Intelligent teaching system based on cross-modal embedding

By introducing cross-modal embedding technology in the intelligent question-and-answer system, combining BERT model and knowledge graph, the existing system's in-depth analysis and personalized guidance in the handling of complex problems is solved, and more accurate semantic understanding and personality are achieved Recommended learning paths to improve students' learning effectiveness.

CN119067816APending Publication Date: 2024-12-03NORTHWEST NORMAL UNIVERSITY
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411297015.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-18
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing intelligent question-and-answer system lacks in-depth analysis and personalized guidance when dealing with complex problems, and cannot provide detailed background information and related knowledge points, resulting in students' in-depth understanding of knowledge.

Method used

Using an intelligent teaching system based on cross-modal embedding, combining BERT model, knowledge graph and deep learning model, capture context information through a bidirectional encoder, calculate word correlation by self-attention mechanism, feedforward neural network performs nonlinear transformation, and path search The algorithm queries knowledge points and expresses the relationship between knowledge points through embedded methods. The deep learning model predicts the learning path for personalized recommendations.

Benefits of technology

The system has improved the semantic understanding of students' questions, provided in-depth analysis and personalized learning path recommendations, helping students to deeply understand knowledge and improve learning efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure SMS_1
    Figure SMS_1
  • Figure SMS_3
    Figure SMS_3
  • Figure SMS_4
    Figure SMS_4
Patent Text Reader

Abstract

The invention provides an intelligent teaching system and method based on cross-modal embedding, a Transform architecture and a BERT model are utilized to improve the semantic understanding ability of the system for questions of students, through context analysis and emotion recognition, the intention and confusion points of the questions of the students are accurately mastered, the knowledge graph is constructed and applied, and the knowledge points of all subjects are analyzed to improve the knowledge points of all subjects, so that the intent and confusion points of the questions of the students can be accurately understood, and the intent and confusion points of the questions of the students can be accurately understood. And constructing a knowledge graph, and associating the relationship among the knowledge points. After students ask questions, deep analysis is provided through the knowledge graph, related knowledge points and internal relations thereof are displayed, the students are helped to deeply understand and self-adaptively learn path recommendation, learning paths and resources are recommended in a personalized mode according to historical learning data and current question content of the students, and the students are helped to systematically solve problems in learning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing and relates to an intelligent teaching system and method based on cross-modal embedding. Background Art

[0002] Intelligent question answering systems (QA systems) are widely used in modern education, aiming to answer students' questions through natural language processing (NLP) technology. These systems can respond to students' questions quickly, but often lack in-depth analysis and personalized guidance. Currently, many QA systems mainly rely on keyword matching-based technology, which works well when dealing with simple questions, but its performance seems to be inadequate when facing complex questions. In addition, such systems usually fail to provide detailed background information and the association of relevant knowledge points during the answering process, resulting in students' insufficient understanding of the knowledge. Summary of the invention

[0003] The purpose of the present invention is to provide an intelligent teaching device based on cross-modal embedding and reinforcement learning in order to solve the problems existing in the prior art.

[0004] To this end, the present invention adopts the following technical solutions: An intelligent teaching system based on cross-modal embedding, comprising: BERT model, which is used to capture contextual relationships in sentences and accurately understand the semantics of student question sentences. The BERT model includes a bidirectional encoder, a self-attention mechanism, a feedforward neural network, and a layer normalization module. The bidirectional encoder is used to capture contextual information, the self-attention mechanism is used to calculate the correlation between words, the feedforward neural network is used to receive the output from the self-attention mechanism and perform nonlinear transformation, and the layer normalization module is numerically stable; A knowledge graph, which can find relevant knowledge points through a path search algorithm and parse them, and represent the knowledge points and relationships as vectors through an embedding method; A deep learning model is used to predict students' learning paths and make personalized recommendations.

[0005] Furthermore, the path search algorithm is a priority search BFS Or Depth First Search DFS.

[0006] Furthermore, the deep learning model adopts RNN or LSTM Model.

[0007] The present invention also provides an intelligent teaching method based on cross-modal embedding, which adopts the above-mentioned intelligent teaching system based on cross-modal embedding, comprising the following steps: (1) The BERT model in the Transformer architecture is used to capture the contextual information in the semantics of students’ questions through a bidirectional encoder; (2) Constructing a knowledge graph containing knowledge points of various disciplines, wherein the knowledge graph includes multiple knowledge points and the relationships between the knowledge points, then using a path search algorithm to find relevant knowledge points and parse them, and finally using an embedding method to represent the relationships between knowledge points as vectors; (3) Based on students’ learning behavior data, use deep learning models to predict students’ learning paths and make personalized recommendations.

[0008] The beneficial effects of the present invention are: The Transformer architecture and BERT model are used to improve the system's ability to understand the semantics of students' questions. Through context analysis and sentiment recognition, the intention and confusion of students' questions can be accurately grasped. The knowledge graph is constructed and applied based on the knowledge points of each subject to associate the relationship between each knowledge point. After students ask questions, the knowledge graph provides in-depth analysis, displays relevant knowledge points and their internal connections, and helps students to gain a deeper understanding. Adaptive learning path recommendation: Based on students' historical learning data and current question content, personalized learning paths and resources are recommended to help students systematically solve learning problems. DETAILED DESCRIPTION

[0009] The present invention is described in detail below in conjunction with embodiments: Example 1 An intelligent teaching system based on cross-modal embedding, including a BERT model, a knowledge graph and a deep learning model, wherein: The BERT model uses the BERT model in the Transformer architecture. The BERT model is used to capture the contextual relationship in the student's question sentence and accurately understand the semantics of the student's question sentence. Specifically, the BERT model includes a bidirectional encoder, a self-attention mechanism, a feedforward neural network, and a layer normalization module. The bidirectional encoder is used to capture the contextual information in the student's question sentence to improve the accuracy of semantic understanding. The self-attention mechanism is used to flexibly calculate the correlation between words and deal with long-distance dependency problems, further improving the system's ability to understand student questions, so as to provide more accurate answers. At the same time, it can also analyze students' learning behaviors. The feedforward neural network is used for feature conversion, and the layer normalization module is numerically stable. Knowledge graph, based on the semantics of the question sentence obtained by the BERT model, the knowledge graph can find relevant knowledge points through the path search algorithm and parse them, and represent the knowledge points and relationships as vectors through the embedding method for easy calculation and query. The knowledge graph can not only provide preliminary answers, but also provide relevant knowledge background and related information to help students understand the problem in depth. In this embodiment, the path search algorithm adopts the priority search BFS Or Depth First Search DFS.

[0010] Deep learning model. In this embodiment, the deep learning model adopts an RNN or LSTM model. Through the learning path prediction model, based on the semantics of the student question sentences collected by the BERT model and the preliminary answers provided by the knowledge graph, the deep learning model can dynamically recommend personalized learning paths and resources based on the student's learning behavior.

[0011] By applying the BERT model, knowledge graph construction and application, and learning path prediction model, the present invention solves the shortcomings of existing intelligent question-answering systems in semantic understanding and deep analysis. The above algorithm design scheme not only improves the system's accurate understanding and deep analysis capabilities, but also provides personalized learning paths and resource recommendations based on students' learning behaviors, thereby improving students' learning efficiency and effectiveness.

[0012] Example 2 An intelligent teaching method based on cross-modal embedding, using the above-mentioned intelligent teaching system based on cross-modal embedding, comprises the following steps: (1) The BERT model in the Transformer architecture is used to capture the contextual information and relationships in the semantics of the sentences asked by students through a bidirectional encoder to improve the accuracy of semantic understanding. The core algorithm model is as follows: in, It is l The middle layer i The hidden state vector of a word. Specifically, the first l The layer is one of the levels of the Transformer Encoder block stacked inside the BERT model. It is responsible for processing the input data and generating the corresponding hidden state vector. LayerNorm Layer normalization operation, LayerNorm The layer normalization operation is mainly used for numerical stability. LayerNorm Normalizing each sample in each feature dimension makes the data in each dimension more stable during the training process, helps to strengthen the features of the data in each dimension, and enables the model to use these features more effectively for learning and reasoning. By normalizing the features of each sample, LayerNorm It also reduces the distribution variation of the input to each layer, helping to alleviate the problem of internal covariate shift during training. LayerNorm Specifically, normalization is performed channel by channel without involving batch dimensions, so the number of parameters is small. It is more suitable for small batches or single samples, which helps to reduce the complexity and computational complexity of the model and improve the training and reasoning efficiency of the model. Attention It is a self-attention mechanism, which is used to calculate the correlation between words. Attention The algorithm model is as follows: in, Q is the query matrix, K T is the transpose of the key matrix, v is the value matrix, d k is the scaling factor.

[0013] FFN It is a feedforward neural network used for feature conversion. The algorithm model is as follows: in, W 1 and W 2 is the weight matrix, b 1 and b 2 is the bias term.

[0014] Since the bidirectional encoder can capture the contextual information at the same time and improve the accuracy of semantic understanding, the bidirectional encoder of the BERT model used in the present implementation captures the contextual relationship in the sentence, so as to more accurately understand the semantics of the student's question. The self-attention mechanism can flexibly calculate the correlation between words and handle long-distance dependency problems, thereby improving the system's ability to understand students' questions and providing more accurate answers.

[0015] The feedforward neural network layer receives the output from the self-attention mechanism and performs further nonlinear transformations on it. This transformation allows the model to learn complex representations of the input data and extract high-level features that are useful for downstream tasks. In BERT, this helps the model gain a deeper understanding of the semantic information of text data.

[0016] (2) Construct a knowledge graph containing knowledge points of various subjects. The knowledge graph includes multiple knowledge points and the relationships between knowledge points. Each node in the knowledge graph represents a knowledge point, and the edge represents the relationship between knowledge points.

[0017] The construction and query of knowledge graphs are mainly based on the following algorithms: in, e Indicates knowledge points, r ij Representing knowledge points e i And knowledge points e j The relationship vector between Rel Represents a relational calculation function, usually using TransE , TransH In this embodiment, the embedding method is used to express TransE,TransE The model algorithm is as follows: To capture the contextual information and relationships in the semantics of the sentences asked by students, a path search algorithm, such as priority search BFS or depth-first search DFS, is used to find relevant knowledge points and parse them. TransE The embedding method represents knowledge points and the relationships between them as vectors, which is convenient for computer query and provides preliminary answers to students asking questions, as well as relevant knowledge background and associated information, helping students to understand the problems in depth.

[0018] (3) Based on students’ learning behavior data, use deep learning models to predict students’ learning paths and make personalized recommendations.

[0019] The algorithm model of the learning path prediction model is as follows: in: X t For students at time step t The learning behavior feature vector, h t is the time step t The hidden state vector of W xh , W hh and W hy is the weight matrix of the model, b h and b y is the bias term, As the activation function, usually used sigmoid or tanh , softmax The probability distribution used to generate the forecast results.

[0020] Through the learning path prediction model, we can dynamically recommend personalized learning paths and resources based on students' learning behaviors. By applying the BERT model, knowledge graph construction and application, and learning path prediction model, we aim to address the shortcomings of existing intelligent question-answering systems in semantic understanding and deep analysis. The above algorithm design not only improves the system's accurate understanding and deep analysis capabilities, but also provides personalized learning paths and resource recommendations based on students' learning behaviors, thereby improving students' learning efficiency and effectiveness.

Claims

1. An intelligent teaching system based on cross-modal embedding, It is characterized in that include: BERT model, which is used to capture contextual relationships in sentences and accurately understand the semantics of student question sentences. The BERT model includes a bidirectional encoder, a self-attention mechanism, a feedforward neural network, and a layer normalization module. The bidirectional encoder is used to capture contextual information, the self-attention mechanism is used to calculate the correlation between words, the feedforward neural network is used to receive the output from the self-attention mechanism and perform nonlinear transformation, and the layer normalization module is numerically stable; A knowledge graph, which can find relevant knowledge points through a path search algorithm and parse them, and represent the knowledge points and relationships as vectors through an embedding method; A deep learning model is used to predict students' learning paths and make personalized recommendations.

2. According to claim 1, an intelligent teaching system based on cross-modal embedding and reinforcement learning, It is characterized in that The path search algorithm is a priority search BFS Or Depth First Search DFS.

3. According to claim 1, an intelligent teaching system based on cross-modal embedding and reinforcement learning, It is characterized in that The deep learning model adopts RNN or LSTM Model.

4. An intelligent teaching method based on cross-modal embedding, It is characterized in that An intelligent teaching system based on cross-modal embedding according to any one of claims 1 to 3 is adopted, comprising the following steps: (1) The BERT model in the Transformer architecture is used to capture the contextual information in the semantics of students’ questions through a bidirectional encoder; (2) Constructing a knowledge graph containing knowledge points of each subject, wherein the knowledge graph includes multiple knowledge points and the relationships between knowledge points. Then, a path search algorithm is used to find relevant knowledge points and analyze them in response to students' questions. Finally, the relationships between knowledge points are represented as vectors through an embedding method. (3) Based on students’ learning behavior data, use deep learning models to predict students’ learning paths and make personalized recommendations.

Citation Information

Cited By

  • Robot online teaching method and system based on artificial intelligence

    CN119741172A