Online video training method based on large model

Abstract

The invention provides an online video training method based on a large model, and the method comprises the steps: constructing teaching data, and displaying the teaching data to a user in a visual manner; carrying out vectorization processing on the teaching data, dividing the teaching data through the similarity between vectors, and adding knowledge point positioning for the teaching data according to a division result; generating a multi-question-type question bank according to the characteristics of different question types in combination with teaching data with knowledge point positioning, and selecting part of questions from the multi-question-type question bank to interact with a user; semantic migration is carried out on the questions so that the questions interacted with the user each time are not completely the same, part of teaching data is played back according to the interaction result of the user and the questions, and personalized review plans and learning suggestions are generated. The system has the beneficial effects that long videos are divided into short videos, and a quick retrieval module and an autonomous question setting module are added, so that the interactivity is increased, the learning interest of students is improved, the learning efficiency is improved, and the learning effect is improved.

Description

Technical Field

[0001] The invention belongs to the field of computer technology, and in particular relates to an online video training method based on a large model. Background Art

[0002] Online video education and training has developed rapidly in recent years, mainly driven by the popularization and technological progress of mobile terminals, 5G communications and the Internet. Distance learning has become mainstream, and many educational institutions and platforms have transformed to provide a wealth of online courses to meet the diverse needs of learners. This trend not only improves the flexibility of learning, but also creates opportunities for the popularization and globalization of educational resources.

[0003] The current shortcomings of online video training mainly include the following:

[0004] 1. The training video is too long: Some online videos are very long. When the amount of information is huge, it will cause information overload for trainees, making it difficult to grasp the key points, making learning difficult and inefficient.

[0005] 2. Lack of interactivity: Traditional video training is usually a one-way communication method, and students can only passively receive information, which leads to a decrease in learning interest; in traditional question banks, students directly memorize and practice answers during exams, and cannot apply what they have learned to other situations, thus failing to achieve true understanding.

[0006] 3. Limited learning depth: Many traditional video trainings may only provide a superficial introduction, lacking in-depth discussion and detailed analysis. Usually, the information is copied and pasted into search engines such as Baidu, which is messy and requires time to summarize by yourself, or the general large model is used to obtain more knowledge, but it is easy to have knowledge illusions, which may mislead students. These increase the difficulty of acquiring knowledge and often give up further research. Summary of the invention

[0007] In view of this, the present invention aims to propose an online video training method based on a large model, in order to solve at least one of the above-mentioned technical problems.

[0008] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0009] A large model-based online video training method, the method comprising:

[0010] Construct teaching data and present it to users in a visual way;

[0011] Vectorize the teaching data, divide the teaching data by the similarity between vectors, and add knowledge point positioning to the teaching data based on the division results;

[0012] Generate a multi-question question bank based on the characteristics of different question types and teaching data with knowledge point positioning, and select some questions from the multi-question question bank to interact with users;

[0013] By performing semantic transfer on the questions, the questions that users interact with are not exactly the same each time. Based on the results of the interaction between the user and the questions, part of the teaching data is replayed, and personalized review plans and learning suggestions are generated.

[0014] Furthermore, the process of vectorizing the teaching data includes:

[0015] Preset teaching documents and teaching videos, separate the teaching videos to obtain teaching audio, transcribe the teaching audio into target text and perform vectorization processing; wherein, the time nodes of the teaching audio, teaching video and target text are associated;

[0016] The logical hierarchical structure in the teaching document is extracted and output in a structured format to obtain the teaching text and teaching pictures, and the teaching text and teaching pictures are vectorized respectively.

[0017] Furthermore, the process of dividing the teaching data by the similarity between the vectors and adding knowledge point positioning to the teaching data according to the division result includes:

[0018] Calculate the cosine similarity between vectors to obtain the segmentation boundary of knowledge points, divide the teaching data once according to the distribution of knowledge points, and use recursive processing to divide the result of the first division twice to obtain multiple data segments that meet the teaching requirements;

[0019] Add corresponding labels to each data segment, obtain the timestamp between two adjacent data segments, and add knowledge point locations in the teaching data based on the timestamp.

[0020] Furthermore, in the process of selecting some questions to interact with the user:

[0021] According to the knowledge points corresponding to the questions, a priority is set for each question, the questions are sorted according to the priority, and questions with higher priorities are preferentially selected to interact with users; wherein the selection priorities are from high to low: teaching text, target text, teaching picture.

[0022] Furthermore, after setting the priority:

[0023] Record the user's interaction traces and dynamically adjust the questions based on the interaction traces. For the same knowledge point, use different types of questions to interact with the user, and the question types in two consecutive interactions are different.

[0024] Furthermore, the generation process of the multi-question type question bank includes:

[0025] Use retrieval-enhanced generation technology to set different answer generation strategies for different question types, and input the answer generation strategies and teaching data into the large language model to generate corresponding questions and answers for different knowledge points in the teaching data.

[0026] Furthermore, the answer generation strategy includes:

[0027] If the question is a true or false question, the generation strategy is to judge the truth or falsehood of the statement based on its content, replace the content of the knowledge point with synonyms or heteronyms, and generate binary options;

[0028] If the question is a multiple-choice question, the generation strategy is to give multiple answers based on the content of the statement, and at least one of the answers is the correct option corresponding to the knowledge point;

[0029] If the question is a fill-in-the-blank question, the generation strategy is to fill in the missing parts according to the context or prompts, delete part of the content of the knowledge point, and record the deleted content as the answer;

[0030] If the question type is a question-and-answer question, the answer given by the user will be recorded and wait for manual processing.

[0031] Furthermore, the process of semantically migrating the topic includes:

[0032] According to the vectorization processing results, the content with high repetition rate in the same knowledge point is selected as the keywords, the keywords are searched on the Internet, and the keywords are replaced with synonyms according to the search results;

[0033] By screening out the contents with high repetition rate in different knowledge points as common words, the common words are removed from the keywords.

[0034] Furthermore, the process of replaying part of the teaching data according to the interaction result between the user and the question includes:

[0035] If the result of the interaction is that the user does not answer the question within the specified time or answers the question incorrectly, the teaching data will be rolled back to the location of the previous knowledge point and will be visually displayed to the user again.

[0036] Furthermore, the process of generating a personalized review plan and learning suggestions includes:

[0037] Questions with incorrect answers as a result of user interaction are selected, and the knowledge points corresponding to the questions are taken as the user's weaknesses. The weaknesses are counted and a review plan and learning suggestions based on the weaknesses are provided to the user.

[0038] Compared with the prior art, the online video training method based on a large model described in the present invention has the following beneficial effects:

[0039] Based on documents and training models, online long videos are intelligently segmented into short videos according to knowledge points, and corresponding question banks are generated based on courseware and video understanding. After users watch each key segment, the system automatically pops up the corresponding questions to help reinforce learning;

[0040] If you encounter knowledge points that you don’t understand, you can directly select them in the system for intelligent search and summary, saving students’ time;

[0041] Dividing long videos into short videos, adding a quick search module and an independent question-setting module, and increasing interactivity will help improve students' interest in learning, increase learning efficiency, and improve learning outcomes. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0043] Figure 1 This is a flow chart of the online video training method according to an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the target text vectorization process according to an embodiment of the present invention;

[0045] Figure 3 A schematic diagram of the vectorization process of teaching texts and teaching pictures according to an embodiment of the present invention;

[0046] Figure 4 A schematic diagram of a process of storing vectorized teaching data into a vector database according to an embodiment of the present invention;

[0047] Figure 5 This is a flow chart of the training question bank model M1 according to an embodiment of the present invention;

[0048] Figure 6 A schematic diagram of the process of generating a question bank according to an embodiment of the present invention;

[0049] Figure 7 This is a schematic diagram of the optimization process of the question bank according to an embodiment of the present invention;

[0050] Figure 8 A schematic diagram of a visualized interactive interface according to an embodiment of the present invention;

[0051] Fig. 9 A schematic diagram of a visualized interactive interface according to an embodiment of the present invention;

[0052] Fig.10 This is a schematic diagram of a visualized interactive interface according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0054] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0055] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0056] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0057] A large model-based online video training method, comprising:

[0058] S1. Construct teaching data and display it to users in a visual way. The screenshot of the visual interactive interface is as follows: Figure 8-10 shown.

[0059] S2. Vectorize the teaching data, divide the teaching data by the similarity between vectors, and add knowledge point positioning to the teaching data based on the division results.

[0060] In this embodiment, the execution process of step S2 includes:

[0061] Preset teaching documents and teaching videos, separate the teaching videos to obtain teaching audio, transcribe the teaching audio into target text and perform vectorization processing; wherein, the time nodes of the teaching audio, teaching video and target text are associated;

[0062] The logical hierarchical structure in the teaching document is extracted and output in a structured format to obtain the teaching text and teaching pictures, and the teaching text and teaching pictures are vectorized respectively.

[0063] Calculate the cosine similarity between vectors to obtain the segmentation boundary of knowledge points, divide the teaching data once according to the distribution of knowledge points, and use recursive processing to divide the result of the first division twice to obtain multiple data segments that meet the teaching requirements;

[0064] Add corresponding labels to each data segment, obtain the timestamp between two adjacent data segments, and add knowledge point locations in the teaching data based on the timestamp.

[0065] In the process of selecting some topics to interact with the user:

[0066] According to the knowledge points corresponding to the questions, set the priority for each question, sort the questions according to the priority, and give priority to the questions with higher priority to interact with the user; wherein, the selection priority is from high to low: teaching text, target text, teaching picture; after setting the priority:

[0067] Record the user's interaction traces and dynamically adjust the questions based on the interaction traces. For the same knowledge point, use different types of questions to interact with the user, and the question types in two consecutive interactions are different.

[0068] In some other embodiments, an execution flow of step S2 is as follows:

[0069] The course lecturer decides the course topic K1, writes the teaching document W1, and records the course video V1. The course lecturer transfers the materials of this course K1: video V1, teaching document W1 (supports unstructured data in rich formats such as markdown, pdf, doc, html, ppt, etc.) to the course topic K1.

[0070] like Figure 2 As shown, after the upload is completed, the audio and video separation task is executed to separate the audio and video of the video file to obtain audio A1, and the audio text transcription task is performed to convert audio A1 into target text VT1, and the text vectorization task is performed to vectorize the target text VT1 into Vec1, and the time points of the target text VT1, Vec1, audio A1, and video V1 are associated;

[0071] like Figure 3As shown, a general document parsing service is provided, which supports extracting logical hierarchical structures such as titles and paragraphs from unstructured documents (text, tables, pictures, etc.) and outputting them in a structured format. The content of the teaching document W1 is parsed and obtained into text WT1 and picture WP1.

[0072] Perform text vectorization tasks, vectorize the target text WT1 into Vec2, and vectorize the target image WP1 into Vec3. There are two ways to execute the image vectorization task. One is the image content understanding service (imagetovec1), which uses a multimodal large model to parse and understand the image content and recognize text. The parsed text vector is stored in the vector database. The other is the image text recognition service (imagetovec2), which uses OCR capabilities to recognize image text, and the parsed text vector is stored in the vector database.

[0073] like Figure 4 As shown, the above Vec1, Vec2, and Vec3 are sliced ​​and stored in the vector database, and the knowledge points are segmented semantically to segment the long video into short videos. The segmentation boundary of the knowledge points is determined by calculating the cosine similarity between the vectors. When the cosine similarity drops significantly, it means that it may be the beginning of a new knowledge point. When the similarity is matched, the threshold can be dynamically adjusted to ensure the rationality of the segmentation granularity and ensure that overly similar content will not be segmented into different knowledge points.

[0074] For some knowledge points, they can be further refined through recursive processing, that is, further retrieval and segmentation. For example, electrical knowledge points can be further divided into finer sub-knowledge points such as current, voltage, and resistance.

[0075] The knowledge point vector obtained by the above method is used to regenerate the short video information. Here, only the short video title SV needs to be generated. SV corresponds to the corresponding video duration interval. Assuming that the long video course is 2 hours and there are 20 knowledge points, the data of the short video collection is: K1 = {SV1, SV2, SV3, ..., SV19}. Since the start time of each short video is recorded, there is no need to physically cut the long video, and the knowledge point video can be located in the player device of the present invention.

[0076] S3. Generate a multi-question type question bank based on the characteristics of different question types and teaching data with knowledge point positioning, and select some questions from the multi-question type question bank to interact with users.

[0077] In this embodiment, the execution process of step S3 includes:

[0078] Use retrieval-enhanced generation technology to set different answer generation strategies for different question types, and input the answer generation strategy and teaching data into the large language model to generate corresponding questions and answers for different knowledge points in the teaching data. The answer generation strategy is as follows:

[0079] If the question is a true or false question, the generation strategy is to judge the truth or falsehood of the statement based on its content, replace the content of the knowledge point with synonyms or heteronyms, and generate binary options;

[0080] If the question is a multiple-choice question, the generation strategy is to give multiple answers based on the content of the statement, and at least one of the answers is the correct option corresponding to the knowledge point;

[0081] If the question is a fill-in-the-blank question, the generation strategy is to fill in the missing parts according to the context or prompts, delete part of the content of the knowledge point, and record the deleted content as the answer;

[0082] If the question type is a question-and-answer question, the answer given by the user will be recorded and wait for manual processing.

[0083] The process of semantically migrating the topic includes:

[0084] According to the vectorization processing results, the content with high repetition rate in the same knowledge point is selected as the keywords, the keywords are searched on the Internet, and the keywords are replaced with synonyms according to the search results;

[0085] By screening out the contents with high repetition rate in different knowledge points as common words, the common words are removed from the keywords.

[0086] Retrieval-Augmented Generation (RAG) is an artificial intelligence method that combines retrieval technology and generation technology to improve the relevance, accuracy and diversity of model-generated content. When dealing with generation tasks, RAG first retrieves the most relevant fragments of the input in a large amount of external data or knowledge base, and then inputs the retrieved information together with the original input into the large language model (LLM) as a prompt or context to guide the model to generate more accurate and rich answers.

[0087] In some other embodiments, an execution process of step S3 is as follows:

[0088] like Figure 5 As shown, the training question bank model M1: set the question type QT, including true or false questions QT1, multiple choice questions QT2, fill-in-the-blank questions QT3, and essay questions QT4. The detailed training processes are as follows:

[0089] True or False Questions QT1: In the retrieval phase, the model will retrieve the definition of true or false questions: it is necessary to judge whether the statement is true or false based on its content; in the generation phase, the model summarizes the characteristics of true or false questions based on the retrieved information: there are usually only two options (True / False), which test direct judgments on a fact or knowledge point.

[0090] Multiple-choice questions QT2: In the retrieval stage, the characteristics of multiple-choice questions are retrieved: including multiple options for one question, with only one or more correct answers; in the generation stage, the model generates a description of the multiple-choice questions: the multiple-choice questions are highly structured and test the ability to identify and select knowledge points.

[0091] Fill-in-the-blank questions QT3: In the retrieval stage, the form and usage of fill-in-the-blank questions are retrieved; in the generation stage, a summary is generated. Fill-in-the-blank questions require candidates to fill in the missing parts based on the context or prompts, testing the application and accuracy of knowledge.

[0092] Question and answer question QT4: In the retrieval stage, the model retrieves characteristic information about the question and answer: candidates are generally required to give extended answers based on the questions; in the generation stage, it is summarized that question and answer questions are usually used to test deep understanding and analytical ability, and the answers are flexible and there is no standard answer.

[0093] like Figure 6 As shown, a large amount of different types of question data are collected, including multiple-choice questions, short-answer questions, fill-in-the-blank questions, and true-or-false questions, so that the RAG model can associate the corresponding questions and answers for learning, and input the preprocessed knowledge points into the LLM model to generate a preliminary question bank model framework; through multiple iterations and feedback, the generation of question bank exercises can be improved.

[0094] Batch input of knowledge point content SV, through the question bank model M1, can get several types of question bank QD, stored in the vector database, again using the large model for tuning, using human language to correct continuity.

[0095] like Figure 7 As shown in the figure, the model is continuously optimized through user feedback (such as the accuracy of answering questions, the length of time to answer questions, user evaluation, etc.). Through the automatic evaluation of questions and the return of user feedback data, the RAG model can continuously fine-tune the generation algorithm to improve the quality and difficulty matching of questions; based on feedback, the model can enhance the control ability of certain question types (such as the quality of generating distractors for multiple-choice questions and the simplicity of answers for short-answer questions).

[0096] After obtaining the question bank QD, students are prevented from memorizing questions when questions are given in real time. The question-setting model M2 is used to generate dynamic questions based on the existing question bank QD. The capabilities of the model M2 are as follows: in each assessment, the numerical value or context of the question is different, the position of the answer options is disrupted, and the question type is dynamically switched. The corresponding answer requires a corresponding question-setting model, which enriches the variety of questions, saves question-setting time, and improves user experience.

[0097] S4. By performing semantic transfer on the questions, the questions that interact with the user each time are not exactly the same. According to the results of the interaction between the user and the questions, part of the teaching data is replayed, and a personalized review plan and learning suggestions are generated.

[0098] In this embodiment, the specific execution process of step S4 includes:

[0099] If the result of the interaction is that the user does not answer the question within the specified time or answers the question incorrectly, the teaching data will be rolled back to the location of the previous knowledge point and will be visually displayed to the user again.

[0100] Questions with incorrect answers as a result of user interaction are selected, and the knowledge points corresponding to the questions are taken as the user's weaknesses. The weaknesses are counted and a review plan and learning suggestions based on the weaknesses are provided to the user.

[0101] In some other embodiments, a specific execution process of step S4 is as follows:

[0102] After the administrator completes the creation of the question bank, it must undergo strict review to form the final question bank set Q = {Q1, Q2, Q3, ..., Q10}; this question bank covers multiple key knowledge points to ensure the comprehensiveness and coverage of the exam.

[0103] The system performs limited semantic transfer processing on each question in the question bank based on the Retrieval Augmented Generation model (RAG); this process ensures that the questions encountered by each student during the learning process will not be exactly the same, thereby improving the pertinence and personalization of the questions and avoiding repetitive learning.

[0104] The playback device of the present invention supports cross-platform operation and provides rich interfaces, and can be seamlessly integrated with third-party systems. After the user logs in through authorization, the learning performance and training results will be automatically reported to the third-party system through the interface, thereby realizing real-time sharing and feedback of learning results.

[0105] After logging in, the user starts watching the course. The system provides an intuitive user interface. After each knowledge point is played, the system automatically pauses and covers the player to display the answer content. For example, if it is a true or false question, the system will prompt the user to make a true or false judgment on a certain knowledge point. If the user answers incorrectly or fails to give a correct answer within the specified time, the system will automatically rewind to the previous video and play it again.

[0106] The question presentation method is diversified. Even for the same knowledge point, different users will see different question formats, which increases the challenge of learning. If a user gives incorrect answers for multiple times in a row, the system will ask the user whether to skip the current knowledge point or conduct an Internet search (for example, through academic databases, Wikipedia, Zhihu and other knowledge vertical websites) to obtain supplementary information on related knowledge points from external resources, and then re-summarize and present it to the user.

[0107] If the user successfully answers the question through online knowledge, the system will report the correct result to the question bank and update the model in real time, thereby optimizing the generation of subsequent questions and the accuracy of answering questions.

[0108] After all video content has been played, the system will automatically summarize the user's learning performance and generate a personalized review plan for the user's weak knowledge points to help the user conduct targeted learning and consolidation.

[0109] The system generates a comprehensive exam based on the preset question ratio. Users need to complete the exam to evaluate their learning effect. Based on the exam results, users can choose to review weak points or directly enter the completion process. The system will record the final results and provide follow-up learning suggestions.

[0110] Based on documents and training models, the big model is used to intelligently segment online long videos into short videos according to knowledge points, and then a summary of the segmented short videos is generated, and the corresponding question bank is generated based on the courseware and video understanding; after the user watches each key segment, the system automatically pops up the corresponding question to help reinforce learning; when encountering knowledge points that you don’t understand, you can directly select them in the system for intelligent retrieval and summary, saving students’ time; dividing long videos into short videos, adding a quick retrieval module and an independent question-setting module, increasing interactivity, and helping to increase students’ interest in learning, improve learning efficiency, and improve learning outcomes.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A large model-based online video training method, characterized in that: The method comprises: Construct teaching data and present it to users in a visual way; Vectorize the teaching data, divide the teaching data by the similarity between vectors, and add knowledge point positioning to the teaching data based on the division results; Generate a multi-question question bank based on the characteristics of different question types and teaching data with knowledge point positioning, and select some questions from the multi-question question bank to interact with users; By performing semantic transfer on the questions, the questions that users interact with are not exactly the same each time. Based on the results of the interaction between the user and the questions, part of the teaching data is replayed, and personalized review plans and learning suggestions are generated.

2. The online video training method based on a large model according to claim 1, characterized in that: The process of vectorizing the teaching data includes: Preset teaching documents and teaching videos, separate the teaching videos to obtain teaching audio, transcribe the teaching audio into target text and perform vectorization processing; wherein, the time nodes of the teaching audio, teaching video and target text are associated; The logical hierarchical structure in the teaching document is extracted and output in a structured format to obtain the teaching text and teaching pictures, and the teaching text and teaching pictures are vectorized respectively.

3. The online video training method based on a large model according to claim 1, characterized in that: The process of dividing the teaching data by the similarity between vectors and adding knowledge point positioning to the teaching data according to the division results includes: Calculate the cosine similarity between vectors to obtain the segmentation boundary of knowledge points, divide the teaching data once according to the distribution of knowledge points, and use recursive processing to divide the result of the first division twice to obtain multiple data segments that meet the teaching requirements; Add corresponding labels to each data segment, obtain the timestamp between two adjacent data segments, and add knowledge point locations in the teaching data based on the timestamp.

4. The online video training method based on a large model according to claim 2 is characterized in that: In the process of selecting some topics to interact with the user: According to the knowledge points corresponding to the questions, a priority is set for each question, the questions are sorted according to the priority, and questions with higher priorities are preferentially selected to interact with users; wherein the selection priorities are from high to low: teaching text, target text, teaching picture.

5. The online video training method based on a large model according to claim 4 is characterized in that: After setting the priority: Record the user's interaction traces and dynamically adjust the questions based on the interaction traces. For the same knowledge point, use different types of questions to interact with the user, and the question types in two consecutive interactions are different.

6. The online video training method based on a large model according to claim 1, characterized in that: The generation process of the multi-question type question bank includes: Use retrieval-enhanced generation technology to set different answer generation strategies for different question types, and input the answer generation strategies and teaching data into the large language model to generate corresponding questions and answers for different knowledge points in the teaching data.

7. The online video training method based on a large model according to claim 6, characterized in that: The answer generation strategy includes: If the question is a true or false question, the generation strategy is to judge the truth or falsehood of the statement based on its content, replace the content of the knowledge point with synonyms or heteronyms, and generate binary options; If the question is a multiple-choice question, the generation strategy is to give multiple answers based on the content of the statement, and at least one of the answers is the correct option corresponding to the knowledge point; If the question is a fill-in-the-blank question, the generation strategy is to fill in the missing parts according to the context or prompts, delete part of the content of the knowledge point, and record the deleted content as the answer; If the question type is a question-and-answer question, the answer given by the user will be recorded and wait for manual processing.

8. The online video training method based on a large model according to claim 1, characterized in that: The process of semantically migrating the topic includes: According to the vectorization processing results, the content with high repetition rate in the same knowledge point is selected as the keywords, the keywords are searched on the Internet, and the keywords are replaced with synonyms according to the search results; By screening out the contents with high repetition rate in different knowledge points as common words, the common words are removed from the keywords.

9. The online video training method based on a large model according to claim 1, characterized in that: The process of replaying part of the teaching data according to the interaction result between the user and the question includes: If the result of the interaction is that the user does not answer the question within the specified time or answers the question incorrectly, the teaching data will be rolled back to the location of the previous knowledge point and will be visually displayed to the user again.

10. The online video training method based on a large model according to claim 1, characterized in that: The process of generating a personalized review plan and learning suggestions includes: Questions with incorrect answers as a result of user interaction are selected, and the knowledge points corresponding to the questions are taken as the user's weaknesses. The weaknesses are counted and a review plan and learning suggestions based on the weaknesses are provided to the user.

Images