An AI online education intelligent question-answering information processing method

Abstract

The present invention relates to an AI online education intelligent question-answering information processing method, comprising the following steps: S1: collecting teaching materials, preprocessing and annotating, and forming a labeled text data set; S2: constructing a question-answering model based on XLNet; S3: introducing a specific context to optimize the question-answering model for a specific subject or topic, and obtaining an optimized question-answering model; S4: considering the age group hierarchical design based on the labeled text data set, and constructing an auxiliary understanding information knowledge map; S5: the user inputs a query, and obtains the best matching answer based on the optimized question-answering model; then, according to the user age group and the best matching answer, the auxiliary understanding information is obtained from the knowledge map to obtain a complete answer; S6: adjusting the content and difficulty provided through user interaction data, and recommending relevant field knowledge based on user behavior and preference. The present invention not only provides the direct answer required by the user, but also provides supplementary background information to enhance understanding, and presents its needs according to user characteristics.

Description

Technical Field

[0001] The present invention relates to the field of intelligent question and answer, and in particular to an AI online education intelligent question and answer information processing method. Background Art

[0002] Globally, education is undergoing a digital transformation, driven by the popularity of the Internet, the development of online learning platforms, and the digitization of educational resources. The traditional education model faces increasing challenges due to its limitations, such as the inability to provide personalized learning paths and limited interaction between teachers and students. In order to address these issues, online education platforms need to introduce intelligent solutions to support a more adaptive and interactive learning experience.

[0003] With the rapid development of AI and NLP technologies, especially the successful application of deep learning models in natural language processing tasks, online education has ushered in a huge opportunity for innovation. The design goal of the intelligent question-answering system is to extract valuable information from massive educational data by combining AI technology to provide users with real-time and accurate help. Summary of the invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide an AI online education intelligent question and answer information processing method, which can understand the user's question intention and key entities, combine the question and answer model with the knowledge graph, and not only provide the user with the direct answer required, but also provide supplementary background information to enhance understanding, and present their needs in a personalized way according to user characteristics.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] An AI online education intelligent question-answering information processing method comprises the following steps:

[0007] S1: Collect teaching materials from educational platforms, open courses, textbooks, and academic papers, and perform preprocessing and annotation to form a labeled text dataset;

[0008] S2: Build a question-answering model based on XLNet, use the labeled data set to train the model, and perform parameter tuning to improve the accuracy of question-answer matching;

[0009] S3: Introduce specific contexts to optimize the question-answering model for specific subjects or topics to obtain more accurate content understanding and an optimized question-answering model;

[0010] S4: Based on the labeled text dataset, consider the age-group stratification design and build a knowledge graph to assist in understanding information; and link knowledge in different subject areas by building shared nodes;

[0011] S5: The user enters a query, and NLU analyzes the user's question, extracts key intents and entities, and obtains the best matching answer based on the optimized question-answering model. Then, based on the user's age group and the best matching answer, auxiliary understanding information is obtained from the knowledge graph to get a complete answer.

[0012] S6: Through user interaction data, automatically adjust the content and difficulty provided, and recommend related areas based on user behavior and preferences, thereby increasing user enthusiasm and engagement.

[0013] Further, preprocessing and annotation are as follows:

[0014] Set a standard data set for validation, evaluate and improve the reliability of the data set through cross-validation, and filter out noise and redundant information;

[0015] Use semi-supervised learning strategies and data augmentation techniques to expand the coverage of small-scale high-quality data;

[0016] Use supervised NLP tools to automatically annotate samples, and use crowdsourcing platforms to verify the auxiliary automatic annotation results. Based on user feedback and verification results, a labeled text dataset is formed.

[0017] Furthermore, machine learning and NLP techniques are used to automatically filter out noise and redundant information, as follows:

[0018] Use regular expressions and pattern recognition methods to remove redundant information from HTML tags, scripts, and references; use language detection algorithms to ensure that all input data is in the target language; use text normalization techniques to clean and standardize text;

[0019] Use natural language processing technology to automatically detect and filter noise and redundant information in the data set, and apply TF-IDF to evaluate the importance of vocabulary:

[0020] ;

[0021] Among them, D represents the entire document set; represents the frequency of word t appearing in document d; is the inverse document frequency; The value is used to measure the importance of word t in document d;

[0022] Set the upper and lower thresholds of the TF-IDF value to determine the words to keep:

[0023] ;

[0024] in, and are the upper and lower thresholds set;

[0025] Build a preprocessing pipeline for text cleaning, including removing stop words, punctuation, and performing lemmatization.

[0026] Furthermore, we built a question-answering model based on XLNet, trained the model using annotated datasets, and optimized its parameters to improve the accuracy of question-answer matching, as follows:

[0027] Based on the labeled dataset, where each question has a corresponding text paragraph and the correct answer, the data is divided into training set, validation set and test set;

[0028] Initialize the XLNet model as the base model for the question-answering task. The input consists of questions and passages. Concatenate them into a sequence, retaining separators and position information to help the model understand the structure.

[0029] During the training process, we use labeled question-answer pairs to fine-tune XLNet. The loss function uses cross entropy loss and we train the predicted start and end positions separately.

[0030] Use the Adam optimizer to update the weights, set the initial learning rate, which is set to 5e-5 in this embodiment, and use the learning rate scheduler to adjust the optimization process;

[0031] After training is completed, the best model and its configuration are saved to obtain the trained question-answering model, which is then converted to ONNX format and optimized to achieve efficient cross-platform reasoning.

[0032] Furthermore, the XLNet model is as follows:

[0033] The input sequence x has T elements. XLNet considers all permutations z of the sequence based on the dynamics of permutations.

[0034] XLNet considers the different possibilities of previous words and subsequent permutations when calculating the hidden state of each word;

[0035] XLNet retains the basic architecture of Transformer and adjusts the self-attention mechanism to handle sequence arrangement and two-stream mechanism, including content stream and query stream:

[0036]

[0037] in, For content flow, To query the stream; are all the elements before the tth element in the permuted sequence; Transformer uses the self-attention mechanism to capture the dependencies between elements in the sequence; MASK represents a placeholder; is the element to be processed currently;

[0038] Using the interpolation objective function, combined with autoregressive generation and a BERT-style masked language model, not only are the tokens masked during training, but the generative order is also scheduled. The objective function is as follows:

[0039] ;

[0040] in, are model parameters, Is the arrangement All elements before t in ; A sequence of a specific arrangement; To arrange The tth element obtained after is the probability of generating the tth element of the current sequence under a given context given the model parameters θ.

[0041] Furthermore, we introduce specific contexts to optimize the question-answering model for specific subjects or topics to obtain more accurate content understanding, as follows:

[0042] Collect specialized corpora for specific disciplines or topics, including textbooks, papers, and industry reports;

[0043] Use word embedding models to learn distributed representations of domain-specific vocabulary, use topic models to identify implicit topics in text, and capture common terms and concepts in the field;

[0044] Select domain-specific words and phrases from topic models and word vector analysis to build a general vocabulary and specific concept set;

[0045] Construct domain-specific keywords and phrases into contextual prompt phrases and insert them into the input text to strengthen the model's understanding of the domain and inject contextual prompts at the beginning or appropriate position of the input sequence;

[0046] In the domain adaptation fine-tuning stage, data with domain labels are used for further fine-tuning; the loss function L is used in fine-tuning domain Optimize model parameters:

[0047] ;

[0048] Among them, D domain is the domain fine-tuning dataset, x,y are the input and target output respectively;

[0049] By introducing domain feature word weights in attention calculation, the model's attention to important concepts is enhanced:

[0050] ;

[0051] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively; k is the dimension of the key vector; B is the domain feature bias matrix.

[0052] Furthermore, based on the labeled text dataset, a hierarchical design based on age groups is considered to construct a knowledge graph to assist in understanding information; and knowledge in different subject areas is linked through shared nodes, as follows:

[0053] Layer the teaching materials by age group and create knowledge nodes at different levels. Each node contains knowledge points and background information suitable for that age group.

[0054] Use named entity recognition and relation extraction to identify entities and relations in text; use the tool spaCy for automated information extraction;

[0055] Use RDF to represent and store knowledge graphs; nodes represent knowledge entities, and edges represent relationships between entities; and introduce age group labels into the knowledge graph to achieve hierarchical access. Nodes and edges carry metadata to indicate applicable age groups and related subject areas;

[0056] The content of each document and subject is represented as a vector based on topic distribution. The vectorized topic representation measures the similarity between different documents or subjects through cosine similarity. Graph edges are created between the identified nodes with similarities greater than the threshold, and interdisciplinary links are constructed to connect knowledge from different disciplines through graph edges.

[0057] Furthermore, S5 is specifically:

[0058] S51: Clean the text input by the user, including removing noise characters and standardizing the input, using NLU to parse the text and extract the intent and entities in the query;

[0059] S52: Retrieve and match answers based on the optimized question-answering model, and search the corpus to find answers based on the extracted intents and entities;

[0060] S53: Use the user registration information or obtain the user's age group through interaction, and based on the answer, use the SPARQL graph database query language to extract relevant background information and explanations from the auxiliary understanding information knowledge graph, and adjust it according to the user's age to provide explanations of corresponding understanding levels.

[0061] Furthermore, S6 is specifically:

[0062] S61: Record each user interaction, including query content, query frequency, selected topic, browsing time, and answer satisfaction; monitor user clicks, reading time, and answer feedback on different content;

[0063] S62: Use K-means algorithm to cluster user behavior data, identify different preferences and behavior patterns, and use moving average method to predict the trend of user behavior data, the changing trend of subject orientation or knowledge depth;

[0064] S63: Based on user preferences and behavior patterns, a path search algorithm is used to explore the knowledge graph, extract knowledge nodes related to user interests from the knowledge graph, and expand from knowledge nodes related to user interests to adjacent related fields based on the constructed interdisciplinary links.

[0065] Furthermore, the path search algorithm uses the A* algorithm, as follows:

[0066] (1) Based on user preferences and behavior patterns, select nodes related to the identified user interests from the knowledge graph as starting nodes;

[0067] (2) Define the weight of the edge based on the strength of the relationship between nodes, cognitive difficulty, and learning preference;

[0068] (3) Select the similarity between the node and the target knowledge as the heuristic evaluation criterion and define the heuristic function h(n):

[0069] h(n)=Sim(n,Goal);

[0070] Among them, n is the node, Goal is the target knowledge;

[0071] Preferably, similarity calculation can be based on topic tags, content relevance, or domain knowledge approximation;

[0072] (4) After A* explores relevant nodes, it uses interdisciplinary links to expand to adjacent related disciplines;

[0073] (5) Adjust the target node set of the heuristic search algorithm according to the interdisciplinary relationship chain, and adjust the weight of the interdisciplinary edge according to the user's knowledge depth and preference;

[0074] (6) Based on the results of path search, relevant knowledge nodes and content are recommended to users, providing in-depth exploration beyond current interests.

[0075] The present invention has the following beneficial effects:

[0076] 1. The present invention can understand the user's question intention and key entities, and combines the question-answering model with the knowledge graph to not only provide the direct answer required by the user, but also provide supplementary background information to enhance understanding, and present the user's needs in a personalized manner according to the user's characteristics;

[0077] 2. The present invention builds a question-answering model based on XLNet, uses annotated data sets to train the model, builds and annotates high-quality domain-specific and stylized data sets, and fine-tunes the question-answering model with personalization and style as the goal, so as to provide more accurate answers that are suitable for different user needs in terms of content and style; and combines the use of the Adam optimizer and the linearly decayed learning rate scheduler to effectively improve the training efficiency and final performance of the question-answering model based on XLNet;

[0078] 3. The present invention can not only parse and answer user queries, but also provide users with additional understanding context by dynamically calling knowledge graphs. It is especially suitable for users of different age groups, ensuring that users get accurate answers and achieve the best match in content and acceptance, which helps to improve users' learning effects and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0080] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0081] refer to Figure 1 In this embodiment, an AI online education intelligent question-answering information processing method is provided, comprising the following steps:

[0082] S1: Collect teaching materials from educational platforms, open courses, textbooks, and academic papers, and perform preprocessing and annotation to form a labeled text dataset;

[0083] S2: Build a question-answering model based on XLNet, use the labeled data set to train the model, and perform parameter tuning to improve the accuracy of question-answer matching;

[0084] S3: Introduce specific contexts to optimize the question-answering model for specific subjects or topics to obtain more accurate content understanding and an optimized question-answering model;

[0085] S4: Based on the labeled text dataset, consider the age-group hierarchical design and build a knowledge graph to assist in understanding information; and link knowledge in different subject areas through shared nodes to help users conduct interdisciplinary learning and provide richer background information;

[0086] S5: The user enters a query, and NLU analyzes the user's question, extracts key intents and entities, and obtains the best matching answer based on the optimized question-answering model. Then, based on the user's age group and the best matching answer, auxiliary understanding information is obtained from the knowledge graph to get a complete answer.

[0087] S6: Through user interaction data, automatically adjust the content and difficulty provided, and recommend related areas based on user behavior and preferences, thereby increasing user enthusiasm and engagement.

[0088] In this embodiment, preferably, the preprocessing and labeling are as follows:

[0089] Set a standard data set for validation, evaluate and improve the reliability of the data set through cross-validation, and filter out noise and redundant information;

[0090] Use semi-supervised learning strategies and data augmentation techniques to expand the coverage of small-scale high-quality data;

[0091] Use supervised NLP tools to automatically annotate samples, and use crowdsourcing platforms to verify the automatic annotation results to improve the accuracy of annotation. Based on user feedback and verification results, a labeled text dataset is formed.

[0092] In this embodiment, preferably, machine learning and NLP technology are used to automatically filter noise and redundant information, as follows:

[0093] Use regular expressions and pattern recognition methods to remove redundant information from HTML tags, scripts, and references; use language detection algorithms (such as the LangDetect library) to ensure that all input data is in the target language; use text normalization techniques to clean and standardize text;

[0094] Use natural language processing technology to automatically detect and filter noise and redundant information in the data set, apply TF-IDF (term frequency-inverse document frequency) to evaluate the importance of vocabulary, and filter out abnormally frequent or unimportant vocabulary:

[0095] ;

[0096] Among them, D represents the entire document set; represents the frequency of word t appearing in document d; is the inverse document frequency; The value is used to measure the importance of word t in document d;

[0097] Set the upper and lower thresholds of the TF-IDF value to determine the words to keep:

[0098] ;

[0099] in, and are the upper and lower thresholds set;

[0100] Build a preprocessing pipeline for text cleaning, including removing stop words, punctuation, and lemmatization (using NLTK or SpaCy).

[0101] In this embodiment, preferably, a question-answering model is constructed based on XLNet, the model is trained using a labeled data set, and parameter tuning is performed to improve the accuracy of question-answer matching, as follows:

[0102] Based on annotated datasets, such as SQuAD (Stanford Question Answering Dataset), where each question has a corresponding text paragraph and the correct answer, the data is divided into training, validation and test sets;

[0103] Initialize the XLNet model as the base model for the question-answering task. The input consists of questions and passages. Concatenate them into a sequence, retaining separators and position information to help the model understand the structure.

[0104] During the training process, we use labeled question-answer pairs to fine-tune XLNet. The loss function uses cross entropy loss and we train the predicted start and end positions separately.

[0105] Use the Adam optimizer to update the weights, set the initial learning rate, which is set to 5e-5 in this embodiment, and use a learning rate scheduler (such as linear decay) to adjust the optimization process;

[0106] After training is completed, the best model and its configuration are saved to obtain the trained question-answering model, which is then converted to ONNX format and optimized to achieve efficient cross-platform reasoning.

[0107] In this embodiment, the preferred XLNet model is as follows:

[0108] The input sequence x has T elements. XLNet considers all permutations z of the sequence based on the dynamics of permutations.

[0109] XLNet considers the different possibilities of previous words and subsequent permutations when calculating the hidden state of each word;

[0110] XLNet retains the basic architecture of Transformer and adjusts the self-attention mechanism to handle sequence arrangement and two-stream mechanism, including content stream and query stream:

[0111]

[0112] in, For content flow, To query the stream; are all the elements before the tth element in the permuted sequence; Transformer uses the self-attention mechanism to capture the dependencies between elements in the sequence; MASK represents a placeholder; is the element to be processed currently;

[0113] Using the interpolation objective function, combined with autoregressive generation and a BERT-style masked language model, not only are the tokens masked during training, but the generative order is also scheduled. The objective function is as follows:

[0114] ;

[0115] in, are model parameters, Is the arrangement All elements before t in ; A sequence of a specific arrangement; To arrange The tth element obtained after is the probability of generating the tth element of the current sequence under a given context given the model parameters θ.

[0116] In this embodiment, a specific context is introduced for a specific subject or topic to optimize the question-answering model to obtain a more accurate understanding of the content, as follows:

[0117] Collect specialized corpora for specific disciplines or topics (e.g., medicine, law, history), including textbooks, papers, and industry reports;

[0118] Use word vector models (such as Word2Vec, GloVe) to learn the distributed representation of domain-specific vocabulary, and use topic models (such as LDA) to identify implicit topics in text and capture common terms and concepts in the field;

[0119] Select domain-specific words and phrases from topic models and word vector analysis to build a general vocabulary and specific concept set;

[0120] Domain-specific keywords and phrases are constructed into contextual prompt phrases and inserted into the input text to strengthen the model's understanding of the domain, and contextual prompts are injected at the beginning or appropriate position of the input sequence; for example, for medical questions, the prompt "Medical context:..." can be added.

[0121] In the domain adaptation fine-tuning stage, data with domain labels are used for further fine-tuning; the loss function L is used in fine-tuning domain Optimize model parameters:

[0122] ;

[0123] Among them, D domain is the domain fine-tuning dataset, x,y are the input and target output respectively;

[0124] By introducing domain feature word weights in attention calculation, the model's attention to important concepts is enhanced:

[0125] ;

[0126] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively; k is the dimension of the key vector; B is the domain feature bias matrix.

[0127] In this embodiment, based on the labeled text data set, the age group is considered to be hierarchical, and a knowledge graph of auxiliary understanding information is constructed; and the knowledge of different subject areas is linked through shared nodes, as follows:

[0128] Layer the teaching materials by age group and create knowledge nodes at different levels. Each node contains knowledge points and background information suitable for that age group.

[0129] Use named entity recognition (NER) and relation extraction to identify entities and relations in text; use the tool spaCy for automated information extraction;

[0130] Use RDF to represent and store knowledge graphs; nodes represent knowledge entities, and edges represent relationships between entities; and introduce age group labels into the knowledge graph to achieve hierarchical access. Nodes and edges carry metadata to indicate applicable age groups and related subject areas;

[0131] The content of each document and subject is represented as a vector based on topic distribution. The vectorized topic representation measures the similarity between different documents or subjects through cosine similarity. Graph edges are created between the identified nodes with similarities greater than the threshold, and interdisciplinary links are constructed to connect knowledge from different disciplines through graph edges.

[0132] In this embodiment, S5 is specifically:

[0133] S51: Clean the text input by the user, including removing noise characters and standardizing the input, using NLU to parse the text and extract the intent and entities in the query;

[0134] S52: Retrieve and match answers based on the optimized question-answering model, and search the corpus to find answers based on the extracted intents and entities;

[0135] S53: Use the user registration information or obtain the user's age group through interaction, and based on the answer, use the SPARQL graph database query language to extract relevant background information and explanations from the auxiliary understanding information knowledge graph, and adjust it according to the user's age to provide explanations of corresponding understanding levels.

[0136] In this embodiment, S6 is specifically:

[0137] S61: Record each user interaction, including query content, query frequency, selected topic, browsing time, and answer satisfaction; monitor user clicks, reading time, and answer feedback on different content;

[0138] S62: Use K-means algorithm to cluster user behavior data, identify different preferences and behavior patterns, and use moving average method to predict the trend of user behavior data, the changing trend of subject orientation or knowledge depth;

[0139] S63: Based on the user's preferences and behavior patterns, use a path search algorithm (such as the A* algorithm) to explore the knowledge graph, extract knowledge nodes related to the user's interests from the knowledge graph, and expand from the knowledge nodes related to the user's interests to adjacent related fields based on the constructed interdisciplinary links.

[0140] In this embodiment, the path search algorithm uses the A* algorithm, which is as follows:

[0141] (1) Based on user preferences and behavior patterns, select nodes related to the identified user interests from the knowledge graph as starting nodes;

[0142] (2) Define the weight of the edge based on the strength of the relationship between nodes, cognitive difficulty, and learning preference;

[0143] (3) Select the similarity between the node and the target knowledge as the heuristic evaluation criterion and define the heuristic function h(n):

[0144] h(n)=Sim(n,Goal);

[0145] Among them, n is the node, Goal is the target knowledge;

[0146] Preferably, similarity calculation can be based on topic tags, content relevance, or domain knowledge approximation;

[0147] (4) After A* explores relevant nodes, it uses interdisciplinary links to expand to adjacent related disciplines;

[0148] (5) Adjust the target node set of the heuristic search algorithm according to the interdisciplinary relationship chain, and adjust the weight of the interdisciplinary edge according to the user's knowledge depth and preference;

[0149] (6) Based on the results of path search, relevant knowledge nodes and content are recommended to users, providing in-depth exploration beyond current interests.

[0150] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0151] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0152] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0153] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0154] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. An AI online education intelligent question-answering information processing method, characterized in that: The following steps are involved: S1: Collect teaching materials from educational platforms, open courses, textbooks, and academic papers, and perform preprocessing and annotation to form a labeled text dataset; S2: Build a question-answering model based on XLNet, use the labeled data set to train the model, and perform parameter tuning to improve the accuracy of question-answer matching; S3: For education-related subjects or topics, introduce education-related context to optimize the question-answering model to obtain more accurate content understanding and obtain an optimized question-answering model; S4: Based on the labeled text dataset, consider the age-group hierarchical design and build a knowledge graph to assist in understanding information; and link knowledge in different subject areas by building shared nodes; S5: The user enters a query, and NLU analyzes the user's question, extracts key intents and entities, and obtains the best matching answer based on the optimized question-answering model. Then, based on the user's age group and the best matching answer, auxiliary understanding information is obtained from the knowledge graph to get a complete answer. S6: Automatically adjust the content and difficulty provided through user interaction data, and recommend related fields based on user behavior and preferences, thereby increasing user enthusiasm and engagement; The question-answering model is built based on XLNet, and the model is trained using annotated data sets, and parameter tuning is performed to improve the accuracy of question-answer matching, as follows: Based on the labeled dataset, where each question has a corresponding text paragraph and the correct answer, the data is divided into training set, validation set and test set; Initialize the XLNet model as the base model for the question-answering task. The input consists of questions and passages. Concatenate them into a sequence, retaining separators and position information to help the model understand the structure. During the training process, we use labeled question-answer pairs to fine-tune XLNet. The loss function uses cross entropy loss and we train the predicted start and end positions separately. Use the Adam optimizer to update the weights, set the initial learning rate to 5e-5, and use the learning rate scheduler to adjust the optimization process; After training is completed, save the best model and its configuration to obtain the trained question-answering model, convert the trained question-answering model into ONNX format and optimize it to achieve efficient cross-platform reasoning; The XLNet model is as follows: The input sequence x has T elements. XLNet considers all permutations z of the sequence based on the dynamics of permutations. XLNet considers the different possibilities of previous words and subsequent permutations when calculating the hidden state of each word; XLNet retains the basic architecture of Transformer and adjusts the self-attention mechanism to handle sequence arrangement and two-stream mechanism, including content stream and query stream: in, For content flow, To query the stream; are all the elements before the tth element in the permuted sequence; Transformer uses the self-attention mechanism to capture the dependencies between elements in the sequence; MASK represents a placeholder; is the element to be processed currently; Using the interpolation objective function, combined with autoregressive generation and a BERT-style masked language model, not only are the tokens masked during training, but the generative order is also scheduled. The objective function is as follows: ； in, are model parameters, Is the arrangement All elements before t in ; A sequence of preset arrangements; To arrange The tth element obtained after is the probability of generating the tth element of the current sequence under a given context given the model parameters θ.

2. According to claim 1, an AI online education intelligent question-answering information processing method is characterized in that: The preprocessing and labeling are as follows: Set a standard data set for validation, evaluate and improve the reliability of the data set through cross-validation, and filter out noise and redundant information; Use semi-supervised learning strategies and data augmentation techniques to expand the coverage of small-scale high-quality data; Use supervised NLP tools to automatically annotate samples, and use crowdsourcing platforms to verify the auxiliary automatic annotation results. Based on user feedback and verification results, a labeled text dataset is formed.

3. The AI ​​online education intelligent question-answering information processing method according to claim 2 is characterized in that: Use machine learning and NLP techniques to automatically filter out noise and redundant information, as follows: Use regular expressions and pattern recognition methods to remove redundant information from HTML tags, scripts, and references; use language detection algorithms to ensure that all input data is in the target language; use text normalization techniques to clean and standardize text; Use natural language processing technology to automatically detect and filter noise and redundant information in the data set, and apply TF-IDF to evaluate the importance of vocabulary: ; Among them, D represents the entire document set; represents the frequency of word t appearing in document d; is the inverse document frequency; The value is used to measure the importance of word t in document d; Set the upper and lower thresholds of the TF-IDF value to determine the words to keep: ； in, and are the upper and lower thresholds set; Build a preprocessing pipeline for text cleaning, including removing stop words, punctuation, and performing lemmatization.

4. The AI ​​online education intelligent question-answering information processing method according to claim 1, characterized in that: The above-mentioned educational-related subjects or topics introduce an educational context-optimized question-answering model to obtain a more accurate understanding of the content, as follows: Collect specialized corpora for education-related subjects or topics, including textbooks, papers, and industry reports; Use word embedding models to learn distributed representations of education-related vocabulary, use topic models to identify implicit topics in texts, and capture common terms and concepts in this field; Select domain-specific words and phrases from topic models and word vector analysis to build a general vocabulary and education-related concept set; Construct education-related keywords and phrases into contextual prompt phrases and insert them into the input text to strengthen the model's understanding of the domain and inject contextual prompts at the beginning or appropriate position of the input sequence; In the domain adaptation fine-tuning phase, data with domain labels is used for further fine-tuning; Loss function L used in fine-tuning domain Optimize model parameters: ； Among them, D domain is the domain fine-tuning dataset, x,y are the input and target output respectively; By introducing domain feature word weights in attention calculation, the model's attention to important concepts is enhanced: ； Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively; k is the dimension of the key vector; B is the domain feature bias matrix.

5. The AI ​​online education intelligent question-answering information processing method according to claim 1 is characterized in that: The labeled text dataset is designed based on age-group stratification to build a knowledge graph to assist understanding of information; and by building shared nodes, the knowledge in different subject areas is linked, as follows: Layer the teaching materials by age group and create knowledge nodes at different levels. Each node contains knowledge points and background information suitable for that age group. Use named entity recognition and relation extraction to identify entities and relations in text; use the tool spaCy for automated information extraction; Use RDF to represent and store knowledge graphs; nodes represent knowledge entities, and edges represent relationships between entities; And introduce age group labels in the knowledge graph to enable hierarchical access, with nodes and edges carrying metadata indicating applicable age groups and related subject areas; The content of each document and subject is represented as a vector based on topic distribution. The vectorized topic representation measures the similarity between different documents or subjects through cosine similarity. Graph edges are created between the identified nodes with similarities greater than the threshold, and interdisciplinary links are constructed to connect knowledge from different disciplines through graph edges.

6. The AI ​​online education intelligent question-answering information processing method according to claim 1, characterized in that: The S5 is specifically: S51: Clean the text input by the user, including removing noise characters and standardizing the input, using NLU to parse the text and extract the intent and entities in the query; S52: Retrieve and match answers based on the optimized question-answering model, and search the corpus to find answers based on the extracted intents and entities; S53: Use the user registration information or obtain the user's age group through interaction, and based on the answer, use the SPARQL graph database query language to extract relevant background information and explanations from the auxiliary understanding information knowledge graph, and adjust it according to the user's age to provide explanations of corresponding understanding levels.

7. The AI ​​online education intelligent question-answering information processing method according to claim 6, characterized in that: The S6 is specifically: S61: Record each user interaction, including query content, query frequency, selected topic, browsing time, and answer satisfaction; monitor user clicks, reading time, and answer feedback on different content; S62: Use K-means algorithm to cluster user behavior data, identify different preferences and behavior patterns, and use moving average method to predict the trend of user behavior data, the changing trend of subject orientation or knowledge depth; S63: Based on user preferences and behavior patterns, a path search algorithm is used to explore the knowledge graph, extract knowledge nodes related to user interests from the knowledge graph, and expand from knowledge nodes related to user interests to adjacent related fields based on the constructed interdisciplinary links.

8. The AI ​​online education intelligent question-answering information processing method according to claim 7, characterized in that: The path search algorithm uses the A* algorithm, which is as follows: (1) Based on user preferences and behavior patterns, select nodes related to the identified user interests from the knowledge graph as starting nodes; (2) Define the weight of the edge based on the strength of the relationship between nodes, cognitive difficulty, and learning preference; (3) Select the similarity between the node and the target knowledge as the heuristic evaluation criterion and define the heuristic function h(n): h(n)=Sim(n,Goal); Among them, n is the node, Goal is the target knowledge; Similarity calculation is based on topic tags, content relevance, or domain knowledge proximity; (4) After A* explores relevant nodes, it uses interdisciplinary links to expand to adjacent related disciplines; (5) Adjust the target node set of the heuristic search algorithm according to the interdisciplinary relationship chain, and adjust the weight of the interdisciplinary edge according to the user's knowledge depth and preference; (6) Based on the results of path search, relevant knowledge nodes and content are recommended to users, providing in-depth exploration beyond current interests.

Images