Method and system for constructing multi-round ticket booking agent based on large model

Abstract

The present invention belongs to the field of artificial intelligence and intelligent customer service technology. The present invention discloses a method and system for a multi-round ticket booking agent based on a large model. The method includes the following steps: S1, obtaining definitions of relevant elements of a multi-round ticket booking dialogue and using them as the basis for constructing a multi-round ticket booking agent, wherein the definitions of the relevant elements include the definition of multi-round interaction intentions, the jump relationship of multi-round states, and the functional definition of a ticket booking service API; the intention recognition module in the present invention can represent the user's real needs with structured intentions, and then can optimize the pruning of intentions when using a large model for intention recognition reasoning; supervised fine-tuning can effectively reduce the difficulty of intention recognition in dialogue management tasks and optimize the system interaction quality; through the dialogue state update task and API scheduling in the multi-round ticket booking agent system, each round of user dialogue can be replied and real-time flight information can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and intelligent customer service technology, and more specifically, to a method and system for constructing a multi-round ticket booking agent based on a large model. Background Art

[0002] Large models (generative large models, such as LLMs) are breakthroughs in the field of artificial intelligence. They demonstrate exceptional capabilities in text generation, language understanding, and dialogue management. Pre-trained on large amounts of text data, these models are able to capture complex language patterns and generate contextually appropriate responses. Applying LLMs to multi-round ticket booking agents can effectively improve the intelligence level and interaction quality of dialogue systems.

[0003] With the rapid development of Internet technology, online ticket booking services have become a part of people's daily lives, playing an important role in the fields of tourism and transportation.

[0004] However, traditional ticket booking systems are mostly based on single-round interactions, requiring users to manually search and filter information. This process is not only time-consuming but may also lead to errors due to improper user operations, affecting the user experience. In addition, existing dialogue systems based on large models still face challenges in multi-round interactions and specific domain tasks. For example, how to maintain contextual consistency in multi-round dialogues, how to handle complex user intentions and preferences, and how to provide accurate information and decision support during the dialogue process. In addition, there are also limitations in personalized services and user intention prediction.

[0005] In view of this, the present invention proposes a method and system for constructing a multi-round ticket booking agent based on a large model to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention provides the following technical solution: a method for constructing a multi-round ticket booking agent based on a large model, comprising the following steps:

[0007] S1. Obtain definitions of relevant elements of a multi-round ticket booking dialogue and use them as the basis for building a multi-round ticket booking agent. The definitions of relevant elements include the definition of multi-round interaction intent, the jump relationship between multi-round states, and the function definition of the ticket booking service API;

[0008] S2. Training and fine-tuning the large model, which includes task definition, dataset construction, and supervised fine-tuning;

[0009] S3. Establish a prompt word project and use it in combination with the fine-tuned large model. Establishing the prompt word project includes updating the dialogue status and calling the API.

[0010] Furthermore, the definition of the multi-round interaction intention includes:

[0011] Obtain the business requirements for multi-round ticket booking conversations and define the intent labels and related slot information for each round in the multi-round ticket booking system based on the business requirements;

[0012] The question entered by the user corresponds to an intent label, and the candidate slot information in the question entered by the user is assigned a value.

[0013] Furthermore, the jump relationship of the multi-round state includes:

[0014] A mapping relationship is established between each dialogue round and the corresponding dialogue state in a multi-round ticket booking dialogue, wherein the dialogue state of each dialogue round and the dialogue state of the next dialogue round are represented by a mapping relationship.

[0015] Furthermore, the function definition of the ticket booking service API includes:

[0016] Define the API's input and output parameters and interaction methods to adapt them to the large model.

[0017] Furthermore, the task definition includes:

[0018] Obtain and combine the business scenarios of multi-round ticket booking conversations, decompose the multi-round ticket booking conversations into structured data based on the business scenarios, and then obtain the definitions of intents, instruction sets, and large model inference tasks;

[0019] The intent is represented by structured JSON data. For each intent, it needs to include the current conversation state, vertical domain label, intent recognition label, input question, probability of intent recognition label, and slot information related to the intent.

[0020] The instruction set is a normalized representation of the system output format in the dialogue system;

[0021] The large model reasoning task is to include an intent recognition module that combines context information in multiple rounds of ticket booking conversations, and then output structured intent information through the intent recognition module.

[0022] Furthermore, the input of the intention recognition module includes: the user's current state input question, the intention state information of the previous round, and the instruction set information of the previous round, which is expressed as:

[0023] f(query n , Intent n-1 , Skill n-1 )=Intent n

[0024] In the formula, query n Indicates the input question of the user's current status in the nth round of dialogue, Intent n-1 Indicates the intention state information of the n-1th round of dialogue, Skill n-1 Indicates the instruction set information of the n-1th round of dialogue; query n , Intent n-1 , Skill n-1 Intent represents the input of the intent recognition module; n Indicates that the intent recognition module outputs the intent information of the nth round of dialogue.

[0025] Furthermore, the dataset construction includes the following processes:

[0026] Step 1: Use a small number of manually labeled samples as a seed dataset, and use a large model to infer the intent and instruction set representation corresponding to the sample data as the first version of the benchmark dataset;

[0027] Step 2: Use the GPT-4o large model to imitate the input query in the benchmark dataset to generate similar dialogues, and generate corresponding intents and instruction sets for the generated dialogue flow to obtain the dataset generated by the large model;

[0028] Step 3: Manually perform quality inspection on the data in step 2, fix any problems in the generated dataset, and store the manually reviewed and proofread data in the database to obtain the second version of the dataset.

[0029] Step 4: For the second version of the dataset, use the GPT-4o large model to rewrite synonyms, including rewriting the user input query and information in the conversation history to generate a candidate synonym dataset;

[0030] Step 5: Manually perform quality inspection on the data in step 4, fix any problems in the generated dataset, and store the manually reviewed and proofread data in the database to obtain the third version of the dataset.

[0031] Step 6: Repeat the process from step 2 to step 5, and continue to expand the data set until the sample size of the data set reaches the preset expectation.

[0032] Furthermore, the supervised fine-tuning includes:

[0033] Based on the constructed dataset, a parameter-efficient fine-tuning method is used to fine-tune the large model.

[0034] Furthermore, the dialogue state update is a task of identifying dialogue state information in multiple rounds of ticket booking dialogue scenarios;

[0035] By combining the intent recognition module and the dialogue management module in the large model reasoning task into a single reasoning module, this module is used as a dialogue state update task.

[0036] That is, the input of the dialogue state update task needs to include the user's current input, the dialogue state of the previous round, and the instructions completed by the system in the previous round. Then, its expression is:

[0037] Prompt word = character information + instruction description + intent set definition + dialogue state definition + context + guide words;

[0038] The API call is the core module for the interaction between the big model and the external ticket booking service. The big model can be used to call the third-party API in real time.

[0039] Build a multi-round ticket booking agent system based on a large model, including:

[0040] The task definition module obtains the definitions of relevant elements of the multi-round ticket booking dialogue and uses them as the basis for building the multi-round ticket booking agent. The definitions of relevant elements include the definition of multi-round interaction intentions, the jump relationship between multi-round states, and the function definition of the ticket service API;

[0041] Large model fine-tuning module, which trains and fine-tunes large models. The large model fine-tuning includes task definition, dataset construction, and supervised fine-tuning.

[0042] The prompt word project establishment module establishes the prompt word project and uses it in combination with the fine-tuned large model. The establishment of the prompt word project includes dialogue status updates and API calls.

[0043] The technical effects and advantages of the method and system for constructing a multi-round ticket booking agent based on a large model in the present invention are as follows:

[0044] 1. Leveraging a large model to complete the entire multi-round ticket booking process. This model plays a key role in multi-round ticket booking conversations (multi-round ticket booking services). It not only generates coherent and natural multi-round responses to meet user inquiry and interaction needs, but also efficiently dispatches third-party APIs to automate the entire ticket inquiry, reservation, and payment process. This integrated solution significantly improves booking efficiency, optimizes the user experience, and reduces labor costs, demonstrating the potential of AI in customer service.

[0045] 2. Through the intent recognition module, a new paradigm for intent systems is proposed. This can represent users' real needs with structured intents, and can then optimize intent pruning when using large models for intent recognition reasoning. Supervised fine-tuning can effectively reduce the difficulty of intent recognition in dialogue management tasks and optimize the quality of system interaction. Through dialogue state update tasks and API scheduling in the multi-round ticket booking agent system, it can respond to each round of user dialogue and obtain real-time flight information. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flow chart of a method for constructing a multi-round ticket booking agent based on a large model according to the present invention;

[0047] Figure 2 This is a schematic diagram of the actual use process of the present invention;

[0048] Figure 3 Schematic diagram of the system for constructing a multi-round ticket booking agent based on a large model according to the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] Example 1

[0051] See also Figure 1 As shown, the method for constructing a multi-round ticket booking agent based on a large model in this embodiment includes the following steps:

[0052] S1. Obtain the definitions of relevant elements of a multi-round ticket booking conversation (multi-round conversation) and use them as the basis for building a multi-round ticket booking agent. The definitions of relevant elements include the definition of multi-round interaction intent, the jump relationship between multi-round states, and the function definition of the ticket booking service API;

[0053] S2. Train and fine-tune the large model. Large model fine-tuning includes task definition, dataset construction, and supervised fine-tuning.

[0054] S3. Establish a prompt word project and use it in conjunction with the fine-tuned large model. Establishing the prompt word project includes conversation status updates and API calls.

[0055] Furthermore, the definition of multi-round interaction intent includes:

[0056] Obtain the business requirements for multi-round ticket booking conversations and define the intent labels and related slot information for each round in the multi-round ticket booking system based on the business requirements;

[0057] The question entered by the user corresponds to an intent label, and the candidate slot information in the question entered by the user is assigned a value.

[0058] Furthermore, the jump relationship of multiple rounds of states includes:

[0059] Establishing a mapping relationship between the dialogue states of each dialogue round in a multi-round ticket booking dialogue, wherein the dialogue state of each dialogue round is represented by the mapping relationship with the dialogue state of the next dialogue round;

[0060] Specifically, it refers to the process of mapping the state of the current conversation turn (including user input, system response, and contextual information) to the state of the next conversation turn. This process describes how the dialogue system updates and predicts the state of the next conversation turn based on user input and the current conversation context, thereby ensuring the continuity and logic of the conversation, ensuring that the conversation flows reasonably according to the contextual state, avoiding the system from falling into repetitive or unstoppable states, and maintaining a smooth user experience.

[0061] Furthermore, the function definition of the ticket booking service API includes:

[0062] Define the API's input and output parameters and interaction methods to adapt them to the large model;

[0063] Specifically, for example, the system can adapt to the resource services of third-party APIs and provide large models for calling, including ticket query, ticket information retrieval and booking functions to complete the complete multi-round ticket booking process.

[0064] Furthermore, the task definition includes:

[0065] Obtain and combine the business scenarios of multi-round ticket booking conversations, decompose the multi-round ticket booking conversations into structured data based on the business scenarios, and then obtain the definitions of intents, instruction sets, and large model inference tasks;

[0066] Intents are represented as structured JSON data. Each intent must include the current conversation state, vertical domain labels, intent recognition labels, input question, probability of intent recognition labels, and slot information (or slot KV) associated with the intent. The definition of intents involved in multi-round ticket booking is shown in the following table:

[0067]

[0068]

[0069] The instruction set is a normalized representation of the system output format in the dialogue system. In the application scenario of multi-round ticket booking dialogue, it mainly involves the TTS voice reply content and the content of the ticket query result form. Therefore, it can be represented by two instructions. The instruction set parameter set consists of the instruction type and the instruction header. The instruction type can include TTS instructions and form instructions. The instruction set involved in the system is defined as follows:

[0070]

[0071] The large model reasoning task is to include an intent recognition module that combines contextual information in a multi-round ticket booking dialogue (multi-round ticket booking agent task), and then output structured intent information through the intent recognition module;

[0072] The input of the intent recognition module includes: the user's current state input question, the intention state information of the previous round, and the instruction set information of the previous round. Its expression is:

[0073] f(query n , Intent n-1 , Skill n-1 )=Intent n

[0074] In the formula, query n Indicates the input question of the user's current status in the nth round of dialogue, Intent n-1 Indicates the intention state information of the n-1th round of dialogue, Skill n-1 Indicates the instruction set information of the n-1th round (previous round) of dialogue; query n , Intent n-1 , Skill n-1 Intent represents the input of the intent recognition module; n Indicates that the intent recognition module outputs the intent information of the nth round of dialogue.

[0075] Furthermore, dataset construction includes the following processes:

[0076] Step 1: Use a small number of manually labeled samples as a seed dataset, and use a large model to infer the intent and instruction set representation corresponding to the sample data as the first version of the benchmark dataset;

[0077] Step 2: Use the GPT-4o large model to imitate the input query in the benchmark dataset to generate similar dialogues, and generate corresponding intents and instruction sets for the generated dialogue flow to obtain the dataset generated by the large model;

[0078] Step 3: Manually perform quality inspection on the data in step 2, fix any problems in the generated dataset, and store the manually reviewed and proofread data in the database to obtain the second version of the dataset.

[0079] Step 4: For the second version of the dataset, use the GPT-4o large model to rewrite synonyms, including rewriting the user input query and information in the conversation history to generate a candidate synonym dataset.

[0080] Step 5: Manually perform quality inspection on the data in step 4, fix any problems in the generated dataset, and store the manually reviewed and proofread data in the database to obtain the third version of the dataset.

[0081] Step 6: Repeat the process from step 2 to step 5, and continue to expand the data set until the sample size of the data set reaches the preset expectation. The preset expectation can be set to 50,000 samples in the data set.

[0082] Furthermore, supervised fine-tuning includes:

[0083] Based on the constructed dataset, a parameter-efficient fine-tuning method (such as Parameter-Efficient Fine-Tuning, PEFT, Lora algorithm) is used to fine-tune the large model (such as ChatGLM3);

[0084] Specifically, the PEFT method is characterized by fixing most pre-trained parameters, requiring only minor or additional model parameter fine-tuning, significantly reducing computational and storage costs and training time. Furthermore, advanced PEFT technology offers comparable performance to full fine-tuning. This fine-tuning process is primarily designed to improve the inference accuracy of large models for complex tasks such as intent recognition, thereby achieving the goal of industrialized business implementation.

[0085] Furthermore, conversation state update is the task of identifying conversation state information in a multi-round ticket booking conversation scenario;

[0086] By combining the intent recognition module and the dialogue management module in the large model reasoning task into a single reasoning module, this module is used as a dialogue state update task.

[0087] That is, the input of the dialogue state update task needs to include the user's current input, the dialogue state of the previous round, and the instructions completed by the system in the previous round. Then, its expression is:

[0088] Prompt word = character information + instruction description + intent set definition + dialogue state definition + context + guide words.

[0089]

[0090]

[0091]

[0092] Specifically, the prompt word engineering of the present invention is also a key component of the multi-round ticket booking intelligent entity, which can directly affect whether the large model can understand user needs and complete the complete ticket booking business process; the prompt word engineering guides the large model to output unified and efficient instructions according to system requirements, so as to better and intelligently collaborate with users to complete voice interactive ticket booking needs.

[0093] Furthermore, API calls are the core component of the interaction between the big model and external ticket booking services. The big model can be used to call third-party APIs in real time to obtain real-time flight information.

[0094] Specifically, in this invention, the API involved mainly includes the flight information API of XX Airlines. The optimization of the prompt word of the API call mainly infers the corresponding API call parameter information from the result of the conversation state update, and then uses the ability of the large model to call the API to complete the intelligent interaction in the current conversation state. The main API parameters involved in this invention are as follows:

[0095] parameter Parameter Type Parameter Description Example leaveTime String Earliest departure time "2024-08-29 15:00:00” arriveTime String Latest arrival time "2024-08-29 20:00:00” fromCityId Int Departure city ID 12 toCityId Int Destination City Id 23 minPrice Float Lowest Price 500 maxPrice Float Highest Price 1000

[0096] Furthermore, after combining the fine-tuned large model with the prompt word project, the present invention also supports multiple deployment methods, including command line calls, GUI deployment, and API deployment, to build a multi-round ticket booking agent with voice interaction capabilities, providing users with highly intelligent ticket booking services in commercial scenarios.

[0097] It should be noted that the reference Figure 2 For example, if a user wants to book a flight from Hangzhou to Beijing through an intelligent ticket booking system, the user will conduct multiple rounds of booking dialogues with the system through natural language to complete the booking process. The process is as follows:

[0098] The task definition first defines the user's intent system and output structure. For example, the intent of "book a flight" and its related slot information, such as departure point, destination, and date, are defined. For example, if a user says, "Book me a flight to Beijing," the system recognizes this as the intent of "book a flight" and extracts the slot information, such as departure point and destination.

[0099] Using large models (such as GPT-4o) to generate a benchmark dataset, the dataset is then continuously expanded and refined through manual review and modification. For example, the system generates multiple queries like "Book me a flight to Beijing" and annotates these queries with intent and instruction sets.

[0100] Use efficient fine-tuning methods (such as the PEFT method) to fine-tune large models to improve the accuracy of intent recognition. For example, the fine-tuned model can more accurately recognize the intent of user input and generate corresponding instruction sets, such as TTS instructions and form instructions.

[0101] By updating the conversation state with clear prompts, the large model is guided to understand the user's needs and update the conversation state. For example, when the user says "departing from Hangzhou", the prompt guides the model to update the departure slot information and ask the user for the next slot information (such as date).

[0102] Key information, such as prompts, includes persona information, instruction descriptions, and intent definitions. This information helps the model understand its role and tasks. For example, the prompt clearly tells the model that it is an airline ticket salesperson and needs to update slot information based on user input.

[0103] The output format is specified through prompt word engineering to ensure that the output is structured and easy-to-parse instructions. For example, the output dialogue status update results must follow a specific JSON format to facilitate system parsing and execution. This allows the large model to call third-party APIs in real time to obtain real-time flight information.

[0104] In this embodiment, the accuracy of intent recognition is improved by focusing on the construction of the dataset and the fine-tuning of the model through large-scale model fine-tuning; the prompt word engineering focuses on carefully designed prompt words to guide the model to understand and respond to user needs and standardize the output format; supervised learning can improve the accuracy of the model in intent recognition tasks; the standardized output format and clear task guidance of the prompt word engineering can improve the interaction efficiency and experience between users and the system; by continuously expanding the dataset and optimizing the prompt words, the present invention can adapt to more business scenarios and user needs.

[0105] Example 2

[0106] See also Figure 3 As shown, the system for constructing a multi-round ticket booking agent based on a large model described in this embodiment includes:

[0107] The task definition module obtains the definitions of relevant elements of the multi-round ticket booking dialogue and uses them as the basis for building the multi-round ticket booking agent. The definitions of relevant elements include the definition of multi-round interaction intentions, the jump relationship between multi-round states, and the function definition of the ticket service API;

[0108] Large model fine-tuning module, which trains and fine-tunes large models. The large model fine-tuning includes task definition, dataset construction, and supervised fine-tuning.

[0109] The prompt word project establishment module establishes the prompt word project and uses it in combination with the fine-tuned large model. The establishment of the prompt word project includes dialogue status updates and API calls.

[0110] Specifically, in the present invention, the large model can be flexibly configured according to business needs and application performance, and is not limited to a certain model; for example, the large models that support local deployment and supervised fine-tuning mainly include BaiChuan2, ChatGLM4 and Qwen; the invention mainly uses ChatGLM4-6B as the base large model for fine-tuning; if the business needs require calling the reasoning service of the third-party large model in the form of an API (such as Qwen2-72B, etc.), the locally deployed reasoning service can be replaced with the corresponding third-party large model reasoning service API for application; in addition, in the fine-tuning process of the large model, the PEFT algorithm used is not limited to the LoRA algorithm, such as QLoRA, AdaLoRA, P-Tuning, Prefix-Tuning, Prompt-Tuning, etc. are all commonly used efficient fine-tuning algorithms, which can be comprehensively considered and selected according to actual conditions such as fine-tuning data sets, LLM model characteristics, and performance factors to achieve better results;

[0111] It's important to note that through fine-tuning the large model and comprehensive optimization of prompt word engineering, the large model infers each process of multi-round dialogues, gradually completing the entire ticket booking process. This innovation can be widely extended to other multi-round interaction scenarios, such as booking hotels and ordering takeout. Furthermore, in terms of application, this innovation can be effectively applied to various intelligent products, such as smart assistants and digital figurines. Users can interact with the system through voice interaction, providing a complete end-to-end voice ticket booking experience.

[0112] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0113] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only one type. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0114] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

[0115] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a multi-round ticket booking agent based on a large model, characterized in that: The following steps are involved: S1. Obtain definitions of relevant elements of a multi-round ticket booking dialogue and use them as the basis for building a multi-round ticket booking agent. The definitions of relevant elements include the definition of multi-round interaction intent, the jump relationship between multi-round states, and the function definition of the ticket booking service API; S2. Training and fine-tuning the large model, which includes task definition, dataset construction, and supervised fine-tuning; The task definition includes: Obtain and combine the business scenarios of multi-round ticket booking conversations, decompose the multi-round ticket booking conversations into structured data based on the business scenarios, and then obtain the definitions of intents, instruction sets, and large model inference tasks; The intent is represented by structured JSON data. For each intent, it needs to include the current conversation state, vertical domain label, intent recognition label, input question, probability of intent recognition label, and slot information related to the intent. The instruction set is a normalized representation of the system output format in the dialogue system; The large model reasoning task in the multi-round ticket booking dialogue includes an intent recognition module that combines context information and then outputs structured intent information through the intent recognition module; The input of the intention recognition module includes: the user's current state input question, the intention state information of the previous round, and the instruction set information of the previous round, which is expressed as: f(query n ,Intent n-1 ,Skill n-1 )=Intent n In the formula, query n Indicates the input question of the user's current status in the nth round of dialogue, Intent n-1 Indicates the intention state information of the n-1th round of dialogue, Skill n-1 Indicates the instruction set information of the n-1th round of dialogue; query n , Intent n-1 , Skill n-1 Intent represents the input of the intent recognition module; n Indicates that the intention recognition module outputs the intention information of the nth round of dialogue; S3. Establish a prompt word project and use it in combination with the fine-tuned large model. Establishing the prompt word project includes updating the dialogue status and calling the API.

2. The method for constructing a multi-round ticket booking agent based on a large model according to claim 1, characterized in that: The definition of the multi-round interaction intention includes: Obtain the business requirements for multi-round ticket booking conversations and define the intent labels and related slot information for each round in the multi-round ticket booking system based on the business requirements; The question entered by the user corresponds to an intent label, and the candidate slot information in the question entered by the user is assigned a value.

3. The method for constructing a multi-round ticket booking agent based on a large model according to claim 1, characterized in that: The jump relationship of the multi-round state includes: A mapping relationship is established between each dialogue round and the corresponding dialogue state in a multi-round ticket booking dialogue, wherein the dialogue state of each dialogue round and the dialogue state of the next dialogue round are represented by a mapping relationship.

4. The method for constructing a multi-round ticket booking agent based on a large model according to claim 1, characterized in that: The function definition of the ticket booking service API includes: Define the API's input and output parameters and interaction methods to adapt them to the large model.

5. The method for constructing a multi-round ticket booking agent based on a large model according to claim 1, characterized in that: The dataset construction includes the following processes: Step 1: Use a small number of manually labeled samples as a seed dataset, and use a large model to infer the intent and instruction set representation corresponding to the sample data as the first version of the benchmark dataset; Step 2: Use the large model to simulate the input query in the benchmark dataset to generate similar dialogues, and generate corresponding intents and instruction sets for the generated dialogue flow to obtain the dataset generated by the large model; Step 3: Manually perform quality inspection on the data in step 2, fix any problems in the generated dataset, and store the manually reviewed and proofread data in the database to obtain the second version of the dataset. Step 4: For the second version of the dataset, use the large model to rewrite synonyms. This involves rewriting the user's input query and information from the conversation history to generate a candidate dataset of synonyms. Step 5: Manually perform quality inspection on the data in step 4, fix any problems in the generated dataset, and store the manually reviewed and proofread data in the database to obtain the third version of the dataset. Step 6: Repeat the process from step 2 to step 5, and continue to expand the data set until the sample size of the data set reaches the preset expectation.

6. The method for constructing a multi-round ticket booking agent based on a large model according to claim 5, characterized in that: The supervised fine-tuning includes: Based on the constructed dataset, a parameter-efficient fine-tuning method is used to fine-tune the large model.

7. The method for constructing a multi-round ticket booking agent based on a large model according to claim 6, characterized in that: The dialogue state update is a task to identify dialogue state information in a multi-round ticket booking dialogue scenario; By combining the intent recognition module and the dialogue management module in the large model reasoning task into a single reasoning module, this module is used as a dialogue state update task. The API call is the core module for the interaction between the big model and the external ticket booking service. The big model can be used to call the third-party API in real time.

8. A system for constructing a multi-round ticket booking agent based on a large model, applied to the method according to any one of claims 1 to 7, characterized in that: include: The task definition module obtains the definitions of relevant elements of the multi-round ticket booking dialogue and uses them as the basis for building the multi-round ticket booking agent. The definitions of relevant elements include the definition of multi-round interaction intentions, the jump relationship between multi-round states, and the function definition of the ticket service API; Large model fine-tuning module, which trains and fine-tunes large models. The large model fine-tuning includes task definition, dataset construction, and supervised fine-tuning. The prompt word project establishment module establishes the prompt word project and uses it in combination with the fine-tuned large model. The establishment of the prompt word project includes dialogue status updates and API calls.

Images