Zero sample course knowledge extraction and filtering method and device and electronic equipment

Abstract

The invention provides a zero sample curriculum knowledge extraction and filtering method and device and electronic equipment, and the method comprises the steps: carrying out the blocking processing of a text of a to-be-processed teaching material, extracting knowledge points, obtaining a knowledge point hierarchical structure, carrying out the filtering of the knowledge points in the knowledge point hierarchical structure through reasoning judgment, and obtaining a zero sample curriculum knowledge extraction and filtering result. Obtaining a filtered knowledge point hierarchical structure; and based on a predefined knowledge type, classifying and marking the knowledge points in the filtered knowledge point hierarchical structure. According to the method, the integrity of the context is ensured through an iterative semantic partitioning strategy, a large model is guided to extract knowledge points of a hierarchical structure from a teaching perspective, low-quality contents are effectively filtered, and the teaching expressive power of a graph structure is further enhanced through semantic knowledge classification.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus and electronic device for zero-shot course knowledge extraction and filtering. Background Technology

[0002] In recent years, Large Language Models (LLMs) have been widely used for zero-shot knowledge extraction tasks due to their rich knowledge reserves and excellent instruction-following capabilities. For example, the ChatExtract framework can be used to automatically extract structured data through engineered prompts. Although LLMs can effectively identify key information in text, the extracted knowledge points are often scattered and flat, lacking hierarchical relationships such as superior-inferior or subordinate relationships. This makes it difficult for the constructed knowledge graph to support effective knowledge navigation and learning path reasoning, and thus fails to meet the needs of educational scenarios for systematic knowledge structures.

[0003] Knowledge point extraction is inherently challenging, requiring precise identification of knowledge points. To improve extraction quality, some studies have attempted to optimize model performance using existing data. For example, pre-training with structures enhances the language model's understanding of structures; joint named entity recognition and relation extraction methods are employed; or task-related instruction and pattern generation strategies are designed to improve the performance of few-shot relation extraction. Furthermore, self-suggestion frameworks provide contextual guidance to the model by generating synthetic samples, thereby improving the performance of zero-shot relation extraction. While these methods reduce reliance on labeled data to some extent, their core still depends on fine-tuning or data augmentation. When facing entirely new domains or textbooks, the lack of labeled data still limits the generality and flexibility of these methods. Summary of the Invention

[0004] This application provides a zero-shot course knowledge extraction and filtering method, apparatus and electronic device, which can address the challenges of structural deficiencies and low quality in current LLMs for zero-shot knowledge extraction.

[0005] Firstly, a zero-shot course knowledge extraction and filtering method is provided, including: The text of the teaching materials to be processed is divided into blocks and knowledge points are extracted to obtain a hierarchical structure of knowledge points, wherein the hierarchical structure of knowledge points includes knowledge points; By using reasoning and judgment, the knowledge points in the knowledge point hierarchy are filtered to obtain the filtered knowledge point hierarchy. Based on predefined knowledge types, the knowledge points in the filtered knowledge point hierarchy are classified and labeled.

[0006] The method provided in this application addresses issues such as inconsistent granularity, structural gaps, and semantic redundancy in the current course knowledge point extraction process. Based on the textbook's hierarchical structure, it ensures the integrity of the context through an iterative semantic segmentation strategy. Combined with Few-shot prompts and expert role settings, it guides the large model to extract hierarchical knowledge points from a teaching perspective. It introduces a six-dimensional reasoning mechanism to effectively filter low-quality content and further enhances the teaching expressiveness of the graph structure through semantic knowledge classification.

[0007] In one possible implementation, the step of segmenting the text of the teaching material to be processed into blocks and extracting knowledge points to obtain a hierarchical structure of knowledge points includes: An iterative semantic segmentation algorithm is used to segment the text of the teaching materials to be processed according to a preset hierarchical structure to obtain text blocks; Extract the knowledge points from the text block to obtain the knowledge point hierarchy structure, which includes block knowledge points and sub-knowledge points. The sub-knowledge points are conceptual units under the block knowledge points.

[0008] In one possible implementation, an iterative semantic segmentation algorithm is used to segment the text of the teaching material to be processed according to a preset hierarchical structure to obtain text blocks, including: Based on the hierarchical division of the text in the teaching material to be processed, the teaching material to be processed is decomposed into multiple text units; Start traversing each of the text units and determine whether the length of the text unit is greater than a preset length threshold; If so, the text unit is segmented using a recursive algorithm to obtain the text block; Otherwise, the text unit is treated as the text block.

[0009] In one possible implementation, the step of using a recursive algorithm to segment the text units to obtain the text blocks includes: Using semantic delimiters, the text units are segmented to obtain a list of multiple semantic units, which includes multiple semantic units; Traverse the list of semantic units, merge multiple semantic units into an accumulation block, and determine whether the length of the accumulation block is greater than a preset length threshold. If it is greater than the sum of its parts, the most recently merged semantic unit is removed from the accumulated block, and the accumulated block with the most recently merged semantic unit removed is taken as the text block. Otherwise, add the semantic unit to the accumulated block; Repeat the above steps to process all semantic units in the semantic unit list and obtain the text block.

[0010] In one possible implementation, the knowledge point hierarchy includes: in, It is the knowledge structure of the textbook. These are the sub-knowledge points mentioned. This refers to the knowledge points mentioned above. It is a summary of the knowledge points in the block. It is the first Individual knowledge points, This is the title of the sub-knowledge point. This refers to the content description of the sub-knowledge points.

[0011] In one possible implementation, the step of using reasoning to filter the knowledge points in the knowledge point hierarchy to obtain a filtered knowledge point hierarchy includes: The reasoning decision is expressed as a judgment function: in, Indicates the first Height knowledge points , It is the number of the sub-knowledge points. It is the text fragment of the teaching material to be processed corresponding to the sub-knowledge point; The sub-knowledge points are in the preset six dimensions. Evaluation results across multiple dimensions ; The result of logical AND operation on the preset six dimensions, when When all six preset dimensions are true. The value is true; where the preset six dimensions are content matching degree, structural completeness, concept support degree, judgment logic degree, case specificity degree, and learning benefit degree.

[0012] In one possible implementation, the step of classifying and labeling knowledge points in the filtered knowledge point hierarchy based on predefined knowledge types includes: Use the following annotation functions to annotate the knowledge points:

[0013] Indicates step-by-step reasoning This knowledge point represents the type of knowledge. This is the first sub-knowledge point. For context summary, The semantic type matched for the sub-knowledge point belongs to a predefined knowledge type. The predefined knowledge types include definitions, processes, and classifications. The definitions include explanations of the meaning of a term or concept. The processes include descriptions of the operational steps or time sequence of a task, technology, or system. The classifications include characterizing the components, subclasses, or subordinate structures of a whole. The reasoning steps include: understanding the meaning of knowledge points, analyzing summary content, matching predefined types, and outputting the final result. Understanding the meaning of knowledge points includes clarifying the semantic boundaries and teaching-oriented keywords of the sub-knowledge points. Analyzing the summary content includes extracting keywords with language structure features from the text fragments of the teaching materials to be processed corresponding to the sub-knowledge points, matching the keywords with the predefined knowledge types item by item to determine whether they conform to the language and structural features of the definition, process, and classification, and selecting the most matching predefined knowledge type.

[0014] Secondly, a zero-sample course knowledge extraction and filtering device is provided, the device comprising: The extraction module is used to segment the text of the teaching materials to be processed and extract knowledge points to obtain a hierarchical structure of knowledge points, wherein the hierarchical structure of knowledge points includes knowledge points; The filtering module is used to filter the knowledge points in the knowledge point hierarchy using reasoning and judgment to obtain the filtered knowledge point hierarchy. The annotation module is used to classify and annotate the knowledge points in the filtered knowledge point hierarchy based on predefined knowledge types.

[0015] In one possible implementation, the extraction module is used to segment the text of the teaching material to be processed according to a preset hierarchical structure using an iterative semantic segmentation algorithm to obtain text blocks; Extract the knowledge points from the text block to obtain the knowledge point hierarchy structure, which includes block knowledge points and sub-knowledge points. The sub-knowledge points are conceptual units under the block knowledge points.

[0016] In one possible implementation, the extraction module is used to decompose the teaching material to be processed according to the hierarchical division in the text to obtain multiple text units; Start traversing each of the text units and determine whether the length of the text unit is greater than a preset length threshold; If so, the text unit is segmented using a recursive algorithm to obtain the text block; Otherwise, the text unit is treated as the text block.

[0017] In one possible implementation, an extraction module is used to segment the text units using semantic delimiters to obtain a list of multiple semantic units, the list of semantic units including multiple semantic units; Traverse the list of semantic units, merge multiple semantic units into an accumulation block, and determine whether the length of the accumulation block is greater than a preset length threshold. If it is greater than the sum of its parts, the most recently merged semantic unit is removed from the accumulated block, and the accumulated block with the most recently merged semantic unit removed is taken as the text block. Otherwise, add the semantic unit to the accumulated block; Repeat the above steps to process all semantic units in the semantic unit list and obtain the text block.

[0018] In one possible implementation, the knowledge point hierarchy includes:

[0019] in, It is the knowledge structure of the textbook. These are the sub-knowledge points mentioned. This refers to the knowledge points mentioned above. It is a summary of the knowledge points in the block. It is the first Individual knowledge points, This is the title of the sub-knowledge point. This refers to the content description of the sub-knowledge points.

[0020] In one possible implementation, the filtering module is used to represent the inference decision as a decision function:

[0021] in, Indicates the first Height knowledge points , It is the number of the sub-knowledge points. It is the text fragment of the teaching material to be processed corresponding to the sub-knowledge point; The sub-knowledge points are in the preset six dimensions. Evaluation results across multiple dimensions ; The result of logical AND operation on the preset six dimensions, when When all six preset dimensions are true. The value is true; where the preset six dimensions are content matching degree, structural completeness, concept support degree, judgment logic degree, case specificity degree, and learning benefit degree.

[0022] In one possible implementation, the annotation module is used to annotate the knowledge points using the following annotation functions: Indicates step-by-step reasoning This knowledge point represents the type of knowledge. The first sub-knowledge point, For context summary, The semantic type matched for the sub-knowledge point belongs to a predefined knowledge type. The predefined knowledge types include definitions, processes, and classifications. The definitions include explanations of the meaning of a term or concept. The processes include descriptions of the operational steps or time sequence of a task, technology, or system. The classifications include characterizing the components, subclasses, or subordinate structures of a whole. The reasoning steps include: understanding the meaning of knowledge points, analyzing summary content, matching predefined types, and outputting the final result. Understanding the meaning of knowledge points includes clarifying the semantic boundaries and teaching-oriented keywords of the sub-knowledge points. Analyzing the summary content includes extracting keywords with language structure features from the text fragments of the teaching materials to be processed corresponding to the sub-knowledge points, matching the keywords with the predefined knowledge types item by item to determine whether they conform to the language and structural features of the definition, process, and classification, and selecting the most matching predefined knowledge type.

[0023] Thirdly, a computer storage medium is provided that stores instructions which, when executed by a computer, cause the computer to perform the method described in the first aspect.

[0024] Fourthly, a computer program product is provided, the computer program product storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect.

[0025] The beneficial effects of the second to fourth aspects can be referred to the introduction of the beneficial effects of the first aspect above, and will not be repeated here. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating the technical concept of a zero-sample knowledge extraction and filtering method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a zero-sample knowledge extraction and filtering method provided in an embodiment of this application; Figure 3 This is a schematic diagram illustrating a zero-sample knowledge extraction and filtering method provided in an embodiment of this application; Figure 4 This is a schematic diagram of the hierarchical knowledge construction result in a zero-sample knowledge extraction and filtering method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a zero-sample knowledge extraction and filtering device provided in an embodiment of this application; Figure 6This is a schematic diagram of the hardware structure of a zero-sample knowledge extraction and filtering device provided in an embodiment of this application. Detailed Implementation

[0027] The solutions provided in the embodiments of this application will now be described with reference to the accompanying drawings. In the embodiments of this application, "multiple" refers to two or more objects, and "various types" refers to two or more types. Terms such as "first," "second," etc., are only used to distinguish similar objects and are not necessarily used to describe a specific order or number of objects.

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0029] With the continuous development of intelligent education technology, course knowledge graphs are gradually becoming an important foundational data supporting the organization of teaching resources, the planning of learning paths, and personalized teaching. Knowledge points refer to the basic concepts, principles, and laws within a particular subject or field. As the most basic and crucial building blocks of a course knowledge graph, their quality directly determines the graph's expressive power and application effectiveness. However, most existing course knowledge graphs rely on manual construction, which is not only time-consuming and labor-intensive but also inefficient, and struggles to adapt to the rapid expansion needs of multiple disciplines and teaching materials. Therefore, there is an urgent need to develop more efficient automated methods for acquiring knowledge points.

[0030] In recent years, Large Language Models (LLMs) have emerged as a promising technology. Their rich knowledge base and excellent instruction-following capabilities enable zero-shot knowledge extraction with outstanding performance. For example, ChatExtract, an automated data extraction framework, was built using engineered prompt sequences. By constructing engineered prompt words to extract structured data, zero-shot extraction was achieved. While LLMs can identify keywords and phrases in text, the extracted knowledge points are fragmented, flat, and lack a hierarchical system, making it difficult to automatically organize the hierarchical, subordinate, and other structured relationships between knowledge points. This structural deficiency makes it difficult for the constructed knowledge graph to support effective knowledge navigation and learning path reasoning, failing to meet the needs of systematic knowledge systems in educational settings. The fine granularity and hierarchy of knowledge points are crucial for educational applications.

[0031] Knowledge extraction tasks require precise identification of knowledge points, increasing the difficulty and challenge of knowledge extraction. Some studies utilize existing data to improve the accuracy of model extraction. For example, DEEPSTRUCT introduces structural pre-training on a task-independent corpus to enhance the structural understanding of the language model. A joint named entity recognition and relation extraction method can also be used to fine-tune a large pre-trained language model to extract complex scientific knowledge. The performance of few-shot relation extraction can be improved by designing task-relevant instructions and pattern-based data generation strategies, but this still relies on a small number of labeled samples. Other works generate synthetic samples using self-suggestion frameworks to provide rich contextual guidance for LLMs, improving their performance in zero-shot relation extraction tasks. These methods reduce the reliance on manually labeled data to some extent, but their core idea is still to provide the model with more data for fine-tuning or optimization. This lack of labeled data still exists when facing entirely new fields or textbooks, limiting the generality and flexibility of the methods.

[0032] To address the challenges of "lack of hierarchical structure" and "low quality of extraction results" commonly found in current large language models for zero-shot knowledge extraction tasks, this application proposes a zero-shot knowledge extraction and filtering framework based on LLMs. To overcome the problems of fragmented and flat knowledge points, this application guides the model to construct a hierarchical knowledge structure through hierarchical constraint prompting engineering. To solve the problems of insufficient extraction accuracy and content redundancy under zero-shot conditions, this application integrates a six-dimensional cognitive verification module and a semantic type classifier to systematically filter and optimize the initial extraction results, ensuring the high quality and pedagogical applicability of the final knowledge points.

[0033] Figure 1 This illustration shows a schematic diagram of the technical concept of a zero-sample knowledge extraction and filtering method provided in an embodiment of this application. Figure 1 As shown, the architecture of this zero-shot knowledge extraction and filtering method mainly includes: a hierarchical knowledge construction part based on iterative semantic segmentation, knowledge point filtering based on reasoning judgment, and knowledge point optimization based on classification filtering.

[0034] Among them, the hierarchical knowledge construction part based on iterative semantic segmentation divides the textbook into blocks based on semantic units, and uses LLMs to extract knowledge points and construct a tree-like hierarchical structure.

[0035] The hierarchical knowledge construction component based on iterative semantic chunking includes a chunking strategy sub-component and a hierarchical structure knowledge extraction sub-component. The chunking strategy sub-component leverages the "chapter-section-paragraph" structure prevalent in course materials, which exhibits a clear hierarchical organization. Each section or paragraph contains multiple learnable knowledge points, thus chunking the content. The quality of the chunking strategy directly impacts the accuracy and contextual completeness of subsequent knowledge point identification. Traditional chunking methods typically employ mechanical segmentation based on word count, page count, or natural paragraphs, easily leading to semantic breaks and incoherent contexts, resulting in incomplete or even erroneous knowledge point extraction by the model.

[0036] This application proposes an iterative semantic segmentation strategy based on textbook structure. This strategy, while respecting the original teaching organization hierarchy of the textbook, performs iterative segmentation in conjunction with semantic integrity constraints to maximize the preservation of the semantic integrity of teaching units. It adopts an iterative structural segmentation process of "chapter → section → semantic paragraph," while combining the maximum input length limit of LLMs and the principle of semantic coherence to achieve block-level segmentation with dual constraints of "structure and semantics."

[0037] The system intelligently segments lengthy textbook texts into appropriately sized, semantically coherent blocks for subsequent processing by a large language model. The core idea is to first perform a coarse segmentation based on the textbook's inherent structure, and then recursively refine the segments that are still too long.

[0038] The block partitioning strategy sub-section includes an iterative block partitioning algorithm for textbooks, which can be divided into two stages: Phase 1: Initial Division Based on Textbook Structure. The goal of this phase is to perform initial division using the textbook's own "chapter-section" hierarchy. The original textbook text and maximum length constraints are received. First, the entire textbook is divided into chapter units, resulting in a chapter list. Next, the content of each chapter is further divided down to the "section" level, forming a final list containing all sections. After this phase, the original entire textbook is broken down into relatively smaller "section" units.

[0039] Phase Two: Recursive Semantic Blocking. The goal of this phase is to ensure that each "segment" meets the length limit of the large model while maintaining relative semantic integrity. It begins by traversing all the "segments" obtained in Phase One. For each "segment," if its length is less than the maximum length limit, it is directly considered a valid text block and stored in the final result list. Otherwise (i.e., the content of the "segment" is too long), it is sent to the "recursive module" for in-depth processing.

[0040] The recursive module is a process that continuously breaks down large blocks of text into smaller, more compliant blocks. For excessively long input text, the algorithm uses a set of semantic delimiters (such as newline characters \n, periods., and commas) to initially segment it, resulting in a list of sentences or paragraphs.

[0041] Iterate through this list, attempting to merge small semantic units (such as sentences) into a "cumulative block". During merging, continuously check if the length of the "cumulative block" is still less than the maximum length limit. If it is, continue merging the next semantic unit. Once it exceeds the limit, output the currently full "cumulative block" as a final text block, then start a new "cumulative block" starting from the currently processed semantic unit, and repeat the above recursive process.

[0042] Finally, the iterative chunking algorithm for textbooks outputs a list of text chunks, each of which combines structural rationality, semantic coherence, and length compliance, laying a solid foundation for subsequent high-quality knowledge point extraction.

[0043] The following is an example of an iterative block partitioning algorithm used in textbooks:

[0044] The iterative block segmentation algorithm for textbooks first performs initial segmentation based on the textbook structure. If the segmented text exceeds the maximum length that a preset block can handle, secondary structures (such as sub-paragraphs or sentences) are searched within that unit for secondary segmentation until the segmented text is less than the set maximum length limit. Conversely, if during the recursive processing, it is found that merging adjacent paragraphs does not exceed the preset block length limit, they are merged to form a complete semantic block. This segmentation strategy has good flexibility, allowing adjacent paragraphs to be merged to form semantically complete knowledge blocks, provided the structure allows.

[0045] The textbook is divided into blocks using an iterative block-segmentation algorithm tailored to the textbook, respecting its inherent hierarchy (chapter-section) and preserving the organizational structure of knowledge. During fine-grained segmentation, punctuation marks and line breaks are used as boundaries to ensure the semantic integrity of each text block as much as possible, avoiding abrupt cuts from the middle of sentences. A recursive mechanism allows for dynamic adjustment of block size, precisely adapting to the input length limitations of large models. The hierarchical knowledge extraction sub-part, based on the inherent structure of the course textbook, constructs a hierarchical knowledge system of "knowledge points—sub-knowledge points" after text segmentation. It integrates a few-sample example guidance strategy and an expert role definition strategy, enabling the large language model not only to perform information extraction tasks but also to understand and express knowledge content from a pedagogical perspective. The hierarchical knowledge structure of "knowledge points—sub-knowledge points" can be formally represented as follows:

[0046] In the formula: This is a block of knowledge points; Provide a summary; It is a collection of sub-knowledge points; For the knowledge structure of the textbook; Titles for knowledge points; For content description; This represents the structural description of a set of sub-knowledge points.

[0047] In this context, "block knowledge" represents the teaching objectives or content scope of a segment or block. For example, binary tree traversal. "Sub-knowledge points" represent the smallest concept (knowledge) unit under a block knowledge point that can be taught, evaluated, or examined independently. For example, preorder traversal, inorder traversal, and postorder traversal.

[0048] To enhance the model's ability to understand teaching knowledge points, the prompt words are set to represent experienced teachers. This role encourages the model to extract valuable knowledge points and summaries from a teaching perspective, rather than simply listing facts. A typical few-shot learning paradigm is adopted, constructing high-quality input-output examples as prompt components. The prompt word template includes textbook text fragments, a prescribed output format, and a few-shot input-output examples.

[0049] Knowledge point filtering based on reasoning can systematically evaluate knowledge points using a six-dimensional reasoning mechanism, ensuring that the knowledge points ultimately retained have pedagogical guidance significance and knowledge carrying capacity.

[0050]

[0051] Table 1. Six-Dimensional Example Table Since the validity of most knowledge points requires consideration of context, semantic relationships, and structural features, this application introduces a thought chain reasoning mechanism. By deducing the intermediate steps and implicit logic between concepts, the rationality and interpretability of the evaluation process are enhanced. Each knowledge point undergoes item-by-item evaluation across six dimensions. If any dimension fails (resulting in false), the knowledge point is eliminated and does not proceed to the subsequent graph construction process.

[0052] This reasoning process can be formally represented by the following judgment function:

[0053] Indicates the first One knowledge point ;

[0054] This represents the corresponding original textbook excerpt; Indicates the sixth dimension Evaluation results across multiple dimensions ; Represents the result of a logical AND operation in six dimensions, when When true is satisfied in all six dimensions, It is true.

[0055] By introducing a "six-dimensional + thinking chain" reasoning mechanism in the reasoning and judgment stage, the validity of knowledge points is systematically and multi-dimensionally judged from six aspects: content matching, structural integrity, conceptual support, logical relationship, specific case, and learning benefits. This reduces redundancy in the initial extraction results and retains the intermediate explanation chain in the reasoning process, ensuring the traceability and verifiability of the screening decision, thereby effectively improving the quality and reliability of knowledge graph construction.

[0056] The knowledge point optimization part based on classification filtering models the task as a supervised semantic classification problem. It combines a predefined set of knowledge types with the classification method of CoT to accurately identify the knowledge type carried by the knowledge point from the textbook context, and then continues to filter based on the knowledge type.

[0057] A systematic analysis of multiple textbooks revealed that three types of knowledge—definitive explanations, process descriptions, and classification-based elaborations—dominate the logical organization of course knowledge. Table 2 shows three representative types of knowledge.

[0058]

[0059] Table 2 Knowledge Point Types The introduction of predefined knowledge types not only supports the establishment of graph edge structures, but also plays a semantic filtering role: if a knowledge point cannot be classified into any knowledge type, it usually indicates that its teaching relevance is insufficient or its semantic expression is ambiguous, and it can be marked as invalid knowledge to improve the quality of the graph.

[0060] The knowledge type identification task is performed by LLMs. To enhance the stability and interpretability of the judgment process, this application designs a structured prompt word template, integrating role constraints and CoT guidance mechanisms, and includes four core reasoning steps. The core reasoning steps are shown in Table 3.

[0061]

[0062] Table 3 Core Reasoning Steps This process can be formally represented as

[0063] In the formula: Indicates step-by-step reasoning The type of knowledge represented by this knowledge point; Knowledge points; For context summary; The matched semantic type belongs to a predefined set. .

[0064] Based on the above technical concept, hierarchical constraint prompting engineering guides the model to construct a hierarchical knowledge structure. To solve the problems of insufficient extraction accuracy and content redundancy under zero-sample conditions, this application integrates a six-dimensional cognitive verification module and a semantic type classifier to systematically filter and optimize the preliminary extraction results, ensuring the high quality and teaching applicability of the final knowledge points.

[0065] The following describes the specific implementation process of the above technical concept.

[0066] like Figure 2 As shown, this embodiment provides a zero-shot course knowledge extraction and filtering method, including: 110. The text of the teaching materials to be processed is divided into blocks and knowledge points are extracted to obtain a hierarchical structure of knowledge points, wherein the hierarchical structure of knowledge points includes knowledge points; 120. Using reasoning and judgment, filter the knowledge points in the knowledge point hierarchy to obtain the filtered knowledge point hierarchy. 130. Based on predefined knowledge types, classify and label the knowledge points in the filtered knowledge point hierarchy.

[0067] In one possible implementation, the step of segmenting the text of the teaching material to be processed into blocks and extracting knowledge points to obtain a hierarchical structure of knowledge points includes: An iterative semantic segmentation algorithm is used to segment the text of the teaching materials to be processed according to a preset hierarchical structure to obtain text blocks; Extract the knowledge points from the text block to obtain the knowledge point hierarchy structure, which includes block knowledge points and sub-knowledge points. The sub-knowledge points are conceptual units under the block knowledge points.

[0068] In one possible implementation, an iterative semantic segmentation algorithm is used to segment the text of the teaching material to be processed according to a preset hierarchical structure to obtain text blocks, including: Based on the hierarchical division of the text in the teaching material to be processed, the teaching material to be processed is decomposed into multiple text units; Start traversing each of the text units and determine whether the length of the text unit is greater than a preset length threshold; If so, the text unit is segmented using a recursive algorithm to obtain the text block; Otherwise, the text unit is treated as the text block.

[0069] In one possible implementation, the step of using a recursive algorithm to segment the text units to obtain the text blocks includes: Using semantic delimiters, the text units are segmented to obtain a list of multiple semantic units, which includes multiple semantic units; Traverse the list of semantic units, merge multiple semantic units into an accumulation block, and determine whether the length of the accumulation block is greater than a preset length threshold. If it is greater than the sum of its parts, the most recently merged semantic unit is removed from the accumulated block, and the accumulated block with the most recently merged semantic unit removed is taken as the text block. Otherwise, add the semantic unit to the accumulated block; Repeat the above steps to process all semantic units in the semantic unit list and obtain the text block.

[0070] In one possible implementation, the knowledge point hierarchy includes:

[0071] in, It is the knowledge structure of the textbook. These are the sub-knowledge points mentioned. This refers to the knowledge points mentioned above. It is a summary of the knowledge points in the block. It is the first Individual knowledge points, This is the title of the sub-knowledge point. This refers to the content description of the sub-knowledge points.

[0072] In one possible implementation, the step of using reasoning to filter the knowledge points in the knowledge point hierarchy to obtain a filtered knowledge point hierarchy includes: The reasoning decision is expressed as a judgment function:

[0073] in, Indicates the first Height knowledge points , It is the number of the sub-knowledge points. It is the text fragment of the teaching material to be processed corresponding to the sub-knowledge point; The sub-knowledge points are in the preset six dimensions. Evaluation results across multiple dimensions; ; The result of logical AND operation on the preset six dimensions, when When all six preset dimensions are true. The value is true; where the preset six dimensions are content matching degree, structural completeness, concept support degree, judgment logic degree, case specificity degree, and learning benefit degree.

[0074] In one possible implementation, the step of classifying and labeling knowledge points in the filtered knowledge point hierarchy based on predefined knowledge types includes: Use the following annotation functions to annotate the knowledge points: Indicates step-by-step reasoning This knowledge point represents the type of knowledge. This is the first sub-knowledge point. For context summary, The semantic type matched for the sub-knowledge point belongs to a predefined knowledge type. The predefined knowledge types include definitions, processes, and classifications. The definitions include explanations of the meaning of a term or concept. The processes include descriptions of the operational steps or time sequence of a task, technology, or system. The classifications include characterizing the components, subclasses, or subordinate structures of a whole.

[0075] The reasoning steps include: understanding the meaning of knowledge points, analyzing summary content, matching predefined types, and outputting the final result. Understanding the meaning of knowledge points includes clarifying the semantic boundaries and teaching-oriented keywords of the sub-knowledge points. Analyzing the summary content includes extracting keywords with language structure features from the text fragments of the teaching materials to be processed corresponding to the sub-knowledge points, matching the keywords with the predefined knowledge types item by item to determine whether they conform to the language and structural features of the definition, process, and classification, and selecting the most matching predefined knowledge type.

[0076] The following experiments were conducted using the methods described in this application: To comprehensively evaluate the effectiveness of knowledge point extraction methods, this application designs an evaluation system from two dimensions: 1) traditional evaluation indicators based on manual annotation comparison, and 2) a collaborative evaluation mechanism combined with LLMs.

[0077] (1) Traditional evaluation indicators based on manual annotation and comparison A validation set was manually constructed, and three commonly used metrics were used to quantitatively evaluate the extraction results, specifically including: Accuracy: Measures how many of the extracted knowledge points are valid knowledge points that are consistent with human annotations; Recall: Measures how many manually labeled knowledge points are successfully extracted by the model; F1 score (F1-Score): The harmonic mean of precision and recall, used to comprehensively evaluate overall extraction performance.

[0078] By comparing the results with those of manual annotation, the above indicators can be used to verify the model's universality and robustness across different textbook texts.

[0079] (2) Collaborative evaluation mechanism in conjunction with LLMs Considering the high cost and strong subjectivity of manual evaluation, this application introduces an unsupervised LLMs collaborative evaluation method, which verifies the teaching effectiveness of knowledge points through multi-model cross-validation to achieve weakly supervised quality control.

[0080] This mechanism integrates multiple mainstream Chinese learning models, uses a unified prompt word template, and scores each knowledge point from the following three dimensions:

[0081] Table 4 Model Evaluation Elements Three large models—Deepseek-r1, Qwen-plus, and ChatGLM-3-Turbo—were selected for knowledge point evaluation. An aggregation strategy was used for the final evaluation and determination.

[0082] The maximum score is 5 points. If the average score is ≥ 4.0, it will be marked as a "high-quality knowledge point". If a score ≤ 2 appears in any model, it is marked as "disagreement exists" and will be manually reviewed.

[0083] Examples of prompt word templates are shown in Table 5.

[0084]

[0085] Table 5. Prompt Word Templates The following is an evaluation of the experimental results: (1) Results of traditional evaluation indicators To verify the effectiveness of the proposed method, this application uses three computer science textbooks as experimental subjects. One chapter is randomly selected from each of the textbooks, namely "Introduction to Computers", "Computer Networks (8th Edition)" and "Fundamentals of Computer Network Security (5th Edition)," to construct a verification set, totaling 201 entries.

[0086] The experiments were conducted using two methods: first, directly extracting knowledge points using the three models Qwen-Plus, DeepSeek-R1, and ChatGLM-3-Turbo; second, extracting knowledge points using the method proposed in this application. The experimental results are shown in Table 6.

[0087] Table 6 Comparison of knowledge point extraction effects between the proposed method and the baseline model. When directly using Qwen-Plus, DeepSeek-R1, and ChatGLM-3-Turbo to extract knowledge points, they generally exhibit high recall but low precision. Among them, the ChatGLM-3-Turbo model shows a high recall on multiple datasets, indicating that it is relatively aggressive in knowledge point extraction and tends to extract a large number of candidate knowledge points.

[0088] The experimental results show that the proposed method achieved high accuracy in all three textbooks, indicating the removal of redundant information without pedagogical value. This verifies that the method, without requiring professional fine-tuning, is well-suited for knowledge point extraction tasks from different textbooks, demonstrating good versatility and practicality.

[0089] (2) Results of collaborative evaluation of large models In the weakly supervised evaluation stage, this application introduces a collaborative evaluation mechanism using multiple LLMs to comprehensively verify the effectiveness of the extracted knowledge points. The specific process is as follows: First, the method described in this application is used to extract and filter knowledge points from computer science textbooks, obtaining a final set of high-quality knowledge points. Then, this set is input into three different LLMs, which score each point based on accuracy, clarity, and pedagogical relevance, and calculate a comprehensive evaluation score.

[0090]

[0091] Table 7 Evaluation Results of the Major Models As shown in the table, different LLMs exhibit slight differences in performance across the three dimensions of "accuracy," "clarity," and "teaching relevance." Qwen-Plus scored highest in both accuracy and teaching relevance, indicating its strong ability to identify terminology and curriculum fit within teaching materials. This may be attributed to its pre-training advantage on Chinese educational corpora, enabling a more accurate understanding of the language style and teaching objectives of the teaching materials.

[0092] DeepSeek-R1 performed relatively well across the three metrics, with a small gap compared to Qwen-Plus, demonstrating its strong generalization and error tolerance capabilities for Chinese knowledge points. In particular, it showed good stability and tolerance when dealing with knowledge points that are slightly loosely structured but semantically clear.

[0093] While ChatGLM-3-Turbo is similar to DeepSeek-R1 in accuracy, it scores lowest in clarity, indicating a slight conservatism in areas such as the completeness of knowledge point expression and the judgment of definitions, possibly tending to assign lower scores to insufficiently expressed knowledge points. However, it maintains a high level in terms of pedagogical relevance, demonstrating its ability to effectively identify knowledge units closely related to the course content.

[0094] The method for acquiring knowledge points based on a large model proposed in this application has achieved excellent results in experiments. This demonstrates that the method presented in this application has gained relatively consistent recognition for its educational value in knowledge point acquisition tasks.

[0095] Furthermore, considering the high sensitivity of generative LLMs to prompts during task execution, this application introduces two prompt design strategies in the knowledge point extraction stage: Few-shot example guidance and expert role setting, to improve the model's understanding depth and output quality. To verify the effectiveness of the above strategies, this application removes Few-shot and expert role setting without changing the corpus, and directly uses a general prompt for knowledge point extraction. The experiments are still based on three typical textbooks: *Introduction to Computers*, *Computer Networks (8th Edition)*, and *Fundamentals of Computer Network Security (5th Edition)*. One chapter from each textbook is randomly selected to manually construct the validation set. The results are shown in Table 8.

[0096] Table 8. Experimental results of ablation using the structured Prompt strategy The comparison results show that the lack of a structured Prompt design significantly reduced the accuracy and comprehensiveness of knowledge point extraction, with the average F1 score decreasing by more than 14 percentage points compared to the original method. This result clearly demonstrates that the Few-shot guidance and expert role setting in the Prompt play a crucial role in guiding the model to understand the textbook structure and teaching context, significantly improving the accuracy and robustness of the extraction results.

[0097] The overall methodological framework proposed in this application also includes two core functional modules: a knowledge reasoning judgment filtering module and a knowledge type classification filtering module. To evaluate the actual contribution of these two modules to the knowledge point extraction effect, this application designed a module ablation experiment: while keeping other conditions unchanged, the two modules were removed respectively, and a multilingual model collaborative evaluation mechanism was used to score the quality of the results. Multiple LLMs were used to independently score the knowledge point extraction results in terms of accuracy, clarity, and pedagogical relevance, and the comprehensive evaluation score of each model was calculated. Finally, the final comprehensive evaluation score was obtained by averaging. The experimental results are shown in Table 9:

[0098] Table 9. Results of the ablation experiments for reasoning and judgment filtering and knowledge type classification filtering. Experimental results show that: 1) The lack of a knowledge reasoning module significantly impacts performance, reducing the model's ability to logically summarize knowledge points and making it prone to overly simplistic and unbalanced knowledge points. 2) The model's performance deteriorates significantly in the absence of a knowledge type classification module. This module, by retaining most knowledge points that conform to textbook types, has a significant impact on model performance. 3) The effectiveness of the module design in this application's method is verified. The complete framework, while maintaining the generality of the large model, enhances the ability to acquire effective knowledge points, laying the foundation for the subsequent construction of high-quality course knowledge graphs.

[0099] This application is based on Figure 3 The text example illustrates its three-step knowledge extraction and filtering process. First, an iterative semantic segmentation algorithm is used to divide the original text into two units with complete semantic meaning (e.g., ...). Figure 3 (As shown in different colors). Next, through the construction of a hierarchical knowledge system, candidate knowledge with a "block knowledge point - sub-knowledge point" hierarchical structure is extracted from these blocks. Then, the second step involves knowledge reasoning and judgment. In this stage, the system filters candidate knowledge points based on a six-dimensional reasoning mechanism. For example, the knowledge point "encrypted credit card transmission" is filtered out because it is judged to be merely a scenario description and fails to extract a core concept with teaching value. The third step involves further filtering through knowledge type classification, matching according to three preset knowledge types: "definition," "process," and "classification." In this stage, knowledge points like "future prospects of quantum encryption" and "universities offering cybersecurity majors" are ultimately removed because they cannot match any preset teaching knowledge type.

[0100] The results of some experiments in "Introduction to Computers (5th Edition)" using the knowledge point extraction and filtering method of this application are as follows: Figure 4 As shown.

[0101] This application addresses the issues of inconsistent granularity, structural deficiencies, and semantic redundancy in current course knowledge point extraction processes by proposing a zero-shot knowledge point extraction and filtering method integrating LLMs. Based on the hierarchical structure of textbooks, this method ensures contextual integrity through an iterative semantic chunking strategy. Combined with few-shot prompts and expert role settings, it guides the large model to extract hierarchical knowledge points from a pedagogical perspective. A six-dimensional reasoning mechanism is introduced to effectively filter low-quality content. Semantic knowledge classification further enhances the pedagogical expressiveness of the graph structure. Experimental results show that the proposed method significantly outperforms comparative models lacking reasoning or knowledge type recognition capabilities in terms of accuracy, clarity, and pedagogical relevance, demonstrating good adaptability and robustness.

[0102] To achieve the above functions, this application provides a zero-sample course knowledge extraction and filtering device 610. Figure 5A schematic diagram of the structure of a device 610 provided in an embodiment of this application is shown.

[0103] like Figure 5 As shown, the device 610 may include an extraction module 611, a filtering module 612, and a labeling module 613.

[0104] The device includes: The extraction module is used to segment the text of the teaching materials to be processed and extract knowledge points to obtain a hierarchical structure of knowledge points, wherein the hierarchical structure of knowledge points includes knowledge points; The filtering module is used to filter the knowledge points in the knowledge point hierarchy using reasoning and judgment to obtain the filtered knowledge point hierarchy. The annotation module is used to classify and annotate the knowledge points in the filtered knowledge point hierarchy based on predefined knowledge types.

[0105] In one possible implementation, the extraction module is used to segment the text of the teaching material to be processed according to a preset hierarchical structure using an iterative semantic segmentation algorithm to obtain text blocks; Extract the knowledge points from the text block to obtain the knowledge point hierarchy structure, which includes block knowledge points and sub-knowledge points. The sub-knowledge points are conceptual units under the block knowledge points.

[0106] In one possible implementation, the extraction module is used to decompose the teaching material to be processed according to the hierarchical division in the text to obtain multiple text units; Start traversing each of the text units and determine whether the length of the text unit is greater than a preset length threshold; If so, the text unit is segmented using a recursive algorithm to obtain the text block; Otherwise, the text unit is treated as the text block.

[0107] In one possible implementation, an extraction module is used to segment the text units using semantic delimiters to obtain a list of multiple semantic units, the list of semantic units including multiple semantic units; Traverse the list of semantic units, merge multiple semantic units into an accumulation block, and determine whether the length of the accumulation block is greater than a preset length threshold. If it is greater than the sum of its parts, the most recently merged semantic unit is removed from the accumulated block, and the accumulated block with the most recently merged semantic unit removed is taken as the text block. Otherwise, add the semantic unit to the accumulated block; Repeat the above steps to process all semantic units in the semantic unit list and obtain the text block.

[0108] In one possible implementation, the knowledge point hierarchy includes:

[0109] in, It is the knowledge structure of the textbook. These are the sub-knowledge points mentioned. This refers to the knowledge points mentioned above. It is a summary of the knowledge points in the block. It is the first Individual knowledge points, This is the title of the sub-knowledge point. This refers to the content description of the sub-knowledge points.

[0110] In one possible implementation, the filtering module is used to represent the inference decision as a decision function:

[0111] in, Indicates the first Height knowledge points , It is the number of the sub-knowledge points. It is the text fragment of the teaching material to be processed corresponding to the sub-knowledge point; The sub-knowledge points are in the preset six dimensions. Evaluation results across multiple dimensions ; The result of logical AND operation on the preset six dimensions, when When all six preset dimensions are true. The value is true; where the preset six dimensions are content matching degree, structural completeness, concept support degree, judgment logic degree, case specificity degree, and learning benefit degree.

[0112] In one possible implementation, the annotation module is used to annotate the knowledge points using the following annotation functions:

[0113] Indicates step-by-step reasoning This knowledge point represents the type of knowledge. This is the first sub-knowledge point. For context summary, The semantic type matched for the sub-knowledge point belongs to a predefined knowledge type. The predefined knowledge types include definitions, processes, and classifications. The definitions include explanations of the meaning of a term or concept. The processes include descriptions of the operational steps or time sequence of a task, technology, or system. The classifications include characterizing the components, subclasses, or subordinate structures of a whole. The reasoning steps include: understanding the meaning of knowledge points, analyzing summary content, matching predefined types, and outputting the final result. Understanding the meaning of knowledge points includes clarifying the semantic boundaries and teaching-oriented keywords of the sub-knowledge points. Analyzing the summary content includes extracting keywords with language structure features from the text fragments of the teaching materials to be processed corresponding to the sub-knowledge points, matching the keywords with the predefined knowledge types item by item to determine whether they conform to the language and structural features of the definition, process, and classification, and selecting the most matching predefined knowledge type.

[0114] Those skilled in the art will readily recognize that, based on the units and algorithm steps described in conjunction with the embodiments disclosed in this application, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is implemented in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0115] It should be noted that, Figure 5 The division of modules / units is illustrative and represents only one logical functional division; in actual implementation, other division methods are possible. For example, two or more functions can be integrated into a single data acquisition module. The integrated modules described above can be implemented either in hardware or as software functional modules.

[0116] The hardware and logical structures of the cluster nodes provided in this application will be described below using embodiments of the present application.

[0117] refer to Figure 6 , Figure 6 This diagram illustrates the hardware structure of a zero-sample course knowledge extraction and filtering device provided in an embodiment of this application. Figure 6As shown, device 700 includes a processor 701, a memory 702, a network interface 703, and a bus 704. The processor 701, memory 702, and network interface 703 are connected via bus 704. The processor 701 is the control center of device 700 and can be a general-purpose central processing unit (CPU). Processor 701 may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, graphics processing units (GPUs), neural processing units (NPUs), tensor processing units (TPUs), or artificial intelligence chips, data processing units (DPUs), etc. As an example, processor 701 includes one or more CPUs, for example... Figure 6 CPU 0 and CPU 1 are shown in the diagram.

[0118] Furthermore, this application does not limit the number of processor cores in each processor.

[0119] The memory 702 is used to store program instructions or data to be accessed by the application process. The processor 701 can execute the program instructions in the memory 702 to implement the zero-sample course knowledge extraction and filtering method provided in this application embodiment.

[0120] The memory 702 is used to store program instructions or data accessed by the application process. The processor 701 can implement the communication method provided in this application embodiment by executing the program instructions in the memory 702. The memory 702 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DRRAM). Non-volatile memory can be storage class memory (SCM), solid-state drive (SSD), hard disk drive (HDD), etc. Storage class memory can be, for example, non-volatile memory (NVM), phase-change memory (PCM), persistent memory, etc.

[0121] In one possible implementation, the memory 702 exists independently of the processor 701. The memory 702 is connected to the processor 701 via a bus 704 and is used to store data, instructions, or program code. When the processor 701 calls and executes the instructions or program code stored in the memory 702, it can implement the zero-shot course knowledge extraction and filtering method provided in the embodiments of this application.

[0122] In another possible implementation, the memory 702 and the processor 701 are integrated together.

[0123] Network interface 703 is used for device 700 to connect with other devices (such as a second device) via a communication network, which may be Ethernet, radio access network (RAN), wireless local area network (WLAN), etc. Network interface 703 includes a receiving unit for receiving data / messages and a sending module for sending data / messages. In this embodiment, network interface 703 includes an RNIC. The RNIC may be a standalone Peripheral Component Interconnect Express (PCIe) network card, or it may be a network interface unit integrated into a system-on-chip (SoC) chip; there is no limitation on this. Bus 704 may be an industry standard architecture (ISA) bus, a peripheral component interconnect (PCI) bus, a high-speed serial computer interconnect express (PCIe) bus, a compute express link (CXL) bus, or an extended industry standard architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0124] It should be pointed out that, Figure 6 The structure shown does not constitute a limitation on device 700, except... Figure 6 In addition to the components shown, the device 700 includes more than Figure 6 This may indicate more or fewer components, or combinations of certain components, or different component arrangements.

[0125] This application also provides a computer program product and a computer-readable storage medium for storing the computer program product. The computer program product may include one or more program instructions, which, when executed by one or more processors, can provide the functions or part of the functions described in the above figures.

[0126] In some examples, a communication device for performing the method may be configured to provide various operations, functions, or actions in response to one or more program instructions stored in a computer-readable storage medium.

[0127] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When executed on a computer and when the computer execution instructions are executed, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks, SSDs).

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of this application.

Claims

1. A zero-shot course knowledge extraction and filtering method, characterized in that, The method includes: The text of the teaching materials to be processed is divided into blocks and knowledge points are extracted to obtain a hierarchical structure of knowledge points, wherein the hierarchical structure of knowledge points includes knowledge points; By using reasoning and judgment, the knowledge points in the knowledge point hierarchy are filtered to obtain the filtered knowledge point hierarchy. Based on predefined knowledge types, the knowledge points in the filtered knowledge point hierarchy are classified and labeled.

2. The method according to claim 1, characterized in that, The process involves segmenting the text of the teaching materials to be processed into blocks and extracting knowledge points to obtain a hierarchical structure of knowledge points, including: An iterative semantic segmentation algorithm is used to segment the text of the teaching materials to be processed according to a preset hierarchical structure to obtain text blocks; Extract the knowledge points from the text block to obtain the knowledge point hierarchy structure, which includes block knowledge points and sub-knowledge points. The sub-knowledge points are conceptual units under the block knowledge points.

3. The method according to claim 2, characterized in that, The iterative semantic segmentation algorithm is used to segment the text of the teaching material to be processed according to a preset hierarchical structure to obtain text blocks, including: Based on the hierarchical division of the text in the teaching material to be processed, the teaching material to be processed is decomposed into multiple text units; Start traversing each of the text units and determine whether the length of the text unit is greater than a preset length threshold; If so, the text unit is segmented using a recursive algorithm to obtain the text block; Otherwise, the text unit is treated as the text block.

4. The method according to claim 3, characterized in that, The step of using a recursive algorithm to segment the text units to obtain the text blocks includes: Using semantic delimiters, the text units are segmented to obtain a list of multiple semantic units, which includes multiple semantic units; Traverse the list of semantic units, merge multiple semantic units into an accumulation block, and determine whether the length of the accumulation block is greater than a preset length threshold. If it is greater than the sum of its parts, the most recently merged semantic unit is removed from the accumulated block, and the accumulated block with the most recently merged semantic unit removed is taken as the text block. Otherwise, add the semantic unit to the accumulated block; Repeat the above steps to process all semantic units in the semantic unit list and obtain the text block.

5. The method according to claim 3, characterized in that, The hierarchical structure of the knowledge points includes: in, It is the knowledge structure of the textbook. These are the sub-knowledge points mentioned. This refers to the knowledge points mentioned above. It is a summary of the knowledge points in the block. It is the i-th sub-knowledge point. This is the title of the sub-knowledge point. This refers to the content description of the sub-knowledge points.

6. The method according to claim 2, characterized in that, The step of using reasoning to filter knowledge points in the knowledge point hierarchy to obtain a filtered knowledge point hierarchy includes: The reasoning decision is expressed as a judgment function: in, Indicates the first Height knowledge points , It is the number of the sub-knowledge points. It is the text fragment of the teaching material to be processed corresponding to the sub-knowledge point; The sub-knowledge points are in the preset six dimensions. d Evaluation results across multiple dimensions ; The result of logical AND operation on the preset six dimensions, when When all six preset dimensions are true. The value is true; where the preset six dimensions are content matching degree, structural completeness, concept support degree, judgment logic degree, case specificity degree, and learning benefit degree.

7. The method according to claim 2, characterized in that, The step of classifying and labeling knowledge points in the filtered knowledge point hierarchy based on predefined knowledge types includes: Use the following annotation functions to annotate the knowledge points: Indicates step-by-step reasoning This knowledge point represents the type of knowledge. This is the first sub-knowledge point. For context summary, The semantic type matched for the sub-knowledge point belongs to a predefined knowledge type. The predefined knowledge types include definitions, processes, and classifications. The definitions include explanations of the meaning of a term or concept. The processes include descriptions of the operational steps or time sequence of a task, technology, or system. The classifications include characterizing the components, subclasses, or subordinate structures of a whole. The reasoning steps include: understanding the meaning of knowledge points, analyzing summary content, matching predefined types, and outputting the final result. Understanding the meaning of knowledge points includes clarifying the semantic boundaries and teaching-oriented keywords of the sub-knowledge points. Analyzing the summary content includes extracting keywords with language structure features from the text fragments of the teaching materials to be processed corresponding to the sub-knowledge points, matching the keywords with the predefined knowledge types item by item to determine whether they conform to the language and structural features of the definition, process, and classification, and selecting the most matching predefined knowledge type.

8. A zero-sample course knowledge extraction and filtering device, characterized in that, The device includes: The extraction module is used to segment the text of the teaching materials to be processed and extract knowledge points to obtain a hierarchical structure of knowledge points, wherein the hierarchical structure of knowledge points includes knowledge points; The filtering module is used to filter the knowledge points in the knowledge point hierarchy using reasoning and judgment to obtain the filtered knowledge point hierarchy. The annotation module is used to classify and annotate the knowledge points in the filtered knowledge point hierarchy based on predefined knowledge types.

9. A computer storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by one or more processors, implements the zero-sample course knowledge extraction and filtering method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes a memory and one or more processors, wherein the memory stores a computer program that, when executed by the one or more processors, implements the zero-sample course knowledge extraction and filtering method as described in any one of claims 1 to 7.

Images