A knowledge point extraction method for textbooks and related equipment

Abstract

The present invention provides a method for extracting knowledge points from textbooks and related equipment, which relate to the field of computer technology. The method includes: dividing the target textbook into blocks according to the directory data of the target textbook to obtain text blocks; extracting knowledge points from the text blocks, and saving the obtained knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map; generating text-question-answer pairs for each knowledge point in the tree-like hierarchical knowledge point map according to the type information of the knowledge points; constructing a training set based on the text blocks, the type information and the question-answer pairs, and training a preset large language model provided with an adapter mechanism according to instruction fine-tuning through the training set to obtain a target knowledge point extraction model; inputting the text blocks of the textbook to be extracted into the target knowledge point extraction model to obtain target knowledge points corresponding to the textbook to be extracted.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method for extracting knowledge points from teaching materials and related equipment. Background Art

[0002] Textbooks are important carriers of knowledge. Currently, large language models face numerous challenges when it comes to extracting knowledge points from textbooks. Traditional methods rely heavily on manual analysis of textbook knowledge points, which is extremely inefficient and difficult to meet the demands of large-scale textbook processing. Existing large language model-based knowledge point extraction methods, due to a lack of targeted training, lack accuracy and generalization capabilities when faced with the diverse representations and knowledge architectures of different textbooks. For example, when extracting the same knowledge point from different versions of textbooks, general large language models often miss, incorrectly extract, or fail to accurately understand the hierarchical relationships between knowledge points. Summary of the Invention

[0003] In order to address the deficiencies in the prior art, the present invention provides a method for extracting knowledge points from textbooks, so as to solve the problem of insufficient accuracy and generalization ability of large language models in the prior art when processing textbook knowledge point extraction tasks.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions.

[0005] In a first aspect, the present application provides a method for extracting knowledge points from a textbook, comprising the following steps:

[0006] Divide the target teaching material into blocks according to the catalogue data of the target teaching material to obtain text blocks;

[0007] Extracting knowledge points from the text block to obtain target knowledge points, and saving the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map;

[0008] Generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge point;

[0009] Constructing a training set based on the text block, the type information, and the question-answer pair, and using the training set to train a preset large language model provided with an adapter mechanism in a manner of fine-tuning instructions to obtain a target knowledge point extraction model;

[0010] The text block of the teaching material to be extracted is input into the target knowledge point extraction model to obtain the target knowledge points corresponding to the teaching material to be extracted.

[0011] Optionally, the step of dividing the target teaching material into blocks according to the catalog data of the target teaching material to obtain text blocks includes:

[0012] Segmenting the target teaching material into chapter-level text blocks, section-level text blocks, and sub-item-level text blocks according to the content information, level information, and page number information in the catalog data;

[0013] The chapter-level text block, the section-level text block and the sub-item-level text block are all regarded as the text block, and the text block includes a hierarchical ID obtained based on the hierarchical information, a title obtained based on the content information and a text range obtained based on the page number information.

[0014] Optionally, the step of extracting knowledge points from the text block includes:

[0015] Identifying the number of characters contained in the text block, and if the number of characters is greater than a threshold, using the threshold as a maximum number of characters, dividing the text block according to preset punctuation marks to obtain initial sub-text blocks with a number of characters less than or equal to the threshold;

[0016] Correcting the boundary of the initial sub-text block by a semantic similarity calculation method based on sliding window sampling to obtain a target sub-text block;

[0017] Extracting knowledge points based on the target sub-text block;

[0018] If the number of characters is less than or equal to the threshold, knowledge points are extracted from the text block.

[0019] Optionally, the step of extracting knowledge points from the text block further includes:

[0020] Identifying the number of characters contained in the text block, and if the number of characters is greater than a threshold, dividing the text block into paragraphs using the threshold as the maximum number of characters to obtain secondary sub-text blocks having a number of characters less than or equal to the threshold;

[0021] Correcting the boundary of the secondary sub-text block by a semantic segmentation method based on LLM to obtain a target sub-text block;

[0022] Extracting knowledge points based on the target sub-text block;

[0023] If the number of characters is less than or equal to the threshold, knowledge points are extracted from the text block.

[0024] Optionally, the step of extracting knowledge points from the text block to obtain target knowledge points includes:

[0025] Performing part-of-speech tagging on the words in the text block, and extracting the words in the text block that are marked as noun parts of speech based on the tagging results to obtain target vocabulary;

[0026] Calculating the co-occurrence weights of the target words based on the TextRank algorithm, sorting the target words according to the co-occurrence weights to obtain a sorting result, and screening knowledge points from all target words according to the sorting result to obtain a first knowledge point;

[0027] Based on the semantic enhancement technology of LLM, knowledge points are extracted from the text block through the hint engineering to obtain the second knowledge points;

[0028] The first knowledge point is supplemented based on the second knowledge point to obtain a target knowledge point.

[0029] Optionally, the step of saving the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map further includes:

[0030] Filtering the target knowledge points to obtain filtered target knowledge points, wherein the filtering includes frequency filtering based on the number of occurrences of the target knowledge points in the target textbook, length filtering based on the number of characters in the target knowledge points, and dictionary filtering based on common words in the target knowledge points using a domain dictionary of the subject area of ​​the target textbook;

[0031] Based on WordNet, synonyms in the filtered target knowledge points are merged to obtain merged knowledge points;

[0032] Annotating the merged knowledge point using the type information of the merged knowledge point and the extended attribute information of the merged knowledge point as annotation information to obtain an annotated knowledge point, and sending the annotated knowledge point to a display device corresponding to the attribute feature according to a preset sending rule based on the attribute feature of the target teaching material;

[0033] Request information for the annotated knowledge point is obtained, and the annotated knowledge point is updated based on the request information to obtain an updated knowledge point, and the updated knowledge point is saved according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map.

[0034] Optionally, the step of generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point graph according to the type information of the target knowledge point includes:

[0035] According to the type information of the target knowledge point, a question template corresponding to the type information is selected from a preset question template library, and the target knowledge point is filled into the question template to obtain a first round of questions;

[0036] Using the content of the target knowledge point as the answer data for the first round of questions to obtain a first round of question-answer pairs;

[0037] Generate an extended question based on the name of the target knowledge point and the hierarchical data of the target knowledge point in the tree-like hierarchical knowledge point map;

[0038] Using the content of the target knowledge point as answer data for the extended question to obtain an extended question-answer pair;

[0039] Extracting entity data from the content of the target knowledge point, and generating progressive follow-up questions based on the entity data through LLM's dynamic questioning technology;

[0040] Using the content of the target knowledge point as answer data for the progressive follow-up question to obtain a follow-up question-answer pair;

[0041] The first-round question-answer pairs, the expanded question-answer pairs, and the follow-up question-answer pairs are used as question-answer pairs for the target knowledge points in the tree-like hierarchical knowledge point map.

[0042] On the other hand, the present application provides a knowledge point extraction system for textbooks, the system comprising:

[0043] A data receiving module, configured to divide the target teaching material into blocks according to the catalogue data of the target teaching material to obtain text blocks;

[0044] A graph generation module is used to extract knowledge points from the text block to obtain target knowledge points, and save the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point graph;

[0045] A question-answering module, configured to generate question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge point;

[0046] A training module is configured to construct a training set based on the text block, the type information, and the question-answer pairs, and to train a preset large language model provided with an adapter mechanism by fine-tuning the instructions using the training set to obtain a target knowledge point extraction model;

[0047] The extraction module is used to input the text block of the teaching material to be extracted into the target knowledge point extraction model to obtain the target knowledge point corresponding to the teaching material to be extracted.

[0048] In a third aspect, the present application also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for extracting knowledge points from textbooks when executing the computer program.

[0049] In a fourth aspect, the present application further provides a computer-readable storage medium, which stores a computer program, and is characterized in that when the computer program is executed by a processor, it implements the above-mentioned knowledge point extraction method for textbooks.

[0050] Beneficial effects:

[0051] This application divides the target teaching material into blocks according to the catalog data of the target teaching material to obtain text blocks; extracts knowledge points from the text blocks to obtain target knowledge points, and saves the target knowledge points according to the hierarchical relationship of the catalog data to obtain a tree-like hierarchical knowledge point map; generates question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge point; constructs a training set based on the text blocks, the type information and the question-answer pairs, and uses the training set to train a preset large language model with an adapter mechanism in a fine-tuned manner according to instructions to obtain a target knowledge point extraction model; inputs the text blocks of the teaching material to be extracted into the target knowledge point extraction model to obtain the target knowledge points corresponding to the teaching material to be extracted. By constructing a special data set based on the knowledge points of the teaching material and fine-tuning the large language model based on it, the model has the ability to automatically extract knowledge points from different teaching materials efficiently and accurately, thereby improving the automation level and accuracy of educational knowledge processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flowchart of a method for extracting knowledge points from textbooks in an embodiment of the present invention;

[0053] Figure 2 Schematic diagram of a knowledge point extraction system for textbooks according to an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0055] Figure 4 It is a structural diagram of a computer-readable storage medium provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0057] The embodiments described in this application are only part of the embodiments of the present invention, not all of the embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0058] See also Figure 1 As shown, the present invention provides a method for extracting knowledge points from textbooks, comprising the following steps:

[0059] S110, dividing the target textbook into blocks according to the catalog data of the target textbook to obtain text blocks;

[0060] S120, extracting knowledge points from the text block to obtain target knowledge points, and saving the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map;

[0061] S130, generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge point;

[0062] S140: constructing a training set based on the text block, the type information, and the question-answer pair, and using the training set to train a preset large language model provided with an adapter mechanism in a manner of fine-tuning instructions to obtain a target knowledge point extraction model;

[0063] S150 , inputting the text block of the teaching material to be extracted into the target knowledge point extraction model to obtain the target knowledge points corresponding to the teaching material to be extracted.

[0064] In a possible implementation, the step of dividing the target teaching material into blocks according to the catalog data of the target teaching material to obtain text blocks includes:

[0065] Segmenting the target teaching material into chapter-level text blocks, section-level text blocks, and sub-item-level text blocks according to the content information, level information, and page number information in the catalog data;

[0066] The chapter-level text block, the section-level text block and the sub-item-level text block are all regarded as the text block, and the text block includes a hierarchical ID obtained based on the hierarchical information, a title obtained based on the content information and a text range obtained based on the page number information.

[0067] For example, we parse the textbook's chapter directory (e.g., a three-level structure of "chapter → section → sub-item") and construct a tree-like hierarchical model: TreeNode{level ID, title, text range}. This divides the textbook into three levels of text blocks: chapter-level, section-level, and sub-item-level. For example, in "High School Physics Compulsory 1," "Chapter 1: Description of Motion" is a chapter-level block, "1.1 Point Mass Reference System" is a section-level block, and "Definition of Point Mass" is a sub-item-level block.

[0068] In a possible implementation, the step of extracting knowledge points from the text block includes:

[0069] Identifying the number of characters contained in the text block, and if the number of characters is greater than a threshold, using the threshold as a maximum number of characters, dividing the text block according to preset punctuation marks to obtain initial sub-text blocks with a number of characters less than or equal to the threshold;

[0070] Correcting the boundary of the initial sub-text block by a semantic similarity calculation method based on sliding window sampling to obtain a target sub-text block;

[0071] Extracting knowledge points based on the target sub-text block;

[0072] If the number of characters is less than or equal to the threshold, knowledge points are extracted from the text block.

[0073] Exemplarily, for an extremely long text block, such as a single section with more than 2,000 (threshold) words, the threshold is used as the maximum number of characters, and the text block is divided according to preset punctuation marks (for example, period / question mark / exclamation mark) to obtain an initial sub-text block with a character number less than or equal to the threshold. The boundary of the initial sub-text block is corrected by a semantic similarity calculation method based on sliding window sampling, for example, sentences with a cosine similarity <0.6 are determined to be boundary sentences to obtain a target sub-text block.

[0074] In a possible implementation, the step of extracting knowledge points from the text block further includes:

[0075] Identifying the number of characters contained in the text block, and if the number of characters is greater than a threshold, dividing the text block into paragraphs using the threshold as the maximum number of characters to obtain secondary sub-text blocks having a number of characters less than or equal to the threshold;

[0076] Correcting the boundary of the secondary sub-text block by a semantic segmentation method based on LLM to obtain a target sub-text block;

[0077] Extracting knowledge points based on the target sub-text block;

[0078] If the number of characters is less than or equal to the threshold, knowledge points are extracted from the text block.

[0079] For example, through the LLM-based semantic segmentation method, specifically, through the LLM semantic segmentation prompt (such as "Please split the semantically complete sub-paragraph here"), it is ensured that the segmentation does not destroy the integrity of the knowledge points.

[0080] In a possible implementation, the step of extracting knowledge points from the text block to obtain target knowledge points includes:

[0081] Performing part-of-speech tagging on the words in the text block, and extracting the words in the text block that are marked as noun parts of speech based on the tagging results to obtain target vocabulary;

[0082] Calculating the co-occurrence weights of the target words based on the TextRank algorithm, sorting the target words according to the co-occurrence weights to obtain a sorting result, and screening knowledge points from all target words according to the sorting result to obtain a first knowledge point;

[0083] Based on the semantic enhancement technology of LLM, knowledge points are extracted from the text block through the hint engineering to obtain the second knowledge points;

[0084] The first knowledge point is supplemented based on the second knowledge point to obtain a target knowledge point.

[0085] For example, NLTK is used for part-of-speech tagging to extract noun phrases (e.g., “mass point”, “reference system”);

[0086] The co-occurrence weights of words are calculated based on the TextRank algorithm, and high-frequency terms are selected as candidate knowledge points (such as phrases in the top 30% of weight).

[0087] LLM semantically enhanced extraction:

[0088] For implicit knowledge points that are difficult to cover with rules (such as "applicable scenarios of the mass model"), LLM is guided to generate structured outputs through prompt engineering, for example:

[0089] Input text: "When studying the motion of an object, if the size and shape of the object have negligible effects on the problem being studied, it can be treated as a point mass."

[0090] Task: Extract core physics concepts and their definitions

[0091] Output format: {"concept": "point mass", "definition": "When studying the motion of an object, if the effects of size and shape can be ignored, it can be considered a point with mass"}.

[0092] In a possible implementation, the step of saving the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map further includes:

[0093] Filtering the target knowledge points to obtain filtered target knowledge points, wherein the filtering includes frequency filtering based on the number of occurrences of the target knowledge points in the target textbook, length filtering based on the number of characters in the target knowledge points, and dictionary filtering based on common words in the target knowledge points using a domain dictionary of the subject area of ​​the target textbook;

[0094] Based on WordNet, synonyms in the filtered target knowledge points are merged to obtain merged knowledge points;

[0095] Annotating the merged knowledge point using the type information of the merged knowledge point and the extended attribute information of the merged knowledge point as annotation information to obtain an annotated knowledge point, and sending the annotated knowledge point to a display device corresponding to the attribute feature according to a preset sending rule based on the attribute feature of the target teaching material;

[0096] Request information for the annotated knowledge point is obtained, and the annotated knowledge point is updated based on the request information to obtain an updated knowledge point, and the updated knowledge point is saved according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map.

[0097] For example, frequency filtering processing: eliminate low-frequency terms that appear less than 2 times in the textbook (such as the abbreviation of experimental equipment); length filtering processing: retain complete knowledge units of 2-50 words, and split over-long sentences (such as splitting a sentence containing 3 definitions into 3 knowledge points); domain dictionary filtering processing: perform precise / fuzzy matching through pre-built subject terminology libraries (such as the physics library contains "inertia" and "acceleration"), and filter common vocabulary (such as "analysis" and "discussion").

[0098] For example, WordNet or LLM is used to identify synonyms (e.g., “constant velocity linear motion” → “uniform linear motion”) and merge them into standard terms.

[0099] Exemplarily, the updated knowledge points are saved according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map, recording the parent-child relationship (such as "mass point" → parent node "1.1 Mass Point Reference System", child node "Application Scenario of Mass Point").

[0100] For example, let's take "Chapter 1: Description of Motion" from the People's Education Press's "High School Physics Compulsory 1" as an example:

[0101] The hierarchical blocks are: chapter-level blocks, section-level blocks, and sub-item-level blocks.

[0102] Chapter-level block: "Chapter 1: Description of Movement" (including the entire chapter text);

[0103] Section-level blocks: "1.1 Mass Reference System" "1.2 Time and Displacement";

[0104] Sub-item-level blocks: In "1.1 Point Mass Reference System," separate "Definition of Point Mass" and "Principles for Selecting a Reference System" into independent sub-item blocks. For the "Definition of Point Mass" paragraph (over 500 words), use the LLM prompt "Split semantically complete sub-paragraphs here" to split it into two sub-item blocks: "Physical Meaning of Point Mass" and "Applicability Conditions of Point Mass Model."

[0105] NLTK+TextRank locates high-frequency terms such as "point mass", "reference system", and "idealized model";

[0106] LLM supplements the extraction of implicit knowledge points. For example, if the input "When studying the motion of an object, if the size and shape of the object have negligible effects on the problem being studied, it can be regarded as a point mass" generates {"concept":"point mass","definition":"A point mass whose size and shape are ignored is used to simplify motion research"}.

[0107] In a possible implementation, the step of generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point graph according to the type information of the target knowledge point includes:

[0108] According to the type information of the target knowledge point, a question template corresponding to the type information is selected from a preset question template library, and the target knowledge point is filled into the question template to obtain a first round of questions;

[0109] Using the content of the target knowledge point as the answer data for the first round of questions to obtain a first round of question-answer pairs;

[0110] Generate an extended question based on the name of the target knowledge point and the hierarchical data of the target knowledge point in the tree-like hierarchical knowledge point map;

[0111] Using the content of the target knowledge point as answer data for the extended question to obtain an extended question-answer pair;

[0112] Extracting entity data from the content of the target knowledge point, and generating progressive follow-up questions based on the entity data through LLM's dynamic questioning technology;

[0113] Using the content of the target knowledge point as answer data for the progressive follow-up question to obtain a follow-up question-answer pair;

[0114] The first-round question-answer pairs, the expanded question-answer pairs, and the follow-up question-answer pairs are used as question-answer pairs for the target knowledge points in the tree-like hierarchical knowledge point map.

[0115] For example, basic questions are generated based on descriptive tags (type information), such as: DEF tag → "What is {knowledge point}?" / PRC tag → "What are the main steps of {knowledge point}?" Extended questions are generated by combining hierarchical information (the hierarchical data of the target knowledge point within the tree-like hierarchical knowledge point map), such as "Based on {chapter title}, explain the physical meaning of {knowledge point}."

[0116] Based on the first round of question-answer pairs, we generate progressive follow-up questions by extracting new entities (entity data, such as "point mass model" mentioning "idealized assumptions") from the answers, for example:

[0117] Q1: "What is a point mass?"

[0118] A1: "A point mass is a point with mass whose size and shape are ignored and is used to simplify the study of motion."

[0119] Q2: “What is the theoretical basis of the point mass model?”

[0120] The format of each question-answer pair is standardized, that is, each question-answer pair contains text block traceability (such as source_block:CH01.S01.S01) and hierarchical metadata (such as level: SUBSECTION), thus forming a structured dataset.

[0121] Specifically, the process of generating the question-answer pair is as follows:

[0122] {

[0123] "knowledge_point": "Point of mass",

[0124] "multiturn_dialogue": [

[0125] {"turn": 1, "question": "What is a point mass?", "answer": "…", "tags": ["DEF"]},

[0126] {"turn": 2, "question": "What are the applicable scenarios for the mass point model?", "answer": "…","tags": ["CHILD_DETAIL"]} ]

[0128] }.

[0129] For example, for the "mass point" knowledge point annotation, the description tag (type information) of "mass point" is: DEF;

[0130] Extended attributes: Difficulty level = "Required for high school", Related test points = "Basics of Physics Kinematics for College Entrance Examination";

[0131] Improve the definition: "In some cases, the size and shape of an object can be ignored and the object can be simplified to a point with mass, which is called a point mass."

[0132] Question-Answer Pair Generation:

[0133] In the first round of questions, the DEF tag triggers the question template to get Q1: "What is a point mass?"

[0134] Combined with the hierarchical information (the hierarchical data of the target knowledge point within the tree-like hierarchical knowledge point map), an extended question Q2 is generated: "Based on the content of Chapter 1, what is the core function of the mass point model?"

[0135] Based on A1, "Point-masses are used to simplify motion research," we generate a follow-up question, Q3: "What are the theoretical limitations of the point-mass model?"

[0136] Structured output:

[0137] {

[0138] "knowledge_point": "Point of mass",

[0139] "multiturn_dialogue": [

[0140] {"turn": 1, "question": "What is a point mass?", "answer": "…", "metadata": {"level": "SUBSECTION", "path": "CH01.S01.S01"}},

[0141] {"turn": 2, "question": "What are the applicable scenarios for the mass point model?", "answer": "…","metadata": {"level": "SUBSECTION", "path": "CH01.S01.S01"}} ]

[0143] }.

[0144] In one possible implementation, when training a preset large language model, multi-granularity training data is used to train the preset large language model. Specifically, the Chinese character encoding of the textbook text is used as the training data of the first granularity data to train the preset large language model, the label of the target knowledge point is encoded as a one-hot vector as the training data of the second granularity data to train the preset large language model, and the knowledge point hierarchical tree is converted into an embedding vector through a graph neural network (GNN) as the third granularity data, which is input into the attention layer of the preset large language model.

[0145] In one possible implementation, when training a pre-set large language model, data from different textbook versions (e.g., the People's Education Press edition and the Shandong University of Science and Technology edition) is alternately fed into the model. This allows the model to learn common features across textbooks (e.g., the hierarchical presentation of knowledge points and subject terminology) while mitigating version-specific differences (e.g., chapter order and case selection). An adapter mechanism is employed: a lightweight cross-textbook adapter module is added to the model to dynamically adjust parameters based on the language style of the new textbook (e.g., classical Chinese vs. vernacular Chinese), eliminating the need to retrain the entire model.

[0146] In one possible implementation, when the large language model is pre-set to a small-scale model (such as BERT-Base), all parameters are updated with a learning rate of 1e-5. This is suitable for scenarios with sufficient computing resources and ensures that the model deeply fits the textbook knowledge patterns. When the large language model is pre-set to a large-scale model (such as LLaMA-7B), LoRA technology is used, and only the low-rank matrix of the attention layer (approximately 0.1% of the parameters) is fine-tuned with a learning rate of 5e-4. This maintains the model's versatility while incorporating educational domain knowledge.

[0147] In one possible implementation, multiple rounds of question-answer pairs in the dataset are converted into a "command-response" format, such as:

[0148] Instruction: Extract the formula and physical meaning of "uniformly accelerated linear motion"

[0149] Response: The formula is \(v = v_0 + at\), which physically describes the uniform change of velocity over time.

[0150] In one possible implementation, reinforcement learning (RLHF) is used to optimize the understanding and generation capabilities of a preset large language model for instructions in the educational field.

[0151] In one possible implementation, the accuracy of question-answer pair generation is optimized through the cross-entropy loss function; the hierarchy of knowledge points output by the model is forced to match the hierarchy of the input text block (for example, chapter-level questions must not output sub-item-level details) through the hierarchical consistency loss function; and a domain adversarial loss function is set to align the model output distribution with the textbook text distribution through adversarial training to avoid generating general knowledge base content.

[0152] In a possible implementation, the step of inputting the text block of the teaching material to be extracted into the target knowledge point extraction model to obtain the target knowledge point corresponding to the teaching material to be extracted further includes:

[0153] Hierarchical path prediction: Output the directory path of the knowledge point (such as "CH02.S03.S01") to ensure that the extraction results can be mapped to the hierarchical structure of the target textbook;

[0154] Type validation rules: Use regular expressions to verify the output format. For example, formula-type knowledge points must contain mathematical symbols (such as "=" and "∑"), and definition-type knowledge points must contain keywords such as "refers to" and "called" to improve the standardization of extraction results.

[0155] In one possible implementation, when new teaching materials are extracted for the first time, the results are fed back to the training dataset, and high-confidence samples are selected through active learning for automatic labeling, continuously optimizing the model's adaptability to emerging teaching materials, forming a closed loop of "extraction → feedback → iteration".

[0156] For example, in the new textbook: the Lu Ke edition of the high school physics textbook "Description of Motion" chapter states that "To describe the motion of an object, it is necessary to select a standard object, called a reference frame." This is used as the input data of the trained preset large prediction model. The trained preset large prediction model outputs:

[0157] Hierarchical prediction: Identified as section-level knowledge points, path "CH01.S02" (corresponding to the directory structure of the Luke version);

[0158] Type verification: Match the "called" keyword through regular expressions and mark it as DEF (define class);

[0159] Structured results:

[0160] Knowledge point content: Reference frame is the object used as a standard when describing the motion of an object.

[0161] Type information: DEF

[0162] Hierarchical Path: Chapter 1 Description of Motion - 1.2 Reference and Coordinate Systems

[0163] Application of incremental learning: The new term "reference body" (synonymous with "reference system") in the extraction results of the Lu Ke version is fed back into the dataset, automatically annotated and then updated to improve the extraction accuracy of subsequent similar teaching materials.

[0164] On the other hand, Figure 2 As shown, the present application provides a knowledge point extraction system for textbooks, the system comprising:

[0165] The data receiving module 201 is used to divide the target teaching material into blocks according to the catalog data of the target teaching material to obtain text blocks;

[0166] A graph generation module 202 is configured to extract knowledge points from the text block to obtain target knowledge points, and save the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point graph;

[0167] A question-answer module 203 is configured to generate question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point graph according to the type information of the target knowledge point;

[0168] A training module 204 is configured to construct a training set based on the text block, the type information, and the question-answer pairs, and train a preset large language model provided with an adapter mechanism by fine-tuning the instructions using the training set to obtain a target knowledge point extraction model;

[0169] The extraction module 205 is used to input the text block of the teaching material to be extracted into the target knowledge point extraction model to obtain the target knowledge points corresponding to the teaching material to be extracted.

[0170] In one possible implementation, Figure 3 As shown, an embodiment of the present application provides a terminal device 300, comprising: a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, the following is achieved: dividing the target teaching material into blocks according to the directory data of the target teaching material to obtain text blocks; extracting knowledge points from the text blocks to obtain target knowledge points, and saving the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map; generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge points; constructing a training set based on the text blocks, the type information, and the question-answer pairs, and training a preset large language model provided with an adapter mechanism by fine-tuning instructions through the training set to obtain a target knowledge point extraction model; inputting the text blocks of the teaching material to be extracted into the target knowledge point extraction model to obtain target knowledge points corresponding to the teaching material to be extracted.

[0171] In one possible implementation, Figure 4 As shown, an embodiment of the present application provides a computer-readable storage medium 400, on which a computer program 411 is stored, which, when executed by a processor, implements: dividing the target teaching material into blocks according to the catalog data of the target teaching material to obtain text blocks; extracting knowledge points from the text blocks to obtain target knowledge points, and saving the target knowledge points according to the hierarchical relationship of the catalog data to obtain a tree-like hierarchical knowledge point map; generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge points; constructing a training set based on the text blocks, the type information and the question-answer pairs, and training a preset large language model provided with an adapter mechanism by fine-tuning the instructions through the training set to obtain a target knowledge point extraction model; inputting the text blocks of the teaching material to be extracted into the target knowledge point extraction model to obtain target knowledge points corresponding to the teaching material to be extracted.

[0172] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of this application. Their specific functions and technical effects can be found in the method embodiment section and will not be repeated here.

[0173] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0174] If the integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a camera / terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, removable hard drives, magnetic disks, or optical disks. In some jurisdictions, based on legislation and patent practice, computer-readable media cannot be electric carrier signals or telecommunication signals.

[0175] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0176] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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.

[0177] In the embodiments provided in this application, it should be understood that the disclosed devices / network equipment and methods can be implemented in other ways. For example, the device / network equipment embodiments described above are merely illustrative. For example, the division of the modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0178] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0179] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

[0180] The above disclosure is merely a preferred embodiment of the present invention and certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for extracting knowledge points from textbooks, characterized in that: The following steps are involved: Divide the target teaching material into blocks according to the catalogue data of the target teaching material to obtain text blocks; Extracting knowledge points from the text block to obtain target knowledge points, and saving the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map; The step of saving the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map further includes: Filtering the target knowledge points to obtain filtered target knowledge points, wherein the filtering includes frequency filtering based on the number of occurrences of the target knowledge points in the target textbook, length filtering based on the number of characters in the target knowledge points, and dictionary filtering based on common words in the target knowledge points using a domain dictionary of the subject area of ​​the target textbook; Based on WordNet, synonyms in the filtered target knowledge points are merged to obtain merged knowledge points; Annotating the merged knowledge point using the type information of the merged knowledge point and the extended attribute information of the merged knowledge point as annotation information to obtain an annotated knowledge point, and sending the annotated knowledge point to a display device corresponding to the attribute feature according to a preset sending rule based on the attribute feature of the target teaching material; Obtaining request information for the annotated knowledge point, and updating the annotated knowledge point based on the request information to obtain an updated knowledge point, and saving the updated knowledge point according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map; Generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge point; The step of generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point graph according to the type information of the target knowledge point includes: According to the type information of the target knowledge point, a question template corresponding to the type information is selected from a preset question template library, and the target knowledge point is filled into the question template to obtain a first round of questions; Using the content of the target knowledge point as the answer data for the first round of questions to obtain a first round of question-answer pairs; Generate an extended question based on the name of the target knowledge point and the hierarchical data of the target knowledge point in the tree-like hierarchical knowledge point map; Using the content of the target knowledge point as answer data for the extended question to obtain an extended question-answer pair; Extracting entity data from the content of the target knowledge point, and generating progressive follow-up questions based on the entity data through LLM's dynamic questioning technology; Using the content of the target knowledge point as answer data for the progressive follow-up question to obtain a follow-up question-answer pair; The first-round question-answer pair, the expanded question-answer pair, and the follow-up question-answer pair are used as question-answer pairs for the target knowledge point in the tree-like hierarchical knowledge point graph; Constructing a training set based on the text block, the type information, and the question-answer pair, and using the training set to train a preset large language model provided with an adapter mechanism in a manner of fine-tuning instructions to obtain a target knowledge point extraction model; The text block of the teaching material to be extracted is input into the target knowledge point extraction model to obtain the target knowledge points corresponding to the teaching material to be extracted.

2. A method for extracting knowledge points from textbooks according to claim 1, characterized in that: The step of dividing the target teaching material into blocks according to the catalogue data of the target teaching material to obtain text blocks includes: Segmenting the target teaching material into chapter-level text blocks, section-level text blocks, and sub-item-level text blocks according to the content information, level information, and page number information in the catalog data; The chapter-level text block, the section-level text block and the sub-item-level text block are all regarded as the text block, and the text block includes a hierarchical ID obtained based on the hierarchical information, a title obtained based on the content information and a text range obtained based on the page number information.

3. The method for extracting knowledge points from textbooks according to claim 1, characterized in that: The step of extracting knowledge points from the text block includes: Identifying the number of characters contained in the text block, and if the number of characters is greater than a threshold, using the threshold as a maximum number of characters, dividing the text block according to preset punctuation marks to obtain initial sub-text blocks with a number of characters less than or equal to the threshold; Correcting the boundary of the initial sub-text block by a semantic similarity calculation method based on sliding window sampling to obtain a target sub-text block; Extracting knowledge points based on the target sub-text block; If the number of characters is less than or equal to the threshold, knowledge points are extracted from the text block.

4. The method for extracting knowledge points from textbooks according to claim 1, wherein: The step of extracting knowledge points from the text block further includes: Identifying the number of characters contained in the text block, and if the number of characters is greater than a threshold, dividing the text block into paragraphs using the threshold as the maximum number of characters to obtain secondary sub-text blocks having a number of characters less than or equal to the threshold; Correcting the boundary of the secondary sub-text block by a semantic segmentation method based on LLM to obtain a target sub-text block; Extracting knowledge points based on the target sub-text block; If the number of characters is less than or equal to the threshold, knowledge points are extracted from the text block.

5. The method for extracting knowledge points from textbooks according to claim 1, characterized in that: The step of extracting knowledge points from the text block to obtain target knowledge points includes: Performing part-of-speech tagging on the words in the text block, and extracting the words in the text block that are marked as noun parts of speech based on the tagging results to obtain target vocabulary; Calculating the co-occurrence weights of the target words based on the TextRank algorithm, sorting the target words according to the co-occurrence weights to obtain a sorting result, and screening knowledge points from all target words according to the sorting result to obtain a first knowledge point; Based on the semantic enhancement technology of LLM, knowledge points are extracted from the text block through the hint engineering to obtain the second knowledge point; The first knowledge point is supplemented based on the second knowledge point to obtain a target knowledge point.

6. A knowledge point extraction system for textbooks, characterized by: The system comprises: A data receiving module, configured to divide the target teaching material into blocks according to the catalogue data of the target teaching material to obtain text blocks; The graph generation module is used to extract knowledge points from the text block to obtain target knowledge points, and save the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point graph; the step of saving the target knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point graph also includes: filtering the target knowledge points to obtain filtered target knowledge points, and the filtering includes frequency filtering for the number of times the target knowledge points appear in the target textbook, length filtering for the number of characters in the target knowledge points, and filtering for common words in the target knowledge points by comparing with the domain dictionary of the subject area of ​​the target textbook. Implemented dictionary filtering processing; merging synonyms in filtered target knowledge points based on WordNet to obtain merged knowledge points; using type information of the merged knowledge points and extended attribute information of the merged knowledge points as annotation information, annotating the merged knowledge points to obtain annotated knowledge points, and based on the attribute characteristics of the target teaching material, sending the annotated knowledge points to a display device corresponding to the attribute characteristics according to a preset sending rule; obtaining request information for the annotated knowledge points, and updating the annotated knowledge points based on the request information to obtain updated knowledge points, and saving the updated knowledge points according to the hierarchical relationship of the directory data to obtain a tree-like hierarchical knowledge point map; The question-answer module is used to generate question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge point; the step of generating question-answer pairs for each target knowledge point in the tree-like hierarchical knowledge point map according to the type information of the target knowledge point comprises: selecting a question template corresponding to the type information in a preset question template library according to the type information of the target knowledge point, and filling the target knowledge point into the question template to obtain a first-round question; using the content of the target knowledge point as the answer data for the first-round question to obtain a first-round question-answer pair; based on the target knowledge point The name of the target knowledge point and the hierarchical data of the target knowledge point in the tree-like hierarchical knowledge point map are used to generate an extended question; the content of the target knowledge point is used as the answer data for the extended question to obtain an extended question-answer pair; the entity data in the content of the target knowledge point is extracted, and according to the entity data, a progressive follow-up question is generated through the dynamic questioning technology of LLM; the content of the target knowledge point is used as the answer data for the progressive follow-up question to obtain a follow-up question-answer pair; the first-round question-answer pair, the extended question-answer pair and the follow-up question-answer pair are used as the question-answer pair for the target knowledge point in the tree-like hierarchical knowledge point map; A training module is configured to construct a training set based on the text block, the type information, and the question-answer pairs, and to train a preset large language model provided with an adapter mechanism by fine-tuning the instructions using the training set to obtain a target knowledge point extraction model; The extraction module is used to input the text block of the teaching material to be extracted into the target knowledge point extraction model to obtain the target knowledge point corresponding to the teaching material to be extracted.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for extracting knowledge points from teaching materials as described in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for extracting knowledge points from teaching materials as described in any one of claims 1 to 5 is implemented.

Images