Intelligent student question and answer method and system based on large language model and rapid retrieval

Abstract

The invention relates to the technical field of student intelligent questioning and answering, and discloses a student intelligent questioning and answering method and system based on a large language model and rapid retrieval, and the method comprises the steps: combining a student portrait to generate question vectors, rejecting noise points of a question vector set, obtaining hot questions, and storing the hot questions into a hot knowledge base after the hot questions are expanded by experts; generating a summary of the knowledge document by using a large language model, and adding the summary to each text block; generating a block summary for each text block, and storing the block summary and the text blocks into a universal knowledge base; inputting questions proposed by students into the large language model to generate a thinking chain; for sub-questions needing to be retrieved, firstly retrieving the hotspot knowledge base, if the sub-questions are not matched with the hotspot knowledge base, retrieving the block summary of the general knowledge base, and obtaining corresponding text blocks through mapping to obtain retrieval results; and guiding the big language model to fuse the thinking chain and the retrieval result to generate a structured answer. According to the method, the summary information is introduced after the text blocks are segmented, so that the information loss in the text segmentation process is reduced, and the reasonability of knowledge base construction is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering technology for students, specifically to an intelligent question answering method for students based on a large language model and fast retrieval. Background Technology

[0002] To improve the teaching quality of the online learning platform, it provides students with an intelligent question-and-answer assistant. During the learning process, students often encounter questions that are difficult to understand or require immediate clarification. The intelligent question-and-answer assistant provides real-time answers, preventing learning delays caused by waiting for teacher responses. Furthermore, the intelligent teacher assistant can guide students to further explore relevant knowledge points, stimulating their learning interest and creating a more autonomous and personalized learning model.

[0003] To effectively meet students' diverse and personalized question-and-answer needs across multiple domains, general-purpose large language models are typically used to implement intelligent question-and-answer assistants. These models possess powerful natural language understanding and generation capabilities, covering a wide range of knowledge domains to satisfy students' diverse learning needs. Through precise contextual analysis and user intent recognition, general-purpose large language models can provide more personalized and accurate answers.

[0004] When implementing a student intelligent question-answering assistant using a generalized large language model, considering the high accuracy requirements of student question-answering scenarios, it is necessary to address the illusion problem of the large language model (i.e., generating inaccurate or false answers). To solve this problem, a retrieval augmentation approach can be adopted. That is, before the large language model answers the student's question, the most relevant text is retrieved from an external knowledge base and fed into the large language model along with the student's question to assist the model in answering. Combining the generalized knowledge of the large language model with the specific knowledge of the external knowledge base can significantly improve the accuracy and reliability of the answers, providing students with more accurate and credible responses.

[0005] However, current student question-answering assistants based on large language models and external knowledge base retrieval have the following problems: (1) The external knowledge base is poorly constructed. In the process of constructing an external knowledge base, excessively large files are often directly split into a specified size. This violent splitting method may lead to the breakage of semantic information, making the text blocks lack contextual information, which affects the accuracy of the knowledge base and the retrieval effect.

[0006] (2) Low efficiency of external knowledge base retrieval. In order to effectively assist students in answering questions, it is necessary to retrieve content that is highly relevant to students' questions from external knowledge bases. This process is often time-consuming and affects students' actual experience.

[0007] (3) Lack of process guidance in the answers. When answering students' questions, the general language model tends to give students the conclusion directly, while ignoring the guidance for students. This opaque answering method is not only not conducive to cultivating students' logical thinking ability and problem-solving ability, but also leads students to doubt the source and rationality of the answer, affecting the final answering effect.

[0008] Existing technical solutions, when facing the specific scenario of intelligent question answering for students, often rely on general-purpose large language models and external knowledge base retrieval. This approach suffers from problems such as unreasonable external knowledge base construction, low retrieval efficiency, and a lack of process guidance in the answering process. Improving the rationality of external knowledge base construction and retrieval efficiency, and enhancing the guidance capabilities of large language models in answering questions, remains a challenging task. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a student intelligent question-answering method and system based on a large language model and fast retrieval.

[0010] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a student intelligent question-answering method based on a large language model and fast retrieval, comprising: The system collects questions raised by students, generates question vectors based on student profiles, removes noise points from the question vector set using the DBSCAN algorithm, obtains hot topics using the K-means clustering algorithm, and stores them in the hot topic knowledge base after being expanded by experts. The knowledge document is generated using a large language model. The knowledge document is divided into blocks and the overview is added to each text block. A block overview is generated for each text block and stored together with the text blocks in a general knowledge base and a mapping relationship is established. The questions raised by students are input into the large language model to generate a thought chain that includes sub-questions, dependent knowledge points, and external retrieval requirements. For the sub-question to be retrieved, first search the hot knowledge base. If no match is found, search the block overview of the general knowledge base. Obtain the corresponding text block through mapping to get the search results. The large language model is guided to integrate thought processes and retrieval results to generate structured answers that include problem-solving strategies, final responses, and confidence levels.

[0011] In one embodiment, the process of collecting questions raised by students, generating a question vector based on student profiles, removing noise points using the DBSCAN algorithm, and then obtaining hot topics using the K-means clustering algorithm specifically includes: Collect students' grade information and the questions they raised; use word embedding interface to represent the question text as a text vector, use one-hot encoding to represent the student's grade information as a grade vector, and concatenate the text vector and grade vector to form the question vector; after standardizing the set of question vectors, use the DBSCAN algorithm to remove noise points, and use the K-means clustering algorithm to perform clustering operation on the resulting noise-free set, and take the question vector corresponding to the question closest to the cluster center in each cluster as the hot question.

[0012] In one embodiment, the step of storing the expanded hot topic knowledge base by experts specifically includes: Experts expand on hot topics based on their teaching experience and develop standard answers for these topics. The hot topics and their corresponding standard answers are then stored in a hot topic knowledge base.

[0013] In one embodiment, the step of generating an overview of a knowledge document using a large language model involves dividing the knowledge document into blocks and adding an overview to each text block; generating a block overview for each text block, storing it together with the text blocks in a general knowledge base, and establishing a mapping relationship, specifically including: The knowledge document is fed into a general large language model, which guides the model to generate an overview of the knowledge document; the knowledge document is divided into blocks to obtain multiple text blocks; the overview of the knowledge document is added to each text block; and the text blocks with the added overview are input into the general large language model, which guides the model to generate a block overview. Both text blocks and their corresponding block summaries are stored in a general knowledge base, along with the mapping relationship between text blocks and block summaries.

[0014] In one embodiment, for the sub-question to be retrieved, the hot topic knowledge base is first searched; if no match is found, the block overview of the general knowledge base is searched, and the corresponding text block is obtained through mapping to obtain the search results. Specifically, this includes: For sub-questions requiring external retrieval, a hot topic knowledge base is first used for retrieval. If a hot topic knowledge base is found with a similarity greater than a set threshold to the sub-question, the hot topic knowledge base and its corresponding standard answer are used as the retrieval results. If no hot topic knowledge base is found with a similarity greater than the set threshold to the sub-question, a block overview in a general knowledge base is retrieved, and the block overview with the highest similarity to the sub-question and its corresponding text block are used as the retrieval results.

[0015] In a second aspect, the present invention provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any embodiment of the first aspect.

[0016] Compared with the prior art, the beneficial technical effects of the present invention are: This invention introduces overview information after text block segmentation, reducing information loss during text segmentation and improving the rationality of knowledge base construction. By leveraging hot topic knowledge base retrieval and overview information retrieval, the efficiency of knowledge base retrieval is effectively improved. Furthermore, by using task decomposition, a general-purpose large language model is guided to generate structured answers according to the thought chain, ensuring the transparency and completeness of the answers. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method in an embodiment of the present invention. Detailed Implementation

[0018] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] like Figure 1 As shown, this invention presents a student intelligent question-answering method based on a large language model and fast retrieval. First, a hot topic knowledge base and a general knowledge base are constructed using clustering algorithms and a general large language model. Second, the general large language model is guided to generate a thought chain for answering student questions, and knowledge base retrieval is performed based on the thought chain information. Finally, the general large language model is guided to generate a structured answer. This method requires an online education platform and specifically includes the following steps: S1, Hot Topic Knowledge Base Construction: Collect questions submitted by students to the question-and-answer assistant within a certain period and combine them with student profiles to achieve question vector modeling. Utilize DBSCAN and K-means clustering algorithms to eliminate low-frequency questions and identify hot topics. Experts expand the hot topics list and develop standard answers, which are then stored in the hot topic knowledge base.

[0020] S2, General Knowledge Base Construction: Knowledge documents are processed using a general large language model to obtain their overviews. The knowledge documents are then divided into blocks according to storage requirements, and the overviews are added to each text block. For each processed text block, the general large language model is used to obtain a block overview. The text blocks and their corresponding block overviews are then stored together in the general knowledge base.

[0021] S3, Mind Chain Acquisition: Input the questions raised by students into the large language model to generate a mind chain that includes sub-questions, dependent knowledge points, and external retrieval requirements.

[0022] S4, Knowledge Base Retrieval: For the sub-question in S3 that requires external retrieval, the first step is to use a hot knowledge base for retrieval. If no documents with high similarity are found in the hot knowledge base, the search will switch to a general knowledge base.

[0023] S5, General Large Language Model Question Answering: Guides the general large language model to integrate its own knowledge with the content retrieved from the knowledge base, generate structured answers, and provide students with confidence information.

[0024] In one embodiment, step S1 specifically includes: Hot Topic Knowledge Base Construction: The hot topic knowledge base is essentially a vector database, using high-performance storage devices as the storage medium to store the questions and answers that students are most likely to ask, enabling fast retrieval. To effectively achieve the goal of high performance, Redis, an in-memory database, can be used, which provides vector storage and vector retrieval capabilities.

[0025] Using an online education platform, questions submitted by students through the question-and-answer assistant are collected, and the grade information of the students who asked the questions is recorded. For each question, the question text is represented as a question vector q using the open-source word embedding interface of OpenAI. For the grade of the student who asked the question, one-hot encoding is used to represent it as a grade vector g. For example, if there are six grades, each grade vector has six dimensions. The i-th bit of the grade vector corresponding to grade i is 1, and the remaining bits are 0. For example, the grade vector corresponding to grade 3 is [0,0,1,0,0,0], and the grade vector corresponding to grade 5 is [0,0,0,0,1,0]. Finally, the question text vector q and the grade vector g are concatenated to obtain the question vector. The core formula is as follows: ; The collected n questions were used to form a vector set. .

[0026] Perform Z-score normalization on set Q: ; in, and These represent the mean and standard deviation of each dimension of all vectors. In this way, the standardized vector set can be obtained. , Let n be the nth standardized question vector. Then, the DBSCAN algorithm is used to process the set. Perform clustering operations. During this process, two core parameters are set. and .in, Represents the neighborhood radius, used to determine whether two samples are neighbors; This represents the minimum number of neighbors, used to determine whether a sample is noise.

[0027] In the DBSCAN algorithm, if a certain point of The number of points in the field is less than ,Right now ; in, This indicates the computation of two vectors. If the Euclidean distance between the points is zero, then this point does not belong to any cluster; that is, this point is labeled as noise (low-frequency data). Remove vector sets. Given points marked as noise, obtain a noise-free set containing m data points. .

[0028] Set the desired number of hot topics, t, as the cluster center number, and use the K-means clustering algorithm to cluster the set. Perform clustering. After clustering, the question text corresponding to the vector closest to the cluster center in each cluster is identified as the hottest question. During this process, if the cluster center vector of a certain cluster... The problem vector is Then calculate its distance. The formula is: .

[0029] Using the above method, for the t clusters obtained through clustering, t hot topics can be identified. These hot topics are then handed over to subject matter experts for processing, allowing them to expand upon the hot topics based on their teaching experience and develop standard answers. Finally, the hot topics and their answers are stored in a hot topic knowledge base.

[0030] In one embodiment, step S2 specifically includes: General Knowledge Base Construction: The general knowledge base is essentially a vector database, using high-capacity storage devices as the storage medium to store the massive amounts of teaching resources collected by the online teaching platform, enabling precise retrieval. To effectively achieve the goal of high-capacity storage, Milvus, a distributed database, can be used, which provides vector storage and vector retrieval functions.

[0031] Knowledge documents collected from online education platforms are fed into a general-purpose large language model (such as OpenAI's ChatGPT, Baidu's Wenxin Yiyan, and iFlytek's Xinghuo) to guide the model in generating an overview of the knowledge documents. The prompts are constructed as follows: ; Here, "connect" is a conjunction that ensures the fluency of the sentence; "knowledge" refers to knowledge documents. (Note 1) The template is as follows: = "You are a teaching expert. Please provide an overview of the knowledge document [" + knowledge + "]. The overview should be concise, professional, and effectively describe the document's content. The overview should not exceed 24 characters."

[0032] The knowledge document's summary (knowledge_summary) can be obtained using the above method. Then, according to database storage requirements, the knowledge document is segmented into blocks. For example, if the database requires that a document not exceed 1024 characters, the knowledge document needs to be divided into text blocks of 1000 characters each (reserving characters for later concatenation with the summary). For each knowledge document, a set of e segmented text blocks is obtained. For each text block, an overview of the knowledge document is added to the text block, resulting in a processed set of text blocks. : ; in, For the i-th text block, For the i-th processed text block, , This represents the total number of text blocks.

[0033] For each processed text block This information is then fed into a general-purpose large language model, which guides the model to generate its block summary. The formula for constructing prompts is as follows: ; Here, connect is a conjunction that ensures the fluency of the statement; This indicates the text block after the above processing.

[0034] Prompt word two The template is as follows: = "You are a teaching expert, please comment on [" + The + "] text block provides an overview. The overview should be concise and professional, effectively conveying the document's content. There is no length limit to the overview.

[0035] Each text block can be obtained using the above method. Block Overview This constitutes an overview collection. Finally, the text block With block overview Encode it into vector form and store it in a general knowledge base. This process also requires additional storage blocks (see overview). With text blocks The mapping relationship between blocks is established to ensure that the corresponding text block can be queried through the block overview.

[0036] In one embodiment, step S3 specifically includes: The query from the student's question is fed into a general-purpose large language model, which generates a structured thought process chain for solving the problem. This chain includes the sub-problems involved in solving the student's question and whether external knowledge retrieval is needed. The prompt word construction formula is as follows: ; Here, "connect" is a conjunction that ensures the fluency of the statement; "query" represents the student's question.

[0037] Prompt word two The template is as follows: = "You are a large language model skilled in teaching support. For a student's question [" + query + "], please break it down into a structured thought process chain, with each step including sub-questions and whether external knowledge retrieval is needed. Please output in standard JSON format, with an example output below:" { "Mind Chain": [ { Step number: 1, Subproblems: "...", "External retrieval required": true / false }, ... ] }".

[0038] In this way, the thought chain needed to solve student problems can be obtained, in the format of a JSON array.

[0039] In one embodiment, step S4 specifically includes: For each sub-problem `child` in the thought chain, if its "Does it need external retrieval?" field is false, this field is directly removed. If its "Does it need external retrieval?" field is true, the content of the "sub-problem" field `child` is used as the retrieval target, and the retrieval operation is performed. First, a similarity search is performed in the hot knowledge base. During this process, a similarity threshold of 0.9 is set, meaning that documents in the hot knowledge base with a similarity greater than or equal to 0.9 with the content of `child` are used as search results. If content with sufficiently high similarity is found in the hot knowledge base, it is used as a search result. If no content with sufficiently high similarity is found, the search proceeds to the general knowledge base. The overviews in the general knowledge base are searched, and the overview with the highest similarity to the student question is retained. By utilizing the mapping relationship between text blocks and block summaries stored in S2, the desired information can be obtained. The corresponding text block The two pieces of information are combined to form the search result. The formula for the search result is as follows: .

[0040] For the subquestion where the value of the "Does external retrieval need to be performed?" field is true, after obtaining the retrieval result, the "Does external retrieval need to be performed?" field is removed, and the "External knowledge" field is added, containing the retrieval result. Through these operations, the thought process chain containing knowledge from the knowledge base can be obtained. .

[0041] In one embodiment, step S5 specifically includes: This involves querying student questions and incorporating the thought process of knowledge from the knowledge base. The data is fed into a general-purpose large language model to guide the model in generating the final structured answer and supplementing the answer confidence information.

[0042] The formula for constructing a prompt is as follows: ; Here, connect is a conjunction that ensures the fluency of the statement; query represents the student's question; This represents the processed thought chain content, which includes knowledge retrieved from the knowledge base.

[0043] Prompt word four The template is as follows: = "You are an intelligent question-and-answer assistant. A student has asked you a question [" + query + "]. To better answer the student's question, you have engaged in deep thinking and obtained a thought chain [" + The + "] section contains step-by-step information and supplementary knowledge for answering student questions. Now, please generate a "Problem-Solving Approach" and a "Final Answer" for the student based on the above content to help them solve the problem. Furthermore, considering that students have a low tolerance for hallucination questions in their question-and-answer scenarios, please supplement the answer with "Answer Confidence" information for the student's reference. This information includes three options: "Low," "Medium," and "High." Please judge based on the completeness of the reasoning path and the sufficiency of knowledge. Please output in standard JSON format. An example output is as follows: { "Problem-solving approach": [ { Step number: 1, Subproblems: "...", "Reasoning process": "...", "Knowledge Used":"..." }, ... ], Final answer: "...", "Response confidence level": "High / Medium / Low" }".

[0044] By using the methods described above, the final answer to the student's question can be obtained and then fed back to the student as the result.

[0045] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0046] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0047] In one embodiment, the present invention provides a computer system, which may be a server. The computer system includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data used in the methods described above. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the methods described above.

[0048] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0049] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0050] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A student intelligent question-answering method based on a large language model and fast retrieval, characterized in that, include: The system collects questions raised by students, generates question vectors based on student profiles, removes noise points from the question vector set using the DBSCAN algorithm, obtains hot topics using the K-means clustering algorithm, and stores them in the hot topic knowledge base after being expanded by experts. The knowledge document is generated using a large language model. The knowledge document is divided into blocks and the overview is added to each text block. A block overview is generated for each text block and stored together with the text blocks in a general knowledge base and a mapping relationship is established. The questions raised by students are input into the large language model to generate a thought chain that includes sub-questions, dependent knowledge points, and external retrieval requirements. For the sub-question to be retrieved, first search the hot knowledge base. If no match is found, search the block overview of the general knowledge base. Obtain the corresponding text block through mapping to get the search results. The large language model is guided to integrate thought processes and retrieval results to generate structured answers that include problem-solving strategies, final responses, and confidence levels.

2. The intelligent question-answering method for students based on a large language model and fast retrieval as described in claim 1, characterized in that, The process of collecting student-submitted questions, generating question vectors based on student profiles, removing noise points from the question vector set using the DBSCAN algorithm, and then identifying hot topics using the K-means clustering algorithm includes: Collect students' grade information and the questions they raised; use word embedding interface to represent the question text as a text vector, use one-hot encoding to represent the student's grade information as a grade vector, and concatenate the text vector and grade vector to form the question vector; after standardizing the set of question vectors, use the DBSCAN algorithm to remove noise points, and use the K-means clustering algorithm to perform clustering operation on the resulting noise-free set, and take the question vector corresponding to the question closest to the cluster center in each cluster as the hot question.

3. The intelligent question-answering method for students based on a large language model and fast retrieval as described in claim 1, characterized in that, The information, after being expanded by experts and stored in the hot topic knowledge base, specifically includes: Experts expand on hot topics based on their teaching experience and develop standard answers for these topics. The hot topics and their corresponding standard answers are then stored in a hot topic knowledge base.

4. The intelligent question-answering method for students based on a large language model and fast retrieval as described in claim 1, characterized in that, The process of generating an overview of a knowledge document using a large language model involves dividing the knowledge document into blocks and adding an overview to each text block; generating a block overview for each text block, storing it along with the text blocks in a general knowledge base, and establishing a mapping relationship. Specifically, this includes: The knowledge document is fed into a general large language model, which guides the model to generate an overview of the knowledge document; the knowledge document is divided into blocks to obtain multiple text blocks; the overview of the knowledge document is added to each text block; and the text blocks with the added overview are input into the general large language model, which guides the model to generate a block overview. Both text blocks and their corresponding block summaries are stored in a general knowledge base, along with the mapping relationship between text blocks and block summaries.

5. The intelligent question-answering method for students based on a large language model and fast retrieval as described in claim 1, characterized in that, For each sub-question to be retrieved, the hot topic knowledge base is searched first. If no match is found, the block overview of the general knowledge base is searched. The corresponding text block is obtained through mapping to obtain the search results, which specifically include: For sub-questions requiring external retrieval, a hot topic knowledge base is first used for retrieval. If a hot topic knowledge base is found with a similarity greater than a set threshold to the sub-question, the hot topic knowledge base and its corresponding standard answer are used as the retrieval results. If no hot topic knowledge base is found with a similarity greater than the set threshold to the sub-question, a block overview in a general knowledge base is retrieved, and the block overview with the highest similarity to the sub-question and its corresponding text block are used as the retrieval results.

6. A computer system comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

Images