A teaching method and system based on cross-media dynamic knowledge graph

Through the teaching method based on the cross-media dynamic knowledge graph, students' learning characteristics are dynamically perceived, and personalized subject knowledge Q&A, resource recommendation and friend recommendation are provided, which solves the problem that existing teaching solutions cannot be formulated in a targeted manner, and improves teaching efficiency and Students’ learning experience.

CN111753098BActive Publication Date: 2024-11-22SHANGHAI HANSHI EDUCATION TECHNOLOGY CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202010581312.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-06-23
Publication Date
2024-11-22
Estimated Expiration
2040-06-23

AI Technical Summary

Technical Problem

The existing teaching plans cannot be formulated in a targeted manner based on students' individual learning characteristics, resulting in inefficient teaching.

Method used

Using a teaching method based on cross-media dynamic knowledge graph, by constructing a cross-media dynamic knowledge graph, storing entity-related information, conducting subject knowledge questions and answers, knowledge point recommendations and friend recommendations, dynamically perceive students' cognitive characteristics, and provide personalized learning resources and recommend.

Benefits of technology

It has achieved teaching according to the aptitude, improved teaching level and learning efficiency, enhanced students' learning interest and memory effects, and reduced their sense of loneliness and anxiety in learning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN111753098B_ABST
    Figure CN111753098B_ABST
Patent Text Reader

Abstract

A teaching method and system based on a cross-media dynamic knowledge graph, the method includes: constructing a cross-media dynamic knowledge graph; storing entity-related information in the knowledge graph, and returning the corresponding attribute value of the entity through subject knowledge questions and answers; obtaining feedback information from students and knowledge points, calculating the knowledge point recommendation degree and the learning resource recommendation value score, thereby constructing a knowledge point recommendation list and a learning resource recommendation list and returning them to the learner; constructing a course knowledge card; recommending friends, and recommending other users with similar learning progress and learning conditions to the current user. The teaching system includes a data layer, a data analysis layer, an application layer, and a user layer. The teaching method of the present invention can be implemented through a computer-readable storage medium. The present invention solves the problem in the prior art that the teaching plan cannot be specifically formulated according to the learning characteristics, and can teach students in accordance with their aptitude and effectively improve the teaching level.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and educational informatization, and specifically relates to a teaching method and system based on a cross-media dynamic knowledge graph. On the basis of the cross-media dynamic knowledge graph, artificial intelligence technology is used to realize functions such as subject knowledge question and answer, intelligent learning resource recommendation, course knowledge cards, and friend recommendation, thereby improving the teaching level. Background Art

[0002] With the rapid development of technologies such as artificial intelligence and deep learning, educational informatization has entered the 2.0 era with artificial intelligence technology as its main feature. The smart education model supported by information technology such as artificial intelligence has become a trend in the development of educational informatization and a hot research direction in academia. As an important part of artificial intelligence technology, knowledge graphs provide possibilities for smart education and promote its innovation and development. As a technical means of iteratively extracting structured knowledge from a large amount of complex data, knowledge graphs have shown broad application prospects in a large number of vertical fields.

[0003] Constructing a cross-media dynamic knowledge graph is conducive to reconstructing the connection between knowledge from a large number of disordered subject teaching resources (such as question banks, videos, etc.), integrating the fragmented knowledge of the subject, and integrating multimodal resources. It can effectively avoid the emergence of "information islands" and get rid of the shortcomings of the traditional knowledge system that is not updated in time. In the traditional teaching process, whether online or offline, teachers cannot formulate different teaching plans according to each person's learning situation, and all students are given the same education. Therefore, it has become a very important issue to teach students in accordance with their individual differences and improve learning efficiency. With the continuous development of artificial intelligence technology, it has provided a way to solve this problem. Compared with the traditional teaching system, the teaching system built based on the cross-media dynamic knowledge graph can dynamically perceive the various cognitive characteristics of the current learners, and provide learners with corresponding learning content and smart education services based on these characteristics. In short, the characteristics and relationships of cross-media knowledge graphs, teaching resources, and teaching behaviors determine that a teaching system can be built based on cross-media dynamic knowledge graphs, providing unprecedented development space and application prospects for the realization of precise teaching, smart learning, etc., and helping the informatization of education. Summary of the invention

[0004] The purpose of the present invention is to address the problem in the above-mentioned prior art that the teaching plan cannot be specifically formulated according to the learning characteristics, and to provide a teaching method and system based on a cross-media dynamic knowledge graph to teach students in accordance with their aptitude, thereby improving the teaching level.

[0005] In order to achieve the above object, the present invention has the following technical solutions:

[0006] A teaching method based on cross-media dynamic knowledge graph, including:

[0007] -Build a cross-media dynamic knowledge graph;

[0008] -Store entity-related information in the knowledge graph and return the corresponding attribute values ​​of the entity through subject knowledge questions and answers;

[0009] - Obtain feedback from students and knowledge points, calculate the knowledge point recommendation degree and learning resource recommendation value score, and then construct a knowledge point recommendation list and a learning resource recommendation list to return to the learner;

[0010] -Build course knowledge cards: Build different knowledge cards for different types of entities. Through user query of entities and matching in the knowledge graph, the corresponding entities and their attribute values ​​are returned in the form of knowledge cards;

[0011] -Friend recommendation: Recommend other users who have similar learning progress and status to the current user.

[0012] Preferably, constructing a cross-media dynamic knowledge graph includes:

[0013] Use Python tools to crawl relevant information of the required courses, including text, images, videos, exercises and PPT; clean the crawled data and manually annotate the text data using the BIO method; extract entities and attributes and relationships between entities in the processed data set, integrate the extracted entity sets, relationship sets, images and videos to form a cross-media knowledge graph, which is jointly edited by students and teachers after authorization by system personnel, and feedback and supplement the knowledge points found to be missing in the graph during the learning process to form a cross-media dynamic knowledge graph.

[0014] Furthermore, a dictionary of subject-specific terms is constructed, and it is added to Jieba's custom vocabulary during word segmentation, and the built-in traditional dictionary is integrated to segment the text; Word2vec is used as the word vector training model, and the organized subject field data set is used as the training corpus of the model to train it; the BILSTM+CRF neural network model is used for entity extraction, and the neural network model is trained through labeled text data, and the word vector representation is used as the input of the BILSTM network, and two LSTM networks are used to obtain the context information of the word vector to extract features for modeling, and then the calculation of the transfer matrix in CRF is integrated to finally predict the probability of the word as an entity to complete the entity extraction; a template-based approach is used for attribute extraction and relationship extraction. Attribute extraction is to describe the entity and express the entity completely through attributes; relationship extraction is to extract the internal logical relationship between entities, connect the scattered entities, and form a semantic network together; the extracted entity set, relationship set, image, and video are integrated and imported into Neo4j in batches to complete knowledge storage and form a cross-media knowledge graph.

[0015] Preferably, the process of subject knowledge question answering includes:

[0016] In the knowledge graph, entity-related information is stored in the form of "entity-attribute-attribute value" triples. When a question is asked, firstly, the entity in the question is extracted through named entity recognition, linked to the knowledge graph, and the corresponding entity is found. Then, the question information is combined with the pre-set rule template to match the entity attribute and return the corresponding attribute value. Further, the question is segmented and word vector trained. In the process of segmentation, the jieba tool is used and the pre-set subject-specific dictionary is added; the BILSTM+CRF model is used to extract entities from natural language questions; combined with the constructed knowledge graph, the extracted entities are matched with the entities in the graph. If there is a matching entity in the knowledge graph, the entity is selected. If there is no matching entity in the knowledge graph, the cosine similarity algorithm is used to calculate the semantic similarity and select the closest entity; the selected entity and attribute are matched with the triple in the knowledge graph; the corresponding attribute value is returned and provided to the user as the answer to the question.

[0017] Preferably, the process of constructing the knowledge point recommendation list and the learning resource recommendation list includes:

[0018] Firstly, by constructing the learner-resource type matrix, the learner-knowledge point rating matrix, and the knowledge point-resource type matrix, the explicit and implicit feedback information of students and knowledge points is obtained, including learners' learning interest preferences, learning status, and resource types contained in knowledge points; then, the method combining collaborative filtering and neural network is used to calculate the knowledge point recommendation degree and learning resource recommendation value rating, and a list is constructed according to the rating.

[0019] Furthermore, the learner set is defined as: U = {u1,u2,u3,…,u m}, the resource type set is: R = {r1, r2, r3, ..., r q}, thus we get the learner-resource type matrix, where the row vectors represent students, the column vectors represent resource types, and the matrix elements p ij represents the usage rate of resource type j by student i, i = {1, 2, 3, ..., m}, j = {1, 2, 3, ..., q}; to avoid the cold start problem, the number of times a learner uses each resource type is initialized to 1, and then the number of times is updated according to subsequent learning behavior;

[0020] The resource type usage calculation formula is as follows:

[0021]

[0022] Among them, v ij represents the number of times user i uses resource type j;

[0023] Define the learner set as: U = {u1,u2,u3,…,u m}, the knowledge point set is: K = {k1, k2, k3, ..., k n}, thus we get the learner-knowledge point scoring matrix, where the row vectors represent students, the column vectors represent knowledge points, and the matrix elements represent the student's score rate for the knowledge point; through the test test, we count the student's score for each knowledge point; among them, for subjective questions, we calculate a similarity value by the similarity between the standard answer and the student's answer, and then give the student's score according to the total score of the question; normalize the student's score for each knowledge point, map the obtained score to the interval [0, 1], and get the student's score rate for the knowledge point;

[0024] Construct a candidate recommended knowledge point list K', and screen the candidate recommended knowledge points based on the learner-knowledge point scoring matrix; set a threshold T. When the student's score rate for a knowledge point is less than the threshold, it indicates that the student's grasp of the knowledge point is relatively weak, and the knowledge point is used as a candidate recommended knowledge point; determine the recommended knowledge point list K'1 based on the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point in the knowledge graph; use BFS breadth-first search to calculate the shortest path between the most recently learned knowledge point and the candidate recommended knowledge point. The smaller the shortest path distance, the stronger the relevance of the knowledge point and the higher the degree of recommendation. Use d(o,t) to represent the shortest path distance between the most recently learned knowledge point o and the candidate recommended knowledge point t; determine the recommended knowledge point list K'2 based on the similarity between the student's most recently learned knowledge point and the candidate recommended knowledge point; according to the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point stored in the knowledge graph, the recommended knowledge point list K'1 is determined based on the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point. The attribute values ​​of each knowledge point are used to calculate the knowledge point similarity. The greater the knowledge point similarity, the more worthy of recommendation. The WV-CNN model is used to calculate the text similarity, and sim(o,t) is used to represent the text similarity between the most recently learned knowledge point o and the candidate recommended knowledge point t. The WV-CNN model consists of three parts: the first part is to represent the input text in terms of words, and obtain the word vector of the text by segmenting, encoding, and generating vectors for the input Chinese sentences, and use it as the input of the convolutional neural network (CNN); the second part is CNN, which has four layers: convolution layer, Dropout layer, pooling layer, and Flatten layer. The weight matrix in the network is continuously adjusted by the gradient descent method to obtain the optimized parameters and results; the third part is the output, that is, the result of the text semantic similarity calculation;

[0025] Linearly combine the generated recommended knowledge point list K'1 and the recommended knowledge point list K'2 to generate the final knowledge point recommendation list, and select the first F knowledge points based on the knowledge point recommendation degree to recommend to the learner;

[0026] Use l it represents the knowledge point recommendation degree of candidate recommended knowledge point t for user i, and the calculation formula is as follows:

[0027]

[0028] Among them, α and β are adjustment factors, and their values ​​must satisfy α·β=1;

[0029] Introduce the knowledge point-resource type matrix and define the knowledge point set as: K = {k1, k2, k3, …, k n}, the resource type set is: R = {r1, r2, r3, ..., r q}, the row vector represents the knowledge point, the column vector represents the resource type, and the matrix element a tjIndicates whether knowledge point t contains resource type j, t = {1, 2, 3, ..., n}, j = {1, 2, 3, ... q}; if it contains, it is recorded as 1, otherwise it is recorded as 0;

[0030] The learning resource recommendation list is obtained by integrating learners' resource type preferences, knowledge point recommendation degree, and resource types included in the knowledge points. The learning resource recommendation value calculation formula is as follows:

[0031] score itj =p ij ·l it ·a tj

[0032] Among them, score itj represents the recommendation value score of user i for type j resource in knowledge point t, p ij represents the usage rate of resource type j by user i, l it represents the recommendation degree of knowledge point t for user i, a tj Indicates whether knowledge point t contains resource type j; selects the top F learning resources based on the learning resource recommendation value score and recommends them to learners.

[0033] Preferably, the process of making friend recommendations includes:

[0034] Get the node position of the most recent learning knowledge point of the current student u and other students v in the knowledge graph. The current student position is represented as A and the other students' positions are represented as B. Combined with the feature of the student's most recent learning position based on the knowledge graph, calculate the position similarity between the current student u and other students v. The position similarity calculation formula is as follows:

[0035]

[0036] Among them, d(A, B) refers to the shortest path distance between node A and node B in the knowledge graph, which is calculated by BFS breadth-first search. That is, the farther the distance between A and B, the lower the similarity, and the lower the relevance of the recently learned knowledge; the closer the distance, the higher the similarity, and the higher the relevance of the recently learned knowledge;

[0037] Introduce the learner-knowledge point bipartite graph association model and define the learner set as: U = {u1,u2,u3,…,u m}, the knowledge point set is: K = {k1, k2, k3, ..., k n}, thus we get the learner-knowledge point matrix, where the row vectors represent students and the column vectors represent students’ learning behaviors about knowledge points, where the learned knowledge points are marked as 1 and the rest are 0;

[0038] Calculate the similarity between the current student u and other students v in their learning situation. The calculation formula is as follows:

[0039]

[0040] where u ki and v ki Respectively represent the current student u and other students v's understanding of knowledge point k i learning behavior, n is the total number of knowledge points. The greater the calculated similarity result, the more knowledge points the two students have learned and the more similar their learning situations are.

[0041] Calculate the comprehensive similarity between users; perform linear weighted fusion of location similarity and learning situation similarity, and use parameter α to balance the weight relationship between the two. The comprehensive similarity calculation formula is as follows:

[0042] sim(u,v i )=α·sim(place(u,v i ))+(1-α)·sim(LS(u,v i ))

[0043] Return the top N students in terms of comprehensive similarity and recommend them to the current student as friends.

[0044] The present invention also provides a teaching system based on a cross-media dynamic knowledge graph, including a data layer, a data analysis layer, an application layer and a user layer; the data layer covers the resource types required for constructing a cross-media dynamic knowledge graph, and data of different media types are multi-source and heterogeneous in form and semantically related to each other; the data analysis layer is used to process the crawled teaching resources, and the processing methods include entity extraction, relationship extraction, attribute extraction, and semantic similarity calculation; the application layer provides knowledge services for users, and the knowledge services include subject knowledge questions and answers, learning resource recommendations, course knowledge cards, and friend recommendations; the user layer divides users into three categories, namely students, teachers, and system personnel.

[0045] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of a teaching method based on a cross-media dynamic knowledge graph.

[0046] Compared with the prior art, the present invention has the following beneficial effects: constructing a cross-media dynamic knowledge graph, providing feedback and supplementing knowledge points that are found to be missing in the graph during the learning process, and continuously improving the knowledge graph by students and teachers. Subject knowledge questions and answers based on knowledge graphs can return the information required by users more conveniently and accurately. The present invention proposes an intelligent learning resource recommendation method based on knowledge graphs, which returns two parts of recommendations to learners: knowledge point recommendations and learning resource recommendations. The knowledge point recommendation list and the learning resource recommendation list combine the learner's learning interest preferences, learning situation, knowledge point inclusion resource types, etc., calculate the knowledge point recommendation degree and learning resource recommendation value score to construct the recommendation list, and the accuracy and reference value are high. The present invention constructs different knowledge cards for different types of entities, and the knowledge cards can structure, visualize and logically memorize a single knowledge point or character of the course, so that users can fully understand and learn each knowledge point after decomposition, strengthen memory and improve learning effects. The present invention recommends friends to users to jointly form a study group, which can enable students to learn together, stimulate learning interest, and better share and disseminate learning results. The present invention can well improve the teaching level.

[0047] Furthermore, the present invention is based on studying students' learning behaviors, that is, combining the location information of students' most recently learned knowledge points in the graph with students' learning situations to recommend friends among students. Among them, students' learning situations are taken into account for learning assistance, and a learner-knowledge point matrix is ​​introduced. Only whether the knowledge point has been learned is used as the basis for similarity calculation. This can avoid the situation where the recommended friends have mastered all or none of the knowledge points brought about by the learner-knowledge point scoring matrix, thereby achieving the effect of learning assistance and common progress. This method improves the commonality and accuracy of students and recommended friends, and makes the availability of information sharing by students higher. At the same time, it also reduces the loneliness and anxiety in the learning process, and can solve problems in the learning process together with friends. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The present invention constructs a flow chart of a cross-media dynamic knowledge graph;

[0049] Figure 2 The flow chart of the subject knowledge question and answer of the present invention;

[0050] Figure 3 Flow chart of the present invention for intelligent learning resource recommendation;

[0051] Figure 4 A flow chart of friend recommendation according to the present invention;

[0052] Figure 5 Schematic diagram of the system architecture of the teaching system based on cross-media dynamic knowledge graph. DETAILED DESCRIPTION

[0053] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0054] See also Figure 1 The specific steps of constructing a cross-media dynamic knowledge graph in the present invention are:

[0055] Step 1: Acquire subject area knowledge; use Python tools to crawl required course-related information through multiple channels, including text, images, videos, exercises, PPT, etc.

[0056] Step 2: Data cleaning and preprocessing: Manually clean the crawled data to ensure data accuracy.

[0057] Step 3: Dataset annotation: BIO method is used to manually annotate text data.

[0058] The BIO annotation method refers to annotating each element in a data set as "BX", "IX" or "O". Among them, "BX" means that the segment where the element is located belongs to type X and the element is at the beginning of the segment, "IX" means that the segment where the element is located belongs to type X and the element is in the middle of the segment, and "O" means that it does not belong to any type. In an embodiment of the present invention, the annotation type may include terms, names of people, places, names of organizations, etc.

[0059] Step 4: Knowledge processing: Use natural language processing techniques such as word segmentation, named entity recognition, and relationship extraction to process the data set, extract the entities and relationships between entities, and clean and integrate the extracted knowledge data.

[0060] The present invention proposes to construct a dictionary of subject-specific terms and add it to Jieba's custom word library during word segmentation, integrating the built-in traditional dictionary to segment the text to ensure the accuracy of the word segmentation results and provide a basis for the completeness and accuracy of the subsequent acquisition of knowledge points.

[0061] The present invention adopts Word2vec as a word vector training model, and uses the sorted subject field data set as the training corpus of the model to train it. Word2vec is a prediction model for efficient learning of word embedding, comprising a three-layer network structure of input layer, hidden layer and output layer.

[0062] The present invention adopts BILSTM+CRF neural network model for entity extraction. This method requires the dataset marked in step 3 above to train the neural network model, and uses the word vector representation as the input of the BILSTM network, uses two LSTM networks to obtain the context information of the word vector to extract features for modeling, and then integrates the calculation of the transfer matrix in CRF, and finally predicts the probability of the word as an entity to complete the entity extraction.

[0063] The present invention adopts a template-based approach to extract attributes and relationships. Attribute extraction mainly describes entities, and entities can be fully expressed through attributes. Relationship extraction extracts the inherent logical relationships between entities, thereby connecting scattered entities to form a semantic network.

[0064] Step 5: Storage and visualization. Integrate the extracted entity sets, relationship sets, images, videos, etc., and import them into Neo4j in batches to complete knowledge storage and form a cross-media knowledge graph. The knowledge graph stores entity-related information in the form of triples of "entity-attribute-attribute value" and "entity-relationship-entity".

[0065] Step 6: Authorization. After being authorized by the system personnel, the knowledge graph can be jointly edited by students and teachers, who can provide feedback and supplement the knowledge points that are missing from the graph during the learning process, continuously improve the knowledge graph, and form a cross-media dynamic knowledge graph.

[0066] See also Figure 2 The specific steps of the present invention for conducting subject knowledge question and answer are:

[0067] Step 1: Analyze the natural language questions raised by the user;

[0068] a. Perform word segmentation and word vector training on the question. In the word segmentation process, use the jieba tool and add pre-set subject-specific dictionaries to improve the accuracy of word segmentation;

[0069] b. Use the question information combined with the pre-set rule template to match entity attributes and select the corresponding attributes of the entity queried by the question.

[0070] Step 2: Use the BILSTM+CRF model to extract entities from natural language questions.

[0071] Step 3: Combine the constructed knowledge graph and match the extracted entities with the entities in the graph.

[0072] a. If there is a matching entity in the knowledge graph, select it;

[0073] b. If there is no matching entity in the knowledge graph, the closest entity is selected using semantic similarity calculation.

[0074] The present invention uses the cosine similarity algorithm to calculate semantic similarity. The cosine similarity calculation is expressed as follows:

[0075]

[0076] Among them, X represents the entity extracted from the question, Y represents the entity in the knowledge graph, and n represents the dimension of the word vector. By calculating the similarity, the entity in the knowledge graph with the highest similarity is selected as the match.

[0077] Step 4: Match the selected entities and attributes with the triples in the knowledge graph;

[0078] Step 5: Return the corresponding attribute value and provide it to the user as the answer to the question.

[0079] See also Figure 3 The specific steps of the present invention for intelligent learning resource recommendation are as follows:

[0080] Step 1: Introduce the learner-resource type matrix. The usage rate of each type of resource by students can reflect the student's resource type usage preference. Define the learner set as: U = {u1,u2,u3,…,u m}, the resource type set is: R = {r1, r2, r3, ..., r q}, thus we can get the learner-resource type matrix, where the row vector represents the students, the column vector represents the resource type, and the matrix element p ij Represents the usage rate of resource type j by student i, i = {1, 2, 3, ..., m}, j = {1, 2, 3, ..., q}. To avoid the cold start problem, the number of times a learner uses each resource type is initialized to 1, and then the number of times is updated according to subsequent learning behavior.

[0081] The resource type usage calculation formula is as follows:

[0082]

[0083] Among them, v ij Indicates the number of times user i uses resource type j.

[0084] Step 2: Introduce the learner-knowledge point scoring matrix. The student's test score for each knowledge point can reflect the student's mastery of the knowledge point. Define the learner set as: U = {u1,u2,u3,…,u m}, the knowledge point set is: K = {k1, k2, k3, ..., k n}, thus we can get the learner-knowledge point scoring matrix, where the row vectors represent students, the column vectors represent knowledge points, and the matrix elements represent the students' scoring rates for the knowledge points.

[0085] a. Through the test test, the student's score for each knowledge point is counted. For subjective questions, a similarity value is obtained by calculating the similarity between the standard answer and the student's answer, and then the student's score is given according to the total score of the question;

[0086] b. Normalize the student's score for each knowledge point, map the obtained score to the interval [0, 1], and obtain the student's score rate for the knowledge point.

[0087] Step 3: Construct a candidate recommended knowledge point list K′, and select candidate recommended knowledge points based on the learner-knowledge point scoring matrix. Set a threshold T. When the student's score rate for a knowledge point is less than the threshold, it indicates that the student's grasp of the knowledge point is relatively weak, and the knowledge point is used as a candidate recommended knowledge point.

[0088] Step 4: Determine the recommended knowledge point list K'1 based on the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point in the knowledge graph. The present invention uses BFS breadth-first search to calculate the shortest path between the most recently learned knowledge point and the candidate recommended knowledge point. The smaller the shortest path distance, the stronger the relevance of the knowledge point, and the higher the degree of recommendation. Use d(o,t) to represent the shortest path distance between the most recently learned knowledge point o and the candidate recommended knowledge point t.

[0089] Step 5: Determine the recommended knowledge point list K'2 based on the similarity between the student's most recently learned knowledge point and the candidate recommended knowledge point. The knowledge point similarity is calculated based on the attribute value of each knowledge point stored in the knowledge graph. The greater the knowledge point similarity, the more worthy of recommendation. In the embodiment of the present invention, the WV-CNN model is used to calculate the text similarity, and sim(o,t) is used to represent the text similarity between the most recently learned knowledge point o and the candidate recommended knowledge point t.

[0090] The WV-CNN Chinese text semantic similarity calculation model consists of three parts:

[0091] The first part is to represent the input text in word vector form. The word vector of the text is obtained by segmenting, encoding and generating vectors for the input Chinese sentences, and it is used as the input of the convolutional neural network (CNN). The second part is CNN, which has four layers: convolutional layer, dropout layer, pooling layer and flatten layer. The weight matrix in the network is continuously adjusted through the gradient descent method to obtain optimized parameters and results. The third part is the output, which is the result of the text semantic similarity calculation.

[0092] Step 6: Generate a knowledge point recommendation list. Linearly combine the recommended knowledge point list K'1 and the recommended knowledge point list K'2 generated in steps 4 and 5 to generate the final knowledge point recommendation list. Select the first F knowledge points based on the knowledge point recommendation degree and recommend them to the learner. it It represents the recommendation degree of candidate recommended knowledge point t for user i. The calculation formula is as follows:

[0093]

[0094] Among them, α and β are adjustment factors, and their values ​​must satisfy α·β=1.

[0095] Step 7: Introduce the knowledge point-resource type matrix. Define the knowledge point set as: K = {k1, k2, k3, ..., k n}, the resource type set is: R = {r1, r2, r3, ..., r q}, the row vector represents the knowledge point, the column vector represents the resource type, and the matrix element a tj Indicates whether knowledge point t contains resource type j, t = {1, 2, 3, ..., n}, j = {1, 2, 3, ..., q}. If it contains, it is recorded as 1, otherwise it is recorded as 0.

[0096] Step 8: Calculate the recommended learning resource score. The recommended learning resource list is obtained by integrating the learner's preference for resource type, the recommended degree of knowledge points, and the resource types contained in the knowledge points. That is, the recommended learning resource score is calculated using the vectors and matrices obtained in steps 1, 6, and 7. The formula for calculating the recommended learning resource score is as follows:

[0097] score itj =p ij ·l it ·a tj

[0098] score itj represents the recommendation value score of user i for type j resource in knowledge point t, p ij represents the usage rate of resource type j by user i, l it represents the recommendation degree of knowledge point t for user i, a tj Indicates whether knowledge point t contains resource type j.

[0099] Step 9: Generate a recommended list of learning resources.

[0100] According to the learning resource recommendation value score, the top F learning resources are selected and recommended to learners.

[0101] like Figure 4 As shown, the specific steps of the present invention for friend recommendation are:

[0102] Step 1: Get the node positions of the most recent knowledge points learned by the current student u and other students v in the knowledge graph. The current student position is represented by A, and the other students' positions are represented by B;

[0103] Step 2: Combined with the feature of the student's most recent learning position based on the knowledge graph, calculate the position similarity between the current student u and other students v. The position similarity calculation formula is as follows:

[0104]

[0105] Among them, d(A, B) refers to the shortest path distance between node A and node B in the knowledge graph, which is calculated by BFS breadth-first search. That is, the farther the distance between A and B, the lower the similarity, and the lower the relevance of the recently learned knowledge; the closer the distance, the higher the similarity, and the higher the relevance of the recently learned knowledge.

[0106] Step 3: Introduce the learner-knowledge point bipartite graph association model and define the learner set as: U = {u1,u2,u3,…,u m}, the knowledge point set is: K = {k1, k2, k3, ..., k n}, we can get the learner-knowledge point matrix, where the row vectors represent students and the column vectors represent students’ learning behaviors about knowledge points. The learned knowledge points are marked as 1 and the rest are marked as 0.

[0107] Step 4: Calculate the similarity between the current student u and other students v in their learning situation. The calculation formula is as follows:

[0108]

[0109] where u ki and v ki Respectively represent the current student u and other students v's understanding of knowledge point k i The learning behavior of n is the total number of knowledge points. That is, the greater the calculated similarity result, the more knowledge points the two students have learned and the more similar their learning situations are.

[0110] Step 5: Calculate the comprehensive similarity between users. Perform linear weighted fusion of location similarity and learning status similarity, and use parameter α to balance the weight relationship between the two. The formula for calculating comprehensive similarity is as follows:

[0111] sim(u,v i )=α·sim(place(u,v i ))+(1-α)·sim(LS(u,v i ))

[0112] Step 6: Return the top N students in terms of comprehensive similarity and recommend them to the current student as friends.

[0113] See also Figure 5 The present invention also provides a teaching system based on a cross-media dynamic knowledge graph, which includes a data layer, a data analysis layer, an application layer and a user layer. Specifically:

[0114] The data layer covers the types of resources needed to build a cross-media dynamic knowledge graph, such as text, images, videos, PPT, etc. Data of different media types are multi-source heterogeneous in form and semantically related to each other. Collecting and processing these multimodal data can realize the integration of multi-source cross-media fragmented knowledge and effectively utilize teaching resources.

[0115] The data analysis layer includes a series of artificial intelligence algorithms, which are used to process crawled teaching resources and provide technical support for the implementation of upper-level applications, such as entity extraction, relationship extraction, attribute extraction, semantic similarity calculation, etc.

[0116] At the application layer, the system mainly provides users with four types of knowledge services, namely subject knowledge Q&A, intelligent learning resource recommendations, course knowledge cards, and friend recommendations.

[0117] At the user level, the system divides users into three categories: students, teachers, and system personnel.

[0118] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the teaching method based on a cross-media dynamic knowledge graph of the present invention.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to impose any limitation on the present invention. Those skilled in the art should understand that, without departing from the spirit and principles of the present invention, the technical solution may be subjected to several simple modifications and substitutions, and these modifications and substitutions are also within the scope of protection covered by the claims.

Claims

1. A teaching method based on cross-media dynamic knowledge graph, characterized in that: include: -Build a cross-media dynamic knowledge graph; -Store entity-related information in the knowledge graph and return the corresponding attribute values ​​of the entity through subject knowledge questions and answers; - Obtain feedback from students and knowledge points, calculate the knowledge point recommendation degree and learning resource recommendation value score, and then construct a knowledge point recommendation list and a learning resource recommendation list to return to the learner; -Build course knowledge cards: Build different knowledge cards for different types of entities. Through user query of entities and matching in the knowledge graph, the corresponding entities and their attribute values ​​are returned in the form of knowledge cards; -Friend recommendation: recommend other users who have similar learning progress and learning status to the current user; The process of constructing a knowledge point recommendation list and a learning resource recommendation list includes: First, by constructing a learner-resource type matrix, a learner-knowledge point rating matrix, and a knowledge point-resource type matrix, we obtain explicit and implicit feedback information about students and knowledge points, including learners' learning interest preferences, learning status, and resource types included in knowledge points. Then, we use a method that combines collaborative filtering with neural networks to calculate the knowledge point recommendation degree and learning resource recommendation value rating, and construct a list based on the rating. Define the learner set as: U = {u1,u2,u3,…,u m }, the resource type set is: R = {r1, r2, r3, ..., r q }, thus we get the learner-resource type matrix, where the row vectors represent students, the column vectors represent resource types, and the matrix elements p ij represents the usage rate of resource type j by student i, i = {1, 2, 3, ..., m}, j = {1, 2, 3, ..., q}; to avoid the cold start problem, the number of times the learner uses each resource type is initialized to 1, and then the number of times is updated according to subsequent learning behavior; the resource type usage rate calculation formula is as follows: Among them, v ij represents the number of times user i uses resource type j; Define the learner set as: U = {u1,u2,u3,…,u m }, the knowledge point set is: K = {k1, k2, k3, ..., k n }, thus we get the learner-knowledge point scoring matrix, where the row vectors represent students, the column vectors represent knowledge points, and the matrix elements represent the student's score rate for the knowledge point; through the test test, we count the student's score for each knowledge point; among them, for subjective questions, we calculate a similarity value by the similarity between the standard answer and the student's answer, and then give the student's score according to the total score of the question; normalize the student's score for each knowledge point, map the obtained score to the interval [0, 1], and get the student's score rate for the knowledge point; Construct a candidate recommended knowledge point list K', and screen the candidate recommended knowledge points based on the learner-knowledge point scoring matrix; set a threshold T. When the student's score rate for a knowledge point is less than the threshold, it indicates that the student's grasp of the knowledge point is relatively weak, and the knowledge point is used as a candidate recommended knowledge point; determine the recommended knowledge point list K'1 based on the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point in the knowledge graph; use BFS breadth-first search to calculate the shortest path between the most recently learned knowledge point and the candidate recommended knowledge point. The smaller the shortest path distance, the stronger the relevance of the knowledge point and the higher the degree of recommendation. Use d(o,t) to represent the shortest path distance between the most recently learned knowledge point o and the candidate recommended knowledge point t; determine the recommended knowledge point list K'2 based on the similarity between the student's most recently learned knowledge point and the candidate recommended knowledge point; according to the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point stored in the knowledge graph, the recommended knowledge point list K'1 is determined based on the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point. The attribute values ​​of each knowledge point are used to calculate the knowledge point similarity. The greater the knowledge point similarity, the more worthy of recommendation. The WV-CNN model is used to calculate the text similarity, and sim(o,t) is used to represent the text similarity between the most recently learned knowledge point o and the candidate recommended knowledge point t. The WV-CNN model consists of three parts: the first part is to represent the input text in terms of words, and obtain the word vector of the text by segmenting, encoding, and generating vectors for the input Chinese sentences, and use it as the input of the convolutional neural network (CNN); the second part is CNN, which has four layers: convolution layer, Dropout layer, pooling layer, and Flatten layer. The weight matrix in the network is continuously adjusted by the gradient descent method to obtain the optimized parameters and results; the third part is the output, that is, the result of the text semantic similarity calculation; Linearly combine the generated recommended knowledge point list K'1 and the recommended knowledge point list K'2 to generate the final knowledge point recommendation list, and select the first F knowledge points based on the knowledge point recommendation degree to recommend to the learner; Use l it represents the knowledge point recommendation degree of candidate recommended knowledge point t for user i, and the calculation formula is as follows: Among them, α and β are adjustment factors, and their values ​​must satisfy α·β=1; Introduce the knowledge point-resource type matrix and define the knowledge point set as: K = {k1, k2, k3, …, k n }, the resource type set is: R = {r1, r2, r3, ..., r q }, the row vector represents the knowledge point, the column vector represents the resource type, and the matrix element a tj Indicates whether knowledge point t contains resource type j, t = {1, 2, 3, ..., n}, j = {1, 2, 3, ... q}; if it contains, it is recorded as 1, otherwise it is recorded as 0; The learning resource recommendation list is obtained by integrating learners' resource type preferences, knowledge point recommendation degree, and resource types included in the knowledge points. The learning resource recommendation value calculation formula is as follows: score itj =p ij ·l it ·a tj Among them, score itj represents the recommendation value score of user i for type j resource in knowledge point t, p ij represents the usage rate of resource type j by user i, l it represents the recommendation degree of knowledge point t for user i, a tj Indicates whether knowledge point t contains resource type j; selects the top F learning resources based on the learning resource recommendation value score and recommends them to learners.

2. The teaching method based on cross-media dynamic knowledge graph according to claim 1 is characterized in that: Building a cross-media dynamic knowledge graph includes: using Python tools to crawl relevant information of the required courses, including text, images, videos, exercises and PPT; cleaning the crawled data and manually annotating the text data using the BIO method; extracting entities and attributes and relationships between entities in the processed data set, integrating the extracted entity sets, relationship sets, images and videos to form a cross-media knowledge graph. The cross-media knowledge graph is edited by students and teachers after being authorized by the system personnel, and feedback and supplementation are provided for knowledge points missing from the graph during the learning process to form a cross-media dynamic knowledge graph.

3. The teaching method based on cross-media dynamic knowledge graph according to claim 2 is characterized in that: Construct a dictionary of subject-specific terms, add it to Jieba's custom word library during word segmentation, and integrate the built-in traditional dictionary to segment the text; use Word2vec as the word vector training model, and use the sorted subject field data set as the training corpus of the model to train it; use the BILSTM+CRF neural network model for entity extraction, and train the neural network model with labeled text data, and use the word vector representation as the input of the BILSTM network, use two LSTM networks to obtain the context information of the word vector to extract features for modeling, and then integrate the calculation of the transfer matrix in CRF, and finally predict the probability of the word as an entity to complete the entity extraction; use a template-based approach to extract attributes and relationships. Attribute extraction is to describe the entity and express the entity completely through attributes; relationship extraction is to extract the internal logical relationship between entities, connect the scattered entities, and form a semantic network together; by integrating the extracted entity set, relationship set, images, and videos, and importing them into Neo4j in batches, knowledge storage is completed to form a cross-media knowledge graph.

4. The teaching method based on cross-media dynamic knowledge graph according to claim 1 is characterized in that: The process of subject knowledge question answering includes: storing entity-related information in the knowledge graph in the form of "entity-attribute-attribute value" triples. When a question is asked, first use named entity recognition to extract the entities in the question, link them with the knowledge graph, find the corresponding entities, and then use the question information combined with the pre-set rule template to match the entity attributes and return the corresponding attribute values.

5. The teaching method based on cross-media dynamic knowledge graph according to claim 4 is characterized in that: Segment the questions and train word vectors. During the segmentation process, use the jieba tool and add the pre-set subject-specific dictionary. Use the BILSTM+CRF model to extract entities from natural language questions. Combined with the constructed knowledge graph, match the extracted entities with the entities in the graph. If there is a matching entity in the knowledge graph, select the entity. If there is no matching entity in the knowledge graph, use the cosine similarity algorithm to calculate the semantic similarity and select the closest entity. Match the selected entities and attributes with the triples in the knowledge graph. Return the corresponding attribute value and provide it to the user as the answer to the question.

6. The teaching method based on cross-media dynamic knowledge graph according to claim 1 is characterized in that: The process of friend recommendation includes: obtaining the node positions of the most recent learning knowledge points of the current student u and other students v in the knowledge graph, with the current student position represented by A and the other students positions represented by B; combining the feature of the student's most recent learning position based on the knowledge graph, calculating the position similarity between the current student u and other students v. The position similarity calculation formula is as follows: Among them, d(A, B) refers to the shortest path distance between node A and node B in the knowledge graph, which is calculated by BFS breadth-first search. That is, the farther the distance between A and B, the lower the similarity, and the lower the relevance of the recently learned knowledge; the closer the distance, the higher the similarity, and the higher the relevance of the recently learned knowledge; Introduce the learner-knowledge point bipartite graph association model and define the learner set as: U = {u1,u2,u3,…,u m }, the knowledge point set is: K = {k1, k2, k3, ..., k n }, thus we get the learner-knowledge point matrix, where the row vectors represent students and the column vectors represent students’ learning behaviors about knowledge points, where the learned knowledge points are marked as 1 and the rest are 0; Calculate the similarity between the current student u and other students v in their learning situation. The calculation formula is as follows: where u ki and v ki Respectively represent the current student u and other students v's understanding of knowledge point k i learning behavior, n is the total number of knowledge points. The greater the calculated similarity result, the more knowledge points the two students have learned and the more similar their learning situations are. Calculate the comprehensive similarity between users; perform linear weighted fusion of location similarity and learning situation similarity, and use parameter α to balance the weight relationship between the two. The comprehensive similarity calculation formula is as follows: sim(u,v i )=α·sim(place(u,v i ))+(1-α)·sim(LS(u,v i )) Return the top N students in terms of comprehensive similarity and recommend them to the current student as friends.

7. A teaching system based on a cross-media dynamic knowledge graph, characterized by: It includes data layer, data analysis layer, application layer and user layer; the data layer covers the resource types required to build a cross-media dynamic knowledge graph. Data of different media types are multi-source and heterogeneous in form and semantically related to each other; the data analysis layer is used to process the crawled teaching resources, and the processing methods include entity extraction, relationship extraction, attribute extraction, and semantic similarity calculation; The application layer provides knowledge services for users, including subject knowledge questions and answers, learning resource recommendations, course knowledge cards, and friend recommendations; The user layer divides users into three categories, namely students, teachers and system personnel; The process of constructing a knowledge point recommendation list and a learning resource recommendation list includes: First, by constructing a learner-resource type matrix, a learner-knowledge point rating matrix, and a knowledge point-resource type matrix, we obtain explicit and implicit feedback information about students and knowledge points, including learners' learning interest preferences, learning status, and resource types included in knowledge points. Then, we use a method that combines collaborative filtering with neural networks to calculate the knowledge point recommendation degree and learning resource recommendation value rating, and construct a list based on the rating. Define the learner set as: U = {u1,u2,u3,…,u m }, the resource type set is: R = {r1, r2, r3, ..., r q }, thus we get the learner-resource type matrix, where the row vectors represent students, the column vectors represent resource types, and the matrix elements p ij represents the usage rate of resource type j by student i, i = {1, 2, 3, ..., m}, j = {1, 2, 3, ..., q}; to avoid the cold start problem, the number of times the learner uses each resource type is initialized to 1, and then the number of times is updated according to subsequent learning behavior; the resource type usage rate calculation formula is as follows: Among them, v ij represents the number of times user i uses resource type j; Define the learner set as: U = {u1,u2,u3,…,u m }, the knowledge point set is: K = {k1, k2, k3, ..., k n }, thus we get the learner-knowledge point scoring matrix, where the row vectors represent students, the column vectors represent knowledge points, and the matrix elements represent the student's score rate for the knowledge point; through the test test, we count the student's score for each knowledge point; among them, for subjective questions, we calculate a similarity value by the similarity between the standard answer and the student's answer, and then give the student's score according to the total score of the question; normalize the student's score for each knowledge point, map the obtained score to the interval [0, 1], and get the student's score rate for the knowledge point; Construct a candidate recommended knowledge point list K', and screen the candidate recommended knowledge points based on the learner-knowledge point scoring matrix; set a threshold T. When the student's score rate for a knowledge point is less than the threshold, it indicates that the student's grasp of the knowledge point is relatively weak, and the knowledge point is used as a candidate recommended knowledge point; determine the recommended knowledge point list K'1 based on the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point in the knowledge graph; use BFS breadth-first search to calculate the shortest path between the most recently learned knowledge point and the candidate recommended knowledge point. The smaller the shortest path distance, the stronger the relevance of the knowledge point and the higher the degree of recommendation. Use d(o,t) to represent the shortest path distance between the most recently learned knowledge point o and the candidate recommended knowledge point t; determine the recommended knowledge point list K'2 based on the similarity between the student's most recently learned knowledge point and the candidate recommended knowledge point; according to the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point stored in the knowledge graph, the recommended knowledge point list K'1 is determined based on the positional relationship between the student's most recently learned knowledge point and the candidate recommended knowledge point. The attribute values ​​of each knowledge point are used to calculate the knowledge point similarity. The greater the knowledge point similarity, the more worthy of recommendation. The WV-CNN model is used to calculate the text similarity, and sim(o,t) is used to represent the text similarity between the most recently learned knowledge point o and the candidate recommended knowledge point t. The WV-CNN model consists of three parts: the first part is to represent the input text in terms of words, and obtain the word vector of the text by segmenting, encoding, and generating vectors for the input Chinese sentences, and use it as the input of the convolutional neural network (CNN); the second part is CNN, which has four layers: convolution layer, Dropout layer, pooling layer, and Flatten layer. The weight matrix in the network is continuously adjusted by the gradient descent method to obtain the optimized parameters and results; the third part is the output, that is, the result of the text semantic similarity calculation; Linearly combine the generated recommended knowledge point list K'1 and the recommended knowledge point list K'2 to generate the final knowledge point recommendation list, and select the first F knowledge points based on the knowledge point recommendation degree to recommend to the learner; Use l it represents the knowledge point recommendation degree of candidate recommended knowledge point t for user i, and the calculation formula is as follows: Among them, α and β are adjustment factors, and their values ​​must satisfy α·β=1; Introduce the knowledge point-resource type matrix and define the knowledge point set as: K = {k1, k2, k3, …, k n }, the resource type set is: R = {r1, r2, r3, ..., r q }, the row vector represents the knowledge point, the column vector represents the resource type, and the matrix element a tj Indicates whether knowledge point t contains resource type j, t = {1, 2, 3, ..., n}, j = {1, 2, 3, ... q}; if it contains, it is recorded as 1, otherwise it is recorded as 0; The learning resource recommendation list is obtained by integrating learners' resource type preferences, knowledge point recommendation degree, and resource types included in the knowledge points. The learning resource recommendation value calculation formula is as follows: score itj =p ij ·l it ·a tj Among them, score itj represents the recommendation value score of user i for type j resource in knowledge point t, p ij represents the usage rate of resource type j by user i, l it represents the recommendation degree of knowledge point t for user i, a tj Indicates whether knowledge point t contains resource type j; selects the top F learning resources based on the learning resource recommendation value score and recommends them to learners.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the teaching method based on a cross-media dynamic knowledge graph as described in any one of claims 1 to 6 are implemented.

Citation Information

Cited By

  • A knowledge graph generation method for English teaching

    CN123031822A