Intelligent teaching method and system based on large model, server and medium

Abstract

The embodiment of the invention discloses an intelligent teaching method and system based on a large model, a server and a medium. The method comprises the following steps: constructing a plurality of agents according to the intention of a user; obtaining a historical dialogue record and a current question of a user to form an input text; identifying a user intention by adopting an intention identification agent to obtain an intention label; and determining an intention type according to the intention label, and calling the corresponding large model agent to perform question and answer response. According to the embodiment of the invention, the large language model and the multiple agents are combined, so that the problems of non-personalized teaching and intelligent adaptation teaching at present are solved, and the problems of large subsequent vector retrieval result error, tedious knowledge graph construction and low reasoning efficiency caused by inaccurate intention recognition of a general large model in the education field at present are solved; therefore, the intellectualization, personalization and online learning efficiency of online education are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of online intelligent education, and in particular to an intelligent teaching method, system, server and medium based on a large model. Background Art

[0002] With the rapid development of artificial intelligence technology, large language model (LLM) technology has become one of the core technologies to promote the widespread application of AI. Large language models have deep semantic understanding and reasoning capabilities by learning massive amounts of data and knowledge, which enables AI to process information and make decisions in a way closer to humans. However, the existing general large models are more suitable for tasks with simple processes such as machine translation and knowledge question answering, and perform poorly on tasks involving complex processes in a specific field.

[0003] As an important way to inherit human culture, education has always faced problems such as uneven resource allocation and difficulty in implementing personalized teaching. Under the traditional teaching method, it is difficult for teachers to provide targeted teaching based on the characteristics of each student, which restricts the personalized development of students. How to integrate AI technology into the traditional education model to achieve intelligent and personalized education is a hot topic and difficulty in current research. Based on this background, developing an intelligent teaching system based on a large model so that it can adapt to a variety of teaching scenarios, truly realize "intelligent adaptive teaching", and provide personalized learning plans for different students has important practical significance and broad application prospects.

[0004] For example, the Chinese patent application with publication number CN117437099A discloses a smart teaching system and teaching method based on a large language model, which combines knowledge graphs with large models to realize an intelligent teaching system that can prepare lessons before class, guide discussions during class, and conduct post-class testing and evaluation. However, this method requires the extraction of triples in advance to construct a knowledge graph. When the amount of knowledge graph data is large, querying the knowledge graph will be time-consuming and inefficient, resulting in slow feedback from the teaching system. In addition, in the same subdivision field, many problems are homogeneous. At this time, the large model is often unable to accurately identify the user's intentions. The large model is better at identifying problems and intentions in different fields in terms of intent recognition. Finally, the patent only uses vector similarity to retrieve related texts. Vector retrieval tends to semantic understanding, but users' questions are often very short keywords, and this keyword can derive multiple semantics. Therefore, it would be better to use keyword retrieval at this time.

[0005] In summary, the defects of the prior art are as follows:

[0006] Non-personalized teaching: It is difficult to provide personalized teaching plans based on students' abilities, interests and learning habits under the traditional teaching model, and all classes are taught in a unified manner.

[0007] The construction of a knowledge graph requires the precise extraction of triples, which is rather cumbersome and requires manual intervention.

[0008] Large models cannot achieve accurate intent recognition in segmented fields because the problems in segmented fields are highly homogeneous.

[0009] Vector search is not as effective as keyword search when the user input is very short or very long. Summary of the invention

[0010] The purpose of the embodiments of the present invention is to provide an intelligent teaching method, system, server and medium based on a large model to solve the current non-personalized intelligent adaptive teaching, as well as the problems in the existing technical solutions that the knowledge graph construction is cumbersome and requires manual intervention, the large model cannot accurately classify homogeneity in segmented fields, and the vector retrieval has a low recall rate on some ultra-short and ultra-long questions.

[0011] To achieve the above objectives, in a first aspect, an embodiment of the present invention provides an intelligent teaching method based on a large model, comprising:

[0012] Build multiple agents based on the user's intentions, including intention recognition agents, learning automatic planning agents, legal knowledge question-answering agents, legal practice consulting agents, legal examination question answering agents, study report interpretation agents, question recommendation agents, course recommendation agents, and other agents;

[0013] Obtaining the user's historical conversation records and current question, and forming an input text according to the user's historical conversation records and current question;

[0014] Inputting the input text into the intention recognition agent to recognize the user's intention and obtain an intention label;

[0015] The intent type is determined according to the intent label, and the corresponding large model agent is called to perform question-and-answer response according to the intent type.

[0016] As a specific implementation of the present application, multiple agents are constructed according to the user's intention as follows:

[0017] Cluster the user's intentions into multiple categories, each category corresponds to an agent;

[0018] The intention recognition agent is used to determine the user's intention, and the remaining agents are used to answer user questions. The user's intention type corresponds to the agent one by one.

[0019] The intent recognition agent includes a large model, a prompt word, and a callable tool. The callable tool is an intent recognition model API tool. The intent recognition model is obtained by stacking Transformer encoders for training. The training process of the intent recognition model includes:

[0020] Process the user's historical conversation records and current questions;

[0021] The processed text is segmented by a word segmenter to obtain the text index corresponding to each word; the processed text includes the user's historical conversation records and current questions;

[0022] The text index after word segmentation is queried in the word vector table to obtain the text vector corresponding to the word segmentation, and the position vector is added to form the input vector;

[0023] Input the input vector into the Transformer encoder to obtain an output vector;

[0024] The output vector is input into the fully connected layer FFN and activated by the Tanh activation function;

[0025] After the softmax function operation, the intent classification and recognition are performed, the loss is calculated by the cross entropy loss function, and the model is trained by back propagation.

[0026] As a specific implementation of this application, the intent tag is obtained as follows:

[0027] The intent recognition agent calls the intent recognition model API tool to perform intent recognition on the input text, and obtains the top three intent labels with the highest confidence probability;

[0028] If the confidence probability corresponding to the first-ranked intent label is greater than the threshold, the corresponding intent recognition label is directly output. Otherwise, the user is actively asked whether his intent is the intent model corresponding to the intent label. If so, the corresponding intent recognition label is directly output.

[0029] The same method is used to determine the second and third ranked intent labels;

[0030] If the top three intent labels with the highest confidence probability do not contain the user's true intention, the intent recognition agent will directly prompt the user to enter more relevant information, and then call the intent recognition model API tool again for recognition.

[0031] As a specific implementation of the present application, determining the intent type according to the intent label, and calling the corresponding large model agent to perform question-answering response according to the intent type, including:

[0032] According to the intent label, query the intent label and intent type mapping table to obtain the intent type;

[0033] Input historical chat records and user questions into the corresponding agent;

[0034] The agent decides whether to call the tool to obtain additional information and responds.

[0035] In the second aspect, an embodiment of the present application also provides an intelligent teaching system based on a large model, including a client and a server, wherein the client is used to send the user's historical chat records, current questions and user basic information to the server; the server includes a processor, an input device, an output device and a memory, and the processor, input device, output device and memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the method of the above-mentioned first aspect.

[0036] In a third aspect, the embodiment of the present application provides a server, which is suitable for intelligent teaching based on a large model, including:

[0037] A construction unit, which is used to construct multiple agents according to the user's intention, including an intention recognition agent, a learning automatic planning agent, a legal knowledge question-answering agent, a legal practice consulting agent, a legal examination question answering agent, a learning report interpretation agent, a question recommendation agent, a course recommendation agent, and other agents;

[0038] An acquisition unit, used to acquire a user's historical conversation record and a current question, and form an input text according to the user's historical conversation record and the current question;

[0039] A recognition unit, used for inputting the input text into the intention recognition agent to recognize the user's intention and obtain an intention label;

[0040] The recognition unit is also used to determine the intent type according to the intent label, and call the corresponding large model agent to perform question and answer response according to the intent type.

[0041] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium storing a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes the method of the first aspect.

[0042] By implementing the embodiments of the present invention, by combining a large language model with multiple agents, the current problems of non-personalized teaching and intelligent adaptive teaching are solved, as well as the current general large model's inaccurate intent recognition in the education field, which leads to large errors in subsequent vector retrieval results, as well as cumbersome knowledge graph construction and low reasoning efficiency. This improves the intelligence, personalization and online learning efficiency of online education. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the specific implementation of the present invention or the technical solution in the prior art, the drawings required for use in the specific implementation or the description of the prior art are briefly introduced below.

[0044] Figure 1 is a schematic diagram of an application environment of an intelligent teaching method based on a large model in an embodiment of the present invention;

[0045] Figure 2 is a flow chart of an intelligent teaching method based on a large model in an embodiment of the present invention;

[0046] Figure 3 It is a mapping table between intent labels and intent types;

[0047] Figure 4 This is the schematic diagram of the intent recognition agent;

[0048] Figure 5 yes Figure 2 Sub-flow chart of step S2;

[0049] Figure 6 is an example diagram of model reasoning of an intelligent teaching method based on a large model in an embodiment of the present invention;

[0050] Figure 7 is an example diagram of model training data of an intelligent teaching method based on a large model in an embodiment of the present invention;

[0051] Figure 8 It is the model training flow chart;

[0052] Fig. 9 yes Figure 2 Sub-flow chart of step S3;

[0053] Fig.10 is a schematic diagram of a computer device in an embodiment of the present invention;

[0054] Fig.11 It is the structural diagram of the agent module. DETAILED DESCRIPTION

[0055] 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 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.

[0056] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0057] In order to enable those skilled in the art to better understand the technical solutions provided by the embodiments of the present application, some technical terms involved in the embodiments of the present application are briefly explained below.

[0058] 1. Large model, here refers to large language model, which is a neural network model based on the Transformer decoder architecture. The number of parameters ranges from billions to hundreds of billions. The larger the number of parameters, the more training data, and the stronger the model's capabilities. The model's capabilities mainly include context learning ability, instruction following ability, and thinking chain ability.

[0059] 2. Agent: The agent is composed of three parts: a big model, prompt words, and callable tools. The prompt words clearly indicate the role, positioning, skills, and when to call which tools of the current big model; the callable tools indicate which API tools the agent can independently decide to call. The big model is a large language model, exemplified by Qwen-plus, Qwen-max, Doubao function-call series, ChatGLM, etc.

[0060] It should be noted that the embodiments of the present invention intend to provide an intelligent teaching solution based on a large model, which, on the one hand, solves the problems of intelligent adaptation and personalization of online education, and on the other hand, solves the problems that some current intelligent teaching methods require manual construction of knowledge graphs, as well as the inaccurate intent recognition of large models in segmented fields and the inaccurate retrieval results of vector retrieval when the input is extremely short or extremely long.

[0061] The intelligent teaching method based on a large model provided by the embodiment of the present invention can be applied in Figure 1In the application environment, the client can be a mobile phone, tablet, computer and other terminals, and the server can be a local server, cloud server cluster, etc. The client sends the user's historical chat records, current questions and basic user information to the server through the network. The large model-based intelligent teaching method deployed on the server processes and answers the client's request and then returns it to the client.

[0062] In one embodiment, if Figure 2 As shown in the figure, an intelligent teaching method based on a large model is provided, which is applied in Figure 1 The server in the example is used as an example to illustrate the following steps:

[0063] S1, builds a multi-agent based on the user’s intention, user conversation history and current question input.

[0064] Among them, to construct a multi-agent based on the user's intention, it is first necessary to cluster the user's questions according to the intention. Exemplary, common text clustering models, such as Kmeans, KNN, BERTopic, etc., preferably use BERTopic. In this embodiment, Figure 3 The diagram labels and intent type mapping relationship table shows that user intentions can be clustered into the following eight categories: automatic learning planning, legal knowledge Q&A, legal practice consultation, answers to legal examination questions, interpretation of study reports, topic recommendations, course recommendations, and others.

[0065] The intelligent agent in this embodiment includes:

[0066] The learning automatic planning agent can call the user portrait API tool. Through this tool, the agent can obtain the user's gender, age, whether the test goal is objective questions or subjective questions in the judicial examination, undergraduate major, whether it is law or non-law, whether it is the first time to take the exam, and the expected study time per day. The agent can customize a detailed personalized learning plan based on the user portrait information;

[0067] The legal knowledge question-answering agent can call the legal knowledge retrieval API tool to obtain legal knowledge documents related to the user's question;

[0068] Legal practice consulting agent, which can call case consulting API tool to obtain relevant case information;

[0069] The legal examination question answering agent can obtain question-related information by calling the question retrieval API tool, such as the question stem, options, standard answers, relevant knowledge points, relevant textbooks and handouts, common questions and answers, etc.

[0070] The learning report interpretation agent can call the learning data retrieval API tool to obtain the user's learning time, class time, number of questions answered, accuracy rate and other data;

[0071] The topic recommendation agent can obtain relevant topics by calling the topic recommendation API;

[0072] Course recommendation agent, which can obtain relevant courses through the course recommendation API tool;

[0073] Other intent agents can call the online retrieval API tool to obtain relevant information.

[0074] In addition to the 8 agents corresponding to the above 8 intents, an intent recognition agent is also required. Specifically, Figure 4 As shown, the diagram illustrates the large model, prompt words and callable tools corresponding to the intention recognition agent in this embodiment. It is used to identify the user's current intention, and then route the current question to one of the eight agents for answering according to the user's intention.

[0075] The user's historical chat records will increase as the chat progresses. When the text length of the historical chat records exceeds the context processing capability of the large model, the historical chat records need to be truncated. Therefore, a common processing method is to select the first N rounds of historical chat records for the current question. One question and one answer is one round of chat records, and the preferred N=3.

[0076] S2, based on the user’s historical conversation records and current questions, calls the intent recognition agent to identify the user’s intent and obtain the corresponding intent label.

[0077] like Figure 5 As shown, step S2 specifically includes:

[0078] S21, the intent recognition agent calls the intent recognition tool to perform intent recognition and obtain the top 3 intent labels and corresponding confidence probabilities.

[0079] For example, if a user asks: "What is the crime of theft?", the intent recognition model API tool returns the following result: {1: 0.85, 2: 0.1, 7: 0.05}, where 1: 0.85 means that the first-ranked intent label is 1, with a corresponding confidence probability of 0.85, the second-ranked intent label is 2, with a corresponding confidence probability of 0.1, and the third-ranked intent label is 7, with a corresponding confidence probability of 0.05.

[0080] S22, determining whether the confidence probability of the first-ranked intention is greater than or equal to a threshold value p.

[0081] Optionally, the threshold is set to 0.8.

[0082] S23, if it is greater than the threshold, the corresponding agent is directly assigned; if it is less than the threshold, the user is asked to confirm using the large model.

[0083] According to the top 3 results obtained in S21 in this embodiment, it is determined that the first ranked intent label is 1. Figure 3 The intent label and intent type mapping table shows that the corresponding intent type is legal knowledge question and answer, and the corresponding confidence probability is 0.85, which is greater than the threshold p=0.8. At this time, the user’s current intent label will be directly judged as 1, and the intent recognition intelligent body will directly output the user’s intent label 1.

[0084] If the confidence probability corresponding to the user's first-ranked intention label is less than the threshold p, then the intention recognition intelligent body will ask the user whether the user's current intention is to consult relevant legal knowledge. If the user answers yes, the intention recognition intelligent body will only output intention label 1. If the user answers no, the intention recognition intelligent body will ask the user whether the current intention is the second-ranked intention type. In this embodiment, the intention type corresponding to intention label 2 is legal practice consultation. The intention type corresponding to the third-ranked intention label will be asked in turn until the user's affirmative answer is obtained. If the user's answers are all negative answers, it proves that the current intention recognition result does not hit the user's true intention. At this time, the intention recognition intelligent body will request the user to enter more user-related information and then repeat the above intention recognition steps.

[0085] Furthermore, the intent recognition agent obtains the user's top 3 intent tags and corresponding confidence probabilities by calling the intent recognition model API tool. Specifically, the intent recognition model API tool is the deployed intent recognition model inference service API, such as Figure 6 As shown, the figure indicates the reasoning process of the intent recognition model.

[0086] S216, initializing the intention recognition model parameters.

[0087] Specifically, the intent recognition model is a neural network model formed by stacking Transformer encoders, which can be obtained through training. Figure 7 As shown, the user's historical chat records and current questions are labeled, and the intent label is one of the eight intent labels. After labeling, training data is obtained, which can be used to train the intent recognition model.

[0088] Train the above intent recognition model. The training process is as follows: Figure 8 As shown, including:

[0089] S210, processing user conversation record and current question text.

[0090] S211, use a tokenizer to tokenize the text, the text includes the previous N rounds of historical chat records and the current question input by the user, when N=0, only the current question of the user. The tokenizer used has different tokenizers according to different models. Preferably, in this embodiment, the tokenizer of the BERT model is used as the tokenizer, and the index corresponding to each word is obtained after tokenization.

[0091] S212, the text index after word segmentation is queried in the word vector table to obtain the text vector token_embedding corresponding to the word segmentation, and the position vector pos_embedding is added to form the input vector X. The word vector table is an initialized embedding table, and the word vector embedding table can be obtained through training. The position vector, optionally, such as Sinusoidal position coding, RoPE position coding, Sinusoidal position coding is used in this embodiment. The position coding can convert the corresponding subscript of each word segmentation in the input text sequence into a position vector, and then add it to the text vector obtained by querying the word vector embedding table to obtain the input vector X.

[0092] X=token_embedding+pos_embedding

[0093] S213, input the input vector into the Transformer encoder. After S212, the input vector X is obtained and input into a module with multiple Transformer encoders stacked. The Transformer encoder has a multi-head attention mechanism, which can obtain the self-attention weight of the input text and has a good effect in text encoding. After this step, the output vector Y after text encoding is obtained.

[0094] Y = TransformerEncoder(X)

[0095] S214, input the output vector Y into the fully connected layer FFN and activate it through the Tanh activation function to obtain the output vector output.

[0096] output = Tanh(FNN(Y))

[0097] S215, the output is operated by the softmax function to obtain the prediction result pred, and the loss loss is calculated by the cross entropy loss function (CrossEntropyLoss), and the model is trained by back propagation.

[0098]

[0099] Where pred represents the model output, y represents the true intent label of the sample, n represents the number of samples, and C represents the number of intent labels.

[0100] The above steps S210-S215 involve the training steps of the intent recognition model, which are also steps included in the intelligent teaching method based on a large model.

[0101] S217, deploy the intent recognition model as an API tool for the intent recognition agent to call, and the deployment can adopt various frameworks, illustrative examples are flask, fastapi, etc. In this embodiment, after fastapi is deployed, the http access address of the API is obtained, such as http: / / localhost:8080 / intent_recognition, and the input parameter structure is as follows:

[0102]

[0103] Here, history represents historical chat records, user represents users, assistant represents large models, and content represents questions or answers from users or large models.

[0104] S218, the intent recognition agent inputs the user's historical chat records and current questions by calling the intent recognition model tool API.

[0105] S219, the intent recognition model API tool receives the user's historical chat records and current questions sent by the intent recognition agent and inputs them into the intent recognition model, and returns the top 3 intentions and corresponding probabilities to the intent recognition agent.

[0106] The above S216, S217, S218, and S219 are the reasoning process of the intention recognition agent calling the intention recognition model tool.

[0107] S3, determines the intent type based on the intent label and calls the corresponding large model agent to respond.

[0108] like Fig. 9 As shown, step S3 includes:

[0109] S31, according to the intent label, query the intent label and intent type mapping relationship table to obtain the intent type.

[0110] In this embodiment, the query Figure 3 From the mapping relationship table between intent labels and intent types, we can see that the intent type corresponding to intent label 1 is legal knowledge question and answer.

[0111] S32, input the historical chat records and the user’s current questions into the corresponding agent.

[0112] In this embodiment, the user's question is "What is the crime of theft?", which is input into the corresponding legal knowledge question-answering agent.

[0113] S33, the agent decides whether to call the tool to obtain other information and gives a response.

[0114] In this embodiment, according to the user's current question, the answer result is as follows: "Theft refers to the act of infringing on the property ownership of others for the purpose of illegal possession. In criminal law, theft is a common property crime. According to the provisions of the Chinese Criminal Law, theft of public or private property may be sentenced to different penalties depending on the severity of the circumstances."

[0115] By implementing the embodiments of the present invention, by combining a large language model with multiple agents, the current problems of non-personalized teaching and intelligent adaptive teaching are solved, as well as the current general large model's inaccurate intent recognition in the education field, which leads to large errors in subsequent vector retrieval results, as well as cumbersome knowledge graph construction and low reasoning efficiency. This improves the intelligence, personalization and online learning efficiency of online education.

[0116] Based on the same inventive concept, the present invention also provides an intelligent teaching system based on a large model, including a client and a server, wherein the client is used to send the user's historical chat records, current questions and user basic information to the server; the server processes and answers the client's request based on the intelligent teaching method of the large model, and returns the result to the client.

[0117] The server can be understood as a Fig.10 The computer device shown may be a cloud server or a local server. The computer device includes a processor, an I / O bridge, a memory bus, an input / output bus, a non-volatile storage medium, and the like. The processor is connected to the I / O bridge via a system bus, the I / O bridge is connected to the memory bus and the input / output bus respectively, and the non-volatile storage medium is connected to the input / output bus via the memory bus. A USB adapter, a graphics adapter, and a network adapter are connected to the input / output bus. The USB adapter is connected to a keyboard and a mouse, the graphics adapter is connected to a display, and the network adapter is connected to a network card.

[0118] The database of the computer device is used to store the data used in the above-mentioned intelligent teaching method based on the large model, including training data, reasoning data, etc. The network card of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the intelligent teaching method based on the large model is implemented.

[0119] Accordingly, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the steps of the above-mentioned large-model-based intelligent teaching method are implemented.

[0120] The computer-readable storage medium may be an internal storage unit of the system described in any of the foregoing embodiments, such as a hard disk or memory of the system. The computer-readable storage medium may also be an external storage device of the system, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the system. Furthermore, the computer-readable storage medium may also include both an internal storage unit of the system and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by the system. The computer-readable storage medium may also be used to temporarily store data that has been output or is to be output.

[0121] Optionally, the server may also include a construction unit, an acquisition unit, an identification unit and multiple agent modules.

[0122] Among them, the construction unit is used to construct multiple agents according to the user's intention, including intention recognition agent, learning automatic planning agent, legal knowledge question and answer agent, legal practice consultation agent, legal examination question answering agent, study report interpretation agent, question recommendation agent, course recommendation agent and other agents;

[0123] An acquisition unit, used to acquire a user's historical conversation record and a current question, and form an input text according to the user's historical conversation record and the current question;

[0124] The recognition unit is used to input the input text into the intention recognition agent to recognize the user's intention, obtain the intention label, determine the intention type according to the intention label, and call the corresponding large model agent to perform question and answer response according to the intention type.

[0125] Specifically, Fig.11 As shown, the multiple agent modules are:

[0126] The intent recognition agent module 10 obtains the user's top 3 intent labels and corresponding confidence probabilities by calling the intent recognition model API tool, accurately identifies the user's current intent label by interacting with the user, and determines the corresponding intent type based on the user's intent label. The intent label and intent type are in a one-to-one correspondence. The corresponding relationship can be referred to Figure 8As shown, according to the user's current intention type, the corresponding agent is assigned to answer the user's specific questions.

[0127] The learning automatic planning intelligent agent module 20 is used to solve the user's learning planning problem. It can call the user portrait API tool to obtain the user's basic personal information, such as gender, age, whether the test goal is objective questions or subjective questions of the judicial examination, undergraduate major, whether it is law or non-law, whether it is the first time to take the exam, the expected study time per day and other information. The intelligent agent formulates a personalized learning plan suitable for the user based on the user portrait.

[0128] The legal knowledge question-and-answer agent module 30 is used to answer users' legal expertise questions. The agent can retrieve question-related background knowledge based on the user's historical chat records and current questions through the legal knowledge retrieval API tool, and then use the contextual learning ability of the large model to provide accurate answers.

[0129] The legal practice consulting agent module 40 is used to answer users' case consulting questions. The agent mainly retrieves the cases included in the user's questions through the case consulting retrieval API, and the big model organizes the language to answer based on the retrieved results.

[0130] The legal examination question answering intelligent agent module 50 is used to answer users' questions about legal examination questions, including objective questions and subjective questions. It mainly obtains relevant information of the questions through the question retrieval API tool, including the question stem, options, questions, standard answers, relevant knowledge points, relevant teaching materials, common questions and answers, etc.

[0131] The learning report interpretation agent module 60 is used to answer users' learning data inquiries. It mainly obtains users' learning data through learning data retrieval API tools, including learning indicators such as learning time, lecture time, number of questions answered, accuracy rate, etc. By analyzing various learning data of users, it points out the advantages and disadvantages of users and gives follow-up learning suggestions.

[0132] The question recommendation intelligent agent module 70 is used to recommend questions to users, including objective questions and subjective questions. It mainly obtains relevant questions through the question recommendation API tool. According to user needs, including the type of question, whether it is an objective question or a subjective question, relevant knowledge points, difficulty of the relevant question, whether it is a real question and other information, the question recommendation API tool obtains relevant questions and recommends them to users.

[0133] The course recommendation agent module 80 recommends courses to other users, including courses on judicial examination-related subjects such as criminal law, civil law, and criminal procedure law. According to the user's personal needs, such as teacher preferences, subjects studied, relevant chapters, knowledge points and other information, the corresponding courses are obtained and recommended to the user.

[0134] Other agent modules 90 are mainly used to answer various other questions from users. When the user's current intention type is judged to be other, the user's current chat will be assigned to this agent for answering. This agent can independently decide whether to call the search engine API tool to search for relevant information online, and summarize and answer the user's questions.

[0135] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An intelligent teaching method based on a large model, characterized in that: include: Build multiple agents based on the user's intentions, including intention recognition agents, learning automatic planning agents, legal knowledge question-answering agents, legal practice consulting agents, legal examination question answering agents, study report interpretation agents, question recommendation agents, course recommendation agents, and other agents; Obtaining the user's historical conversation records and current question, and forming an input text according to the user's historical conversation records and current question; Inputting the input text into the intention recognition agent to recognize the user's intention and obtain an intention label; The intent type is determined according to the intent label, and the corresponding large model agent is called to perform question-and-answer response according to the intent type.

2. The large model-based intelligent teaching method according to claim 1, characterized in that: Constructing multiple agents based on the user's intentions is as follows: Cluster the user's intentions into multiple categories, each category corresponds to an agent; The intention recognition agent is used to determine the user's intention, and the remaining agents are used to answer user questions. The user's intention type corresponds to the agent one by one.

3. The intelligent teaching method based on a large model as claimed in claim 1, characterized in that: The intention recognition agent includes three parts: a large model, prompt words and callable tools. The callable tool is an intention recognition model API tool, and the intention recognition model is obtained by stacking training of Transformer encoders.

4. The large model-based intelligent teaching method according to claim 3, characterized in that: The training process of the intent recognition model includes: Process the user's historical conversation records and current questions; The processed text is segmented by a word segmenter to obtain the text index corresponding to each word; the processed text includes the user's historical conversation records and current questions; The text index after word segmentation is queried in the word vector table to obtain the text vector corresponding to the word segmentation, and the position vector is added to form the input vector; Input the input vector into the Transformer encoder to obtain an output vector; The output vector is input into the fully connected layer FFN and activated by the Tanh activation function; After the softmax function operation, the intent classification and recognition are performed, the loss is calculated by the cross entropy loss function, and the model is trained by back propagation.

5. The large model-based intelligent teaching method according to claim 3, characterized in that: The intent tag is: The intent recognition agent calls the intent recognition model API tool to perform intent recognition on the input text, and obtains the top three intent labels with the highest confidence probability; If the confidence probability corresponding to the first-ranked intent label is greater than the threshold, the corresponding intent recognition label is directly output. Otherwise, the user is actively asked whether his intent is the intent model corresponding to the intent label. If so, the corresponding intent recognition label is directly output. The same method is used to determine the second and third ranked intent labels; If the top three intent labels with the highest confidence probability do not contain the user's true intention, the intent recognition agent will directly prompt the user to enter more relevant information, and then call the intent recognition model API tool again for recognition.

6. The large model-based intelligent teaching method according to claim 1, characterized in that: Determine the intent type according to the intent label, and call the corresponding large model agent to perform question-answering response according to the intent type, including: According to the intent label, query the intent label and intent type mapping table to obtain the intent type; Input historical chat records and user questions into the corresponding agent; The agent decides whether to call the tool to obtain additional information and responds.

7. An intelligent teaching system based on a large model, including a client and a server, characterized in that: The client is used to send the user's historical chat records, current questions and basic user information to the server; the server includes a processor, an input device, an output device and a memory, and the processor, input device, output device and memory are interconnected, wherein the memory is used to store a computer program, and the computer program includes program instructions, and the processor is configured to call the program instructions to execute the large model-based intelligent teaching method as described in any one of claims 1-6.

8. A server, suitable for intelligent teaching based on a large model, characterized in that: The server includes: A construction unit, which is used to construct multiple agents according to the user's intention, including an intention recognition agent, a learning automatic planning agent, a legal knowledge question-answering agent, a legal practice consulting agent, a legal examination question answering agent, a learning report interpretation agent, a question recommendation agent, a course recommendation agent, and other agents; An acquisition unit, used to acquire a user's historical conversation record and a current question, and form an input text according to the user's historical conversation record and the current question; A recognition unit, used for inputting the input text into the intention recognition agent to recognize the user's intention and obtain an intention label; The recognition unit is also used to determine the intent type according to the intent label, and call the corresponding large model agent to perform question and answer response according to the intent type.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 6.

Images