Intelligent question answering method and device for preschool education and storage medium

Abstract

The invention relates to the technical field of intelligent questioning and answering, and particularly relates to a preschool education intelligent questioning and answering method and device and a storage medium, and the method comprises the steps: constructing a knowledge base fusing user-child dynamic portraits; after a question proposed by a user side is received, firstly, a state perception model is called to analyze the intention and emotion of the question, and an integrated portrait is formed in combination with historical portrait information; then, performing enhanced context retrieval based on the integrated portrait so as to accurately recall related documents; then, obtaining reliable evidence fragments from the recalled documents by adopting an extraction type model; and finally, fusing the original question by a personalized answer generation model, integrating portrait information and all evidence fragments, synthesizing a section of personalized and contextualized final answer, and outputting the personalized and contextualized final answer. Through the hybrid processing architecture, the professionalism, pertinence and the like of answering are remarkably improved, and the application effect in scenes such as homework and garden cooperation is optimized.

Description

Technical Field

[0001] This invention relates to the field of intelligent question-and-answer technology, and in particular to an intelligent question-and-answer method, device and storage medium for preschool education. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent question-answering systems have been widely used in many industries. They can automatically answer questions posed by users, greatly improving the efficiency of information acquisition. In the field of education, intelligent question-answering systems have also shown great potential, assisting teachers in answering questions and guiding students' learning.

[0003] However, the field of preschool education has its unique characteristics and complexities. Questions raised by parents or teachers are often not simple knowledge queries, but rather comprehensive requests for help encompassing complex situations, children's specific behaviors, and the questioner's own anxiety and confusion. Existing general-purpose intelligent question-answering systems or traditional knowledge graph-based and FAQ-based question-answering robots have significant shortcomings in handling such questions. On the one hand, they lack in-depth professional knowledge of preschool education, making it difficult to provide scientific and rigorous answers; on the other hand, they cannot perceive and understand the deeper intentions and emotional states behind users' questions, nor do they incorporate personalized information such as the child's age and developmental characteristics. This results in answers that are often "one-size-fits-all" templates, lacking specificity and empathy, failing to truly resolve users' actual confusion, and may even have negative consequences due to inappropriate advice. Therefore, how to construct a system that can deeply understand the complex questions in preschool education scenarios and, in conjunction with the dynamic situations of users and children, provide high-quality answers that are professional, personalized, and contextualized is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] The technical problem to be solved by this application is to provide a preschool education intelligent question-answering method, device and storage medium, which aims to overcome the problems of insufficient professionalism, lack of personalization and context awareness in the answers of existing question-answering systems.

[0005] To achieve the above objectives, according to one aspect of the present invention, a preschool education intelligent question-answering method is provided, comprising: Construct a knowledge base containing a user-child dynamic profile database; Receive questions from users; The state-aware model is invoked to parse the question, obtain the current context information of the question, and combine it with the corresponding historical profile information in the user-child dynamic profile database to form integrated profile information; Based on the integrated profile information, contextual documents related to the question are retrieved from the knowledge base; Based on the context document, at least one piece of evidence is extracted using an extractive model. Based on the question, the integrated profile information, and at least one piece of evidence, a generative model is used to synthesize a personalized and contextualized final answer. The final answer will be output to the user.

[0006] As a preferred embodiment of the present invention, the state-aware model parses the problem and obtains the current context information of the problem, including: Identify the questioner's intent and the user's emotions; and Extract key entity information from the problem.

[0007] As a preferred embodiment of the present invention, the formation of the integrated portrait information further includes: Update the historical profile information in the user-child dynamic profile database based on the current context information.

[0008] As a preferred embodiment of the present invention, the step of retrieving and recalling context documents related to the question from the knowledge base based on the integrated profile information includes: Keyword retrieval is performed based on the keywords in the question and the metadata in the integrated profile information; The problem is converted into a query vector, and vector similarity retrieval is performed. The context document is obtained by combining the results of the keyword retrieval and the vector similarity retrieval.

[0009] As a preferred embodiment of the present invention, the step of extracting at least one piece of evidence based on the context document using an extractive model includes: For each context document recalled, calculate the probability that it contains the answer and the probability that each text fragment therein is an answer fragment, and form a confidence score; Text segments with confidence scores greater than a preset threshold are selected as the evidence segments.

[0010] As a preferred embodiment of the present invention, the step of synthesizing a personalized and contextualized final answer using a generative model based on the question, the integrated profile information, and the at least one piece of evidence includes: The question, the integrated profile information, and the at least one piece of evidence are combined to form a structured prompt; The structured prompts are input into the generative model to generate the final answer.

[0011] As a preferred embodiment of the present invention, before outputting the final answer to the user terminal, the method further includes: The final answer is validated; the validation includes calculating the final confidence level of the final answer. When the final confidence level is greater than the preset output threshold, the final answer is output; otherwise, the preset guiding template is output.

[0012] According to another aspect of the present invention, a preschool education intelligent question-and-answer device is provided, comprising: The knowledge base construction module is used to build a knowledge base, which contains a user-child dynamic profile database. The question receiving module is used to receive questions submitted by users. The retrieval and processing module is used to invoke a state-aware model to parse the question, obtain the current context information of the question, and combine it with the corresponding historical profile information in the user-child dynamic profile database to form integrated profile information; based on the integrated profile information, retrieve context documents related to the question from the knowledge base; based on the context documents, extract at least one piece of evidence using an extraction model; and based on the question, the integrated profile information, and the at least one piece of evidence, synthesize a personalized and contextualized final answer using a generative model. The answer output module is used to output the final answer to the user.

[0013] As a preferred embodiment of the present invention, the retrieval processing module is specifically used for: By identifying the questioner's intent and the user's emotions, and extracting key entity information from the question, the current context information can be obtained; Keyword retrieval is performed based on the keywords in the question and the metadata in the integrated profile information. The question is then converted into a query vector for vector similarity retrieval. The two retrieval results are then fused to obtain the context document. The question, the integrated profile information, and the at least one piece of evidence are constructed into a structured prompt, and the structured prompt is input into the generative model to generate the final answer.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the preceding claims.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This application, upon receiving a user's question, first uses a state-aware model to analyze the contextual information such as intent and emotion within the question. This information is then combined with a pre-built user-child dynamic profile database to form a complete profile integrating the current state and historical background. Subsequently, this enhanced integrated profile information guides the knowledge base retrieval process, thereby more accurately recalling highly relevant professional knowledge context. Based on this, instead of directly generating an answer, a reliably relevant evidence fragment is extracted from multiple recalled documents using an extractive model. This evidence, along with the question and the integrated profile information, is then input into a generative model to synthesize a personalized, contextualized answer that is faithful to professional knowledge while fully considering the child's individual situation and the parent's current state. This method ensures that the final output answer is not only based on sound professional content but also closely aligns with the user's specific needs and emotional state in its expression and emphasis of suggestions. This significantly improves the application effectiveness, professionalism, and user trust of intelligent question answering in preschool education scenarios such as home-school collaboration. Attached Figure Description

[0016] Figure 1 This is a flowchart of a preferred embodiment of the present invention; Figure 2 This is a module architecture diagram of a preferred embodiment of the present invention; Illustration: 10. Knowledge base construction module; 11. Question receiving module; 12. Retrieval and processing module; 13. Answer output module. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this embodiment can be combined with each other. The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] Reference Figure 1 The preferred embodiment of the invention provides an intelligent question-answering method for preschool education, including: constructing a knowledge base that integrates user profiles, receiving questions raised by users, retrieving highly relevant contexts through state awareness and profile enhancement, and processing questions and generating personalized and contextualized answers using a hybrid mode that combines extractive reading comprehension and generative models, and finally outputting the answers to users.

[0019] The constructed knowledge base includes: a document library formed by data such as course documents, child development guidelines, and activity plans uploaded by kindergartens, childcare institutions, or related educational organizations; an annotation database formed by manually annotating historical question-and-answer data and educational cases through an annotation system optimized for the characteristics of the preschool education field; a model library formed by building and training models, including core reading comprehension models, state perception models (including intention and emotion recognition, key information extraction), personalized answer generation models, and their pre-training parameters; and a user-child dynamic profile database for recording and iterating dynamic information of each family (user).

[0020] When processing user questions, the system adopts a hybrid processing architecture: first, the user input is parsed by a state-aware model and combined with a dynamic profile; then, the profile-enhanced retrieval module accurately locates the knowledge; next, the reading comprehension model extracts multi-source factual evidence; and finally, the personalized answer generation model integrates all the information to synthesize the final answer.

[0021] In this embodiment, a document library is constructed from document data uploaded by educational institutions. An education-related annotation database is formed through manual annotation and an annotation system. A model is built, and a knowledge base is synthesized from the model library formed by model training. When a user asks a question, a hybrid retrieval method combining keywords and vectors is used. A reading comprehension model is selected, and the corresponding answer content in the document is displayed through semantic analysis and understanding.

[0022] Reference Figure 2 A preferred embodiment of the present invention also provides a preschool education intelligent question-and-answer device, comprising: Knowledge Base Construction Module 10: Used to build the knowledge base required by the intelligent question answering system, which integrates dynamic profiles; Question Receiving Module 11: Used to receive questions raised by users on the client side; Retrieval Processing Module 12: Used for in-depth processing of user questions and generation of answers. Its internal logic has been restructured into a complete process that includes state awareness, profile-enhanced retrieval, multi-evidence extraction, and personalized synthesis. Answer output module 13: Connected to the retrieval processing module 12, it is used to output the final generated personalized and contextualized answers to the user terminal.

[0023] The knowledge base constructed by the knowledge base construction module 10 includes a document library based on preschool education institution document data, an annotation database of education-related knowledge formed by manual annotation through an annotation system, a model library formed by building and training models, and a user-child dynamic profile database.

[0024] The document library is constructed as follows: all unstructured documents (such as PDFs and Word documents) are uniformly converted to plain text format and cleaned (headers, footers, and invalid characters are removed). Next, a recursive character text segmenter is used, with paragraphs as the primary boundary, setting `chunk_size=512` and `chunk_overlap=100`, to segment long documents into semantically coherent knowledge fragments. Each fragment uses a Chinese sentence vector model such as M3E-base to calculate its vector representation, which, along with the original text and metadata (source, chapter, applicable age group, etc.), is stored in an index database combining Elasticsearch and FAISS. Elasticsearch handles keyword indexing and metadata filtering, while FAISS handles efficient vector similarity retrieval.

[0025] The annotation database is built using the open-source annotation platform Docano, with pre-defined annotation task templates. For question-answer pairs, the annotation template includes "Question", "Context", "Answer_text", and "Answer_start". For state awareness, the annotation template includes "Text", "Intent" (pre-defined tags: behavioral intervention, ability development, knowledge inquiry, emotional help), "Emotion" (pre-defined tags: anxiety, confusion, joy, neutral), and "Entities" (using the BIO annotation method to annotate the child's age, specific behavior, and context).

[0026] The user-child dynamic profile database is stored using MongoDB, a NoSQL database. Its data structure is in JSON format. Each child (with the child ID as the primary key) has a document containing static information (such as date of birth and gender) and dynamic information (such as "developmental concerns" tags accumulated through question-and-answer interactions, recent emotional state, interaction history, and records of educational activities recommended by the system).

[0027] It includes a text classification model for recognizing parents' questioning intentions (such as behavioral intervention and skills development) and emotions (such as anxiety and confusion), and a named entity recognition (NER) model for extracting entities such as children's age, specific behaviors, and scenarios. Both models are fine-tuned based on a Chinese pre-trained model (MacBERT-based). The classification model is followed by a fully connected layer and a softmax activation function after its [CLS] output, while the NER model is followed by a conditional random field (CRF) layer after the output of each token to improve the accuracy of entity boundary recognition.

[0028] The appropriate model combination selected by the retrieval processing module 12 is a cascaded hybrid model architecture. The first level is a state-aware model, the second level is a reading comprehension model based on RoBERTa-wwm-ext-large (as an evidence extractor), and the third level is a personalized answer generation model (selecting a Chinese large language model that has been fine-tuned by instructions; in this embodiment, Qwen-7B-Chat is used).

[0029] The problem-solving steps are as follows: S1.1: State Awareness and Profile Integration: After receiving the user's question Q, the system first calls the state-aware model in the model library for parsing, obtaining a structured context object. ; .

[0030] At the same time, the user-child dynamic profile database is queried based on the user ID to obtain the child's historical profile. The currently parsed and The profile database was integrated and updated. The specific update strategy was as follows: for tags such as "development focus", a weighted cumulative method was used, and newly identified entities would have their corresponding tag weights increased; for "emotional state", a sliding time window averaging method was used to record and smooth the emotional values ​​of the three most recent interactions; interaction records were directly appended.

[0031] S1.2: Enhanced Profile Contextual Retrieval: Utilizing the question Q and integrated profile information, an enhanced, weighted hybrid query is constructed. This query consists of two paths: The first path uses Elasticsearch for keyword and metadata queries, and its query body (Query DSL) structure is as follows: { "query": { "bool": { "must": [{"match": {"content": "The core keywords in the user's question"}}], "should": [{"match": {"title": "Core keywords in user questions"}}], "filter": [{"term": {"age_group": "age group in the profile"}}] } } } The second approach involves converting the user's question Q into a vector and performing a Top-N (e.g., N=50) vector similarity retrieval in FAISS. Finally, the results from both approaches are fused using the Reciprocal Rank Fusion (RRF) algorithm to obtain the final Top-K (e.g., K=5) highly relevant document segments. .

[0032] S1.3: Multi-source evidence extraction: Segmenting long documents into paragraphs It performs input preparation, feature extraction, and model inference, but the goal is not to find a single best answer, but rather to select from all highly relevant paragraphs. Extract all possible pieces of evidence. Each piece of evidence Includes a confidence score :

[0033] in, It is a paragraph The probability of including the answer (obtained by activating a binary logits output by the reading comprehension model with a Sigmoid function). yes The probability of a fragment being an answer (calculated from the joint probability of the start and end positions of the answer's logits). Filtering out... Greater than the threshold (like This value is determined by evaluating all evidence using F1 scores on the validation set.

[0034] S1.4: Personalized Answer Synthesis: This involves calling a personalized answer generation model from the model library. The input to this model is a carefully designed structured prompt containing all the information from the previous steps: .

[0035] The specific implementation of this function is to generate a Markdown-formatted string with the following structure: ### Task Description You are an early childhood education expert. Based on the following information, please generate a professional, empathetic, and personalized response for a parent.

[0036] ### Parental Issues {Q} ### Parent Status and Child Information - **Parents' Emotions**: {C_context.Emotion} - **Main Intent**: {C_context.Intent} - **Child's Age**: {P_profile.Age} - **Behavior involved:** {C_context.Entities.Behavior} - **Historical Focus**: {P_profile.FocusTags} ### Related professional knowledge reference When generating your answer, please strictly refer to the following evidence and do not fabricate anything out of thin air: - **Evidence 1 (Confidence level: {Score_1})**: {Ev_1} - **Evidence 2 (Confidence level: {Score_2})**: {Ev_2} ... ### Your answer Based on this input, the generative model synthesizes a completely new, fluent, and personalized response. During generation, a sampling strategy with a low temperature coefficient (e.g., Temperature=0.7) is used to ensure the professionalism and accuracy of the answers.

[0037] S1.5: Answer Validation and Output Decision: For the generated answer... Security and reliability verification is performed. First, a security scan is conducted using a pre-trained text security classifier (which identifies abusive, violent, and inappropriate content). A final confidence score is then calculated. :

[0038] in, It is the average log probability of the token generated by the model. It is the semantic similarity between the generated answer and the source evidence set (the specific calculation method is: ...). And every piece of evidence Encode using sentence vector model, calculate With all (the maximum value of the cosine similarity) It is the balance coefficient (in this embodiment) (Focusing on the fidelity of the answer to the evidence). If Greater than the threshold (In this embodiment) Then the output will be... Otherwise, output a preset, guiding template and prompt the user to contact a human expert.

[0039] The training process for each model in the model library is as follows: S2.1: Train a self-attention-based language model using massive amounts of unstructured text data (such as children's picture books, parenting encyclopedias, and educational WeChat articles, with a total data volume of more than 50GB); S2.2: The pre-trained language model is used as an encoder and incorporated into the model framework designed for the reading comprehension task; then, the reading comprehension model is trained using a large amount of general domain reading comprehension labeled data (this embodiment uses public datasets such as CMRC 2018 and DRCD); S2.3: Fine-tune the reading comprehension model using a small amount of annotated reading comprehension data specific to early childhood education (such as educational question-and-answer pairs with correct answers, teaching case analyses, etc.).

[0040] S2.4: State-Aware Model Training: By collecting historical question-and-answer data and having it manually labeled by educational experts (labeling intent, emotion, and key entities), a text classification model and a sequence labeling (NER) model were fine-tuned based on a pre-trained language model to construct a state-aware model. Five-fold cross-validation was used during training to enhance the model's generalization ability, and Focal Loss was used to handle potential label imbalance issues (e.g., there were far more samples of the emotion "anxiety" than "joy").

[0041] S2.5: Personalized Answer Generation Model Training: A high-quality (input Prompt -> output answer) training dataset is constructed manually or semi-automatically. Specifically, the process from S1.1 to S1.3 is used to process a batch of seed questions, automatically generating preliminary prompts and evidence. Then, educational experts write or refine the final ideal answers, forming thousands of high-quality instruction-response pairs. Using this dataset, a large pre-trained generative model (Qwen) is efficiently fine-tuned using LoRA, enabling it to learn to generate personalized answers that conform to expert tone and early childhood education principles based on structured input.

[0042] In this embodiment, four modules—a knowledge base construction module 10, a question receiving module 11, a retrieval and processing module 12, and an answer output module 13—are used to provide intelligent answers to questions from users (educators or parents).

[0043] The workflow of this invention will be described in detail using a specific scenario: Take the application of the home-school co-education platform as an example.

[0044] Step 1: Knowledge Base Construction and Preparation In the platform's backend, the system pre-builds an enhanced knowledge base specifically for preschool education through the knowledge base construction module 10. This knowledge base includes: Document Repository: This repository contains documents uploaded by kindergartens and childcare institutions, including the "Guidelines for Child Development," the "Handbook for Observing and Analyzing Children's Behavior," various curriculum documents, activity plans, and accumulated home-school communication records (anonymized over the years). All documents are cleaned, segmented, and vectorized using the M3E-base model before being stored in a hybrid Elasticsearch+FAISS index.

[0045] Annotated database: Structured data extracted from historical high-quality question-and-answer pairs and educational case analyses through manual annotation and the use of the Docano system, which is optimized or configured for the characteristics of the preschool education field.

[0046] Model Library (Enhanced): Includes a pre-trained RoBERTa-wwm-ext-large reading comprehension model, a state-aware model fine-tuned based on MacBERT, and a LoRA-tuned Qwen-7B-Chat personalized answer generation model.

[0047] User-Child Dynamic Profile Database: An initial profile record is created for each family on the platform and stored in MongoDB.

[0048] Step Two: User Questions and Responses A parent (user) entered a question (Q) into the question receiving module 11 through the online consultation interface of the parent-teacher communication platform: "My child, Xiaoming, has recently been unwilling to share his new toys with other children at kindergarten. He cries whenever the teacher mentions it. What should I do?" Step 3: Problem Retrieval and Processing After receiving question Q, the retrieval and processing module 12 performs the following operations: S1.1 (State Awareness and Profile Integration): The state awareness model analyzes Q, outputting the intent as "behavioral intervention," the emotion as "anxiety," and the entities as "unwilling to share," "crying," and "kindergarten." The system queries the user ID, retrieves the profile of "Xiaoming," and learns that his age is 3 years and 2 months. This information, along with the results of this analysis, is updated to his dynamic profile, increasing the weight of the "social-sharing" tag in "developmental focus."

[0049] S1.2 (Enhanced Image Search): The system constructs an enhanced hybrid query, strongly matching "share" in Elasticsearch and filtering metadata for "age_group: 3-4 years old," while simultaneously performing vector retrieval in FAISS. Through RRF fusion, relevant paragraphs from the "Guidelines for Child Development" regarding "the development of property rights awareness in 3-4 year old children" are accurately retrieved.

[0050] S1.3 (Multi-source evidence extraction): The reading comprehension model extracts multiple evidence fragments from the recalled paragraphs, and the confidence level of each fragment is higher than the threshold of 0.65, such as: "3-4 years old is the critical period for the development of children's awareness of property rights...", "For children with intense emotions, their emotions should be soothed first...", and "You can try to guide sharing through role-playing games...".

[0051] S1.4 (Personalized Answer Synthesis): The personalized answer generation model receives a Markdown-formatted structured input containing the original question, parent's emotions, child's age, and all evidence fragments, and synthesizes a completely new, personalized answer.

[0052] Step 4: Output the answer The answer verification security scan passed S1.5, and the final confidence level was [not specified]. After the confidence level is met, the answer output module 13 will output the final answer A_final to the parent interface on the user's end. "Hello Xiaoming's mother, I see you are troubled by your child's unwillingness to share toys. Please don't be too anxious. In fact, it is very normal for children around 3 years old like Xiaoming to be unwilling to share. This is a manifestation of their developing self-awareness and sense of ownership."

[0053] When he cries at kindergarten, the first thing we should do is understand and accept his emotions. We can hug him and tell him, "This is your new toy. You like it very much and don't want to share it with other children. Mommy understands." Forcing him to share will only make him feel more insecure.

[0054] At home, you can try some fun ways to guide him, such as playing role-playing games like "opening a shop" or "the little host entertaining guests," so that he can experience the joy of sharing. At the same time, encourage and praise him for every small act of sharing.

[0055] This will take time. Please keep in touch with the teacher and give Xiaoming some time and patience. We hope these suggestions are helpful! If the confidence level of the final answer is below the threshold, the system will output a prompt, such as "I'm sorry, regarding your child's situation, I have found some professional principles, but they are not enough to form a complete recommendation. I suggest you try to communicate more in detail with the class teacher." Through the above process, this invention can intelligently and accurately synthesize relevant educational suggestions from a professional knowledge base to address specific educational concerns raised by parents, effectively reducing the pressure on teachers to provide immediate answers and improving the efficiency and quality of home-school collaboration.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A preschool education intelligent question-and-answer method, characterized in that, include: Construct a knowledge base containing a user-child dynamic profile database; Receive questions from users; The state-aware model is invoked to parse the question, obtain the current context information of the question, and combine it with the corresponding historical profile information in the user-child dynamic profile database to form integrated profile information; Based on the integrated profile information, contextual documents related to the question are retrieved from the knowledge base; Based on the context document, at least one piece of evidence is extracted using an extractive model. Based on the question, the integrated profile information, and at least one piece of evidence, a generative model is used to synthesize a personalized and contextualized final answer. The final answer will be output to the user.

2. The method according to claim 1, characterized in that, The state-aware model parses the problem and obtains the current context information of the problem, including: Identify the questioner's intent and the user's emotions; and Extract key entity information from the problem.

3. The method according to claim 2, characterized in that, The formation of the integrated portrait information also includes: Update the historical profile information in the user-child dynamic profile database based on the current context information.

4. The method according to claim 1, characterized in that, The step of retrieving and recalling context documents related to the question from the knowledge base based on the integrated profile information includes: Keyword retrieval is performed based on the keywords in the question and the metadata in the integrated profile information; The problem is converted into a query vector, and vector similarity retrieval is performed. The context document is obtained by combining the results of the keyword retrieval and the vector similarity retrieval.

5. The method according to claim 1, characterized in that, The step of extracting at least one piece of evidence based on the context document using an extractive model includes: For each context document recalled, calculate the probability that it contains the answer and the probability that each text fragment therein is an answer fragment, and form a confidence score; Text segments with confidence scores greater than a preset threshold are selected as the evidence segments.

6. The method according to claim 1, characterized in that, The process of synthesizing a personalized, contextualized final answer using a generative model based on the question, the integrated profile information, and at least one piece of evidence includes: The question, the integrated profile information, and the at least one piece of evidence are combined to form a structured prompt; The structured prompts are input into the generative model to generate the final answer.

7. The method according to claim 1, characterized in that, Before outputting the final answer to the user, the following steps are also included: The final answer is validated; the validation includes calculating the final confidence level of the final answer. When the final confidence level is greater than the preset output threshold, the final answer is output; otherwise, the preset guiding template is output.

8. A preschool education intelligent question-and-answer device, characterized in that, include: The knowledge base construction module (10) is used to construct a knowledge base, which contains a user-child dynamic profile database; The question receiving module (11) is used to receive questions raised by users. The retrieval processing module (12) is used to call the state-aware model to parse the question, obtain the current context information of the question, and combine it with the corresponding historical profile information in the user-child dynamic profile database to form integrated profile information; Based on the integrated profile information, contextual documents related to the question are retrieved from the knowledge base; based on the contextual documents, at least one piece of evidence is extracted using an extraction model; and based on the question, the integrated profile information, and the at least one piece of evidence, a generative model is used to synthesize a personalized and contextualized final answer. The answer output module (13) is used to output the final answer to the user terminal.

9. The apparatus according to claim 8, characterized in that, The retrieval processing module (12) is specifically used for: By identifying the questioner's intent and the user's emotions, and extracting key entity information from the question, the current context information can be obtained. Keyword retrieval is performed based on the keywords in the question and the metadata in the integrated profile information. The question is then converted into a query vector for vector similarity retrieval. The two retrieval results are then fused to obtain the context document. The question, the integrated profile information, and the at least one piece of evidence are constructed into a structured prompt, and the structured prompt is input into the generative model to generate the final answer.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images