AI-based multi-person interactive science knowledge question and answer platform

Abstract

The present invention relates to the field of intelligent question-answering technology, and discloses an AI-based multi-person interactive popular science knowledge question-answering interactive platform, which incorporates multiple users into an interactive conversation space through a user identity recognition unit, and uses a data module construction unit to construct a multi-dimensional popular science knowledge data module. The user request receiving unit receives questions from multiple people, the question text processing unit pre-processes the questions, and the user intention recognition unit parses the deep semantics of the questions. The candidate answer retrieval unit retrieves candidate answers from the knowledge module, and the interactive answer generation unit combines the multi-person conversation context to screen, combine and adjust the candidate answers to form an interactive reply text. Finally, the user interaction processing unit realizes real-time interaction and question-answering among multiple people. The present invention uses AI technology to enhance the fun and interactivity of popular science, promote knowledge sharing and thinking collision, and effectively improve the efficiency of popular science and user learning experience.

Description

Technical Field

[0001] The present invention relates to the field of intelligent question-answering technology, and specifically to an AI-based multi-person interactive popular science knowledge question-answering interactive platform. Background Art

[0002] In today's era of rapid development of information technology, the dissemination and popularization of popular science knowledge is particularly important. Although traditional popular science methods, such as books, lectures, exhibitions, etc., can meet the public's popular science needs to a certain extent, they are limited by the single form and lack of interactivity, and it is difficult to fully stimulate the public's interest in learning and desire to explore. With the rise of Internet technology, online popular science platforms have gradually become a new channel for the dissemination of popular science knowledge. They deliver popular science information to users in various forms such as text, pictures, and videos, which improves the interest and accessibility of popular science.

[0003] Most of the existing online science popularization platforms adopt a "one-to-many" communication model, that is, the platform releases science popularization content, and users passively accept it, lacking an effective interactive mechanism. In this model, users can only passively accept the information provided by the platform, and it is difficult to conduct in-depth exploration and discussion based on their own interests or questions. At the same time, due to the lack of direct communication channels between users, an environment for knowledge sharing and collision of ideas cannot be formed, thus limiting the depth and breadth of science popularization effects.

[0004] Existing popular science question-and-answer systems can often only handle questions from a single user, and cannot support scenarios where multiple people can ask questions and interact at the same time. In a multi-person conversation space, the sharing of questions and answers, the continuity of follow-up questions, and the real-time interaction between users are all key factors that affect the effectiveness of popular science. However, existing systems are usually unable to effectively handle these complex scenarios, resulting in poor user experience and low efficiency in popular science. Summary of the invention

[0005] The purpose of the present invention is to provide an AI-based multi-person interactive science knowledge question and answer interactive platform to solve the problems raised in the above background technology.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solution: an AI-based multi-person interactive science knowledge question-and-answer interactive platform, the platform comprising:

[0007] A user identification unit, used to identify the identities of multiple users and add each user to the interactive session space;

[0008] Data module construction unit, used to pre-construct multi-dimensional popular science knowledge data modules and classify and label the knowledge;

[0009] A user request receiving unit is used to receive popular science knowledge question and answer requests from multiple users in a multi-person conversation space and obtain an initial question text;

[0010] A question text processing unit, connected to the user request receiving unit, for preprocessing the initial question text, including removing redundant information, word segmentation and part-of-speech tagging, to form a structured question statement;

[0011] The user intention recognition unit is used to input the structured question statement into the pre-trained user intention understanding model. The model is responsible for analyzing the deep semantics of the question statement, identifying the user's query intention and focus, and forming the intention recognition result;

[0012] A candidate answer retrieval unit, connected to the data module construction unit, is used to retrieve relevant information from the multi-dimensional popular science knowledge data module according to the intention recognition result to obtain multiple candidate answers;

[0013] An interactive answer generation unit is used to encapsulate the intention recognition result and multiple candidate answers according to a preset interactive template to obtain an interactive question text, further pass the interactive question text into a pre-trained multi-person interactive model, and use the natural language understanding and generation capabilities of the multi-person interactive model, combined with the conversation context of multiple users, to screen, combine and adjust multiple candidate answers to form an answer containing interactive questions and guiding prompts, and obtain an interactive reply text;

[0014] The user interaction processing unit is used to send the interactive reply text to the multi-person conversation space and receive the user's reply, so as to realize real-time interaction and question-answering among multiple people.

[0015] Preferably, the user intent understanding model is constructed using a long short-term memory network LSTM algorithm.

[0016] Preferably, the step of training the user intent understanding model includes:

[0017] Step 1: Collect and organize a large amount of popular science question and answer data, clean and annotate the data, and form a training data set;

[0018] Step 2: Perform word segmentation, part-of-speech tagging, and named entity recognition on the questions in the training dataset to extract text features;

[0019] Step 3: Build an LSTM network model, including input layer, LSTM layer, fully connected layer and output layer; the input layer receives the text feature vector, the LSTM layer processes the sequence data through memory units and gating mechanism, the fully connected layer maps the output of the LSTM layer to the intent category space, and the output layer uses the softmax function to output the intent probability distribution;

[0020] Step 4: Use the training data set to train the LSTM model, adjust the model parameters through the back propagation algorithm and optimizer, and minimize the loss function;

[0021] Step 5: Evaluate the trained LSTM model on the validation dataset and measure the performance of the model by calculating the accuracy, recall, and F1 value indicators;

[0022] Step 6: Adjust the model structure or hyperparameters based on the evaluation results, and repeat steps 3 to 5 until the model performance meets the expected requirements.

[0023] Preferably, in step 1, the method of cleaning the collected data includes:

[0024] Remove duplicates: Remove identical or highly similar duplicate questions by comparing the question text or question identifiers in the dataset;

[0025] Filter invalid data: delete data containing irrelevant information, garbled characters, illegal characters or incorrect formats;

[0026] Correcting erroneous data: Manually or automatically correcting obviously erroneous or inconsistent data;

[0027] Normalization processing: unify data formats, including date formats and number formats.

[0028] Preferably, in step 1, the cleaned data is labeled and a corresponding label or category is assigned to each piece of data; the specific steps include:

[0029] Intent labeling: assign one or more intent labels to each question based on the content of the question and the user's query intent;

[0030] Entity annotation: Identify key entities in the question and assign corresponding labels to these entities;

[0031] Answer annotation: For each question, provide the corresponding correct answer or answer range, and mark the type and source of the answer.

[0032] Preferably, the multi-person interaction model is constructed based on a universal large language model.

[0033] Preferably, the training step of the multi-person interaction model includes:

[0034] S1, model fine-tuning: Based on the general large language model, fine-tuning is performed using the cleaned and annotated multi-person conversation data to make the model suitable for multi-person conversation scenarios and popular science knowledge question-answering tasks;

[0035] S2, context understanding ability training: Design specific training tasks, including conversation context understanding, conversation state tracking, and user historical conversation content screening, to enhance the model's ability to understand and utilize context information in multi-person conversations;

[0036] S3, interactive strategy learning: Through reinforcement learning methods, the model is trained to learn effective interactive strategies of questioning, guidance and feedback.

[0037] Preferably, the implementation method of the preset interaction template includes: designing multiple types of interaction templates, including multiple-choice question templates, fill-in-the-blank question templates, true-or-false question templates and open-ended discussion templates, each template comprising a question display area, an answer option area, a user input area and a guiding prompt area.

[0038] Preferably, social functions are integrated into the preset interactive template, including likes, comments and sharing to social platforms.

[0039] Preferably, the data modules constructed by the data module construction unit include a popular science knowledge database module, a high-dimensional document vector database module and a knowledge graph module.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention uses AI technology to realize the function of multiple people asking questions and interacting at the same time, breaking the traditional "one-to-many" one-way communication mode. Users can not only receive popular science information, but also communicate and discuss in real time according to their own interests and questions, which significantly enhances the user's sense of participation and interactivity. Through the many-to-many communication method, it helps to form a good environment for knowledge sharing and thinking collision, and enhance the depth and breadth of popular science effects.

[0042] The present invention integrates natural language processing technology and user intention recognition model, can accurately analyze user questions and intentions, quickly retrieve relevant information from the preset multi-dimensional popular science knowledge data module, and generate accurate and personalized answers. Through the intelligent question-and-answer mechanism, the dissemination efficiency and accuracy of popular science knowledge are greatly improved, meeting the diverse learning needs of users. In a multi-person conversation space, real-time interaction and question-and-answer between users not only contribute to the growth of personal knowledge, but also promote the sharing and dissemination of knowledge within the community. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a work flow chart of the present invention;

[0044] Figure 2 Flowchart for building a user intent understanding model;

[0045] Figure 3 Flowchart for training the human interaction model. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] See also Figure 1-3 The present invention provides a technical solution: an AI-based multi-person interactive science knowledge question-and-answer platform, the platform comprising:

[0048] User identity unit: uses identity authentication technology, such as username and password login, third-party account login, etc. to identify the user's identity, and assigns a unique session identifier to each user, adding them to the multi-person session space. This unit maintains the user session status to ensure the user's continuous interaction in the session.

[0049] Data module construction unit: pre-construct a data module containing a wide range of popular science knowledge, covering natural science, technology, engineering, mathematics and other fields. Use classification algorithms and labeling systems to classify and label knowledge for subsequent rapid retrieval. The data modules constructed by the data module construction unit include a popular science knowledge database module, a high-dimensional document vector database module and a knowledge graph module: ① Popular science knowledge database module: This module stores a large amount of popular science knowledge, including knowledge points in multiple fields such as astronomy, biology, and physics. ② High-dimensional document vector database module: This module converts popular science articles or questions into high-dimensional vectors, and calculates the similarity between vectors to achieve accurate content recommendation and user query matching. ③ Knowledge graph module: This module constructs a graph structure of popular science knowledge and displays the correlation between knowledge points.

[0050] User request receiving unit: In a multi-person conversation space, it monitors and receives popular science question and answer requests from multiple users in real time. Through technologies such as WebSocket, it ensures the real-time transmission and reception of user requests and obtains the initial question text.

[0051] Question text processing unit: pre-processes the initial question text, including removing redundant information, word segmentation, part-of-speech tagging, etc. Natural language processing technology, such as NLTK tools, is used to convert the question text into a structured question statement.

[0052] User intention recognition unit: Build a user intention understanding model and perform semantic analysis on the structured question statement. The model recognizes the user's query intention and focus, and forms an intention recognition result.

[0053] Candidate answer retrieval unit: Based on the intention recognition results, relevant information is retrieved from the multi-dimensional popular science knowledge data module, and multiple candidate answers are obtained by combining knowledge classification and labels.

[0054] Interactive answer generation unit: Design a preset interactive template to encapsulate the intent recognition results and multiple candidate answers to form an interactive question text. Based on the pre-trained multi-person interactive model, combined with the conversation context of multiple users, the candidate answers are screened, combined and adjusted to generate answers containing interactive questions and guiding prompts to form interactive reply text. Integrate social functions in the preset interactive template, including likes, comments and sharing to social platforms.

[0055] User interaction processing unit: sends interactive reply text to the multi-person conversation space and receives user replies in real time. By processing user replies and updating the conversation context, continuous interaction and question-answering between multiple people is achieved.

[0056] The present invention will be further described below in conjunction with Examples 1 to 3:

[0057] Embodiment 1:

[0058] Reference Figure 2 This embodiment is used to describe the process of building and training a user intent understanding model. The user intent understanding model is built using the long short-term memory network (LSTM) algorithm, which aims to accurately parse the deep semantics of user questions and identify the user's query intent and focus. The following are the specific steps of model training:

[0059] Step 1: Data collection and cleaning, including:

[0060] Step 1.1, data collection: Collect a large amount of popular science Q&A data from popular science websites, forums, Q&A communities and other channels, including natural sciences, technology, engineering, mathematics and other fields.

[0061] Step 1.2, data cleaning:

[0062] Remove duplicates: Use hashing algorithms or text similarity calculations to compare question texts or question identifiers in the dataset and remove duplicate questions that are exactly the same or highly similar.

[0063] Filter invalid data: Use regular expressions and text processing tools to remove data that contains irrelevant information, garbled text, illegal characters, or format errors.

[0064] Correcting erroneous data: Manually or automatically correct obvious errors or inconsistencies in data, such as spelling errors and grammatical errors.

[0065] Normalization processing: Unify data formats, including date formats (such as YYYY-MM-DD), number formats (such as integers, floating point numbers), etc., to ensure data consistency and processability.

[0066] Step 1.3: Data labeling:

[0067] Intent labeling: Assign one or more intent labels to each question based on the content of the question and the user's query intent, such as "inquiry about physics principles", "chemical experiment steps", etc.

[0068] Entity labeling: Use named entity recognition (NER) technology to identify key entities in the question, such as names of people, places, scientific terms, etc., and assign corresponding labels to these entities.

[0069] Answer annotation: For each question, provide the corresponding correct answer or answer range, and mark the type of answer (such as text, picture, video) and source (such as authoritative science website, expert answer).

[0070] Step 2: Perform word segmentation, part-of-speech tagging, and named entity recognition on the questions in the training dataset to extract text features. Use the NLTK natural language processing tool to convert the question text into a word vector sequence as the input of the LSTM model.

[0071] Step 3: Build an LSTM network model. The model structure includes:

[0072] Input layer: receives text feature vectors, i.e. word vector sequences.

[0073] LSTM layer: processes sequence data through memory units and gating mechanisms (input gate, forget gate, output gate) to capture long-term dependencies in text. The LSTM unit is calculated as follows:

[0074] Forget Gate: ;

[0075] Input Gate: ;

[0076] Output Gate: ;

[0077] in, and denote the weight matrix and bias term respectively, Represents the sigmoid activation function.

[0078] Fully connected layer: maps the output of the LSTM layer to the intent category space, that is, the set of intent labels.

[0079] Output layer: The softmax function is used to output the intent probability distribution, indicating the possibility of each intent label.

[0080] Step 4: Use the training dataset to train the LSTM model, adjust the model parameters through the back-propagation algorithm and optimizer (such as Adam) to minimize the loss function. Set hyperparameters such as learning rate, batch size, and number of training rounds to control the training process.

[0081] Step 5: Evaluate the trained LSTM model on the validation data set and measure the performance of the model by calculating indicators such as accuracy, recall, and F1-Score. Accuracy indicates the proportion of correct predictions, recall indicates the proportion of correctly predicted positive samples, and F1-Score is the harmonic average of accuracy and recall.

[0082] Step 6: Adjust the model structure or hyperparameters based on the evaluation results, such as increasing the number of LSTM layers, adjusting the number of hidden units, changing the learning rate, etc. Repeat steps 3 to 5 and perform multiple iterations of training until the model performance meets the expected requirements, such as an accuracy rate of more than 90% and an F1 value of 0.9.

[0083] Here is a Python code example for building and training a user intent understanding model:

[0084]

[0085] In the sample code, a sequential model is built using the Sequential class.

[0086] Add an embedding layer to convert word indexes into word vectors. Here, the embedding dimension is set to 128.

[0087] Add a spatial dropout layer (SpatialDropout1D) to reduce overfitting.

[0088] Add an LSTM layer with 128 LSTM units and enable dropout and recurrent dropout to further reduce overfitting.

[0089] Add a fully connected layer (Dense) and use the ReLU activation function.

[0090] Add an output layer with the number of units equal to the number of labels and use the softmax activation function to get the probability distribution of each label.

[0091] Use sparse_categorical_crossentropy as the loss function because the labels are integer encoded. Use adam optimizer. Set the evaluation metric to accuracy. Use the fit function to train the model, specifying the training data, number of training rounds (epochs), batch size (batch_size), and validation set ratio. Use the predict function to predict the validation set and get the label probability distribution of each sample.

[0092] Use the argmax function to convert the probability distribution to the predicted label index. Calculate the accuracy, recall, and F1 score to evaluate the performance of the model.

[0093] Embodiment 2:

[0094] Reference Figure 3 This embodiment is used to describe a training method for a multi-person interaction model, and the specific steps include:

[0095] S1: Select a general large language model with excellent performance as the basic model. The model has been pre-trained on large-scale text data and has strong language understanding and generation capabilities. In order to make the model better adapt to the multi-person conversation scenario and the task of popular science question and answer, perform the following fine-tuning steps:

[0096] S101: Collect and clean multi-person conversation data, especially those involving popular science questions and answers. Data cleaning includes removing noise, correcting grammatical errors, and unifying formats. At the same time, the data is annotated, including dialogue roles, dialogue turns, question and answer pairs, and corresponding popular science knowledge points.

[0097] S102: Input the cleaned and annotated multi-person conversation data into the general large language model, and use fine-tuning technology to make the model learn the specific pattern of multi-person conversation and the specific task of popular science question and answer. During the fine-tuning process, by adjusting the model's learning rate, number of training rounds and other hyperparameters, and using appropriate optimization algorithms, the model can better adapt to new tasks and scenarios while maintaining its original language understanding capabilities.

[0098] S2: In order to enhance the model’s ability to understand and utilize contextual information in multi-person conversations, the following specific training tasks are designed:

[0099] S201: Conversation context understanding: Construct a training dataset containing multiple rounds of conversations, each of which contains context information. The training model understands and predicts the possible content or response of the next round of conversation based on the current conversation content. Through this task, the model can learn the coherence and logic of the conversation.

[0100] S202: Dialogue state tracking. In a multi-person conversation, the dialogue state may change as new information is added. Design training tasks that require the model to update the dialogue state in real time based on the dialogue content and accurately track key information such as the role, intention, and topic of each participant.

[0101] S203: Filtering the content of user historical conversations. In a multi-person conversation, a user may mention the same topic or question multiple times. The training model filters out the most relevant information from the user's historical conversation content to answer the user's questions more accurately or provide relevant popular science knowledge.

[0102] S3: In order to improve the interactive ability of the model in multi-person conversations, the reinforcement learning method is used to train the model to learn effective questioning, guidance and feedback interactive strategies. The specific implementation is as follows:

[0103] Define reward function: Define an appropriate reward function based on the model's performance in multi-person conversations. For example, when the model successfully guides users to ask valuable questions, give positive rewards; when the model fails to accurately understand the user's intentions or provides incorrect scientific knowledge, give negative rewards.

[0104] Strategy exploration and optimization: Use reinforcement learning algorithms (such as Q-learning, Deep Q-Network, etc.) to allow the model to explore strategies in a multi-person conversation environment. The model learns how to maximize the reward function by constantly trying different ways of asking questions, guiding, and providing feedback. During the exploration process, the model gradually forms a set of effective interaction strategies.

[0105] Strategy evaluation and update: Regularly evaluate the model's interaction strategy, and update and optimize the strategy based on the evaluation results. Evaluation criteria may include user satisfaction, conversation efficiency, accuracy of popular science knowledge transfer, etc.

[0106] Embodiment 3:

[0107] In order to meet the needs of different popular science content and user preferences, the system has designed a variety of interactive templates, including multiple-choice question templates, fill-in-the-blank question templates, true-or-false question templates, and open discussion templates. Each template is designed with the following areas to provide a rich interactive experience:

[0108] Question display area: used to clearly display popular science questions or topics to ensure that users can accurately understand the core of the question. For example, in the multiple-choice question template, the question display area will present a specific popular science question, such as "Which planet is the largest in the solar system?"

[0109] Answer option area: For multiple-choice and true-false question templates, multiple options are provided for users to choose from. These options are designed based on the correctness of popular science knowledge to guide users to think and judge. In the fill-in-the-blank question template, this area is used to display the blank part that users need to fill in.

[0110] User Input Area: Provide space for users to input answers or opinions. In the open discussion template, users are encouraged to express their thoughts and opinions freely.

[0111] Guidance hint area: In order to reduce the cognitive burden of users, the system sets up a guidance hint area to provide necessary clues to solve the problem or guide users to think more deeply. For example, in the case of difficult or complex problems, this area will provide some keywords or brief explanations of related concepts.

[0112] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0113] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based multi-person interactive science knowledge question-and-answer platform, characterized by: The platform includes: A user identification unit, used to identify the identities of multiple users and add each user to the interactive session space; Data module construction unit, used to pre-construct multi-dimensional popular science knowledge data modules and classify and label the knowledge; A user request receiving unit is used to receive popular science knowledge question and answer requests from multiple users in a multi-person conversation space and obtain an initial question text; A question text processing unit, connected to the user request receiving unit, for preprocessing the initial question text, including removing redundant information, word segmentation and part-of-speech tagging, to form a structured question statement; The user intention recognition unit is used to input the structured question statement into the pre-trained user intention understanding model. The model is responsible for analyzing the deep semantics of the question statement, identifying the user's query intention and focus, and forming the intention recognition result; A candidate answer retrieval unit, connected to the data module construction unit, is used to retrieve relevant information from the multi-dimensional popular science knowledge data module according to the intention recognition result to obtain multiple candidate answers; An interactive answer generation unit is used to encapsulate the intention recognition result and multiple candidate answers according to a preset interactive template to obtain an interactive question text, further pass the interactive question text into a pre-trained multi-person interactive model, and use the natural language understanding and generation capabilities of the multi-person interactive model, combined with the conversation context of multiple users, to screen, combine and adjust multiple candidate answers to form an answer containing interactive questions and guiding prompts, and obtain an interactive reply text; The user interaction processing unit is used to send the interactive reply text to the multi-person conversation space and receive the user's reply, so as to realize real-time interaction and question-answering among multiple people.

2. The AI-based multi-person interactive science knowledge question-and-answer interactive platform according to claim 1 is characterized in that: The user intention understanding model is constructed using the long short-term memory network LSTM algorithm.

3. The AI-based multi-person interactive science knowledge question-and-answer interactive platform according to claim 2 is characterized in that: The steps to train a user intent understanding model include: Step 1: Collect and organize a large amount of popular science question and answer data, clean and annotate the data, and form a training data set; Step 2: Perform word segmentation, part-of-speech tagging, and named entity recognition on the questions in the training dataset to extract text features; Step 3: Build an LSTM network model, including input layer, LSTM layer, fully connected layer and output layer; the input layer receives the text feature vector, the LSTM layer processes the sequence data through memory units and gating mechanism, the fully connected layer maps the output of the LSTM layer to the intent category space, and the output layer uses the softmax function to output the intent probability distribution; Step 4: Use the training data set to train the LSTM model, adjust the model parameters through the back propagation algorithm and optimizer, and minimize the loss function; Step 5: Evaluate the trained LSTM model on the validation dataset and measure the performance of the model by calculating the accuracy, recall, and F1 value indicators; Step 6: Adjust the model structure or hyperparameters based on the evaluation results, and repeat steps 3 to 5 until the model performance meets the expected requirements.

4. The AI-based multi-person interactive science knowledge question-and-answer interactive platform according to claim 3 is characterized in that: In step 1, the method of cleaning the collected data includes: Remove duplicates: Remove identical or highly similar duplicate questions by comparing the question text or question identifiers in the dataset; Filter invalid data: delete data containing irrelevant information, garbled characters, illegal characters or incorrect formats; Correcting erroneous data: Manually or automatically correcting obviously erroneous or inconsistent data; Normalization processing: unify data formats, including date formats and number formats.

5. The AI-based multi-person interactive science knowledge question-and-answer interactive platform according to claim 4 is characterized in that: In step 1, the cleaned data is labeled and a corresponding label or category is assigned to each piece of data; the specific steps include: Intent labeling: assign one or more intent labels to each question based on the content of the question and the user's query intent; Entity annotation: Identify key entities in the question and assign corresponding labels to these entities; Answer annotation: For each question, provide the corresponding correct answer or answer range, and mark the type and source of the answer.

6. The AI-based multi-person interactive science knowledge question-and-answer interactive platform according to claim 1 is characterized by: The multi-person interaction model is constructed based on a general large language model.

7. The AI-based multi-person interactive science knowledge question-and-answer interactive platform according to claim 6, characterized in that: The training steps of the multi-person interaction model include: S1, model fine-tuning: Based on the general large language model, fine-tuning is performed using the cleaned and annotated multi-person conversation data to make the model suitable for multi-person conversation scenarios and popular science knowledge question-answering tasks; S2, context understanding ability training: Design specific training tasks, including conversation context understanding, conversation state tracking, and user historical conversation content screening, to enhance the model's ability to understand and utilize context information in multi-person conversations; S3, interactive strategy learning: Through reinforcement learning methods, the model is trained to learn effective interactive strategies of questioning, guidance and feedback.

8. The AI-based multi-person interactive science knowledge question-and-answer interactive platform according to claim 1, characterized in that: The implementation method of the preset interaction template includes: designing multiple types of interaction templates, including multiple-choice question templates, fill-in-the-blank question templates, true-or-false question templates and open discussion templates, each template comprising a question display area, an answer option area, a user input area and a guiding prompt area.

9. The AI-based multi-person interactive science knowledge question-and-answer interactive platform according to claim 8, characterized in that: Integrate social features into preset interactive templates, including likes, comments, and sharing to social platforms.

10. The AI-based multi-person interactive science knowledge question-and-answer platform according to claim 1, characterized in that: The data modules constructed by the data module construction unit include a popular science knowledge database module, a high-dimensional document vector database module and a knowledge graph module.

Images