Intelligent digital textbook generation method and system based on generative artificial intelligence

Abstract

The application provides a wisdom digital teaching material generation method and system based on generative artificial intelligence, relates to the technical field of generative artificial intelligence, and first collects an initial subject knowledge corpus set and a preset cognitive gradient evolution template, performs latent semantic space mapping and knowledge atomization cutting processing on the corpus set, is reconstructed into a basic cognitive atomic unit, and is given an encoding mark and extracts a semantic feature vector. Then, a cognitive conflict network between the basic cognitive atomic units is constructed, a pre-trained model is called to perform text content deepening and cognitive guidance question generation processing driven by cross-concept thinking conflict, and a strengthened cognitive atomic unit set is obtained. Finally, dynamic cognitive path planning is performed according to the strengthened cognitive atomic unit set and the cognitive conflict network, a wisdom digital teaching material file with a nonlinear thinking transition path is generated, the personalized learning needs of students are met, and the learning effect is improved.

Description

Technical Field

[0001] This invention relates to the field of generative artificial intelligence technology, and more specifically, to a method and system for generating intelligent digital teaching materials based on generative artificial intelligence. Background Technology

[0002] Most existing digital textbooks are simply digitized versions of paper textbooks, lacking personalized design for different learning stages and cognitive levels. When integrating subject knowledge, teachers or textbook writers mainly rely on personal experience and subjective judgment, making it difficult to accurately grasp the order and depth of knowledge presentation, resulting in textbook content that fails to meet the diverse learning needs of students.

[0003] On the other hand, traditional digital textbooks tend to organize knowledge content linearly, lacking the exploration and utilization of cross-conceptual cognitive conflicts. Students struggle to deepen their understanding through comparison and interaction between different concepts, hindering the development of their critical thinking and innovative abilities. Furthermore, existing technologies fail to fully consider the characteristics of knowledge within its latent semantic space when processing subject knowledge, making it difficult to effectively extract and utilize semantic information, resulting in a lack of scientific rigor and systematic approach in textbook content generation. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for generating intelligent digital teaching materials based on generative artificial intelligence, the method comprising:

[0005] An initial set of subject knowledge corpus and a preset cognitive gradient evolution template are collected. The initial set of subject knowledge corpus includes multiple original academic literature units and multiple original teaching case units. The preset cognitive gradient evolution template includes a field defining the hierarchical relationship of concept cognition and a field defining cross-conceptual thinking conflict points.

[0006] The initial subject knowledge corpus is subjected to latent semantic space mapping and knowledge atomization cutting. Based on the concept cognitive hierarchy progression relationship definition field in the preset cognitive gradient evolution template, the original academic literature unit and the original teaching case unit are deconstructed and reconstructed into multiple basic cognitive atomic units. Each basic cognitive atomic unit is assigned a globally unique atomic code identifier. At the same time, the semantic density distribution feature vector of each basic cognitive atomic unit in the latent semantic space is extracted.

[0007] Based on the semantic density distribution feature vector and the cross-conceptual thinking conflict point definition field in the preset cognitive gradient evolution template, a cognitive conflict network between the basic cognitive atomic units is constructed. The cognitive conflict network includes a network topology with the atomic code identifier as the vertex and conflict edges connecting the vertices. The conflict edges have an initial value of conflict intensity calculated based on the semantic repulsion force between the semantic density distribution feature vectors.

[0008] The pre-trained concept generative adversarial neural network model is invoked. Based on the network topology in the cognitive conflict network and the initial value of the conflict intensity, the basic cognitive atomic units are subjected to cross-concept thinking conflict-driven text content deepening processing and cognitive guidance question generation processing to generate a set of enhanced cognitive atomic units containing deepened text content units and a set of associated cognitive guidance questions.

[0009] Dynamic cognitive path planning is performed based on the enhanced cognitive atomic unit set and the cognitive conflict network to generate a smart digital teaching material file with a non-linear thinking leap path. The smart digital teaching material file contains multiple teaching material cognitive module units adaptively arranged according to the concept cognitive hierarchy progression relationship defined fields. Each teaching material cognitive module unit contains a corresponding enhanced cognitive atomic unit and intra-module concept discrimination navigation links and inter-module thinking leap trigger links generated based on the cognitive conflict network.

[0010] Furthermore, embodiments of the present invention also provide a smart digital textbook generation system based on generative artificial intelligence, comprising:

[0011] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described method for generating intelligent digital teaching materials based on generative artificial intelligence by executing the machine-executable instructions.

[0012] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of the intelligent digital textbook generation system based on generative artificial intelligence reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the intelligent digital textbook generation system based on generative artificial intelligence to execute the above-described intelligent digital textbook generation method based on generative artificial intelligence.

[0013] Based on the above, by collecting an initial set of subject knowledge corpora and a pre-set cognitive gradient evolution template, the initial set of subject knowledge corpora undergoes latent semantic space mapping and knowledge atomization segmentation. This accurately deconstructs and reconstructs original academic literature units and original teaching case units into basic cognitive atomic units, assigning them unique atomic codes and extracting semantic density distribution feature vectors. Based on the semantic density distribution feature vectors and the cognitive gradient evolution template, a cognitive conflict network is constructed, which can present cross-conceptual thinking conflict relationships between basic cognitive atomic units. A pre-trained concept generative adversarial neural network model is invoked to drive cross-conceptual thinking conflict processing. The generated enhanced cognitive atomic unit set not only contains deepened text content but also is associated with a set of cognitive guidance questions, helping students to deeply understand knowledge and cultivate thinking skills. Finally, dynamic cognitive path planning is performed based on the enhanced cognitive atomic unit set and the cognitive conflict network. The generated intelligent digital textbook file has a non-linear thinking leap path, including adaptively arranged textbook cognitive module units, intra-module concept differentiation navigation links, and inter-module thinking leap trigger links. This better meets students' personalized learning needs, stimulates students' learning initiative and creativity, and significantly improves learning outcomes. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the execution flow of the intelligent digital textbook generation method based on generative artificial intelligence provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of exemplary hardware and software components of the intelligent digital textbook generation system based on generative artificial intelligence provided in an embodiment of the present invention. Detailed Implementation

[0016] Figure 1 This is a flowchart illustrating a method for generating intelligent digital teaching materials based on generative artificial intelligence, provided in one embodiment of the present invention. A detailed description follows.

[0017] Step S110: Collect an initial set of subject knowledge corpus and a preset cognitive gradient evolution template. The initial set of subject knowledge corpus includes multiple original academic literature units and multiple original teaching case units. The preset cognitive gradient evolution template includes a concept cognitive hierarchy progression relationship definition field and a cross-concept thinking conflict point definition field.

[0018] In this embodiment, the initial subject knowledge corpus was collected by connecting to multiple academic resource platforms through standardized API interfaces, covering multiple disciplines such as natural sciences and social sciences. The original academic literature units include peer-reviewed journal articles, conference proceedings, dissertations, and chapters of professional works. Each literature unit contains complete structured metadata, such as title, author information, publication date, keywords, abstract, and full text. The original teaching case units cover curriculum design schemes, teaching practice records, learning evaluation materials, and descriptions of typical teaching scenarios from basic education to higher education. Case units include elements such as teaching objectives, teaching processes, teaching resources, and feedback. All corpus units have undergone copyright compliance review to ensure compliance with Creative Commons agreements, and data acquisition is performed using encrypted transmission protocols. The pre-defined cognitive gradient evolution template uses an extensible markup language (EXPLAIN) format to define a structured framework. The concept cognitive hierarchy progression definition field describes the subject's conceptual system through a tree-like node structure. Each node includes a unique identifier for the concept, a cognitive depth coefficient, a list of preceding concepts, and a description of the concept's connotation. The cross-concept cognitive conflict point definition field uses a two-dimensional matrix structure to record the conflict type code, conflict concept pair identifiers, conflict description text, and conflict resolution path guidance. The template's development process involved subject matter experts and educational psychologists, and multiple rounds of revisions ensured that the cognitive hierarchy division and conflict type definition conformed to the subject's cognitive patterns.

[0019] Step S120: Perform latent semantic space mapping and knowledge atomization cutting on the initial subject knowledge corpus. According to the concept cognitive hierarchy progression relationship definition field in the preset cognitive gradient evolution template, deconstruct and reconstruct the original academic literature unit and the original teaching case unit into multiple basic cognitive atomic units, and assign a globally unique atomic code identifier to each basic cognitive atomic unit. At the same time, extract the semantic density distribution feature vector of each basic cognitive atomic unit in the latent semantic space.

[0020] In the latent semantic space mapping process, the initial corpus is first preprocessed: a rule-based and machine learning-based word segmentation method is used to segment the text into words, and a bidirectional maximum matching algorithm combined with the word segmentation model is used to improve the accuracy of professional terminology recognition; part-of-speech tagging is used to select nouns, verbs, adjectives, and adverbs, which have actual semantic value, while filtering out prepositions, conjunctions, and other function words; the TF-IDF weighted method is used to calculate the importance of words, retaining the top 80% of core words by weight. The pre-trained word vector model uses the CBOW architecture of Word2Vec, and is fine-tuned using subject-specific corpora on the basis of a general corpus, mapping words to a 300-dimensional high-dimensional vector space. By calculating the word co-occurrence matrix, singular value decomposition is used to retain 90% of the variance contribution, reducing the high-dimensional vector space to a 100-dimensional global latent semantic space. Knowledge atomization segmentation is based on defining fields according to the hierarchical relationship of conceptual cognition. A BiLSTM-CRF sequence labeling model is used to identify concept boundaries in the text. The model input is a sequence of word vectors and part-of-speech features, and the output is concept entity labels. A dependency parser is combined to extract the subject-verb-object core structure, determining the boundaries of the smallest semantic units. Original documents and cases are decomposed into independent basic cognitive atomic units, each containing three elements: the conceptual subject, attribute description, and relational relationship. Atomic encoding identifiers use a cryptographic hash algorithm based on SHA-3-256. A 256-bit hash calculation is performed on the text content of the atomic unit, and the first 16 bits are truncated and converted into a 32-bit hexadecimal string. The encoding format is "subject domain code-cognitive hierarchy code-hash value," ensuring global uniqueness. The process of extracting semantic density distribution feature vectors is as follows: the atomic unit text is segmented into word-level semantic units, which are then converted into 300-dimensional vectors through a pre-trained word vector model; a 10×10×10 three-dimensional grid is divided in the semantic space, the number of vector landing points in each grid is calculated, and the density scalar value is obtained by dividing by the total number of vectors, forming a 1000-dimensional density distribution vector. The vector is then subjected to L2 normalization so that the sum of the values ​​of each dimension is 1.

[0021] Step S121: Perform text noise filtering and unified character encoding conversion on each original academic document unit and each original teaching case unit in the initial subject knowledge corpus to obtain a clean text corpus unit set, which contains continuous semantically clean character stream data.

[0022] The text noise removal process employs a three-level filtering mechanism. The first level removes markers from the text using regular expression matching, including various tags, citation markers, and special control characters. Pattern matching is implemented using Python's `re` module, with a matching rule library containing over 200 common noise patterns. The second level uses rule-based redundancy identification, detecting and deleting duplicate paragraphs through text fingerprint comparison and identifying highly similar redundant sentences using cosine similarity. The third level filters low-quality text through language model scoring, using the GPT-2 model to calculate text fluency scores and filtering abnormal text segments with scores below 0.6. Character encoding conversion adopts the UTF-8 standard, supporting batch conversion of multiple encoding formats. Automatic encoding detection is performed using the `chardet` library. During conversion, a Unicode character mapping table is used to handle rare characters and special symbols; for characters that cannot be mapped, U+FFFD replacement markers are used. The processed clean text corpus units maintain paragraph continuity according to the original document structure. Each unit is stored as a UTF-8 encoded txt file, with the file header containing metadata blocks: the original source URL, processing timestamp, data checksum, and preprocessing step log, ensuring data traceability. The collection of clean text corpus units is organized using a directory tree structure and stored in a three-level classification system by subject area, document type, and year, which facilitates subsequent batch processing.

[0023] Step S122: Call the pre-trained deep semantic coding network to perform global latent semantic space embedding processing on the set of clean text corpus units, map each clean text corpus unit to an initial semantic distribution point in a high-dimensional space, and generate an initial semantic coordinate vector corresponding to each initial semantic distribution point.

[0024] The pre-trained deep semantic encoding network adopts a Transformer-based BERT-Base architecture, containing a 12-layer Transformer encoder with a hidden layer dimension of 768, a multi-head attention mechanism with 12 heads, and a feedforward network dimension of 3072. After pre-training on a general corpus, the model is fine-tuned using subject-specific corpora. The fine-tuning process uses a learning rate of 2e-5, 5 training epochs, and a batch size of 32. Input processing: Clean text corpus units are segmented into text fragments no longer than 512 characters. Input sequences are constructed by adding special tags, and the text is converted into a vocabulary index using WordPiece segmentation, with a vocabulary size of 30522. The embedding layer includes word embeddings, position embeddings, and segment embeddings. These are element-wise summed and then processed by layer normalization and dropout. Encoding process: Each Transformer encoder layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by residual connections and layer normalization. The self-attention mechanism employs scaled dot product attention, calculating the product of the query, key, and value matrices, and obtaining the attention weights through the softmax function. The network output uses the hidden states corresponding to special labels as the initial semantic vectors of text units, which are then normalized by a LayerNorm layer to obtain 768-dimensional initial semantic coordinate vectors. Vector storage uses 32-bit floating-point format, and an approximate nearest neighbor index is constructed using the FAISS library to support fast retrieval of semantically similar units, with a retrieval precision set to 0.9.

[0025] Step S123: Parse the concept cognitive hierarchy progression relationship definition field in the preset cognitive gradient evolution template, obtain the multiple core discipline concept tags contained in the concept cognitive hierarchy progression relationship definition field and the cognitive depth level value corresponding to each core discipline concept tag, as well as the description text of multiple basic cognitive constituent elements under each core discipline concept tag.

[0026] The lxml library is used to parse the XML-formatted cognitive gradient evolution template, constructing a DOM tree structure to traverse concept hierarchy nodes. Core subject concept tags are extracted from the name attribute of concept nodes, using the naming convention "subject code_concept term," with each concept tag corresponding to a unique identifier attribute. Cognitive depth level values ​​are stored in the depth attribute, using an integer scale of 0-5, determined through expert evaluation combined with a concept dependency algorithm. The descriptive text of basic cognitive components is located in the text attribute of the element's child nodes; each core concept contains 3-15 elements, with dependencies defined between elements through preceding nodes. Three levels of validation are performed during parsing: syntax validation ensures the XML structure conforms to the schema definition, semantic validation verifies the uniqueness of concept tags and the acyclicity of hierarchical relationships, and integrity validation ensures all element description texts are not empty. The parsing results are stored as a nested dictionary structure: the top-level key is the core concept tag, and the values ​​include three sub-items: depth, element list, and preceding concept list. Each element in the element list contains an element ID, descriptive text, and importance weight. The parsing results are persisted to the local file system in JSON format, and a concept relationship graph is generated for visualization and conflict analysis.

[0027] Step S124: Perform density-based spatial clustering analysis on the initial semantic distribution points, identify the local maxima regions of semantic density in the high-dimensional space, define each local maxima region of semantic density as a potential cognitive atom nucleus, and calculate the semantic membership probability value of each initial semantic distribution point to all potential cognitive atom nuclei.

[0028] Density-based spatial clustering analysis employs the DBSCAN algorithm, with core parameters including neighborhood radius and minimum sample size. The neighborhood radius is determined using the K-distance method: the distance from each point to its 5th nearest neighbor is calculated, a K-distance curve is plotted, and the distance at the inflection point of the curve is taken as the neighborhood radius. The minimum sample size is set between 5 and 15 depending on the data size. The algorithm execution flow is as follows: Unlabeled points are randomly selected, and the number of points within a neighborhood centered at that point with a neighborhood radius equal to the radius is queried. If the number meets the minimum sample size requirement, it is marked as a core point, and the cluster is expanded to form a cluster; otherwise, it is marked as a noise point. This process is repeated until all points have been processed. After clustering, the semantic density distribution is calculated for each cluster: the kernel density of all initial semantic vectors within the cluster is estimated in a high-dimensional space using a Gaussian kernel function, and the bandwidth is selected using cross-validation. Local maxima of density are the potential cognitive nuclei, determined by finding the peak points of the density function using the gradient ascent method. The semantic membership probability values ​​are calculated using a Gaussian mixture model: It is assumed that the data consists of a mixture of multiple Gaussian distributions, each corresponding to an atomic nucleus; model parameters are estimated using the EM algorithm; the posterior probability of each sample point belonging to each distribution is calculated, resulting in a membership probability vector, where the sum of the vector elements is 1. The probability values ​​are normalized using the Softmax function to ensure that the maximum value corresponds to the most likely atomic nucleus.

[0029] Step S125: Based on the semantic membership probability value of each initial semantic distribution point, each semantically complete sentence unit or paragraph unit in the pure text corpus unit set is assigned to the basic cognitive constituent element description text corresponding to the potential cognitive atom nucleus with the highest membership, forming a cognitive atom prototype unit with the basic cognitive constituent element description text as the semantic aggregation center.

[0030] First, semantic unit segmentation is performed: a hybrid approach based on punctuation and syntactic analysis is adopted. Sentence segmentation uses the Punkt segmenter of NLTK to identify terminal punctuation; long paragraphs are split into sub-paragraphs through deep analysis of syntactic dependency trees, ensuring that each sub-paragraph contains a complete predicate structure. For each sentence or paragraph unit, a semantic vector is generated through a deep semantic encoding network, and its membership probability value to all potential cognitive nuclei is calculated. The nucleus with the highest probability value is selected as the belonging object, and the descriptive text of the basic cognitive constituent elements associated with that nucleus is queried. Element matching uses semantic similarity verification: the cosine similarity between the unit text and the element description text is calculated. If the similarity is less than 0.5, a manual review process is triggered. The attribution results are stored in a list structure, with each entry containing unit ID, original text, nucleus ID, element ID, membership probability, and similarity score. Cognitive atom prototype units are grouped by element ID, and the set of units under each element forms a prototype unit, containing a unit list and aggregated statistical information.

[0031] Step S126: Perform semantic redundancy elimination and logical order reordering on multiple sentence units or paragraph units belonging to the same basic cognitive component description text. Based on the original contextual relationship of the sentence units or paragraph units in the pure text corpus unit set and their semantic coherence scores, merge them to generate basic cognitive atomic units with independent and complete cognitive meanings.

[0032] Semantic redundancy elimination employs a two-step method: first, calculating the cosine similarity of semantic vectors among all units under the same element to construct a similarity matrix; then, applying a greedy algorithm, starting with the unit pair with the highest similarity, retaining units with higher information entropy and deleting redundant units. Logical order rearrangement constructs a directed weighted graph: nodes are sentence or paragraph units, and edge weights are semantic coherence scores, obtained by calculating context prediction probabilities using a language model. An improved PageRank algorithm is used to calculate node importance, combined with topological sorting to determine the optimal order: calculating the in-degree and out-degree of each node to identify the starting node; sorting by node importance, prioritizing high-importance nodes; and finding the longest path through dynamic programming to ensure the integrity of the logical chain. The merging process uses the T5 model for text fusion: the sorted unit list is used as input, and the model generates fused text through instruction prompts. After fusion, manual verification is triggered by the following condition: when the ROUGE-L score of the fused text and the original unit is lower than 0.7, it is marked as requiring manual review. The final generated basic cognitive atomic unit includes an atomic ID, the element to which it belongs, the fused text, a list of source unit IDs, and a creation timestamp.

[0033] Step S127: Generate a globally unique atomic code identifier for each basic cognitive atomic unit based on its semantic content hash calculation. Establish an immutable mapping relationship between the atomic code identifier and the core subject concept label and cognitive depth level value to which the basic cognitive atomic unit belongs.

[0034] The atomic code identifier generation uses the SHA-3-256 hash algorithm: the text content field of the basic cognitive atomic unit is encoded in UTF-8, the hash value is calculated, and the first 16 bytes are truncated and converted into a 32-bit hexadecimal string. The encoding structure is "XX-Y-ZZZZZZZZZZZZZZZZZ", where XX is the subject area code, Y is the cognitive depth level value, and ZZZZ is the first 15 digits of the hash value. The mapping relationship is stored on the blockchain: using the Hyperledger Fabric consortium blockchain, each code corresponds to a transaction record, including atomic ID, core concept tag, cognitive depth value, hash value, and timestamp. Smart contracts implement the functions of adding and querying mapping relationships. Adding requires multi-node consensus to ensure immutability. A local code-concept mapping cache is maintained, with a cache expiration time set to 24 hours to improve query efficiency. After code generation, conflict detection is performed: the blockchain is queried to ensure that the code has not been used. If a conflict is found, a random number is added to the hash value for retry, with a maximum of 5 attempts.

[0035] Step S128: Extract all semantic vector representations contained in each basic cognitive atom unit, perform spatial grid partitioning on all semantic vector representations, count the density of semantic vector landing points in each spatial grid, and generate a semantic density distribution feature vector that can characterize the coverage and distribution pattern of the basic cognitive atom unit in the latent semantic space. The semantic density distribution feature vector contains the density scalar value corresponding to each spatial grid.

[0036] First, the text of basic cognitive atomic units is segmented into word-level semantic units, and core words are retained after filtering out stop words. Each word is converted into a 300-dimensional vector using a pre-trained word vector model, resulting in a vector list. Spatial grid partitioning: A 10×10×10 three-dimensional grid is created in the 300-dimensional semantic space, with the grid side length determined by dividing the feature value range by 10. For each vector, its grid index in each dimension is calculated. The number of vectors within each grid is counted, and the density scalar value is calculated. The generated semantic density distribution feature vector has a dimension of 1000, with each element corresponding to the density scalar value of a grid. The vectors are L1 normalized. To handle sparsity, grids with density values ​​below 1e-5 are zeroed out and stored in a sparse vector format. The feature vectors are stored in association with atomic encoding identifiers for subsequent cognitive conflict network construction.

[0037] Step S130: Based on the semantic density distribution feature vector and the cross-conceptual thinking conflict point definition field in the preset cognitive gradient evolution template, construct a cognitive conflict network between the basic cognitive atomic units. The cognitive conflict network includes a network topology with the atomic code identifier as the vertex and conflict edges connecting the vertices. The conflict edges have an initial value of conflict intensity calculated based on the semantic repulsion force between the semantic density distribution feature vectors.

[0038] The cognitive conflict network is constructed using an undirected weighted graph model. The vertex set consists of atomic codes representing all basic cognitive units, and the edge set represents the conflict relationships between concepts. The network construction process is as follows: Initialize the vertex set; traverse all basic cognitive units, adding atomic codes to the vertex set. Vertex attributes include concept labels, depth values, and semantic density vectors. For all vertex pairs, calculate the semantic repulsion force. Determine the conflict type based on the cross-conceptual conflict point definition field, and then calculate the initial conflict intensity value. If the intensity value exceeds 0.3, add conflict edges. The network is stored using an adjacency list format, with each vertex associated with an edge list. Edge attributes include the target vertex, conflict type, and intensity value. Semantic repulsion force is calculated using Euclidean distance: the distance is the square root of the sum of the squares of the differences in the grid density values ​​corresponding to the two semantic density distribution feature vectors. Conflict type determination: Query the concept label using the atomic code to retrieve the list of conflicting concept pairs. If a match exists, obtain the conflict type code; otherwise, determine it as a potential conflict based on semantic similarity. The initial conflict intensity value is transformed using a non-linear mapping function to ensure the intensity value is within the range of 0-1. After the network is built, connectivity analysis is performed, isolated vertices are deleted, and weakly connected components are manually reviewed.

[0039] Step S131: Obtain the pre-defined multiple conflict type identifiers in the cross-conceptual thinking conflict point definition field. The conflict type identifiers include conceptual connotation opposition conflict type, theoretical paradigm incommensurable conflict type, and empirical conclusion contradiction conflict type. Configure the corresponding conflict intensity nonlinear mapping function parameters for each conflict type identifier.

[0040] Conflict type identifiers are extracted from the conflict type nodes of the cognitive gradient evolution template: conceptual connotation opposition conflict type, theoretical paradigm incommensurable conflict type, and empirical conclusion contradiction conflict type. Each type is configured with independent nonlinear mapping function parameters. The function adopts the general form: intensity value = a / (1+exp(-b*(distance-c)))+d, where a, b, c, and d are type-specific parameters. The parameter determination process involves collecting conflict intensity scores from over 500 pairs of concept pairs from subject matter experts; normalizing the scores to 0-1 as the target intensity value; and fitting the function parameters using the least squares method with distance as input. Parameters for conceptual connotation opposition conflict type: a=0.95, b=2.5, c=0.4, d=0.05; parameters for theoretical paradigm incommensurable conflict type: a=0.9, b=3.0, c=0.5, d=0.1; parameters for empirical conclusion contradiction conflict type: a=0.85, b=2.0, c=0.35, d=0.15. The parameters are stored in a JSON configuration file and can be dynamically adjusted through template updates. After adjustment, the strength values ​​of the affected conflict edges need to be recalculated.

[0041] Step S132: Using the atomic code identifier of each basic cognitive atomic unit as a network vertex, construct an initial blank network vertex set, and add all the atomic code identifiers of all basic cognitive atomic units to the network vertex set to form a complete vertex set.

[0042] The vertex set is implemented using a hash table data structure, with the key being the atomic code identifier and the value being the vertex object. Each vertex object contains attributes: atomic code, core concept label, cognitive depth value, semantic density, inbound edge list, and outbound edge list. The construction process is as follows: initialize an empty hash table; traverse all basic cognitive atomic units to obtain the atomic code; check if the code already exists in the hash table; if so, record a conflict log and trigger atomic code regeneration; create vertex objects and populate attribute values; add the vertex objects to the hash table. Vertex set integrity verification: check if the number of vertices matches the number of basic cognitive atomic units; if not, check for encoding conflicts or omissions. To improve retrieval efficiency, auxiliary indexes are constructed: mappings from concept labels to the vertex list and from cognitive depth values ​​to the vertex list. These indexes are implemented using Redis sorted sets and support range queries.

[0043] Step S133: Traverse every two different atomic coding identifiers in the set of network vertices, obtain the semantic density distribution feature vectors corresponding to the two different atomic coding identifiers, calculate the absolute value of the density difference of the two semantic density distribution feature vectors per spatial grid, and obtain the local density difference value on each spatial grid.

[0044] A double-loop traversal of the vertex set is employed: the outer loop iterates through vertices from the first to the second-to-last, while the inner loop iterates through vertices from the next vertex to the last, ensuring that each unordered pair is processed only once. For each vertex pair: semantic density distribution feature vectors are obtained from the vertex set through atomic encoding, with the vector format being a sparse dictionary; the grid indices of the two vectors are merged to obtain the union of all non-zero grids; for each grid, the local density difference is calculated as the absolute value of the difference between the density values ​​of the two vectors at that grid; the difference values ​​are stored in a temporary array, with the array index being the grid. To optimize computational efficiency, vectorization operations are used: sparse vectors are converted into dense arrays, and the element-wise difference is calculated using NumPy's `abs` function to obtain a dense difference array. For large-scale vertex sets, a block-based processing strategy is adopted: the vertex set is divided into multiple blocks, and the difference values ​​of vertex pairs within and between blocks are calculated using multi-threaded parallel computation. The difference value calculation results are stored as triples, containing two vertex codes and the difference array.

[0045] Step S134: Perform spatial integration and accumulation on the local density difference values ​​to obtain the overall semantic distribution deviation parameter of the two basic cognitive atomic units in the latent semantic space, and use the overall semantic distribution deviation parameter as the preliminary calculation result of semantic repulsion force.

[0046] The spatial integral accumulation employs a weighted summation method: the overall semantic distribution deviation parameter is the sum of the products of the local density differences of all grids and their corresponding weights, where the weights represent the importance of the grid in the semantic space. The weights are determined based on the variance contribution of the grids in principal component analysis: the variance contribution of the semantic density distribution eigenvector is calculated through principal component analysis, and the top 10 principal components are selected; the projection variance of each grid in the principal component space is calculated, with larger variances resulting in higher weights; the weights are normalized so that the sum of all weights is 1. The integral calculation process involves: multiplying the difference array and the weight array element-wise; summing the products to obtain the deviation parameter; and standardizing the parameter by dividing it by the maximum possible distance to ensure the parameter falls within the 0-1 range. To reduce computational load, grids with weights below 1e-4 are ignored. The calculation results are rounded to four decimal places as preliminary results for semantic repulsion.

[0047] Step S135: Based on the core subject concept label to which the basic cognitive atomic unit corresponding to the two different atomic code identifiers belongs in the concept cognitive hierarchy progression relationship definition field and the cognitive depth level value corresponding to the core subject concept label, determine the actual conflict type between the two basic cognitive atomic units. The actual conflict type is one of the conceptual connotation opposition conflict type, the theoretical paradigm incommensurable conflict type, or the empirical conclusion contradiction conflict type.

[0048] The core subject concept tags and cognitive depth level values ​​of vertex objects are queried using atomic encoding identifiers. A list of conflicting concept pairs is retrieved from the cross-conceptual conflict point definition field. This list contains multiple conflicting concept pairs and their corresponding conflict type codes. The conflict type determination process is as follows: Check if a combination of two concept tags exists in the list of conflicting concept pairs; if it exists, the type attribute is directly obtained as the actual conflict type; if it does not exist, calculate the semantic similarity of the two concept tags. If the similarity is less than 0.3, it is determined to be a conflict type of opposing conceptual connotations; if the similarity is between 0.3 and 0.7, and the difference in cognitive depth values ​​is greater than 2, it is determined to be a conflict type of incommensurable theoretical paradigms; otherwise, it is determined to be a conflict type of contradictory empirical conclusions. For the determination results, the decision basis is recorded and stored in the decision basis attribute of the conflict edge. When two atomic units belong to the same core concept tag, no conflict edge is created.

[0049] Step S136: Obtain the nonlinear mapping function parameters of the conflict intensity corresponding to the actual conflict type, input the preliminary calculation result of the semantic repulsion force into the nonlinear mapping function of the conflict intensity for mapping transformation processing, and obtain the initial value of the conflict intensity between the two basic cognitive atomic units.

[0050] Based on the determined actual conflict type, the corresponding nonlinear mapping function parameters are read from the configuration file. The preliminary calculation results of semantic repulsion are substituted into the function for calculation. For example, for the conceptual connotation opposition conflict type with a distance of 0.5: Intensity value = 0.95 / (1+exp(-2.5*(0.5-0.4)))+0.05=0.95 / (1+exp(-0.25))+0.05≈0.95 / 1.284+0.05≈0.739+0.05=0.789. The calculation result is rounded to three decimal places; if the result is greater than 1, it is truncated to 1; if it is less than 0, it is set to 0. Intensity values ​​are filtered by a threshold: conflict edges are created only when the intensity value is greater than 0.3; conflict relationships below the threshold are considered weak conflicts and are not added to the network.

[0051] Step S137: In the set of network vertices, create a conflict edge between the two different atomic code identifiers. The conflict edge connects the two different atomic code identifiers, and stores the calculated initial value of the conflict intensity as the weight attribute of the conflict edge. At the same time, the actual conflict type is also associated and stored as the type label attribute of the conflict edge.

[0052] Conflicting edges are represented using objects, containing attributes such as: source vertex code, target vertex code, initial conflict strength value, conflict type label, conflict type decision criteria, and semantic repulsion force value. In an undirected graph, both directions of an edge are considered as the same edge, and uniqueness is ensured by sorting by code size. Edges are stored using an adjacency list: the edge object is added to both the outgoing edge list of the source vertex and the incoming edge list of the target vertex. Edge attributes support dynamic updates: when the semantic density distribution feature vector is updated, the distance and strength can be recalculated, and the corresponding edge attributes can be updated accordingly. To facilitate conflict strength analysis, each edge also records a calculation timestamp to track the timeliness of the strength value.

[0053] Step S138: After all the different atomic code pairs in the set of network vertices have been processed, a cognitive conflict network containing all vertices and all conflicting edges is generated.

[0054] After processing all vertex pairs, a network integrity check is performed: the number of vertices matches the number of basic cognitive atomic units; the number of edges meets expectations; and the attributes of each conflicting edge are complete. The network is stored in GraphML format, containing graph nodes and edge child nodes, with vertex and edge attributes stored as data child nodes. A network statistics report is also generated, showing the number of vertices, edges, average degree, maximum degree, conflict type distribution, and intensity value distribution. The network files are managed using version control, with a new version number generated after each update, and historical versions retained for backtracking. After the cognitive conflict network is constructed, it is visualized using Gephi software, employing a force-directed layout to display vertices and edges. Vertex size is proportional to degree, edge thickness is proportional to intensity value, and conflict types are distinguished by edge color.

[0055] Step S140: Invoke the pre-trained concept generative adversarial neural network model, and perform cross-concept thinking conflict-driven text content deepening processing and cognitive guidance question generation processing on the basic cognitive atomic units according to the network topology in the cognitive conflict network and the initial value of the conflict intensity, to generate a set of enhanced cognitive atomic units containing the deepened text content units and the set of associated cognitive guidance questions.

[0056] The pre-trained conceptual generative adversarial neural network model employs a generator-discriminator architecture. The generator, based on the Transformer-XL architecture, contains a 6-layer encoder and a 6-layer decoder, with 1024 hidden layers and 16 attention heads, using relative positional encoding. The discriminator uses a hybrid CNN+Transformer architecture, containing 3 convolutional layers and 2 Transformer encoder layers, outputting authenticity probability and conflict consistency scores. Model pre-training is performed on a large-scale subject-specific corpus, using the AdamW optimizer with a learning rate of 5e-5, 100 training epochs, and a batch size of 64. The fine-tuning phase uses a dataset containing cognitive conflict annotations. The input samples are basic cognitive atomic unit text and conflict network information; the output is deepened text and cognitive guidance questions. The text content deepening process is as follows: input is basic cognitive atomic unit text and local conflict topological context; the generator captures conflict relationships through a self-attention mechanism, generating deepened text containing critical analysis; the discriminator scores the authenticity and conflict consistency of the generated text; and the generator is jointly optimized using adversarial loss and content loss. The cognitive-guided question generation employs a template-based and generation-based approach: identifying cognitive tension points from the deepening text, selecting question templates based on conflict types, filling in concept names, and adding intensity modifiers. The set of reinforced cognitive atomic units is organized by atomic coding identifiers; each unit contains the deepening text, a set of questions, a list of conflict sources, and a generation confidence score.

[0057] Step S141: From the cognitive conflict network, extract all adjacent vertices directly connected by conflict edges to each vertex corresponding to the basic cognitive atomic unit, as well as the initial value of the conflict intensity and the conflict type label of the conflict edge connected to each adjacent vertex, and construct a set of local conflict topology context information with the current basic cognitive atomic unit as the core vertex.

[0058] For each vertex of a basic cognitive atomic unit: obtain all conflicting edges of that vertex from the adjacency list of the cognitive conflict network; for each edge, extract the atomic encoding identifier, initial conflict strength value, and conflict type label of the adjacent vertices; collect the semantic density distribution feature vector, core concept label, and cognitive depth value of the basic cognitive atomic unit corresponding to the adjacent vertices; construct a local conflict topological context information set, with the data structure being a dictionary containing lists of core vertices and adjacent vertices. Each element in the adjacent vertex list contains a vertex identifier, strength, type, semantic density, concept label, and depth value. To control the size of the context, adjacent vertices are sorted in descending order of strength, retaining the top 10 strongly conflicting vertices, and only recording the count of the remaining weakly conflicting vertices. The context information set is stored in association with the atomic encoding identifier and serves as one of the inputs to the model.

[0059] Step S142: Perform joint embedding encoding on the semantic density distribution feature vector of the basic cognitive atomic unit and the set of local conflict topological context information to generate a conflict-aware input tensor containing the semantic information of the core vertex and the distribution of its surrounding conflict field.

[0060] The joint embedding encoding process consists of five steps: adjacent vertex feature sorting and normalization, edge attribute vectorization encoding, adjacent vertex feature enhancement, conflict field fusion, and tensor construction. First, adjacent vertex features are sorted by conflict intensity and truncated to a fixed length. Then, conflict intensity and type are encoded into edge attribute vectors. Next, adjacent vertex features and edge attributes are concatenated. Finally, a graph attention network is used to fuse core vertices and adjacent vertex features, ultimately generating a conflict-aware input tensor with dimensions [1, 1024, sequence length]. The tensor is stored in Float32 format and normalized using a LayerNorm layer to ensure numerical stability.

[0061] Step S1421: Extract the semantic density distribution feature vectors of the basic cognitive atomic units corresponding to all adjacent vertices from the local conflict topological context information set, and sort all the semantic density distribution feature vectors of adjacent vertices in descending order of their corresponding initial conflict strength values ​​to obtain the sorted adjacent vertex feature vector sequence.

[0062] Extract the semantic density distribution feature vector and initial conflict strength value of each adjacent vertex from the list of adjacent vertices in the local conflict topological context information set. Create a list of tuples and sort them in descending order of initial conflict strength value using a sorting function. The sorted result retains only the semantic density distribution feature vector, forming a sequence of adjacent vertex feature vectors. If the number of adjacent vertices is 0, the sequence is empty; if the number is more than 15, the first 15 vectors are truncated; if there are fewer than 15, all vectors are retained. The sorting process ensures that the features of strongly conflicting vertices are considered first.

[0063] Step S1422: Perform a fixed maximum length truncation or zero-padding operation on the sorted adjacent vertex feature vector sequence to make the length of the adjacent vertex feature vector sequence uniformly equal to the preset maximum number of adjacent vertices, thus obtaining a length-normalized adjacent vertex feature vector matrix.

[0064] The maximum number of adjacent vertices is preset to 15. For the sorted sequence of adjacent vertex feature vectors: if the sequence length is greater than 15, the first 15 vectors are truncated, and the remaining vectors are discarded; if the length is less than 15, a zero vector is added to the end of the sequence; if the length is equal to 15, it remains unchanged. The processed sequence is converted into an adjacent vertex feature vector matrix of dimension [15, 1000]. Adding zero vectors ensures that the input matrix dimensions of different core vertices are consistent, facilitating batch processing.

[0065] Step S1423: Obtain the initial value of the conflict intensity and the conflict type label of the conflict edge connecting the core vertex and each adjacent vertex. Vectorize the initial value of the conflict intensity and the conflict type label to generate an edge attribute feature vector corresponding to each adjacent vertex. The edge attribute feature vector contains a normalized initial value of the conflict intensity and a binary encoded component representing the conflict type label.

[0066] For each adjacent vertex: obtain the initial conflict intensity value and normalize it; obtain the conflict type label and perform one-hot encoding; the edge attribute feature vector is formed by concatenating the normalized initial conflict intensity value and the one-hot encoded conflict type label, with a total dimension of 4. For example, the edge attribute vector with an intensity of 0.789 and a type of conceptual connotation opposition conflict is [0.789, 1, 0, 0]. If the adjacent vertex is a virtual vertex added with zero padding, then the edge attribute vector is [0, 0, 0, 0].

[0067] Step S1424: The length-normalized adjacent vertex feature vector matrix and the corresponding edge attribute feature vector are concatenated along the feature channel dimension to obtain an enhanced adjacent vertex feature representation containing adjacent vertex semantic information and connecting edge attributes.

[0068] The length-normalized adjacent vertex feature vector matrix has dimensions [15, 1000], and the edge attribute feature vector has dimensions [15, 4]. Concatenation is performed along the feature channel dimension: the enhanced adjacent vertex feature representation has dimensions [15, 1004]. The concatenation operation is implemented using NumPy's `concatenate` function, with `axis=1`. The enhanced feature representation simultaneously includes the semantic density distribution information of adjacent vertices and the attribute information of conflicting edges.

[0069] Step S1425: Perform conflict field fusion processing on the semantic density distribution feature vector of the core vertex and the enhanced adjacent vertex feature representation. The conflict field fusion processing adopts a graph attention network mechanism, using the semantic density distribution feature vector of the core vertex as the attention query benchmark, and the enhanced adjacent vertex feature representation as the attention key and attention value. Calculate the attention coefficient of each adjacent vertex relative to the core vertex, and perform a weighted summation on the enhanced adjacent vertex feature representation to obtain a conflict-aware input tensor that aggregates adjacent vertex information and its conflict intensity.

[0070] The conflict field fusion employs a single-head attention mechanism within a graph attention network: the semantic density distribution feature vector of the core vertex serves as the query vector, while the feature representations of enhanced adjacent vertices serve as the key and value vectors; the attention score is calculated as the product of the query vector and the key vector, processed by an activation function; the attention coefficients are normalized using a softmax function; the aggregated adjacent features are the weighted sum of the attention coefficients and the value vectors; the core vertex vector is concatenated with the aggregated features, and a conflict-aware feature vector is obtained through linear transformation and an activation function; this feature vector is then expanded into a tensor form and used as the model's input. This attention mechanism enables the model to adaptively focus on important adjacent conflicting vertices, enhancing its conflict perception capability.

[0071] Step S143: Input the conflict-aware input tensor into the generator network part of the pre-trained concept generative adversarial neural network model, and perform deep conflict feature extraction and fusion processing through multiple residual convolution modules and self-attention mechanism modules of the generator network part to generate a hidden layer feature map with conflict-driven semantic enhancement.

[0072] The generator network consists of an alternating structure of three residual convolutional modules and three self-attention mechanism modules. The residual convolutional modules pass the input tensor through two 3×3 convolutional layers, each followed by BatchNorm and ReLU; skip connections add the input to the convolutional output. The self-attention mechanism modules include multi-head self-attention, residual connections and normalization, and a feedforward network with residual connections and normalization. The deep conflict feature extraction process involves passing the conflict-aware input tensor first through an embedding layer, then sequentially through the residual convolutional modules and the self-attention modules, with the output feature map dimension remaining unchanged for each module. The final generated hidden feature map contains a deep fusion of core vertex semantics and adjacency conflict features.

[0073] Step S144: The discriminator network of the pre-trained concept generative adversarial neural network model is invoked to perform authenticity and conflict consistency discrimination on the hidden layer feature map. The parameters of the generator network are adjusted in reverse according to the loss gradient returned by the discriminator network. Based on the parameter-adjusted generator network, the deepened text content units for the basic cognitive atomic units are constructed segment by segment in an autoregressive generation manner. The deepened text content units, while fully retaining the original cognitive kernel of the basic cognitive atomic units, introduce critical analysis and integrative thinking guidance content on the related conflict atomic units in the local conflict topological context information set.

[0074] The discriminator network takes as input the concatenated vector of hidden layer feature maps and generated text, and outputs a authenticity score and a conflict consistency score. The loss function is a weighted sum of the standard GAN loss and the conflict consistency loss. Parameter tuning uses backpropagation: calculating the discriminator's score for the generated text; calculating the gradient of the loss function with respect to the generator parameters; and updating the generator parameters using the Adam optimizer. Text generation employs an autoregressive approach: starting with the initial token, predicting the next token each time until the final token is generated. Beam search is used during the generation process to improve text quality. The deepened text content units retain the core concept definitions and basic principles of the original atomic units, add a conflict identification section to analyze the conceptual differences between related conflicting atomic units, and add thought-provoking paragraphs.

[0075] Example Implementation Section:

[0076] Step S145: Analyze each semantic paragraph in the deepened text content unit, identify the cognitive tension points in the semantic paragraph, and the cognitive tension points are manifested as the positions of sentences in the deepened text content unit that simultaneously mention two core subject concept tags that are connected by conflicting edges.

[0077] When identifying cognitive tension points in the deepened text content units, the text is first divided into independent paragraphs according to semantic logic. Then, core concepts are located through terminology extraction. Finally, the co-occurrence positions of conflicting concept pairs are determined using a cognitive conflict network. The specific process is implemented through four levels: semantic paragraph division, concept extraction, conflict pair matching, and coordinate localization, ensuring accurate capture of implicit conceptual conflict clues within the text. Each semantic paragraph unit serves as an independent analysis unit. Core disciplinary concept tags are extracted using natural language processing techniques and compared with conflict relationships in the cognitive conflict network. When a pair of concepts connected by conflicting edges is detected appearing simultaneously, it is marked as a cognitive tension point, and its precise location in the text is recorded. For example, in a text discussing different theoretical systems, if two concept tags with conflicting theoretical paradigms appear simultaneously, the system will automatically identify this location as a cognitive tension point and prepare for generating targeted questions.

[0078] Step S1451: Perform semantic paragraph segmentation on the deepened text content unit to obtain multiple consecutive semantic paragraph units, each semantic paragraph unit containing a set of sentence sequences that express relatively complete semantics.

[0079] Semantic paragraph segmentation is achieved using a hybrid strategy: First, explicit paragraph delimiters are identified through a rule engine, such as consecutive line breaks, indentation markers, and heading numbers (e.g., "1.", "(1)", etc.), which usually correspond to natural segmentation in the text structure. For long texts without explicit markers, a BERT-based boundary prediction model is used for auxiliary segmentation. After fine-tuning on disciplinary corpora, this boundary prediction model can determine paragraph boundaries by analyzing the semantic coherence probability between sentences. Specifically, when implementing, after splitting the text into a sequence of sentences, the boundary probability of adjacent sentence pairs is calculated, and when the probability exceeds 0.85, it is determined as a paragraph boundary. Each semantic paragraph unit after segmentation contains a list of sentences, the starting character offset (relative to the full text), and the ending character offset. For example, a certain paragraph contains 3 sentences, starting from the 120th character of the full text and ending at the 380th character. To ensure the semantic integrity of paragraphs, post-processing verification is also required: paragraphs longer than 500 characters are split again, and paragraphs containing multiple topics are subject to topic clustering through the LDA topic model to ensure that each paragraph focuses on a single topic. The segmentation results are stored as a list of paragraph objects, and each object contains attributes such as id, text, start_pos, and end_pos.

[0080] Step S1452: Perform named entity recognition processing and term extraction processing on each semantic paragraph unit, and extract all core disciplinary concept labels that appear in the semantic paragraph unit and the exact character position offsets of each core disciplinary concept label in the semantic paragraph unit.

[0081] The BERT-CRF named entity recognition model is used for concept extraction. This BERT-CRF named entity recognition model is trained on a disciplinary corpus containing labeled samples and can identify concept entities (with the label CONCEPT) in the text, outputting the text content of the concept and its relative position within the paragraph. At the same time, exact matching is performed through a domain term library, which contains all core disciplinary concept labels extracted from the cognitive gradient evolution template, and the AC automata algorithm is used to achieve efficient matching. For the identified concepts, standardization processing is required: synonyms are merged (e.g., "electromagnetic field" and "electromagnetic field" are unified into a standard term), and concept normalization is performed through WordNet. The calculation of position offsets needs to be converted to the absolute coordinates of the full text: relative starting position within the paragraph + starting position of the paragraph = starting position of the full text, relative ending position within the paragraph + starting position of the paragraph = ending position of the full text. For example, if a certain concept starts at the 20th character within a paragraph and the paragraph starts at the 120th character of the full text, then its starting position in the full text is 140. The extraction results are stored as a list of concept objects, and each object contains term (term text), start (starting position of the full text), end (ending position of the full text), confidence (recognition confidence). Concepts with a confidence lower than 0.7 need to be manually reviewed and confirmed.

[0082] Step S1453: Traverse all extracted core subject concept tags. For each core subject concept tag, search for a list of other core subject concept tags that are connected to it by conflicting edges in the cognitive conflict network.

[0083] To accelerate queries, a concept conflict index table is constructed: the conflict edge relationships in the cognitive conflict network are converted into a dictionary structure, with the key being the core concept label and the value being a list of conflicting concepts containing conflict type and intensity value. For each extracted core concept label c, the conflicting concept list `conflict_list` is obtained by querying the index table, and the list is sorted in descending order of conflict intensity. For example, the conflicting list of concept c might contain [(c1, 0.85, CT01), (c2, 0.72, CT02), ...], where the tuple elements are the conflicting concept label, intensity value, and conflict type, respectively. If the concept does not exist in the index table, it is supplemented by semantic similarity calculation: the word vector of concept c is compared with the word vectors of all concepts in the network using cosine similarity, and concepts with a similarity lower than 0.3 are taken as potential conflicting concepts, added to `conflict_list`, and marked as "potential conflict". Query results are cached in a Redis database for 24 hours to improve the efficiency of repeated queries. For concept tags containing polysemous words, disambiguation needs to be performed in conjunction with the context. The semantic vector of the concept in the current paragraph is calculated by the BERT context encoder, and then compared with the concept vector in the conflict network. The most similar concept is selected as the target conflict concept.

[0084] Step S1454: For the current semantic paragraph unit, construct a set of core subject concept tags that appear in the corresponding paragraph, traverse each pair of tags in the set of core subject concept tags, and determine whether the pair of tags exists in the conflict edge connection relationship of the cognitive conflict network.

[0085] Construct a set of paragraph concepts \(S = \{c_1, c_2, \ldots, c_n\}\), and use a combination algorithm to generate all unordered tag pairs \((c_i, c_j)\) where \(i < j\) to avoid duplicates. For each tag pair, determine the conflict relationship through the following steps: First, check whether \((c_i, c_j)\) exists in the conflict edge set of the cognitive conflict network, which can be achieved by querying the conflict index table constructed in step S1453; if it does not exist, then check whether \((c_j, c_i)\) exists (symmetry of the undirected graph); if it still does not exist, calculate the semantic similarity of the two concepts, and when the similarity is less than 0.3 and both are sub - concepts of the same super - concept, it is determined as a potential conflict pair. To improve efficiency, use a matrix to store the determination results of conflict pairs: construct an \(n\times n\) boolean matrix \(M\), where \(M[i][j] = True\) indicates that there is a conflict relationship between \((c_i, c_j)\). For a paragraph containing \(n\) concepts, \(n(n - 1) / 2\) determinations need to be made, and vectorized operations are used for parallel processing. The determination results are stored in an associated manner with information such as the conflict intensity and type of the concept pair.

[0086] Step S1455: If this pair of tag combinations exists in the conflict edge connection relationship of the cognitive conflict network, it is determined that there is a cognitive tension point in the semantic paragraph unit, and the name of this pair of tag combinations and the exact character position offsets of each of this pair of tag combinations in the semantic paragraph unit are recorded.

[0087] When the tag pair \((c_i, c_j)\) is determined to have a conflict relationship, create a cognitive tension point object with the following attributes: id (unique identifier), concept_pair \((c_i, c_j)\), positions \([(s_i, e_i), (s_j, e_j)]\), where \(s_i / e_i\) is the start / end position of \(c_i\), and \(s_j / e_j\) is the start / end position of \(c_j\), conflict_type (conflict type), intensity (conflict intensity). The position information needs to be accurate to the character level. For example, \(c_i\) is located at characters 140 - 155 in the full text, and \(c_j\) is located at characters 180 - 195. For multiple co - occurrences of the same concept pair within a paragraph, cognitive tension point objects need to be created separately and numbered in the order of appearance. If there is a nested relationship between two concepts in the text (such as one concept is included in the terms of another concept), only the position of the outermost concept is recorded. Concept pairs with a distance exceeding 500 characters are excluded during the recording process to avoid reduced relevance due to excessive intervals. All cognitive tension point objects are stored in a paragraph - level list and associated with the paragraph ID.

[0088] Step S1456: Based on the offset of the end character position of the first core subject concept tag and the offset of the start character position of the second core subject concept tag in the tag pair, calculate the core text span in the semantic paragraph unit that mentions both concepts simultaneously, and use the start and end positions of the core text span as the positioning coordinates of the cognitive tension point in the deepened text content unit.

[0089] The core text span calculation follows these rules: the starting position is the earlier starting position of the two concepts (min(s_i, s_j)), and the ending position is the later ending position (max(e_i, e_j)), forming the coordinate interval [start, end]. For example, if c_i starts and ends at [140, 155], and c_j starts and ends at [180, 195], then the core text span is [140, 195], covering the complete text area from c_i to c_j. If the two concepts overlap (e.g., c_i [140, 160], c_j [150, 170]), then the span is [140, 170]. The positioning coordinates are stored as tuples (start, end) and associated with cognitive tension point objects. For easy visualization, the relative position of this span within the paragraph (the offset relative to the paragraph start) and the coordinate mapping relationship during page rendering also need to be calculated. Location coordinates are used to highlight the text area containing conflicting concept pairs when the teaching materials are presented, and to associate cognitive guidance questions in the vicinity of that area.

[0090] Step S146: For each identified cognitive tension point, based on the two core subject concept labels mentioned at the cognitive tension point and their corresponding initial conflict intensity values, generate a cognitive guidance question to guide learners to think about the relationship between the two. The cognitive guidance question is presented in the form of a question and points to the conflict edge of the two core subject concept labels in the cognitive conflict network.

[0091] Cognitive-guided question generation is a crucial step in connecting cognitive conflict networks with learning interactions. By combining conflict type, intensity, and contextual information, targeted, thought-provoking questions are generated. The specific process includes five steps: conflict information extraction, template matching, concept filling, intensity modification, and positional association. This ensures that the questions not only reflect the essence of the conflict between concepts but also guide learners to engage in deeper thinking. Question generation should avoid simple yes-or-no questions, employing guiding phrases such as "how," "why," and "how to understand" to promote learners' metacognition of conflict relationships. Each question corresponds one-to-one with a cognitive tension point and is dynamically bound to the text content through location coordinates, achieving a three-in-one association of "text-conflict-question."

[0092] Step S1461: Obtain the first core subject concept label and the second core subject concept label mentioned at the cognitive tension point, and extract the initial value of the conflict intensity and the conflict type label of the conflict edge connecting the first core subject concept label and the second core subject concept label from the cognitive conflict network.

[0093] The `concept_pair` attribute is extracted from the cognitive tension point object to obtain the first core concept label `c_a` and the second core concept label `c_b`. The adjacency list of the cognitive conflict network is used to query the conflicting edge connecting `c_a` and `c_b`, obtaining the `intensity` (initial conflict intensity value) and `type` (conflict type label) from the edge attributes. If multiple conflicting edges exist (such as concept conflicts at different levels), the edge with the highest intensity value is selected as the primary conflicting edge. For example, `c_a` and `c_b` may have two conflicting edges, CT01 (intensity 0.85) and CT03 (intensity 0.62). In this case, CT01 is selected as the conflict type of the current cognitive tension point. The extracted results are added to the `conflict_details` attribute of the cognitive tension point object, containing information such as `intensity`, `type`, and `edge_id`.

[0094] Step S1462: Based on the conflict type tag, retrieve the interrogative sentence template that matches the conflict type tag from the preset question template library. The interrogative sentence template includes a first concept placeholder for accommodating the first core subject concept tag and a second concept placeholder for accommodating the second core subject concept tag.

[0095] The pre-set question template library is stored categorized by conflict type, with each type containing 5-8 different sentence templates. For example, templates for CT01 (conceptual connotation opposition) include: "How to understand the contradictory relationship between {concept A} and {concept B} in terms of connotation?", "Analyze the core differences between {concept A} and {concept B}?", and "What impact does the opposition between {concept A} and {concept B} have on the development of the discipline?". {concept A} and {concept B} in the templates are placeholders for the first and second concepts, identified by curly braces. Template selection employs a two-layer strategy: first, candidate templates are matched based on conflict type labels; then, the semantic similarity between the template and the current paragraph text is calculated using the BERT model, selecting the template with the highest similarity. For example, if a paragraph contains descriptions related to "applicable conditions," templates containing terms such as "scope of application" and "preconditions" are prioritized. The template library supports dynamic expansion; teachers can add new templates and label them with applicable conflict types through the management interface.

[0096] Step S1463: Fill the first core subject concept tag's complete name text into the first concept placeholder position of the question template, and fill the second core subject concept tag's complete name text into the second concept placeholder position of the question template to generate a preliminarily formatted question text.

[0097] Concept filling is achieved using a string replacement mechanism: The template text is traversed, and placeholders for {Concept A} and {Concept B} are identified and replaced with the full names c_a and c_b, respectively. The name text must use standardized terminology, such as expanding "Newton's Second Law of Motion" to "Newton's Second Law of Motion". During the filling process, cases where concept names contain special characters, such as parentheses and quotation marks, must be handled to ensure the grammatical correctness of the replaced statements. For example, the template "What are the differences between the theoretical foundations of {Concept A} and {Concept B}?" might become "What are the differences between the theoretical foundations of quantum mechanics and classical mechanics?" after filling. The initially formatted question text undergoes grammatical validation, using NLTK's syntax checker to detect issues such as subject-verb agreement and article usage, and automatically corrects simple grammatical errors.

[0098] Step S1464: Based on the magnitude of the initial value of the conflict intensity, select a degree adverb from the preset intensity modifier lexicon that corresponds to the numerical range of the initial value of the conflict intensity, insert the degree adverb into the initially formatted question text to modify the position of the relationship strength between the first core subject concept label and the second core subject concept label, and generate a cognitive guidance question with intensity awareness.

[0099] The intensity modifier library is divided into three levels based on the conflict intensity value range: high intensity range (0.7-1.0) corresponds to adverbs such as "significant," "high," and "fundamental"; medium intensity range (0.4-0.7) corresponds to adverbs such as "to a certain extent," "obvious," and "partial"; and low intensity range (0.3-0.4) corresponds to adverbs such as "slight," "potential," and "possible." The corresponding adverb is selected based on the initial conflict intensity value and inserted into the position describing the relationship in the question text. For example, the question with intensity 0.85 (high intensity) "What are the differences between the theoretical foundations of quantum mechanics and classical mechanics?" can be modified by inserting "fundamental" to become "What are the fundamental differences between the theoretical foundations of quantum mechanics and classical mechanics?". The insertion position is determined by the template. <modifier>Tags specify, such as template "{Concept A} and {Concept B}". <modifier>In what aspects do the differences lie? <modifier>The marker specifies the adverb insertion position. If no marker is provided in the template, it is inserted by default at the first adjective position after the concept pair.

[0100] Step S1465: Associate and store the generated cognitive guidance question with the positioning coordinates of the cognitive tension point in the deepened text content unit so that the cognitive guidance question can be displayed near the corresponding position in the deepened text content unit during final presentation.

[0101] Create a question-location association table. Each record contains `question_id` (unique question identifier), `question_text` (question text), `conflict_intensity` (conflict intensity), `conflict_type` (conflict type), `position` (location coordinates (start, end)), and `paragraph_id` (paragraph ID). The association storage uses JSON format and is stored in the same file as the refined text content units, using `position` coordinates for fast location. During textbook rendering, the question display position is calculated based on the location coordinates: typically, a question icon floats to the right of the text span; hovering the mouse reveals the full question text; clicking it brings up an interactive answer interface. To support accessibility, ARIA tags need to be added to the questions to describe the relationship between the question and the text. The association storage also needs to record metadata such as question generation time and generated model version for quality traceability and model optimization.

[0102] Step S147: Sort and combine all the generated cognitive guidance questions according to their order of appearance in the deepened text content units to form a set of related cognitive guidance questions.

[0103] For all cognitive guidance questions within the same enhanced text content unit, they are sorted in ascending order by their location coordinates' start values ​​to ensure the question order matches the text reading order. During sorting, if the location coordinates of two questions overlap (the overlapping area exceeds 50% of the smaller span), they are merged into a question group and sorted in descending order of conflict intensity. For example, questions Q1 (span [140, 195]) and Q2 (span [160, 210]) overlap and are merged into question group [Q1, Q2] (assuming Q1 has higher intensity). The data structure of the associated cognitive guidance question set includes: questions (sorted list of questions / list of question groups), total_count (total number of questions), conflict_type_distribution (conflict type distribution statistics), and avg_intensity (average conflict intensity). The set also needs to contain metadata for each question, such as the generated confidence score (based on discriminator scoring) and associated conflict edge ID. Finally, it is stored in JSON array format as one of the attributes of the enhanced cognitive atomic unit.

[0104] Step S148: Encapsulate and combine the deepened text content unit and the set of related cognitive guidance questions to generate reinforced cognitive atomic units corresponding to the basic cognitive atomic units. All the reinforced cognitive atomic units corresponding to the basic cognitive atomic units together constitute a set of reinforced cognitive atomic units.

[0105] The enhanced cognitive atomic units are encapsulated in JSON-LD format, with core attributes including: @id (atomic code identifier), originalAtomicId (original basic cognitive atomic unit ID), deepenedText (deepened text content unit, including a list of paragraphs and location information), guidanceQuestions (a set of related cognitive guidance questions), conflictSources (a list of conflict source vertices, including conflict atomic unit codes and strengths), generationConfidence (generation confidence score, ranging from 0 to 1, determined by the average score output by the discriminator), and timestamp (generation timestamp). Integrity verification is required during the encapsulation process: checking whether the deepened text retains the original cognitive kernel (calculated through text similarity with basic cognitive atomic units, threshold > 0.8), and whether the question set corresponds one-to-one with cognitive tension points. The enhanced cognitive atomic unit set is organized into a directory structure according to subject area and cognitive depth level, such as "domain / phys / hierarchy / l1 / atomic_units.jsonl", with each file containing multiple JSON-LD records for enhanced cognitive atomic units. After the set is generated, the quality is evaluated by sampling review (randomly selecting 5% of the units), with a focus on checking the accuracy of the text, the relevance of the problem, and the correctness of the conflict description. If the pass rate is less than 90%, the model will be retrained or the parameters will be adjusted.

[0106] Step S150: Perform dynamic cognitive path planning processing based on the enhanced cognitive atomic unit set and the cognitive conflict network to generate a smart digital teaching material file with a non-linear thinking leap path. The smart digital teaching material file contains multiple teaching material cognitive module units adaptively arranged according to the concept cognitive hierarchy progression relationship defined fields. Each teaching material cognitive module unit contains a corresponding enhanced cognitive atomic unit and intra-module concept discrimination navigation links and inter-module thinking leap trigger links generated based on the cognitive conflict network.

[0107] By integrating and reinforcing cognitive atomic units and cognitive conflict network information, a multi-dimensional learning path network is constructed. This process first divides the learning into modules based on cognitive depth levels, then guides the paths within each module based on conflict intensity, and finally establishes cross-level transition links, forming a two-dimensional navigation structure of "hierarchical progression + conflict-driven." The teaching materials adopt the EPUB3.3 standard format, supporting rich media content and interactive navigation, enabling learners to learn in a traditional linear order while also autonomously jumping between modules based on the intensity of cognitive conflict, achieving personalized exploratory learning.

[0108] Step S151: Parse the concept cognitive hierarchy progression relationship definition field in the preset cognitive gradient evolution template, and obtain the cognitive depth hierarchy division structure defined in the concept cognitive hierarchy progression relationship definition field. The cognitive depth hierarchy division structure includes basic cognitive level identifiers, advanced cognitive level identifiers and cutting-edge cognitive level identifiers, as well as a list of core subject concept labels expected to be covered under each cognitive depth hierarchy identifier.

[0109] Parsing the cognitive gradient evolution template using XPath <hierarchy>The nodes are used to extract the cognitive depth hierarchy structure: basic cognitive level (labeled L1, depth value 0.0-1.5), intermediate cognitive level (labeled L2, depth value 1.6-3.5), and advanced cognitive level (labeled L3, depth value 3.6-5.0). Each level label corresponds to... <level>Child nodes, containing<expected_concepts> This list records the core subject concept tags that each level should cover. For example, L1 includes concepts such as "basic definitions" and "basic principles," while L3 includes concepts such as "cutting-edge theories" and "interdisciplinary applications." The level division also includes... <prerequisite>Child nodes define the dependencies between levels, such as L2 requiring mastery of 80% of the concepts in L1. The parsed results are stored as a list of hierarchical objects, each containing attributes such as id (hierarchical identifier), depth_range (depth value range), expected_concepts (expected concept list), prerequisites (prerequisites list), and learning_hours (suggested learning time).

[0110] Step S152: Traverse each enhanced cognitive atom unit in the set of enhanced cognitive atom units, and based on the atomic code identifier of the basic cognitive atom unit corresponding to each enhanced cognitive atom unit, trace back the core subject concept label and cognitive depth level value of its mapping association, and then determine the cognitive depth level identifier to which each enhanced cognitive atom unit should belong.

[0111] For each enhanced cognitive atomic unit, it is associated with the basic cognitive atomic unit through `originalAtomicId`. Then, the cognitive depth level value (the Y part in the encoding structure) is extracted from the atomic code identifier of the basic cognitive atomic unit. The hierarchical identifier is determined based on the range of the value (e.g., 1.2 belongs to L1, 3.0 belongs to L2). For example, in the atomic code "PH-2-XXXXXX", "2" represents a depth value of 2.0, belonging to L2. The attribution process verifies whether the core concept label of the unit is in the expected_concepts list of the hierarchy. If not, it is marked as "pending classification" and manual review is triggered. The attribution result is recorded in the hierarchy attribute of the enhanced cognitive atomic unit, including `level_id` (hierarchical identifier), `depth_value` (depth value), and `concept_matched` (concept matching status). Finally, all units are grouped according to the hierarchical identifier, forming three unit lists: L1_units, L2_units, and L3_units.

[0112] Step S153: According to the cognitive depth hierarchy division structure, perform internal path planning processing on all enhanced cognitive atomic units belonging to the same cognitive depth hierarchy identifier. The internal path planning processing determines the nonlinear recommended exploration order of enhanced cognitive atomic units within the same cognitive depth hierarchy based on the initial value of the conflict intensity and the conflict type label of the conflict edges between the vertices corresponding to these enhanced cognitive atomic units in the cognitive conflict network, and generates the concept discrimination navigation link within the module. The concept discrimination navigation link within the module contains the link address pointing to other enhanced cognitive atomic units within the same cognitive depth hierarchy that have strong conflict edges connected to the current enhanced cognitive atomic unit.

[0113] Intra-module path planning is key to achieving hierarchical nonlinear learning, guiding the exploration of connections between concepts through conflict networks. First, a hierarchical induced subgraph is constructed, edge weights are calculated, an exploration sequence is generated using a greedy algorithm, and finally, navigation links pointing to strongly conflicting units are generated for each unit, forming a mesh-like internal navigation structure.

[0114] For example, step S1531: Obtain the induced subgraph in the cognitive conflict network of all vertices corresponding to the enhanced cognitive atomic units belonging to the same cognitive depth level identifier, the induced subgraph containing these vertices and all conflict edges between these vertices.

[0115] The process begins by selecting a set of vertices, V_level, belonging to the target level from the cognitive conflict network (determined by the level identifier encoded in atomic codes). Then, all conflicting edges E_level between these vertices are extracted to form an induced subgraph G_level = (V_level, E_level). For example, the induced subgraph for the L2 level contains all vertices corresponding to L2 enhanced cognitive atomic units and their conflicting edges. The induced subgraph is stored using an adjacency list, where each vertex is associated with its list of conflicting edges within the level. Edge attributes retain information such as conflict strength and type. To reduce computational complexity, weak conflicting edges with a strength below 0.3 are filtered out, retaining only strong conflict relationships. The induced subgraph also requires calculation of basic topological indices, such as average degree and the number of connected components. If an isolated vertex (degree 0) exists, it is added to the nearest connected component.

[0116] Step S1532: In the induced subgraph, a path planning weight value is calculated for each conflict edge. The path planning weight value is positively correlated with the initial value of the conflict intensity of the conflict edge and is obtained by multiplying it by the preset type weight coefficient corresponding to the conflict type label of the conflict edge.

[0117] The preset type weight coefficients reflect the navigation priority of different conflict types: conceptual contradiction (CT01) has a coefficient of 1.2, theoretical incommensurability (CT02) has a coefficient of 1.5, and empirical contradiction (CT03) has a coefficient of 1.0. The formula for calculating the path planning weight value is: weight(e) = intensity(e) * type_coeff(e.type), where intensity(e) is the initial value of conflict intensity, and type_coeff is the type coefficient. For example, the edge weight with intensity 0.75 and type CT02 is 0.75 * 1.5 = 1.125. The weight value is rounded to three decimal places to measure the importance of the edge to the path planning. After calculation, the weight of each edge is normalized (divided by the maximum weight value) so that the weight is in the range of 0-1, which facilitates comparison between different levels.

[0118] Step S1533: Using all vertices in the induced subgraph as traversal nodes and the path planning weight value as the attraction index for the transfer between nodes, a greedy traversal algorithm based on maximizing conflict intensity is adopted. Starting from a preset starting vertex, at each step, the vertex pointed to by the edge with the largest path planning weight value among the conflict edges connected to the current vertex and not visited is selected as the next visited vertex, until all vertices have been visited once, thereby generating a traversal sequence as the nonlinear recommended exploration order.

[0119] The default starting vertex is the vertex corresponding to the core concept within the level, specified by domain experts based on the subject knowledge structure (e.g., the starting vertex for level L1 is the "basic definition" concept unit). The greedy traversal algorithm is as follows: Initialize the visit set visited={start_vertex}, and the exploration sequence sequence=[start_vertex]; the current vertex current=start_vertex; when the size of visited is less than the size of V_level, execute: get all unvisited adjacent vertices of current, calculate the edge weights corresponding to these vertices, select the vertex next_vertex with the largest weight and add it to visited and sequence, update current=next_vertex; repeat until all vertices have been visited. If there are edges with the same weight, prioritize the vertex with the lower cognitive depth value (to promote exploration from basic to complex). The generated non-linear recommended exploration order is dynamic and needs to be recalculated when the cognitive conflict network is updated.

[0120] Step S1534: According to the nonlinear recommended exploration order, for each enhanced cognitive atom unit in the sequence, extract all conflict edges between it and other enhanced cognitive atom units in the sequence in the inducible subgraph, and generate a link for each conflict edge pointing to the enhanced cognitive atom unit corresponding to the other end of the conflict edge, as part of the concept discrimination navigation link within the module.

[0121] For each unit u in the sequence, find all its conflicting edges in the induced subgraph G_level, and filter out edges whose other endpoint v is also in the sequence. Generate a concept discrimination navigation link within the module for each eligible edge. The link attributes include: target_id (target unit atomic code), conflict_intensity (conflict intensity), conflict_type (conflict type), display_text (display text, formatted as "discrimination: {v.concept_label}"), and sort_order (sorting priority, in descending order of conflict intensity). The link list for each unit is limited to a maximum of 5 links to avoid information overload. Links are embedded in the unit text as HTML hyperlinks, typically marked when the concept first appears, and clicking them navigates to the target unit's text location. Link styles are differentiated based on conflict type (e.g., red for CT01, blue for CT02), and intensity is indicated by the font weight (higher intensity means bolder font).

[0122] Step S154: For each enhanced cognitive atom unit, analyze the conflict edges in the cognitive conflict network that connect the vertex of the enhanced cognitive atom unit to vertices in other cognitive depth levels and whose initial conflict intensity exceeds a preset conflict threshold. Extract the atomic encoding identifiers of the target enhanced cognitive atom units in other cognitive depth levels pointed to by these conflict edges, and generate inter-module thought transition trigger links from the current enhanced cognitive atom unit to the target enhanced cognitive atom unit.

[0123] A default conflict threshold of 0.6 (adjustable via configuration file) is used to filter cross-level conflict relationships with sufficient intensity. For each enhanced cognitive atom unit u: traverse all conflict edges in its cognitive conflict network, and filter edges that meet the following conditions: the level identifier of the target vertex v is different from u, and the edge strength > threshold. For each selected edge, extract the atom code identifier of v and its level, and generate inter-module thought leap trigger links. Attributes include: target_id (target unit atom code), target_level (target level identifier), conflict_intensity (conflict intensity), trigger_text (trigger text, formatted as "leap to {target_level}: {v.concept_label}"), and recommended_prerequisite (recommended prerequisite knowledge, a list of core concepts in the level where v is located). Links are sorted by target level (prioritizing the same level → higher level → lower level), and within the same level, they are sorted in descending order of intensity. The trigger link appears as a group of floating buttons on the right side of the textbook interface. When the mouse hovers over it, the conflict intensity and recommended prerequisite knowledge are displayed. Clicking it will perform a cross-level jump and update the navigation path.

[0124] Step S155: Create a corresponding textbook cognitive module unit for each cognitive depth level identifier. All reinforced cognitive atomic units belonging to the cognitive depth level identifier are sequentially used as internal sub-modules of the textbook cognitive module unit according to the non-linear recommended exploration order. The concept discrimination navigation links within the module and the thinking leap trigger links between modules are embedded as dynamic interactive metadata into the data structure of each reinforced cognitive atomic unit.

[0125] The cognitive module unit is the basic building block of the smart digital textbook, with each level corresponding to one module. The module data structure includes: module_id (module identifier, such as "M_L1_PHYS"), level_id (cognitive depth level identifier), title (module title, such as "Basic Cognitive Module"), learning_objectives (list of learning objectives, extracted from the cognitive gradient evolution template), submodules (list of internal submodules, i.e., reinforced cognitive atomic units arranged in a non-linear recommended exploration order), and interactive_metadata (dynamic interactive metadata, containing indices of links within / between modules). The dynamic interactive metadata is stored in JSON format and linked to the link list through unit atomic codes, for example: {"atomic_code":"PH-L2-XXXX", "intra_links":[...], "inter_links":[...]}. The module unit also includes navigation configuration information, such as the default number of links displayed and the conflict intensity visualization threshold. The submodule arrangement strictly follows the non-linear recommended exploration order, while also allowing learners to adjust the sorting to alphabetical order or publication time through settings.

[0126] Step S156: All created textbook cognitive module units are non-linearly associated and combined according to the progressive relationship of the basic cognitive level identifier, the advanced cognitive level identifier, and the cutting-edge cognitive level identifier to generate a smart digital textbook file with a network-like but non-linear structure. The text interaction area inside each enhanced cognitive atom unit in the smart digital textbook file is dynamically bound to the concept discrimination navigation link within the module and the thinking leap trigger link between the modules, so as to realize the learner's exploratory learning of autonomous thinking leap based on the intensity of cognitive conflict during the learning process.

[0127] The non-linear relationships between modules are achieved through a master-controlled interactive graph, ultimately generating a smart digital teaching material file that supports dynamic navigation. This smart digital teaching material file integrates a dual navigation logic of hierarchical progression and conflict-driven approaches.

[0128] For example, step S1561: Create an empty master interaction graph data structure, which is used to store all textbook cognitive module units and their relationships.

[0129] The main control interaction graph adopts an attribute graph model, containing a node set and an edge set. The node set contains unit nodes of the textbook's cognitive modules, with attributes such as module_id, level_id, title, submodule_count (number of submodules), and completion_rate (suggested completion rate). The edge set contains inter-module connection edges, divided into default progressive edges (progressive by level) and cross-level leap edges (based on conflict relationships), with edge attributes such as source_module, target_module, link_type, avg_intensity (average conflict intensity), and conflict_type_distribution (conflict type distribution). The graph is stored using the Neo4j graph database and supports path analysis using the Cypher query language, such as "find all conflict-driven paths from L1 to L3".

[0130] Step S1562: Use all the created textbook cognitive module units as module nodes in the main control interaction graph data structure. Each module node is associated with its corresponding cognitive depth level identifier and all the enhanced cognitive atomic units within the corresponding module.

[0131] When creating a module node, the attributes of the textbook cognitive module units need to be mapped to the graph node attributes, and a sub-module index needs to be established: a many-to-many relationship is constructed through the atomic codes of the sub-modules: module_id→atomic_code. For example, module M_L1_PHYS is associated with the codes of all L1-level reinforcement cognitive atomic units. To accelerate queries, a level_id index and a title full-text index are created. The node also stores the module's metadata, such as creation time, version number, and update log.

[0132] Step S1563: Traverse all module nodes. For each module node, analyze the inter-module thinking transition trigger links carried by each enhanced cognitive atom unit within it. The inter-module thinking transition trigger links point to target enhanced cognitive atom units under other cognitive depth level identifiers.

[0133] Iterate through the submodules list of the module nodes. For each enhanced cognitive atomic unit u, extract the inter-module cognitive leap trigger links from its inter_links attribute. For each link, find the corresponding target unit v through target_id. Query the module node v_module to which v belongs. Record the conflict intensity and type of the module pair (current module, v_module) and the link. For example, if unit u in module M_L1 points to unit v in module M_L2, then record the (M_L1, M_L2) module pair and related conflict attributes.

[0134] Step S1564: Determine the target textbook cognitive module unit to which the target reinforced cognitive atomic unit pointed to by the inter-module thinking leap trigger link belongs, and in the master control interaction graph data structure, create a cross-layer leap association edge between the current module node and the module node corresponding to the target textbook cognitive module unit. The cross-layer leap association edge carries the initial value of the original conflict intensity and conflict type label information that triggers the leap.

[0135] For each module pair (u_module, v_module): Summarize all link strengths from u_module to v_module and calculate the average as the edge strength; statistically analyze the conflict type distribution and select the most frequent type as the primary conflict type; create cross-layer transition edges in the master interaction graph with the following attributes: type="jump", intensity=average strength, main_conflict_type=primary type, source=u_module.id, target=v_module.id, link_count=total number of links. For example, the edge strength from M_L1 to M_L2 is 0.72, the primary conflict type is CT02, and the total number of links is 15. Cross-layer edges are stored in descending order of strength to prioritize the display of strong conflict transition paths.

[0136] Step S1565: Based on the progressive relationship between the basic cognitive level identifier, the advanced cognitive level identifier, and the cutting-edge cognitive level identifier, a default progressive learning association edge is preset between module nodes belonging to different cognitive depth level identifiers in the master control interaction graph data structure. The default progressive learning association edge represents the standard learning flow from basic to advanced to cutting-edge.

[0137] By default, progressive learning edges are created in the order L1→L2→L3, with the following attributes: type="default", intensity=1.0 (maximum intensity), direction="forward", and recommended_path=true. For example, a default edge is created between M_L1 and M_L2, and between M_L2 and M_L3. In the navigation, the default edges are represented by a "Next" button, guiding learners to learn in a traditional linear order. Reverse default edges (L3→L2→L1) are also supported for review paths, with the attribute direction="backward".

[0138] Step S1566: Serialize the master interactive graph data structure containing all module nodes, cross-level transition links, and default progressive learning links into an interactive mesh navigation data file. Dynamically bind all text content, differentiation navigation links, and transition trigger links within each textbook cognitive module unit to the corresponding nodes and edges in the mesh navigation data file. Finally, generate a smart digital textbook file that allows learners to make non-linear jumps and explorations between different cognitive depth levels based on the intensity of cognitive conflict.

[0139] The master interactive graph is serialized into a JSON-formatted mesh navigation data file (navigation.json), containing a nodes array (module nodes) and an edges array (associated edges). Dynamic binding is implemented as follows: the positions of cognitive tension points in the text are bound to cognitive guidance questions via DOM element IDs; intra-module links are bound to HTML anchor points of target units via target_id; inter-module links are bound to edge attributes in the navigation data file via source_module and target_module. The smart digital textbook file adopts EPUB3.3 format and includes: text content files (HTML), navigation files (nav.xhtml), resource files (images, videos, etc.), and mesh navigation data files. Interactive functions are implemented through JavaScript: relevant links are highlighted based on the current reading position, the navigation path is updated based on learner clicks, and the learning trajectory is recorded for personalized recommendations. The final generated textbook supports online and offline reading, adaptively displays on different devices, and achieves the goal of "non-linear learning driven by cognitive conflict".

[0140] In one exemplary embodiment, a smart digital textbook generation system based on generative artificial intelligence is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2 As shown, this intelligent digital textbook generation system based on generative artificial intelligence includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements an intelligent digital textbook generation method based on generative artificial intelligence. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of the intelligent digital textbook generation system based on generative artificial intelligence, or an external keyboard, touchpad, or mouse, etc.

[0141] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.< / prerequisite> < / level> < / hierarchy> < / modifier> < / modifier> < / modifier>

Claims

1. A method for generating a smart digital textbook based on generative artificial intelligence, characterized by, include: An initial set of subject knowledge corpus and a preset cognitive gradient evolution template are collected. The initial set of subject knowledge corpus includes multiple original academic literature units and multiple original teaching case units. The initial subject knowledge corpus is subjected to latent semantic space mapping and knowledge atomization cutting. Based on the concept cognitive hierarchy progression relationship definition field in the preset cognitive gradient evolution template, the original academic literature unit and the original teaching case unit are deconstructed and reconstructed into multiple basic cognitive atomic units. Each basic cognitive atomic unit is assigned a globally unique atomic code identifier. At the same time, the semantic density distribution feature vector of each basic cognitive atomic unit in the latent semantic space is extracted. Based on the semantic density distribution feature vector and the cross-conceptual thinking conflict point definition field in the preset cognitive gradient evolution template, a cognitive conflict network between the basic cognitive atomic units is constructed. The pre-trained concept generative adversarial neural network model is invoked. Based on the network topology in the cognitive conflict network and the initial value of the conflict intensity, the basic cognitive atomic units are subjected to cross-concept thinking conflict-driven text content deepening processing and cognitive guidance question generation processing to generate a set of enhanced cognitive atomic units containing deepened text content units and a set of associated cognitive guidance questions. Dynamic cognitive path planning is performed based on the enhanced cognitive atomic unit set and the cognitive conflict network to generate a smart digital teaching material file with a non-linear thinking leap path.

2. The intelligent digital textbook generation method based on generative artificial intelligence according to claim 1, characterized in that, The initial subject knowledge corpus undergoes latent semantic space mapping and knowledge atomization processing. Based on the concept cognitive hierarchy progression relationship definition field in the preset cognitive gradient evolution template, the original academic literature units and the original teaching case units are deconstructed and reconstructed into multiple basic cognitive atomic units. Each basic cognitive atomic unit is assigned a globally unique atomic code identifier. Simultaneously, the semantic density distribution feature vector of each basic cognitive atomic unit in the latent semantic space is extracted, including: Each original academic document unit and each original teaching case unit in the initial subject knowledge corpus are subjected to text noise filtering and unified character encoding conversion to obtain a set of clean text corpus units, which contains continuous semantically clean character stream data. A pre-trained deep semantic coding network is invoked to perform global latent semantic space embedding processing on the set of clean text corpus units, mapping each clean text corpus unit to an initial semantic distribution point in a high-dimensional space, and generating an initial semantic coordinate vector corresponding to each initial semantic distribution point. The concept cognitive hierarchy progression relationship definition field in the preset cognitive gradient evolution template is parsed to obtain multiple core discipline concept tags and the cognitive depth level value corresponding to each core discipline concept tag contained in the concept cognitive hierarchy progression relationship definition field, as well as the description text of multiple basic cognitive components under each core discipline concept tag. Density-based spatial clustering analysis is performed on the initial semantic distribution points to identify local maxima regions of semantic density in the high-dimensional space. Each local maxima region of semantic density is defined as a potential cognitive atom nucleus, and the semantic membership probability value from each initial semantic distribution point to all potential cognitive atom nuclei is calculated. Based on the semantic membership probability value of each initial semantic distribution point, each semantically complete sentence unit or paragraph unit in the pure text corpus unit set is assigned to the basic cognitive constituent element description text corresponding to the potential cognitive atomic nucleus with the highest membership degree, forming a cognitive atomic prototype unit with the basic cognitive constituent element description text as the semantic aggregation center. Semantic redundancy elimination and logical order reordering are performed on multiple sentence units or paragraph units belonging to the same basic cognitive component description text. Based on the original contextual relationship of the sentence units or paragraph units in the pure text corpus unit set and the semantic coherence scores between them, they are merged to generate basic cognitive atomic units with independent and complete cognitive meaning. A globally unique atomic code identifier is generated for each basic cognitive atomic unit based on its semantic content hash calculation. An immutable mapping relationship is established between the atomic code identifier and the core subject concept label and cognitive depth level value to which the basic cognitive atomic unit belongs. Extract all semantic vector representations contained in each basic cognitive atom unit, perform spatial grid partitioning on all semantic vector representations, count the density of semantic vector landing points in each spatial grid, and generate a semantic density distribution feature vector that can characterize the coverage and distribution pattern of the basic cognitive atom unit in the latent semantic space. The semantic density distribution feature vector contains the density scalar value corresponding to each spatial grid.

3. The intelligent digital textbook generation method based on generative artificial intelligence according to claim 1, characterized in that, The construction of the cognitive conflict network between the basic cognitive atomic units, based on the semantic density distribution feature vector and the cross-conceptual cognitive conflict point definition field in the preset cognitive gradient evolution template, includes: Obtain multiple conflict type identifiers pre-defined in the cross-conceptual thinking conflict point definition field. The conflict type identifiers include conceptual connotation opposition conflict type, theoretical paradigm incommensurable conflict type, and empirical conclusion contradiction conflict type. Configure corresponding conflict intensity nonlinear mapping function parameters for each conflict type identifier. Using the atomic code identifier of each basic cognitive atomic unit as a network vertex, an initial blank network vertex set is constructed. All atomic code identifiers of all basic cognitive atomic units are added to the network vertex set to form a complete vertex set. Traverse every two different atomic coding identifiers in the set of network vertices, obtain the semantic density distribution feature vectors corresponding to the two different atomic coding identifiers, and calculate the absolute value of the density difference of the two semantic density distribution feature vectors per spatial grid to obtain the local density difference value on each spatial grid. The local density difference values ​​are spatially integrated and accumulated to obtain the overall semantic distribution deviation parameter of the two basic cognitive atomic units in the latent semantic space, and the overall semantic distribution deviation parameter is used as the preliminary calculation result of semantic repulsion force. Based on the core subject concept label and the cognitive depth level value corresponding to the core subject concept label of the basic cognitive atomic unit corresponding to the two different atomic coding identifiers in the concept cognitive hierarchy progression relationship definition field, the actual conflict type between the two basic cognitive atomic units is determined. The actual conflict type is one of the conceptual connotation opposition conflict type, the theoretical paradigm incommensurable conflict type, or the empirical conclusion contradiction conflict type. Obtain the parameters of the nonlinear mapping function for conflict intensity corresponding to the actual conflict type, input the preliminary calculation result of the semantic repulsion force into the nonlinear mapping function for mapping transformation, and obtain the initial value of the conflict intensity between the two basic cognitive atomic units. In the set of network vertices, a conflict edge is created between the two different atomic code identifiers. The conflict edge connects the two different atomic code identifiers, and the calculated initial value of the conflict intensity is stored as the weight attribute of the conflict edge. At the same time, the actual conflict type is also stored as the type label attribute of the conflict edge. After all the different atomic code pairs in the set of network vertices have been processed, a cognitive conflict network containing all vertices and all conflicting edges is generated.

4. The intelligent digital textbook generation method based on generative artificial intelligence according to claim 1, characterized in that, The pre-trained concept-generative adversarial neural network model is invoked, and based on the network topology and initial conflict intensity values ​​in the cognitive conflict network, the basic cognitive atomic units undergo cross-conceptual thinking conflict-driven text content deepening processing and cognitive guidance question generation processing to generate a reinforced cognitive atomic unit set containing deepened text content units and a set of associated cognitive guidance questions, including: From the cognitive conflict network, for each vertex corresponding to a basic cognitive atom unit, extract all its adjacent vertices directly connected by conflict edges, as well as the initial value of the conflict intensity and the conflict type label of the conflict edges connected to each adjacent vertex, and construct a set of local conflict topology context information with the current basic cognitive atom unit as the core vertex. The semantic density distribution feature vector of the basic cognitive atomic unit and the set of local conflict topological context information are jointly embedded and encoded to generate a conflict-aware input tensor containing the semantic information of the core vertex and the distribution of its surrounding conflict field. The conflict-aware input tensor is input into the generator network part of the pre-trained concept generative adversarial neural network model. Deep conflict feature extraction and fusion processing are performed through multiple residual convolution modules and self-attention mechanism modules of the generator network part to generate a hidden layer feature map with conflict-driven semantic enhancement. The discriminator network of the pre-trained conceptual generative adversarial neural network model is invoked to perform authenticity and conflict consistency discrimination on the hidden layer feature map. The parameters of the generator network are adjusted in reverse according to the loss gradient returned by the discriminator network. Based on the parameter-adjusted generator network, the deepened text content units for the basic cognitive atomic units are constructed segment by segment in an autoregressive generation manner. The deepened text content units, while fully retaining the original cognitive kernel of the basic cognitive atomic units, introduce critical analysis and integrative thinking guidance content on the related conflict atomic units in the local conflict topological context information set. Analyze each semantic paragraph in the deepened text content unit, identify the cognitive tension points in the semantic paragraph, and the cognitive tension points are manifested as the positions of sentences that simultaneously mention two core subject concept tags that are connected by conflicting edges within the deepened text content unit; For each identified cognitive tension point, based on the two core subject concept labels mentioned at the same time and their corresponding initial conflict intensity values, a cognitive guidance question is generated to guide learners to think about the relationship between the two. The cognitive guidance question is presented in the form of a question and points to the conflict edge of the two core subject concept labels in the cognitive conflict network. All generated cognitive guidance questions are sorted and combined according to their order of appearance in the deepened text content units to form a set of related cognitive guidance questions; The enhanced text content unit and the set of related cognitive guidance questions are encapsulated and combined to generate a reinforced cognitive atom unit corresponding to the basic cognitive atom unit. All the reinforced cognitive atom units corresponding to the basic cognitive atom units together constitute a set of reinforced cognitive atom units.

5. The intelligent digital textbook generation method based on generative artificial intelligence according to claim 1, characterized in that, The process of generating a smart digital teaching material file with a non-linear thinking leap path by performing dynamic cognitive path planning based on the enhanced cognitive atom unit set and the cognitive conflict network includes: The concept cognitive hierarchy progression relationship definition field in the preset cognitive gradient evolution template is parsed to obtain the cognitive depth hierarchy division structure defined in the concept cognitive hierarchy progression relationship definition field. The cognitive depth hierarchy division structure includes basic cognitive level identifiers, advanced cognitive level identifiers and cutting-edge cognitive level identifiers, as well as a list of core subject concept labels expected to be covered under each cognitive depth hierarchy identifier. Traverse each enhanced cognitive atom unit in the set of enhanced cognitive atom units, and based on the atomic code identifier of the basic cognitive atom unit corresponding to each enhanced cognitive atom unit, trace back the core subject concept label and cognitive depth level value of its mapping association, and then determine the cognitive depth level identifier to which each enhanced cognitive atom unit should belong. According to the cognitive depth hierarchy, all enhanced cognitive atomic units belonging to the same cognitive depth hierarchy are subjected to internal path planning. The internal path planning process determines the nonlinear recommended exploration order of enhanced cognitive atomic units within the same cognitive depth hierarchy based on the initial value of the conflict intensity and the conflict type label of the conflict edges between the vertices corresponding to these enhanced cognitive atomic units in the cognitive conflict network, and generates the concept discrimination navigation link within the module. The concept discrimination navigation link within the module contains the link address pointing to other enhanced cognitive atomic units within the same cognitive depth hierarchy that have strong conflict edges connected to the current enhanced cognitive atomic unit. For each enhanced cognitive atom unit, analyze the conflict edges in the cognitive conflict network that connect the vertex of the enhanced cognitive atom unit to vertices in other cognitive depth levels and whose initial conflict intensity exceeds a preset conflict threshold. Extract the atom code identifiers of the target enhanced cognitive atom units in other cognitive depth levels pointed to by these conflict edges, and generate inter-module thinking leap trigger links from the current enhanced cognitive atom unit to the target enhanced cognitive atom unit. Create a corresponding textbook cognitive module unit for each cognitive depth level identifier. All enhanced cognitive atomic units belonging to the cognitive depth level identifier are sequentially used as internal sub-modules of the textbook cognitive module unit according to the non-linear recommended exploration order. The concept discrimination navigation links within the module and the thinking leap trigger links between modules are embedded as dynamic interactive metadata into the data structure of each enhanced cognitive atomic unit. All created cognitive module units of the teaching materials are non-linearly associated and combined according to the progressive relationship of the basic cognitive level identifier, the advanced cognitive level identifier, and the cutting-edge cognitive level identifier to generate a smart digital teaching material file with a network-like but non-linear structure. The text interaction area inside each enhanced cognitive atom unit in the smart digital teaching material file is bound to the concept analysis navigation link within the module and the thinking leap trigger link between the modules, so as to realize the exploratory learning of learners to make autonomous thinking leaps based on the intensity of cognitive conflict during the learning process.

6. The intelligent digital textbook generation method based on generative artificial intelligence according to claim 4, characterized in that, The step of jointly embedding and encoding the semantic density distribution feature vector of the basic cognitive atomic unit and the set of local conflict topological context information to generate a conflict-aware input tensor containing the semantic information of the core vertices and the distribution of their surrounding conflict fields includes: Extract the semantic density distribution feature vectors of the basic cognitive atomic units corresponding to all adjacent vertices from the local conflict topological context information set, and sort all the semantic density distribution feature vectors of adjacent vertices in descending order of their corresponding initial conflict strength values ​​to obtain the sorted adjacent vertex feature vector sequence. The sorted adjacent vertex feature vector sequence is truncated to a fixed maximum length or padded with zeros at the end, so that the length of the adjacent vertex feature vector sequence is uniformly equal to the preset maximum number of adjacent vertices, thus obtaining a length-normalized adjacent vertex feature vector matrix. Obtain the initial value of the conflict intensity and the conflict type label of the conflict edge connecting the core vertex and each adjacent vertex. Vectorize the initial value of the conflict intensity and the conflict type label to generate an edge attribute feature vector corresponding to each adjacent vertex. The edge attribute feature vector contains a normalized initial value of the conflict intensity and a binary encoded component representing the conflict type label. The length-normalized adjacent vertex feature vector matrix and the corresponding edge attribute feature vector are concatenated along the feature channel dimension to obtain an enhanced adjacent vertex feature representation containing adjacent vertex semantic information and connecting edge attributes. The semantic density distribution feature vector of the core vertex and the enhanced adjacent vertex feature representation are subjected to conflict field fusion processing. The conflict field fusion processing adopts a graph attention network mechanism, using the semantic density distribution feature vector of the core vertex as the attention query benchmark, and the enhanced adjacent vertex feature representation as the attention key and attention value. The attention coefficient of each adjacent vertex relative to the core vertex is calculated, and the enhanced adjacent vertex feature representation is weighted and summed to obtain a conflict-aware input tensor that aggregates the adjacent vertex information and its conflict intensity.

7. The intelligent digital textbook generation method based on generative artificial intelligence according to claim 4, characterized in that, The process of analyzing each semantic paragraph in the deepened text content unit and identifying cognitive tension points within the semantic paragraphs includes: The deepened text content units are subjected to semantic paragraph segmentation to obtain multiple consecutive semantic paragraph units, each of which contains a sequence of sentences expressing relatively complete semantics; Named entity recognition and term extraction are performed on each semantic paragraph unit to extract all core subject concept tags appearing in the semantic paragraph unit and the precise character position offset of each core subject concept tag appearing in the semantic paragraph unit. Iterate through all the extracted core subject concept tags. For each core subject concept tag, search for a list of other core subject concept tags that are connected to it by conflicting edges in the cognitive conflict network. For the current semantic paragraph unit, a set of core subject concept tags that appear in the corresponding paragraph is constructed. Each pair of tags in the set of core subject concept tags is traversed to determine whether the pair of tags exists in the conflict edge connection relationship of the cognitive conflict network. If the pair of tags exists in the conflict edge connection relationship of the cognitive conflict network, then it is determined that there is a cognitive tension point in the semantic paragraph unit, and the name of the pair of tags and the precise character position offset of each pair of tags in the semantic paragraph unit are recorded. Based on the offset of the ending character of the first core subject concept tag and the offset of the starting character of the second core subject concept tag in the tag pair, the core text span that mentions both concepts simultaneously in the semantic paragraph unit is calculated, and the starting and ending positions of the core text span are used as the positioning coordinates of the cognitive tension point in the deepened text content unit.

8. The intelligent digital textbook generation method based on generative artificial intelligence according to claim 4, characterized in that, For each identified cognitive tension point, based on the two core subject concept tags mentioned simultaneously at that cognitive tension point and their corresponding initial conflict intensity values, a cognitive guidance question is generated to guide learners to think about the relationship between the two, including: Obtain the first core subject concept label and the second core subject concept label mentioned at the cognitive tension point, and extract the initial value of the conflict intensity and the conflict type label of the conflict edge connecting the first core subject concept label and the second core subject concept label from the cognitive conflict network; Based on the conflict type tag, a question template matching the conflict type tag is retrieved from a preset question template library. The question template includes a first concept placeholder for accommodating the first core discipline concept tag and a second concept placeholder for accommodating the second core discipline concept tag. Fill the first core subject concept tag's full name text into the first concept placeholder position of the question template, and fill the second core subject concept tag's full name text into the second concept placeholder position of the question template to generate a preliminary formatted question text; Based on the magnitude of the initial conflict intensity value, a degree adverb corresponding to the numerical range of the initial conflict intensity value is selected from a preset intensity modifier word library. The degree adverb is inserted into the initially formatted question text to modify the position of the relationship strength between the first core subject concept label and the second core subject concept label, thereby generating a cognitive guidance question with intensity awareness. The generated cognitive guidance question is associated with and stored in the positioning coordinates of the cognitive tension point in the deepened text content unit, so that the cognitive guidance question can be displayed near the corresponding position in the deepened text content unit during final presentation.

9. A smart digital textbook generation system based on generative artificial intelligence, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the intelligent digital textbook generation method based on generative artificial intelligence as described in any one of claims 1 to 8 by executing the machine-executable instructions.

10. A computer program product, characterized in that, The computer program product includes machine-executable instructions stored in a computer-readable storage medium. The processor of the intelligent digital textbook generation system based on generative artificial intelligence reads the machine-executable instructions from the computer-readable storage medium and executes the machine-executable instructions, causing the intelligent digital textbook generation system based on generative artificial intelligence to perform the intelligent digital textbook generation method based on generative artificial intelligence as described in any one of claims 1 to 8.

Images