Unlabeled course knowledge point graph construction method and system based on GLM-4

Abstract

The present invention provides a method and system for constructing a non-labeled course knowledge point map based on GLM-4, and the method includes: generating a first hierarchical structure of a knowledge point map; performing a document segmentation operation on chapters and subordinate titles at all levels according to the first hierarchical structure to obtain chapter independent units and hierarchical title independent units; performing a data cleaning operation to obtain pre-processed chapter independent units and pre-processed title independent units; performing a knowledge point acquisition operation, calling an application programming interface API of the GLM-4 model, and giving a prompt instruction, thereby using the GLM-4 model to extract knowledge points that meet the outline from the pre-processed chapter independent units and pre-processed title independent units; performing a noise removal operation using the GLM-4 model to obtain denoised data; performing a comparison using a cosine similarity algorithm; and constructing a knowledge map. The present invention solves the technical problem that the knowledge point map is too refined in granularity, resulting in difficult-to-understand knowledge.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph construction, and in particular to a method and system for constructing an unlabeled course knowledge point graph based on GLM-4 (Generalized Linear Model-4). Background Art

[0002] In the solution of integrating knowledge graph technology with course content, the existing knowledge point graph is too fine in granularity. Although this overly fine granularity can show more knowledge point details, it may also cause students to focus on tedious knowledge point fragments during the learning process, making it difficult to form an understanding of knowledge. Too many details can easily distract learners' attention, making it difficult for them to grasp the core concepts, which in turn affects their deep understanding of knowledge and their ability to apply it.

[0003] The existing invention patent application document "A knowledge point dynamic recommendation and teaching feedback system based on knowledge graph and its implementation method" with publication number CN117668239A, the knowledge point subgraph module of the existing system contains the data generated by the interaction between the user and the system, and the subgraph formed by the constructed knowledge graph. The user completes the question answering, and the knowledge points linked to the question constitute the knowledge point subgraph, reflecting the user's current mastery of the knowledge point status. The knowledge point recommendation module divides the recommendation into consolidation recommendation and expansion recommendation according to the user's answering situation. The consolidation recommendation recommends knowledge points from the knowledge point subgraph according to the recommendation weight recorded in the knowledge point subgraph module, and the expansion recommendation recommends knowledge points from the external neighborhood according to the knowledge point recommendation weight, and finally obtains the recommended knowledge points. The exercise recommendation and visualization module matches and extracts from the exercise library according to the recommended knowledge points, and presents the matched exercises and knowledge subgraphs to the user visually, realizing the recommendation and visualization of knowledge points. However, the aforementioned prior art does not utilize the understanding ability of the large model, and the accuracy and efficiency of constructing the knowledge graph are not high. At the same time, in the process of constructing the knowledge graph, the knowledge graph of the aforementioned prior art obtains knowledge points as entities, defines relationships to construct the graph, and the knowledge graph constructed in this way lacks hierarchical relationships, and the positioning of knowledge points is not clear enough.

[0004] The existing invention patent application document with publication number CN112287037A, "A method, device and storage medium for constructing a multi-entity hybrid knowledge graph", includes: adding knowledge points, students, and teaching resources as entities to the knowledge graph. Through the OCR (Optical Character Recognition) algorithm, LSTM (Long Short-Term Memory) algorithm, and TFIDF (term frequency-inverse document frequency) algorithm, the corresponding test knowledge points are matched to the teaching resources, and the relationship between the teaching resources and the knowledge points is constructed. Combined with the students' learning behavior on different teaching resources, the students' mastery of the knowledge points is obtained, and the preference for learning resources is recorded. However, the aforementioned existing scheme uses manual methods to construct knowledge graphs, which is not only time-consuming and labor-intensive, but also has low operation accuracy.

[0005] The existing public document "Application of Large Language Model in Automatic Construction of Subject Knowledge Graph" uses a large language model to automatically construct the subject knowledge graph of colleges and universities. By integrating the advantages of LLM (Large Language Model), an optimized automatic construction process is designed. Taking the subject of colleges and universities as an example, the knowledge entities and knowledge relationships based on textbooks, slides, and teaching syllabi are automatically extracted to form the final subject knowledge system. By combing the subject knowledge framework, students can understand the relationship between knowledge points in various professional courses, quickly master the key points of course knowledge, and accurately trace back the knowledge learning context to check for omissions. However, the aforementioned existing document directly incorporates the teaching syllabus into the knowledge graph, which is prone to introduce more redundant information. This existing technology simply incorporates the teaching syllabus into the knowledge graph, making it difficult to screen out knowledge points closely related to the teaching syllabus, and it is difficult to avoid the interference of redundant information and irrelevant content. This existing document does not perform sharding on the text data, which will result in too much data being provided to the large language model, which is prone to generate a lot of noise.

[0006] In summary, the existing technology has the technical problem that the knowledge point graph is too fine in granularity, making the knowledge difficult to understand. Summary of the invention

[0007] The technical problem to be solved by the present invention is: how to solve the technical problem in the prior art that the knowledge point graph is too fine in granularity, making the knowledge difficult to understand.

[0008] The present invention adopts the following technical solutions to solve the above technical problems: A method for constructing an unlabeled course knowledge point map based on GLM-4 includes:

[0009] S1. Perform a framework generation operation to extract the course content hierarchy from the document directory as the first hierarchy of the knowledge point map;

[0010] S2, obtaining each chapter and its subordinate titles at all levels in the document, and performing a document segmentation operation on the chapters and their subordinate titles at all levels according to the first hierarchical structure to obtain chapter independent units and hierarchical title independent units;

[0011] S3, performing data cleaning operations on the chapter independent units and the graded title independent units to obtain pre-processed chapter independent units and pre-processed title independent units;

[0012] S4, perform knowledge point acquisition operation, call the application programming interface API (Application Programming Interface) of the GLM-4 model, and give prompt instructions to extract knowledge points that meet the outline from the pre-processed chapter independent units and pre-processed title independent units;

[0013] S5, perform noise removal operation, send the knowledge points that meet the outline into the GLM-4 model for processing to remove noise, supplement missing data, and obtain denoised data;

[0014] S6, performing a data verification operation, comparing the denoised data with the content of the outline through a cosine similarity algorithm, so that all knowledge points in the outline are extracted and the verified data is obtained;

[0015] S7. Construct a knowledge graph, use Python tools to batch process the verified data, and store it in the Neo4j database to construct a knowledge point graph.

[0016] In the knowledge point map constructed by the present invention, each entity corresponds to a complete knowledge point rather than a subdivided part of the knowledge point, which can make the granularity of the knowledge point map more moderate and avoid the limitations caused by over-refinement; the knowledge point extraction process completely relies on the GLM-4 model, which has strong understanding and analysis capabilities to ensure the accuracy of the knowledge points and the efficiency of the extraction operation; in order to further simplify the manual review process, the present invention uses the GLM-4 model to conduct a secondary review of the data that meets the outline knowledge points, effectively removes data noise, and improves data purity; in order to ensure that there are no omissions in the outline knowledge points, the present invention uses the GLM-4 model for secondary processing, and finally uses similarity comparison technology to accurately determine whether all knowledge points in the outline have been extracted, thereby ensuring the integrity of the knowledge point map.

[0017] In a more specific technical solution, in S1, the directory of the document is parsed using a preset large model, and the corresponding Cypher language is generated according to the demand information;

[0018] Using Python tools, regular expressions are used to match corresponding contents from the directory and store them in the Neo4j database as a course content hierarchy.

[0019] The present invention ensures the logical rigor and structural clarity of the knowledge point graph, which not only improves the efficiency of data processing, but also reduces the need for manual intervention, providing a solid foundation for building the framework of the entire knowledge management system.

[0020] In a more specific technical solution, in S4, the prompt command is used to instruct the GLM-4 model to capture significant knowledge points according to the outline at the secondary title level, and simultaneously extract the definition of the significant knowledge points, the original text content, and the example questions;

[0021] In the absence of a second-level title, the prompt command is used to guide the GLM-4 model to extract knowledge points at the first-level title level;

[0022] The GLM-4 model returns structured knowledge point data according to the preset format, which is then processed to obtain knowledge points that conform to the syllabus.

[0023] In a more specific technical solution, in S5, the knowledge points that meet the outline are submitted to the GLM-4 model, and the prompt instructions corresponding to the knowledge points that meet the outline are submitted;

[0024] Using the GLM-4 model, identify and remove redundant information in the document;

[0025] Verify the accuracy of knowledge point definitions, original text expressions and examples in the document.

[0026] The prompt instruction used in the present invention triggers the GLM-4 model to review the data in the knowledge points that meet the outline, and uses the review operation to ensure the integrity of the JSON (JavaScript Object Notation) data, without missing any important knowledge points, and ensure that the definitions, original texts and examples corresponding to the knowledge points are complete; identify and remove all redundant information that is not related to the original document; verify the accuracy of the data, including: verifying the accuracy of the knowledge point definitions, original text expressions and examples. By using the GLM-4 model to perform the aforementioned self-checking process, the purity and reliability of the data can be improved.

[0027] In a more specific technical solution, S6 includes:

[0028] S61. Build a BERT (Bidirectional Encoder Representations from Transformers) model, use CourseKG (Course Knowledge Graph) to optimize the BERT model, convert the Chinese text in the syllabus into word vectors with text information, process them to obtain word embedding vectors, input the word embedding vectors into the BERT model for feature extraction, and obtain semantic feature sequence vectors;

[0029] S62. Use the recurrent neural network (RNN) to capture the dependencies in the semantic feature sequence vector. In CourseKG, combine the forward GRU (Gated Recurrent Unit) neural network model and the backward GRU neural network model to obtain the BiGRU (Bidirectional, Gated Recurrent Unit) model. Use the BiGRU model to extract the key features of the education record specified entity and process it to obtain the BiGRU output information.

[0030] S63. Using the MHSA (Multi-Head Self-Attention Mechanism) model, the semantic feature sequence vectors at different levels and angles are combined according to the MHSA mechanism to obtain a comprehensive interactive representation; wherein, according to the MHSA mechanism, the entity recognition feature information is extracted from the BiGRU output information, wherein the result matrix generated by the BiGRU model includes: query matrix , key matrix and the value matrix , , Represents the current memory cell, is the batch parameter, where , Represents the total number of memory cells; SoftMax operation is performed based on the result matrix;

[0031] S64. Multiply the result matrix by the value matrix Derived self-attention interaction aggregate information to represent the context of the document;

[0032] S65, performing a multi-self-attention head aggregation operation according to the self-attention interaction aggregation information;

[0033] S66, using a CRF (Constant Rate Factor) decoder, derive an optimal prediction sequence based on the links between adjacent entity labels;

[0034] S67. According to the cosine similarity algorithm, use the following logic to calculate the data of the outline In line with the syllabus knowledge points to determine the knowledge point status of the outline:

[0035]

[0036] in, represents the cosine function, Data representing an outline In line with the syllabus knowledge points Similar parameters.

[0037] In the data verification operation of the present invention, the knowledge points that have been noise-free and meet the outline are processed to ensure the consistency between the extracted knowledge points and the outline content. The similarity comparison technology is used to compare and analyze the processed outline data with the knowledge points that have been noise-free and meet the outline. The present invention accurately verifies whether the knowledge points in the outline have been fully extracted, thereby ensuring the comprehensiveness and accuracy of the knowledge point map construction operation in terms of content.

[0038] In a more specific technical solution, the recursive neural network RNN ​​in S62 includes: an embedding layer, a GRU layer and a softmax layer; wherein the embedding layer generates a word embedding vector for each word using the word2vec algorithm, and in the GRU layer, for each input word, the GRU unit is cyclically used according to the following logic:

[0039]

[0040]

[0041]

[0042]

[0043] in, is the input of the GRU unit, is the hidden state output of the GRU unit, is the position in the word sequence, is the reset gate corresponding to the position in the word sequence, is the update gate corresponding to the position in the word sequence, is the hidden state output of the previous position of the position in the word sequence of the GRU unit, It is the update gate The weight matrix of It is the reset gate The weight matrix of is the hidden state gate The weight matrix of It is the update gate The weight matrix from the input layer to the hidden layer; It is the reset gate The weight matrix from the input layer to the hidden layer; is the hidden state gate The corresponding weight matrix from the input layer to the hidden layer, is the sigmoid function, represents the hyperbolic tangent activation function, is the element-wise product, represents the candidate hidden layer.

[0044] In a more specific technical solution, in S64, the following logic is used to represent the context :

[0045]

[0046] in, represents a single self-attention head, represents the query matrix, Represents the key matrix and Value matrix, represents the self-attention head, is the activation function, is the batch parameter, Represents the transpose of the key matrix.

[0047] In a more specific technical solution, in S65, the query matrix , key matrix and the value matrix Connect the output of the projection to complete the multi-self-attention head aggregation operation and obtain the self-attention head:

[0048]

[0049] in, represents the query matrix corresponding to each attention head; represents the key matrix corresponding to each attention head; represents the value matrix corresponding to each attention head, It means to aggregate the query matrix, key matrix, and value matrix corresponding to each attention head. , represents the connection of the self-attention head, Indicates Pay attention to the head.

[0050] In a more specific technical solution, in S66, the scoring function of the CRF decoder is determined using the following logic: :

[0051]

[0052] in, is the input sequence , Indicates Inputs, is the label sequence of the predicted output, is the label number, For the The observation matrix corresponding to the label sequence , Indicates Tag sequence , No. Tag sequence The corresponding transfer matrix;

[0053] For a given input sequence , using the following logic, according to the scoring function , expression tag sequence The conditional probability :

[0054]

[0055] in, represents all feasible label sequences for a given sentence, Represents the scoring function Find the natural exponential function, Represents the input sequence and candidate labels The natural exponential function of

[0056] In the decoding stage, the following logic is used to express the optimal sequence label :

[0057]

[0058] in, The function for finding the maximum subscript of a sequence, Represents an optimization function.

[0059] In a more specific technical solution, the unlabeled course knowledge point graph construction system based on GLM-4 includes:

[0060] The framework generation module is used to perform the framework generation operation and extract the course content hierarchy from the document directory as the first hierarchy of the knowledge point map;

[0061] The document segmentation module is used to obtain each chapter and its subordinate titles at all levels in the document, and perform document segmentation operations on the chapters and their subordinate titles at all levels according to the first hierarchical structure to obtain chapter independent units and hierarchical title independent units. The document segmentation module is connected to the framework generation module;

[0062] A data cleaning module is used to perform data cleaning operations on the chapter independent unit and the graded title independent unit to obtain pre-processed chapter independent unit and pre-processed title independent unit. The data cleaning module is connected to the document segmentation module;

[0063] The knowledge point acquisition module is used to perform knowledge point acquisition operations, call the application programming interface API of the GLM-4 model, and give prompt instructions to extract knowledge points that meet the outline from the pre-processed chapter independent units and pre-processed title independent units. The knowledge point acquisition module is connected to the data cleaning module;

[0064] The noise removal module is used to perform noise removal operations and send the knowledge points that meet the outline into the GLM-4 model for processing to remove noise, supplement missing data, and obtain denoised data. The noise removal module is connected to the knowledge point acquisition module;

[0065] The data checking module is used to perform data checking operations, and compares the denoised data with the content of the outline through the cosine similarity algorithm, so that all knowledge points in the outline are extracted and the checked data is obtained. The data checking module is connected to the noise removal module;

[0066] The knowledge graph construction module is used to construct the knowledge graph. It uses Python tools to batch process the verified data and store it in the Neo4j database to construct a knowledge point graph. The knowledge graph construction module is connected to the data verification module.

[0067] Compared with the prior art, the main technical difference of the present invention is that the present invention uses the GLM-4 large model to construct a knowledge point map. By using the strong understanding ability of the large model, knowledge points are obtained from textbooks and a knowledge point map is constructed, which is more accurate and efficient. In terms of knowledge point map construction, the knowledge map of the prior art obtains knowledge points as entities, defines relationships to construct a map, and the knowledge map constructed in this way lacks a hierarchical relationship. The present invention adds a hierarchical relationship such as a first-level title and a second-level title in the process of constructing a knowledge point map, so that the positioning of the knowledge points is clearer.

[0068] The present invention uses the GLM large model to construct a knowledge graph in an unsupervised manner, fully utilizing the understanding ability of the large model to automatically extract chapters, first-level titles, second-level titles and knowledge points from the textbooks, and finally automatically stores them in the Neo4j database.

[0069] Compared with the document "Application of Large Language Model in Automatic Construction of Subject Knowledge Graph", which directly incorporates the syllabus into the knowledge graph, no redundant information will be introduced. The present invention does not simply incorporate the syllabus into the knowledge point map, but uses the syllabus to constrain the GLM-4 model, thereby screening out knowledge points closely related to the syllabus, effectively avoiding the interference of redundant information and irrelevant content. The document "Application of Large Language Model in Automatic Construction of Subject Knowledge Graph" does not mention the processing of text data, but the present invention first performs data segmentation and then provides it to the GLM-4 large model for knowledge point extraction. This method can effectively solve the problem of a large amount of noise generated by too much data provided to the large language model. The knowledge point map constructed by the present invention not only includes knowledge points, but also integrates teaching resources, including: examples. This design can better assist students in learning, help students understand and master knowledge points, and provide targeted exercises, thereby improving learning effects.

[0070] Compared with the prior art, the present invention has the following advantages:

[0071] In the knowledge point map constructed by the present invention, each entity corresponds to a complete knowledge point rather than a subdivided part of the knowledge point, which can make the granularity of the knowledge point map more moderate and avoid the limitations caused by over-refinement; the knowledge point extraction process completely relies on the GLM-4 model, which has strong understanding and analysis capabilities to ensure the accuracy of the knowledge points and the efficiency of the extraction operation; in order to further simplify the manual review process, the present invention uses the GLM-4 model to conduct a secondary review of the data that meets the outline knowledge points, effectively removes data noise, and improves data purity; in order to ensure that there are no omissions in the outline knowledge points, the present invention uses the GLM-4 model for secondary processing, and finally uses similarity comparison technology to accurately determine whether all knowledge points in the outline have been extracted, thereby ensuring the integrity of the knowledge point map.

[0072] The present invention ensures the logical rigor and structural clarity of the knowledge point graph, which not only improves the efficiency of data processing, but also reduces the need for manual intervention, providing a solid foundation for building the framework of the entire knowledge management system.

[0073] The prompt instruction used in the present invention triggers the GLM-4 model to review the data in the knowledge points that meet the outline, and uses the review operation to ensure the integrity of the JSON data, without missing any important knowledge points, and ensure that the definitions, original texts and examples corresponding to the knowledge points are complete; identify and remove all redundant information that is not related to the original document; verify the accuracy of the data, including: verifying the accuracy of the knowledge point definitions, original text expressions and examples. By using the GLM-4 model to perform the aforementioned self-checking process, the purity and reliability of the data can be improved.

[0074] In the data verification operation of the present invention, the knowledge points that have been noise-free and meet the outline are processed to ensure the consistency between the extracted knowledge points and the outline content. The similarity comparison technology is used to compare and analyze the processed outline data with the knowledge points that have been noise-free and meet the outline. The present invention accurately verifies whether the knowledge points in the outline have been fully extracted, thereby ensuring the comprehensiveness and accuracy of the knowledge point map construction operation in terms of content.

[0075] The present invention solves the technical problem in the prior art that the knowledge point graph is too fine in granularity, making the knowledge difficult to understand. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 A schematic diagram of basic steps of a method for constructing an unlabeled course knowledge point graph based on GLM-4 according to Example 1 of the present invention;

[0077] Figure 2 This is a data flow processing diagram of a method for constructing an unlabeled course knowledge point graph based on GLM-4 according to Example 1 of the present invention;

[0078] Figure 3 A schematic diagram of a knowledge point graph schema design for the method of Embodiment 1 of the present invention;

[0079] Figure 4 This is a schematic diagram of the specific steps of performing a data verification operation in Example 2 of the present invention. DETAILED DESCRIPTION

[0080] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] Example 1

[0082] like Figure 1 and Figure 2 As shown in the figure, the method for constructing an unlabeled course knowledge point map based on GLM-4 includes the following basic steps:

[0083] S1. Perform a framework generation operation to extract the course content hierarchy from the document directory as the first hierarchy of the knowledge point map;

[0084] In this embodiment, through the framework generation operation, the first hierarchical structure of the knowledge point map can be preliminarily generated according to the directory in the document, laying the foundation for the subsequent knowledge point extraction operation and knowledge point map construction operation;

[0085] As shown Figure 3 in the figure, in this embodiment, through the framework generation operation, the hierarchical structure of the course content is extracted from the document directory, including but not limited to: chapters, first-level headings, and second-level headings. On this basis, in the Neo4j graph database, according to the hierarchical relationship that chapters contain first-level headings and first-level headings contain second-level headings, the first-level structure of the knowledge point graph is constructed. Through the above preliminary construction process, the logical rigor and structural clarity of the knowledge point graph are ensured, providing a basis for the overall framework construction operation of the knowledge management system.

[0086] In this embodiment, the basic hierarchical structure of the knowledge point graph is carefully constructed. The above process can utilize the understanding ability of the large model to parse the document directory and generate the corresponding Cypher language according to requirements. To avoid introducing noise and simplify the traditional manual verification process, in this embodiment, for example, Python can be used to process and directly match the corresponding content from the directory through regular expressions. For example, by using the regular expression "(?<=第[\d一二三四五六七八九十]+章\s).*", the chapter information can be accurately matched, and the complete data can be obtained and directly stored in the Neo4j database.

[0087] S2. Obtain each chapter and its subordinate headings in the document, and perform document segmentation operations on the chapters and their subordinate headings according to the first-level structure to obtain chapter independent units and hierarchical heading independent units;

[0088] In this embodiment, through the document segmentation operation, according to the first-level structure, each chapter and its subordinate headings in the document are effectively segmented to ensure the independence of knowledge points;

[0089] In this embodiment, through the document segmentation operation, the text content in the document is accurately segmented according to the first-level structure constructed in the Neo4j database. Specifically, the text is divided into chapter independent units and hierarchical heading independent units bounded by chapters, first-level headings, and second-level headings, reducing the amount of data sent to the GLM-4 model, enabling the GLM-4 model to extract knowledge points more efficiently and accurately. In addition, by maintaining the first-level structure, the orderliness of the subsequent data processing process is ensured, avoiding data chaos, and laying the first-level data foundation for the accurate construction of the knowledge point graph.

[0090] S3. Perform data cleaning operations on the chapter independent units and hierarchical heading independent units to obtain preprocessed chapter independent units and preprocessed heading independent units;

[0091] In this embodiment, the data is pre-processed through the aforementioned data cleaning operation to remove irrelevant information, thereby creating conditions for the knowledge point extraction operation;

[0092] In the data cleaning process of this embodiment, data segmentation and deep cleaning are performed to remove elements corresponding to irrelevant information and non-text information. Specifically, non-text information includes but is not limited to: pictures, tables. For mathematical formula expressions in documents, mathematical formula expressions are uniformly converted to, for example, Latex format to improve the standardization of text data. Through the implementation of the aforementioned data cleaning operation, formula display errors in the text are reduced, for example: avoid converting the number: 2 2 Incorrectly displayed as normal text: 22, thereby reducing the probability of misunderstanding and ensuring the accuracy and reliability of the knowledge extraction process.

[0093] S4, perform knowledge point acquisition operation, call the application programming interface API of the GLM-4 model, and give prompt instructions to extract knowledge points that meet the outline from the pre-processed chapter independent units and pre-processed title independent units;

[0094] In the knowledge point acquisition operation of this embodiment, a directive prompt instruction is provided to guide the GLM-4 model to accurately identify and extract key knowledge points in the data parsing process. The prompt specifically instructs the GLM-4 model to capture knowledge points according to the outline at the secondary title level, and simultaneously extract the definition of the aforementioned knowledge points, the original text content, and related examples. In the case of the lack of a secondary title, the prompt quality guides the GLM-4 model to extract knowledge points at the primary title level. In this embodiment, the GLM-4 model uses, for example, JSON format to return screened and structured knowledge points that meet the outline.

[0095] S5, perform noise removal operation, send the knowledge points that meet the outline into the GLM-4 model for processing to remove noise, supplement missing data, and obtain denoised data;

[0096] In this embodiment, in the aforementioned noise removal operation, the JSON data initially output by the GLM-4 model is checked. In view of the possibility of residual noise data in the initial extraction process, as well as the high cost and time consumption of manual review, in this embodiment, the data is submitted to the GLM-4 model again with the corresponding prompt instruction.

[0097] In this embodiment, the JSON data initially output by the GLM-4 model is subjected to an in-depth inspection. The noise data remaining in the initial extraction process includes, but is not limited to, irrelevant information, incomplete information, plain text information, missing data, omitted data, and examples lacking a problem-solving process, in order to avoid manual review that consumes costs and time. In this embodiment, the data in accordance with the outline knowledge points are submitted to the GLM-4 model again, combined with the corresponding prompt instructions. The GLM-4 model is controlled to review the data in accordance with the outline knowledge points. Specifically, during the aforementioned review process, the integrity of the JSON data is ensured, no knowledge points are omitted, and the integrity of the definitions, original texts, and examples corresponding to the knowledge points is ensured. For example: instruct the GLM-4 model to obtain the plain text corresponding to the complete knowledge points, and the examples include a complete problem-solving process; identify and remove all redundant information irrelevant to the original document; verify the accuracy of the data, and verify the accuracy of the knowledge point definitions, original text expressions, and examples.

[0098] Through the aforementioned self-checking process, the GLM-4 model is used to improve the data purity and reliability.

[0099] S6, performing a data verification operation, comparing the denoised data with the content of the outline through a cosine similarity algorithm, so that all knowledge points in the outline are extracted and the verified data is obtained;

[0100] like Figure 4 As shown, in this embodiment, step S6 of performing the data verification operation also includes the following specific steps:

[0101] S61. Build a BERT (Bidirectional Encoder Representations from Transformers) model, use CourseKG to optimize the BERT model, convert the Chinese characters in the outline into word vectors with text information, process them to obtain word embedding vectors, input the word embedding vectors into the BERT model for feature extraction, and obtain semantic feature sequence vectors;

[0102] In this embodiment, in view of the scarcity of corpora in the field of education, this embodiment uses the fine-tuning tool in CourseKG to tune the BERT model to improve the data processing accuracy in the field of education. In the aforementioned tuning process, only words related to education are masked to ensure that the word embedding vector can accurately reflect the text semantic information of the outline. The word embedding vector is input into the BERT model for feature extraction to obtain a semantic feature sequence vector.

[0103] S62. Using the recurrent neural network RNN, the dependency relationship in the semantic feature sequence vector is captured; in CourseKG, the forward GRU neural network model and the backward GRU neural network model are combined to obtain the BiGRU model, and the key feature education record designated entity is extracted using the BiGRU model, and the BiGRU output information is obtained by processing;

[0104] In this embodiment, the recursive neural network RNN ​​includes: an embedding layer, a GRU layer and a softmax layer. Specifically. The embedding layer generates a word embedding vector for each word using the word2vec algorithm, wherein the text corpus is constructed based on the input data. The GRU layer includes: a GRU unit, which adds and deletes gate structure control information, and the gate structure includes but is not limited to: an update gate and an output gate. In the GRU layer, the model cyclically uses the GRU unit for each input word, and the specific implementation formula is as follows:

[0105]

[0106]

[0107]

[0108]

[0109] However, the simple GRU layer cannot fully utilize the contextual information in the education records. In CourseKG, the BiGRU model is used to extract the key features of the specified entities in the education records.

[0110] S63, using the MHSA model, combining the semantic feature sequence vectors of different levels and angles according to the MHSA mechanism, to obtain a comprehensive interactive representation; wherein, according to the MHSA mechanism, entity recognition feature information is extracted from the output information of the BiGRU model, wherein the result matrix generated by the BiGRU model includes: query matrix , key matrix and the value matrix , , Represents the current memory cell, is the batch parameter, where , Represents the total number of memory cells; SoftMax operation is performed based on the result matrix;

[0111] S64. Multiply the result matrix by the value matrix Derived self-attention interaction aggregate information to represent the context of the document;

[0112] In this embodiment, for a single self-attention head and a group , representing the context :

[0113]

[0114] S65, performing a multi-self-attention head aggregation operation according to the self-attention interaction aggregation information;

[0115] In this embodiment, the query matrix , key matrix and the value matrix Connect the output of the projection to complete the multi-self-attention head aggregation operation and obtain the self-attention head:

[0116]

[0117] In this embodiment, the MHSA module includes: no less than 2 MHSA layers. In the MHSA model, long-range characteristics and global information in the denoised data are fully captured.

[0118] S66, using the CRF decoder, deriving the optimal prediction sequence according to the links between adjacent entity labels;

[0119] In this embodiment, the CRF decoder is used to autonomously learn restrictive rules to reduce the probability of non-logical sequences appearing in the predicted sequence. In this embodiment, the scoring function of the CRF decoder is determined by the following formula: :

[0120]

[0121] For a given input sequence , using the following logic, according to the scoring function , expression tag sequence The conditional probability :

[0122]

[0123] In the decoding phase, the optimal sequence marker The expression is as follows:

[0124]

[0125] S67. Calculate the outline data based on the cosine similarity algorithm In line with the syllabus knowledge points similarity to determine the knowledge point status of the outline;

[0126] In this embodiment, the cosine similarity calculation formula is as follows:

[0127] .

[0128] In this embodiment, the following table 1 shows the data of the outline In line with the syllabus knowledge points Cosine similarity of:

[0129] Table 1 Knowledge point similarity table

[0130]

[0131] According to Table 1 above, it can be judged whether the knowledge points obtained by the aforementioned GLM-4 model completely include the knowledge points in the syllabus.

[0132] S7. Construct a knowledge graph, use Python tools to batch process the verified data, and store it in the Neo4j database to construct a knowledge point graph.

[0133] In the knowledge point graph construction operation in this embodiment, the processed knowledge points in the verified data are added to the Neo4j database. Since the entities are complete knowledge points, the relationships between entities include: inclusion relationships, and the example relationship corresponding to the knowledge point includes: the subordinate relationship between the example and the knowledge point.

[0134] In this embodiment, a knowledge point node is created in the Neo4j database based on the processed knowledge point, and the node label can be set to, for example, Knowledge. The corresponding secondary title and primary title are found in the first hierarchical structure generated in the Neo4j database from the framework generation operation, and an inclusion relationship is created between the knowledge point and the corresponding secondary title and primary title. The definition of the knowledge point and the corresponding plain text are both attributes of the knowledge point node. A node with a node label Node Label of Example is created separately for the example question, and the relationship between the example question and the corresponding knowledge point includes: the subordinate relationship between the example question and the knowledge point. According to the above process, batch processing is performed using Python and stored in the Neo4j database.

[0135] In summary, in the knowledge point map constructed by the present invention, each entity corresponds to a complete knowledge point rather than a subdivided part of the knowledge point, which can make the granularity of the knowledge point map more moderate and avoid the limitations caused by over-refinement; the knowledge point extraction process completely relies on the GLM-4 model, which has strong understanding and analysis capabilities to ensure the accuracy of the knowledge points and the efficiency of the extraction operation; in order to further simplify the manual review process, the present invention uses the GLM-4 model to conduct a secondary review of the data that meets the outline knowledge points, effectively removes data noise, and improves data purity; in order to ensure that there are no omissions in the outline knowledge points, the present invention uses the GLM-4 model for secondary processing, and finally uses the similarity comparison technology to accurately determine whether all the knowledge points in the outline have been extracted, thereby ensuring the integrity of the knowledge point map.

[0136] The present invention ensures the logical rigor and structural clarity of the knowledge point graph, which not only improves the efficiency of data processing, but also reduces the need for manual intervention, providing a solid foundation for building the framework of the entire knowledge management system.

[0137] The prompt instruction used in the present invention triggers the GLM-4 model to review the data in the knowledge points that meet the outline, and uses the review operation to ensure the integrity of the JSON data, without missing any important knowledge points, and ensure that the definitions, original texts and examples corresponding to the knowledge points are complete; identify and remove all redundant information that is not related to the original document; verify the accuracy of the data, including: verifying the accuracy of the knowledge point definitions, original text expressions and examples. By using the GLM-4 model to perform the aforementioned self-checking process, the purity and reliability of the data can be improved.

[0138] In the data verification operation of the present invention, the knowledge points that have been noise-free and meet the outline are processed to ensure the consistency between the extracted knowledge points and the outline content. The similarity comparison technology is used to compare and analyze the processed outline data with the knowledge points that have been noise-free and meet the outline. The present invention accurately verifies whether the knowledge points in the outline have been fully extracted, thereby ensuring the comprehensiveness and accuracy of the knowledge point map construction operation in terms of content.

[0139] The present invention solves the technical problem in the prior art that the knowledge point graph is too fine in granularity, making the knowledge difficult to understand.

[0140] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, 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 embodiments of the present invention.

Claims

1. A method for constructing an unlabeled course knowledge point map based on GLM-4, characterized in that: The method comprises: S1. Perform a framework generation operation to extract the course content hierarchy from the document directory as the first hierarchy of the knowledge point map; S2, obtaining each chapter and its subordinate titles at all levels in the document, and performing a document segmentation operation on the chapters and their subordinate titles at all levels according to the first hierarchical structure to obtain chapter independent units and hierarchical title independent units; S3, performing data cleaning operations on the chapter independent units and the graded title independent units to obtain pre-processed chapter independent units and pre-processed title independent units; S4, perform knowledge point acquisition operation, call the application programming interface API of the GLM-4 model, and give prompt instructions to extract knowledge points that meet the outline from the pre-processed chapter independent units and pre-processed title independent units; S5, perform noise removal operation, send the knowledge points that meet the outline into the GLM-4 model for processing to remove noise, supplement missing data, and obtain denoised data; S6, performing a data verification operation, comparing the denoised data with the content of the outline through a cosine similarity algorithm, so that all knowledge points in the outline are extracted and the verified data is obtained; In S6, a BERT model is constructed, CourseKG is used to optimize the BERT model, and feature extraction is performed to obtain a semantic feature sequence vector; Use recurrent neural network (RNN) to capture dependencies in semantic feature sequence vectors; Using the MHSA model, the semantic feature sequence vectors at different levels and angles are combined according to the MHSA mechanism to obtain a comprehensive interactive representation; According to the cosine similarity algorithm, the outline data is calculated In line with the syllabus knowledge points similarity to determine the knowledge point status of the outline; S7. Construct a knowledge graph, use Python tools to batch process the verified data, and store it in the Neo4j database to construct a knowledge point graph.

2. The method for constructing an unlabeled course knowledge point map based on GLM-4 according to claim 1, characterized in that: The S1 includes: S11, parsing the directory of the document using a preset large model, and generating a corresponding Cypher language according to the demand information; S12. Using the Python tool, match corresponding content from the directory through regular expressions, and store the corresponding content in the Neo4j database as the course content hierarchy.

3. The method for constructing an unlabeled course knowledge point map based on GLM-4 according to claim 1, characterized in that: The S4 includes: S41, using the prompt command to instruct the GLM-4 model to capture significant knowledge points according to the outline at the secondary title level, and simultaneously extract the definition of the significant knowledge points, the original text content, and the example questions; S42, when the secondary title is missing, using the prompt instruction to guide the GLM-4 model to extract knowledge points at the primary title level; S43. The GLM-4 model returns structured knowledge point data according to a preset format, and processes the data to obtain the knowledge points that conform to the syllabus.

4. The method for constructing an unlabeled course knowledge point map based on GLM-4 according to claim 1, characterized in that: The S5 includes: S51, submitting the knowledge points that meet the syllabus to the GLM-4 model, and submitting the prompt instructions corresponding to the knowledge points that meet the syllabus; S52, using the GLM-4 model to identify and remove redundant information in the document; S53, verifying the accuracy of the knowledge point definitions, original text expressions and examples in the document.

5. The method for constructing an unlabeled course knowledge point map based on GLM-4 according to claim 1, characterized in that: The S6 includes: S61, converting the Chinese text in the outline into a word vector with text information, processing to obtain a word embedding vector, inputting the word embedding vector into the BERT model for feature extraction, and obtaining the semantic feature sequence vector; S62, in the CourseKG, combining the forward GRU neural network model and the backward GRU neural network model to obtain a BiGRU model, using the BiGRU model to extract key feature education record designated entities, and processing to obtain BiGRU output information; S63, extracting entity recognition feature information from the BiGRU output information according to the MHSA mechanism, wherein the result matrix generated by the BiGRU model includes: query matrix , key matrix and the value matrix , , Represents the current memory cell, is the batch parameter, where , represents the total number of memory cells; performing SoftMax operation according to the result matrix; S64, multiplying the result matrix by the value matrix deriving self-attention interaction aggregate information to characterize the context of the document; S65, performing a multi-self-attention head aggregation operation according to the self-attention interaction aggregation information; S66, using the CRF decoder, deriving the optimal prediction sequence according to the links between adjacent entity labels; S67, according to the cosine similarity algorithm, using the following logic to calculate the data of the outline The knowledge points that are consistent with the syllabus to determine the knowledge point status of the outline: In the formula, represents the cosine function, Data representing the outline The knowledge points that are consistent with the syllabus Similar parameters.

6. The method for constructing an unlabeled course knowledge point map based on GLM-4 according to claim 5, characterized in that: The recursive neural network RNN ​​in S62 includes: an embedding layer, a GRU layer and a softmax layer; wherein the embedding layer generates the word embedding vector for each word using the word2vec algorithm, and in the GRU layer, for each input word, the GRU unit is cyclically used according to the following logic: In the formula, is the input of the GRU unit, is the hidden state output of the GRU unit, is the position in the word sequence, is the reset gate corresponding to the position in the word sequence, is the update gate corresponding to the position in the word sequence, is the hidden state output of the previous position of the position in the word sequence of the GRU unit, It is the update gate The weight matrix of It is the reset gate The weight matrix of is the hidden state gate The weight matrix of is the update gate The weight matrix from the input layer to the hidden layer; is the reset gate The weight matrix from the input layer to the hidden layer; is the hidden state gate The corresponding weight matrix from the input layer to the hidden layer, is the sigmoid function, represents the hyperbolic tangent activation function, is the element-wise product, represents the candidate hidden layer.

7. The method for constructing an unlabeled course knowledge point map based on GLM-4 according to claim 5, characterized in that: In S64, the following logic is used to characterize the context : In the formula, represents a single self-attention head, represents the query matrix, Represents the key matrix and Value matrix, represents the self-attention head, is the activation function, is the batch parameter, represents the transpose of the key matrix.

8. The method for constructing an unlabeled course knowledge point map based on GLM-4 according to claim 5, characterized in that: In S65, the query matrix , the bond matrix And the value matrix The output of the projection is connected to complete the multi-self-attention head aggregation operation and obtain the self-attention head: In the formula, represents the query matrix corresponding to each attention head; represents the key matrix corresponding to each of the self-attention heads; represents the value matrix corresponding to each self-attention head, represents aggregation of the query matrix, the key matrix, and the value matrix corresponding to each attention head, , represents the connection of the self-attention head, Indicates Each of them pays attention to the head.

9. The method for constructing an unlabeled course knowledge point map based on GLM-4 according to claim 1, characterized in that: In S66, the scoring function of the CRF decoder is determined using the following logic: : In the formula, is the input sequence , Indicates Inputs, is the label sequence of the predicted output, is the label number, For the The observation matrix corresponding to the label sequence , Indicates The tag sequence , No. The tag sequence The corresponding transfer matrix; For the given input sequence , using the following logic, according to the scoring function , expressing the tag sequence The conditional probability : In the formula, represents all feasible label sequences for a given sentence, Represents the scoring function Find the natural exponential function, Represents the input sequence and candidate labels The natural exponential function of In the decoding stage, the following logic is used to express the optimal sequence label : In the formula, The function for finding the maximum subscript of a sequence, Represents an optimization function.

10. A system for constructing an unlabeled course knowledge point map based on GLM-4, used to execute the method for constructing an unlabeled course knowledge point map based on GLM-4 as claimed in any one of claims 1 to 9, characterized in that: The system comprises: The framework generation module is used to perform the framework generation operation and extract the course content hierarchy from the document directory as the first hierarchy of the knowledge point map; A document segmentation module, used to obtain each chapter and subordinate titles at all levels in the document, and perform document segmentation operations on the chapters and subordinate titles at all levels according to the first hierarchical structure to obtain chapter independent units and hierarchical title independent units, wherein the document segmentation module is connected to the framework generation module; A data cleaning module, used for performing data cleaning operations on the chapter independent units and the graded title independent units to obtain pre-processed chapter independent units and pre-processed title independent units, the data cleaning module being connected to the document segmentation module; A knowledge point acquisition module is used to perform knowledge point acquisition operations, call the application programming interface API of the GLM-4 model, and give a prompt instruction to extract knowledge points that meet the outline from the pre-processed chapter independent unit and the pre-processed title independent unit. The knowledge point acquisition module is connected to the data cleaning module; A noise removal module is used to perform a noise removal operation, and send the knowledge points that meet the outline into the GLM-4 model for processing to remove noise, supplement missing data, and obtain denoised data. The noise removal module is connected to the knowledge point acquisition module; A data checking module is used to perform a data checking operation, and compares the denoised data with the content of the outline through a cosine similarity algorithm, so that all knowledge points in the outline are extracted to obtain the checked data, and the data checking module is connected to the noise removal module; The knowledge graph construction module is used to construct the knowledge graph, batch process the verified data using Python tools, and store them in the Neo4j database to construct a knowledge point graph. The knowledge graph construction module is connected to the data verification module.

Images