elements, this application first presets the basic weight range according to the node type, and then uses the XGBoost model trained based on historical teaching case data as the dynamic weight adjustment sub-model. The teaching scenario feature elements are used as the input of the dynamic weight adjustment sub-model, and the output is the weight adjustment coefficient. The final relation edge weight is the product of the basic weight and the adjustment coefficient.
[0086] As can be seen, the embodiments of this application make implicit relationships in text explicit by modeling graph data composed of entity nodes, attribute nodes, and relation edges. It not only retains the information of individual elements, but also quantifies the constraint strength between elements through relation edge weights. The structured representation enables subsequent semantic parsing to accurately capture deep logic, solving the semantic loss problem caused by element isolation in traditional text parsing.
[0087] Please continue to refer to Figure 1 In this embodiment of the application, the online generation method of teaching content based on knowledge base further includes step S102. In step S102, a semantic parsing model based on CNN-BiLSTM is used to perform semantic parsing on the graph data. The graph data is converted into a fusion matrix by concatenating the adjacency matrix and the feature matrix through channel concatenation. Local correlation features between nodes are extracted by multi-scale convolution kernel of CNN. Then, the local feature sequence is globally semantically fused by BiLSTM model to output structured retrieval conditions containing knowledge attribute labels, difficulty thresholds, and presentation format identifiers.
[0088] Therefore, this embodiment combines the advantages of Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory Network (BiLSTM) to perform semantic parsing on graph data in order to output structured retrieval conditions.
[0089] In the CNN feature extraction module, square convolution kernels and asymmetric convolution kernels are used to perform convolution operations on the fusion matrix. Each convolution kernel corresponds to a preset stride parameter and activation function, which can specifically capture the node association patterns in different ranges in the graph data, such as the symmetric association of close nodes and the asymmetric association of mid-distance nodes, and then extract the local association features between nodes.
[0090] In the BiLSTM semantic fusion module, this embodiment receives the local features output by the CNN and converts them into a local feature sequence. The BiLSTM model processes the local feature sequence and can simultaneously consider the past and future contextual information in the sequence to achieve global semantic fusion. Three parallel attention branches are introduced into the hidden layer of the BiLSTM, which correspond to the generation of knowledge attribute labels, difficulty thresholds, and presentation format identifiers, respectively. Each attention branch calculates the attention of the feature sequence through a weight matrix and fuses the attention weights with the hidden state of the BiLSTM to generate a goal-oriented intermediate feature vector.
[0091] After concatenating the intermediate feature vectors generated by the three attention branches, the input is fed into the fully connected layer for processing, and the initial output includes knowledge attribute labels, difficulty thresholds, and presentation format identifiers.
[0092] Subsequently, this application embodiment introduces a search condition verification module, which constructs a feature sequence-valid search condition mapping library based on historical generated data, calibrates the preliminary output results, and finally outputs accurate structured search conditions.
[0093] This application utilizes a collaborative mechanism that combines CNN to extract local associations with BiLSTM to fuse global semantics, resulting in higher accuracy of the final output structured retrieval conditions. Specifically, the multi-scale convolutional kernels of CNN specifically capture node associations of different ranges, solving the problem of insufficient adaptation of single-scale models to complex graph structures. BiLSTM performs temporal modeling of local feature sequences, making up for the shortcomings of CNN in long-distance semantic dependencies and achieving complete analysis of local details and global logic.
[0094] Step S103: Based on the structured retrieval conditions, call the preset knowledge graph to perform associated retrieval; wherein, the knowledge graph contains knowledge point entity nodes and causal, progressive and supplementary relationship edges between nodes, locate the core knowledge point nodes through entity recognition, traverse the subordinate child nodes of the core knowledge point nodes to obtain the corresponding associated knowledge units, and filter out knowledge units that meet the characteristics of the student profile according to the difficulty threshold to form a candidate knowledge set.
[0095] This application achieves logic-driven associative retrieval through relational edges of a pre-defined knowledge graph: based on hierarchical traversal of core nodes and child nodes, it ensures that the retrieval results focus on the subdivided content of the target knowledge point and avoids an overly broad scope; further, by combining difficulty thresholds and student profiles for screening, it solves the problem of mismatch between knowledge units and student levels in traditional retrieval.
[0096] Step S104: Calculate the weighted values of the knowledge graph relationship edges and the matching degree with the teaching objectives to generate a unique sorted sequence of knowledge units; integrate the sorted knowledge units through the Transformer model, simultaneously detect semantic conflicts between the generated content and the expressions in the knowledge base, label the conflicting knowledge units with credibility scores, and form the teaching content.
[0097] Please continue to refer to Figure 1 According to a preferred embodiment of this application, the teaching content generation method of this application embodiment further includes step S105. In step S105, during the stage of students previewing teaching content, dynamic evaluation data is collected through a real-time perception module to construct a temporary cognitive profile of the students. The evaluation data includes the time spent by students when previewing teaching content, the marking of key knowledge points, and the content of immediate questions. The temporary profile is then integrated with the students' historical learning data, out-of-syllabus knowledge units are removed, and the presentation format of the retained units is adjusted to output personalized teaching content.
[0098] In this embodiment, during the process of students previewing teaching content (such as online courseware, videos, exercises, etc.), the system captures the students' interactive behavior data throughout the process through a real-time perception module, including dwell time, marking of key knowledge points, and immediate questions. Among them, dwell time is used to represent the students' difficulty in understanding this part of the knowledge points; marking of key knowledge points reflects the students' active attention to the knowledge points; and immediate questions reflect the students' confusion about specific content.
[0099] In constructing the temporary profile of the learner, based on the collected dynamic evaluation data, the system constructs a temporary cognitive profile, focusing on the immediate characteristics exposed by the learner during the current preview process, rather than long-term stable historical characteristics; the temporary profile includes: immediate understanding of difficulties: located by knowledge units whose dwell time exceeds a threshold; active focus on key points: core knowledge points extracted through marked behavior; real-time points of confusion: specific questions analyzed through immediate question content; the temporary cognitive profile in this embodiment has obvious time-sensitive characteristics and is only used to reflect the learner's dynamic state in this preview;
[0100] To avoid the randomness of temporary data (such as excessively long dwell time due to accidental operations), the system integrates temporary cognitive profiles with students' historical learning data to form a more comprehensive cognitive model. Historical data includes records of past incorrect answers, past knowledge mastery tags, and learning style preferences;
[0101] Based on the fused cognitive model, this application embodiment adjusts the initially generated teaching content.
[0102] Furthermore, such as Figure 3 As shown, the steps for extracting local correlation features between nodes using multi-scale convolutional kernels of CNN include:
[0103] S301. Configure at least two different sizes of convolution kernels, including square convolution kernels and asymmetric convolution kernels. Each convolution kernel corresponds to a preset stride parameter and activation function (such as ReLU function) to match node association patterns in different ranges in graph data, including symmetric association of close nodes and asymmetric association of mid-distance nodes.
[0104] S302. Calculate the association asymmetry index of graph data through the semantic complexity analysis module. The association asymmetry index includes the ratio of the number of entity nodes to the number of attribute nodes (total number of entity nodes / total number of attribute nodes) and the difference in the length of indirect association paths.
[0105] In this embodiment, the imbalance between core elements and parameter elements in the graph data is quantified by calculating the ratio of the number of entity nodes to the number of attribute nodes. For example, if there are 5 entity nodes and 2 attribute nodes in a graph, the ratio is 5÷2=2.5. The larger the ratio, the denser the entity nodes are relative to the attribute nodes, which may indicate an asymmetric association pattern.
[0106] S303. If the correlation asymmetry index exceeds the preset threshold, the asymmetric convolution kernel generation mechanism is activated. The weight parameters of the asymmetric convolution kernel are optimized by the generative adversarial network (GAN), and the asymmetric convolution kernel is used to adapt the complex asymmetric correlation region in the graph data.
[0107] S304. Apply convolution kernels of each scale to local regions of the fusion matrix respectively, and extract node association features within the corresponding regions through sliding window calculation to generate multiple sets of local feature mapping matrices.
[0108] S305. Channel concatenation is performed on the local feature mapping matrices output by convolutional kernels of different scales to form an aggregated feature matrix containing multi-scale correlation information. The aggregated feature matrix is used as the input of the BiLSTM model.
[0109] In one exemplary embodiment of this application, a fusion matrix of dimension N×N is defined as X, and two convolutional kernels of different scales are configured; wherein, the square small-scale convolutional kernel is represented as Step size is The weight matrix is Asymmetric large-scale convolution kernels are represented as follows: ,in, Step size is The weight matrix is Weight matrix It was obtained through training and optimization using a Generative Adversarial Network (GAN).
[0110] Furthermore, for symmetric associations of nearby nodes, the calculation is performed using a small-scale square convolution kernel, and can be expressed as follows:
[0111]
[0112] Furthermore, for the asymmetric association of mid-distance nodes, it is calculated through sliding asymmetric large-scale convolution kernels, and can be represented as:
[0113]
[0114] In the formula, Represents a small-scale feature mapping matrix. This represents the activation function. , The term represents the bias term, i and j represent the indexes of the feature map; p represents the row index of the convolution kernel and the local region of the fusion matrix, and q represents the column index of the convolution kernel and the local region of the fusion matrix. This represents the side length of the square convolution kernel. , These represent the number of rows and columns of the asymmetric convolution kernel, respectively. , This indicates the stride of the corresponding convolutional kernel; , This represents the weight matrix corresponding to the convolution kernel;
[0115] Furthermore, in this embodiment, the square kernel feature mapping matrix F1 and the asymmetric kernel feature mapping matrix F2 are concatenated along the channel dimension to form the aggregated feature F, which is represented as: ,in, This indicates a channel cascading operation.
[0116] Furthermore, in this embodiment of the application, the steps of performing global semantic fusion on local feature sequences using a BiLSTM model to output structured retrieval conditions containing knowledge attribute labels, difficulty thresholds, and presentation format identifiers include: converting the aggregated feature matrix output by the CNN into a one-dimensional local feature sequence by rows or columns, where each sequence element corresponds to a local associated feature, and adding a position code, which is used to characterize the position of the feature in the global sequence;
[0117] Three parallel attention branches are introduced into the hidden layer of BiLSTM to generate knowledge attribute labels, difficulty thresholds, and presentation format identifiers, respectively. Each attention branch calculates the attention of the feature sequence through a weight matrix and fuses the attention weights with the hidden state of BiLSTM to generate a goal-oriented intermediate feature vector.
[0118] After concatenating the intermediate feature vectors of the three branches, the concatenation is input into the fully connected layer, which initially outputs knowledge attribute labels, difficulty thresholds, and presentation format identifiers.
[0119] Furthermore, this embodiment introduces a search condition verification module, which constructs a feature sequence-valid search condition mapping library based on historical generated data, calibrates the preliminary output, and finally outputs calibrated structured search conditions.
[0120] As can be seen, this application achieves high-precision semantic parsing through the collaborative mechanism of CNN and BiLSTM: CNN is configured with two types of convolutional kernels, square and asymmetric, and calculates the association asymmetry index through the semantic complexity analysis module. When the index exceeds the threshold, the GAN-optimized asymmetric convolutional kernel is activated to accurately capture complex asymmetric associations; BiLSTM introduces three parallel attention branches, combined with the retrieval condition verification module, which not only solves the problem of insufficient adaptation of single-scale convolutional kernels to complex graph structures, but also makes up for the shortcomings of traditional LSTM in target-oriented semantic fusion, thus significantly improving the output accuracy of structured retrieval conditions.
[0121] Furthermore, such as Figure 4 As shown, in the step of calling a preset knowledge graph for association retrieval in this embodiment, the method of constructing the knowledge graph also includes step S401;
[0122] In step S401, based on the subject knowledge system of the knowledge base, core knowledge points and subordinate sub-knowledge points are extracted as entity nodes, and common student error patterns are extracted from the student's historical answer data as error pattern nodes, together forming a hierarchical node system.
[0123] In this embodiment of the application, the core knowledge point node corresponds to the key concept in the subject, the subordinate child nodes contain the detailed content of the knowledge point, the error pattern node is associated with the knowledge point node that is prone to the error, and each node is assigned a unique identifier.
[0124] Step S402: Based on the logical relationships between nodes, construct causal relationship edges, progressive relationship edges, supplementary relationship edges, as well as analytical relationship edges and induced relationship edges between error mode nodes and knowledge units;
[0125] In this embodiment, the causal relationship edge is used to connect knowledge points that have causal deductions, the progressive relationship edge is used to connect knowledge points that have a learning sequence, the supplementary relationship edge is used to connect knowledge points that complement each other, the parsing relationship edge is used to connect knowledge units that can explain the error pattern, and the inducing relationship edge is used to connect knowledge units that are prone to causing the error pattern.
[0126] Step S403: Label each relation edge with a weight value. The weight of the parsed relation edge is set based on the historical effectiveness of the knowledge unit in correcting the error, and the weight of the induced relation edge is set based on the association frequency between the knowledge unit and the error. Establish a mapping relationship between each knowledge point node, error pattern node and corresponding knowledge unit. Each knowledge unit is associated with at least one node and an attribute label is attached to realize the association binding of knowledge point node-error pattern node-knowledge unit-attribute feature.
[0127] Step S404: Introduce a dynamic update module for the knowledge graph. By collecting new knowledge points, relationships, knowledge units, and new student error patterns from teaching practice, the node system, relationship edge weights, and knowledge unit mappings are updated regularly.
[0128] The knowledge graph of this application achieves logic-driven association retrieval through a hierarchical system of core knowledge point nodes, subordinate child nodes, and error pattern nodes, combined with causal, progressive, and supplementary relationship edges and the parsing and inducing relationship edges of error pattern nodes: after locating the core node through entity recognition, the child nodes are traversed to obtain the associated knowledge units, and at the same time, the parsing units corresponding to common student errors are matched first based on the error pattern nodes.
[0129] Furthermore, this embodiment of the invention combines difficulty thresholds with student profiles to ensure that knowledge units not only focus on the subdivided content of the target knowledge points to avoid being too broad, but also match the student's current level and weaknesses, thus solving the problem of mismatch between knowledge units and student levels and lack of targeted errors in traditional retrieval.
[0130] Furthermore, such as Figure 5 As shown, in this embodiment, the steps of fusing and sorting the knowledge units using the Transformer model include:
[0131] Step S501: Obtain the weighted calculation-generated knowledge unit sorting sequence, which contains several knowledge units sorted according to their relevance strength and matching degree with the teaching objectives;
[0132] Step S502: Extract teaching style feature vectors, which are obtained by training based on the instructor's historical teaching content;
[0133] The step of extracting teaching style feature vectors includes: collecting teachers' historical teaching materials and classifying and labeling them; extracting text features of various materials using a pre-trained BERT model, and constructing a style feature library by combining the speech rate and pause features of the speech-to-text; clustering the style feature library using the K-means clustering algorithm to generate style labels; and dynamically updating the corresponding teaching style feature vectors based on the feature matching degree of the teacher's latest teaching materials.
[0134] Step S503: Convert each knowledge unit into a standardized feature vector, fuse it with the teaching style feature vector, input it into the self-attention layer of the Transformer encoder, and calculate the semantic association weight and style adaptation weight between units.
[0135] Step S504: Achieve multi-dimensional fusion optimization through multi-head attention parallel processing, wherein the multi-dimensional aspects include terminology consistency, logical coherence, and style consistency;
[0136] Step S505: Based on the fusion feature vector output by the encoder, the decoder generates basic text content that conforms to the teaching style. According to the presentation format identifier in the structured retrieval conditions, the multimodal conversion module is called to convert the basic text content into teaching content of the corresponding modality. The multimodality includes text, static illustrations, dynamic demonstration videos and audio narration.
[0137] Step S506: The final generated teaching content is multimodal structured content that integrates teaching styles.
[0138] This application generates a unique knowledge unit ranking sequence by weighted calculation of the matching degree between the knowledge graph relation edge weights and the teaching objectives. When the Transformer model is fused, the instructor's teaching style feature vector is first extracted. After fusing the knowledge unit features and style features, the consistency of terminology, logical coherence, and style uniformity are optimized in parallel through multi-head attention. Then, the multimodal transformation module generates content in the form of text, static illustrations, dynamic videos, etc., avoiding the problem of the relevance and teaching objectives being disconnected in traditional ranking. Furthermore, through style transfer and multimodal output, the generated content not only retains the core information of the knowledge unit and forms a coherent logical chain, but also adapts to the instructor's teaching habits and the student's learning preferences.
[0139] In addition, in the step of using the K-means clustering algorithm to cluster the style feature library and generate style labels in this application, firstly, quantifiable style features are extracted from historical teaching materials (such as lesson plans, courseware, videos, etc.), and the resulting feature vector includes text features, visual features, and interaction features. Among them, text features include term density, sentence complexity, and example type; visual features include image-text ratio, color preference, and layout style; and interaction features include question frequency and interaction form.
[0140] Further steps include standardizing the data and handling missing values. Then, the elbow rule and silhouette coefficient are used to determine the optimal number of clusters K. The sum of squared errors (SSE) for different K values (e.g., K=2-10) are calculated, and the K-SSE curve is plotted. The inflection point of the curve (where the slope drops sharply) is selected as the candidate K. Then, the silhouette coefficient of the candidate K is calculated (the value range is [-1,1], the closer it is to 1, the better the clustering quality). Finally, the optimal K is determined.
[0141] Then, clustering is performed using a determined K value, specifically including: randomly initializing K cluster centers; calculating the Euclidean distance from each sample to each cluster center and assigning it to the nearest center; recalculating the center of each cluster (mean of the feature vector); repeating the above steps until convergence (the center no longer changes or the change is less than the threshold).
[0142] To generate style labels for each cluster, the following method is used: compare the feature values of each cluster center and extract significant feature differences; then, formulate labels based on these feature differences.
[0143] According to another embodiment of this application, an online teaching content generation system based on a knowledge base is provided.
[0144] Please refer to Figure 6 The online teaching content generation system provided in this application includes the following modules:
[0145] The data conversion module 601 is used to respond to the teaching content generation request triggered by the teacher and convert the teaching content generation request into graph data.
[0146] The retrieval condition construction module 602 is used to perform semantic parsing on the graph data using a semantic parsing model based on CNN-BiLSTM. The graph data is converted into a fusion matrix by concatenating the adjacency matrix and the feature matrix through channel concatenation. Local correlation features between nodes are extracted by multi-scale convolution kernels of CNN. Then, the local feature sequence is globally semantically fused by the BiLSTM model to output structured retrieval conditions containing knowledge attribute labels, difficulty thresholds, and presentation format identifiers.
[0147] The associated retrieval module 603 is used to call a preset knowledge graph to perform associated retrieval based on the structured retrieval conditions; wherein, the knowledge graph includes knowledge point entity nodes and causal, progressive and supplementary relationship edges between nodes, the core knowledge point nodes are located by entity recognition, the subordinate child nodes of the core knowledge point nodes are traversed to obtain the corresponding associated knowledge units, and knowledge units that meet the characteristics of the student profile are selected according to the difficulty threshold to form a candidate knowledge set.
[0148] The content generation module 604 is used to perform weighted calculations based on the matching degree between the weight values of the knowledge graph relation edges and the teaching objectives, generating a unique sorted sequence of knowledge units; the sorted knowledge units are fused through the Transformer model, and semantic conflict detection is performed on the generated content and the expressions in the knowledge base simultaneously, and the credibility scores of conflicting knowledge units are marked to form teaching content.
[0149] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.
[0150] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.
[0151] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers.
[0152] Electronic devices can also refer to various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.
[0153] Electronic devices include a computing unit, which can perform various appropriate actions and processes based on a computer program stored in read-only memory or loaded from a storage unit into random access memory. RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.
[0154] Multiple components in the device are connected to the I / O interface, including: input units, such as keyboards and mice; output units, such as various types of displays and speakers; storage units, such as disks and optical discs; and communication units, such as network cards, modems, and wireless transceivers.
[0155] The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.
[0156] The computing unit can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities.
[0157] Examples of computing units include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any suitable processors, controllers, microcontrollers, etc.
[0158] The computing unit performs the various methods and processes described above, such as a knowledge-based online method for generating instructional content. For example, in some embodiments, a knowledge-based online method for generating instructional content may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit.
[0159] In some embodiments, part or all of the computer program may be loaded into and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the computing unit, one or more steps of the knowledge base-based online generation method for instructional content described above may be performed.
[0160] Alternatively, in other embodiments, the computing unit may be configured, by any other suitable means (e.g., by means of firmware), to perform a knowledge base-based online generation method for instructional content.
[0161] Various implementations of the systems and techniques described above in this document can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays, application-specific integrated circuits, application-specific standard products, systems-on-a-chip systems, payload programmable logic devices, computer hardware, firmware, software, and / or combinations thereof.
[0162] These various implementations may include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, and can receive data and instructions from a storage system, at least one input device and at least one output device, and transmit data and instructions to the storage system, the at least one input device and the at least one output device.
[0163] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.
[0164] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.
Claims
1. A method for online generation of teaching content based on a knowledge base, characterized in that, Includes the following steps: In response to a teaching content generation request triggered by an instructor, the teaching content generation request is converted into graph data. This includes: extracting multimodal elements from the teaching content generation request using word segmentation tools, speech recognition models, and handwritten feature extraction tools in natural language processing, obtaining entity elements, attribute elements, teaching scene feature elements, speech features, and handwritten annotation features; converting entity elements into entity nodes of graph data, and attribute elements into attribute nodes of graph data, establishing node connections through relational edges to obtain graph data; wherein, entity nodes have a unique code in the subject knowledge base, and attribute nodes are attached to at least one entity node; the weight values of relational edges are determined through a dynamic relational weight self-adjustment mechanism; wherein, the dynamic relational weight self-adjustment mechanism includes: constructing a multimodal feature vector from the extracted text scene feature elements, speech features, and handwritten annotation features, inputting it into a weight allocation model based on an attention mechanism, calculating the influence weight of each modality feature, and determining the dynamic weight values of relational edges based on the weighted fusion result; The graph data is semantically parsed using a semantic parsing model based on CNN-BiLSTM. The graph data is transformed into a fusion matrix by concatenating the adjacency matrix and the feature matrix through channel concatenation. Local correlation features between nodes are extracted by multi-scale convolution kernels of CNN. Then, the local feature sequence is globally semantically fused by the BiLSTM model to output structured retrieval conditions containing knowledge attribute labels, difficulty thresholds, and presentation format identifiers. Based on the structured retrieval conditions, a preset knowledge graph is invoked for associated retrieval. The knowledge graph includes knowledge point entity nodes and causal, progressive, and supplementary relationship edges between nodes. Core knowledge point nodes are located through entity recognition, and their subordinate child nodes are traversed to obtain corresponding associated knowledge units. Knowledge units matching the student profile characteristics are then selected based on a difficulty threshold to form a candidate knowledge set. The knowledge graph construction method includes the following steps: Based on the subject-domain knowledge system of the knowledge base, core knowledge points and their subordinate child knowledge points are extracted as entity nodes. Simultaneously, common student error patterns are extracted from historical student answer data as error pattern nodes, forming a hierarchical node system. Core knowledge point nodes correspond to key concepts within the subject, their subordinate child nodes contain detailed content of that knowledge point, and error pattern nodes are associated with knowledge point nodes prone to producing that error. Each node is assigned a unique identifier. For the logical relationships between nodes, causal relationship edges, progressive relationship edges, supplementary relationship edges, and error pattern nodes are constructed with knowledge units. The system includes analytical and induced relationship edges between elements. Causal relationship edges connect knowledge points with causal derivations, progressive relationship edges connect knowledge points with a learning sequence, supplementary relationship edges connect mutually complementary knowledge points, analytical relationship edges connect knowledge units that can explain the error pattern, and induced relationship edges connect knowledge units that are prone to causing the error pattern. Each relationship edge is labeled with a weight value. The weight of analytical relationship edges is set based on the historical effectiveness of the knowledge unit in correcting the error, while the weight of induced relationship edges is set based on the frequency of association between the knowledge unit and the error. A mapping relationship is established between each knowledge point node, error pattern node, and corresponding knowledge unit. Each knowledge unit is associated with at least one node and has an attribute label attached, achieving the association binding of knowledge point node-error pattern node-knowledge unit-attribute features. A dynamic update module for the knowledge graph is introduced, which periodically updates the node system, relationship edge weights, and knowledge unit mapping by collecting new knowledge points, relationships, knowledge units, and new student error patterns from teaching practice. The knowledge graph relation edge weights are weighted and calculated based on their matching degree with the teaching objectives to generate a unique sorted sequence of knowledge units. The sorted knowledge units are then integrated using a Transformer model, and semantic conflict detection is performed on the generated content and the expressions in the knowledge base. Conflicting knowledge units are labeled with credibility scores to form the teaching content.
2. The online method for generating teaching content based on a knowledge base according to claim 1, characterized in that, It also includes the following steps: During the preview of teaching content by students, dynamic evaluation data is collected through the real-time perception module to construct a temporary cognitive profile of students. The evaluation data includes the time students spend previewing teaching content, the key knowledge points they mark, and the questions they ask in real time. By integrating temporary profiles with students' historical learning data, outdated knowledge units are removed and the presentation format of the remaining units is adjusted to output personalized teaching content.
3. The online method for generating teaching content based on a knowledge base according to claim 2, characterized in that, The characteristic elements of a teaching scenario are used to characterize the type of teaching scenario; Speech features include speech rate and intonation variations, used to characterize the urgency of a request or the emphasis of a particular content; handwritten annotation features include underlines and asterisks, used to characterize core needs.
4. The online generation method for teaching content based on a knowledge base according to claim 3, characterized in that, The steps for extracting local correlation features between nodes using multi-scale convolutional kernels of CNNs include: Configure at least two different sizes of convolution kernels, including square convolution kernels and asymmetric convolution kernels. Each convolution kernel corresponds to a preset stride parameter and activation function, which are used to match node association patterns in different ranges in graph data, including symmetric association of close nodes and asymmetric association of mid-distance nodes. The semantic complexity analysis module calculates the association asymmetry index of graph data, which includes the ratio of the number of entity nodes to the number of attribute nodes and the difference in the length of indirect association paths. If the correlation asymmetry index exceeds a preset threshold, the asymmetric convolution kernel generation mechanism is activated. The weight parameters of the asymmetric convolution kernel are optimized through a generative adversarial network (GAN), and the asymmetric convolution kernel is used to adapt complex asymmetric correlation regions in the graph data. Convolutional kernels of different scales are applied to local regions of the fusion matrix, and node association features within the corresponding regions are extracted by a sliding window calculation to generate multiple sets of local feature mapping matrices. The local feature mapping matrices output by convolutional kernels of different scales are channel-concatenated to form an aggregated feature matrix containing multi-scale correlation information. The aggregated feature matrix is used as the input of the BiLSTM model.
5. The online generation method for teaching content based on a knowledge base according to claim 4, characterized in that, Define a fusion matrix of dimension N×N as X, and configure two convolutional kernels of different scales; where the small-scale square convolutional kernel is represented as... Step size is The weight matrix is Asymmetric large-scale convolution kernels are represented as follows: Step size is The weight matrix is Weight matrix It was obtained through training and optimization using a Generative Adversarial Network (GAN). For symmetric associations of nearby nodes, the calculation is performed by sliding a small-scale square convolution kernel, and is expressed as follows: ； For asymmetric associations of mid-distance nodes, the calculation is performed by sliding an asymmetric large-scale convolution kernel, and is expressed as follows: ； In the formula, Represents a small-scale feature mapping matrix. This represents the activation function. , The term represents the bias term, i and j represent the indexes of the feature map; p represents the row index of the convolution kernel and the local region of the fusion matrix, and q represents the column index of the convolution kernel and the local region of the fusion matrix. This represents the side length of the square convolution kernel. , These represent the number of rows and columns of the asymmetric convolution kernel, respectively. , This indicates the stride of the corresponding convolutional kernel; , This represents the weight matrix corresponding to the convolution kernel; The square kernel feature mapping matrix F1 and the asymmetric kernel feature mapping matrix F2 are concatenated along the channel dimension to form the aggregated feature F, which is represented as: ,in, This indicates a channel cascading operation.
6. The online generation method for teaching content based on a knowledge base according to claim 5, characterized in that, The steps involve using a BiLSTM model to perform global semantic fusion on local feature sequences, outputting structured retrieval conditions that include knowledge attribute labels, difficulty thresholds, and presentation format identifiers. The aggregated feature matrix output by the CNN is converted into a one-dimensional local feature sequence by rows or columns. Each sequence element corresponds to a local associated feature, and a positional code is attached to represent the position of the feature in the global sequence. Three parallel attention branches are introduced into the hidden layer of BiLSTM to generate knowledge attribute labels, difficulty thresholds, and presentation format identifiers, respectively. Each attention branch calculates the attention of the feature sequence through a weight matrix and fuses the attention weights with the hidden state of BiLSTM to generate a goal-oriented intermediate feature vector. After concatenating the intermediate feature vectors of the three branches, the concatenation is input into the fully connected layer, which initially outputs knowledge attribute labels, difficulty thresholds, and presentation format identifiers. A retrieval condition verification module is introduced, which constructs a feature sequence-valid retrieval condition mapping library based on historical generated data, calibrates the preliminary output, and finally outputs calibrated structured retrieval conditions.
7. The online method for generating teaching content based on a knowledge base according to claim 6, characterized in that, The steps involved in fusing the sorted knowledge units using the Transformer model include: Obtain a weighted sequence of knowledge units, which contains several knowledge units sorted by their relevance strength and their matching degree with the teaching objectives. Extract teaching style feature vectors, which are obtained by training based on the instructor's historical teaching content; Each knowledge unit is transformed into a standardized feature vector and fused with the teaching style feature vector. This is then input into the self-attention layer of the Transformer encoder to calculate the semantic association weights and style adaptation weights between units. Multi-dimensional fusion optimization is achieved through parallel processing of multi-head attention, whereby the multi-dimensional aspects include terminology consistency, logical coherence, and style consistency. Based on the fusion feature vector output by the encoder, the decoder generates basic text content that conforms to the teaching style. According to the presentation format identifier in the structured retrieval conditions, the multimodal conversion module is called to convert the basic text content into teaching content of the corresponding modality. The multimodality includes text, static illustrations, dynamic demonstration videos and audio narration. The final generated teaching content is a multimodal structured content that integrates teaching styles.
8. The online generation method for teaching content based on a knowledge base according to claim 7, characterized in that, The steps for extracting teaching style feature vectors include: Collect historical teaching materials from instructors and categorize and label them; Text features of various materials are extracted using a pre-trained BERT model, and a style feature library is constructed by combining the speech rate and pause features of the speech-to-text text. The K-means clustering algorithm is used to cluster the style feature library to generate style labels; Based on the feature matching degree of the instructor's latest teaching materials, the corresponding teaching style feature vector is dynamically updated.
9. A generation system for implementing the online generation method of knowledge base-based teaching content as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The data conversion module is used to respond to the teaching content generation request triggered by the instructor and convert the teaching content generation request into graph data; The retrieval condition construction module is used to perform semantic parsing on the graph data using a semantic parsing model based on CNN-BiLSTM. The graph data is transformed into a fusion matrix that is concatenated by concatenating the adjacency matrix and the feature matrix through a channel concatenation method. Local correlation features between nodes are extracted by multi-scale convolution kernels of CNN. Then, the local feature sequence is globally semantically fused by the BiLSTM model to output structured retrieval conditions containing knowledge attribute labels, difficulty thresholds, and presentation format identifiers. The associated retrieval module is used to call a preset knowledge graph to perform associated retrieval based on the structured retrieval conditions. The knowledge graph includes knowledge point entity nodes and causal, progressive, and supplementary relationship edges between nodes. The core knowledge point nodes are located by entity recognition, and the subordinate child nodes of the core knowledge point nodes are traversed to obtain the corresponding associated knowledge units. The knowledge units that meet the characteristics of the student profile are selected according to the difficulty threshold to form a candidate knowledge set. The content generation module is used to perform weighted calculations based on the matching degree between the edge weights of the knowledge graph and the teaching objectives, generating a unique sorted sequence of knowledge units. The sorted knowledge units are then fused using the Transformer model, and semantic conflict detection is performed on the generated content and the expressions in the knowledge base. Conflicting knowledge units are labeled with credibility scores to form the teaching content.