Multi-course integrated knowledge graph construction and real question analysis method and system

Abstract

The present invention provides a method and system for constructing a multi-course integrated knowledge graph and analyzing real questions, and the method includes: performing a knowledge point alignment operation, integrating cross-course knowledge points, using a cosine similarity calculation method to quantitatively analyze knowledge points, identifying and aligning synonymous knowledge points in different courses; performing a knowledge point disambiguation operation, processing situations where knowledge points have the same name but different meanings, and obtaining integrated knowledge points and integrated knowledge graphs; performing a knowledge point normalization operation on the integrated knowledge points, and performing a standardized processing on the integrated knowledge graph to obtain standardized knowledge points and standardized knowledge graphs; converting and extracting the standardized knowledge points to obtain key features, and obtaining a semantic feature sequence vector according to the key feature processing; matching to obtain matching results and applicable knowledge points; verifying and evaluating the matching results, and mounting the questions to the corresponding knowledge points. The present invention solves the technical problems of one-sided knowledge point extraction results and low result accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge point extraction, and specifically to a method and system for constructing a multi-course integrated knowledge graph and analyzing real questions. Background Art

[0002] With the rapid development of information technology, online learning models supported by the Internet and information technology have become one of the main ways of teaching and learning. The development of online learning can overcome regional differences, solve the problem of uneven distribution of educational resources, and improve the quality of teaching and learning. The organization of knowledge can be divided into four levels, from top to bottom: disciplines, knowledge fields, knowledge units, and knowledge points. In the field of education, the mastery of the knowledge system can be decomposed into the mastery of knowledge points and the logical relationship between knowledge points, and learning in units of knowledge points conforms to the laws of human cognition. Learning resources are constantly being created and changed, while the knowledge points corresponding to the disciplines in which the learning resources are located remain basically unchanged. Therefore, the design of learning resources should be carried out around the overall knowledge system of the discipline to cultivate students' systematic knowledge and comprehensive abilities. However, in the existing massive learning resources, the annotation of knowledge points is generally insufficient, which makes it difficult for students to accurately identify and understand the specific knowledge points involved in the questions when facing real questions, thus affecting the learning effect and learning efficiency. Especially in comprehensive examinations such as job examinations, civil service examinations, and postgraduate entrance examinations, candidates need to master the knowledge points of multiple courses, which are usually highly systematic and cross-cutting. Exam questions involve the comprehensive application of multiple knowledge points, which places higher demands on students' knowledge system. If the knowledge points in the learning resources are not clearly marked, it will be difficult for students to prepare for the exam effectively, which will affect the overall learning effect.

[0003] At present, there are relatively few studies on the annotation of test knowledge points in learning resources. Chen Yewang et al. analyzed the semantic environment of domain ontology knowledge and the resource document structure, used the semantic environment information expressed by the domain ontology, and semantically annotated various document resources in the agricultural field, and proposed an improved method for document semantic annotation based on ontology. Wei Wei et al. introduced the concept of text semantics, considered the degree of semantic matching between the test questions to be annotated and the learning rules, and proposed an automatic annotation method for test knowledge points based on semantic association rules. The limitation of this method is that the semantic understanding of the test questions is not deep enough, so the overall accuracy is still low, and there is still a certain gap with the expected annotation accuracy.

[0004] The existing invention patent application document "A method and device for aligning graph knowledge points with videos" with publication number CN117033665A, the existing method includes: obtaining teaching videos and identifying the text content with timestamp marks in the teaching videos; calculating the first vector of each sentence in the text content; obtaining the knowledge points of the corresponding courses of the teaching videos; constructing an initial knowledge graph based on the knowledge points; calculating the second vector of each knowledge point; determining the corresponding teaching video information for each knowledge point according to the matching value of the first vector and the second vector; associating the corresponding teaching video information for each knowledge point on the initial knowledge graph to generate a final knowledge graph. The aforementioned existing scheme has the problem of many-to-many matching. In addition, the existing technology does not adopt an adaptive threshold mechanism, and cannot automatically adjust the matching standard according to the semantic density of the knowledge point, and the alignment accuracy is low. The model used in the existing technology is difficult to achieve accurate word sense disambiguation. The aforementioned existing scheme mainly focuses on the alignment of knowledge points and videos. In the process of acquiring knowledge points, the acquired knowledge points are not processed, the problem of effectively integrating the knowledge points of multiple courses cannot be effectively integrated, and the uniqueness of the knowledge points cannot be guaranteed. Definition errors are prone to occur, and it is difficult to ensure the consistency of the knowledge points and the accuracy of matching.

[0005] The existing invention patent application document "Test question generation system method device and storage medium based on knowledge graph" with publication number CN115048506A, the test question generation server in the existing method is used to build a knowledge graph; obtain test questions in the field corresponding to the knowledge graph in network resources, and generate a test question bank; extract initial test knowledge points from the test questions in the test question bank, and obtain the test knowledge point bank; obtain each knowledge point in the current test question outline, and determine the subdivided knowledge points to be tested corresponding to each knowledge point from the test knowledge point bank; based on the subdivided knowledge points to be tested, determine the test content corresponding to the subdivided knowledge points to be tested from the test question bank; send the test content corresponding to the subdivided knowledge points to be tested to the user terminal for the question setter to select; receive the question setter's operation on the selection of the test content, and generate the corresponding test question content. The aforementioned existing solutions mainly focus on the construction of knowledge graphs in a single field, lack the function of cross-course integration, which may lead to confusion of knowledge points. At the same time, the test question generation system does not integrate information across courses, resulting in inconsistent definitions of knowledge points. The parsing process of the aforementioned test question generation system is relatively simple, making it difficult to deeply explore the potential characteristics of the questions. The lack of a systematic verification step may result in a mismatch between the generated test questions and the knowledge points.

[0006] The existing public document "Design and Implementation of Intelligent Learning Guidance System Based on Knowledge Graph" states that:

[0007] The main tasks include designing and implementing a test question recommendation algorithm based on a knowledge graph, combining the study of the correlation between knowledge points, and designing how to recommend suitable test questions to users. By analyzing the user's test data on the platform, we can understand the user's mastery of knowledge points, analyze the prerequisite knowledge points of the user's weak knowledge points based on the correlation between knowledge points, and recommend corresponding test question data. The intelligent guidance system disclosed in the aforementioned prior art mainly relies on named entity recognition technology, and only focuses on the identification of single-discipline related characters, concepts, and experiments, and lacks cross-domain integration capabilities. The recommendation mechanism of the existing intelligent guidance system is not comprehensive and systematic, and it is difficult to deeply analyze the semantic information of the questions. There is a lack of verification steps, and it is easy for the recommended test questions to not match the knowledge points.

[0008] In summary, the existing technology has technical problems such as one-sided knowledge point extraction results and low accuracy of results. Summary of the invention

[0009] The technical problem to be solved by the present invention is: how to solve the technical problems in the prior art that the knowledge point extraction results are one-sided and the result accuracy is not high.

[0010] The present invention adopts the following technical solutions to solve the above technical problems: the method for constructing a multi-course fusion knowledge graph and analyzing real questions includes:

[0011] S1. Through the knowledge point alignment algorithm, the knowledge points across courses are integrated, and the cosine similarity calculation method is used to quantitatively analyze the knowledge points to obtain the cosine similarity calculation results, and to identify and align synonymous knowledge points in different courses;

[0012] S2. Perform knowledge point disambiguation operations, analyze the semantic differences of knowledge points in different contexts, use natural language processing technology and preset machine learning models to perform word sense disambiguation, handle situations where knowledge points have the same name but different meanings, and obtain integrated knowledge points and integrated knowledge graphs;

[0013] S3, perform knowledge point normalization operation on the integrated knowledge points, perform standardization processing on the integrated knowledge graph, and obtain standardized knowledge points and standardized knowledge graph;

[0014] S4, converting and extracting the normalized knowledge points to obtain key features, and processing the key features to obtain a semantic feature sequence vector;

[0015] S5. Match the standardized knowledge points and questions to obtain matching results and applicable knowledge points;

[0016] S6. Verify and evaluate the matching results, and mount the questions to the corresponding knowledge points according to the matching results, the corresponding relationship between the applicable knowledge points and the questions.

[0017] The present invention integrates the knowledge graphs of multiple courses, processes the synonymous relationship between knowledge points, and ensures that synonymous knowledge points in different courses can be aligned. Through the knowledge point disambiguation operation, the knowledge points with the same name but different meanings are processed to ensure the uniqueness and accuracy of the knowledge points in all courses in the knowledge graph. The integrated knowledge points are semantically checked to correct possible semantic errors and ensure that the definition, representation and relationship of each knowledge point in the knowledge graph meet the predetermined standards and specifications.

[0018] The question information obtained by the present invention includes keywords, features, question types, difficulty, grammar and semantic information, which will be dynamically integrated and matched with the processed knowledge points. Verification and optimization are performed based on the matching results through a variety of evaluation indicators to ensure system reliability and accuracy. The verified matching results are mounted on the standardized knowledge graph for further teaching and learning analysis. It is ensured that semantic information and structured information are taken into account during the extraction and matching of knowledge points, so that the results are more comprehensive and accurate.

[0019] In a more specific technical solution, in S1, a cosine similarity calculation method is used to quantitatively analyze the multi-dimensional features of knowledge points to obtain a cosine similarity calculation result, and the similarity between knowledge points is evaluated according to the cosine similarity calculation result. Based on the cosine similarity calculation result, a synonymous relationship detection technology is used to identify and align synonymous knowledge points, wherein in the knowledge point list of each textbook, the cosine similarity calculation result is calculated for each pair of knowledge points to generate a similarity matrix;

[0020] Set a similarity threshold, and determine whether two knowledge points are aligned based on the similarity threshold and cosine similarity calculation results;

[0021] According to the similarity matrix and similarity threshold, the knowledge points of different courses are matched to generate a list of aligned knowledge points; among them, the course knowledge graph is converted into attribute triples , relation triples ,in, a Represents attributes, r To represent relations, we unify the entity embedding space through predicate and attribute embedding learning, capture the similarity between entities in the predicate knowledge graph and the attribute knowledge graph, and aggregate at least two entity types using an attention-based function based on the predicate similarity in the entity similarity.

[0022] The present invention improves the accuracy and reliability of knowledge points through knowledge point alignment operation, and can also enhance the structuring and standardization of the knowledge base. In course construction, the aligned knowledge points help to clarify the teaching content, avoid duplication and confusion, thereby improving the course quality and teaching effect, and ensuring that students can systematically master each knowledge point.

[0023] In a more specific technical solution, the following logic is used to calculate the i Attention weights for entity types :

[0024]

[0025] in, Indicates i Embedding of entity types The trainable weight matrix, x Indicates the knowledge point definition embedding vector, Represents the transpose of the orthogonal semantic analysis matrix corresponding to the embedding vector of the knowledge point definition, x Indicates the knowledge point definition embedding vector, softmax is the activation function;

[0026] The predicate embedding is calculated by minimizing the following objective function:

[0027]

[0028]

[0029] in, PE represents predicate embedding, represents the predicate embedding objective function, Represents a predicate triple element, represents a predicate triple, Representation predicate P The corresponding randomly selected predicate triplet element, Representation predicate P The corresponding randomly selected predicate triples, is the margin hyperparameter, max means maximization, h is the entity parameter, p Represents a predicate, Indicates the entity child node corresponding to the current entity parameter and predicate. Represents the entity child node corresponding to the current triple element and predicate, represents a predicate triple element function, A function representing an arbitrarily chosen element of a predicate triple;

[0030] Attribute triples , relation triples ,Using the translation-based method TransE, the structural information of the entity is preserved;

[0031] The alignment state of an entity is calculated using the following equation:

[0032]

[0033] in, represents the alignment result, Represents the entity to be aligned, represents an entity in a given graph, represents the cosine similarity between the entity to be aligned and the entity in the given graph, argmax Find the function for the maximum subscript, Represents a given graph.

[0034] In a more specific technical solution, based on the translation method TransE, the following logic is used to minimize the margin-based objective function: :

[0035]

[0036]

[0037]

[0038]

[0039] in, represents the L2 norm of the vector, Represents a graph, represents a random selection of entities, represents a randomly selected predicate triplet element, Represents entities and collections of elements;

[0040] The structural embedding is calculated by minimizing the following objective function:

[0041]

[0042]

[0043] in, represents the weight vector of the gating mechanism, Represents the relationship count, represents a triple;

[0044] Minimize the following objective function , learn attribute embeddings:

[0045]

[0046]

[0047]

[0048] in, Represents the elements of an attribute triple, represents an element of a randomly selected attribute triplet, represents a randomly selected attribute triple, represents the attribute triple element function, represents an arbitrarily selected attribute triplet element function, Represents the feature vector function of the knowledge point in the standard knowledge base corresponding to the current attribute, Represents the orthogonal semantic analysis matrix The corresponding graph, Represents the orthogonal semantic analysis matrix The corresponding attribute set, Indicates randomly selecting a node.

[0049] In a more specific technical solution, S2 includes:

[0050] S21, extracting context features of knowledge points; comparing and classifying the context features to determine the specific meaning of each knowledge point;

[0051] S22, compare the knowledge points by combining the preset dictionary and knowledge base;

[0052] S23. Constructing a hierarchical label tree , used to represent the relationship between knowledge points. According to the relationship between knowledge points, a multi-label ranking model is trained to retrieve related entities as candidate entities;

[0053] S24, using the scorer, for each candidate entity Assign ranking points , k Represents the entity serial number, and selects the candidate entity with the highest score;

[0054] S25. Determine the relationship between two knowledge points by searching for attribute correlations, merge the knowledge points, and determine whether they have the same name or different meanings, so as to obtain integrated knowledge points and an integrated knowledge graph.

[0055] In a more specific technical solution, in S22, the preset entity L As a space of labeled entities, the properties of knowledge points are represented according to a given document:

[0056]

[0057] in, Indicates the knowledge point definition corresponding to the embedding vector i tags, |W| is the total number of tags, X Represents a collection of attributes;

[0058] Use the following logic to express the mention information of knowledge points:

[0059] .

[0060] in, m Indicates a sequence of mentioned information. |m| Indicates the number of times the information is mentioned. Indicates the tag corresponding to the knowledge point definition embedding vector, Indicates the sequence of mentioned information The corresponding information s.

[0061] The present invention ensures that the meaning of each knowledge point is clear and consistent through comparison operations, thereby improving the overall reliability and accuracy of the system and providing a basis for knowledge point alignment and question matching.

[0062] In a more specific technical solution, S25 includes:

[0063] S251, using unigram vectors and bigram TF-IDF vectors as bag-of-words instance features to represent the mention information and attribute information of knowledge points;

[0064] S252: performing a connection operation on the mentioned information and the attribute information, wherein the mentioned information and the attribute information are connected using the following logic:

[0065]

[0066] Among them, the mention function corresponding to the mention information is used And knowledge point mention function The text features of the knowledge point attributes are converted into bigram TF-IDF vectors, where c represents the clustering operation, d is the feature dimension, R represents the set of real numbers, express d dimensional real number set;

[0067] S253, establishing a clustering-based tag tree, performing hierarchical semantic indexing on tag entities, and using the clustering-based tag tree to calculate the relevance scores of all tag entities to perform relevance ranking;

[0068] Aggregate the reference entity of the label entity and the mention features matched by the label entity to calculate the label entity Label features :

[0069]

[0070] in, Indicates the node corresponding to the label entity , Represents the L2 norm of the node corresponding to the label entity,m Indicates a sequence of mentioned information. Represents the set of mention information corresponding to the tag entity, The tokens corresponding to the sequence of mentioned information;

[0071] S254. Perform hierarchical clustering to create a label tree for the entity H ;

[0072] S255, in the tag tree H In the node Representative entities Ranking is performed, where for each entity child node , Represents a node set, based on the node and entity child nodes u The corresponding model weight , parameterized linear sequencer Ranking:

[0073]

[0074] in, x Indicates the knowledge point definition embedding vector, Represents the model weight The matrix transpose of ;

[0075] S256, learning model weights through linear SVM , parameterized linear sequencer , determine the loss function:

[0076]

[0077]

[0078] in, Representing Entities y Whether it is an entity child node u Coverage, the entity subnodes of the label entity set L satisfy: , Represents knowledge point definition embedding vector x ,entity y , entity child node u ,node Corresponding entity , the corresponding loss function, D Represents a set of knowledge point definition embedding vectors and entity corresponding points;

[0079] S257, using the sigmoid function, get the node Entity child nodes u The probability score :

[0080]

[0081] in, Indicates i Knowledge points define embedding vectors x The corresponding parameterized linear sorter.

[0082] In a more specific technical solution, in S4, a question parsing operation and a text embedding operation are performed to convert the normalized knowledge points into text information word vectors, and key features in the question are extracted based on the text information word vectors;

[0083] Use the preset large model to process key features and obtain semantic feature sequence vectors;

[0084] The attention mechanism is used to obtain the attention weight, which represents the relevance of knowledge and text. Based on the attention mechanism, the following logic is used to assign the attention weight to the existing textbook knowledge and obtain text-aware knowledge:

[0085]

[0086] in, is the feature vector of the question, is the feature vector of the selected existing textbook knowledge, is the text-aware knowledge feature vector, softmax is the activation function, The eigenvector of the problem corresponds to the transpose of the matrix, The transpose of the matrix corresponding to the eigenvector representing the selected existing textbook knowledge.

[0087] The present invention extracts key information and features from the questions by parsing them, so that the big model can better understand and process the questions, thereby improving the accuracy of matching the real questions with the knowledge points. By conducting an in-depth analysis of the grammar and semantics of the questions, the type, difficulty and keywords of the questions are identified, and the questions are structured using natural language processing technology. In addition, the question parsing operation also includes the identification and processing of special symbols, formulas and charts in the questions, so that the big model can fully understand the content of the questions. Through the question parsing operation, it is ensured that all details of the questions can be accurately captured and expressed, thereby providing more contextual information in the knowledge point matching process, improving the accuracy and precision of the matching results, and optimizing the overall performance and reliability of the system.

[0088] In a more specific technical solution, in S5, a feature vector of each question is extracted;

[0089] Convert the normalized knowledge points into vector representations to match the feature vectors of the questions: Use the following logic to process the single vector of knowledge , the feature vector of the question Each representation in the knowledge vector Connection, processing to obtain the weight vector of the gating mechanism :

[0090]

[0091] in, is a learnable parameter, The eigenvector of the problem corresponds to the transpose of the matrix, Represents a join operation.

[0092] The present invention gives priority to the keywords and features of the questions, and refers to the question type and difficulty level to ensure the comprehensiveness and accuracy of the matching process. In addition, the pre-trained model and semantic similarity algorithm are also used in the matching operation to improve the accuracy and efficiency of the matching. Through the matching operation, the knowledge points involved in the question can be automatically identified, which can assist teachers in accurate teaching and assessment analysis, and provide students with personalized learning suggestions to improve learning effects and efficiency.

[0093] In a more specific technical solution, the multi-course fusion knowledge graph construction and real question analysis system includes:

[0094] The knowledge point alignment module is used to integrate knowledge points across courses through the knowledge point alignment algorithm, and to quantitatively analyze the knowledge points using the cosine similarity calculation method to obtain the cosine similarity calculation results, and to identify and align synonymous knowledge points in different courses;

[0095] The knowledge point disambiguation module is used to perform knowledge point disambiguation operations, analyze the semantic differences of knowledge points in different contexts, use natural language processing technology and preset machine learning models to perform word sense disambiguation, handle situations where knowledge points have the same name but different meanings, and obtain integrated knowledge points and integrated knowledge graphs. The knowledge point disambiguation module is connected to the knowledge point alignment module;

[0096] The knowledge point normalization module is used to perform knowledge point normalization operations on the integrated knowledge points and standardize the integrated knowledge graph to obtain standardized knowledge points and standardized knowledge graphs. The knowledge point normalization module is connected to the knowledge point disambiguation module.

[0097] The question parsing module is used to convert and extract the normalized knowledge points to obtain key features, and obtain the semantic feature sequence vector according to the key features. The question parsing module is connected with the knowledge point normalization module;

[0098] The knowledge point and question matching module is used to match the standardized knowledge points and questions to obtain matching results and applicable knowledge points. The knowledge point and question matching module is connected with the question parsing module;

[0099] The result verification and evaluation module is used to verify and evaluate the matching results. According to the corresponding relationship between the matching results, applicable knowledge points and questions, the questions are mounted to the corresponding knowledge points. The result verification and evaluation module is connected to the knowledge point and question matching module.

[0100] Compared with the prior art, the present invention has the following advantages:

[0101] The present invention integrates the knowledge graphs of multiple courses, processes the synonymous relationship between knowledge points, and ensures that synonymous knowledge points in different courses can be aligned. Through the knowledge point disambiguation operation, the knowledge points with the same name but different meanings are processed to ensure the uniqueness and accuracy of the knowledge points in all courses in the knowledge graph. The integrated knowledge points are semantically checked to correct possible semantic errors and ensure that the definition, representation and relationship of each knowledge point in the knowledge graph meet the predetermined standards and specifications.

[0102] The question information obtained by the present invention includes keywords, features, question types, difficulty, grammar and semantic information, which will be dynamically integrated and matched with the processed knowledge points. Verification and optimization are performed based on the matching results through a variety of evaluation indicators to ensure system reliability and accuracy. The verified matching results are mounted on the standardized knowledge graph for further teaching and learning analysis. It is ensured that semantic information and structured information are taken into account during the extraction and matching of knowledge points, so that the results are more comprehensive and accurate.

[0103] The present invention improves the accuracy and reliability of knowledge points through knowledge point alignment operation, and can also enhance the structuring and standardization of the knowledge base. In course construction, the aligned knowledge points help to clarify the teaching content, avoid duplication and confusion, thereby improving the course quality and teaching effect, and ensuring that students can systematically master each knowledge point.

[0104] The present invention ensures that the meaning of each knowledge point is clear and consistent through comparison operations, thereby improving the overall reliability and accuracy of the system and providing a basis for knowledge point alignment and question matching.

[0105] The present invention extracts key information and features from the questions by parsing them, so that the big model can better understand and process the questions, thereby improving the accuracy of matching the real questions with the knowledge points. By conducting an in-depth analysis of the grammar and semantics of the questions, the type, difficulty and keywords of the questions are identified, and the questions are structured using natural language processing technology. In addition, the question parsing operation also includes the identification and processing of special symbols, formulas and charts in the questions, so that the big model can fully understand the content of the questions. Through the question parsing operation, it is ensured that all details of the questions can be accurately captured and expressed, thereby providing more contextual information in the knowledge point matching process, improving the accuracy and precision of the matching results, and optimizing the overall performance and reliability of the system.

[0106] The present invention gives priority to the keywords and features of the questions, and refers to the question type and difficulty level to ensure the comprehensiveness and accuracy of the matching process. In addition, the pre-trained model and semantic similarity algorithm are also used in the matching operation to improve the accuracy and efficiency of the matching. Through the matching operation, the knowledge points involved in the question can be automatically identified, which can assist teachers in accurate teaching and assessment analysis, and provide students with personalized learning suggestions to improve learning effects and efficiency.

[0107] The present invention solves the technical problems in the prior art that the knowledge point extraction results are one-sided and the result accuracy is not high. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Figure 1 A schematic diagram of the basic steps of the method for constructing a multi-course fusion knowledge graph and analyzing real questions in Example 1 of the present invention;

[0109] Figure 2 This is a schematic diagram of specific steps for performing knowledge point disambiguation in Example 1 of the present invention;

[0110] Figure 3 This is a schematic diagram of specific steps for distinguishing entities with the same name but different meanings according to Example 1 of the present invention. DETAILED DESCRIPTION

[0111] 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.

[0112] Example 1

[0113] like Figure 1 As shown, the multi-course fusion knowledge graph construction and real question analysis method provided by the present invention includes the following basic steps:

[0114] S1. Through the knowledge point alignment algorithm, the knowledge points across courses are integrated, and the cosine similarity calculation method is used to quantitatively analyze the knowledge points to obtain the cosine similarity calculation results, and to identify and align synonymous knowledge points in different courses;

[0115] In this embodiment, the cosine similarity calculation method is used to quantitatively analyze the multidimensional features of the knowledge points to obtain the cosine similarity calculation results, and the similarity between the knowledge points is evaluated according to the cosine similarity calculation results. Based on the cosine similarity calculation results, the synonymous knowledge points are identified and aligned using the synonymy relationship detection technology. In step S1, in the knowledge point list of each textbook, the cosine similarity calculation results are calculated for each pair of knowledge points to generate a similarity matrix;

[0116] Set a similarity threshold, and determine whether two knowledge points are aligned based on the similarity threshold and cosine similarity calculation results;

[0117] According to the similarity matrix and similarity threshold, the knowledge points of different courses are matched to generate a list of aligned knowledge points; among them, the course knowledge graph is converted into attribute triples , relation triples ,in, a Represents attributes, r To represent relations, we unify the entity embedding space through predicate and attribute embedding learning, capture the similarity between entities in the predicate knowledge graph and the attribute knowledge graph, and aggregate at least two entity types using an attention-based function based on the predicate similarity in the entity similarity.

[0118] In this embodiment, through the knowledge point alignment operation, in the process of integrating multiple courses, the synonymous relationship between knowledge points is processed, and the synonymous knowledge points in different courses are aligned. If a pair of knowledge points can be aligned, a similarity relationship is added between the two knowledge points, and the pair of knowledge points is added to the knowledge graph.

[0119] In this embodiment, the cosine similarity is calculated for each pair of knowledge points in the knowledge point list of each textbook to generate a similarity matrix. The cosine similarity formula is:

[0120]

[0121] In this embodiment, a similarity threshold of 0.9 is set to determine whether two knowledge points can be aligned. According to the similarity matrix and the set similarity threshold, the knowledge points of multiple courses are matched to generate an aligned knowledge point list.

[0122] In this embodiment, the course knowledge graph is regarded as an attribute triple and relation triples , learn a unified entity embedding space by embedding predicates and attributes, capturing and relation triples The similarity between entities of the corresponding two knowledge graphs. For predicate similarity, an attention-based function is used to aggregate no less than 2 entity types. In this embodiment, the calculation of i Attention weights for entity types :

[0123]

[0124] In this embodiment, the predicate embedding is calculated by minimizing the following objective function:

[0125]

[0126]

[0127] In this embodiment, the attribute triple , relation triples , using the translation method TransE to preserve the structural information of the entity; using the following logic to minimize the margin-based objective function , learn structural embeddings:

[0128]

[0129]

[0130]

[0131]

[0132] In this example, the following objective function is minimized to calculate the structural embedding:

[0133]

[0134]

[0135] In this embodiment, the following objective function is minimized , learn attribute embeddings:

[0136]

[0137]

[0138]

[0139] In this embodiment, the alignment state of the entity is calculated by the following equation:

[0140]

[0141] In this embodiment, for a given graph, the similarity between the entity to be aligned and all entities in the given graph is calculated; Represents the expected pair of aligned entities. In this embodiment, the similarity threshold is used to filter out entity pairs that cannot be aligned due to dissimilarity.

[0142] S2. Perform knowledge point disambiguation operations, analyze the semantic differences of knowledge points in different contexts, use natural language processing technology and preset machine learning models to perform word sense disambiguation, handle situations where knowledge points have the same name but different meanings, and obtain integrated knowledge points and integrated knowledge graphs;

[0143] In the knowledge point disambiguation operation of this embodiment, the problem of knowledge points with the same name but different meanings is solved by analyzing the semantic differences of knowledge points in different contexts and performing word sense disambiguation using natural language processing technology and machine learning models.

[0144] like Figure 2 As shown, in this embodiment, step S2 of performing knowledge point disambiguation also includes the following specific steps:

[0145] S21, extracting context features of knowledge points; comparing and classifying the context features to determine the specific meaning of each knowledge point;

[0146] S22. Combine the preset dictionary and knowledge base to compare the knowledge points to further improve the accuracy and efficiency of disambiguation;

[0147] In this embodiment, the preset entity L As a space of labeled entities, the properties of knowledge points are represented according to a given document:

[0148]

[0149] Use the following logic to express the mention information of knowledge points:

[0150] .

[0151] S23. Constructing a hierarchical label tree , used to represent the relationship between knowledge points. According to the relationship between knowledge points, a multi-label ranking model is trained to retrieve related entities as candidate entities;

[0152] S24, using the scorer, for each candidate entity Assign ranking points , k Represents the entity serial number, and selects the candidate entity with the highest score;

[0153] S25, determining the relationship between two knowledge points by searching for attribute relevance, merging the knowledge points, determining whether the knowledge points have the same name or different meanings, and obtaining integrated knowledge points and an integrated knowledge graph;

[0154] In this embodiment, the knowledge points of the text type test questions in the learning resources are automatically annotated. Effective information is added to the relevant data information through annotation. In the process of establishing the annotation, the association between the annotated data and the annotated data is obtained. The annotation methods include: network annotation, public annotation and semantic annotation. The annotation tasks include: word part-of-speech tagging, semantic role tagging, document category tagging, image tagging, user characteristic tagging and social tagging.

[0155] like Figure 3 As shown, in this embodiment, if two knowledge points are similar, they are merged into one knowledge point, otherwise it is explained that the two knowledge points are synonyms, and a distinguishing label is added to one of the entities to distinguish entities with the same name but different meanings. Step S25 includes:

[0156] S251, using unigram vectors and bigram TF-IDF vectors as bag-of-words instance features to represent the mention information and attribute information of knowledge points;

[0157] S252: performing a connection operation on the mentioned information and the attribute information, wherein the mentioned information and the attribute information are connected using the following logic:

[0158]

[0159] S253, establishing a clustering-based tag tree, performing hierarchical semantic indexing on tag entities, and using the clustering-based tag tree to calculate the relevance scores of all tag entities to perform relevance ranking;

[0160] In this embodiment, the reference entity of the tag entity and the mention feature matched by the tag entity are aggregated to calculate the tag entity Label features :

[0161]

[0162] S254. Perform hierarchical clustering to create a label tree for the entity H ;

[0163] In this embodiment, it is assumed that the tag tree The root node represents The balanced K-Means algorithm recursively transforms nodes into The representative entity L is divided into child nodes K clusters , until |L| meets the stopping criterion B. Depend on nodes, including the leaf nodes of the |L| label entities and other non-leaf nodes that implicitly collect entities with similar semantics.

[0164] S255, in the tag tree H In the node Representative entities Ranking is performed, where for each entity child node , Represents a node set, based on the node and entity child nodes u The corresponding model weight , parameterized linear sequencer Ranking:

[0165]

[0166] S256, learning model weights through linear SVM , parameterized linear sequencer , determine the loss function:

[0167]

[0168]

[0169] S257, using the sigmoid function, get the node Entity child nodes u The probability score :

[0170] .

[0171] S3, perform knowledge point normalization operation on the integrated knowledge points, perform standardization processing on the integrated knowledge graph, and obtain standardized knowledge points and standardized knowledge graph;

[0172] In this embodiment, a knowledge point normalization operation is performed, the text is preprocessed to perform semantic checks on the integrated knowledge points, the integrated knowledge graph is standardized, and the standardized knowledge points and the standardized knowledge graph are obtained. The BERT model is used to process the integrated knowledge point definition embedding vector and the standard knowledge base definition embedding vector. The cosine similarity calculation result is used to verify whether the text pair matches, identify and correct definition errors, and merge the standardized knowledge points.

[0173] In this embodiment, the integrated knowledge graph is further standardized to ensure the consistency and accuracy of each integrated knowledge point in the knowledge graph. In the knowledge point standardization operation, the definition, expression form and semantic relationship of the knowledge point are checked and corrected to eliminate possible polysemous words, synonyms and differences in different versions of teaching materials. In the knowledge point standardization operation, natural language processing technology is used to perform semantic analysis and correction on the description of the knowledge point to ensure the accuracy and consistency of the knowledge point in different contexts. Through these processing steps, the structuring degree of the knowledge graph is enhanced to ensure that all knowledge points are represented with a unified standard, thereby improving the overall quality and reliability of the knowledge graph.

[0174] In this embodiment, a cleaning operation is performed on the data in the integrated knowledge points, and the cleaning operation includes but is not limited to: removing duplicate items, standardizing terms, and unifying formats.

[0175] In this embodiment, knowledge points of different forms are unified into a standardized expression form through text preprocessing technology, including: denoising, lemmatization and stemming.

[0176] In this embodiment, the records of knowledge points are unified into JSON format, which contains clear fields. In this embodiment, a unique identifier is assigned to each knowledge point and set in the knowledge point ID field. All knowledge points follow this identifier system to avoid duplication and conflict, maintain consistent labels and classifications, and facilitate subsequent processing and analysis.

[0177] In this embodiment, the BERT (Bidirectional Encoder Representations from Transformers) model is used to obtain the embedding vectors of the definitions of knowledge points and the definitions in the standard knowledge base, and then the cosine similarity is used to verify whether the text pairs match, and the inconsistencies and errors in the definitions are identified and corrected. The same knowledge points in different expressions are identified and merged. For example: the basic components of a computer system and the components of a computer system are merged into a unified basic component entry of a computer system. In this embodiment, the cosine similarity formula is:

[0178]

[0179] In the formula, cosine represents the cosine function, express, u is the feature vector of the knowledge point in the graph, v is the feature vector of the knowledge point in the standard knowledge base, n is the vector dimension, Represents an entity child node, Represents a node, represents the L1 norm of the feature vector of the knowledge point in the graph, The L1 norm of the feature vector representing the knowledge point in the standard knowledge base, Indicates i feature vectors of the knowledge points in the graph, Indicates i The feature vectors of the knowledge points in the standard knowledge base.

[0180] In this embodiment, a similarity threshold of 0.9 is set to determine whether two knowledge points can be aligned. According to the similarity matrix and the set threshold, the knowledge points of multiple courses are matched to generate an aligned knowledge point list. As shown in Table 1:

[0181] Table 1 Text similarity score table

[0182]

[0183] S4, converting and extracting the normalized knowledge points to obtain key features, and processing the key features to obtain a semantic feature sequence vector;

[0184] In this embodiment, the questions are analyzed through real question analysis operations to extract key information and features in the questions so that the big model can better understand and process the questions, thereby improving the accuracy of matching the questions with the knowledge points.

[0185] In this embodiment, a model based on the Transformer network structure is used to perform deep feature extraction on the question text, and each question is converted into a word vector with text information, thereby achieving efficient embedding of the text. In view of the scarcity of corpora in the field of education, the present invention refers to the fine-tuned BERT model in CourseKG to perform model optimization operations to improve the accuracy in the field of education. In this process, only words related to education are masked to ensure that the word embedding vector can accurately reflect the semantic information of the outline text. The word embedding vector is input into the BERT model for feature extraction to obtain a semantic feature sequence vector.

[0186] In this embodiment, the attention mechanism is used to obtain the attention weight, which represents the relevance between knowledge and text. The higher the attention weight, the more important the knowledge is. The attention mechanism is used to allocate the attention weight to the existing textbook knowledge to obtain text-aware knowledge, which is expressed as:

[0187]

[0188] For example, using the attention mechanism, the attention weights of a question and the corresponding external known knowledge are 1.32, 0.69, 2.46, 1.74, and 0.33, respectively. The third textbook knowledge point has the highest relevance to the text, and the fifth has the lowest, so more attention is paid to the third textbook knowledge point.

[0189] S5. Match the standardized knowledge points and questions to obtain matching results and applicable knowledge points;

[0190] In this embodiment, knowledge points and real questions are matched, and the parsed question information includes: keywords, features, question types, difficulty, grammar and semantic information. The processed knowledge points and real questions are input into the big model to match the knowledge points;

[0191] In this embodiment, the parsed question information, including but not limited to: keywords, features, question types, difficulty, grammar and semantic information, is input into the big model together with the aforementioned knowledge points for matching to identify applicable knowledge points. In the process of matching knowledge points with real questions, deep learning and natural language processing technology are used to combine the contextual information of knowledge points and questions to calculate the similarity between applicable knowledge points and real questions.

[0192] In this embodiment, the feature vector of each question is extracted; the normalized knowledge points are converted into vector representations to match the feature vectors of the question: the following logic is used to process the single knowledge vector , the feature vector of the question Each representation in the knowledge vector Connection, processing to obtain the weight vector of the gating mechanism :

[0193]

[0194] Using weight vector Reconstructing knowledge representation and text representation.

[0195] In this embodiment, a prompt is constructed for the GLM-4 large model, the format is: title + knowledge point, and the output is: {title:..., knowledge point: {1}{2}...}. The prompt is used to guide the model to learn how to match the title with the corresponding knowledge point. The constructed prompt is input into the GLM-4 large model to allow the model to learn and match. Through multiple rounds of training, the GLM-4 large model gradually learns how to match according to the characteristics of the title and the characteristics of the knowledge point, and improves the accuracy of the matching operation. The GLM-4 large model outputs the matching results and obtains a list of knowledge points corresponding to each question, in the format: {title:..., knowledge point: {1}{2}...}.

[0196] S6. Verify and evaluate the matching results, and mount the questions to the corresponding knowledge points according to the matching results, the corresponding relationship between the applicable knowledge points and the questions.

[0197] In this embodiment, the matching results are verified and evaluated to ensure the reliability and accuracy of the system. During the result verification and evaluation operation, a variety of evaluation indicators, such as accuracy, recall rate and F1 score, are used to comprehensively measure the effect of matching knowledge points and questions. According to the correspondence between the processed knowledge points and questions, the corresponding questions are hung under the corresponding knowledge points to form a knowledge graph. The verification process includes manual verification and automated testing to ensure that each matching result meets expectations. In addition, during the result verification and evaluation operation, user feedback is collected, iterative optimization is performed, and the performance of the matching algorithm is continuously improved. Through result verification and evaluation, the matching ability of the system is monitored and improved to ensure that the association between knowledge points and questions is accurate, providing a solid foundation for subsequent teaching and learning analysis.

[0198] In this embodiment, according to the matching results, the corresponding knowledge points are found in the basic knowledge system framework generated in the knowledge graph, and the Exam relationship between the real questions and the knowledge points is established. This information is batch processed and stored in the knowledge graph through Python scripts. Through the above operations, the matching accuracy of the knowledge points and the questions is ensured, and the questions and knowledge points are structured and stored in the database.

[0199] In the process of integrating multiple course knowledge points, the present invention adopts a technical solution based on deep semantic analysis. The knowledge points of each course are text-vectorized by natural language processing (NLP) technology, and the context-related embedding representation is obtained by using the GLM-4 model. The semantic similarity of knowledge points in different courses is calculated by cosine similarity, and the alignment path is optimized in combination with the dynamic programming algorithm to solve the problem of many-to-many matching. In addition, an adaptive threshold mechanism is introduced to automatically adjust the matching criteria according to the semantic density of the knowledge points to ensure high-precision alignment. The present invention uses the GLM-4 model to embed the knowledge points in context, and captures the semantic differences of vocabulary in different contexts through a dynamic multi-head self-attention mechanism, thereby achieving accurate word sense disambiguation. At the same time, a technology based on Word Sense Disambiguation (WSD) is introduced, combined with prior knowledge in the knowledge graph, to further improve the accuracy of disambiguation. The present invention improves the uniqueness of knowledge points through disambiguation technology, corrects definition errors, ensures the consistency of knowledge points, and enhances the application ability of knowledge graphs in actual examinations and learning in question analysis. Through natural language processing technology, cosine similarity and attention mechanism, the relationship between knowledge points and questions is quantified, the matching accuracy is improved, and a systematic processing flow is established so that each step has clear goals and outputs, improving the overall efficiency and effectiveness of the system.

[0200] The present invention emphasizes the integration of knowledge points between different courses, focuses on the alignment and disambiguation of knowledge points, and ensures the unity of synonymous knowledge points. The present invention introduces the knowledge point disambiguation operation, analyzes the semantic differences of knowledge points in different contexts, and uses natural language processing technology to solve the problem of different meanings of the same name. The present invention normalizes and standardizes the knowledge points, uses the GLM-4 model to generate embedding vectors, and uses the similarity verification mechanism to perform consistency verification on the knowledge point definition in combination with the Euclidean distance. The present invention performs deep feature extraction on the question text based on the Transformer network structure, and uses a multi-layer Transformer encoder to capture the key semantic information in the question. The present invention uses the cosine similarity algorithm to match the question feature vector with the pre-constructed knowledge point embedding vector. In order to enhance the robustness of the matching, the present invention introduces the contrastive learning technology, and ensures the precise alignment of the question features with the knowledge point definition by constructing a positive and negative sample pair optimization model. The present invention adds a verification and evaluation mechanism after matching the knowledge points and questions to ensure the reliability of the matching results.

[0201] The present invention can extract key features from questions and match them with standardized knowledge points through question parsing operations. It can be more comprehensive and systematic, and can deeply analyze the semantic information of questions. The present invention introduces a mechanism for verifying and evaluating matching results, and can obtain a reliable correspondence between recommended test questions and knowledge points, avoiding the situation where recommended test questions do not match knowledge points.

[0202] In summary, the present invention integrates the knowledge graphs of multiple courses, processes the synonymous relationship between knowledge points, and ensures that synonymous knowledge points in different courses can be aligned. Through the knowledge point disambiguation operation, the knowledge points with the same name but different meanings are processed to ensure the uniqueness and accuracy of the knowledge points in all courses in the knowledge graph. The integrated knowledge points are semantically checked to correct possible semantic errors and ensure that the definition, representation and relationship of each knowledge point in the knowledge graph meet the predetermined standards and specifications.

[0203] The question information obtained by the present invention includes keywords, features, question types, difficulty, grammar and semantic information, which will be dynamically integrated and matched with the processed knowledge points. Verification and optimization are performed based on the matching results through a variety of evaluation indicators to ensure system reliability and accuracy. The verified matching results are mounted on the standardized knowledge graph for further teaching and learning analysis. It is ensured that semantic information and structured information are taken into account during the extraction and matching of knowledge points, so that the results are more comprehensive and accurate.

[0204] The present invention improves the accuracy and reliability of knowledge points through knowledge point alignment operation, and can also enhance the structuring and standardization of the knowledge base. In course construction, the aligned knowledge points help to clarify the teaching content, avoid duplication and confusion, thereby improving the course quality and teaching effect, and ensuring that students can systematically master each knowledge point.

[0205] The present invention ensures that the meaning of each knowledge point is clear and consistent through comparison operations, thereby improving the overall reliability and accuracy of the system and providing a basis for knowledge point alignment and question matching.

[0206] The present invention extracts key information and features from the questions by parsing them, so that the big model can better understand and process the questions, thereby improving the accuracy of matching the real questions with the knowledge points. By conducting an in-depth analysis of the grammar and semantics of the questions, the type, difficulty and keywords of the questions are identified, and the questions are structured using natural language processing technology. In addition, the question parsing operation also includes the identification and processing of special symbols, formulas and charts in the questions, so that the big model can fully understand the content of the questions. Through the question parsing operation, it is ensured that all details of the questions can be accurately captured and expressed, thereby providing more contextual information in the knowledge point matching process, improving the accuracy and precision of the matching results, and optimizing the overall performance and reliability of the system.

[0207] The present invention gives priority to the keywords and features of the questions, and refers to the question type and difficulty level to ensure the comprehensiveness and accuracy of the matching process. In addition, the pre-trained model and semantic similarity algorithm are also used in the matching operation to improve the accuracy and efficiency of the matching. Through the matching operation, the knowledge points involved in the question can be automatically identified, which can assist teachers in accurate teaching and assessment analysis, and provide students with personalized learning suggestions to improve learning effects and efficiency.

[0208] The present invention solves the technical problems in the prior art that the knowledge point extraction results are one-sided and the result accuracy is not high.

[0209] 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. The method of constructing a multi-course integrated knowledge graph and analyzing real questions is characterized by: The method comprises: S1. Through the knowledge point alignment algorithm, the knowledge points across courses are integrated, and the cosine similarity calculation method is used to quantitatively analyze the knowledge points to obtain the cosine similarity calculation results, and to identify and align synonymous knowledge points in different courses; S2. Perform knowledge point disambiguation operations, analyze the semantic differences of knowledge points in different contexts, use natural language processing technology and preset machine learning models to perform word sense disambiguation, handle situations where knowledge points have the same name but different meanings, and obtain integrated knowledge points and integrated knowledge graphs; S3, perform knowledge point normalization operation on the integrated knowledge points, perform standardization processing on the integrated knowledge graph, and obtain standardized knowledge points and standardized knowledge graph; S4, converting and extracting the normalized knowledge points to obtain key features, and processing the key features to obtain a semantic feature sequence vector; Among them, the question parsing operation and text embedding operation are performed to convert the standardized knowledge points into text information word vectors, and the key features in the question are extracted according to the text information word vectors; Use the preset large model to process key features and obtain semantic feature sequence vectors; S5. Match the standardized knowledge points and questions to obtain matching results and applicable knowledge points; S6. Verify and evaluate the matching results, and mount the questions to the corresponding knowledge points according to the matching results, the corresponding relationship between the applicable knowledge points and the questions.

2. The method for constructing a multi-course fusion knowledge graph and analyzing real questions according to claim 1 is characterized in that: In S1, the cosine similarity calculation method is used to perform the quantitative analysis on the multidimensional features of the knowledge points to obtain a cosine similarity calculation result, the similarity between the knowledge points is evaluated according to the cosine similarity calculation result, and based on the cosine similarity calculation result, the synonymous knowledge points are identified and aligned using a synonymy relationship detection technology, wherein in the knowledge point list of each textbook, the cosine similarity calculation result is calculated for each pair of the knowledge points to generate a similarity matrix; Setting a similarity threshold, and determining whether two knowledge points are aligned according to the similarity threshold and the cosine similarity calculation result; According to the similarity matrix and the similarity threshold, the knowledge points of different courses are matched to generate an aligned knowledge point list; wherein the course knowledge graph is converted into attribute triples , relation triples ,in, a Represents attributes, r To represent relations, the entity embedding space is unified through predicate and attribute embedding learning, and the similarity between entities in the predicate knowledge graph and the attribute knowledge graph is captured. According to the predicate similarity in the similarity between entities, an attention-based function is used to aggregate no less than two entity types.

3. The method for constructing a multi-course fusion knowledge graph and analyzing real questions according to claim 2 is characterized in that: Using the following logic, calculate i The attention weights of the entity types : In the formula, Indicates i embeddings of said entity type The trainable weight matrix, x Indicates the knowledge point definition embedding vector, represents the transpose of the orthogonal semantic analysis matrix corresponding to the embedding vector of the knowledge point definition, softmax is the activation function; The predicate embedding is calculated by minimizing the following objective function: In the formula, PE represents predicate embedding, represents the predicate embedding objective function, Represents a predicate triple element, represents a predicate triple, Representation predicate P The corresponding randomly selected predicate triplet element, Represents the predicate P The corresponding randomly selected predicate triples, is the margin hyperparameter, max means maximization, h is the entity parameter, p Represents a predicate, Indicates the current entity parameter and the entity child node corresponding to the predicate, represents the entity child node corresponding to the current triple element and the predicate, represents a predicate triple element function, A function representing an arbitrarily chosen element of a predicate triple; For the attribute triple , the relationship triple ,Using the translation-based method TransE, the structural information of the entity is preserved; The alignment state of the entity is calculated by the following equation: In the formula, represents the alignment result, Represents the entity to be aligned, represents an entity in a given graph, represents the cosine similarity between the entity to be aligned and the entity in the given graph, argmax Find the function for the maximum subscript, Represents a given graph.

4. The method for constructing a multi-course fusion knowledge graph and analyzing real questions according to claim 3 is characterized in that: In the translation method TransE, the following logic is used to minimize the margin-based objective function : In the formula, represents the L2 norm of the vector, Represents a graph, represents a random selection of entities, represents a randomly selected predicate triplet element, Represents entities and collections of elements; The structural embedding is calculated by minimizing the following objective function: In the formula, represents the weight vector of the gating mechanism, Represents the relationship count, represents a triple; Minimize the following objective function , learn attribute embeddings: In the formula, represents the element of the attribute triple, represents a randomly selected element of the attribute triple, represents the attribute triple selected randomly, represents the attribute triple element function, represents an arbitrarily selected attribute triplet element function, Represents the feature vector function of the knowledge point in the standard knowledge base corresponding to the current attribute, Represents the orthogonal semantic analysis matrix The corresponding graph, Represents the orthogonal semantic analysis matrix The corresponding attribute set, Indicates randomly selecting a node.

5. The method for constructing a multi-course fusion knowledge graph and analyzing real questions according to claim 1 is characterized in that: The S2 includes: S21, extracting context features of the knowledge points; comparing and classifying the context features to determine the specific meaning of each of the knowledge points; S22, comparing the knowledge points in combination with a preset dictionary and a knowledge base; S23. Constructing a hierarchical label tree , used to represent the relationship between knowledge points, according to which the relationship between knowledge points is trained to obtain a multi-label ranking model, and retrieve related entities as candidate entities; S24, using a scorer, for each candidate entity Assign ranking points , k represents the entity serial number, and selects the candidate entity with the highest score; S25. Determine the relationship between the two knowledge points by retrieving attribute correlations, so as to merge the knowledge points and determine whether they have the same name or different meanings, and obtain the integrated knowledge points and the integrated knowledge graph.

6. The method for constructing a multi-course integrated knowledge graph and analyzing real questions according to claim 5 is characterized in that: In S22, the preset entity L As a space of label entities, it represents the attributes of the knowledge point according to a given document: In the formula, Indicates the knowledge point definition corresponding to the embedding vector i tags, |W| is the total number of tags, X Represents a collection of attributes; Use the following logic to express the mention information of the knowledge point: In the formula, m represents the sequence of mentioned information, |m| Indicates the number of times the mentioned information is mentioned. Indicates the tag corresponding to the embedding vector defined by the knowledge point, Represents the sequence of mentioned information The corresponding information s.

7. The method for constructing a multi-course integrated knowledge graph and analyzing real questions according to claim 5 is characterized in that: The S25 includes: S251, using unigram vectors and bigram TF-IDF vectors as bag-of-words instance features to represent the mention information and attribute information of the knowledge point; S252: performing a connection operation on the mentioned information and the attribute information, wherein the mentioned information and the attribute information are connected using the following logic: In the formula, the mention function corresponding to the mention information is used And knowledge point mention function The text features of the knowledge point attributes are converted into the bigram TF-IDF vector, where: c represents the clustering operation, d is the feature dimension, R represents the set of real numbers, express d dimensional real number set; S253, establishing a clustering-based tag tree, performing hierarchical semantic indexing on tag entities, and using the clustering-based tag tree to calculate relevance scores of all the tag entities to perform relevance ranking; Aggregate the reference entity of the label entity and the mention features matched by the label entity to calculate the label entity Label features : In the formula, Indicates the node corresponding to the label entity , represents the L2 normal form of the node corresponding to the label entity, m Indicates a sequence of mentioned information. represents the set of mention information corresponding to the tag entity, Indicates the tag corresponding to the mentioned information sequence; S254. Perform hierarchical clustering to create a label tree for the entity H ; S255, in the label tree H In the above example, Representative entities Ranking is performed, where for each entity child node , Represents a node set, according to the node and the entity child node u The corresponding model weight , parameterized linear sequencer Ranking: In the formula, x Indicates the knowledge point definition embedding vector, Represents the model weight The matrix transpose of ; S256, learning the model weights by linear SVM , the parameterized linear sequencer , determine the loss function: In the formula, Representing Entities y Entity child node u In the covered state, the entity subnodes of the label entity set L satisfy: , Indicates that the knowledge point defines the embedding vector x , the entity y , the entity child node u , the node Corresponding entity , the corresponding loss function, D Represents a set of embedding vectors defining the knowledge points and corresponding points of the entities; S257, using the sigmoid function, obtain the node The entity child node u The probability score : In the formula, Indicates i The knowledge points define embedding vectors x The corresponding parameterized linear sequencer.

8. The method for constructing a multi-course integrated knowledge graph and analyzing real questions according to claim 1 is characterized in that: In S4, the attention mechanism is used to obtain the attention weight, which represents the relevance of knowledge and text. Based on the attention mechanism, the following logic is used to assign the attention weight to the existing teaching material knowledge, and the text perception knowledge is obtained: ； in, is the feature vector of the question, is the feature vector of the existing textbook knowledge selected, is the text-aware knowledge feature vector, softmax is the activation function, The eigenvector of the problem corresponds to the transpose of the matrix, The transpose of the matrix corresponding to the eigenvector of the selected existing teaching material knowledge is represented.

9. The method for constructing a multi-course integrated knowledge graph and analyzing real questions according to claim 1 is characterized in that: In S5, extracting a feature vector of each question; Convert the normalized knowledge points into vector representations to match the feature vectors of the question: Use the following logic to process the single vector of knowledge points , the feature vector of the question Each of the representations is a single vector with the knowledge Connection, processing to obtain the weight vector of the gating mechanism : In the formula, is a learnable parameter, The eigenvector of the problem corresponds to the transpose of the matrix, Represents a join operation.

10. A system for constructing a multi-course integrated knowledge graph and analyzing real questions, used to execute the method for constructing a multi-course integrated knowledge graph and analyzing real questions as described in any one of claims 1 to 9, characterized in that: The system comprises: The knowledge point alignment module is used to integrate the knowledge points across courses through the knowledge point alignment algorithm, and to quantitatively analyze the knowledge points using the cosine similarity calculation method to obtain the cosine similarity calculation results, thereby identifying and aligning the synonymous knowledge points in different courses. A knowledge point disambiguation module is used to perform knowledge point disambiguation operations, analyze the semantic differences of the knowledge points in different contexts, use natural language processing technology and a preset machine learning model to perform word sense disambiguation, and handle situations where knowledge points have the same name but different meanings, thereby obtaining integrated knowledge points and integrated knowledge graphs. The knowledge point disambiguation module is connected to the knowledge point alignment module; A knowledge point normalization module, used to perform knowledge point normalization operations on the integrated knowledge points, perform standardization processing on the integrated knowledge graph, and obtain standardized knowledge points and standardized knowledge graphs. The knowledge point normalization module is connected to the knowledge point disambiguation module; A question parsing module, used to convert and extract the normalized knowledge points to obtain key features, and obtain a semantic feature sequence vector according to the key features. The question parsing module is connected to the knowledge point normalization module; A knowledge point and question matching module is used to match the standardized knowledge points and questions to obtain matching results and applicable knowledge points. The knowledge point and question matching module is connected to the question parsing module; The result verification and evaluation module is used to verify and evaluate the matching result, and mount the question to the corresponding knowledge point according to the matching result, the corresponding relationship between the applicable knowledge point and the question. The result verification and evaluation module is connected to the knowledge point and question matching module.

Images