Airline ticket booking system and method based on large language model

Abstract

The invention discloses an air ticket booking system and method based on a large language model, and the system comprises a data collection module which is used for collecting the current natural language query, historical dialogue records, real-time position information, historical behavior data and user portrait tags of a user after the authorization of the user is obtained; the query parameter structured processing module is used for carrying out space-time semantic analysis processing on the acquired data of the data acquisition module and outputting structured query parameters containing a departure place identifier, a destination identifier and departure time information; and the intention recognition module is used for performing intention classification reasoning through a pre-trained large language model based on the current natural language query, the historical dialogue record and the structured query parameters of the user, and outputting a fine-grained intention category comprising at least one of air ticket query, special-price air tickets, multi-day air tickets, price prediction and price reduction reminding. According to the system and the method, the query accuracy, the service suitability and the interaction efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of aviation travel technology, and in particular to a flight ticket booking system and method based on a large language model. Background Technology

[0002] With the rapid development of online travel services (OTAs) and airline direct sales platforms, users have become accustomed to searching for and booking air tickets through digital channels. Traditional air ticket booking systems are usually based on structured form input, requiring users to explicitly fill in fixed "slot" information such as departure city, arrival city, and departure date.

[0003] To improve convenience, some systems have introduced simple Natural Language Processing (NLP) techniques or keyword matching mechanisms to attempt to understand user text input. However, these existing technologies still have significant limitations when dealing with the complex and ambiguous travel intentions of users in the real world, mainly in the following aspects: Insufficient understanding of fuzzy and unstructured queries: Users often use natural language containing vague timeframes, geographical preferences, and complex intentions when querying, such as "Where in the south is cheap after National Day?" or "Take the kids to the beach for summer vacation." Existing technical solutions mostly rely on keyword dictionaries and rule matching based on fixed patterns, making it difficult to accurately interpret the specific date ranges or precise geographical locations corresponding to non-standard expressions such as "after National Day," "summer vacation," "southern," and "beach." This often leads to the system failing to understand the query, returning incorrect results, or directly requiring the user to re-enter the information, causing query interruptions and a degraded user experience.

[0004] The current solutions suffer from coarse-grained intent recognition and limited scenario adaptability: They typically only identify the general intent of "searching for flights," failing to differentiate between specific scenarios. For example, they cannot determine whether a user wants to find fixed flights for a specific date, understand price trends over a certain period, actively seek special offers, or monitor price fluctuations for alerts. Due to imprecise intent recognition, the system provides a generic flight list, unable to offer differentiated data services and presentations based on the user's true purpose, leading to either information overload or information insufficiency.

[0005] Lack of integration and utilization of multi-dimensional context: Users' travel decisions are often influenced by a combination of historical behavior, real-time context, and personal preferences. Existing technical solutions typically process single queries in isolation, failing to effectively integrate and utilize multi-source contextual information such as user history, search and order data, real-time location, and user profile tags. For example, when a user queries "Let's go to Shanghai," if the system cannot combine it with the previous sentence "flights from Beijing to Guangzhou," it cannot understand that "Shanghai" is a new destination. This lack of context awareness leads to a lack of coherence and personalization in recommendation results.

[0006] The presentation of search results is rigid and inefficient: even when a user's query is successfully parsed, existing systems typically present flight results in a traditional, information-dense list format. This presentation lacks targeted optimization for different query scenarios, requiring users to manually filter and compare the information they need, resulting in a long decision-making path and low interaction efficiency. Furthermore, the process from browsing results to placing an order is fragmented, with insufficient conversion guidance. Summary of the Invention

[0007] To address the technical problems existing in the background art, this invention proposes a flight ticket booking system and method based on a large language model.

[0008] This invention proposes a flight booking system based on a large language model, comprising: The data acquisition module is used to collect the user's current natural language query, historical dialogue records, real-time location information, historical behavior data and user profile tags after obtaining user authorization; The query parameter structuring module is used to perform spatiotemporal semantic parsing on the data collected by the data acquisition module and output structured query parameters containing departure location identifier, destination identifier and departure time information. The intent recognition module is used to perform intent classification reasoning based on the user's current natural language query, historical dialogue records and structured query parameters, through a pre-trained large language model, and output fine-grained intent categories including at least one of flight query, discounted flight, multi-day flight, price prediction and price reduction reminder. The contextualized service routing module is used to route structured query parameters to the corresponding contextualized data service interface based on fine-grained intent categories. The data service and card rendering module has multiple data service interfaces and rendering engines corresponding to different intent categories. It is used to call the interface to obtain flight ticket data based on the routing results and generate front-end interactive card data using the corresponding rendering template. The interactive booking guidance module is used to send front-end interactive card data to the client for display and guide users to complete the booking process through card interaction events.

[0009] Preferably, the query parameter structuring module is specifically used for: The origin identifier is determined based on the collected data. The priority of determining the origin identifier is as follows: directly extract it from the user's current natural language query, or extract commonly used origins from historical behavior data, or use the city corresponding to the real-time location information by default. Process the natural language destination descriptions in the collected data; If a standard city code can be matched, it will be used directly as the destination identifier; otherwise, the natural language destination description will be matched with a predefined city cluster label library, and the set of city codes associated with the successfully matched labels will be used as the destination identifier. The natural language departure time description in the collected data is parsed, and based on the preset date calculation rules and holiday rules, the natural language departure time description is converted into a specific departure date or date range as departure time information.

[0010] Preferably, the city cluster tag library is a multi-dimensional, dynamically updatable knowledge base used to map fuzzy geographical preference semantics to a specific set of city codes.

[0011] Preferably, in the intent recognition module, the intent classification reasoning process includes: The user's current natural language query, contextual information from historical dialogue records, and structured query parameters are combined into a formatted input according to a preset prompt template. The formatted input is fed into a pre-trained large language model, which identifies the user's intent based on the input content and outputs a fine-grained intent category.

[0012] Preferably, the pre-trained large language model is configured to distinguish and output different fine-grained intent categories based on the destination identifier type, departure time information type, and semantic features in the user's current natural language query in the structured query parameters.

[0013] Preferably, the scenario-based service routing module is specifically used for: Receive fine-grained intent categories from the intent recognition module and structured query parameters from the query parameter structuring module; Based on the preset intent-interface mapping relationship, fine-grained intent categories are mapped to the corresponding contextualized data service interfaces; The structured query parameters are distributed to the contextualized data service interface that is mapped to trigger subsequent data queries; The intent-interface mapping relationship includes at least the following: mapping the intent to purchase discounted air tickets to the discounted air ticket query interface, mapping the intent to query air tickets to the real-time flight query interface, mapping the intent to purchase air tickets for multiple days to the multi-day price trend query interface, mapping the intent to predict prices to the price prediction analysis interface, and mapping the intent to remind users of price reductions to the price reduction monitoring and subscription interface.

[0014] Preferably, the data service and card-based rendering module is specifically used for: Based on the routing results provided by the scenario-based service routing module, call the specific data service interface bound to the fine-grained intent category to obtain a set of flight ticket data results that conform to the structured query parameters; The flight ticket data result set is input into the rendering template corresponding to the fine-grained intent category. The rendering engine performs structured transformation and style rendering to generate the front-end interactive card data containing core flight information and booking guidance controls.

[0015] Preferably, the interactive booking guidance module is specifically used for: The front-end interactive card data is sent to the client, which then renders and displays the corresponding visual interactive card. In response to user interaction with the booking guidance controls within the card, a deep booking link containing structured query parameters is triggered; The system's booking interface is invoked. Based on the structured query parameters and user identity information, the system generates and redirects the user to the corresponding flight class selection and order form filling page to guide the user through the subsequent booking process.

[0016] Preferably, the system further includes a price reduction monitoring service module for responding to price reduction alerts in fine-grained intent categories: Create a monitoring task for structured query parameters; The system initiates a background process to periodically compare prices and sends a notification message through the user's pre-selected notification channel when the preset price reduction conditions are met.

[0017] This invention proposes a flight booking method based on a large language model, comprising the following steps: S1. Collect the user's current natural language query, historical dialogue records, real-time location information, historical behavior data, and user profile tags; S2. Perform spatiotemporal semantic parsing on the collected data to generate structured query parameters containing departure location identifier, destination identifier, and departure time information; S3. Based on the user's current natural language query, historical dialogue records, and structured query parameters, perform intent classification reasoning through a pre-trained large language model and output fine-grained intent categories. S4. Based on the fine-grained intent category, route the structured query parameters to the corresponding contextualized data service interface; S5. Call the corresponding scenario-based data service interface to obtain flight ticket data, and use the rendering template bound to the fine-grained intent category to generate front-end interactive card data; S6. Send the front-end interactive card data to the client for display, and guide the user to complete the booking process through card interaction events.

[0018] This invention proposes a flight booking system and method based on a large language model. By introducing deep fusion of a pre-trained large language model and multi-source contextual data, it constructs a complete intelligent link from fuzzy natural language queries to precise flight booking services. This solves the technical bottleneck of traditional solutions' difficulty in understanding unstructured and fuzzy spatiotemporal semantic queries, improving the accuracy and robustness of query parsing. Through fine-grained intent recognition and intelligent scene routing, it achieves precise matching and on-demand allocation of service functions, optimizing the utilization efficiency of backend computing resources. The use of scenario-based and card-based result rendering and interactive guidance shortens the user's path from information acquisition to decision-making, reducing interaction complexity and cognitive load. This improves the intelligence level, response efficiency, and user experience of the flight query and booking process. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system architecture of a flight booking system based on a large language model proposed in this invention; Figure 2 This is a schematic diagram illustrating the workflow of a flight booking method based on a large language model proposed in this invention. Detailed Implementation

[0020] Reference Figure 1 and Figure 2 The present invention proposes a flight booking system based on a large language model, comprising: The data acquisition module is used to collect the user's current natural language query, historical dialogue records, real-time location information, historical behavior data, and user profile tags after obtaining user authorization.

[0021] It should be noted that the user's current natural language query is the text entered by the user in this instance, such as "cheap flights to Sanya next month". The historical dialogue record is the content of the most recent rounds (e.g., the last 5 rounds) of dialogue in the current session, used to understand the context. Real-time location information is the latitude and longitude and city name of the user's current location, obtained through device GPS or IP positioning. Historical behavior data includes the user's flight search records (departure city, destination, date) and completed historical flight order data over a past period (e.g., 90 days). User profile tags are tags derived from the analysis of the user's historical behavior, such as "price-sensitive", "business travel preference", "preferred airline", etc.

[0022] The query parameter structured processing module is used to perform spatiotemporal semantic parsing on the data collected by the data acquisition module, and output structured query parameters containing departure location identifier, destination identifier and departure time information.

[0023] In this embodiment, the query parameter structuring module is specifically used for: determining the departure point identifier based on the collected data, with the priority of determining the departure point identifier being: directly extracting it from the user's current natural language query, or extracting commonly used departure points from historical behavior data, or using the city corresponding to the real-time location information by default; processing the natural language destination description in the collected data; if it can match a standard city code, it is directly used as the destination identifier; otherwise, the natural language destination description is matched with a predefined city cluster tag library, and the set of city codes associated with the successfully matched tags is used as the destination identifier; parsing the natural language departure time description in the collected data, and converting the natural language departure time description into a specific departure date or date range based on preset date calculation rules and holiday rules, so as to serve as departure time information.

[0024] Specifically, the city cluster tag library is a multi-dimensional and dynamically updated knowledge base used to map fuzzy geographical preference semantics to a specific set of city codes.

[0025] Specifically, the structured query parameter is a data structure containing three key fields: departure point identifier, destination identifier, and departure time information. Its processing flow is as follows: Determination of departure point: A three-tier priority strategy is adopted. First, an attempt is made to directly parse the explicit departure point from the user's current natural language query (e.g., "Beijing" in "flying from Beijing to Shanghai"). If not found, historical behavior data is queried to extract the most frequently used departure city. If still no result is found, the city corresponding to the user's real-time location information is used as the default departure point. The final result needs to be converted into the three-letter city code specified by the International Air Transport Association (IATA) (e.g., Beijing: PEK).

[0026] Destination Identifier Determination: The processing logic consists of two steps. First, it attempts to precisely match the destination description in the query (e.g., "Sanya") with a standard three-letter code library for cities. If the match is successful, the three-letter code is directly used as the destination identifier. Second, if the match fails (e.g., the user inputs "seaside" or "summer resort city"), fuzzy matching is initiated. The system maintains a multi-dimensional, dynamically updated city cluster tag library. This knowledge base contains tags across multiple dimensions, such as "coastal cities," "tourist cities," "historical cities," "southern cities," and "summer resort cities." Each tag is associated with a specific list of three-letter codes for cities (e.g., the "coastal cities" tag is associated with [SYX (Sanya), XMN (Xiamen), TAO (Qingdao)...]). The system calculates the semantic similarity between the user's fuzzy description and all tags, selects the tag with the highest matching degree, and uses its associated list of cities as the destination identifier. This is a list, not a single code.

[0027] Determining departure time information: For vague time expressions such as "next week", "summer vacation", "after National Day", etc., the module has a built-in set of date calculation rules and holiday rules.

[0028] It should be noted that the rule uses the current date as the base date T for calculation. For example: "Next week": The output date range is [Monday of the week following T, Sunday of the week following T].

[0029] "Summer Vacation": Based on preset rules (such as July 1 to August 31), output the specific date range [202X-07-01, 202X-08-31].

[0030] "After National Day": This refers to the first week after the National Day holiday (October 1st to 7th), namely [202X-10-08, 202X-10-14].

[0031] "End of month": This refers to the last 5 days of the current month.

[0032] By parsing the departure time description in natural language, the final output is a specific date (such as "202X-12-25") or a date range (such as [202X-07-01, 202X-07-07]) as the departure time information.

[0033] The intent recognition module is used to perform intent classification and reasoning based on the user's current natural language query, historical dialogue records, and structured query parameters through a pre-trained large language model, and outputs a fine-grained intent category including at least one of the following: flight search, discounted flight, multi-day flight, price prediction, and price reduction reminder.

[0034] In this embodiment, the intent classification reasoning process in the intent recognition module includes: combining the user's current natural language query, contextual information from historical dialogue records, and structured query parameters into a formatted input according to a preset prompt template; sending the formatted input into a pre-trained large language model, which identifies the user's intent based on the input content and outputs a fine-grained intent category.

[0035] In this embodiment, the pre-trained large language model is configured to distinguish and output different fine-grained intent categories based on the destination identifier type, departure time information type and semantic features in the user's current natural language query in the structured query parameters.

[0036] Specifically, after training, the large language model establishes a mapping relationship between its output and input features as follows: When the destination identifier in the structured query parameters is a specific city code and the departure time information is a single date, the "flight search" intent is output. When the destination is identified by a city cluster label, or when the user's current natural language query contains the semantic keyword "discounted airfare", output the intent "discounted airfare". When the destination is identified as a specific city code and the departure time information is a date range, output the intent of "multi-day flight". When the destination is identified by a specific city code, the departure time is a single date, and the user's current natural language query contains the semantic keyword "price prediction", output the intent "price prediction". When the destination is identified by a specific city code, the departure time is a single date, and the user's current natural language query contains the semantic keyword for reminder subscription, output the intent of "price reduction reminder".

[0037] Specifically, the input to the intent recognition module is the user's current natural language query, the user's historical dialogue records, and the structured query parameters output by the query parameter structuring module. The processing procedure of the intent recognition module includes: Input Formatting: First, the above multi-source inputs are concatenated according to the preset prompt word template to form a well-formatted input. For example: "User's current query: {query text}. Historical dialogues: {last 3 rounds of dialogue}. Parsed information: origin: {origin identifier}, destination: {destination identifier}, time: {departure time information}. Please determine the user's intent."

[0038] Model Inference: Formatted input is fed into a pre-trained large language model, which establishes a complex mapping between input features and intent categories. Its inference logic is as follows: When the destination identifier in the structured query parameters is a single city's three-letter code and the departure time information is a single date, the model tends to judge it as a precise "flight search" intent. When the destination identifier is a list of city cluster tags, or when the query text contains special offer semantic keywords such as "special offer" or "cheap," the model judges it as a "discounted flight" intent. When the departure time information is a date range, it is judged as a "multi-day flight" intent. If, based on a single city and a single date, the query text contains price prediction semantic keywords such as "when to buy cheap" or "price trend," it is judged as a "price prediction" intent. If it contains reminder subscription semantic keywords such as "remind me when prices drop" or "notify me," it is judged as a "price drop reminder" intent. The model ultimately outputs a definite, fine-grained intent category.

[0039] It should be noted that this application uses a large-scale language model based on a Decoder-only Transformer architecture as its foundation. To adapt it for the flight ticket intent recognition task, a two-stage training process was implemented: Pre-training: The model is first pre-trained on a massive amount of general internet text corpus to obtain general language understanding and generation capabilities.

[0040] Domain Adaptation and Fine-tuning: Subsequently, supervised fine-tuning was performed using a specially constructed airline ticketing domain dataset. This dataset consists of two parts: (a) an airline ticketing domain corpus: text collected from airline websites, OTA platforms, and travel communities, containing information such as flight details, routes, fares, and destination descriptions; and (b) labeled query-intent pairs: a large number of de-identified real-world user query conversations, labeled with corresponding fine-grained intent category labels (such as flight search, discounted tickets, etc.) by professionals according to the category system defined by this invention. During fine-tuning, the cross-entropy loss function was used, with a learning rate of 3e-5, a batch size of 16, and 5 training epochs, aiming to enable the model to learn the mapping relationship between fuzzy queries and specific intents.

[0041] Intent reasoning process: In the application phase, the above multi-source inputs are first concatenated into a formatted input according to a preset prompt template, for example: "User query: [query text]. Historical context: [last 3 rounds of dialogue]. Known information: departure point: [identifier], destination: [identifier], time: [information]. Please determine the intent." This formatted input is then fed into the trained model. The model uses its internal parameters to comprehensively calculate the features in the input (such as whether the destination identifier is a specific city or a list of city clusters, whether the departure time is a point or a range, and whether the query text contains specific semantic keywords), and finally outputs a definite fine-grained intent category. This process is essentially a classification reasoning performed by the model using the complex patterns of semantics and user behavior patterns in the air ticket domain learned during the training phase.

[0042] The contextualized service routing module is used to route structured query parameters to the corresponding contextualized data service interface based on fine-grained intent categories.

[0043] In this embodiment, the scenario-based service routing module is specifically used to: receive fine-grained intent categories from the intent recognition module and structured query parameters from the query parameter structuring module; map the fine-grained intent categories to the corresponding scenario-based data service interfaces according to the preset intent-interface mapping relationship; and distribute the structured query parameters to the mapped scenario-based data service interfaces to trigger subsequent data queries; wherein, the intent-interface mapping relationship includes at least: mapping the special offer ticket intent to the special offer ticket query interface, mapping the ticket query intent to the real-time flight query interface, mapping the multi-day ticket intent to the multi-day price trend query interface, mapping the price prediction intent to the price prediction analysis interface, and mapping the price reduction reminder intent to the price reduction monitoring and subscription interface.

[0044] Specifically, the contextualized service routing module directs queries to different service terminals based on the identified intent. Internally, the contextualized service routing module maintains an intent-interface mapping configuration table. The processing steps of the contextualized service routing module include: Reception and Mapping: The contextualized service routing module receives fine-grained intent categories from the intent recognition module and structured query parameters from the query parameter structuring module. Based on the mapping table, it maps the intents to the corresponding contextualized data service interfaces. For example: "Special offer airfare" Discounted flight ticket search interface; "Flight Search" Intent Real-time flight query interface; "Multi-day flight" intention Multi-day price trend query interface; "Price prediction" intention Price forecasting and analysis interface; "Price Reduction Alert" Intent Price drop monitoring and subscription interface.

[0045] Parameter distribution: Subsequently, the scenario-based service routing module distributes the structured query parameters as request parameters to the mapped backend interface, triggering the subsequent data acquisition process.

[0046] The data service and card-based rendering module has multiple data service interfaces and rendering engines corresponding to different intent categories. It is used to call the interface to obtain flight ticket data based on the routing results and generate front-end interactive card data using the corresponding rendering template.

[0047] In this embodiment, the data service and card rendering module is specifically used to: call a specific data service interface bound to a fine-grained intent category based on the routing results provided by the scenario-based service routing module to obtain a set of flight ticket data results that conform to the structured query parameters; input the set of flight ticket data results into the rendering template corresponding to the fine-grained intent category, and perform structured transformation and style rendering through the rendering engine to generate front-end interactive card data containing core flight information and booking guidance controls.

[0048] Specifically, the data service and card-based rendering module is responsible for acquiring data and transforming its presentation format. The processing steps of the data service and card-based rendering module include: Data Acquisition: Based on the routing results provided by the scenario-based service routing module, the corresponding data service interfaces are invoked. For example, when invoking the discounted flight ticket query interface, the destination identifier containing a list of city clusters and the date range are passed in. The interface returns a list of the best discounted flights from the departure point to each destination within the cluster within that range, forming a flight ticket data result set.

[0049] Card-based rendering: The system predefines different rendering templates for each fine-grained intent category. The template defines the card's layout, style, and the data fields to be displayed. The rendering engine populates the corresponding intent's template with data from the flight ticket data set, performs structured transformation and style rendering, and finally generates a front-end interactive card data set. This front-end interactive card data set is typically a JSON or HTML fragment containing core flight information (time, price, airline) and booking guidance controls such as "Book," "View Details," and "Subscribe for Price Reduction."

[0050] The interactive booking guidance module is used to send front-end interactive card data to the client for display and guide users to complete the booking process through card interaction events.

[0051] In this embodiment, the interactive booking guidance module is specifically used to: send front-end interactive card data to the client, drive the client to render and display the corresponding visual interactive card; respond to the user's operation on the booking guidance control in the card, trigger the deep booking link containing structured query parameters; call the system booking interface, generate and jump to the corresponding flight cabin selection and order filling page according to the structured query parameters and user identity information, so as to guide the user to complete the subsequent booking operation.

[0052] Specifically, the processing steps of the interactive reservation guidance module include: Card Display: The front-end interactive card data is sent to the user's client. The client then renders a visual interactive card based on this data and displays it to the user.

[0053] Booking Guide: When a user clicks the "Book" button on the card, the module captures this interaction event. It then generates a deep booking link that encodes the structured query parameters (such as flight number and date). By calling the system's unified booking interface, carrying this link and the user's identity information, the system automatically redirects to a pre-filled flight class selection and order form page, seamlessly guiding the user through the entire booking and payment process.

[0054] In this embodiment, the system also includes a price reduction monitoring service module, used to respond to price reduction alerts in fine-grained intent categories: Create a monitoring task for structured query parameters; The system initiates a background process to periodically compare prices and sends a notification message through the user's pre-selected notification channel when the preset price reduction conditions are met.

[0055] Example 1

[0056] In this example, the user input is: "Where is the cheapest place to go to the beach with my child during summer vacation?"

[0057] The system processing procedure corresponding to this embodiment includes: Data acquisition module: Collects the query text. Simultaneously, it obtains the user's location (Beijing), historical orders (previously booked flights to Sanya), and the "family travel" profile tag.

[0058] Query parameter structure processing module: Origin: The user did not specify. Historically, Beijing has been the most frequently used origin, and the location is also Beijing. Therefore, the origin identifier is determined to be PEK.

[0059] Destination: The query "seaside" could not find a specific city. Matching with the city cluster tag library, the semantics highly match the "coastal city" tag. This tag is associated with a list of cities such as [SYX (Sanya), XMN (Xiamen), TAO (Qingdao), ...]. This list will be used as the destination identifier.

[0060] Departure time: "Summer vacation" is interpreted as a date range [2024-07-01, 2024-08-31] according to the rules, and is used as departure time information.

[0061] Intent Recognition Module: The original query and parsed parameters (PEK, [list of coastal cities], summer travel range) are formatted and input into the large language model. Model analysis features: destination is tagged as a city cluster, query contains the keyword "cheap," and user has a family tag. After comprehensive judgment, the fine-grained intent category is output as "discounted airfare."

[0062] Contextualized service routing module: Based on the mapping relationship, the intent of "discounted air tickets" is routed to the discounted air ticket query interface, and the parameters {departure point: PEK, destination list: [SYX, XMN, ...], time range: [2024-07-01, 2024-08-31]} are distributed there.

[0063] Data service and card-based rendering module: Calls the discounted flight ticket query interface to obtain a list of discounted flights from Beijing to various coastal cities during the summer vacation. Using the rendering template corresponding to "Discounted Flight Tickets," it generates a front-end interactive card. The card's top title is "Summer Beach Special Offers," and below it, multiple discounted flight options to destinations such as Sanya, Xiamen, and Qingdao are displayed in list format, with a "Book" button after each option.

[0064] Interactive booking guidance module: The card is presented in the user's app or system interface. When the user clicks the "Book" button under the "Sanya" flight, the system automatically redirects to the order page, where flight and date information is automatically filled in. The user only needs to add passenger information to complete the payment.

[0065] Reference Figure 1 and Figure 2The present invention proposes a flight booking method based on a large language model, comprising the following steps: S1. Collect the user's current natural language query, historical dialogue records, real-time location information, historical behavior data, and user profile tags; S2. Perform spatiotemporal semantic parsing on the collected data to generate structured query parameters containing departure location identifier, destination identifier, and departure time information; S3. Based on the user's current natural language query, historical dialogue records, and structured query parameters, perform intent classification reasoning through a pre-trained large language model and output fine-grained intent categories. S4. Based on the fine-grained intent category, route the structured query parameters to the corresponding contextualized data service interface; S5. Call the corresponding scenario-based data service interface to obtain flight ticket data, and use the rendering template bound to the fine-grained intent category to generate front-end interactive card data; S6. Send the front-end interactive card data to the client for display, and guide the user to complete the booking process through card interaction events.

[0066] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A flight booking system based on a large language model, characterized in that, include: The data acquisition module is used to collect the user's current natural language query, historical dialogue records, real-time location information, historical behavior data and user profile tags after obtaining user authorization; The query parameter structuring module is used to perform spatiotemporal semantic parsing on the data collected by the data acquisition module and output structured query parameters containing departure location identifier, destination identifier and departure time information. The intent recognition module is used to perform intent classification reasoning based on the user's current natural language query, historical dialogue records and structured query parameters, through a pre-trained large language model, and output fine-grained intent categories including at least one of flight query, discounted flight, multi-day flight, price prediction and price reduction reminder. The contextualized service routing module is used to route structured query parameters to the corresponding contextualized data service interface based on fine-grained intent categories. The data service and card rendering module has multiple data service interfaces and rendering engines corresponding to different intent categories. It is used to call the interface to obtain flight ticket data based on the routing results and generate front-end interactive card data using the corresponding rendering template. The interactive booking guidance module is used to send front-end interactive card data to the client for display and guide users to complete the booking process through card interaction events.

2. The flight booking system based on a large language model according to claim 1, characterized in that, The query parameter structuring module is specifically used for: The origin identifier is determined based on the collected data. The priority of determining the origin identifier is as follows: directly extract it from the user's current natural language query, or extract commonly used origins from historical behavior data, or use the city corresponding to the real-time location information by default. Process the natural language destination descriptions in the collected data; If a standard city code can be matched, it will be used directly as the destination identifier; otherwise, the natural language destination description will be matched with a predefined city cluster label library, and the set of city codes associated with the successfully matched labels will be used as the destination identifier. The natural language departure time description in the collected data is parsed, and based on the preset date calculation rules and holiday rules, the natural language departure time description is converted into a specific departure date or date range as departure time information.

3. The airline ticket booking system based on a large language model according to claim 2, characterized in that, The city cluster tag library is a multi-dimensional, dynamically updated knowledge base used to map fuzzy geographical preference semantics to a specific set of city codes.

4. The flight booking system based on a large language model according to claim 1, characterized in that, In the intent recognition module, the intent classification reasoning process includes: The user's current natural language query, contextual information from historical dialogue records, and structured query parameters are combined into a formatted input according to a preset prompt template. The formatted input is fed into a pre-trained large language model, which identifies the user's intent based on the input content and outputs a fine-grained intent category.

5. The airline ticket booking system based on a large language model according to claim 4, characterized in that, The pre-trained large language model is configured to distinguish and output different fine-grained intent categories based on the destination identifier type, departure time information type, and semantic features in the user's current natural language query in the structured query parameters.

6. The airline ticket booking system based on a large language model according to claim 1, characterized in that, The scenario-based service routing module is specifically used for: Receive fine-grained intent categories from the intent recognition module and structured query parameters from the query parameter structuring module; Based on the preset intent-interface mapping relationship, fine-grained intent categories are mapped to the corresponding contextualized data service interfaces; The structured query parameters are distributed to the contextualized data service interface that is mapped to trigger subsequent data queries; The intent-interface mapping relationship includes at least the following: mapping the intent to purchase discounted air tickets to the discounted air ticket query interface, mapping the intent to query air tickets to the real-time flight query interface, mapping the intent to purchase air tickets for multiple days to the multi-day price trend query interface, mapping the intent to predict prices to the price prediction analysis interface, and mapping the intent to remind users of price reductions to the price reduction monitoring and subscription interface.

7. The airline ticket booking system based on a large language model according to claim 1, characterized in that, The data service and card-based rendering module is specifically used for: Based on the routing results provided by the scenario-based service routing module, call the specific data service interface bound to the fine-grained intent category to obtain a set of flight ticket data results that conform to the structured query parameters; The flight ticket data result set is input into the rendering template corresponding to the fine-grained intent category. The rendering engine performs structured transformation and style rendering to generate the front-end interactive card data containing core flight information and booking guidance controls.

8. The airline ticket booking system based on a large language model according to claim 1, characterized in that, The interactive booking guidance module is specifically used for: The front-end interactive card data is sent to the client, which then renders and displays the corresponding visual interactive card. In response to user interaction with the booking guidance controls within the card, a deep booking link containing structured query parameters is triggered; The system's booking interface is invoked. Based on the structured query parameters and user identity information, the system generates and redirects the user to the corresponding flight class selection and order form filling page to guide the user through the subsequent booking process.

9. The airline ticket booking system based on a large language model according to claim 1, characterized in that, The system also includes a price reduction monitoring service module, used to respond to price reduction alerts in fine-grained intent categories: Create a monitoring task for structured query parameters; The system initiates a background process to periodically compare prices and sends a notification message through the user's pre-selected notification channel when the preset price reduction conditions are met.

10. A flight booking method based on a large language model, characterized in that, Includes the following steps: S1. Collect the user's current natural language query, historical dialogue records, real-time location information, historical behavior data, and user profile tags; S2. Perform spatiotemporal semantic parsing on the collected data to generate structured query parameters containing departure location identifier, destination identifier, and departure time information; S3. Based on the user's current natural language query, historical dialogue records, and structured query parameters, perform intent classification reasoning through a pre-trained large language model and output fine-grained intent categories. S4. Based on the fine-grained intent category, route the structured query parameters to the corresponding contextualized data service interface; S5. Call the corresponding scenario-based data service interface to obtain flight ticket data, and use the rendering template bound to the fine-grained intent category to generate front-end interactive card data; S6. Send the front-end interactive card data to the client for display, and guide the user to complete the booking process through card interaction events.

Images