Intelligent note system training method based on artificial intelligence

Abstract

The invention provides an intelligent note system training method based on artificial intelligence, which comprises the following steps: analyzing file data in various formats, performing word segmentation and NER by using tools such as Jieba and Spacy, and constructing a graph database in combination with TF-IDF to realize mapping and association of learning resources and notes so as to form a structured knowledge system. The system can generate targeted practice examination questions, collect student data to establish a learner portrait, generate a personalized review plan based on an Ebbinghaus forgetting curve, and continuously iterate the model according to new student learning data. According to the method, the problems of incomplete knowledge point extraction, poor learning resource relevance and poor personalized recommendation in the prior art are effectively solved, and the learning efficiency and effect are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and educational technology, and in particular to an intelligent note-taking system based on artificial intelligence technology and its application in computer science education. Background Art

[0002] With the rapid development of information technology and artificial intelligence technology, the education industry is gradually transforming towards digitalization and intelligence. Traditional teaching models and learning methods are difficult to meet the requirements of modern education for personalized learning. Through the intelligent note-taking system, students can manage learning materials more efficiently and obtain personalized learning suggestions, thereby improving learning results. However, the existing intelligent note-taking systems still have shortcomings in terms of the accuracy of information extraction, the relevance of data structures, and the effectiveness of personalized recommendations.

[0003] Existing technologies mainly rely on students to manually record and organize their study notes, which is not only inefficient but also prone to omissions of important knowledge points. At the same time, traditional note-taking systems usually lack the ability to parse and integrate data in multiple file formats (such as PDF, DOCX, TXT, etc.), and cannot fully utilize the various learning resources uploaded by students. In addition, when constructing a knowledge system, existing technologies often use simple keyword extraction or word frequency-based statistical methods, and fail to fully utilize modern natural language processing and machine learning technologies to improve the accuracy and systematicness of knowledge point extraction.

[0004] In order to overcome the shortcomings of the prior art, the present invention proposes a training method for an intelligent note-taking system based on artificial intelligence. The method parses data in various file formats uploaded by students, uses Jieba Chinese word segmentation library and SpacyNLP toolkit for word segmentation, and then combines TF-IDF and named entity recognition (NER) technology to extract knowledge points and key information in the text, and stores them in a structured graph database. By establishing a structured knowledge system, effective mapping and association of learning resources and student notes are achieved, thereby generating personalized learning suggestions and review plans for students.

[0005] Compared with the prior art, the intelligent note-taking system of the present invention has the following advantages: First, the present invention can efficiently parse data in files of various formats and accurately extract knowledge points, thus avoiding the problem of incomplete information extraction in traditional note-taking systems. Secondly, the present invention improves the systematicness and relevance of the knowledge system by constructing a graph database to perform structured management of learning resources and student notes. Finally, the present invention dynamically generates a review plan through a personalized recommendation algorithm in combination with the student's learning situation and characteristics, thereby improving the learning effect.

[0006] In summary, the training method of the artificial intelligence-based intelligent note-taking system proposed in the present invention makes full use of natural language processing and machine learning technology, has significant advantages in knowledge point extraction, learning resource association and personalized recommendation, and provides a new solution for achieving more efficient personalized learning. Summary of the invention

[0007] In view of this, the purpose of the present invention is a training method for an intelligent note-taking system based on artificial intelligence, which aims to solve the problems of incomplete knowledge point extraction, poor correlation of learning resources, and poor personalized recommendation effect in the prior art. It can efficiently extract knowledge points from a variety of file formats (PDF, DOCX, TXT) and build a structured knowledge system. The system can achieve effective mapping and association between learning resources and student notes, and generate questions at the same time, thereby providing students with a dynamic review plan.

[0008] The training method of the intelligent note-taking system proposed by the present invention comprises the following steps:

[0009] Multi-format file data extraction: By parsing the various file formats (PDF, DOCX, TXT) uploaded by students, extracting text, pictures, icons and other data, converting the file content into plain text data and storing it in categories.

[0010] Knowledge points and key information extraction: Based on the extracted plain text data, Jieba Chinese word segmentation library and SpacyNLP toolkit are used for word segmentation, and TF-IDF and named entity recognition (NER) technology are combined to identify and extract high-weight words and named entities in the text. These knowledge points and key information are stored in a structured way in the graph database.

[0011] Build a structured knowledge system: Associate the processed learning resources with the student notes and store them in the graph database in the form of nodes and edges to form a structured knowledge system. By calculating the similarity between texts, use the RoBERTa model to find the relevant points between knowledge points and notes, and map them to the knowledge system.

[0012] Generate targeted practice test questions: Based on the structured knowledge system, use the pre-trained model in the Transformers library to generate preliminary practice test content, including multiple-choice questions, fill-in-the-blank questions, short-answer questions, etc. Use the Generative Adversarial Network (GANs) model to optimize the generated questions and manage students' answer records.

[0013] Learner portrait and personalized review plan: By analyzing students' question-answering records and notes, we optimize the mapping of the knowledge system and build learner portraits based on students' learning habits and characteristics. Based on the Ebbinghaus forgetting curve model, we generate and dynamically adjust the review plan, optimize the review interval, and improve learning results.

[0014] The method of the present invention can realize efficient parsing and knowledge point extraction of multi-format file data, build a structured knowledge system, and provide students with personalized learning suggestions and review plans. By utilizing modern artificial intelligence technology, the present invention significantly improves the efficiency and learning effect of the intelligent note-taking system, overcoming the shortcomings of the prior art.

[0015] According to another aspect of the present invention, a training method for an intelligent note-taking system based on artificial intelligence is proposed, comprising:

[0016] The data collection and preprocessing module, knowledge extraction and annotation module, knowledge system module, question generation module, learner portrait module, and continuous update module are connected in sequence;

[0017] The data acquisition and preprocessing module extracts multi-format file data and parses the various file formats uploaded by students to realize the basis of system intelligence;

[0018] The knowledge extraction and annotation model uses Jieba Chinese word segmentation library and SpacyNLP toolkit for word segmentation, and then combines TF-IDF and NER to extract knowledge points and key information in text data to build a graph database;

[0019] The knowledge system module maps and associates learning resources and student learning notes based on a graph database, stores them in the form of nodes and edges, and forms a structured knowledge system;

[0020] The question generation module generates targeted practice test questions based on the structured knowledge system and obtains students' question-taking records;

[0021] The learner portrait module optimizes the mapped knowledge system based on the student's question record and note content, and establishes a learner portrait;

[0022] The continuous updating module collects student learning data for optimizing learner portraits, review plans, Ebbinghaus forgetting curve models and knowledge systems;

[0023] In the above technical solution, knowledge points can be efficiently extracted from a variety of file formats (PDF, DOCX, TXT), and a structured knowledge system can be constructed. The system can achieve effective mapping and association between learning resources and student notes, and generate questions at the same time, so as to provide students with a dynamic review plan, solve the problems of incomplete knowledge point extraction, poor correlation of learning resources, and poor personalized recommendation effect in the prior art, and the present invention provides a personalized learning experience by utilizing advanced machine learning algorithms. The system can dynamically adjust the learning plan based on the real-time feedback of students and the learning data collected daily to ensure that the learning plan matches the actual needs of students. In the specific implementation process, the present invention adopts high-quality and diverse training data sets, covering student data of different backgrounds, disciplines and learning styles. The use of this data set significantly improves the accuracy of the artificial intelligence model and reduces the deviation problem caused by the use of limited or biased data in traditional technologies. The present invention can provide more accurate and fairer teaching effects, which has significant advantages over the prior art. In addition, through the application of reinforcement learning algorithms, the system can continuously optimize learning strategies. The system will continuously monitor the learning progress and feedback of students, and adjust the teaching content and review plan according to real-time data. This dynamic adjustment not only improves students' learning efficiency, but also enables more accurate delivery and consolidation of knowledge points in personalized teaching.

[0024] In some embodiments, the data acquisition and preprocessing module includes:

[0025] The system supports a wide range of file formats, including PDF, DOCX, and TXT. These files contain text, images, charts, and code snippets. Advanced file parsing tools Apache Tika and Tesseract are used to deeply parse non-text data and convert it into processable plain text and structured data. The word embedding technology RoBERTa is used to convert the parsed text data into a context-sensitive high-dimensional vector representation. The data is processed by a batch processor to form batches suitable for deep learning models, preparing for subsequent training and inference.

[0026] In some embodiments, the knowledge extraction and annotation model includes:

[0027] Jieba Chinese word segmentation library and SpacyNLP toolkit are used to segment the extracted plain text data, identify word boundaries, and perform part-of-speech tagging. The weight of each word in the text is calculated, and high-weight words are identified. Then, TF-IDF and NER are combined to identify important concepts, algorithms, code examples, and named entities in these words in the book, and stored in the database in a structured manner. For professional terms such as algorithm names and technical terms, Jieba's dictionary is expanded to improve the accuracy of word segmentation, and a knowledge base containing common algorithm names is built. During the recognition process, the candidate words extracted from the text are matched with the knowledge base to further improve the accuracy of algorithm name recognition. TF-IDF is used to calculate word weights and the frequency of each word in the text. Considering the distribution of words in the entire document collection, rare words are given higher weights. Spacy and HanLP are used to perform named entity recognition on Chinese and English texts, respectively, to identify names of people, places, institutions, and proper nouns.

[0028] In some embodiments, the knowledge system module includes:

[0029] Using the graph database management system Neo4j, each knowledge point and key information will be stored as a node in the graph database. Nodes include name, type (algorithm, concept, code example), description, and source. The association information between different knowledge points is stored as relationships in the graph database. For example, an algorithm may be associated with a specific code example, or a concept may form a subordinate relationship with other related concepts. Relationships contain attributes such as association type, weight, and superior-female relationship. Then calculate the similarity between texts, use the RoBERTa model to calculate the similarity, set a threshold to find the knowledge points and notes related points, and add them to the database to form a knowledge system, which is used to generate review plans and construct exercises.

[0030] In some embodiments, the topic generation module includes:

[0031] Based on the structured knowledge system, the text generation variant in the BERT series is used and fine-tuned. T5 (Text-To-Text Transfer Transformer) unifies various NLP tasks into the form of "text to text" to complete various question generation tasks. For fine-tuning, materials extracted from the book are used as question data sets, including multiple-choice questions, fill-in-the-blank questions, and short-answer questions, and the knowledge points corresponding to the questions are marked. For example, the multiple-choice question, "What is the order of pre-order traversal of a binary tree? A. Left root right B. Root left and right C. Left and right root" corresponds to the knowledge point "binary tree traversal", the fill-in-the-blank question, "In Python, the keyword used to create anonymous functions is ______." corresponds to the knowledge point "Python anonymous function", and the short-answer question, "Please explain the basic idea of ​​the quick sort algorithm." corresponds to the knowledge point "Quick sort", and then the GANs model is used to optimize the generated questions. Specifically, the fine-tuned T5 model is used as the generator to generate questions. A discriminator is trained based on RNN to distinguish the difference between the generated questions and the real questions. By alternately training the generator and the discriminator, the generator can generate questions that are closer and closer to the quality of the real questions. Design a database to store and manage students' answer records, including student ID, question ID, answer time, answer, correct answer, score, to help build learner portraits.

[0032] In some embodiments, the learner portrait module includes:

[0033] The students' note data are classified and sorted according to subjects, chapters, and knowledge points, and cleaned and preprocessed to eliminate noise, fill missing values, convert data types, and ensure the accuracy and consistency of the data. In combination with the test results, the mastery of each knowledge point is analyzed according to the structured knowledge system, and student characteristics are collected as the basis for establishing learner portraits. Learner characteristics are collected, including knowledge level, learning habits, learning style, and interest preferences. Learners are clustered according to similar characteristics to form different learner groups. On the basis of cluster analysis, a multi-layer perceptron MLP model is constructed by combining decision trees and neural networks to construct and refine learner portraits. The specific steps include optimizing model parameters using the back propagation algorithm for more refined learning and prediction, evaluating the performance of the model through cross-validation, ensuring that the model can accurately identify and classify learner characteristics, and then combining K-means to realize automatic identification and classification of learner characteristics, and using the DAX function in PowerBI to create dynamic labels to display learners' feature labels under different conditions. The learner portrait is displayed in word cloud charts and scatter plots.

[0034] In some embodiments, the continuous updating module includes:

[0035] Based on the learning data in the student characteristics, the Ebbinghaus forgetting curve is used to calculate the forgetting rate of each knowledge point, and the interval time is optimized according to the student feedback.

[0036] For each knowledge point, the exponential decay model of the Ebbinghaus forgetting curve is used to calculate the forgetting rate according to the time after learning (in days). -λt , where R(t) is the forgetting rate after time t (usually between 0 and 1, 0 means complete memory, and 1 means complete forgetting), λ is the forgetting constant (the larger the λ value, the faster the forgetting; conversely, the slower the forgetting), and t is the time from the beginning of learning (in days). The forgetting rate is applied to each knowledge point learning record of the students to estimate their memory retention of the knowledge points at different time points. By continuously acquiring new student learning data, learner portraits, review plans, Ebbinghaus forgetting curve models, and knowledge systems are iteratively optimized.

[0037] Learning data includes: learning time preference, learning speed, accuracy of answering questions, homework completion status, knowledge level, learning habits, learning style, and interest preferences. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to further explain the embodiments of the present invention or the implementation schemes in the prior art, the mentioned drawings will be briefly introduced below. It should be clear that these drawings are only used to show certain specific implementation schemes of the present invention. They do not represent all possible forms of the present invention. For ordinary technicians in this field, more variants and expanded schematic diagrams can be derived based on these drawings without additional creative efforts.

[0039] Figure 1 It is a flowchart of one embodiment of the training method of the intelligent note-taking system based on artificial intelligence of the present invention;

[0040] Figure 2 It is a schematic diagram of the module connection flow of the second embodiment of the training method of the intelligent note-taking system based on artificial intelligence of the present invention;

[0041] Figure 3 It is a flow chart of a data acquisition module of Embodiment 3 of the training method of the intelligent note-taking system based on artificial intelligence of the present invention;

[0042] Figure 4 It is a flowchart of collecting student characteristics and establishing learner portraits in the fourth embodiment of the training method of the intelligent note-taking system based on artificial intelligence of the present invention; DETAILED DESCRIPTION

[0043] The following will describe in detail multiple embodiments of the present invention in conjunction with the accompanying drawings, in order to further illustrate its technical details and features. It should be clear that the examples listed are only exemplary in nature and are intended to illustrate rather than limit the scope of the present invention. Any other implementation methods conceived by those skilled in the art based on the core concept of the present invention without the need for creative contributions are deemed to fall within the scope of protection claimed by the present invention.

[0044] The method of the present invention can achieve efficient parsing and knowledge point extraction of multi-format file data, build a structured knowledge system, and provide students with personalized learning suggestions and review plans. When constructing a knowledge system, the prior art often uses simple keyword extraction or statistical methods based on word frequency, and fails to make full use of modern natural language processing and machine learning technologies to improve the accuracy and systematicity of knowledge point extraction. By utilizing modern artificial intelligence technology, the present invention parses file data in multiple formats and accurately extracts knowledge points, thereby avoiding the problem of incomplete information extraction in traditional note-taking systems. Secondly, the present invention improves the systematicity and relevance of the knowledge system by constructing a graph database to perform structured management of learning resources and student notes. Finally, the present invention uses a personalized recommendation algorithm, combined with students' learning conditions and characteristics, to dynamically generate a review plan to improve learning outcomes. The problems of the shortcomings of the prior art are overcome.

[0045] Embodiment 1

[0046] See also Figure 1 , Figure 1 FIG. 1 is a flow chart of one embodiment of the training method of the intelligent note-taking system based on artificial intelligence of the present invention. It should be noted that if there is substantially the same result, the method of the present invention is not limited to the training method of the intelligent note-taking system based on artificial intelligence. Figure 1 The process sequence shown is limited. Figure 1 As shown, the method comprises the following steps:

[0047] S101, extracting multi-format file data;

[0048] In this embodiment, the file formats include PDF, DOCX, and TXT files;

[0049] In this embodiment, the system supports a wide range of file formats, including PDF, DOCX, and TXT. These files contain text, pictures, and code snippets. Advanced file parsing tools Apache Tika and Tesseract are used, where Apache Tika is used to parse text and metadata, and Tesseract is used to extract text information from images, perform deep parsing of non-text data, and convert it into processable plain text and structured data. The RoBERTa model is used and fine-tuned for code snippets. First, the vocabulary of RoBERTa is expanded, which contains common keywords, function names, and variable names in programming languages. For example, if is a keyword in Python for conditional judgment, and def is a keyword for defining functions in Python. The model is fine-tuned using open source projects on GitHub so that it can understand the grammatical structure and semantic information of the code. In the fine-tuning process, CodeBERT is introduced to further enhance the model's ability to understand the code. The parsed text data is converted into a context-sensitive high-dimensional vector representation. The data is processed by a batch processor to form a batch suitable for a deep learning model, ready for subsequent training and reasoning.

[0050] S102, extracting knowledge points and key information from text data to help build a graph database;

[0051] In this embodiment, for Chinese text, the system uses Jieba Chinese word segmentation library for accurate word segmentation, and enhances the recognition ability of professional terms (such as algorithm names and technical vocabulary such as depth-first search and quick sort) by expanding Jieba's dictionary. This step is done by using the jieba.load_userdict('path_to_custom_dict.txt') function in Python to load the custom dictionary, and then calling the jieba.cut(text, cut_all=False) function to perform accurate word segmentation on the text.

[0052] For English text, the system uses the SpaCy library for word segmentation and part-of-speech tagging. By loading SpaCy's English model (en_core_web_sm), the system automatically processes the English text and extracts the word segmentation results and part-of-speech information. This step is also implemented in Python, using SpaCy's nlp = spacy.load ('en_core_web_sm') and doc = nlp (text) functions.

[0053] To evaluate the importance of each word in the text, the system calculates the TF-IDF (term frequency-inverse document frequency) value. For the Chinese and English word segmentation results, the system uses the sklearn.feature_extraction.text.TfidfVectorizer class to calculate the TF-IDF matrix. By adjusting the parameters of this class, such as max_df, min_df, and stop_words, the system can optimize the TF-IDF calculation results and thus more accurately evaluate the importance of words.

[0054] In order to further identify important concepts, algorithm names, technical terms, and named entities in the text, the system combines NER (named entity recognition) technology. For Chinese text, the system uses HanLP and the extended SpaCy for named entity recognition to identify names of people, places, organizations, and custom algorithm names and technical terms. For English text, SpaCy can already identify multiple entity types in the text, such as names of people (PERSON), places (GPE), and organizations (ORG).

[0055] Finally, the system stores all the identified information in a structured manner in the database. The database selected is Neo4j (NoSQL graph database). In the database design, the system sets the document ID, word segmentation result list, TF-IDF value mapping, and NER result list fields to facilitate subsequent query, analysis, and application.

[0056] The following are the steps for calculating TF-IDF word weights:

[0057] Step 1: Calculate the frequency of a word in the document

[0058]

[0059] Step 2: Calculate a measure of the general importance of the words.

[0060]

[0061] |D| represents the total number of documents in the corpus

[0062] {d∈D:t∈d} represents the number of documents containing word t.

[0063] (3) TF-IDF value: Multiply TF and IDF to get the TF-IDF value.

[0064] TF-IDF(t,d,D)=TF(t,d)×IDF(t,D)

[0065] When the similarity score TF-IDF(t,d,D)>0.7, the knowledge point is considered to be relevant to the question.

[0066] The knowledge point mapping matrix M is constructed by the similarity scores between relevant knowledge points and questions. ij , as the basis for calculating the mastery of knowledge points.

[0067] For each question i, according to the matching results between keywords and knowledge points, the corresponding position M in the matrix M ij Fill in the similarity score. If a knowledge point j does not match question i, then M ij =0, otherwise M ij =1.

[0068] S103, map and associate learning resources and students’ learning notes to form a structured knowledge system;

[0069] In this embodiment, Neo4j, a graph database management system, is used, and the system stores each knowledge point and key information as a node in the graph database. Each node contains attributes such as name, type (such as algorithm, concept, code example, etc.), description, and source.

[0070] For example:

[0071] Algorithm node: name, type (algorithm), description, complexity, application scenario, and implementation language.

[0072] Concept node: name, type (concept), definition, historical background, related formulas.

[0073] Code example node: code language, functional description, implementation details, and associated algorithms.

[0074] These attributes are used to support efficient retrieval and query. The association information between different knowledge points is stored as relationships in the graph database. Relationships are directional and contain attributes such as association type, weight, and hierarchical relationship. For example, an algorithm node may form a "use" relationship with a related code example node, or a concept node may form a "subordinate" relationship with other related concept nodes. These relationships play a key role in subsequent reasoning and knowledge expansion.

[0075] The system uses the RoBERTa model to calculate the similarity of texts to accurately identify the association between the note content and related knowledge points. In the similarity calculation of the model, cosine similarity is used to measure the similarity score of text pairs, and the optimal threshold of the model is set to 0.75, which can strike a balance between 90% recall and 85% precision.

[0076] Specifically, when the similarity score between the note content and the knowledge point is higher than 0.75, the system will mark the knowledge point as related to the note content and associate and map it in the user's knowledge system. For text pairs with similarity scores below the threshold, the system will further apply an analysis method based on named entity recognition (NER) to ensure that potential related knowledge points are not missed.

[0077] The system uses named entity recognition (NER) and topic modeling (LDA) to capture contextual information in the text. The output results can be fused with the similarity score of the RoBERTa model, which accounts for 70% of the weight, while NER and LDA each account for 15% to improve the final judgment accuracy to dynamically adjust the similarity threshold in different learning fields. For example, in the field of mathematics, the model sets the threshold to 0.8 to ensure strict matching, while in the field of literature, the threshold is reduced to 0.7 to tolerate greater semantic changes. In this way, the system can adapt to various types of learning content more accurately, improve the accuracy and coverage of related knowledge points, and add these related points to the graph database to further expand and improve the knowledge system. This knowledge system supports the generation of review plans and is also used to construct exercises for students' personalized needs. By analyzing the correlation between knowledge points, the system can generate targeted review materials and design exercises covering various knowledge points to help students consolidate what they have learned more effectively.

[0078] S104. Generate targeted practice test questions based on the structured knowledge system;

[0079] In this embodiment, in order to enhance students' understanding and memory of knowledge points, this method generates targeted practice test questions based on a structured knowledge system. The specific steps are as follows:

[0080] Based on the structured knowledge system, the text generation variants in the BERT series are used and fine-tuned. T5 (Text-To-Text Transfer Transformer) unifies various NLP tasks into the form of "text to text" to complete various question generation tasks. For fine-tuning, materials extracted from the book are used as question data sets, including multiple-choice questions, fill-in-the-blank questions, and short-answer questions, and the knowledge points corresponding to the questions are marked. For example, the multiple-choice question, "What is the pre-order traversal order of a binary tree? A. Left root right B. Root left and right C. Left and right root" corresponds to the knowledge point "Binary tree traversal", the fill-in-the-blank question, "In Python, the keyword used to create anonymous functions is ______." corresponds to the knowledge point "Python anonymous function", and the short-answer question, "Please explain the basic idea of ​​the quick sort algorithm." corresponds to the knowledge point "Quick sort".

[0081] The loss function is designed as follows:

[0082] Sequence generation loss: For sequence-to-sequence tasks such as title generation, cross-entropy loss is used. It measures the difference between the sequence generated by the model and the real sequence.

[0083] The real sequence is y=(y1,y2,...,y t ), the sequence generated by the model is The cross entropy loss is expressed as:

[0084]

[0085] L is the value of the cross entropy loss function, which is used to measure the difference between the sequence generated by the model and the true sequence.

[0086] T is the total length of the sequence, i.e. the number of time steps in the sequence.

[0087] t is the index of the current time step, t=1,2,...,T.

[0088] The predicted value generated by the model at time step t.

[0089] y t The true target value, that is, the expected output at time step t.

[0090] The sequence of all forecast values ​​generated before time step t.

[0091] The context or knowledge point description of the x input is the conditional input based on which the model generates the sequence.

[0092] It represents the probability that the model predicts the current time step to be the true label given the input and the generated sequence.

[0093] The steps in the formula are:

[0094] For each time step t, we calculate the model’s performance given the input x and the previously generated sequence Next, predict the output of the current time step t and the true value y t The probability of matching. Calculate the logarithmic probability at each time step. Average the logarithmic probabilities of all time steps to get the cross entropy loss value L of the entire sequence. The smaller this loss value is, the closer the sequence generated by the model is to the true sequence, and vice versa. By minimizing the cross entropy loss, the model will learn to generate outputs that match the true sequence as much as possible given the input context.

[0095] The GANs model is then used to optimize the generated questions. Specifically, the fine-tuned T5 model is used as the generator to generate questions. A discriminator is trained based on RNN to distinguish the difference between the generated questions and the real questions.

[0096] The specific contents of RNN training are as follows:

[0097] The model design includes:

[0098] Network structure: LSTM is used as the basic unit of RNN, and multiple layers are stacked to enhance feature extraction capabilities.

[0099] Input: real question sequence y=(y1,y2,...,y t ) or a sequence of questions generated by the generator

[0100] Output: A scalar D = (y) or Represents the probability that the input sequence is a real question.

[0101] The training objectives include:

[0102] Discriminator: maximize logD(y) (the probability of correct judgment of the real question) and minimize (Probability of misjudgment of generated questions).

[0103] Comprehensive objectives:

[0104] D(y) represents the probability of the discriminator correctly judging the true question sequence yyy.

[0105] Represents the sequence of questions generated by the generator.

[0106] Represents the discriminator to generate a sequence of questions probability of misjudgment.

[0107] Indicates the distribution of real question data p data The expected value under (y).

[0108] Represents the conditional distribution in the generator The expected value below.

[0109] p data (y) represents the distribution of real question data.

[0110] Indicates that the generator generates a question under a given input x distribution.

[0111] By alternately training the generator and the discriminator, the generator can generate questions that are closer and closer to the quality of real questions. The specific steps are as follows:

[0112] Step 1: Fix the discriminator and train the generator (maximize the probability that the discriminator misclassifies the questions generated by the generator as real questions, that is, minimize

[0113] Through the back-propagation algorithm, the parameters of the generator are adjusted according to the output of the discriminator on the generated questions, so that the generated questions are closer to the distribution of real questions.

[0114] The loss function of the generator is (Note: In order to be consistent with the training objectives of the discriminator, use format for description).

[0115] Step 2: Fix the generator and train the discriminator (to improve the discriminator's ability to distinguish between real questions and generated questions)

[0116] Use the real questions and the questions generated by the generator as input, calculate the output of the discriminator respectively, and adjust the parameters of the discriminator according to the output.

[0117] The loss function of the discriminator is as described above, that is,

[0118] Iteration: Repeat steps 1 and 2 until the performance of the discriminator on the validation set no longer improves significantly.

[0119] S105. Collect students’ test records and notes;

[0120] In this embodiment, by collecting students' exercise records and notes, we can better understand students' learning situation, establish learner portraits, and continuously update the operation of the module. The specific steps are as follows:

[0121] Categorize students' notes by subject, chapter, and knowledge point, for example:

[0122] Taking computer science as an example, "Computer Science" is directly used as the first-level classification and then further subdivided into computer basics, programming languages ​​and algorithms, operating systems, computer networks, database systems, software engineering, artificial intelligence and machine learning according to the chapters of the textbook. Under each chapter, it is further divided according to specific knowledge points. The following is a specific example under the "Programming Languages ​​and Algorithms" chapter:

[0123] Python programming basics: variables and data types, control structures (if-else, for, while), function definition and calling, lists, tuples, dictionaries and other data structures.

[0124] Through such classification and organization, we can clearly understand the distribution of different chapters and specific knowledge points in the discipline of computer science, which provides a basis for subsequent data cleaning, preprocessing and analysis.

[0125] After classification and sorting, clean and preprocess the data to eliminate noise, fill missing values, convert data types, and ensure data accuracy and consistency. The specific steps are as follows:

[0126] Step 1: Use a moving average algorithm to smooth the data and remove noise.

[0127] Through a time series data X = [x1, x2, ..., x n ], smoothed using a moving average with a window size of k:

[0128]

[0129] represents the i-th data point after smoothing.

[0130] k represents the window size of the moving average, which determines the The number of adjacent data points used when calculating the noise level is generally such as k = 5, that is, the window size is 5, the noise will be significantly reduced.

[0131] j represents the index used to represent the position of the data point within the window, from 0 to k-1.

[0132] Step 2: Use the z-score method to identify and handle outliers. For each data point x i , calculate the z-score:

[0133]

[0134] z i Represents the z-score of the ith data point, which is used to measure the degree of deviation of the data point from the mean of the data set.

[0135] x i represents the i-th original data point.

[0136] μ represents the mean of the data set X, which indicates the central tendency of the data.

[0137] σ represents the standard deviation of the data set X, which indicates the degree of dispersion of the data.

[0138] When |zi |>3 then x i are outliers and should be deleted or replaced.

[0139] Step 3: For missing values, you can use the mean to fill in. Suppose in the data set there is a column X = [x1, x2, ..., x n ] has missing values, use the mean μ of the column x To fill:

[0140]

[0141] μ x It represents the mean of the dataset X after removing missing values.

[0142] n represents the total size of the dataset X.

[0143] m represents the number of missing values ​​in the dataset X.

[0144] M represents an index set, indicating the location of missing values ​​in the data set.

[0145] j represents the index position where the missing value is located.

[0146] For each missing data point x i (j∈M) with μ x replace.

[0147] Step 4: Use linear interpolation to fill in missing values. Given a data point x i-1 and x i+1 Adjacent to the missing value x i , then the interpolation formula is:

[0148]

[0149] x i-1 Represents the i-1th data point, i.e., the missing value x i A previous data point.

[0150] x i+1 Represents the i+1th data point, i.e., the missing value x i One data point later.

[0151] x i Indicates missing values, which are filled using interpolation methods.

[0152] Step 5: To ensure the consistency of the data, the data is usually standardized and converted into a standard normal distribution with a mean of 0 and a standard deviation of 1. Given data x i , the standardized data is:

[0153]

[0154] It represents the i-th data point after standardization, which has the properties of mean 0 and standard deviation 1.

[0155] μ represents the mean of the data set X and is used when calculating standardization.

[0156] σ represents the standard deviation of the data set X.

[0157] Step 6: Standardize the data and use the Min-Max normalization formula to convert the data to a range between 0 and 1. You can use the minimum-maximum normalization:

[0158]

[0159] Represents the normalized i-th data point, ranging from 0 to 1.

[0160] min(X) represents the minimum value in the data set X.

[0161] max(X) represents the maximum value in the data set X.

[0162] Improve the performance of the model and the accuracy of predictions through the above six steps.

[0163] Design a database to store and manage students' answer records, including student ID, question ID, answer time, answer, correct answer, score, to help build learner portraits.

[0164] S106. Use student learning data to build learner profiles based on student characteristics;

[0165] In this embodiment, in order to better understand the learning status and needs of students, a learner portrait is established, and the mastery of each knowledge point is analyzed based on the structured knowledge system and knowledge point mapping matrix M in combination with the test results, including:

[0166] Q={q1,q2,...,q m}Question set, including m questions.

[0167] K={k1,k2,...,k m}Knowledge point set, including n knowledge points.

[0168] A={a1,a2,...,a m}Students’ answers, where a i Indicates the student's answer to question q i The answer score is usually 0 or 1 (i.e. correct or not).

[0169] ai It can also mean that students are in question q i The score percentage on , a floating point number between 0 and 1.

[0170] K={k1,k2,...,k n}Students' mastery of each knowledge point ranges from 0 to 1.

[0171] Through the question-knowledge point mapping matrix M and the answer situation A, calculate the student’s understanding of each knowledge point k j Mastery

[0172]

[0173] k j ≥0.8 indicates that students have a good understanding of knowledge point k j Good mastery.

[0174] k j <0.8 indicates that students are familiar with knowledge point k j Needs further consolidation.

[0175] k j <0.5 means that the student has a good understanding of the knowledge point k j The grasp of the content is poor and needs to be reviewed in depth.

[0176] Through multi-level feature collection, a solid foundation is provided for the construction of learner portraits. The collected student characteristics include multi-dimensional information such as knowledge level, learning habits, learning style, interest preferences, etc. First, the system uses K-means cluster analysis to preliminarily process these features, identify similar learner groups, and reduce the dimension and noise of the data. Then, based on cluster analysis, the system combines decision trees and multi-layer perceptron (MLP) models to further refine the learner portraits and enhance the modeling ability of nonlinear features. In order to ensure the accuracy of the portrait, the system uses the back propagation algorithm to optimize the parameters of the MLP model, and comprehensively evaluates the model through cross-validation to prevent overfitting and ensure its accuracy and robustness in identifying and classifying learner characteristics. In addition, the system uses the DAX function in PowerBI to create dynamic labels, which can be automatically updated according to different conditions, providing a flexible and intuitive analysis perspective. Through visualization tools such as word cloud diagrams and scatter plots, learner portraits will display the multi-dimensional characteristics of students in an intuitive way, helping teachers and education managers to quickly understand students' personalized needs and thus develop more effective teaching strategies. The specific contents are as follows:

[0177] Step 1: Standardize the data and use the Min-Max normalization formula to convert the data to a range between 0 and 1. You can use the minimum-maximum normalization:

[0178]

[0179] x represents the original feature value, which represents a specific data point in the student feature, such as a student's test score or homework score.

[0180] It indicates that the normalized eigenvalues ​​are between 0 and 1, which is used to eliminate the dimensional differences between features.

[0181] min(X) represents the minimum value of the feature among all student feature data.

[0182] max(X) represents the maximum value of this feature among all student feature data.

[0183] Step 2: Use the K-means algorithm to cluster student features and divide students into different groups of learners. The goal of K-means clustering is to minimize the sum of the squared distances from each data point to the center of the cluster to which it belongs.

[0184] The objective function of means is:

[0185]

[0186] μ i Represents the center of the ith cluster. Represents the feature mean vector of the ith student group.

[0187] x j Represents the feature vector of students belonging to the i-th cluster. Represents the feature data of a student in the i-th student group.

[0188] C i represents the i-th cluster. represents the i-th student group obtained after K-means algorithm clustering.

[0189] ||x j -μ j || represents the Euclidean distance between the student's feature vector and the cluster center. represents the sum of the squared distances between a student's feature vector and the cluster center to which it belongs, and is used to measure the similarity between the student and the cluster center.

[0190] Step 3: Based on cluster analysis, use the multi-layer perceptron (MLP) model to construct and refine the learner profile. The structure of the MLP model is as follows:

[0191] y=f(W2·σ(W1·x+b1)+b2)

[0192] x represents the input feature vector. represents the student's characteristic data, such as knowledge level, learning habits, learning style, etc.

[0193] W1 represents the first layer weight matrix. It represents the connection weight from the input layer to the hidden layer, which is used to linearly transform the input features.

[0194] b1 represents the first layer bias vector. It is used to add an offset between the input layer and the hidden layer to increase the flexibility of the model.

[0195] σ represents the activation function. The commonly used ReLU (Rectified Linear Unit) activation function is used to increase the nonlinear ability of the model. The formula is:

[0196] σ(z)=max(0,z)

[0197] W2 represents the second layer weight matrix. It represents the connection weight from the hidden layer to the output layer, which is used to linearly transform the output of the hidden layer.

[0198] b2 represents the second layer bias vector. It is used to add an offset between the hidden layer and the output layer.

[0199] f represents the activation function of the output layer. The commonly used softmax function is used to convert the results of the output layer into probability distribution. The formula is:

[0200]

[0201] z i Represents the value of the i-th node in the output layer.

[0202] Step 4: Use the back propagation algorithm to optimize the parameters of the MLP model. The loss function is selected as the cross entropy loss:

[0203]

[0204] Represents the loss function value. Represents the difference between the model prediction result and the true label, and the goal is to minimize this value.

[0205] N represents the number of samples. represents the total number of samples in the training data.

[0206] C represents the number of categories. It represents the total number of categories in the classification task, such as different learner groups or labels.

[0207] y ij Represents the jth class true label of the i-th sample. Indicates whether sample i belongs to category j, if yes, it is 1, otherwise it is 0.

[0208] Represents the predicted probability of the jth class for the i-th sample. Represents the probability that the model predicts that sample i belongs to class j.

[0209] Step 5: Evaluate the performance of the model through cross-validation. Using k-fold cross-validation, divide the data into k subsets, train the model with k-1 subsets in turn, and test the model with the remaining subset. The final evaluation indicators (accuracy, AUC) take the average of all test results.

[0210] Step 6: Use DAX functions in Power BI to create dynamic labels for each student. Example of DAX functions:

[0211] StudentTag=IF(Student[Score]>80,"High Performer","NeedsImprovement")

[0212] Student[Score] indicates the student's score. It indicates the score of a student in a test or assignment.

[0213] "High Performer" is the label name. It indicates the label for students whose scores are higher than 80.

[0214] "Needs Improvement" indicates the label name. It indicates the label for students whose scores are less than or equal to 80.

[0215] Step 7: Use word cloud and scatter plots to display learner portraits. For example, word cloud maps are used to display students’ interest preference keywords, and scatter plots are used to display students’ mastery of different knowledge points and their learning progress.

[0216] S107. Generate a review plan using the Ebbinghaus forgetting curve model;

[0217] In order to solve the problem of students forgetting knowledge points, this embodiment uses the Ebbinghaus forgetting curve model to generate a review plan. The specific steps are as follows:

[0218] For each knowledge point, the exponential decay model of the Ebbinghaus forgetting curve is used to calculate the forgetting rate based on the time after learning (in days). Using the formula:

[0219] R(t)=-e -λt

[0220] R(t) represents the forgetting rate after time t (usually between 0 and 1, 0 means complete memory and 1 means complete forgetting).

[0221] λ represents the forgetting constant (the larger the λ value, the faster the forgetting; conversely, the slower the forgetting).

[0222] t represents the time from the beginning of learning (in days). The forgetting rate is applied to each knowledge point learning record of the students to estimate their memory retention of the knowledge points at different time points.

[0223] To determine the key review points, use the following formula:

[0224] T n =T0+λ·log(n)

[0225] T n Indicates the time of the nth review node (in days).

[0226] T0 represents the initial review time point, which is usually a fixed time after the first study.

[0227] λ represents the time growth factor, which is used to control the interval between review time points and is adjusted according to the student's learning progress and forgetting speed.

[0228] n represents the serial number of the review node.

[0229] Based on the principle of Ebbinghaus's forgetting curve, the review interval is gradually extended to improve memory effect. By calculating each key review node, a personalized review plan is generated.

[0230] S108. Continuously update learner portraits, review plans, Ebbinghaus forgetting curve models and knowledge systems based on student data;

[0231] In this embodiment, a continuous learning method is designed to more accurately understand the learning status and needs of students. This method can regularly update the model training corpus by continuously acquiring students' learning data to keep the learner profile, review plan, Ebbinghaus forgetting curve model and knowledge system up to date.

[0232] Embodiment 2

[0233] See also Figure 2 , a training method for an intelligent note-taking system based on artificial intelligence, comprising:

[0234] The data collection and preprocessing module, knowledge extraction and annotation module, knowledge system module, question generation module, learner portrait module, and continuous update module are connected in sequence;

[0235] The data acquisition and preprocessing module extracts multi-format file data and parses the various file formats uploaded by students to realize the basis of system intelligence;

[0236] The knowledge extraction and annotation model uses Jieba Chinese word segmentation library and SpacyNLP toolkit for word segmentation, and then combines TF-IDF and NER to extract knowledge points and key information in text data to build a graph database;

[0237] The knowledge system module maps and associates learning resources and student learning notes based on a graph database, stores them in the form of nodes and edges, and forms a structured knowledge system;

[0238] The question generation module generates targeted practice test questions based on the structured knowledge system and obtains students' question-taking records;

[0239] The learner portrait module optimizes the mapped knowledge system based on the student's question record and note content, and establishes a learner portrait;

[0240] The continuous updating module collects student learning data on a daily basis to optimize learner portraits and knowledge systems;

[0241] In this embodiment, knowledge points can be efficiently extracted from a variety of file formats (such as PDF, DOCX, TXT, etc.), and a structured knowledge system can be constructed. The system can achieve effective mapping and association between learning resources and student notes, and generate questions at the same time, so as to provide students with a dynamic review plan, solve the problems of incomplete knowledge point extraction, poor correlation of learning resources, and poor personalized recommendation effect in the prior art, and the present invention provides a personalized learning experience by utilizing advanced machine learning algorithms. The system can dynamically adjust the learning plan based on the real-time feedback of students and the learning data collected daily to ensure that the learning plan matches the actual needs of students. In the specific implementation process, the present invention adopts high-quality and diverse training data sets, covering student data of different backgrounds, disciplines and learning styles. The use of this data set significantly improves the accuracy of the artificial intelligence model and reduces the deviation problem caused by the use of limited or biased data in traditional technology. The present invention can provide more accurate and fairer teaching effects, which has significant advantages over the prior art. In addition, through the application of reinforcement learning algorithms, the present system can continuously optimize learning strategies. The system will continuously monitor the learning progress and feedback of students, and adjust the teaching content and review plan according to real-time data. This dynamic adjustment not only improves students' learning efficiency, but also enables more accurate delivery and consolidation of knowledge points in personalized teaching.

[0242] Embodiment 3

[0243] See Figure 3 , a data acquisition module flow chart of a training method for an intelligent note-taking system based on artificial intelligence, it should be noted that if there are substantially the same results, the method of the present invention is not based on Figure 1 The process sequence shown is limited. The method comprises the following steps:

[0244] In this embodiment, the system supports a wide range of file formats, including PDF, DOCX, and TXT. These files contain text, pictures, and code snippets. Advanced file parsing tools Apache Tika and Tesseract are used, where Apache Tika is used to parse text and metadata, and Tesseract is used to extract text information from images, perform deep parsing of non-text data, and convert it into processable plain text and structured data. The RoBERTa model is used and fine-tuned for code snippets. First, the vocabulary of RoBERTa is expanded, which contains common keywords, function names, and variable names in programming languages. For example, if is a keyword in Python for conditional judgment, and def is a keyword for defining functions in Python. The model is fine-tuned using open source projects on GitHub so that it can understand the grammatical structure and semantic information of the code. In the fine-tuning process, CodeBERT is introduced to further enhance the model's ability to understand the code. The parsed text data is converted into a context-sensitive high-dimensional vector representation. The data is processed by a batch processor to form a batch suitable for a deep learning model, ready for subsequent training and reasoning.

[0245] Embodiment 4

[0246] See Figure 4 , a flow chart of collecting student characteristics and establishing learner portraits in a training method of an intelligent note-taking system based on artificial intelligence. It should be noted that if there are substantially the same results, the method of the present invention is not based on Figure 1 In this embodiment, in order to better understand the learning status and needs of students, by collecting students' exercise records and notes, we can better understand students' learning situation, establish learner portraits and continuously update the operation of the module.

[0247] Figure 4 Corresponding to the implementation steps S105 and S106 in the training method of the intelligent note-taking system based on artificial intelligence in one of the above modules and embodiments, the specific principles and usage methods of each module will not be elaborated here.

[0248] The advantages of the above four examples include the following five points:

[0249] 1. The present invention can support data parsing of multiple file formats (PDF, DOCX, TXT), and use advanced file parsing tools to deeply parse non-text data into processable plain text and structured data. This capability avoids the problem of incomplete information extraction in traditional note-taking systems and improves the efficiency and accuracy of information processing.

[0250] 2. Use Jieba Chinese word segmentation library and SpacyNLP toolkit for word segmentation, combined with TF-IDF and named entity recognition (NER) technology, to efficiently extract high-weight words and named entities from text data, such as important concepts, algorithms, code examples, etc. This precise extraction ensures the comprehensiveness and accuracy of knowledge points.

[0251] 3. Based on the structured knowledge system, the pre-trained model (T5) in the Transformers library is used to generate preliminary question content, and then optimized through the GANs model. There are various types of questions, including multiple-choice questions, fill-in-the-blank questions, short-answer questions, etc., which help students consolidate the knowledge they have learned. At the same time, the management of students' answer records provides data support for subsequent analysis and optimization.

[0252] 4. By analyzing students' test records and notes, combining students' use of Jieba Chinese word segmentation library and SpacyNLP toolkit for word segmentation, and combining TF-IDF and named entity recognition (NER) technology, high-weighted words and named entities can be efficiently extracted from text data, such as important concepts, algorithms, code examples, etc. This precise extraction ensures the comprehensiveness and accuracy of knowledge points.

[0253] The learning habits and characteristics (such as knowledge level, learning habits, learning style, and interest preferences) of students are used to build learner portraits. Based on the Ebbinghaus forgetting curve model, the review plan is dynamically generated and adjusted to optimize the review interval, thereby improving learning results. This personalized recommendation algorithm can provide students with more accurate and effective learning suggestions.

[0254] 5. The present invention continuously collects students' learning data (learning time preference, learning speed, correct answer rate, etc.), updates and optimizes learner portraits and knowledge systems every day. At the same time, it uses reinforcement learning algorithms to continuously optimize review plans and Ebbinghaus forgetting curve models to ensure that the learning plan matches the actual needs of students.

Claims

1. A training method for an intelligent note-taking system based on artificial intelligence, characterized in that: include: Extract multi-format file data and parse various file formats uploaded by students to realize the basis of system intelligence; Based on the file data, Jieba Chinese word segmentation library and SpacyNLP toolkit are used for word segmentation, and TF-IDF and NER are combined to extract knowledge points and key information in the text data and build a graph database; Based on the graph database, learning resources and student learning notes are mapped and associated, stored in the form of nodes and edges, forming a structured knowledge system; Generate targeted practice test questions based on the structured knowledge system and obtain students' test-taking records; Collect students' exercise records and notes, and build learner portraits; Collect student learning data to optimize learner portraits, review plans, Ebbinghaus forgetting curve models and knowledge systems.

2. The training method of the intelligent note-taking system based on artificial intelligence according to claim 1, characterized in that: include: The data collection and preprocessing module, knowledge extraction and annotation module, knowledge system module, question generation module, learner portrait module, and continuous update module are connected in sequence; The data acquisition and preprocessing module extracts multi-format file data and parses the various file formats uploaded by students to realize the basis of system intelligence; The knowledge extraction and annotation model uses Jieba Chinese word segmentation library and SpacyNLP toolkit for word segmentation, and then combines TF-IDF and NER to extract knowledge points and key information in text data to build a graph database; The knowledge system module maps and associates learning resources and student learning notes based on a graph database, stores them in the form of nodes and edges, and forms a structured knowledge system; The question generation module generates targeted practice test questions based on the structured knowledge system and obtains students' question-taking records; The learner portrait module optimizes the mapped knowledge system based on the student's question record and note content, and establishes a learner portrait; The continuous updating module collects student learning data for optimizing learner portraits and knowledge systems.

3. The training method of the intelligent note-taking system based on artificial intelligence according to claim 2, characterized in that: The data acquisition and preprocessing module comprises: Extract text, pictures, icons and other data from student notes, authoritative subject materials, and academic papers, including files in PDF, DOCX, TXT and other formats, parse the files into plain text data, and store them in categories.

4. The training method of the intelligent note-taking system based on artificial intelligence according to claim 2, characterized in that: The knowledge extraction and annotation module includes: Jieba Chinese word segmentation library and Spacy NLP toolkit are used to segment the extracted plain text data, identify word boundaries, perform part-of-speech tagging, calculate the weight of each word in the text, identify high-weight words, and then combine TF-IDF and NER to identify important concepts, algorithms, code examples and named entities in these words in the book, and store them in the database in a structured way, which is the basis for building a graph database.

5. The training method of the intelligent note-taking system based on artificial intelligence according to claim 2, characterized in that: The knowledge system module includes: The processed plain text learning data is classified, associated and indexed, stored in the form of nodes and edges, and a graph database is constructed. The similarity between texts is then calculated, and the RoBERTa model is used to calculate the similarity. A threshold is set to find the relevant points between knowledge points and notes, and add them to the database to form a knowledge system, which is the basis for generating targeted test questions.

6. The training method of the intelligent note-taking system based on artificial intelligence according to claim 2, characterized in that: The topic generation module comprises: Based on the structured knowledge system, the pre-trained model in the Transformers library is used to generate preliminary question content. The generated question types can include multiple-choice questions, fill-in-the-blank questions, short-answer questions, etc. The GANs model is used to optimize the generated questions, store and manage students' answer records, and help build learner portraits.

7. The training method of the intelligent note-taking system based on artificial intelligence according to claim 2, characterized in that: The learner portrait module includes: The students' note data are classified and organized according to subjects, chapters, and knowledge points, and combined with the test-taking situation. According to the structured knowledge system, the mastery of each knowledge point is analyzed. At the same time, student characteristics are collected as the basis for establishing learner portraits. Learner characteristics collected include knowledge level, learning habits, learning style, and interest preferences. Learners are clustered according to similar characteristics to form different learner groups. A learner portrait model is built using a multi-layer perceptron constructed using decision trees and neural networks, and then combined with K-means to achieve automatic recognition and classification of learner characteristics.

8. The training method of the intelligent note-taking system based on artificial intelligence according to claim 2, characterized in that: The continuous updating module comprises: Students’ learning data are collected every day, and the learner portrait, review plan, Ebbinghaus forgetting curve model and knowledge system are updated every 15 days based on the students’ learning data. The learning data include: learning time preference, learning speed, accuracy of answering questions, homework completion status, knowledge level, learning habits, learning style and interest preferences.

9. The training method of the intelligent note-taking system based on artificial intelligence according to claim 8, characterized in that The Ebbinghaus Forgetting Curve Model generates a review plan and iterates the learner profile, review plan, Ebbinghaus Forgetting Curve Model and knowledge system through student learning data, including: Based on the learning records in the learning feature and the Ebbinghaus forgetting curve, the forgetting rate of each knowledge point is calculated. According to students' feedback and new student learning data, the learner portrait, review plan, Ebbinghaus forgetting curve model and knowledge system are optimized and iterated.

Images