Method and system for training content generation based on deep learning model, and storage medium

Abstract

The application relates to the technical field of data generation, and discloses a training content generation method and system based on a deep learning model and a storage medium. The method comprises the following steps: obtaining basic content data, extracting text, tables and format content to generate structured data; performing semantic analysis on the structured data by using a deep learning model, identifying the content structure, extracting training knowledge elements, constructing a knowledge graph by combining the correlation between the elements, calculating the number of occurrences, the content length and the core degree of the training knowledge elements to obtain a difficulty coefficient; obtaining training parameters, calculating a knowledge coverage coefficient and a knowledge depth coefficient, combining a first text framework and a large language model to generate initial content; and performing full-content verification and correction on the initial content by using a second text framework to obtain training content. The application realizes automatic generation of training content, solves problems such as low efficiency, high cost and unstable quality of manual test question setting, and improves the scientificity and pertinence of the training content.

Description

Technical Field

[0001] This application relates to the field of data generation technology, and in particular to a method, system and storage medium for generating training content based on a deep learning model. Background Technology

[0002] Online education and digital teaching have been widely integrated into various scenarios, including daily teaching in schools, professional skills training in hospitals, courses in vocational training institutions, and on-the-job training for professional development. As a core means of evaluating teaching and testing training effectiveness, the scientific nature, relevance, and suitability of test questions directly impact learners' mastery of knowledge and the efficiency of achieving learning goals. High-quality questions not only accurately reflect learning weaknesses but also guide learners to systematically consolidate knowledge, making them a crucial element in ensuring the quality of digital teaching.

[0003] Currently, training questions in various teaching scenarios still mainly rely on professional question setters to create them. These question setters need to thoroughly study the latest teaching syllabus, core textbook content, and skills training requirements in the corresponding field. They then manually design question types, draft question stems, select assessment knowledge points, and provide precise answers and explanations, ensuring that the entire process fully considers both the difficulty level of the questions and the effectiveness of the assessment.

[0004] However, the manual question-generating model has significant limitations: It is inefficient; a high-quality question that accurately covers knowledge points, is logically rigorous, and has an appropriate level of difficulty takes an average of 15-30 minutes from conception to finalization, making it difficult to meet the needs of large-scale teaching for rapid expansion of the question bank; it is costly; professional question creators need solid subject backgrounds and rich experience, and long-term reliance on such human resources leads to high labor costs for question bank construction and maintenance; it has unstable quality; the quality of questions directly depends on the individual professional competence and experience of the question creators, and different individuals have different standards, making it difficult to form a unified and standardized question bank system; it lacks comprehensive knowledge coverage; manual selection of knowledge points is easily influenced by subjective factors, making it difficult to systematically cover all core knowledge points and potential test points in textbooks or syllabi, easily leading to omissions of key points; and it is difficult to update; when textbook content, syllabi, or industry skill standards are updated, the time required for manual revision of the existing question bank and the writing of new questions is long, resulting in a lag in question bank content updates, further increasing the time and labor costs of question bank maintenance. Summary of the Invention

[0005] To improve the quality and efficiency of training content, this application provides a method, system, and storage medium for generating training content based on a deep learning model.

[0006] Firstly, this application provides a method for generating training content based on a deep learning model, employing the following technical solution:

[0007] A method for generating training content based on a deep learning model includes the following steps:

[0008] Acquire basic content data, and extract text content, table content, and formatted content from the basic content data to generate structured data;

[0009] Based on a pre-defined deep learning model, semantic analysis is performed on structured data to identify content structure from the structured data, training knowledge elements are extracted from the structured data according to the content structure, the relationships between training knowledge elements are identified, and a knowledge graph is constructed based on the training knowledge elements and the relationships.

[0010] Calculate the position of training knowledge elements in structured data, and calculate the frequency of occurrence, content length, and coreness of training knowledge elements in the knowledge graph based on their positions; calculate the difficulty coefficient based on the frequency of occurrence, content length, and coreness.

[0011] Obtain the input training parameters, including the number of training samples, training difficulty, and training type; calculate the knowledge coverage coefficient based on the number of training samples and the knowledge graph, and calculate the knowledge depth coefficient based on the training difficulty, training type, and knowledge graph; apply the preset first text framework, and generate initial content based on the large language model using structured data according to the number of training samples, training type, knowledge coverage coefficient, and knowledge depth coefficient;

[0012] The initial content is fully validated using a pre-defined second text framework to generate validation data. The initial content is then corrected based on the validation data to obtain the training content.

[0013] By adopting the above technical solution, basic content data is transformed into structured data. Semantic analysis is carried out using deep learning models to construct a knowledge graph. The difficulty coefficient is accurately calculated by combining the occurrence characteristics of training knowledge elements with the core degree of the knowledge graph. Then, the knowledge coverage coefficient and knowledge depth coefficient are determined based on the input training parameters. The initial content generation and full content verification and correction are realized through a dual-text framework. This not only effectively solves the problems of low efficiency, high cost, unstable quality, incomplete knowledge coverage and delayed updates of manual question generation, but also ensures the scientificity and pertinence of training content in terms of knowledge coverage, difficulty adaptation and type matching. It significantly improves the efficiency and quality of training content generation and reduces the cost of question bank construction and maintenance.

[0014] Optionally, the step of extracting text content, table content, and formatted content from the basic content data to generate structured data also includes the following sub-steps:

[0015] Obtain the type and format of basic content data;

[0016] If the type format is within the preset first format range, the basic content data is converted into image sequence data. The image sequence data retains the page layout information of the basic content data. The graphic data in the image sequence data is extracted, and the text content, table content, and format content are identified from the graphic data.

[0017] If the type format is within the preset second format range, then the text content, table content, and format content are directly separated from the basic content data;

[0018] If the type format is outside the first format range and the second format range, a conversion request is issued, the converter corresponding to the conversion request is obtained, and the type format is converted using the converter to obtain a temporary format. If the temporary format is within the first format range or the second format range, the type format is added to the first format range, and the first format range is associated with the converter.

[0019] By adopting the above technical solution, the basic content data type and format are first identified, and differentiated processing strategies are adopted for different format ranges. Data in the first format range is converted into image sequences to retain page layout information and extract relevant content. Data in the second format range is directly segmented and the content is extracted. Formats exceeding the two ranges are converted by a converter and then included in the corresponding format range and associated with the converter. This not only achieves comprehensive adaptation and flexible processing of basic content data of various types and formats, effectively expanding the adaptation range of basic content data, but also ensures data integrity and accuracy during the extraction of text, tables and formatted content.

[0020] Optionally, the steps of performing semantic analysis on structured data based on a pre-defined deep learning model, identifying content structure from the structured data, and extracting training knowledge elements from the structured data based on the content structure further include the following sub-steps:

[0021] The deep learning model is used to identify the chapters, sections and subsections of a document as its content structure. The headings and paragraphs at all levels in the content structure are extracted, and the relationships between the structures in the content structure are analyzed. The relationships include progressive, parallel and subordinate.

[0022] Based on the order of association, definitions, theorems, formulas and rules are extracted from the content structure as core knowledge points, explanations of the core knowledge points are generated, keywords or terms related to the core knowledge points are extracted as standardized names of the core knowledge points, and training knowledge elements are obtained by binding the explanations with the standardized names.

[0023] By adopting the above technical solution, the deep learning model accurately identifies the content structure of document chapters, sections, and subsections, and analyzes the progressive, parallel, and subordinate relationships between the structures. Core knowledge points such as definitions, theorems, formulas, and rules are extracted in the order of the relationships, and corresponding explanations are generated and bound to standardized names to form training knowledge elements. This not only realizes the structured and logical extraction of core knowledge points and ensures the accuracy, standardization, and relevance of training knowledge elements, but also provides a high-quality knowledge foundation with clear logic and well-defined definitions for the subsequent construction of knowledge graphs.

[0024] Optionally, the step of calculating the difficulty coefficient based on the frequency of occurrence, content length, and core content also includes the following sub-steps:

[0025] Select standard elements from the training knowledge elements, and use the frequency of occurrence, content length, and coreness of the standard elements as the standard frequency, standard length, and standardness;

[0026] Based on the training knowledge elements that need to be calculated, the ratio of the number of occurrences to the standard number of occurrences is calculated as the frequency ratio, the ratio of the content length to the standard length is calculated as the length ratio, and the ratio of coreness to standardness is calculated as the core ratio.

[0027] The difficulty coefficient is calculated by weighting the ratio of frequency, length, and core content.

[0028] By adopting the above technical solution, a unified benchmark is established by selecting standard elements in the training knowledge elements. The corresponding ratios of the current training knowledge elements to the standard elements in terms of frequency of occurrence, content length, and coreness are calculated. The difficulty coefficient is then obtained by weighted averaging. This not only realizes the quantification and standardization of difficulty calculation and effectively avoids the interference of subjective factors in difficulty judgment, but also takes into account the occurrence characteristics of knowledge elements and the core value in the knowledge graph through the comprehensive consideration of multi-dimensional indicators, making the calculation of the difficulty coefficient more objective, accurate, and reasonable.

[0029] Optionally, the steps of calculating the knowledge coverage coefficient based on the number of training samples and the knowledge graph, and calculating the knowledge depth coefficient based on the training difficulty, training type, and knowledge graph, also include the following sub-steps:

[0030] Obtain the number of clusters of training knowledge elements extracted from the knowledge graph, and calculate the ratio of the number of training elements to the number of clusters as the knowledge coverage coefficient.

[0031] Obtain the difficulty coefficient of the training knowledge element in the knowledge graph, calculate the average of all difficulty coefficients as the reference difficulty, and calculate the ratio of the training difficulty to the reference difficulty as the difficulty ratio.

[0032] Obtain the main type of the training knowledge element in the knowledge graph. The main type corresponds to a type amplitude. Calculate the average of all type amplitudes as the reference amplitude. Calculate the ratio of the type amplitude of the training type to the reference amplitude as the type ratio.

[0033] The knowledge depth coefficient is calculated by weighting the difficulty ratio and the type ratio.

[0034] By adopting the above technical solution, the knowledge coverage coefficient is quantified by the ratio of the number of clusters of training knowledge elements in the knowledge graph to the number of training elements. Combined with the average difficulty coefficient and the average amplitude of the main type of training knowledge elements in the knowledge graph, the ratio of training difficulty to reference difficulty and the ratio of training type amplitude to reference amplitude are calculated respectively. Then, the knowledge depth coefficient is obtained by weighted averaging. This not only achieves accurate quantitative calculation of knowledge coverage coefficient and knowledge depth coefficient, but also ensures the objectivity and rationality of coefficient calculation by relying on the knowledge graph. It takes into account the comprehensiveness and depth adaptability of the knowledge coverage of training content and reduces the interference of subjective factors on parameter determination.

[0035] Optionally, the step of generating initial content using structured data based on a large language model according to the number of training samples, training type, knowledge coverage coefficient, and knowledge depth coefficient also includes the following sub-steps:

[0036] Extract the first question stem information from the first text frame, and use the first question stem information to guide the large language model to generate initial content for selection;

[0037] The character count of the previous first question stem is the first character count, and the character count of the current first question stem is the second character count;

[0038] The ratio of the second character count to the first character count is calculated as the character count ratio. The weight of the difficulty ratio is adjusted based on the negative correlation of the character count ratio.

[0039] By adopting the above technical solution, the large language model is guided to generate selection-based initial content by extracting the first question stem information from the first text framework, ensuring the standardization and directionality of the initial content generation. At the same time, by calculating the character ratio of adjacent first question stem information, dynamic monitoring of the length of the question stem description is achieved. Furthermore, relying on the negative correlation adjustment mechanism between the character ratio and the difficulty ratio weight, the influence of difficulty factors on the initial content can be flexibly adjusted according to changes in the character count of the question stem, effectively avoiding the difficulty judgment deviation caused by the question stem description being too long or too short, and further optimizing the balance between the standardization of the question stem description and the difficulty suitability of the initial content.

[0040] Optionally, the step of generating initial content using structured data based on a large language model according to the number of training samples, training type, knowledge coverage coefficient, and knowledge depth coefficient also includes the following sub-steps:

[0041] Extract the second question stem information from the first text frame, and use the second question stem information to guide the large language model to generate short-answer type initial content;

[0042] The number of characters in the current first question stem is the second number of characters;

[0043] Obtain the distinguishing characters between the previous first question stem information and the current first question stem information, and calculate the proportion of the number of different characters in the number of second characters as the distinguishing ratio;

[0044] The weight of the type ratio is adjusted based on the positive correlation between the difference ratio and the type ratio.

[0045] By adopting the above technical solution, the large language model is guided to generate short-answer initial content by extracting the second question stem information from the first text framework, ensuring the standardization and question type adaptability of the generated short-answer training content. At the same time, by calculating the difference ratio of the number of distinguishing characters in adjacent first question stem information to the number of characters in the current question stem, the degree of differentiation in question stem expression is accurately quantified. Furthermore, relying on the positive correlation adjustment mechanism between the difference ratio and the type ratio weight, the influence weight of question type factors on the initial content can be flexibly adjusted according to the degree of differentiation in question stem expression. This effectively avoids excessive repetition of question stem expressions of the same type of question, improves the diversity and relevance of the initial content, and further optimizes the balance between question type adaptability and expression richness in the short-answer initial content.

[0046] Optionally, the step of applying a preset second text framework to perform full content validation on the initial content, generating validation data, and then correcting the initial content based on the validation data to obtain the training content further includes the following sub-steps:

[0047] Based on the second text framework, at least one verification dimension is matched from the preset verification feature library;

[0048] If the verification dimension includes content correctness, the correlation between the initial content and the basic content data is verified using a pre-defined preliminary large language model. If the correlation is within a pre-defined range, the verification passes; otherwise, the verification fails. Here, 40% ≤ correlation range ≤ 100%.

[0049] If the verification dimension includes content duplication, the cosine similarity between the current initial content and the historically stored content is calculated based on the vector algorithm. If the cosine similarity is less than the preset reference similarity, the verification passes; otherwise, the verification fails.

[0050] Correct or regenerate any parts of the initial content that failed verification.

[0051] By adopting the above technical solution, and relying on the second text framework to match at least one verification dimension from the preset verification feature library, multi-dimensional standardized full-process verification of the initial content is achieved. Regarding the content correctness dimension, a preliminary large language model is used to verify the correlation between the initial content and the basic content data. By limiting the correlation range to 40% ≤ correlation ≤ 100%, the close connection between the initial content and core knowledge points is effectively ensured, avoiding deviation from the core content of the teaching syllabus and textbooks. This also prevents the copying of basic content data, ensuring the originality and scientific rigor of the training content. Regarding the content repetition dimension, the cosine similarity between the current initial content and historically stored content is accurately calculated based on a vector algorithm. A similarity threshold is used to effectively identify and avoid duplicate content, improving the diversity of the training content. Simultaneously, targeted correction or regeneration of the failed verification parts of the initial content can accurately address the quality defects present in the initial content.

[0052] Secondly, this application provides a training content generation system based on a deep learning model, which adopts the following technical solution:

[0053] A training content generation system based on a deep learning model includes a processor, wherein the processor performs the steps of the training content generation method based on a deep learning model as described in any one of the preceding claims.

[0054] Thirdly, this application provides a storage medium, which adopts the following technical solution:

[0055] A storage medium storing a program that, when executed by a processor, implements the steps of the deep learning model-based training content generation method described above.

[0056] In summary, this application includes at least one of the following beneficial technical effects: This application efficiently solves the core pain points of low efficiency, high cost, unstable quality, incomplete knowledge coverage, and lagging updates under the manual question generation model, significantly improves the scientific nature, relevance, originality, and diversity of training content, effectively guarantees the overall quality level of the content, greatly reduces the human resource investment and time cost of question bank construction and maintenance, and can accurately adapt to the needs of various large-scale and diversified digital teaching and skills training scenarios such as online education and professional skills training. Attached Figure Description

[0057] Figure 1 This is a step diagram of a method for generating training content based on a deep learning model.

[0058] Figure 2 This diagram illustrates the sub-steps involved in extracting text, table, and formatted content from basic data to generate structured data.

[0059] Figure 3 It is a diagram of the sub-steps for extracting training knowledge elements from structured data based on the content structure. Detailed Implementation

[0060] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0061] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0062] This application discloses a method for generating training content based on a deep learning model. The method utilizes a large language model that can automatically capture or obtain resources such as newly published textbooks and industry reference materials as basic data sources, either directly or according to user instructions. Figure 1 Specifically, it includes the following steps:

[0063] Acquire basic content data (including new textbooks, reference materials, etc. acquired automatically / by command), extract text content, tabular data, and formatting information, and generate structured data through standardization processing to ensure data integrity and standardization;

[0064] Based on a pre-defined deep learning model adapted to the semantic features of multiple disciplines, semantic analysis is performed on structured data to accurately identify the content structure of document chapter-section-subsection levels. Training knowledge elements such as definitions and theorems are extracted from the structure, and the progressive and parallel relationships between elements are sorted out to build a logically clear knowledge graph.

[0065] The specific location of training knowledge elements in structured data is determined, and the frequency of their occurrence and the length of the core explanation content are statistically analyzed. Combined with their correlation with the core nodes in the knowledge graph, an objective difficulty coefficient is obtained through multi-dimensional quantitative calculation.

[0066] The system obtains the training parameters input by the user (including the number of training sessions, target training difficulty, and question type, etc.), calculates the knowledge coverage coefficient by combining the knowledge cluster distribution in the knowledge graph, and calculates the knowledge depth coefficient based on the adaptability of the training difficulty, question type, and knowledge graph. It then calls the preset first text framework, and the large language model combines the structured data, training parameters, and the two types of coefficients to generate initial content that meets the requirements.

[0067] The system applies a pre-defined second text framework to perform full-dimensional verification (including correctness, repetition, etc.) on the initial content and generates verification data. For parts that fail verification, the system automatically or manually corrects them, ultimately obtaining qualified training content.

[0068] The basic content data (including newly added resources) is transformed into structured data. Semantic analysis is carried out using deep learning models to construct a knowledge graph. The difficulty coefficient is accurately calculated by combining the occurrence characteristics of training knowledge elements with the core degree of the knowledge graph. Then, the knowledge coverage coefficient and knowledge depth coefficient are determined based on the input training parameters. The initial content generation and full content verification and correction are realized through a dual-text framework. This not only effectively solves the problems of low efficiency, high cost, unstable quality, incomplete knowledge coverage and delayed content updates caused by manual question generation, but also flexibly adapts to the update needs of new textbooks and new reference materials. It ensures the scientificity and pertinence of training content in terms of knowledge coverage, difficulty adaptation and type matching, significantly improves the efficiency and quality of training content generation, and greatly reduces the manpower and time costs of question bank construction and maintenance.

[0069] Reference Figure 2 The steps of extracting text content, table content, and formatted content from basic content data to generate structured data also include the following sub-steps:

[0070] The basic content data type and format are obtained by combining file extension parsing, file header information recognition, and format feature detection. The first preset format range includes non-editable scanned PDFs, image files (JPG / PNG / BMP), encrypted read-only documents, etc. The second preset format range includes editable PDFs, Word (DOC / DOCX), TXT text, Excel (XLS / XLSX), and other formats that support content segmentation and extraction.

[0071] If the format falls within the preset first format range, the basic content data is converted page by page into a high-definition image sequence of 300 DPI. The image sequence data is stored in PNG lossless compression format to ensure complete preservation of the page layout information of the basic content data, including the position coordinates, hierarchical relationships, and layout styles of text, tables, and charts. Subsequently, the graphic data in the image sequence data is extracted using a graphic data separation algorithm. A high-precision OCR engine with an accuracy of ≥98% is used to recognize the graphic data. A table structure detection model is used to identify table borders and cell distribution to extract table content. A text region positioning model is used to identify text content and corresponding font, font size, color, and other format attributes, achieving accurate separation of text content, table content, and formatted content.

[0072] If the type format is within the preset second format range, the content is directly segmented based on the document's native markup language (such as Word's XML structure and PDF's object flow structure), the text content is extracted through the tag parsing algorithm, the table's row and column data and header information are extracted through the table object recognition module, and the format attributes such as paragraph spacing, indentation, bold / italic are obtained through the format attribute reading interface. No additional format conversion steps are required, ensuring the high efficiency of content extraction.

[0073] If the format type falls outside the first and second format ranges, a conversion request is issued by calling a preset format conversion interface library. This library integrates mainstream cross-format conversion tools, such as the LibreOffice conversion interface and online format conversion APIs, to obtain a converter that matches the format type. The converter is then used to convert the format type to a temporary format, prioritizing conversion to editable PDF format. The format verification module checks whether the temporary format falls within the first or second format range. If the verification passes, the format type is added to the adaptation list of the first format range, and the system's format-converter mapping database is associated with the format type and the corresponding converter's call path and parameter configuration. Subsequently, when basic content data of the format type is obtained again, the associated converter can be automatically called to complete the format conversion without repeatedly initiating a conversion request.

[0074] It achieves comprehensive adaptation and flexible processing of basic content data in various formats, effectively expanding the scope of adaptation of basic content data, and ensuring data integrity and accuracy during the extraction of text, tables and formatted content.

[0075] Reference Figure 3 The steps of performing semantic analysis on structured data based on a pre-defined deep learning model, identifying content structure from the structured data, and extracting training knowledge elements from the structured data based on the content structure also include the following sub-steps:

[0076] The deep learning model is used to identify the chapters, sections and subsections of a document as its content structure. The headings and paragraphs at all levels in the content structure are extracted, and the relationships between the structures in the content structure are analyzed. The relationships include progressive, parallel and subordinate.

[0077] The pre-selected deep learning model is preferably the BERT multi-layer bidirectional semantic understanding model, which has been pre-trained and fine-tuned with massive amounts of labeled data from educational and professional training documents. This model has accurate hierarchical structure recognition and semantic association analysis capabilities, and can effectively adapt to the document structure characteristics of different disciplines and fields. In the process of recognizing the chapter, section, and subsection structure of a document, the model will combine document format identifiers (such as heading styles, indentation levels, and numbering rules) and semantic features (such as logical boundaries of content and topic switching nodes) for dual judgment to ensure the accuracy of structure recognition. At the same time, the identified chapter, section, and subsection structures are hierarchically numbered and structured for storage, forming a clear document hierarchy tree. This facilitates subsequent knowledge point extraction and correlation analysis. When extracting headings and content paragraphs at all levels, the complete text of headings at each level is extracted first as hierarchical identifiers. Key sentences are filtered for content paragraphs, and transitional statements, repetitive expressions, and irrelevant annotations are removed through semantic redundancy calculation, retaining only core knowledge expression paragraphs to improve the efficiency and accuracy of subsequent processing. When analyzing the relationships between structures, in addition to identifying the three basic relationships of progression, parallelism, and subordination, the preceding relationships of chapter structures are further constructed to clarify the basic prerequisite chapters required to learn the knowledge of the current chapter. At the same time, the corresponding knowledge level is marked for each chapter structure to form a complete chapter relationship map.

[0078] Based on the order of association, definitions, theorems, formulas and rules are extracted from the content structure as core knowledge points, explanations of the core knowledge points are generated, keywords or terms related to the core knowledge points are extracted as standardized names of the core knowledge points, and training knowledge elements are obtained by binding the explanations with the standardized names.

[0079] Based on the logical order of chapter structure relationships, a hybrid extraction strategy combining rule extraction and deep learning model recognition is employed to accurately extract definitions, theorems, formulas, and rules as core knowledge points from the content structure: Definition knowledge points are identified through semantic boundary recognition using feature keywords such as "definition," "refers to," and "that is"; Theorem knowledge points are identified through logical derivation structures using identifiers such as "theorem," "axiom," and "law"; Formula knowledge points are identified through high-precision OCR recognition and LaTeX format conversion using mathematical symbols and formula formatting to ensure the accuracy and structured storage of formulas; Rule knowledge points are identified through industry standard expression features using keywords such as "rule," "standard," and "requirement." When generating explanations for core knowledge points, two levels of explanation content are generated for each core knowledge point: the first level is a concise description, limited to 50 words, for quick understanding of the core meaning of the knowledge point; the second level is a detailed explanation, fully elaborating on the specific content, applicable conditions, derivation process, and related explanations of the knowledge point, ensuring the accuracy, completeness, and comprehensibility of the explanation content. When extracting standardized names for core knowledge points, the principles of conciseness, uniqueness, and professionalism are followed, prioritizing the use of general subject-specific and industry-standard terms. If multiple related terms exist, semantic similarity calculations are used to select the term with the highest match to the knowledge point's connotation as the standardized name, and other related terms are stored as synonyms for easy subsequent knowledge retrieval and matching. When obtaining training knowledge elements, in addition to binding the explanations of core knowledge points with standardized names, a further association between each training knowledge element and other training knowledge elements is constructed. This includes the prerequisite knowledge elements required by the training knowledge element, parallel knowledge elements at the same level, the component knowledge elements, and the actual application scenario knowledge elements. Simultaneously, a basic difficulty coefficient is initially labeled for each training knowledge element, forming a complete structured training knowledge element.

[0080] To illustrate the above implementation process more clearly, a specific example of a mathematics teaching document will be used for explanation. The chapter structure and the relationship between the training knowledge elements in this example are shown below:

[0081] ├─3.1 The Concept of Functions (Basic Training Knowledge Elements)

[0082] │├─Prerequisite knowledge elements: sets, mappings

[0083] │├─Core knowledge points: Definition, domain, range, and correspondence rule of functions

[0084] │├─Normalized Name: The Concept of Functions

[0085] │├─Brief Description: The correspondence between variables based on sets and mappings.

[0086] │└─Initial Difficulty Level: 0.3

[0087] ├─3.2 Properties of Functions (Intermediate Training Knowledge Elements)

[0088] │├─Prerequisite knowledge element: 3.1 The concept of a function

[0089] │├─Core knowledge points: Monotonicity, parity, periodicity

[0090] │├─Normalized Name: Properties of Functions

[0091] │├─Brief Description: The basic characteristics and rules of change of a function within its domain

[0092] │└─Initial Difficulty Level: 0.6

[0093] └─3.3 Basic Elementary Functions (Intermediate Training Knowledge Elements)

[0094] ├─Prerequisite knowledge elements: 3.1 The concept of a function, 3.2 Properties of functions

[0095] ├─Core knowledge points: Power functions, exponential functions, logarithmic functions, trigonometric functions

[0096] ├─ Normalized Name: Basic Elementary Functions

[0097] ├─Brief description: The most basic and commonly used type of function in mathematics

[0098] └─Initial Difficulty Level: 0.5

[0099] By using deep learning models to accurately identify the content structure of documents at the chapter, section, and subsection levels, and to analyze the progressive, parallel, and subordinate relationships between these structures, core knowledge points such as definitions, theorems, formulas, and rules are extracted in the order of these relationships. Corresponding explanations are generated and bound to standardized names to form training knowledge elements. This not only achieves the structured and logical extraction of core knowledge points and ensures the accuracy, standardization, and relevance of training knowledge elements, but also provides a high-quality knowledge foundation with clear logic and well-defined definitions for the subsequent construction of knowledge graphs.

[0100] In this embodiment, the step of calculating the difficulty coefficient based on the frequency of occurrence, content length, and core content also includes the following sub-steps:

[0101] The selection of standard elements follows three principles: disciplinary consensus, fundamental universality, and data typicality. Priority is given to selecting training knowledge elements from the training knowledge elements that are at the most basic level of the knowledge graph, support the learning of most subsequent knowledge elements, and exhibit typical occurrence characteristics in structured data, serving as unified standard benchmarks. Specifically, standard elements can be determined through a combination of expert annotation and algorithmic selection. For example, in the training knowledge element set of mathematics, "basic concepts of sets" are selected as standard elements. Simultaneously, three indicators for standard elements are precisely defined and quantified: standard frequency is the number of independent occurrences of the standard element in the entire structured data document, excluding elements that are excluded during the statistical process. Non-core mentions such as annotations, citations, and repetitions are excluded; the standard length is the total number of characters of the core explanation content corresponding to the standard element, removing redundant content such as format marks, transition sentences, and example auxiliary explanations; the standard degree is the core metric value of the standard element in the knowledge graph. Its core degree calculation comprehensively considers the position in the knowledge graph (core nodes are assigned a high value, and extended nodes are assigned a low value), the abstraction level of the concept (the lower the abstraction level, the stronger the basic attribute, and the higher the basic core degree value), and the dependency depth (the more subsequent training knowledge elements that depend on this standard element, the greater the core degree weight ratio). It is calculated through a preset weighted algorithm, and the value range is uniformly standardized as [0,1].

[0102] Based on the training knowledge elements that need to be calculated, the ratio of the number of occurrences to the standard number of occurrences is calculated as the frequency ratio, the ratio of the content length to the standard length is calculated as the length ratio, and the ratio of coreness to standardness is calculated as the core ratio.

[0103] Among them, the three original indicators for the current training knowledge element are all calculated using the ratio method corresponding to the standard element's indicator, achieving normalization processing without unitization, effectively eliminating the dimensional differences between different indicators, and providing a unified numerical benchmark for subsequent weighted average calculation. The specific statistical rules are consistent with the standard element, that is, the occurrence frequency of the current training knowledge element is its independent occurrence frequency in the structured data, the content length is the total number of characters in its core explanation content, and the coreness is its comprehensive quantitative value in the knowledge graph. At the same time, in order to avoid the interference of extreme outliers on the calculation results, uniform upper and lower limit threshold constraints are set for the obtained frequency ratio, content length ratio, and coreness ratio, limiting them to a reasonable range of [0.2, 5.0]. If the calculation result exceeds this range, the critical value of the range is automatically taken as the final ratio.

[0104] The difficulty coefficient is calculated by weighting the ratio of frequency, length, and core content.

[0105] The final calculation of the difficulty coefficient is performed using a weighted average algorithm. A scientific weight allocation strategy is preset, and users can customize the adjustment according to different teaching scenarios to adapt to diverse difficulty assessment standards. Specifically, this implementation prioritizes a basic weight allocation scheme: the weight of the frequency ratio is 0.2, the weight of the length ratio is 0.3, and the weight of the core ratio is 0.5. This weight allocation fully reflects the dominant influence of the core value of knowledge elements in the knowledge graph on the difficulty, while also taking into account their appearance characteristics and the auxiliary role of the explanation length. The weighted average calculation formula is: Difficulty coefficient = (frequency ratio × weight 1 + length ratio × weight 2 + core ratio × weight 3) / (weight 1 + weight 2 + weight 3). Furthermore, the calculated difficulty coefficient is normalized twice and mapped to the standard interval [0,1], where 0 represents the simplest basic training knowledge element and 1 represents the most difficult advanced training knowledge element, which facilitates accurate adaptation of the difficulty of subsequent training content.

[0106] To illustrate the above implementation process more clearly, specific embodiments are described below:

[0107] Standard element selection: The "basic concept of sets" in mathematics is selected as the standard element, with a standard degree of 12, a standard length of 800 characters, and a standard degree of 0.6;

[0108] Current element index: "The concept of function" is selected as the training knowledge element whose difficulty coefficient needs to be calculated. It appears 8 times independently, the core explanation is 1200 characters long, and the coreness is 0.8.

[0109] Ratio calculation: Frequency ratio = 8 / 12 ≈ 0.67, Length ratio = 1200 / 800 = 1.5, Core ratio = 0.8 / 0.6 ≈ 1.33;

[0110] Weighted average calculation: Using the basic weight allocation scheme (0.2, 0.3, 0.5), the difficulty coefficient = (0.67×0.2+1.5×0.3+1.33×0.5) / (0.2+0.3+0.5)≈1.21;

[0111] Normalization: Mapping 1.21 to the interval [0,1], the final difficulty coefficient of the "concept of function" is 0.605.

[0112] By establishing a unified benchmark by selecting standard elements from the training knowledge elements, the corresponding ratios of the current training knowledge elements to the standard elements in terms of frequency of occurrence, content length, and coreness are calculated. The difficulty coefficient is then obtained by weighted averaging. This not only achieves the quantification and standardization of difficulty calculation and effectively avoids the interference of subjective factors in difficulty judgment, but also takes into account the occurrence characteristics of knowledge elements and the core value in the knowledge graph through the comprehensive consideration of multi-dimensional indicators, making the calculation of the difficulty coefficient more objective, accurate, and reasonable.

[0113] In this embodiment, the steps of calculating the knowledge coverage coefficient based on the number of training sessions and the knowledge graph, and calculating the knowledge depth coefficient based on the training difficulty, training type, and knowledge graph, further include the following sub-steps:

[0114] Obtain the number of clusters of training knowledge elements extracted from the knowledge graph, and calculate the ratio of the number of training elements to the number of clusters as the knowledge coverage coefficient.

[0115] The clustering of training knowledge elements strictly follows the principle of "core priority and logical connection," directly serving the knowledge point coverage priority strategy: core training knowledge elements (such as basic subject concepts and core theorems) in the knowledge graph are independently divided into first-level clusters, with each first-level cluster corresponding to a must-test point; extended training knowledge elements (such as derivative applications and auxiliary explanations) are integrated into second-level clusters according to their subordinate / progressive relationship with the core elements, ensuring that each core knowledge point corresponds to at least one independent cluster; the number of clusters is the total number of first-level and second-level clusters after division, and redundant clusters with duplicate associations are removed during statistics to ensure data accuracy. The core significance of the knowledge coverage coefficient lies in quantifying the coverage ability of the number of training sessions for knowledge clusters: when the coefficient is ≥1, it indicates that the number of training sessions is sufficient to cover all knowledge clusters, and the system will trigger the strategy of "core cluster priority allocation + extended cluster supplementation as needed," allocating at least one training content to each first-level cluster, and allocating the remaining training slots according to the cluster importance ratio; when the coefficient is <1, it indicates that the number of training sessions is limited, and the system will prioritize the coverage of first-level clusters, selectively covering second-level clusters according to knowledge relevance, ensuring that no core knowledge points are omitted. Meanwhile, to facilitate the execution of subsequent strategies, the knowledge coverage coefficient is normalized to the [0,2] interval, where [0,1) corresponds to "basic coverage" and [1,2] corresponds to "full coverage".

[0116] Obtain the difficulty coefficient of the training knowledge element in the knowledge graph, calculate the average of all difficulty coefficients as the reference difficulty, and calculate the ratio of the training difficulty to the reference difficulty as the difficulty ratio.

[0117] The difficulty coefficient of the training knowledge elements is the standardized quantification result from the aforementioned steps (values ​​[0,1]). This implementation combines a difficulty grading generation strategy, mapping it to a three-level difficulty system: easy questions (difficulty range [0,0.3], corresponding to basic concepts and definitions); medium questions (difficulty range [0.3,0.7], corresponding to conceptual understanding and simple application); and difficult questions (difficulty range [0.7,1], corresponding to comprehensive application and in-depth understanding). The reference difficulty is the arithmetic mean of the difficulty coefficients of all training knowledge elements in the knowledge graph, objectively reflecting the overall basic difficulty level of the current knowledge system, and serving as the training input for the user. Difficulty provides a unified comparison benchmark (the training difficulty input by the user must correspond to this three-level system; for example, inputting "medium" will map to 0.5). The difficulty ratio is calculated by "user's target difficulty / reference difficulty," achieving de-unitized normalization: when the ratio > 1, it indicates that the user's required difficulty is higher than the overall basic level, and the system will dynamically adjust the proportion of difficulty levels (e.g., lowering the proportion of basic and increasing the proportion of difficult); when the ratio < 1, it indicates that the user's required difficulty is lower than the overall basic level, and the system will increase the proportion of easy questions and decrease the proportion of difficult questions; when the ratio = 1, the preset basic proportion is maintained (30% easy, 50% medium, and 20% difficult).

[0118] Obtain the main type of the training knowledge element in the knowledge graph. The main type corresponds to a type amplitude. Calculate the average of all type amplitudes as the reference amplitude. Calculate the ratio of the type amplitude of the training type to the reference amplitude as the type ratio.

[0119] The training knowledge elements are categorized into four main types, each linked to an optimal question type: conceptual knowledge (multiple choice, true / false, and fill-in-the-blank), comprehension knowledge (short answer and open-ended questions), application knowledge (calculation and comprehensive questions), and material analysis knowledge (question stem questions, i.e., reading material plus multiple sub-questions). To quantify these types, each category is assigned a unique amplitude, positively correlated with the depth and complexity of the question type: 1 for conceptual knowledge, 2 for comprehension knowledge, 3 for application knowledge, and 4 for material analysis knowledge, ensuring that type differences can be quantified numerically. The reference amplitude is the arithmetic mean of the amplitudes of all training knowledge element categories, reflecting the overall level of question type adaptation within the current knowledge system. The training type input by the user (such as "short answer") corresponds to the type amplitude of its respective subject type ("short answer" corresponds to comprehension knowledge, amplitude 2). The type ratio = training type amplitude / reference amplitude, realizing quantitative comparison between different types: when the ratio is >1, the depth of the question type representing the user's needs is higher than the overall basic level, and the system will strengthen the application layer examination of the corresponding knowledge when generating; when the ratio is <1, the depth of the question type representing the user's needs is lower, and the system focuses on the basic concept layer examination.

[0120] The knowledge depth coefficient is calculated by weighting the difficulty ratio and the type ratio.

[0121] The knowledge depth coefficient is calculated using a weighted average algorithm, with a pre-set scientific basic weight allocation scheme that allows users to customize adjustments based on teaching scenarios: the difficulty ratio weight is 0.6 (emphasizing the dominant influence of difficulty on knowledge depth), and the type ratio weight is 0.4 (taking into account the auxiliary role of question type suitability on depth). The calculation formula is: Knowledge Depth Coefficient = (Difficulty Ratio × 0.6 + Type Ratio × 0.4) / (0.6 + 0.4). Furthermore, the calculation results are normalized to the standard interval [0,1], where [0,0.3] corresponds to "basic depth" (emphasizing concept memorization), (0.3,0.7] corresponds to "medium depth" (emphasizing understanding and application), and (0.7,1] corresponds to "advanced depth" (emphasizing comprehensive innovation). This coefficient directly serves as the core depth basis for subsequent initial content generation, ensuring that the depth of the training content is highly matched with user needs and knowledge characteristics.

[0122] To illustrate the implementation process more clearly, specific examples are provided below:

[0123] Knowledge coverage coefficient calculation: There are 5 first-level clusters and 3 second-level clusters in the knowledge graph, with a total of 8 clusters; the user inputs 12 training questions, and the knowledge coverage coefficient = 12 / 8 = 1.5 (1.5 after normalization), which is considered "sufficient coverage". The system assigns 1 question to each of the 5 first-level clusters, and the remaining 7 questions are assigned according to the importance of the clusters.

[0124] Difficulty ratio calculation: The average difficulty coefficient of all training knowledge elements in the knowledge graph is 0.45 (reference difficulty). The user inputs the training difficulty as "medium" (mapped to 0.5). The difficulty ratio = 0.5 / 0.45≈1.11;

[0125] Type ratio calculation: The average type amplitude of all training knowledge elements in the knowledge graph is 2.2 (reference amplitude). When the user inputs the training type "Comprehensive Question" (corresponding to applied knowledge, type amplitude 3), the type ratio = 3 / 2.2 ≈ 1.36.

[0126] Knowledge depth coefficient calculation: (1.11×0.6+1.36×0.4)=1.202, normalized to 0.601, corresponding to "medium depth". The system will generate comprehensive questions that focus on understanding and application, and the difficulty will be slightly higher than the overall basic level.

[0127] By quantifying the knowledge coverage coefficient using the ratio of the number of clusters of training knowledge elements in the knowledge graph to the number of training elements, and combining the average difficulty coefficient and the average amplitude of the subject type of the training knowledge elements in the knowledge graph, the ratio of training difficulty to reference difficulty and the ratio of training type amplitude to reference amplitude are calculated respectively. Then, a weighted average is used to obtain the knowledge depth coefficient. This not only achieves precise quantitative calculation of the knowledge coverage coefficient and the knowledge depth coefficient, but also ensures the objectivity and rationality of the coefficient calculation by relying on the knowledge graph. It takes into account the comprehensiveness and depth adaptability of the knowledge coverage of the training content, reduces the interference of subjective factors on parameter determination, and provides quantitative support for the implementation of the three generation strategies.

[0128] In this embodiment, the step of generating initial content using structured data based on a large language model according to the number of training samples, training type, knowledge coverage coefficient, and knowledge depth coefficient further includes the following sub-steps:

[0129] Extract the first question stem information from the first text frame, and use the first question stem information to guide the large language model to generate initial content for selection;

[0130] The first text framework is a structured template library specifically adapted for generating choice-based content. It contains multiple standardized question templates validated in educational scenarios. Each template includes four core modules: question type identifier, knowledge point association, difficulty matching, and expression standardization. The first question stem information is the target template content extracted from this template library. The extraction logic is as follows: based on the previously calculated knowledge coverage coefficient, knowledge depth coefficient, and the user-input training type (choice-based), the optimal template with a knowledge point matching degree ≥85% and a difficulty matching range consistent with the knowledge depth coefficient is selected from the template library. To accurately guide the large language model in generating compliant initial choice-based content, the first question stem information will be integrated into a standardized Prompt template, as shown in the following example:

[0131] "You are a professional question-setting expert in the relevant field. Please generate one multiple-choice training question based on the following parameters:"

[0132] Core knowledge points: {Standardized name of training knowledge element + concise description};

[0133] Target difficulty: {Difficulty level based on knowledge depth coefficient mapping, such as easy / medium / hard};

[0134] Knowledge point coverage requirement: It is necessary to associate {number of associated knowledge points corresponding to the knowledge coverage coefficient} related knowledge elements;

[0135] Expression guidelines: The question stem should be 80-150 characters long, and there should be 4 options (1 correct answer + 3 distractors). The distractors should be similar to the correct answer but without logical ambiguity, and should not contain expressions that directly conflict with the basic content data.

[0136] Structured output requirements: Output in the format of "Question stem + Option A + Option B + Option C + Option D + Correct answer";

[0137] The preferred large language model is a specialized model fine-tuned from educational multiple-choice question generation data, such as the fine-tuned version of GPT-4 or the industry-customized version of Llama3. During generation, the core knowledge point explanations and related knowledge element information in the structured data are used as supplementary context inputs. At the same time, the model's "expression standard constraint plugin" is enabled to force the generated content to conform to the format requirements and expression standards of the first text framework.

[0138] The character count of the previous first question stem is the first character count, and the character count of the current first question stem is the second character count;

[0139] The character count strictly follows a unified standard: only valid characters in the core expression of the question stem are counted, excluding non-substantive expression characters such as punctuation marks, format marks, and special symbols. At the same time, fixed expression characters such as "Which of the following statements is correct?" and "According to the above knowledge points, which of the following options is incorrect?" are removed. Only the differentiated expression parts related to knowledge points are counted to ensure that the character count of the question stems before and after are comparable. If it is the first multiple-choice initial content generated in the current training task, the standard character count of the template of the difficulty level in the first text frame (e.g., the standard character count of the medium difficulty template is 120 characters) is taken as the first character count to provide a benchmark for the subsequent character count ratio calculation.

[0140] The ratio of the second character count to the first character count is calculated as the character count ratio. The weight of the difficulty ratio is adjusted based on the negative correlation of the character count ratio.

[0141] The character count ratio is calculated using division to two decimal places. Its core function is to dynamically reflect the difference between the current question stem length and the historical / standard question stem length. The core logic of the "negative correlation adjustment" mechanism is as follows: the longer the question stem has characters, the higher the cognitive load required for the user to understand it. If the original difficulty ratio weight is maintained, the overall difficulty of the actual training content will exceed expectations. Therefore, the weight of the difficulty ratio needs to be reduced to balance the overall difficulty. Conversely, the shorter the question stem has characters, the lower the cognitive load. The weight of the difficulty ratio can be appropriately increased to avoid the overall difficulty being too low.

[0142] This implementation method presupposes specific negative correlation adjustment rules:

[0143] When the character count ratio is in the range of [0.8, 1.2] (the length of the question stem is the same as the benchmark): the difficulty ratio weight remains at the initial value of 0.6;

[0144] When the character ratio is ∈ (1.2, 1.5] (the question is relatively long): the difficulty ratio weight is dynamically adjusted according to "initial value × (2 - character ratio)". For example, when the character ratio is 1.3, the weight = 0.6 × (2 - 1.3) = 0.42;

[0145] When the character count ratio is ∈ (1.5, +∞) (the question stem is too long): the difficulty ratio weight is fixedly reduced to 0.3, and the expression simplification instruction of the large language model is triggered, requiring the question stem to be regenerated and the character count to be controlled;

[0146] When the character ratio is ∈ [0.5, 0.8) (the question is relatively short): the difficulty ratio weight is dynamically adjusted upwards according to "initial value × (1 + (0.8 - character ratio))". For example, when the character ratio is 0.6, the weight = 0.6 × (1 + (0.8 - 0.6)) = 0.72;

[0147] When the character count ratio is ∈ (0, +∞, 0.5) (the question stem is too short): the difficulty ratio weight is fixedly increased to 0.8, and the expression supplementation instruction of the large language model is triggered at the same time, requiring the logical chain of the question stem to be enriched without redundancy;

[0148] The adjusted difficulty ratio weight will be fed back to the generation parameters of the large language model in real time to optimize the difficulty control of subsequent multiple-choice initial content, ensuring that the "question length + actual difficulty" of each content comprehensively adapts to user needs.

[0149] To illustrate the above implementation process more clearly, specific embodiments are described below:

[0150] First question information extraction: The user needs to generate a multiple-choice question of medium difficulty in mathematics, with a knowledge depth coefficient of 0.66 (mapped to medium difficulty). Extract the corresponding template from the first text frame as the first question information (standard character count 120 characters, initial difficulty ratio weight 0.6).

[0151] First content generation: Since there is no historical question stem, the default first character count is 120 characters, the second character count of the current first question stem information is 110 characters, the character count ratio is 110 / 120≈0.92 (∈[0.8,1.2]), the difficulty ratio weight remains at 0.6, and the large language model generates the initial content of question stem + 4 options according to the Prompt template;

[0152] The second content generation: extract the new first question stem information (match the next training knowledge element), the second character count = 140 characters, the first character count = 110 characters (the effective character count of the previous question stem), the character count ratio = 140 / 110≈1.27 (∈(1.2,1.5]), the difficulty ratio weight = 0.6×(2-1.27)=0.438, the large language model generates difficulty-adapted initial content according to the adjusted weight, to avoid the overall difficulty being too high due to the increased length of the question stem.

[0153] By extracting the first question stem information from the first text framework to guide the large language model in generating selection-based initial content, the standardization and directionality of the initial content generation are ensured. At the same time, by calculating the character ratio of adjacent first question stem information, dynamic monitoring of the length of the question stem description is achieved. Furthermore, relying on the negative correlation adjustment mechanism between the character ratio and the difficulty ratio weight, the influence of difficulty factors on the initial content can be flexibly adjusted according to changes in the character size of the question stem, effectively avoiding the difficulty judgment deviation caused by the question stem description being too long or too short, and further optimizing the balance between the standardization of question stem description and the difficulty suitability of the initial content.

[0154] In this embodiment, the step of generating initial content using structured data based on a large language model according to the number of training samples, training type, knowledge coverage coefficient, and knowledge depth coefficient further includes the following sub-steps:

[0155] Extract the second question stem information from the first text frame, and use the second question stem information to guide the large language model to generate short-answer type initial content;

[0156] The first text framework includes a standardized template library for short-answer questions. The templates are divided into three sub-libraries based on knowledge depth coefficients: basic, comprehension, and application. Each sub-library's templates contain four core modules: knowledge point anchoring fields, question logic fields, answer requirement fields, and length constraint fields. This ensures that the generated content aligns with the core requirements of short-answer questions: "testing conceptual understanding, logical explanation, and simple application." The second question stem information extraction logic is as follows: based on the user's input training type (short-answer), knowledge coverage coefficient (number of associated knowledge points), and knowledge depth coefficient (difficulty level), the optimal template with a fit of ≥90% is selected from the corresponding sub-library. Furthermore, a "random rotation mechanism" is used when using templates to avoid the continuous use of the same set of fixed question formats.

[0157] To accurately guide the large language model in generating short-answer initial content that meets the requirements, the information in the second question stem will be integrated into a structured Prompt template, as shown in the following example:

[0158] "You are a question-setting expert in the {corresponding field}. Please generate one short-answer practice question based on the following parameters:"

[0159] Core knowledge points: {Standardized names of training knowledge elements + detailed explanations of key information};

[0160] Related knowledge points: must cover {number corresponding to the knowledge coverage coefficient} related secondary knowledge elements;

[0161] Target difficulty: {Levels based on knowledge depth coefficient mapping, such as basic / understanding / application};

[0162] Expression guidelines: The question stem should be presented in the form of a question (such as "Briefly describe the core features of XX" or "Analyze the applicable conditions of XX"), and the number of characters should be controlled between 60 and 100. Avoid overly complicated background information and directly point to the core knowledge being tested.

[0163] Answer requirements: Clearly state the number of core scoring points in the answer (2-4), guide users to answer around the essence of the knowledge points, and avoid irrelevant statements;

[0164] Structured output requirements: Generate in the format of "question stem + key scoring point hints (implicit output for subsequent verification)";

[0165] The preferred large language model is a specialized model fine-tuned from educational short-answer question generation data (such as Claude3 Education Customized Edition or ERNIE4.0 Industry Fine-tuned Edition). During generation, the logical derivation process of core knowledge points in the structured data and the connection relationship of related knowledge elements are used as supplementary context inputs. At the same time, the "question logic optimization plugin" is enabled to ensure that the question stem is accurate, unambiguous, and conforms to the subject expression habits.

[0166] The number of characters in the current first question stem is the second number of characters;

[0167] The character count follows a unified standard for effective expression: only characters in the question stem that are directly related to the knowledge points are counted, excluding fixed question prefixes such as "Please answer briefly," "Short description," and "Analyze and explain," punctuation marks, special marks, and redundant characters for repetition and emphasis, ensuring that the character count accurately reflects the core expression length of the question stem; if the current question is the first short-answer type initial content generated in this training task, the standard character count of the corresponding knowledge depth coefficient template (e.g., the standard character count of the comprehension template is 80 characters) is taken by default as the reference benchmark for the second character count, providing a basis for subsequent difference ratio calculation.

[0168] Obtain the distinguishing characters between the previous first question stem information and the current first question stem information, and calculate the proportion of the number of different characters in the number of second characters as the distinguishing ratio;

[0169] The acquisition of distinguishing characters is achieved through a high-precision string comparison algorithm (such as an optimized version of the Levenshtein distance algorithm): the system compares the effective description of the previous first question stem with the current first question stem character by character, identifies semantically differentiated characters (excluding non-substantial differences such as synonym substitution and word order fine-tuning), and counts the total number of these characters as the distinguishing character count; the formula for calculating the distinguishing ratio is: distinguishing ratio = distinguishing character count ÷ second character count (the result is rounded to two decimal places). Its core function is to quantify the degree of difference between the previous and subsequent question stem descriptions, providing data support for subsequent type ratio weight adjustment; if it is the initial content of the first short answer type, the default distinguishing ratio is 1.0 (assuming there is no historical description to compare, it is considered completely differentiated), to avoid logical gaps in adjustment.

[0170] The weight of the type ratio is adjusted based on the positive correlation between the difference ratio and the type ratio.

[0171] The core logic of "positive correlation adjustment" is as follows: the higher the difference ratio, the greater the difference between the statements in the preceding and following questions. It is not necessary to rely too much on the type ratio weight to ensure content diversity. The type ratio weight can be appropriately reduced to focus on knowledge coverage and difficulty matching. The lower the difference ratio, the higher the repetition of the statements in the preceding and following questions. It is necessary to increase the type ratio weight and strengthen the differentiated guidance of question type adaptability to avoid redundancy in expression and overlap in examination direction.

[0172] This implementation method pre-defines a clear positive correlation adjustment rule, as follows:

[0173] When the difference ratio ∈ [0.8, 1.0] (representing high differentiation): the type ratio weight maintains the initial base value of 0.4, prioritizing the adaptation of the knowledge coverage coefficient and the knowledge depth coefficient;

[0174] When the difference ratio ∈ [0.5, 0.8) (moderate differentiation in expression): the weight of the type ratio is increased to 0.5, and the risk of repetition is further reduced by strengthening the differentiated design of question type adaptation;

[0175] When the difference ratio ∈ [0.3, 0.5) (representing low degree of differentiation): the type ratio weight is increased to 0.6, and the "question logic switching instruction" of the large language model is triggered, requiring a change in the question angle, such as switching from "briefly describe the features" to "analyze the causes";

[0176] When the difference ratio ∈ [0, 0.3) (serious repetition of expression): the type ratio weight is increased to 0.7, and the template reselection mechanism is triggered to re-select templates with different question logics from the sub-library corresponding to the first text framework and generate a new expression of the question stem;

[0177] The adjusted type ratio weights will be synchronized to the generation parameters of the large language model in real time, ensuring that the initial content of each short answer question not only meets the requirements of the question type, but also effectively avoids repetition of expressions, thereby improving the diversity and relevance of the overall content.

[0178] To illustrate the above implementation process more clearly, specific embodiments are described below:

[0179] Second question stem information extraction: Users need to generate comprehension-based short answer content for physics, with a knowledge coverage coefficient of 2 (associating with 2 secondary knowledge points) and a knowledge depth coefficient of 0.6 (mapped to comprehension difficulty). Templates are extracted from the comprehension sub-library of the first text framework, and the second question stem information is integrated into the corresponding Prompt template. The effective character count (second character count) of the current question stem is 90 characters.

[0180] Historical question stem comparison: The previous short answer question stem had 85 effective characters, and the string comparison algorithm identified 34 different characters.

[0181] Discrimination ratio calculation: Discrimination ratio = 34 ÷ 90 ≈ 0.38 (∈ [0.3, 0.5), indicating low degree of differentiation).

[0182] Weight adjustment: Adjust the type ratio weight to 0.6 according to the rules, and trigger the "question logic switching instruction" at the same time;

[0183] Initial content generation: The large language model generates short-answer initial content based on the adjusted weights and instructions, with the question logic of "analyze the causes and core influencing factors of the XX phenomenon". This not only meets the difficulty requirements of the comprehension type, but also avoids repetition with the previous question "briefly describe the characteristics of the XX phenomenon".

[0184] By extracting information from the second question stem from the first text framework to guide the large language model in generating short-answer initial content, the standardization and question type adaptability of the generated short-answer training content are ensured. At the same time, by calculating the difference ratio of the number of distinguishing characters in adjacent first question stem information to the number of characters in the current question stem, the degree of differentiation in question stem expression is accurately quantified. Furthermore, relying on the positive correlation adjustment mechanism between the difference ratio and the type ratio weight, the influence weight of question type factors on the initial content can be flexibly adjusted according to the degree of differentiation in question stem expression. This effectively avoids excessive repetition of question stem expressions of the same type of question, improves the diversity and relevance of the initial content, and further optimizes the balance between question type adaptability and expression richness in the short-answer initial content.

[0185] In this embodiment, the step of applying a preset second text framework to perform full content verification on the initial content, generating verification data, and then correcting the initial content based on the verification data to obtain the training content further includes the following sub-steps:

[0186] Based on the second text framework, at least one verification dimension is matched from the preset verification feature library;

[0187] The second text framework is a standardized verification process template library adapted to multiple training types (selective, short answer, calculation, judgment, etc.). It has built-in core configuration items such as verification priority sorting, judgment criteria for each dimension, verification data output format, and correction process trigger conditions, which can realize differentiated adaptation and unified management of verification processes for different question types. The preset verification feature library includes five core verification dimensions, namely content correctness, content repetition, question rationality, expression clarity, and knowledge adaptability. At the same time, it supports users to customize and add special verification dimensions according to specific teaching scenario needs (such as K12 basic education, higher education professional courses, vocational skills training, etc.), such as parsing completeness dimension and question type compliance dimension. The matching logic for verification dimensions adopts a dual mechanism of "question type adaptation + priority sorting": First, basic verification dimensions are automatically matched based on the training type of the initial content. Then, at least one core verification dimension is selected for verification according to a fixed priority of "content correctness > content repetition > question rationality > expression clarity > knowledge adaptability". For example, multiple-choice initial content is matched by default with four core dimensions: "content correctness + content repetition + question rationality + expression clarity". Short-answer initial content is matched by default with four core dimensions: "content correctness + content repetition + expression clarity + knowledge adaptability", ensuring that the core quality indicators of the training content are verified first.

[0188] If the verification dimension includes content correctness, the correlation between the initial content and the basic content data is verified using a pre-defined preliminary large language model. If the correlation is within a pre-defined range, the verification passes; otherwise, the verification fails. Here, 40% ≤ correlation range ≤ 100%.

[0189] The pre-set preliminary large language model is preferably a high-precision language model that has been fine-tuned specifically for "quality verification of educational training content," such as the GPT-4Turbo educational verification customized version or the ERNIE4.0 quality control fine-tuned version. This type of model has higher accuracy and reliability for tasks such as knowledge point anchoring, answer correctness verification, and option rationality analysis of training content. This implementation uses a "two-stage cross-validation method" to verify the correctness of the content. The first stage is knowledge point anchoring verification. Based on the core knowledge point information in the basic content data, including standardized names, detailed explanations, applicable conditions, and related knowledge points, the model determines whether the question stem of the initial content accurately tests the target knowledge point and whether there is a core error of "question not matching the target." The second stage is answer and expression cross-validation. The model applies differentiated verification standards for different training types to ensure that the answers, options, and explanations of the initial content meet quality requirements.

[0190] During the verification process, the system automatically integrates the content to be verified with the basic reference data into a standardized Prompt template, which is then input into the preparatory large language model. A specific example of the Prompt template is as follows:

[0191] As a professional knowledge verification expert in the corresponding subject area, please strictly verify the correctness of the initial training content according to the following requirements, and point out any problems found:

[0192] Basic reference data: {Standardized names, detailed explanations, applicable conditions, and core related knowledge points of the target training knowledge elements}

[0193] Initial content to be verified:

[0194] Title: {generated_question};

[0195] Answer: {generated_answer};

[0196] Analysis: {generated_explanation};

[0197] Options: {generated_options} (Only optional content is available);

[0198] Verification points:

[0199] (1) Whether the question accurately tests the target knowledge point and whether the expression is clear and unambiguous;

[0200] (2) Whether the answer is correct and whether it is consistent with the basic reference data; for calculation problems, it is necessary to further verify the logic of the calculation process and the accuracy of the calculation results;

[0201] (3) If it is a multiple choice question, it must meet the configuration requirement of 4 options (A / B / C / D), including at least 1 correct answer and 1 incorrect option. The incorrect option must be sufficiently distracting, without obvious logical errors, and there must be no multiple correct answers or no correct answer.

[0202] (4) Whether the analysis is complete and clear, whether it can accurately explain the basis of the correct answer, and whether it can clearly point out the reason for the error of the wrong option;

[0203] Output requirements:

[0204] (1) Quantify the correlation between the initial content and the basic reference data, with a value range of 0%-100%;

[0205] (2) Provide a clear verification conclusion: pass or fail;

[0206] (3) If the verification fails, the specific problem type must be clearly indicated and targeted correction suggestions must be provided.

[0207] The quantification of relevance strictly follows preset calculation rules. For multiple-choice questions, in a four-option question, a correct answer is recorded as 1, a correct option as 1, and an incorrect answer as 0. Since the question is always correct and at least one option is correct, the relevance between the correct option and the question is at least 2 / 5 (40%) and at most 100%. In this implementation, the relevance range is limited to 40% ≤ relevance range ≤ 100%. Only when the relevance of the initial content falls within this range can the content be considered to have passed the correctness verification. This effectively ensures the close connection between the initial content and the core knowledge points, avoiding deviation from the core content of the teaching syllabus and textbooks, and also prevents the copying of basic content data, thus ensuring the originality and scientific nature of the training content.

[0208] If the verification dimension includes content duplication, the cosine similarity between the current initial content and the historically stored content is calculated based on the vector algorithm. If the cosine similarity is less than the preset reference similarity, the verification passes; otherwise, the verification fails.

[0209] In this embodiment, the vector algorithm used is a high-precision semantic embedding model based on the Transformer architecture, such as a fine-tuned version of Sentence-BERT for the education field. This model can transform information such as the question stem, core examination points, and options (selective content) of the initial content into high-dimensional semantic vectors (768-dimensional semantic vectors are preferred in this embodiment), which can more accurately capture the semantic information of the content and avoid the omission problem of "different forms but the same semantics" in traditional string comparison methods. The calculation logic of cosine similarity is as follows: First, the current initial content is transformed into a standardized semantic vector. Then, the semantic vectors of historical content with the same knowledge point, question type, and difficulty level are extracted from the system's historical training content repository. The cosine similarity between the current vector and the historical vector is calculated one by one. Finally, the highest cosine similarity is taken as the basis for determining the repetition of the initial content. In this embodiment, the preset reference similarity threshold is 85%. Users can customize and adjust this threshold within the range of 75%-90% according to the actual training content diversity requirements. When the calculated cosine similarity is less than the preset reference similarity, the content duplication verification is deemed to have passed. When the cosine similarity is greater than or equal to the preset reference similarity, the content duplication verification is deemed to have failed, and it is considered that the current initial content and the historical stored content have semantic duplication.

[0210] To improve the efficiency and accuracy of content duplication verification, the system has built a dedicated "semantic deduplication index library." This index library categorizes and manages the semantic vectors of historical training content according to three dimensions: knowledge point, question type, and difficulty level. When performing similarity calculations, it only needs to search and compare within the corresponding category index, avoiding meaningless cross-category comparisons and significantly improving verification efficiency. Furthermore, this index library supports real-time updates; after newly generated training content passes verification, its semantic vectors are added to the corresponding category index in real time, ensuring the accuracy of subsequent verifications.

[0211] To further improve the overall quality of the training content, this implementation method adds two derivative verification dimensions—question rationality verification and expression clarity verification—to the core verification dimensions, as a supplement to the core verification dimensions, as follows:

[0212] Question validity verification: Strictly adhere to the relevant requirements for question quality control, and implement differentiated verification standards for different training types; for multiple-choice questions, check whether there are obvious errors in the options, whether there are multiple correct answers, and whether the distractors are sufficiently distracting; for true / false questions, check whether the question stem contains absolute statements, such as words like "certainly," "absolutely," "inevitably," and "must," and whether the logic is rigorous; for fill-in-the-blank questions, check whether the answer is unique, and if there are multiple reasonable answers, whether this is clearly stated in the explanation; for short-answer questions, check whether the question stem clearly states the answer requirements and whether sufficient known conditions are provided.

[0213] Clarity of Expression Verification: Check whether the wording of the initial content, including the question stem, options, answers, and explanations, is clear, accurate, and unambiguous; check whether the use of terminology is consistent with the standardized names in the basic content data, and whether there are any typos or grammatical errors; check whether the wording of the question stem is concise and clear, whether there is any lengthy or redundant background information, and whether the core examination intent is clear; check whether the wording of the explanation is logically coherent, whether the steps are clear, and whether it is easy for learners to understand and master.

[0214] Correct or regenerate any parts of the initial content that fail validation (either manually or automatically).

[0215] This implementation method employs a collaborative correction strategy of "automatic correction first, manual correction as a backup" to ensure the efficiency and quality of the initial content correction.

[0216] Automatic Correction: When initial content verification fails, the system first triggers an automatic correction process. A pre-set large language model, based on the problem description, failure reasons, and correction suggestions in the verification data, performs targeted corrections on the failed parts of the initial content. For example, for multiple-choice questions that fail content correctness verification, the model will regenerate the question stem, adjust the logic of the options, and correct the answer or explanation. For questions that fail content repetition verification, the model will adjust the wording of the question stem, change the examples in the options, optimize the focus of the examination, and generate differentiated content. For questions that fail to verify reasonableness or clarity of expression, the model will directly correct ambiguous statements in the question stem, replace absolute terms, and supplement necessary information. After automatic correction, the system will re-perform a full-dimensional verification operation on the corrected content until verification passes. If the content still fails verification after three consecutive automatic corrections, the system will mark the initial content as "requiring manual intervention" and trigger a manual correction process.

[0217] Manual correction: The system pushes the initial content of the verification failure along with a detailed problem report to the user through a dedicated online editing platform. This report includes information such as the verification dimensions, reasons for failure, similarity calculation results, automatic correction records of the model, and suggested correction schemes. Users can comprehensively modify the initial content through the online editing platform, including the question stem, options, answers, and explanations. The platform supports real-time preview, allowing users to view the effect of the modifications in real time. Simultaneously, the platform provides a quality scoring function, assigning a quality score of 0-10 points and a corresponding difficulty tag to the modified content based on the compliance status of each verification dimension, assisting users in optimizing content quality.

[0218] Regardless of whether automatic or manual correction is used, the corrected training content must be output in a unified JSON format to ensure structured storage and subsequent retrieval and management of the content. The specific JSON output format is as follows:

[0219] {

[0220] "question":"The revised question content",

[0221] "options":["A. Revised option A","B. Revised option B","C. Revised option C","D. Revised option D"],#Required for multiple-choice questions

[0222] "answer":"The corrected answer",

[0223] "explanation":"Revised and detailed answer analysis",

[0224] "difficulty": "Revised actual difficulty label",

[0225] "knowledge_points":["core knowledge points","related knowledge point 1","related knowledge point 2"],

[0226] "quality_score": 9.5, # Corrected quality score

[0227] "verification_result":"Verification passed" #Final verification result

[0228] }

[0229] By relying on a second text framework to match at least one verification dimension from a pre-set verification feature library, multi-dimensional standardized full-process verification of initial content is achieved. For the content correctness dimension, a pre-prepared large language model is used to verify the correlation between the initial content and the basic content data. By limiting the correlation range to 40% ≤ correlation ≤ 100%, the close connection between the initial content and core knowledge points is effectively ensured, avoiding deviation from the core content of the teaching syllabus and textbooks. This also prevents the copying of basic content data, guaranteeing the originality and scientific rigor of the training content. For the content repetition dimension, the cosine similarity between the current initial content and historically stored content is accurately calculated based on a vector algorithm. A similarity threshold is used to effectively identify and avoid duplicate content, improving the diversity of training content. Simultaneously, targeted corrections or regeneration of failed verification parts in the initial content can accurately address quality defects. The final output training content can be directly incorporated into an online question bank, supporting functions such as filtering by question type and difficulty, online preview, and online editing, significantly improving the quality of training content generation and the efficiency of question bank construction.

[0230] This application also discloses a training content generation system based on a deep learning model, including a processor, wherein the processor executes the steps of the training content generation method based on a deep learning model as described in any of the preceding embodiments.

[0231] This application also discloses a storage medium storing a program that, when executed by a processor, implements the steps of the training content generation method based on a deep learning model described above.

[0232] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for generating training content based on a deep learning model, characterized in that, Includes the following steps: Acquire basic content data, and extract text content, table content, and formatted content from the basic content data to generate structured data; Based on a pre-defined deep learning model, semantic analysis is performed on structured data to identify content structure from the structured data, training knowledge elements are extracted from the structured data according to the content structure, the relationships between training knowledge elements are identified, and a knowledge graph is constructed based on the training knowledge elements and the relationships. Calculate the position of training knowledge elements in structured data, and calculate the frequency of occurrence, content length, and coreness of training knowledge elements in the knowledge graph based on their position. The difficulty level is calculated based on the frequency of occurrence, length of content, and core content. Obtain the input training parameters, which include the number of training samples, training difficulty, and training type. The knowledge coverage coefficient is calculated based on the number of training sessions and the knowledge graph, and the knowledge depth coefficient is calculated based on the training difficulty, training type, and knowledge graph. The system applies a pre-defined first text framework and generates initial content based on the large language model using structured data, taking into account the number of training sessions, training type, knowledge coverage coefficient, and knowledge depth coefficient. The initial content is fully validated using a pre-defined second text framework to generate validation data. The initial content is then corrected based on the validation data to obtain the training content. The steps of calculating the knowledge coverage coefficient based on the number of training samples and the knowledge graph, and calculating the knowledge depth coefficient based on the training difficulty, training type, and knowledge graph, also include the following sub-steps: Obtain the number of clusters of training knowledge elements extracted from the knowledge graph, and calculate the ratio of the number of training elements to the number of clusters as the knowledge coverage coefficient. Obtain the difficulty coefficient of the training knowledge element in the knowledge graph, calculate the average of all difficulty coefficients as the reference difficulty, and calculate the ratio of the training difficulty to the reference difficulty as the difficulty ratio. Obtain the main type of the training knowledge element in the knowledge graph. The main type corresponds to a type amplitude. Calculate the average of all type amplitudes as the reference amplitude. Calculate the ratio of the type amplitude of the training type to the reference amplitude as the type ratio. The knowledge depth coefficient is calculated by weighting the difficulty ratio and the type ratio.

2. The method for generating training content based on a deep learning model according to claim 1, characterized in that, The steps of extracting text content, table content, and formatted content from basic content data to generate structured data also include the following sub-steps: Obtain the type and format of basic content data; If the type format is within the preset first format range, the basic content data is converted into image sequence data. The image sequence data retains the page layout information of the basic content data. The graphic data in the image sequence data is extracted, and the text content, table content, and format content are identified from the graphic data. If the type format is within the preset second format range, then the text content, table content, and format content are directly separated from the basic content data; If the type format is outside the first format range and the second format range, a conversion request is issued, the converter corresponding to the conversion request is obtained, and the type format is converted using the converter to obtain a temporary format. If the temporary format is within the first format range or the second format range, the type format is added to the first format range, and the first format range is associated with the converter.

3. The method for generating training content based on a deep learning model according to claim 2, characterized in that, The steps of performing semantic analysis on structured data based on a pre-defined deep learning model, identifying content structure from structured data, and extracting training knowledge elements from structured data based on content structure also include the following sub-steps: The deep learning model is used to identify the chapters, sections and subsections of a document as its content structure. The headings and paragraphs at all levels in the content structure are extracted, and the relationships between the structures in the content structure are analyzed. The relationships include progressive, parallel and subordinate. Based on the order of association, definitions, theorems, formulas and rules are extracted from the content structure as core knowledge points, explanations of the core knowledge points are generated, keywords or terms related to the core knowledge points are extracted as standardized names of the core knowledge points, and training knowledge elements are obtained by binding the explanations with the standardized names.

4. The method for generating training content based on a deep learning model according to claim 3, characterized in that, The steps for calculating the difficulty coefficient based on frequency of occurrence, length of content, and core relevance also include the following sub-steps: Select standard elements from the training knowledge elements, and use the frequency of occurrence, content length, and coreness of the standard elements as the standard frequency, standard length, and standardness; Based on the training knowledge elements that need to be calculated, the ratio of the number of occurrences to the standard number of occurrences is calculated as the frequency ratio, the ratio of the content length to the standard length is calculated as the length ratio, and the ratio of coreness to standardness is calculated as the core ratio. The difficulty coefficient is calculated by weighting the ratio of frequency, length, and core content.

5. The method for generating training content based on a deep learning model according to claim 4, characterized in that, The steps for generating initial content using structured data based on a large language model, considering the number of training samples, training type, knowledge coverage coefficient, and knowledge depth coefficient, also include the following sub-steps: Extract the first question stem information from the first text frame, and use the first question stem information to guide the large language model to generate initial content for selection; The character count of the previous first question stem is the first character count, and the character count of the current first question stem is the second character count; The ratio of the second character count to the first character count is calculated as the character count ratio. The weight of the difficulty ratio is adjusted based on the negative correlation of the character count ratio.

6. The method for generating training content based on a deep learning model according to claim 4, characterized in that, The steps for generating initial content using structured data based on a large language model, considering the number of training samples, training type, knowledge coverage coefficient, and knowledge depth coefficient, also include the following sub-steps: Extract the second question stem information from the first text frame, and use the second question stem information to guide the large language model to generate short-answer type initial content; The number of characters in the current first question stem is the second number of characters; Obtain the distinguishing characters between the previous first question stem information and the current first question stem information, and calculate the proportion of the number of different characters in the number of second characters as the distinguishing ratio; The weight of the type ratio is adjusted based on the positive correlation between the difference ratio and the type ratio.

7. The method for generating training content based on a deep learning model according to claim 6, characterized in that, The process of applying a pre-defined second text framework to perform full content validation on the initial content, generating validation data, and then correcting the initial content based on the validation data to obtain the training content also includes the following sub-steps: Based on the second text framework, at least one verification dimension is matched from the preset verification feature library; If the verification dimension includes content correctness, the correlation between the initial content and the basic content data is verified using a pre-defined preliminary large language model. If the correlation is within a pre-defined range, the verification passes; otherwise, the verification fails. Here, 40% ≤ correlation range ≤ 100%. If the verification dimension includes content duplication, the cosine similarity between the current initial content and the historically stored content is calculated based on the vector algorithm. If the cosine similarity is less than the preset reference similarity, the verification passes; otherwise, the verification fails. Correct or regenerate any parts of the initial content that failed verification.

8. A training content generation system based on a deep learning model, characterized in that, The method includes a processor that performs the steps of the training content generation method based on a deep learning model as described in any one of claims 1-7.

9. A storage medium, characterized in that, The storage medium stores a program that, when executed by a processor, implements the steps of the training content generation method based on a deep learning model as described in any one of claims 1-7.

Images