AI intelligent question answering system and method based on multiple GPT models

Abstract

The invention provides an AI intelligent question answering system based on multiple GPT models, and the system comprises a GPT model layer which comprises a plurality of GPT model instances, and each GPT model instance is finely adjusted according to different knowledge fields; the data management layer comprises a data collection module, a data cleaning module, a data labeling module and a data storage module; the platform layer comprises a user management module, an authority management module and an account binding module; and the integration and communication layer is used for data interaction between the system and an external platform. According to the method and the device, the GPT model instances for different knowledge fields are constructed, and the most suitable GPT model instance can be automatically selected to answer questions according to the questioning content and learning background of the user, so that the individual requirements of different users are met, and the accuracy and the user satisfaction of the intelligent question answering system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of AI intelligent Q&A, and particularly relates to an AI intelligent Q&A system and method based on multiple GPT models. Background Art

[0002] With the development of artificial intelligence technology, the current education field is rapidly digitizing, and intelligent Q&A systems are increasingly widely used in the education field. The main purpose of these systems is to provide instant learning support for students through automated means. However, existing intelligent Q&A systems still have deficiencies in many aspects.

[0003] First of all, most existing Q&A systems rely on rules and a preset answer library, and these answers are usually extracted from a fixed knowledge base. However, this method cannot fully consider the differences between individual students, such as learning progress, interests, and comprehension abilities. Therefore, these systems are difficult to provide truly personalized learning support, resulting in students being unable to obtain effective answers when encountering complex problems or problems not covered in the knowledge base.

[0004] Secondly, data privacy and security issues are another major challenge faced by existing systems. As these systems collect more and more student data, how to protect the privacy and security of this data has become an important issue. Improper data management may not only lead to data leakage but also violate the privacy rights of students. Many existing systems have imperfect measures in aspects such as data encryption and access control, resulting in data security risks.

[0005] Thirdly, existing systems generally lack an effective user permission management mechanism. The users of intelligent Q&A systems include not only students but also users of different roles such as teachers, parents, and administrators. The permissions of these users in the system should be different to ensure the security of the system and the reasonable use of resources. However, many systems fail to provide fine-grained enough permission control, resulting in the security and usability of the system being affected.

[0006] Finally, there are also deficiencies in the integration and scalability of the system. Many existing systems are difficult to integrate with other education platforms or tools, such as learning management systems (LMS) or communication tools (such as enterprise WeChat). This not only limits the functions and application scope of the system but also affects the user experience.

[0007] In summary, existing educational intelligent Q&A systems have many deficiencies in personalized services, data privacy protection, user permission management, and system integration and scalability, and there is an urgent need for a new solution to improve these problems. Summary of the Invention

[0008] The present invention provides an AI intelligent Q&A system and method based on multiple GPT models, which solves the problems such as the lack of personalized services in the existing educational intelligent Q&A systems.

[0009] The technical solution of the present invention is realized as follows:

[0010] In the first aspect of the present invention, an AI intelligent Q&A system based on multiple GPT models is provided, including:

[0011] The GPT model layer includes multiple GPT model instances, and each GPT model instance is fine-tuned according to different knowledge domains; the GPT model layer is used to automatically select the optimal GPT model instance for answering questions according to the user's question data and context information;

[0012] The data management layer includes a data collection module, a data cleaning module, a data annotation module, and a data storage module. The data collection module is used to collect a large amount of teaching data and learning data of different users. The data cleaning module is used to clean the collected data and remove noise and irrelevant information; the data annotation module is used to annotate the cleaned data; the data storage module is used to encrypt and store all the data;

[0013] The platform layer includes a user management module, a permission management module, and an account binding module. The user management module is used for user account registration / login, user information management, and user behavior monitoring; the permission management module is used to assign different system permissions to user accounts of different roles; the account binding module is used to bind each user account to a specific GPT model instance;

[0014] The integration and communication layer is used for data interaction between the system and external platforms.

[0015] Specifically, the fine-tuning method of the GPT model instance includes the following steps:

[0016] Collect a large amount of teaching data in a specific domain and preprocess the collected data;

[0017] Annotate the preprocessed teaching data, and annotate the category information and keyword information of the data;

[0018] Use the annotated teaching data to train and test the GPT model, and adjust the parameters of the GPT model according to the test results to make the question answering accuracy of the GPT model in a specific domain meet the set requirements.

[0019] Specifically, the method for selecting the optimal GPT model instance includes the following steps:

[0020] Keyword extraction: Use the TF-IDF algorithm to extract keywords from the user's question data. The calculation formula is as follows:

[0021] TF-IDF(t, d) = TF(t, d) × IDF(t);

[0022] Among them, t represents the keyword; d represents the document, that is, the user's question data; TF(t, d) represents the frequency of the keyword t appearing in the document d; IDF(t) represents the inverse document frequency of the keyword t in all documents;

[0023]

[0024] Among them, N represents the total number of documents, D represents the set of all documents, and |{d ∈ D: t ∈ d}| represents the number of documents containing the keyword t;

[0025] Topic extraction: Use the LDA model to extract the topic of the user's question data. The calculation formula is as follows:

[0026]

[0027] Among them, p(z|d) represents the probability that the document d belongs to the topic z, p(d|z) represents the probability of generating the document d given the topic z, p(z) represents the prior probability of the topic z, and p(d) represents the marginal probability of the document d;

[0028] Sentiment analysis: Use the sentiment analysis model to analyze the sentiment tendency of the user's question data. The sentiment score output by the sentiment analysis model is in the range of [-1, 1], indicating the user's sentiment tendency from negative to positive;

[0029] Model matching: Combine the extracted keywords, topics, and sentiment tendencies, and use a preset matching algorithm to select the GPT model instance with the highest matching degree from multiple GPT model instances;

[0030] Model call: Combine the user's question data and context information to call the selected optimal GPT model instance to generate an answer text.

[0031] Furthermore, the method of the model matching includes the following steps:

[0032] Feature vectorization: Convert the extracted keyword, topic, and sentiment tendency features into vector representations, and calculate the comprehensive feature vector V features :

[0033] V features = αV keywords + βV topics + γV sentiment ;

[0034] Among them, V keywords, V topics , V sentiment respectively represent the vector representations of the extracted keywords, themes, and sentiment tendencies, and α, β, and γ respectively represent the weight coefficients of the keyword vector, theme vector, and sentiment tendency vector;

[0035] Model similarity calculation. Calculate the similarity between the calculated comprehensive feature vector and the feature vectors of each GPT model instance. The calculation formula is as follows:

[0036]

[0037] where, V model is the feature vector of the GPT model instance, and ||V|| represents the norm of the vector V;

[0038] Optimal model selection. Select the GPT model instance with the highest similarity to the comprehensive feature vector from multiple GPT model instances as the optimal model M * :

[0039]

[0040] where, represents the similarity between the comprehensive feature vector and the i-th GPT model instance, and n is the number of GPT model instances.

[0041] Furthermore, the method for model invocation includes the following steps:

[0042] Context understanding. Integrate the user's question data with the context information to ensure that the model accurately understands the user's needs;

[0043] Input standardization. Convert the user's question data and context information into a standardized input format acceptable to the model to ensure the consistency of the input format;

[0044] Model inference. Pass the standardized input to the selected GPT model instance, and the model performs inference and generates a preliminary answer;

[0045] Answer post-processing. Revise and optimize the preliminary answer generated by the model to ensure the coherence, accuracy, and relevance of the answer;

[0046] Feedback and adjustment. Fine-tune the generated answer according to the user's feedback. If the user is not satisfied with the preliminary answer, the system adjusts the answer according to the feedback or re-selects a better GPT model instance to regenerate the answer.

[0047] Specifically, the context information includes at least one or more of the user's historical question data, current learning progress, and understanding ability level.

[0048] Specifically, the data collection includes:

[0049] User interaction logs, collecting all behavior logs of users in the system;

[0050] Online education resources, integrating teaching resources from multiple online education platforms;

[0051] Textbooks and supplementary materials, collecting and digitally processing traditional paper textbooks and supplementary materials;

[0052] Teaching data provided by IP teachers, integrating teaching data of IP teachers with top teaching quality rankings across the network;

[0053] The data cleaning includes: removing duplicate data, correcting error data, data standardization processing, and denoising processing;

[0054] The data annotation includes:

[0055] Classification annotation, classifying and annotating data according to the categories of different knowledge fields;

[0056] Important information annotation, annotating important information and keywords in the data;

[0057] Difficulty level annotation, grading and annotating according to the difficulty and complexity of the data;

[0058] The data storage includes:

[0059] Encrypted storage, encrypting data using the AES-256 algorithm;

[0060] Access control, restricting unauthorized users' access to system data by setting data access permissions;

[0061] Data backup, backing up data regularly.

[0062] Specifically, the user management module includes:

[0063] User registration and login unit, used for registering or logging in user accounts;

[0064] User information management unit, used for users to manage personal information;

[0065] User role management unit, used for administrators to manage user role information in the system;

[0066] User behavior recording unit, used for recording users' behavior logs in the system;

[0067] The permission management module includes:

[0068] Permission setting unit, used for administrators to set corresponding levels of access permissions in the system according to the roles of different user accounts;

[0069] A permission review unit, where the user administrator reviews and approves permission change requests submitted by users;

[0070] A permission change unit, used for administrators to adjust users' access permissions in the system;

[0071] A permission log unit, used to record logs of all permission changes;

[0072] The account binding module includes:

[0073] An instance selection unit, used for users to select a suitable GPT model instance in the account settings;

[0074] An instance binding unit, used to bind the GPT model instance selected by the user to the corresponding user account;

[0075] An instance switching unit, used for users to switch the bound GPT model instance;

[0076] An instance management unit, used for administrators to manage GPT model instances in the system.

[0077] Specifically, the external platform includes one or more of app, mini-program, client, and website.

[0078] The second aspect of the present invention provides an AI intelligent question answering method based on multiple GPT models, including the following steps:

[0079] Collect a large amount of teaching resource data and classify it according to the knowledge fields it belongs to;

[0080] Clean and label the classified data to obtain training data;

[0081] Build multiple GPT models, and use different categories of training data to train different GPT models respectively to obtain multiple GPT model instances;

[0082] According to the user's question data and context information, automatically select the optimal GPT model instance to answer questions and generate a preliminary answer;

[0083] Post-process the preliminary answer generated by the model to ensure the coherence, accuracy, and relevance of the answer;

[0084] Fine-tune the post-processed answer according to the user's feedback to generate an answer that meets the user's needs.

[0085] Compared with the prior art, the present invention has the following beneficial effects:

[0086] (1) By constructing multiple GPT model instances for different knowledge domains, the present invention can automatically select the most suitable GPT model instance for answering questions according to the content of the user's question and the learning background, thereby meeting the personalized needs of different users and improving the accuracy and user satisfaction of the intelligent question-answering system;

[0087] (2) By encrypting the storage and controlling the access rights of user data, the present invention ensures the privacy and security of user data, and can effectively prevent data leakage and improper use;

[0088] (3) By docking the data of the intelligent question-answering system with an external platform, the present invention realizes convenient user communication and system notification functions. Users can obtain learning support and system notifications through familiar platforms, improving the usability and interactivity of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0090] Figure 1 It is a schematic diagram of the architecture of an AI intelligent question-answering system based on multiple GPT models according to the present invention;

[0091] Figure 2 It is a schematic diagram of the flow of an AI intelligent question-answering method based on multiple GPT models according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0092] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0093] Referring to Figure 1 , the first aspect of the embodiment of the present invention provides an AI intelligent question-answering system based on multiple GPT models, including:

[0094] The GPT model layer includes multiple GPT model instances, and each GPT model instance is fine-tuned according to different knowledge domains; the GPT model layer is used to automatically select the optimal GPT model instance for answering questions according to the user's question data and context information;

[0095] The data management layer includes a data collection module, a data cleaning module, a data annotation module, and a data storage module. The data collection module is used to collect a large amount of teaching data and learning data of different users. The data cleaning module is used to clean the collected data to remove noise and irrelevant information. The data annotation module is used to annotate the cleaned data. The data storage module is used to encrypt and store all the data.

[0096] The platform layer includes a user management module, a permission management module, and an account binding module. The user management module is used for user account registration / login, user information management, and user behavior monitoring. The permission management module is used to assign different system permissions to user accounts of different roles. The account binding module is used to bind each user account to a specific GPT model instance.

[0097] The integration and communication layer is used for data interaction between the system and external platforms.

[0098] In this embodiment, the GPT model instances may include, but are not limited to, the following three types:

[0099] Plagiarism Master: It mainly processes popular videos on the Internet. This model can disassemble and refine professional scripts, create new ones, transform the original popular categories into English categories, and generate popular short video scripts suitable for English categories. Through training on a large amount of video script data, Plagiarism Master can identify the key elements of popular videos and generate new and attractive English short video scripts.

[0100] Copywriting Assistant: It can generate copy suitable for the Internet according to the user's own product characteristics. Through training on a large amount of advertising and marketing data, Copywriting Assistant can generate creative and attractive copy to help users improve the product promotion effect. Copywriting Assistant can understand the core selling points of the product and generate high-quality marketing content in combination with market trends and user needs.

[0101] English Helper: A special model trained based on all the English teaching and research systems and real exam course answers of the top three English IP teachers on the Internet. English Helper can provide professional English answers, covering aspects such as vocabulary, grammar, reading comprehension, and writing guidance. Through training on a large amount of English teaching data, English Helper can provide accurate English learning support to help users improve their English level.

[0102] Specifically, the fine-tuning method of the GPT model instance includes the following steps:

[0103] Collect a large amount of teaching data in a specific field (such as mathematics textbooks, exercise questions and answers, historical documents and papers, etc.), and preprocess the collected data (including cleaning the collected data, removing noise and irrelevant information to ensure the quality of the data);

[0104] Annotate the preprocessed teaching data, and annotate important information such as the category information and keywords of the data;

[0105] Use the annotated teaching data to train and test the GPT model, and adjust the parameters of the GPT model according to the test results to make the accuracy of answering questions of the GPT model in a specific field meet the set requirements;

[0106] Evaluate the trained model, test its performance on different questions, and further adjust the model parameters according to the evaluation results.

[0107] Specifically, the method for selecting the optimal GPT model instance includes the following steps:

[0108] Keyword extraction, use the TF-IDF algorithm to extract keywords from the user's question data. TF-IDF measures the importance of a word to a document, and the calculation formula is as follows:

[0109] TF-IDF(t,d) = TF(t,d) × IDF(t);

[0110] Among them, t represents the keyword; d represents the document, that is, the user's question data; TF(t,d) represents the frequency of the keyword t appearing in the document d; IDF(t) represents the inverse document frequency of the keyword t in all documents;

[0111]

[0112] Among them, N represents the total number of documents, D represents the set of all documents, and |{d∈D: t∈d}| represents the number of documents containing the keyword t;

[0113] Topic extraction, use the LDA model to extract the topic of the user's question data. LDA is a generative model, assuming that each document is a mixture of several topics, and each topic is defined by a set of vocabulary distributions. The calculation formula is as follows:

[0114]

[0115] Among them, p(z|d) represents the probability that the document d belongs to the topic z, p(d|z) represents the probability of generating the document d given the topic z, p(z) represents the prior probability of the topic z, and p(d) represents the marginal probability of the document d;

[0116] Sentiment analysis uses sentiment analysis models (such as TextBlob or VADER) to analyze the sentiment tendency of user question data and determine the emotional state of the questioner. The sentiment score output by the sentiment analysis model is in the range of [-1, 1], indicating the sentiment tendency of the user from negative to positive;

[0117] Model matching combines the extracted keywords, themes, and sentiment tendencies, and uses a preset matching algorithm (the matching algorithm can use techniques such as TF-IDF and word vectors for feature matching) to select the GPT model instance with the highest matching degree from multiple GPT model instances;

[0118] Model invocation combines the user's question data and context information to invoke the selected optimal GPT model instance to generate an answer text.

[0119] Furthermore, the method of model matching includes the following steps:

[0120] Feature vectorization converts the extracted keyword, theme, and sentiment tendency features into vector representations and calculates the comprehensive feature vector V features :

[0121] V features = αV keywords + βV topics + γV sentiment ;

[0122] Among them, V keywords 、V topics 、V sentiment respectively represent the vector representations of the extracted keywords, themes, and sentiment tendencies, and α, β, and γ respectively represent the weight coefficients of the keyword vector, theme vector, and sentiment tendency vector;

[0123] Model similarity calculation calculates the similarity between the calculated comprehensive feature vector and the feature vectors of each GPT model instance. The calculation formula is as follows:

[0124]

[0125] Among them, V model is the feature vector of the GPT model instance, and ||V|| represents the modulus of the vector V;

[0126] Optimal model selection selects the GPT model instance with the highest similarity to the comprehensive feature vector from multiple GPT model instances as the optimal model M * :

[0127]

[0128] Among them, represents the similarity between the comprehensive feature vector and the i-th GPT model instance, and n is the number of GPT model instances.

[0129] Further, the method for model invocation includes the following steps:

[0130] Context understanding: Use the integrate_context() function to integrate the user's question data with context information (including information such as the user's question, history, and current situation) to ensure that the model accurately understands the user's needs;

[0131] Input standardization: Use the standardize_input() function to convert the user's question data and context information into a standardized input format acceptable to the model to ensure the consistency of the input format;

[0132] Model inference: Pass the standardized input to the selected GPT model instance, and the model performs inference and generates a preliminary answer (using the model_instance.infer() method to execute the inference process of the model);

[0133] Answer post-processing: Use the post_process_answer() function to correct and optimize the preliminary answer generated by the model (including spelling check, grammar correction, format adjustment, etc.) to ensure the coherence, accuracy, and relevance of the answer;

[0134] Feedback and adjustment: Use the adjust_answer_or_switch_model() function to fine-tune the generated answer according to the user's feedback. If the user is not satisfied with the preliminary answer, the system adjusts the answer according to the feedback or re-selects a better GPT model instance to regenerate the answer.

[0135] Through this detailed algorithm logic and pseudocode, the system can achieve precise invocation of GPT model instances and generate answers that meet the user's needs and context. This step ensures the intelligence and flexibility of the system.

[0136] The system of the present invention can understand the questions raised by users, analyze the semantic structure of the questions, and generate answers that conform to logic and grammar. During this process, the system can also adjust the answering style and depth according to the user's historical question records to better meet the user's needs.

[0137] The semantic understanding and generation function of the system is implemented through the GPT model. Through learning from large-scale text data, the GPT model can understand and generate natural language texts that conform to grammar and logic. After receiving the user's question, the system first performs semantic analysis on the question to determine the theme and key content of the question, and then generates an answer that conforms to the context.

[0138] Specifically, the context information includes at least one or more of the user's historical question data, current learning progress, and comprehension ability level.

[0139] The following uses a specific case to illustrate how the system of the present invention automatically selects the most suitable GPT model instance and generates an answer:

[0140] Case 1: The user asks "How to improve English writing skills?"

[0141] Step 1, Model Selection

[0142] The system performs semantic analysis on the question and extracts the keywords "improve" and "English writing skills";

[0143] According to keyword matching, the system selects the "English Assistant" model instance because this instance is optimized for English learning;

[0144] Step 2, Semantic Analysis and Generation

[0145] The system analyzes the question, determines the theme as "English writing skills", and the key content as "how to improve";

[0146] Combined with the user's historical learning records (such as the user has previously consulted English grammar and vocabulary questions), the system understands the user's learning background;

[0147] The "English Assistant" model instance generates an answer: "To improve English writing skills, you can start from the following aspects: 1. Increase the amount of reading and accumulate vocabulary and sentence patterns; 2. Practice writing more and pay attention to logic and structure; 3. Find writing partners and revise and give feedback to each other; 4. Study classic model essays and learn writing skills from them."

[0148] Case 2: The user requests to generate an advertising copy

[0149] Step 1, Model Selection

[0150] The system performs semantic analysis on the request and extracts the keywords "generate" and "advertising copy";

[0151] According to keyword matching, the system selects the "Copywriting Assistant" model instance because this instance is optimized for advertising and marketing copy;

[0152] Step 2, Semantic Analysis and Generation

[0153] The system analyzes the request, determines the theme as "advertising copy", and the key content as "generate";

[0154] Combined with the user's product features and market demands, the system generates an advertising copy with creativity and attractiveness;

[0155] The "Copywriting Assistant" model instance generates the copy: "Experience the brand-new product, enhance your life quality, purchase now and enjoy the time-limited discount!"

[0156] Case 3: The user requests to generate a new video script by paraphrasing.

[0157] Step 1, model selection

[0158] The system performs semantic analysis on the request and extracts the keywords "paraphrasing" and "video script".

[0159] Based on keyword matching, the system selects the "Paraphrasing Master" model instance because this instance is optimized for paraphrasing video scripts.

[0160] Step 2, semantic analysis and generation

[0161] The system analyzes the request, determines that the theme is "video script", and the key content is "paraphrasing".

[0162] The system understands the key elements and style of the original video script and re-creates it through the Paraphrasing Master.

[0163] The "Paraphrasing Master" model instance generates a new English category video script, retaining the original attractiveness while meeting the new category requirements.

[0164] Specifically, the system in this embodiment is specially trained for English learning. The system relies on the top three English IP teachers across the network and has a large amount of high-quality English learning data as the training basis. In addition, a large number of new users learn on the system platform every month, generating rich learning process data. This data includes users' learning progress, wrong answers, improvement suggestions, etc., greatly enriching the system's training dataset and enabling the system to continuously optimize and improve its Q&A ability.

[0165] In the training of English learning, it mainly includes the following aspects:

[0166] Vocabulary learning: The system provides rich vocabulary learning materials, including the definitions, usages, examples, and synonyms of words. Through training on a large amount of English vocabulary data, the system can provide accurate vocabulary explanations and usage guidance when users ask questions.

[0167] Grammar learning: The system covers English grammar rules from basic to advanced. Through training on a large amount of grammar data, the system can answer various questions about grammar from users and provide detailed grammar analysis and practice questions.

[0168] Reading Comprehension: The system has collected a large number of reading comprehension materials, including short passages, novels, news reports, etc. Through training on these materials, the system can help users improve their reading comprehension ability, provide reading skills and answer analysis;

[0169] Writing Skills: The system provides writing guidance and model essays. Through training on writing data, the system can provide writing suggestions and evaluations for users to help them improve their writing ability.

[0170] The present invention utilizes the data resources of the top three English IP teachers across the network. These data cover a wide range of English learning content, ensuring the high quality of training data. In particular, the newly added user learning process data each month provides the system with the latest and most practical learning situation, helping the system continuously optimize the GPT model and making its performance in answering questions in the field of English learning more excellent. Through these targeted trainings, the system performs outstandingly in the field of English learning, can provide high-quality English learning support for users, and helps users improve their English level.

[0171] Specifically, the data collection includes:

[0172] User interaction logs, collecting all the behavior logs of users in the system, including questions, learning, evaluations, etc. By analyzing these log data, the system can understand users' learning habits and needs and further optimize the question-answering service;

[0173] Online education resources, integrating teaching resources from multiple online education platforms, including video courses, e-books, question banks, etc., to provide diverse learning materials for the system;

[0174] Textbooks and tutoring materials, collecting and digitally processing traditional paper textbooks and tutoring materials and converting them into electronic data available for the system to use;

[0175] Teaching data provided by IP teachers, integrating the teaching data of the top three English IP teachers across the network. These data include teachers' teaching videos, lecture notes, exercises and answers, etc., to provide high-quality English learning data for the system.

[0176] The data cleaning includes:

[0177] Removing duplicate data, checking and deleting duplicate entries in the dataset to prevent data redundancy;

[0178] Correcting incorrect data, identifying and correcting errors in the data, such as spelling mistakes, incomplete information, etc.;

[0179] Data standardization processing, performing standardization processing on the data to unify the data format and unit for convenient subsequent analysis and processing;

[0180] Denoising processing, filtering out noise information in the data, such as irrelevant characters, tags, etc., to ensure the purity of the data.

[0181] The data annotation includes:

[0182] Classification annotation, classifying and annotating the data according to the categories of different knowledge fields, such as vocabulary, grammar, reading, writing, etc., to provide clear learning goals for the model;

[0183] Important information annotation, annotating important information and keywords in the data, such as key vocabulary, grammar rules, problem-solving ideas, etc., to help the model understand and generate more accurate answers;

[0184] Difficulty level annotation, grading and annotating according to the difficulty and complexity of the data to ensure that the model can handle problems at different levels.

[0185] The data storage includes:

[0186] Encrypted storage, using the AES-256 algorithm to encrypt the data to prevent data leakage and tampering during storage and ensure the high security of the data;

[0187] Access control, restricting unauthorized users' access to system data by setting data access permissions. Only authorized users and system modules can access specific data; through fine-grained permission settings, ensure the rationality and security of data use;

[0188] Data backup, regularly backing up the data to prevent data loss and damage, and adopting a multiple backup strategy to ensure the integrity and recoverability of the data.

[0189] Specifically, the system of the present invention supports the management of multiple user types, including students, teachers, parents, head teachers, teaching assistants, big IP teachers, and sales consultants. Each user type has different permissions and function modules. For example, students can ask questions and view learning records, teachers can view students' learning progress and answer quality, parents can obtain their children's learning reports, head teachers can track the overall learning situation of class students, teaching assistants can assist head teachers in teaching, big IP teachers can provide high-quality teaching content and answers, sales consultants can manage users' purchases and service situations, and administrators are responsible for the overall management and maintenance of the system.

[0190] Specifically, the user management module includes:

[0191] User registration and login unit, used to register or log in to a user account. The registration process requires providing basic personal information and verification information;

[0192] User Information Management Unit, which is used for users to manage their personal information. Users can manage their personal information in the system, including modifying passwords, updating personal profiles, etc.;

[0193] User Role Management Unit, which is used for administrators to manage the role information of users in the system. The system supports multiple user roles, and each role has different permissions and functions;

[0194] User Behavior Record Unit, which is used to record the behavior logs of users in the system, including login records, operation records, etc., facilitating system monitoring and auditing by administrators.

[0195] Specifically, the Permission Management Module assigns different system permissions to users with different roles; for example, students can only access public resources and their own learning data, while teachers can access more educational resources and students' learning records. Permission management ensures the security of the system and the reasonable use of resources.

[0196] The said Permission Management Module includes:

[0197] Permission Setting Unit, which is used for administrators to set corresponding levels of access permissions according to the roles of different user accounts in the system, ensuring that users with different roles can only access resources and functions within their permission scope;

[0198] Permission Review Unit, where user administrators review and approve permission change requests submitted by users. When users apply for higher permissions, they need to go through the review and approval of administrators;

[0199] Permission Change Unit, which is used for administrators to adjust the access permissions of users in the system. The permissions of users can be changed according to actual needs;

[0200] Permission Log Unit, which is used to record the logs of all permission changes, facilitating auditing and traceability by administrators.

[0201] Specifically, each user account can be bound to a specific GPT model instance. This means that each question of the user will be processed by the most suitable GPT model instance, providing personalized Q&A services. Users can select or change the bound GPT instance in the account settings to adapt to their changing learning needs.

[0202] The said Account Binding Module includes:

[0203] Instance Selection Unit, which is used for users to select a suitable GPT model instance in the account settings. The system provides multiple instances for users to choose from;

[0204] Instance Binding Unit, which is used to bind the GPT model instance selected by the user to the corresponding user account. When the system processes the questions of the user, it will automatically select the corresponding GPT model instance;

[0205] An instance switching unit for users to switch the bound GPT model instance. Users can switch the bound GPT model instance at any time according to their needs, and the system will adjust the answering strategy according to the user's selection.

[0206] An instance management unit for administrators to manage GPT model instances in the system, including adding, deleting, and updating instances, to ensure the flexibility and scalability of the system.

[0207] Specifically, the external platform includes one or more of app, mini-program, client, and website.

[0208] In this embodiment, the external platform can adopt Enterprise WeChat. The integration of the system with Enterprise WeChat enables users to receive system notifications and answering results through Enterprise WeChat. For example, when a student asks a question, the system can send the answering result to the student through Enterprise WeChat. In addition, teachers can also receive system notifications through Enterprise WeChat, such as students' learning progress reports or system update information.

[0209] The Enterprise WeChat integration function module includes the following aspects:

[0210] Message push module: The system can push messages to users through Enterprise WeChat, including answering results, system notifications, etc. Users can view the messages in Enterprise WeChat, which is convenient and fast.

[0211] Notification management module: Administrators can set the rules and content of message notifications in the system to ensure the timeliness and accuracy of message pushing.

[0212] User interaction module: Users can interact with the system through Enterprise WeChat, including asking questions, giving feedback, etc. The system will process users' requests in real time and feedback the results through Enterprise WeChat.

[0213] Data synchronization module: A data synchronization mechanism between the system and Enterprise WeChat to ensure that users' operations in Enterprise WeChat can be reflected in the system in real time.

[0214] Due to the relatively dense time points of information interaction in the education industry, the system needs to be able to handle a large number of information interaction requirements during peak periods. These information interactions include not only text but also various forms such as pictures and voices. The system design fully considers these characteristics to ensure that it can efficiently process various types of information and maintain stability and efficiency in a high-concurrency environment.

[0215] Diversity processing of information types: The system supports the processing of multiple information types, including text, pictures, voices, etc. Through advanced natural language processing and image recognition technologies, the system can accurately understand and process users' various information requests.

[0216] Peak processing of information interaction: The system designs an efficient message queue and load balancing mechanism to ensure quick response and processing of user requests during peak information interaction periods, avoiding response delays caused by system congestion;

[0217] Guarantee of information integrity: The system ensures the complete transmission and processing of all interaction information, avoiding information loss or damage. Through data backup and recovery mechanisms, the system can quickly recover in case of anomalies, ensuring the continuity and reliability of the user experience.

[0218] Refer to Figure 2 , the second aspect of the embodiment of the present invention provides an AI intelligent question answering method based on multiple GPT models, including the following steps:

[0219] Collect a large amount of teaching resource data and classify it according to the knowledge fields it belongs to;

[0220] Clean and label the classified data to obtain training data;

[0221] Construct multiple GPT models, and use different categories of training data to train different GPT models respectively to obtain multiple GPT model instances;

[0222] According to the user's question data and context information, automatically select the optimal GPT model instance to answer questions and generate a preliminary answer;

[0223] Post-process the preliminary answer generated by the model to ensure the coherence, accuracy, and relevance of the answer;

[0224] Fine-tune the post-processed answer according to the user's feedback to generate an answer that meets the user's needs.

[0225] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. An AI intelligent question answering system based on multiple GPT models, characterized in that: include: The GPT model layer includes multiple GPT model instances, each of which is fine-tuned according to different knowledge domains; The GPT model layer is used to automatically select the optimal GPT model instance for answering questions based on the user's question data and context information; The data management layer includes a data collection module, a data cleaning module, a data annotation module and a data storage module. The data collection module is used to collect a large amount of teaching data and learning data of different users. The data cleaning module is used to clean the collected data and remove noise and irrelevant information. The data annotation module is used to annotate the cleaned data. The data storage module is used to encrypt and store all data. The platform layer includes a user management module, a permission management module and an account binding module. The user management module is used for user account registration / login, user information management and user behavior monitoring; the permission management module is used to assign different system permissions to user accounts of different roles; The account binding module is used to bind each user account to a specific GPT model instance; Integration and communication layer, used for data interaction between the system and external platforms.

2. The AI ​​intelligent question answering system based on multiple GPT models as claimed in claim 1, characterized in that: The fine-tuning method of the GPT model instance comprises the following steps: Collect a large amount of teaching data in a specific field and pre-process the collected data; Label the preprocessed teaching data, label the category information and keyword information of the data; The GPT model is trained and tested using labeled teaching data, and the parameters of the GPT model are adjusted based on the test results so that the GPT model's accuracy in answering questions in specific fields meets the set requirements.

3. The AI ​​intelligent question answering system based on multiple GPT models as claimed in claim 1, characterized in that: The method for selecting the optimal GPT model instance includes the following steps: Keyword extraction: Use the TF-IDF algorithm to extract keywords from user question data. The calculation formula is as follows: TF-IDF(t,d)=TF(t,d)×IDF(t); Where t represents a keyword; d represents a document, i.e., user question data; TF(t,d) represents the frequency of keyword t appearing in document d; IDF(t) represents the inverse document frequency of keyword t in all documents; Where N represents the total number of documents, D represents the set of all documents, and |{d∈D:t∈d}| represents the number of documents containing keyword t; Topic extraction, using the LDA model to extract the topic of user question data, the calculation formula is as follows: Where p(z|d) represents the probability that document d belongs to topic z, p(d|z) represents the probability of generating document d given topic z, p(z) represents the prior probability of topic z, and p(d) represents the marginal probability of document d; Sentiment analysis: Use the sentiment analysis model to analyze the sentiment tendency of user question data. The sentiment score output by the sentiment analysis model is in the range of [-1,1], indicating the user's sentiment tendency from negative to positive; Model matching: combining the extracted keywords, topics, and sentiment tendencies, and using a preset matching algorithm to select the GPT model instance with the highest matching degree from multiple GPT model instances; Model call, combining the user's question data and context information to call the selected optimal GPT model instance to generate the question answer text.

4. The AI ​​intelligent question answering system based on multiple GPT models as claimed in claim 3, characterized in that: The model matching method comprises the following steps: Feature vectorization: convert the extracted keywords, topics, and sentiment features into vector representations and calculate the comprehensive feature vector V features : V features =αV keywords +βV topics +γV sentiment ; Among them, V keywords 、V topics 、V sentiment They represent the vector representations of the extracted keywords, topics, and sentiment tendencies, respectively; α, β, and γ represent the weight coefficients of the keyword vector, topic vector, and sentiment tendency vector, respectively; Model similarity calculation: the calculated comprehensive feature vector is similar to the feature vector of each GPT model instance. The calculation formula is as follows: Among them, V model is the feature vector of the GPT model instance, ||V|| represents the modulus of vector V; Optimal model selection: select the GPT model instance with the highest similarity to the comprehensive feature vector from multiple GPT model instances as the optimal model M * : in, It represents the similarity between the comprehensive feature vector and the i-th GPT model instance, and n is the number of GPT model instances.

5. The AI ​​intelligent question answering system based on multiple GPT models as claimed in claim 3, characterized in that: The model calling method comprises the following steps: Contextual understanding: integrating user question data with contextual information to ensure that the model accurately understands user needs; Input standardization: converting user question data and context information into a standardized input format acceptable to the model to ensure consistency of the input format; Model inference, passing the standardized input to the selected GPT model instance, the model performs inference and generates preliminary answers; Answer post-processing: correcting and optimizing the preliminary answers generated by the model to ensure the coherence, accuracy, and relevance of the answers; Feedback and adjustment: Fine-tune the generated answers based on user feedback. If the user is not satisfied with the initial answer, the system adjusts the answer based on the feedback or reselects a better GPT model instance to regenerate the answer.

6. The AI ​​intelligent question answering system based on multiple GPT models as claimed in claim 1, characterized in that: The context information includes at least one or more of the user's historical question data, current learning progress, and comprehension ability level.

7. The AI ​​intelligent question answering system based on multiple GPT models as claimed in claim 1, characterized in that: The data collection includes: User interaction logs, which collect all user behavior logs in the system; Online education resources, integrating teaching resources from multiple online education platforms; Teaching materials and tutoring materials: collect and digitize traditional paper teaching materials and tutoring materials; The teaching data provided by IP teachers integrates the teaching data of IP teachers with the highest teaching quality rankings on the entire network; The data cleaning includes: removing duplicate data, correcting erroneous data, data standardization and denoising; The data annotation includes: Classification and labeling: classify and label data according to categories in different knowledge fields; Important information annotation: annotate important information and keywords in the data; Difficulty level labeling: graded labeling based on the difficulty and complexity of the data; The data storage includes: Encrypted storage, using AES-256 algorithm to encrypt data; Access control, which limits unauthorized users’ access to system data by setting data access permissions; Data backup, back up data regularly.

8. The AI ​​intelligent question answering system based on multiple GPT models as claimed in claim 1, characterized in that: The user management module comprises: User registration and login unit, used to register or log in to a user account; User information management unit, used for users to manage personal information; User role management unit, used by administrators to manage user role information in the system; User behavior recording unit, used to record the user's behavior log in the system; The rights management module includes: The permission setting unit is used by the administrator to set the corresponding level of access rights according to the roles of different user accounts in the system; Permission review unit: User administrators review and approve permission change requests submitted by users; The permission change unit is used by administrators to adjust the user's access rights in the system; Permission log unit, used to record logs of all permission changes; The account binding module includes: The instance selection unit is used by users to select a GPT model instance that suits them in account settings; An instance binding unit, used to bind the GPT model instance selected by the user to the corresponding user account; An instance switching unit, used by users to switch bound GPT model instances; The instance management unit is used by administrators to manage GPT model instances in the system.

9. The AI ​​intelligent question answering system based on multiple GPT models as claimed in claim 1, characterized in that: The external platform includes one or more of an app, a mini-program, a client, and a website.

10. An AI intelligent question answering method based on multiple GPT models, characterized in that: The following steps are involved: Collect a large amount of teaching resource data and classify them according to the knowledge field they belong to; Clean and label the classified data to obtain training data; Build multiple GPT models, use different categories of training data to train different GPT models, and obtain multiple GPT model instances; Automatically select the best GPT model instance to answer questions and generate preliminary answers based on the user's question data and context information; Post-process the initial answers generated by the model to ensure coherence, accuracy, and relevance of the answers; The post-processed answers are fine-tuned based on user feedback to generate answers that meet user needs.

Images