Travel slot extraction method and device, storage medium and computer program

Abstract

The invention relates to a travel slot extraction method and device, a storage medium and a computer program. The method comprises the following steps: acquiring initial travel information input by a user; calling the slot extraction model and the slot existence discrimination model in parallel to obtain a slot value extraction result and a slot label detection result; if the slot position labels corresponding to the slot position values included in the slot position value extraction result are all effective slot position labels and each effective slot position label in the slot position label detection result is covered by the slot position label, performing semantic and logic verification on each slot position value in the slot position value extraction result based on the initial travel information; and if the slot position value verification result indicates that all slot position values in the slot position value extraction result are not abnormal, generating target travel information based on the slot position value extraction result. According to the scheme provided by the invention, the efficiency and accuracy of business travel slot position extraction can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to methods, apparatus, storage media and computer programs for extracting travel cabin space. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent dialogue systems have been widely applied across various industries, especially in the business travel service sector. Business travel agents, acting as an interactive bridge between users and business travel service platforms, play a crucial role in understanding user intent, extracting key information, and completing operations such as travel applications, ride-hailing requests, bookings, inquiries, and rescheduling. Among these, slot extraction is one of the core technologies in the natural language understanding module. Its goal is to identify and extract semantic units related to specific tasks from the natural language statements input by users, providing structured input for subsequent dialogue management and execution. In business travel scenarios, user expressions are highly diverse, colloquial, contain ellipses, and involve multiple intertwined intents. For example, a user might input "Book me flights from Beijing to Shanghai tomorrow, economy class, for two people," which includes multiple slots such as time, departure point, destination, cabin class, and number of passengers.

[0003] In related technologies, generative large language models are directly applied to slot extraction tasks, typically using hints or few-shot learning methods. However, generative large language models still suffer from slow inference speed, high computational resource consumption, and insufficient output accuracy in industrial applications such as travel agents, failing to provide a smooth user experience in travel slot extraction tasks. Summary of the Invention

[0004] To address or partially address the problems existing in related technologies, this application provides a method, apparatus, storage medium, and computer program for extracting travel cabin space, which can improve the efficiency and accuracy of travel cabin space extraction.

[0005] This application provides a method for extracting travel slots, comprising: acquiring initial travel information input by a user; concurrently invoking a slot extraction model and a slot existence discrimination model, extracting slot values ​​from the initial travel information according to a target slot value sequence based on the slot extraction model, and obtaining a slot value extraction result; determining the existence probability of each preset slot tag from the initial travel information based on the slot existence discrimination model, and obtaining a slot tag detection result, wherein the existence probability of valid slot tags included in the slot tag detection result is greater than or equal to a preset probability threshold; if the slot tags corresponding to the slot values ​​included in the slot value extraction result are all valid slot tags, and each valid slot tag in the slot tag detection result is covered by the slot tag, then performing semantic and logical verification on each slot value in the slot value extraction result based on the initial travel information; if the slot value verification result indicates that all slot values ​​in the slot value extraction result are normal, then generating target travel information based on the slot value extraction result.

[0006] A second aspect of this application provides a travel slot extraction device, comprising: an acquisition module for acquiring initial travel information input by a user; a first processing module for concurrently calling a slot extraction model and a slot existence discrimination model, and extracting slot values ​​from the initial travel information according to a target slot value sequence based on the slot extraction model, thereby obtaining a slot value extraction result; and a second processing module for determining the existence probability of each preset slot tag from the initial travel information based on the slot existence discrimination model, thereby obtaining a slot tag detection result, wherein the valid slot tag detection result includes... The probability of the existence of a slot tag is greater than or equal to a preset probability threshold; the verification module is used to perform semantic and logical verification on each slot value in the slot value extraction result based on the initial travel information if all the slot tags corresponding to the slot values ​​included in the slot value extraction result are valid slot tags, and each valid slot tag in the slot tag detection result is covered by the slot tag; the generation module is used to generate target travel information based on the slot value extraction result if the slot value verification result indicates that all the slot values ​​in the slot value extraction result are normal.

[0007] A third aspect of this application provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method described above.

[0008] A fourth aspect of this application provides a computer program product comprising computer instructions that, when executed by a processor, implement the method described above.

[0009] The technical solution provided in this application may include the following beneficial results: This application obtains initial travel information input by the user, and then calls the slot extraction model and the slot existence discrimination model in parallel. On the one hand, based on the slot extraction model, the initial travel information is extracted according to the target slot value sequence to obtain the slot value extraction result. On the other hand, based on the slot existence discrimination model, the existence probability of each preset slot label is determined from the initial travel information to obtain the slot label detection result. The existence probability of the valid slot labels included in the slot label detection result is greater than or equal to the preset probability threshold. Further, if the slot labels corresponding to the slot values ​​included in the slot value extraction result are all valid slot labels, and each valid slot label in the slot label detection result is covered by slot labels, then the semantic and logical verification of each slot value in the slot value extraction result is performed based on the initial travel information. Finally, if the slot value verification result indicates that all slot values ​​in the slot value extraction result are normal, the target travel information is generated based on the slot value extraction result. The processing method based on parallel call slot extraction model and slot existence discrimination model can improve the efficiency of travel slot extraction, and the consistency discrimination based on slot value extraction result and slot label detection result can improve the accuracy of slot extraction.

[0010] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0011] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0012] Figure 1 This is a schematic flowchart illustrating the travel slot extraction method according to an embodiment of this application; Figure 2 This is another schematic flowchart illustrating the travel slot extraction method shown in the embodiments of this application; Figure 3 This is another schematic flowchart illustrating the travel slot extraction method shown in the embodiments of this application; Figure 4 This is another schematic flowchart illustrating the travel slot extraction method shown in the embodiments of this application; Figure 5 This is a schematic flowchart of the slot existence determination model shown in the embodiments of this application; Figure 6 This is another schematic flowchart illustrating the travel slot extraction method shown in the embodiments of this application; Figure 7 This is a schematic diagram of the travel cabin extraction device shown in the embodiments of this application. Detailed Implementation

[0013] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0014] With the rapid development of artificial intelligence technology, intelligent dialogue systems have been widely applied across various industries, especially in the business travel service sector. Business travel agents, acting as an interactive bridge between users and business travel service platforms, play a crucial role in understanding user intent, extracting key information, and completing operations such as travel applications, ride-hailing requests, bookings, inquiries, and rescheduling. Among these, slot extraction is one of the core technologies in the natural language understanding module. Its goal is to identify and extract semantic units related to specific tasks from the natural language statements input by users, providing structured input for subsequent dialogue management and execution. In business travel scenarios, user expressions are highly diverse, colloquial, contain ellipses, and involve multiple intertwined intents. For example, a user might input "Book me flights from Beijing to Shanghai tomorrow, economy class, for two people," which includes multiple slots such as time, departure point, destination, cabin class, and number of passengers.

[0015] In related technologies, generative large language models are directly applied to slot extraction tasks, typically using hints or few-shot learning methods. However, generative large language models still suffer from slow inference speed, high computational resource consumption, and insufficient output accuracy in industrial applications such as travel agents, failing to provide a smooth user experience in travel slot extraction tasks. To address these issues, this application provides a travel slot extraction method that improves the efficiency and accuracy of travel slot extraction.

[0016] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic flowchart illustrating the travel slot extraction method according to an embodiment of this application. See also... Figure 1 In a specific embodiment, this travel space extraction method can be applied to electronic devices, such as smartphones, tablets, and laptops. The specific process of this travel space extraction method is as follows: S101, Obtain the initial travel information input by the user.

[0018] In this embodiment, within the travel service domain, the travel agent acts as an interactive bridge between the user and the travel service platform, undertaking the crucial tasks of understanding user intent, extracting key information, and completing operations such as travel applications, ride-hailing applications, bookings, inquiries, and rescheduling. Slot extraction is one of the core technologies in the Natural Language Understanding (NLU) module. Its goal is to identify and extract semantic units related to specific tasks from the user's input natural language statements, providing structured input for subsequent dialogue management and execution. In travel scenarios, user expressions are highly diverse, colloquial, contain ellipses, and involve multiple intertwined intents. For example, a user might input "Book me flights from Beijing to Shanghai tomorrow, economy class, for two people," which includes multiple slots such as time, departure point, destination, cabin class, and number of passengers.

[0019] An intelligent agent is an entity or program that can autonomously perceive, make decisions, and execute actions in a specific environment to achieve its goals. Its core characteristics include autonomy, environmental interactivity, and goal orientation. For example, in a business travel intelligent agent scenario, the user only needs to input a sentence, and the intelligent agent can automatically generate business trip applications, official business applications, hotel reservations, etc., which includes automated processes such as intent recognition, slot extraction, form generation, and database querying.

[0020] Specifically, obtaining the initial travel information input by the user is the interaction entry point through the travel intelligence agent system in the travel slot extraction process, collecting all original travel-related information submitted by the user in natural language. This step is not limited to a single input channel and can cover various interaction forms such as application form input, text input in chat window dialogues, and voice-to-text input, thereby completely capturing the user's original expression of travel needs and providing unprocessed basic data for subsequent slot extraction, verification, and correction.

[0021] The initial travel information input by users is essentially unstructured natural language text. Its core characteristics are the diversity, colloquialism, flexibility, and implicit information inherent in travel scenarios. It covers key semantic units required throughout the entire travel process, including but not limited to information corresponding to preset slots such as departure point, destination, departure time, arrival time, mode of transportation, cabin class, number of travelers, reason for travel, and special requirements (such as vehicle preference and seating requirements). It may also contain redundant expressions or corrective statements unrelated to travel. Acquiring the initial travel information input by users serves as the data source for the slot extraction process. It is crucial to ensure the integrity and originality of the information collected, avoiding pre-processing semantic filtering or structural optimization of user expressions. This ensures that subsequent models can perform accurate analysis based on complete original information, preventing slot extraction bias due to information loss.

[0022] In some embodiments, the input format of initial travel information exhibits significant diversity, taking into account users' actual expression habits in business travel scenarios. Examples include standard complete expressions, colloquial abbreviations, mixed multi-trip expressions, negation and correction expressions, ambiguous and special format expressions, and expressions containing noise and typos.

[0023] Optionally, the standard complete expression type refers to a user clearly and completely stating their core travel needs, with no omissions or ambiguities. For example, "Please book economy class tickets for two people departing from Shanghai Hongqiao Airport to Beijing Daxing Airport on October 20, 2025. The reason for the business trip is to attend an industry summit, and the return date is the afternoon of October 23." or "Initiate a business trip application: Fly from Guangzhou Baiyun Airport to Chengdu Shuangliu Airport on November 5, and return by high-speed rail on November 8. The reason for the business trip is to visit a client. There are three people traveling with me, and we need to book business class seats." The standard complete expression type input includes complete slot information such as departure point, destination, time, mode of transportation, number of people, and reason for business trip. The expression is standardized and is the ideal form of initial business travel information.

[0024] Optionally, colloquial abbreviations refer to expressions used by users in everyday conversation, characterized by omissions or flexible word order. For example, "Flying to Shenzhen for a meeting tomorrow, returning the day after tomorrow" omits the departure point and specific mode of transportation, requiring further inference based on context or user history data. Similarly, "Help me hail a taxi to a long-distance meeting, there are many people," without directly mentioning the vehicle type, implying a need for a "business vehicle," thus representing implicit slot information. Another example is "Going to Hangzhou next Wednesday, returning next Friday for project coordination," which omits the departure point and mode of transportation, requiring the model to identify the core slot and reserve space for supplementation. Colloquial abbreviation input aligns with everyday communication habits, but the information is not explicit enough, placing higher demands on the semantic understanding capabilities of subsequent slot extraction.

[0025] Optionally, multi-trip mixed expression type refers to a user's sentence containing multiple consecutive trips, with slot information presented in an overlapping manner. For example, "First, I book the high-speed train from Harbin to Beijing tomorrow morning, then I book a flight from Beijing to Guangzhou tonight. The reason for the business trip is project inspection, and I'm traveling alone the whole time." or "Next Monday, I'll fly to Xi'an, then from Xi'an to Lanzhou by train on Wednesday, and return to Shanghai from Lanzhou on Friday. All of these are business trips." For multi-trip mixed expression type inputs, the model needs to distinguish the slot correspondence between different trips to avoid slot confusion.

[0026] Optionally, the negation and correction expression category refers to user statements containing self-correcting or negative statements, requiring the model to capture the final, true need. For example, "Book me a ticket to Shanghai—oh wait, I mean Hangzhou, departing next Tuesday," or "I want to book a flight from Wuhan to Qingdao on October 15th, not economy class, but business class, for training." The input for negation and correction expressions may contain temporarily changed slot information, and the initial expression may contain invalid information. The model needs to filter out erroneous information and extract the corrected, valid slots.

[0027] Optionally, fuzzy and special format expressions refer to users using non-standard time and location descriptions or special symbols. For example, "Going to Beijing on the first day after National Day, returning the following Wednesday, business trip for bidding," where "first day after National Day" and "Beijing" are fuzzy time and city aliases; another example, "Flying to Xiamen Port the following Friday to attend an industry exhibition," where "Xiamen Port" needs to be mapped to the specific city "Xiamen," and "the following Friday" is a non-standard time expression. Inputs for fuzzy and special format expressions require the model to have the ability to parse fuzzy semantics, aliases, and special formats to ensure accurate extraction of slot information.

[0028] Optionally, the "noisy and misspelled expression" category refers to user input containing irrelevant information, misspellings, or redundant expressions. For example, "By the way, are there any economy class seats available for the Shanghai-Beijing flight tomorrow? I need them for a business trip," contains the misspellings "Shanghai," "Beijing," and "economy class," as well as the redundant information "By the way." Another example is "Help me initiate a business trip application. Um, I'm going from Chengdu to Chongqing next Monday for technical support, about 3 days," which contains the interjection "um," the misspellings "Chengdu," and "Chongqing." Inputs containing noisy and misspelled expressions are subject to noise interference, requiring the model to have anti-interference capabilities, filtering out invalid information while correcting erroneous expressions and accurately extracting core slots.

[0029] Understandably, ensuring the integrity and originality of the initial travel information input by users provides a data foundation for high-quality slot extraction; adapts to diverse expression scenarios and improves system compatibility; provides a basis for subsequent verification and correction; reduces data preprocessing costs and improves process efficiency; and ensures the comprehensiveness and robustness of slot extraction.

[0030] S102, the slot extraction model and the slot existence discrimination model are called in parallel. Based on the slot extraction model, the initial travel information is extracted according to the target slot value sequence to obtain the slot value extraction result.

[0031] After obtaining the user's initial travel information, the parallel computing design of the system architecture simultaneously initiates the reasoning process of the slot extraction model and the slot existence discrimination model. The parallel scheduling mechanism can break the traditional serial processing mode, allowing the two models to read the initial travel information and perform independent calculations at the same time, avoiding efficiency loss caused by a single model waiting.

[0032] The slot extraction model is used to accurately extract key semantic units related to travel from unstructured initial travel information, according to a preset target slot value sequence format, and finally output structured slot value extraction results. The extraction logic driven by the target slot value sequence abandons redundant output formats and clearly defines the fixed order of preset slots, such as "departure location, destination, departure time, end time, mode of transportation, reason for business trip". The model only outputs an ordered sequence of pure slot values ​​and does not contain any descriptive redundant data.

[0033] In some embodiments, combining different types of initial travel information in a business travel scenario with the model's extraction logic, the following is a typical example of the slot extraction model extracting information according to the target slot value sequence. The target slot sequence can be "departure location, destination, departure time, end time, mode of transportation, reason for business trip".

[0034] Optionally, let's take the extraction of standard complete expression input as an example. For instance, the initial travel information is "Book me economy class tickets from Shanghai Hongqiao to Beijing Daxing on October 20, 2025, for two people. The reason for the trip is to attend an industry summit, and the return trip is on the afternoon of October 23." The slot extraction process involves the model identifying explicit slot information in the input, sorting it according to a preset sequence, filtering redundant expressions, and outputting only core slot values, such as "two people" and "economy class tickets." If these are not preset in the target sequence, they are not extracted, or they are extracted according to an extended sequence. Thus, the slot value extraction result according to the target slot sequence is "Shanghai Hongqiao, Beijing Daxing, October 20, 2025, October 23, 2025 afternoon, airplane, attending an industry summit." Optionally, let's take the extraction of colloquial abbreviations as an example. For instance, the initial travel information is "Flying to Shenzhen for a meeting tomorrow, returning the day after tomorrow." The slot extraction model, based on its semantic understanding capabilities fine-tuned in the travel domain, completes the omitted information, identifies "fly" as the mode of transportation "airplane," "meeting" as the reason for the trip, and "tomorrow" and "the day after tomorrow" as the departure and arrival times, respectively. Thus, the slot value extraction result according to the target slot sequence is "Beijing, Shenzhen, tomorrow, the day after tomorrow, airplane, meeting."

[0035] Optionally, let's take the extraction of multi-trip mixed-expression input as an example. For instance, the initial travel information is "First book the high-speed train from Harbin to Beijing tomorrow morning, then book the flight from Beijing to Guangzhou tonight; the reason for the business trip is project inspection." The slot extraction process involves the model identifying two consecutive trips in a single sentence, splitting the slot information according to the trip order, corresponding to two sets of target sequences, and using a separator to distinguish between multiple trips, ensuring a one-to-one correspondence between slots and trips. Thus, the slot value extraction result according to the target slot sequence is "Harbin, Beijing, tomorrow morning, none, high-speed train, project inspection; Beijing, Guangzhou, evening, none, flight, project inspection," where no specific end time is marked as "none."

[0036] Optionally, let's take the extraction of negative and modified expressions as an example. For instance, the initial travel information is "Book me a ticket to Shanghai, oh wait, I mean Hangzhou, departing next Tuesday and returning next Thursday by train." The slot extraction model captures the modification intention of "oh wait, I mean Hangzhou," filters out the invalid slot "Shanghai," extracts the modified destination "Hangzhou," and simultaneously identifies the departure time "next Tuesday," the return time "next Thursday," and the mode of transportation "train." Thus, the slot value extraction result according to the target slot sequence is "default departure location, Hangzhou, next Tuesday, next Thursday, train, none."

[0037] Optionally, the extraction of inputs with fuzzy or special formats can be illustrated using an example. For instance, the initial travel information might be "Going to Beijing on the first day after National Day, returning the following Wednesday, reason for the trip is bidding." The slot extraction model, through fine-tuning its learned alias mapping capabilities, transforms "Beijing" into "capital," parses "the first day after National Day" into a specific date (e.g., October 8, 2025), and converts "the following Wednesday" into a standard time expression, outputting according to the target sequence. Thus, the slot value extraction result according to the target slot sequence would be "Default departure point, Beijing, October 8, 2025, the following Wednesday, none, bidding."

[0038] Optionally, let's take the extraction of input containing noise and typos as an example. For instance, the initial travel information is "By the way, are there any economy class seats left for tomorrow's Shanghai-Beijing flight? For business trips." The slot extraction model has the ability to resist disturbances, filtering out redundant information such as "By the way," correcting typos such as "Shanghai" to "Shanghai," "Beijing" to "Beijing," and "economy class" to "economy cabin," identifying the mode of transportation as "airplane," and simplifying the reason for the trip from "for business trips" to "business trip." Thus, the slot value extraction result according to the target slot sequence is "Shanghai, Beijing, tomorrow, none, airplane, business trip."

[0039] Understandably, the parallel invocation mechanism allows slot extraction and existence determination to be performed simultaneously, avoiding the waiting time in serial processing. Combined with the target slot value sequence format, it reduces redundant data, significantly improving the model inference speed, thereby significantly improving extraction efficiency and reducing response latency.

[0040] The target slot value sequence format allows the extracted results to be directly presented as an ordered set of pure slot values, eliminating the need for additional parsing of redundant information and facilitating rapid reading and comparison by subsequent semantic and logical verification modules. Simultaneously, the structured format supports standardized annotation for special cases such as multi-runtime and missing values, providing direct support for automated system processing, achieving structured data output, and reducing subsequent processing costs.

[0041] S103, Based on the slot existence discrimination model, determine the existence probability of each preset slot tag from the initial travel information to obtain the slot tag detection result, wherein the existence probability of the valid slot tags included in the slot tag detection result is greater than or equal to the preset probability threshold.

[0042] Based on the slot existence discrimination model, the probability of existence of each preset slot tag is determined from the initial travel information. The slot tag detection result is a key step in the core verification process of travel slot extraction. It is executed in parallel with the slot extraction model. The core goal is to provide objective and quantitative evidence of the existence of slot extraction results and filter out erroneous extractions.

[0043] It should be noted that the slot existence discrimination model is a lightweight model optimized based on the Transformer Encoder architecture. It adds three fully connected layers and two layer normalization modules to the original architecture. Through a step-by-step process of feature mapping, enhancement, and probability output, it accurately captures the semantic existence features of each slot. The slot existence discrimination model takes semantic expression as input, and after multi-layer processing, outputs a dimension vector matching the preset number of slot labels. Then, a function maps the vector to an existence probability within the (0,1) interval, realizing a quantitative judgment on the existence of each slot. In addition, a preset probability threshold is used as the judgment standard. Slot labels with an existence probability greater than or equal to the preset probability threshold are valid slot labels, indicating that the slot does exist in the initial travel information; those with an existence probability less than the preset probability threshold are invalid slot labels, and the corresponding slot has no semantic support in the original input.

[0044] The preset slot labels are determined based on the core needs of business travel scenarios, including key dimensions such as departure point, destination, departure time, end time, mode of transportation, reason for business trip, number of travelers, and special requirements, ensuring coverage of core information needs throughout the entire business travel process. This step, executed in parallel with the slot extraction model, does not increase the overall process time and provides crucial evidence for subsequent consistency verification.

[0045] In some embodiments, the initial travel information of different types in the travel scenario is combined with the probability calculation logic of the slot existence discrimination model. The preset slot label may include departure point, destination, departure time, end time, mode of transportation, and reason for business trip. The preset probability threshold may be 0.5.

[0046] Optionally, taking the detection of standard complete expression class input as an example, for instance, the initial travel information is "Help me book economy class tickets from Shanghai Hongqiao to Beijing Daxing on October 20, 2025, for two people, the reason for the business trip is to attend an industry summit, and the return time is the afternoon of October 23". During the detection process, the slot existence discrimination model analyzes the explicit semantic features in the input. Each preset slot has a clear corresponding expression. Through feature extraction and probability calculation, the existence probability of each slot is output, thus obtaining the slot label detection results as departure place (0.98), destination (0.99), departure time (0.97), end time (0.96), mode of transportation (0.95), and reason for business trip (0.94). The existence probability of all slots is greater than or equal to 0.5, and all are judged as valid slot labels.

[0047] Optionally, let's take the detection of colloquial abbreviations as an example. For instance, the initial travel information is "Flying to Shenzhen for a meeting tomorrow, returning the day after tomorrow." During the detection process, the slot existence discrimination model identifies "tomorrow" as the departure time, "Shenzhen" as the destination, "fly" as the mode of transportation, "meeting" as the reason for the business trip, and "the day after tomorrow" as the end time. However, the input does not mention the departure location, and there is no information related to the departure location in the semantic features. Thus, the slot label detection results are: departure location (0.12), destination (0.97), departure time (0.98), end time (0.95), mode of transportation (0.96), and reason for the business trip (0.93). Only the probability of the departure location is less than 0.5, which is judged as an invalid slot label, and the rest are valid slot labels.

[0048] Optionally, taking the detection of error extraction scenarios as an example, for instance, the initial travel information is "Help me initiate a business trip application to Beijing now". During the detection process, the user input "now" is a time expression and has no semantic meaning related to the departure location. The slot existence discrimination model identifies the semantic missingness corresponding to the departure location through semantic feature analysis, and at the same time accurately captures the valid semantics of the destination, departure time, and reason for business trip. Thus, the slot label detection results are: departure location (0.08), destination (0.99), departure time (0.96), end time (0.21), mode of transportation (0.35), and reason for business trip (0.92). The probability of the existence of departure location, end time, and mode of transportation is less than 0.5, so they are judged as invalid slot labels; the destination, departure time, and reason for business trip are valid slot labels.

[0049] In travel intelligence systems, while slot extraction models can accurately identify key information such as "departure location," "date," and "reason for travel," they may still generate "ghost slots" (i.e., the model incorrectly predicts slots that do not actually exist) when faced with colloquial expressions, ellipses, noisy inputs, or sparsely labeled samples. For example, if a user inputs "Help me initiate a business trip application from now to Beijing," the model outputs: {"departure location": "now", "destination": "Beijing"} → "departure location = now" is an incorrect extraction. If such errors are not intercepted, they will directly lead to the failure of subsequent application, ticket booking, and rescheduling business logic, seriously affecting user experience and system reliability.

[0050] Optionally, taking the detection of multi-trip mixed expression input as an example, for instance, the initial travel information is "First book the high-speed train from Harbin to Beijing tomorrow morning, then book the flight from Beijing to Guangzhou tonight, the reason for the business trip is project inspection." During the detection process, the slot existence discrimination model identifies the two consecutive trips in a single sentence, captures the slot semantic features corresponding to each trip, and outputs the slot existence probability according to the trip dimension, ensuring that the slot existence judgment is not confused in multi-trip scenarios. Thus, the slot label detection results are: departure location (0.99), destination (0.98), departure time (0.97), end time (0.15), mode of transportation (0.96), reason for business trip (0.95); (second trip): departure location (0.98), destination (0.99), departure time (0.96), end time (0.12), mode of transportation (0.97), reason for business trip (0.95); only the end time of the two trips has an existence probability of less than 0.5, and is judged as invalid slot labels, while the rest are valid slot labels.

[0051] Optionally, taking the detection of fuzzy and specially formatted inputs as an example, for instance, the initial travel information is "Going to Beijing on the first day after National Day, returning the following Wednesday, the reason for the business trip is bidding." During the detection process, the slot existence discrimination model fine-tunes its learned semantic adaptation ability, mapping "Beijing" to "capital", parsing "the first day after National Day" as departure time semantics, and parsing "the following Wednesday" as end time semantics, accurately capturing the implicit semantic features of each slot, thus obtaining the slot label detection results as departure location (0.23), destination (0.97), departure time (0.95), end time (0.94), mode of transportation (0.31), and reason for business trip (0.96); the probability of the departure location and mode of transportation being less than 0.5 is judged as invalid slot labels, and the rest are valid slot labels.

[0052] Optionally, taking the detection of inputs containing noise and misspellings as an example, for instance, the initial travel information is "By the way, are there any economy class seats left for the Shanghai-Beijing flight tomorrow? For a business trip." During the detection process, the slot existence discrimination model has the ability to resist disturbances, filtering out the redundant information "By the way," correcting the semantic features corresponding to the misspellings "Shanghai," "Beijing," and "economy class," and accurately identifying the valid semantics of departure point, destination, departure time, mode of transportation, and reason for business trip. Thus, the slot label detection results are: departure point (0.96), destination (0.95), departure time (0.98), end time (0.27), mode of transportation (0.94), and reason for business trip (0.93). Only the probability of the end time is less than 0.5, which is judged as an invalid slot label, and the rest are valid slot labels.

[0053] Understandably, the slot existence discrimination model, through deep analysis of semantic features, can effectively identify slots that the slot extraction model may incorrectly extract. For example, for incorrect extractions like "departure location = present," the slot existence discrimination model quickly determines it as invalid due to its low existence probability, preventing subsequent business logic failures and improving extraction accuracy. Furthermore, the automated and intelligent judgment of slot existence provides verification of slot extraction results without human intervention. Compared to traditional methods relying on manual review, this significantly reduces labor costs while avoiding subjective errors in human judgment, giving the travel intelligence system stronger autonomous processing capabilities and making it more suitable for large-scale industrial applications.

[0054] S104. If all slot labels corresponding to the slot values ​​included in the slot value extraction result are valid slot labels, and each valid slot label in the slot label detection result is covered by a slot label, then semantic and logical verification is performed on each slot value in the slot value extraction result based on the initial travel information.

[0055] The verification process of the slot value extraction results and the slot label detection results is a prerequisite for judgment. It must meet two core requirements at the same time: first, all slot labels in the slot extraction results are valid labels; second, all valid labels output by the slot existence discrimination model are covered by the extraction results, ensuring that there are no erroneous slots and no missing key information in the extraction results.

[0056] By verifying core dimensions and based on the original semantics of initial travel information, each slot value is reviewed individually from both semantic and logical perspectives. Semantic verification focuses on the rationality of the slot value itself, such as whether it is a valid city and standard time. Logical verification focuses on the rationality of the relationship between slot values, such as whether the departure time is earlier than the end time. In addition, by verifying the execution logic, the semantic validity of an individual slot value is first judged, and then the consistency of slot groups with logical relationships is checked to ensure that the extracted results not only conform to the semantic rules of an individual slot, but also meet the business logic of the travel scenario.

[0057] In some embodiments, the following is an example of semantic and logical verification, which combines different types of extraction results in the business travel scenario. The preset slot labels are departure point, destination, departure time, end time, mode of transportation, and reason for business trip; the preset probability threshold is 0.5.

[0058] Optionally, the verification of the complete extraction results of the standard can be used as an example. For example, the slot value extraction result is "Shanghai Hongqiao, Beijing Daxing, October 20, 2025, October 23, 2025 afternoon, airplane, attending industry summit"; the slot label detection result is that the probability of all slot labels is greater than or equal to 0.95, all of which are valid labels and there are no omissions. Semantic validation verifies the rationality of each slot value one by one. "Shanghai Hongqiao" and "Beijing Daxing" are valid airports and cities. "October 20, 2025" and "October 23, 2025 (afternoon)" are standard time formats. "Airplane" is a valid mode of transportation. "Attending an industry summit" is a reasonable reason for business trip. Logical validation verifies the consistency of related slots. The departure time (October 20, 2025) is earlier than the end time (October 23, 2025 (afternoon)). The departure and destination are different cities. The mode of transportation "airplane" matches the airport name (Shanghai Hongqiao, Beijing Daxing). There is no logical conflict. Therefore, the validation result is that all slot values ​​have no semantic and logical abnormalities and pass the validation.

[0059] Optionally, an example of verifying the extraction result with semantic errors is given. For example, the slot value extraction result is "Shang Hai, Bei Jing, December 20, 2025, October 15, 2025, airplane, attend industry summit"; the slot label detection result is that the existence probability of all slot labels is greater than or equal to 0.9, all are valid labels, and there is no omission. In semantic verification, "Shang Hai" is a misspelling of "Shanghai" (semantically invalid), "Bei Jing" is a misspelling of "Beijing" (semantically invalid), and "December 20, 2025" is an invalid time (December does not conform to the calendar rules, semantically invalid); in logical verification, the departure time (December 20, 2025) is an invalid time, and the end time (October 15, 2025) is earlier than the departure time (assuming the corrected departure time is October 20, 2025), there is a logical conflict, so the verification result is that there are semantic errors (misspellings, invalid times) and logical conflicts (time order reversal), and the verification fails.

[0060] Optionally, an example of verifying the extraction result with implicit semantics is given. For example, the slot value extraction result is "Guangzhou, long-term meeting location, tomorrow, the day after tomorrow, business vehicle, have a long-term meeting"; the slot label detection result is that the existence probability of all slot labels is greater than or equal to 0.85, all are valid labels, and there is no omission. In semantic verification, "Guangzhou" is a valid departure place, "tomorrow" and "the day after tomorrow" are standard time expressions, "business vehicle" is a valid means of transportation, and "have a long-term meeting" is a reasonable reason for a business trip; the "long-term meeting location" needs to be confirmed in combination with the initial travel information and belongs to the destination that is not clear but can be supplemented by the context; in logical verification, the departure time "tomorrow" is earlier than the end time "the day after tomorrow", and the means of transportation "business vehicle" matches the implicit requirement of "a large number of people", there is no logical conflict, so the verification result is that there are no abnormalities in the semantics and logic of all slot values, and the verification passes.

[0061] Optionally, the verification of the results of multi-trip mixed extraction can be used as an example. For example, the slot value extraction result is "Harbin, Beijing, tomorrow morning, none, high-speed rail, project inspection; Beijing, Guangzhou, evening, none, airplane, project inspection"; the slot label detection result is that the probability of the existence of all slot labels in both trips is greater than or equal to 0.9, all of which are valid labels and there are no omissions. In the semantic validation, the first trip's "Harbin" and "Beijing" are valid cities, "tomorrow morning" is a reasonable time, and "high-speed rail" is a valid mode of transportation. In the second trip's "Beijing" and "Guangzhou" are valid cities, "evening" is a reasonable time, "airplane" is a valid mode of transportation, and "project inspection" is a reasonable reason for business trip. The absence of an end time fits the scenario of multiple trips without a specified return time (semantically valid). In the logical validation, the departure point (Harbin) and destination (Beijing) of the first trip are different, and the departure point (Beijing) and destination (Guangzhou) of the second trip are different, and the departure point (Beijing) of the second trip is the destination of the first trip, which is logically consistent. The modes of transportation "high-speed rail" and "airplane" match the trips between cities respectively, without logical conflicts. Therefore, the validation result is that all slot values ​​have no semantic or logical abnormalities and pass the validation.

[0062] Optionally, let's take the validation of extraction results containing logical conflicts as an example. For instance, the slot value extraction result is "Beijing, Beijing, October 20, 2025, October 18, 2025, train, client visit"; the slot label detection result shows that the probability of all slot labels is greater than or equal to 0.88, all of which are valid labels, and none are missing. In semantic validation, "Beijing" is a valid city, "October 20, 2025" and "October 18, 2025" are standard times, "train" is a valid mode of transportation, and "client visit" is a reasonable reason for business trip, all of which are semantically valid. In logical validation, both the departure point and destination are "Beijing," the same city, which does not conform to the logic of a business trip. The departure time (October 20, 2025) is later than the end time (October 18, 2025), the time order is reversed, and there is a logical conflict. Therefore, the validation result is that there is a logical conflict, and the validation fails.

[0063] Understandably, semantic and logical validation can identify implicit issues such as valid tags with incorrect values, including typos, invalid timestamps, and logical conflicts. These issues cannot be detected by the probability detection of the slot existence discrimination model, further improving the accuracy of slot extraction and filtering out implicit errors. The validation process fully integrates the core business rules of the travel scenario, ensuring that the extraction results meet the actual business execution requirements. This avoids subsequent business failures such as ticket booking and business trip applications due to logical errors, improving system reliability and user experience. In addition, semantic validation adapts to ambiguous expressions, implicit information, and typos, while logical validation supports multiple itineraries and special scenarios, enabling the system to handle complex inputs such as colloquialisms, abbreviations, and multiple intentions, adapting to complex semantic expressions and enhancing system robustness. The automation of semantic and logical validation allows for in-depth review of the extraction results without human intervention. Compared to traditional methods relying on manual review, this significantly reduces labor costs while avoiding subjective errors in human judgment, giving the travel intelligent agent system stronger autonomous processing capabilities, reducing human intervention costs, and enhancing industrial application value.

[0064] In some optional embodiments, during the verification process of the slot value extraction result and the slot label detection result, the travel slot extraction method may further include the following steps: if the slot labels corresponding to the slot values ​​included in the slot value extraction result include invalid slot labels, and / or the slot labels do not cover all valid slot labels, then the initial travel information is re-extracted based on a preset model using prompt word engineering to obtain a first corrected slot value extraction result; wherein, the first slot labels corresponding to the first slot values ​​included in the first corrected slot value extraction result are all valid slot labels, and each valid slot label in the slot label detection result is covered by the first slot label, and the probability of the existence of invalid slot labels is less than a preset probability threshold; semantic and logical verification is performed on each first slot value in the first corrected slot value extraction result based on the initial travel information.

[0065] Specifically, if there are invalid slot labels or missing valid slot labels in the slot value extraction results, a precise correction mechanism will be activated to ensure the completeness and validity of the extraction results. The core logic of the correction mechanism includes the following dimensions: First, trigger condition determination, clearly defining two core trigger scenarios: invalid slot labels exist in the slot extraction results, or the slot labels in the extraction results do not completely cover all valid labels output by the slot existence discrimination model. Correction is initiated as long as either scenario is met. Second, correction execution subject, selecting a more powerful preset large model, utilizing its strong context understanding and complex semantic parsing capabilities to compensate for the adaptation limitations of the basic slot extraction model. Third, prompt word engineering design, through structured prompt word templates, clearly informing the large model of the slot definition, valid label range, and correction objectives (removing invalid labels and supplementing missing labels) for the travel scenario, ensuring that the correction results meet the preset requirements. Fourth, correction result constraints, the first corrected slot value extraction results must meet dual standards: all first slot labels are valid labels and completely cover all valid labels in the slot detection results, while all invalid labels are removed. Fifth, secondary verification closed loop, after correction, semantic and logical verification must be performed again based on the initial travel information to ensure that the correction results are not only valid and fully covered by labels, but also that the slot values ​​themselves are reasonable and without logical conflicts. While ensuring efficiency, it addresses extraction defects in complex scenarios and maximizes the reliability of results.

[0066] In some embodiments, combining the two core triggering scenarios of invalid tags and missing valid tags in the business travel scenario, the following is an example of the correction steps, wherein the preset slot tags include departure point, destination, departure time, end time, mode of transportation, and reason for business trip; the preset probability threshold is 0.5.

[0067] Optionally, the correction of scenarios containing invalid slot labels can be illustrated as follows: For example, the initial travel information is "Help me initiate a business trip application to Beijing now, departing tomorrow to attend a product launch"; the slot value extraction result is "now, Beijing, tomorrow, none, none, attend product launch"; the slot label detection result is departure location (0.08), destination (0.99), departure time (0.96), end time (0.21), mode of transportation (0.35), and reason for business trip (0.92).

[0068] The condition for triggering the correction process is that the probability of the tag corresponding to "departure location = now" in the extracted results is less than 0.5 (0.08), indicating an invalid slot tag, thus meeting the trigger condition. During the correction process, a structured prompt is constructed as follows: "Based on the following travel information 'Help me initiate a business trip application to Beijing now, departing tomorrow to attend a product launch,' extracting six slots: departure location, destination, departure time, end time, mode of transportation, and reason for business trip. The known valid slot tags are destination (0.99), departure time (0.96), and reason for business trip (0.92), while the invalid slot tags are departure location (0.08), end time (0.21), and mode of transportation (0.35). Please remove invalid slot tags, supplement any potentially missing valid information, and output in the format 'departure location, destination, departure time, end time, mode of transportation, reason for business trip,' marking invalid or unmentioned slots as 'none.'" After the pre-set large model parses the prompt, it identifies that "now" is not the departure location, is semantically invalid, supplements the implicit departure location, and finally outputs the corrected result. The first corrected slot value extraction result is "Shanghai, Beijing, Tomorrow, None, None, Attend product launch". During the second verification process, semantic verification confirms that "Shanghai" and "Beijing" are valid cities, "Tomorrow" is a reasonable time, and "Attend product launch" is a valid reason for business trip. Logical verification shows no conflict and passes the verification.

[0069] Optionally, the correction of missing valid slot labels can be illustrated as follows: For example, the initial travel information is "Booked a high-speed train from Guangzhou to Shenzhen next Wednesday, returning on Friday to discuss cooperation"; the slot value extraction result is "Guangzhou, Shenzhen, next Wednesday, Friday, high-speed train, none"; the slot label detection result is departure location (0.99), destination (0.98), departure time (0.97), end time (0.96), mode of transportation (0.95), and reason for business trip (0.93).

[0070] The condition for triggering the correction process is that the "Reason for Business Trip" is marked as "None" in the extracted results, but the "Reason for Business Trip" is a valid label in the slot detection results, which is considered a missing valid label and meets the trigger condition. During the correction process, a structured prompt is constructed as "Based on the following travel information 'Booked a high-speed train from Guangzhou to Shenzhen next Wednesday, returning on Friday to discuss cooperation,' extracting six slots: departure location, destination, departure time, end time, mode of transportation, and reason for business trip. Known valid slot labels include departure location (0.99), destination (0.98), departure time (0.97), end time (0.96), mode of transportation (0.95), and reason for business trip (0.93). Please supplement the missing valid slot labels, remove invalid labels, and output in the format 'departure location, destination, departure time, end time, mode of transportation, reason for business trip'." After the preset large model parses the prompt words, it identifies the business trip reason corresponding to "discuss cooperation" and adds it to the extraction results, thus obtaining the first corrected slot value extraction result as "Guangzhou, Shenzhen, next Wednesday, Friday, high-speed rail, discuss cooperation"; the second verification process is to confirm that all slot values ​​are valid through semantic verification and that there are no conflicts through logical verification, and the verification is passed.

[0071] Optionally, the correction of scenarios containing both invalid and missing valid labels can be illustrated as follows: For example, the initial travel information is "Book me a flight to Hangzhou next Monday, and visit a client in Nanjing, returning next Thursday, with Shanghai as the departure point"; the slot value extraction result is "Shanghai, Hangzhou, next Monday, next Thursday, airplane, none"; the slot label detection result is departure point (0.99), destination (0.98), departure time (0.97), end time (0.96), mode of transportation (0.95), and reason for business trip (0.94).

[0072] The condition for triggering the correction process is the omission of "reason for business trip" in the extracted results, i.e., the valid labels are not covered, and the single destination is mistakenly split into "Hangzhou", failing to identify "Nanjing" as the second destination, implying the omission of valid labels. There are no invalid labels, but the trigger condition of missing valid labels is met. During the correction process, a structured prompt is constructed as "Based on the following travel information 'Book me a flight, flying to Hangzhou next Monday, and visiting a client in Nanjing, returning next Thursday, departure point is Shanghai'", extracting six slots: departure point, destination, departure time, end time, mode of transportation, and reason for business trip. All slot labels are known to be valid. Please fill in the missing slot information, correct the destination identification error, and output the multi-trip results in the format 'departure point, destination, departure time, end time, mode of transportation, reason for business trip'." After the pre-set large model parses the prompt, it identifies a single sentence containing two trips, adds the second trip destination "Nanjing", and clarifies the reason for business trip as "visiting a client". The first corrected slot value extraction result is "Shanghai, Hangzhou, next Monday, none, flight, visit customer; Shanghai, Nanjing, none, next Thursday, flight, visit customer"; the second verification process is semantic verification to confirm that all slot values ​​are valid, and logical verification to ensure that there are no conflicts in multiple runs, and the verification is passed.

[0073] Optionally, the correction of scenarios where ambiguous expressions lead to the omission of effective labels can be illustrated as follows: For example, the initial travel information is "Go to Chengdu after the National Day holiday, then transfer to Chongqing, take the high-speed rail, and handle project matters"; the slot value extraction result is "None, Chengdu, after the National Day holiday, none, high-speed rail, handle project matters"; the slot label detection result is departure location (0.32), destination (0.99), departure time (0.97), end time (0.28), mode of transportation (0.96), reason for business trip (0.95), and second destination (0.98).

[0074] The condition for triggering the correction process is that the extracted result is missing the valid label "second trip destination = Chongqing" and the departure point is not specified. However, the departure point is an invalid label, which does not affect the correction trigger and satisfies the trigger condition of missing valid labels. During the correction process, the structured prompt is constructed as follows: "Based on the following travel information 'going to Chengdu after National Day, then transferring to Chongqing, taking the high-speed rail to handle project matters', extract six slots (supporting multiple trips): departure point, destination, departure time, end time, mode of transportation, and reason for business trip (supporting multiple trips). Known valid slot labels include destination (0.99), departure time (0.97), mode of transportation (0.96), reason for business trip (0.95), and second trip destination (0.98). Please add the missing second trip destination, output in multi-trip format, and mark invalid slots as 'none'." After parsing the prompt in the preset large model, the second trip destination "Chongqing" is added, and the departure time "after National Day" is specified. Thus, the first corrected slot value extraction result is "None, Chengdu, after National Day, none, high-speed rail, handling project matters; Chengdu, Chongqing, none, none, high-speed rail, handling project matters"; the second verification process is semantic verification to confirm the slot value is valid, and logical verification of multi-run continuity, and the verification is passed.

[0075] Understandably, this correction process can remove invalid tags, prevent business logic failures, effectively intercept erroneous slots such as "departure location = present," and utilize the semantic understanding capabilities of the large model to remove invalid tags without semantic support, thus preventing errors in subsequent business processes such as business trip applications and ticket bookings due to invalid tags. For valid tags not covered by the basic model, the large model can accurately supplement them, ensuring that the extracted results cover all the key information required in business travel scenarios, thereby supplementing missing tags and ensuring the completeness of the extraction.

[0076] Furthermore, the basic model quickly extracts data to meet the needs of most simple scenarios, only calling the large model for correction when extraction defects occur. This avoids the inefficiency caused by using the large model throughout the process, thus enabling dynamic model switching and balancing efficiency and accuracy. The preset large model has stronger capabilities for parsing complex semantics, handling scenarios that the basic model struggles to adapt to, such as multi-running scenarios, ambiguous expressions, and implicit information. This adapts to complex scenarios and enhances the system's robustness.

[0077] S105, if the slot value verification result indicates that there are no abnormalities in all slot values ​​in the slot value extraction result, then generate the target travel information based on the slot value extraction result.

[0078] After verifying tag validity, coverage completeness, and both semantic and logical validation, the structured slot values, validated through multiple rounds, are transformed into standardized target travel information that meets the needs of business travel scenarios. This provides a direct basis for subsequent automated business processes such as business trip applications, ticket bookings, and itinerary planning. Specifically, the generation of target travel information is only triggered when the slot value validation results clearly show no anomalies in any of the slot values, ensuring the absolute reliability of the input data. Based on preset business travel data specifications, the extracted slot value sequence format is mapped to standardized data structures containing fixed fields, such as forms, database records, and business interface parameters. These fields cover core information such as departure location, destination, departure time, end time, mode of transportation, reason for business trip, number of people, and special requirements, while also supplementing derived fields required by the business, such as trip duration and booking channels corresponding to the mode of transportation. Furthermore, the target travel information can dynamically adjust its output format according to the needs of downstream business scenarios, supporting both form generation and database queries within the travel intelligence system, as well as interface calls to third-party booking platforms, achieving full automation of the extraction, generation, and execution process.

[0079] In some embodiments, combining the extraction results of different types of valid slots in a business travel scenario, the following is an example of generating target business travel information, wherein the preset slot labels include departure location, destination, departure time, end time, mode of transportation, reason for business trip, and number of business travelers; the preset output format is a business form.

[0080] Optionally, taking the generation of target travel information for a standard complete itinerary as an example, the slot value extraction result is "Shanghai Hongqiao, Beijing Daxing, October 20, 2025, 09:30, October 23, 2025, 16:00, by plane, attending an industry summit, 2 people"; the verification result shows that all slot values ​​are semantically valid, logically consistent, and without anomalies. In the process of generating target travel information, the core mapping field is to map the slot values ​​to fixed fields according to business specifications: "Shanghai Hongqiao" is mapped to "departure point", "Beijing Daxing" is mapped to "destination", "October 20, 2025, 09:30" is standardized to "departure time", and "by plane" is mapped to "mode of transportation" with the addition of "booking channel is airline official website or third-party ticketing platform"; the supplementary derived fields are used to calculate the itinerary duration as "3 days, 6 hours, and 30 minutes", and the "cabin class defaults to economy class" is marked; a standardized travel application form is generated to meet the requirements.

[0081] Optionally, taking the generation of target travel information with colloquialized itinerary abbreviations as an example, the slot value extraction result is "Guangzhou, Shenzhen, tomorrow 08:00, the day after tomorrow 18:00, high-speed rail, project coordination, 1 person"; the verification result is that all slot values ​​are semantically valid, logically consistent, and without anomalies. In the process of generating target travel information, the standardized time field is parsed as "tomorrow 08:00" as "2025-10-16T08:00:00" and "the day after tomorrow 18:00" as "2025-10-17T18:00:00"; the supplementary business default fields are "departure details: Guangzhou South Station", "destination details: Shenzhen North Station", "seat class: second class"; and standardized information adapted to the high-speed rail booking interface is generated.

[0082] Optionally, taking the generation of target travel information for multiple mixed trips as an example, the slot value extraction result is "Harbin, Beijing, tomorrow morning 07:30, none, high-speed rail, project inspection, 1 person; Beijing, Guangzhou, evening 19:00, none, flight, project inspection, 1 person"; the verification result is that the slot values ​​of the two trips are semantically valid, logically coherent, and without anomalies. In the process of generating target travel information, splitting the multi-trip structure involves separating the two trips with semicolons, generating sub-trip information for each, and marking them with "Row Procedure Number: 1 / 2"; supplementing missing fields involves marking "End Time" as "Pending Confirmation", supplementing the default booking channels for each trip; and generating form data that supports multi-trip management.

[0083] Understandably, target travel information is generated using fixed fields and standardized formats, eliminating the need for additional parsing or conversion by downstream business modules. It can directly interface with travel application forms, ticketing booking interfaces, and itinerary management systems, significantly reducing cross-module and cross-platform data adaptation costs. This results in standardized output data and lowers downstream business integration costs. All generated target travel information is based on error-free slot extraction results, ensuring the accuracy of core information and preventing issues such as rejected travel applications and failed bookings due to information errors. This guarantees business execution reliability and improves user experience. Furthermore, the output format and fields of target travel information can be dynamically adjusted according to downstream needs. It supports form generation and database storage within internal systems as well as interface calls from external platforms, adapting to all travel scenarios including travel applications, ticketing bookings, hotel reservations, and ride-hailing services. This multi-scenario adaptation enhances system scalability.

[0084] In some optional embodiments, the travel slot extraction method may further include the following steps: if the slot value verification result indicates that there is an abnormal slot value in the slot value extraction result, then the initial travel information is re-extracted based on a preset model using prompt word engineering to obtain a second corrected slot value extraction result, wherein the second slot value included in the second corrected slot value extraction result matches the initial travel information; and target travel information is generated based on the second corrected slot value extraction result.

[0085] It should be noted that the process of correcting semantic and logical validation is to resolve the implicit problem of valid tags but abnormal slot values, and to ensure that the final output of target travel information fully conforms to the semantic rules and business logic of the travel scenario. The core logic includes the following dimensions: First, semantic and logical verification is triggered only when the error is explicitly indicated by the semantic and logical checks, unlike the first correction scenario of invalid tags and / or missing tags, focusing on the rationality of the slot value itself; Second, a more powerful pre-built model is selected to leverage its strong contextual understanding, semantic error correction, and logical reasoning capabilities to address complex error scenarios in the slot extraction model; Third, structured prompt word templates are used to clearly inform the model of the error slot type, the initial travel information context, the slot definition, and the correction standards, guiding the model to accurately locate and correct the error; Fourth, the second corrected slot value extraction result must completely match the initial travel information, meaning the corrected slot value must faithfully reflect the user's original needs without any additional additions or alterations, while also resolving the original semantic and logical errors; Fifth, the corrected second corrected slot value extraction result does not require multiple rounds of verification and can be directly used to generate the target travel information.

[0086] In some embodiments, considering common slot value anomaly types in business travel scenarios, such as semantic invalidity, logical conflict, format error, and unrecognized implicit requirements, the following are examples of the steps for the semantic and logical verification correction process. The preset slot labels include departure location, destination, departure time, end time, mode of transportation, and reason for business trip; the preset probability threshold is 0.5.

[0087] Optionally, let's take the correction of anomalies caused by semantic invalidity as an example. For instance, the initial travel information is "Book me a flight from Guangzhou to Harbin next Tuesday to attend a technical seminar and return next Thursday"; the slot value extraction result is "Guangzhou, Harbin, next Tuesday, next Thursday, airplane, attend technical seminar"; the slot label detection result is that all slot labels are valid and fully covered; the verification result is that "Guangzhou" and "Harbin" are typos and semantically invalid, triggering the correction condition.

[0088] The correction process involves constructing a targeted prompt phrase: "Based on the following travel information, 'Book me a flight from Guangzhou to Harbin next Tuesday to attend a technical seminar, returning next Thursday,' extracting six slots: departure location, destination, departure time, end time, mode of transportation, and reason for travel. It is known that 'Guangzhou' and 'Harbin' in the original extraction results are typos. Please correct the typos to ensure the slot values ​​are semantically consistent with the original input, outputting in the format of 'departure location, destination, departure time, end time, mode of transportation, reason for travel.'" After the pre-set large model parses the prompt phrase, it identifies and corrects the typos, outputting accurate slot values, resulting in the second corrected slot value extraction result: "Guangzhou, Harbin, next Tuesday, next Thursday, airplane, attending a technical seminar." The target travel information is generated directly from the corrected results to create standardized form data, adapted to the flight booking interface.

[0089] Optionally, let's take the correction of anomalies caused by format errors as an example. For instance, the initial travel information is "Book me a high-speed train from Shenzhen to Chengdu on December 8, 2025, and return on December 12, for project coordination"; the slot value extraction result is "Shenzhen, Chengdu, December 8, 2025, December 12, 2025, high-speed train, project coordination"; the slot label detection result is that all slot labels are valid and fully covered; the verification result is that "December 2025" is an invalid time format, there is no 13th month, the semantics are invalid, and the correction condition is triggered.

[0090] The correction process involved constructing a targeted prompt phrase: "Based on travel information 'Book me a high-speed train from Shenzhen to Chengdu on December 8th, 2025, and return on December 12th, for project coordination,' extracting six slots. The original extraction result showed 'December 2025' as an invalid time format. A reasonable time was inferred from the context to ensure standard format and logical consistency, and the output was in the specified format." After parsing the prompt phrase using the pre-set large model, the time format was corrected to a valid date, resulting in the second corrected slot value extraction result: "Shenzhen, Chengdu, January 8th, 2026, January 12th, 2026, high-speed rail, project coordination." The target travel information was generated as high-speed rail booking request parameters based on the standard time format.

[0091] Understandably, for anomalies that the basic model struggles to handle, such as semantic invalidity, logical conflicts, and format errors, the pre-defined large model, through its precise understanding and error correction capabilities, significantly improves the accuracy of corrected slot values, resolves slot value anomalies, and enhances the accuracy of slot extraction results.

[0092] As can be seen from the above, this embodiment obtains the initial travel information input by the user; it calls the slot extraction model and the slot existence discrimination model in parallel, extracts slot values ​​from the initial travel information according to the target slot value sequence based on the slot extraction model, and obtains the slot value extraction result; it determines the existence probability of each preset slot tag from the initial travel information based on the slot existence discrimination model, and obtains the slot tag detection result, wherein the existence probability of the valid slot tags included in the slot tag detection result is greater than or equal to the preset probability threshold; if the slot tags corresponding to the slot values ​​included in the slot value extraction result are all valid slot tags, and each valid slot tag in the slot tag detection result is covered by slot tags, then semantic and logical verification is performed on each slot value in the slot value extraction result based on the initial travel information; if the slot value verification result indicates that all slot values ​​in the slot value extraction result are without abnormalities, then the target travel information is generated based on the slot value extraction result. The processing method based on parallel call slot extraction model and slot existence discrimination model can improve the efficiency of travel slot extraction, and the consistency discrimination based on slot value extraction result and slot label detection result can improve the accuracy of slot extraction.

[0093] To better illustrate the travel slot extraction scheme in the aforementioned embodiments, please refer to 2. Figure 2 This is another schematic flowchart of the travel slot extraction method shown in the embodiments of this application, which includes the following steps: S201, construct a preset model output structure, and obtain an initial slot value dataset based on the preset model output structure. The initial slot value dataset includes real user data, and / or preset model prompt word generation data, and / or code rule generation data.

[0094] In this embodiment, a high-quality, high-coverage, and low-noise target slot value dataset is constructed through a full-link design that includes output structure optimization, multi-source data acquisition, quality filtering, semantic deduplication, data augmentation, and model training. This provides a solid data foundation for the training of subsequent slot extraction and slot existence discrimination models, and can solve the problems of high noise, narrow coverage, and weak adaptability of traditional datasets, ensuring that the model has the accurate semantic understanding and efficient reasoning capabilities required for business travel scenarios.

[0095] The system constructs a pre-defined model output structure and obtains an initial slot value dataset based on this structure. It optimizes the output structure by abandoning traditional redundant formats and adopting a slot value sequence format, outputting only pure slot values ​​in a fixed order to reduce redundant data and lay a structural foundation for efficient model inference. Through multi-source heterogeneous data fusion, it integrates three data sources: real user data, pre-defined model prompt word generation data, and code rule generation data, balancing data authenticity, diversity, and boundary scenario coverage. Based on intelligent quality scoring filtering, a dedicated quality scoring model is trained to quantify and score the initial data, dynamically removing low-quality data and selecting high-quality data to ensure dataset purity. Semantic deduplication is performed by clustering highly similar samples based on semantic similarity calculations to avoid model overfitting and improve training efficiency. Sample diversity is expanded through synonym substitution, sentence transformation, and adversarial perturbation to enhance model robustness. Finally, dual-model collaborative training is conducted. Based on the optimized target slot value dataset, a slot extraction model and a slot existence discrimination model are trained separately to ensure that both models adapt to the semantic features of the travel scenario, forming a collaborative capability for extraction and verification.

[0096] In some embodiments, the entire process of building travel scenario data and training models is combined, wherein the preset slot labels include departure point, destination, departure time, end time, mode of transportation, and reason for business trip; the preset output structure is a slot value sequence.

[0097] Specifically, the output structure is defined as a fixed sequence format that clearly states the departure point, destination, departure time, end time, mode of transportation, and reason for the business trip, without any redundant descriptive information.

[0098] In the multi-source data collection process, real user data is collected from the travel system's production environment, anonymized, and then pre-labeled using a large model and manually verified. For example, if a user inputs "Flying from Shanghai to Beijing tomorrow for a meeting, returning next Friday," it is labeled as "Shanghai, Beijing, Tomorrow, Next Friday, Airplane, Meeting." This means that the production data is first anonymized, and then pre-labeled and manually verified using a large model prompting method according to extraction rules (slot order). For example, "Flying from Shanghai to Beijing tomorrow for a meeting, returning next Friday" → Shanghai, Beijing, Tomorrow, Next Friday, Airplane, Meeting. The information in these slots is: departure location, destination, departure time, return time, mode of transportation, and reason for business trip.

[0099] The preset model prompt word generation data involves designing structured prompt words and calling a preset large model to generate diverse travel query statements. An example prompt word is: "To train a slot extraction model, some data needs to be generated. The application environment of the data is a business trip application. Please submit a business trip application in a natural, conversational style, including departure point, destination, date, mode of transportation, and reason for the trip. Omissions or aliases are allowed. Below is a business trip application example; generate data based on this example:" {"instruction": "Flying to Xiamen today for bidding purposes, returning to Siping on July 17, 2025",} "output": "Siping, Xiamen, Today, 2025-07-17, Aircraft, Bidding"} The output is as follows: {"instruction": "Could you please submit a business trip application for me, flying from Zhangjiakou to Shuozhou the day after tomorrow, returning next Tuesday, for market research?"} "output": "Zhangjiakou, Shuozhou, the day after tomorrow, next Tuesday, by plane, market research"}.

[0100] The code rules generate data through Python scripts based on sentence templates and slot values. In other words, Python scripts are written to automatically generate fine-tuned corpora based on slot value combinations and sentence templates. For example, the Python script might be: templates = [ "Help me initiate a business trip application from {departure point} to {destination point}". "{Departure Time} From {Departure Location} Fly to {Destination} for {Business Trip Reason}", "Please help me submit a business trip application. My departure time is [date] from [place of departure] to [destination], and my return time is [date]. Reason for business trip is [reason]." ] cities = ["Beijing", "Shanghai", "Guangzhou", "Shenzhen", "Chengdu", "Hangzhou"...] dates = ["Tomorrow", "Next Wednesday", "August 15th", "Early Next Month"...] Advantages: Covers boundary combinations (such as less popular cities + special dates), improves recall rate in long-tail scenarios, and can quickly generate a large amount of fine-tuning data for specific sentence structures. Labeling: The program automatically generates corresponding slot labels, 100% accurate.

[0101] S202, score the quality of each slot value in the initial slot value dataset, and filter the initial slot value dataset based on the score results.

[0102] Based on model fine-tuning, the input consists of manually labeled positive samples (high-quality sentences selected by humans, scored 8-10, such as "Fly from Guangzhou to Chengdu on October 20, 2025 to attend an industry summit, and return on October 23"); negative samples (sentences containing noise, grammatical errors, and slot conflicts, scored 0-4, such as "Goed to Shanghai yesterday and will return tomorrow, by car, no reason"); and neutral samples (sentences that are partially vague or colloquial but acceptable, scored 5-7, such as "Fly to Shenzhen next Wednesday for a meeting"). After training, the root mean square error is less than 0.8, ensuring that the absolute deviation between the predicted score and the human score is small. The correlation coefficient of the human scores is greater than 0.85, indicating that the model prediction and human judgment have a high degree of consistency.

[0103] In dynamic threshold filtering, hard filtering removes samples with scores less than 4, such as "Going to Beijing, tomorrow, by plane," which is semantically incomplete and has a score of 3.2. Soft filtering sends samples with scores of 4-6 to the enhancement pool, such as "Flying from Hangzhou to Xi'an the day after tomorrow, business trip," which contains typos and incomplete information and has a score of 5.8. Samples with scores greater than or equal to 7 enter the quality pool, such as "Next Thursday I'll take the high-speed rail from Wuhan to Changsha to meet on a project, and return next Saturday," which has a score of 8.6. Dynamic sampling retains 100% of samples with scores greater than or equal to 8, samples with scores of 7-8 at 80%, and samples with scores of 6-7 at 50% after manual verification of validity.

[0104] The intelligent quality scoring filter in this step specifically includes: Step (1): Construct a large-scale data quality scoring model (Quality Scoring LLM). To uniformly evaluate the three types of data quality, a dedicated "travel statement quality scoring model" Q-Model is trained. Model selection: Fine-tuned based on a lightweight pre-trained model (such as Qwen3-7B). Through expert review of the training corpus, training data with different standards are obtained. Positive samples are manually selected high-quality statements (clearly labeled, semantically complete, and unambiguous), with a score of 8-10; negative samples are statements containing noise, grammatical errors, and slot conflicts, with a score of 0-4; neutral samples are partially ambiguous or colloquial but acceptable statements, with a score of 5-7.

[0105] The input is the original text (e.g., "Tomorrow, Shanghai to Beijing, economy class"). The output is the quality score ∈ [0,10], a regression task, with MSE as the loss function. The evaluation metrics are: on the independent validation set, the model must meet the following two core metrics: (a) Root Mean Square Error (RMSE) < 0.8, ensuring a small absolute deviation between the predicted score and the human rating; (b) Pearson correlation coefficient with the human rating > 0.85, indicating a high degree of consistency between the model prediction and human judgment.

[0106] Step (2), input all samples in the D_raw dataset into the Q-Model for automatic scoring and perform dynamic threshold filtering. Under hard filtering, samples with a score < 4 (obviously low-quality, incorrect, or irrelevant) are excluded. Under soft screening, samples with scores between 4 and 6 are sent to the "pending enhancement pool" for semantic rewriting or manual review; samples with scores ≥ 7 are directly sent to the "high-quality pool". Under the dynamic sampling strategy, for samples with scores ≥ 8, they are retained in the original proportion; for samples with scores between 7 and 8, they are sampled at a proportion of 80%; for samples with scores between 6 and 7 that are manually confirmed to be valid, they are sampled at a proportion of 50% and added for enhancement.

[0107] S203, perform deduplication on the filtered dataset based on semantic similarity.

[0108] Use Sentence-BERT to calculate the semantic similarity of samples, cluster samples with a cosine similarity greater than 0.95 to achieve deduplication, improve the generalization ability of the model, increase training efficiency, and reduce training costs, etc. For example: Sample 1: "Shanghai, Beijing, tomorrow, next Friday, plane, attend a meeting"; Sample 2: "Shanghai, Beijing, tomorrow, next Friday, plane, participate in a meeting"; Processing result: The two samples are highly semantically similar. Retain Sample 1 and exclude Sample 2 to avoid redundant training data.

[0109] S204, perform data enhancement on the deduplicated dataset to obtain the target slot value dataset.

[0110] The data enhancement methods for the deduplicated dataset can include synonym replacement enhancement, sentence pattern transformation enhancement, and adversarial perturbation enhancement.

[0111] Synonym replacement enhancement is used for entity replacement based on a domain synonym dictionary and slot recognition. Sentence pattern transformation enhancement is used to generate diverse sentence patterns with semantic equivalence using large model prompts, and adversarial perturbation enhancement is used to simulate real noise by inserting irrelevant words and injecting typos while retaining correct slot annotations.

[0112] Optionally, synonym replacement enhancement can include replacing city aliases ("Imperial Capital" replaced by "Beijing", "Hu" "Shanghai"), date expression replacement ("the next day" replaced by "tomorrow", "next Wednesday" "Wednesday 7 days later"), transportation mode replacement ("high-speed rail" replaced by "high-speed railway"), and cabin term replacement ("economy class" "Y class", "coach class"). For example, "Shanghai, Beijing, tomorrow, next Friday, plane, attend a meeting" is enhanced to "Hu, Beijing, the next day, next Friday, civil aviation, participate in a meeting".

[0113] Optionally, sentence transformation enhancement is to change the active voice to the passive voice, and / or change the statement to a question. For example, "Book me a high-speed train ticket from Guangzhou to Shenzhen tomorrow" is enhanced to "Please book me a high-speed train ticket from Guangzhou to Shenzhen tomorrow" and "Is there a high-speed train from Guangzhou to Shenzhen tomorrow?".

[0114] Optionally, adversarial perturbation enhancement is to insert irrelevant words and inject typos. For example, inserting the irrelevant word "By the way", we get "By the way, is there still an economy class ticket from Shanghai to Beijing tomorrow?". Replace the correct words in the dataset with typos and add typos to the dataset to obtain "Tomorrow, shanghai flies to beijing economy class", while retaining and annotating the correct slots to improve robustness. "By the way, is there still an economy class ticket from shanghai to beijing tomorrow? For business trip" is annotated as "Shanghai, Beijing, tomorrow, no, airplane, business trip"; "Help me initiate a business trip application from Guangzhou to Harbin, departing next Tuesday" is annotated as "Guangzhou, Harbin, next Tuesday, no, no, business trip".

[0115] S205. Perform model training based on the target slot value dataset to obtain a slot extraction model and a slot existence discrimination model.

[0116] As can be seen from the above, in this embodiment, multi-source data fusion covers real scenarios, boundary cases, and long-tail requirements. Intelligent filtering and deduplication significantly reduce the noise rate of the dataset. Dynamic enhancement greatly improves the sample diversity. Operations such as adversarial perturbation and synonym replacement in the data enhancement link enable the model to accurately extract slots even when facing typos, vague expressions, and interference from irrelevant information, thereby enhancing the model's robustness and adapting to complex business travel scenarios. In addition, the slot extraction model and the slot existence discrimination model trained based on the same target dataset have consistent semantic understanding criteria and can accurately cooperate to complete consistency verification.

[0117] Further, please refer to Figure 3 , Figure 3 which is another process schematic diagram of the business travel slot extraction method shown in the embodiment of the present application, and it includes the following steps: S301. Use the target slot value dataset as the first input data, and the preset slot value sequence corresponding to the target slot value dataset as the first supervision data to form a first training sample pair of the first input data and the first supervision data.

[0118] In this embodiment, the training process of the slot extraction model is a full-link design of sample pair construction, two-stage adaptive hierarchical fine-tuning, training output, and verification optimization. Based on the high-quality target slot value dataset, the Qwen3-0.6B large model is adapted to a business travel scenario-specific slot extraction model. The core goal is to achieve high-precision understanding of business travel semantics and structured slot extraction by the model while ensuring low computational overhead.

[0119] It should be noted that the training sample pair construction uses user natural language commands from the target slot value dataset as the first input data and standardized preset slot value sequences as the first supervision data, forming a one-to-one correspondence between input and supervision training sample pairs to ensure that the model learns the mapping relationship between travel commands and slot extraction results. The first stage of fine-tuning, i.e., adapter module fine-tuning, freezes all original parameters of the Qwen3-0.6B backbone model and trains only the low-rank matrix parameters and slot label embedding parameters of each layer, i.e., the first target parameters, to quickly adapt to travel tasks without destroying pre-trained knowledge. The second stage of fine-tuning only unfreezes the parameters of the preset number of layers at the bottom of the backbone model, such as the last 12 layers, based on their corresponding low-rank matrix parameters. The rank matrix parameters are finely tuned while the remaining layers remain frozen. A top-down learning rate decay mechanism is introduced to enhance the high-level semantic adaptation capability. Model training involves inputting training sample pairs into the second fine-tuned model, using negative log-likelihood and slot classification cross-entropy as the task loss function, and training the model to output a target slot value sequence consistent with the preset slot value sequence. Validation data is selected from the target slot value dataset, such as 10% of the first preset proportion, to evaluate the model's generalization ability. Model optimization iteration is based on the slot value sequence prediction results of the validation data, adjusting hyperparameters such as the learning rate and decay factor to ultimately obtain a high-precision and high-efficiency slot extraction model.

[0120] In other words, fine-tuning the adapter module requires freezing the backbone model parameters, including all the original parameters of Qwen3-0.6B. (Including embedding, all attention / MLP weights, and lm_head) are completely frozen, and only the LoRA block parameters are trained to achieve rapid adaptation to downstream tasks. The AdamW optimizer is used to update only the LoRA parameters and possible embedding layers. This stage converges quickly to avoid destroying pre-trained knowledge. Let the pre-trained language model parameters be... The parameters of the adapter module are The parameters of the slot output layer are The optimization objective during adapter module fine-tuning is:

[0121] in, For user input, These are actual slot labels. () represents the model prediction function. The task loss function is (negative log-likelihood + slot classification cross-entropy). During training, [the following will be removed]. The gradient in the model is the gradient of the frozen language model.

[0122] The trainable parameters are: the low-rank matrix parameters of all LoRA modules, denoted as... The required prompt template or special token for slot generation is embedded. If included, it is denoted as... If not introduced, then It can be considered as a fine-tuning part of lm_head or zero.

[0123] In some embodiments, the entire process of training a business travel scenario model is described, with the first model using Qwen3-0.6B, the preset slot value sequence being departure point, destination, departure time, end time, mode of transportation, and reason for business trip, and the first preset ratio being 10%.

[0124] Optionally, the target slot value dataset contains optimized high-quality samples, and one sample is selected to construct a training pair. The first input data is "By the way, are there any economy class seats left for Shanghai to Beijing tomorrow? For business trips," a realistic scenario expression containing noise and typos; the preset slot value sequence, i.e., the first supervised data, is "Shanghai, Beijing, tomorrow, no, plane, business trip," a standardized slot value sequence with no redundant data. Thus, the first training sample pair is "By the way, are there any economy class seats left for Shanghai to Beijing tomorrow? For business trips" and "Shanghai, Beijing, tomorrow, no, plane, business trip."

[0125] S302, freeze the backbone model parameters in the first model, and perform the first fine-tuning of the first model based on the first target parameters to obtain the first fine-tuned model. The first target parameters include the low-rank matrix parameters and slot label embedding parameters corresponding to each layer of the backbone model parameters.

[0126] During the first fine-tuning of the first model, the model parameter configuration process freezes all original parameters (θ) of the Qwen3-0.6B backbone model. base This includes the embedding layer, the attention / MLP weights of the 24-layer Transformer, and the lm_head, whose gradient propagation is removed. The first target parameter is defined as the low-rank matrix parameter (θ). adapter ), configure a low-rank matrix with rank 8 for each Transformer attention layer; slot label embedding parameters (θ) slot ): Introduce special data embeddings corresponding to 6 slots, with dimensions consistent with the model's hidden layer.

[0127] The training process uses the AdamW optimizer with negative log-likelihood and slot classification cross-entropy as loss functions. It trains on 10 samples and quickly converges to obtain the first fine-tuned model. At this point, the model can initially identify basic slots in travel scenarios, but its accuracy in recognizing complex semantics such as ambiguous time and implicit needs needs improvement.

[0128] S303, Unfreeze the backbone model parameters located at the bottom preset layer in the first fine-tuning model, keep the backbone model parameters other than the preset layer frozen, and perform a second fine-tuning on the first fine-tuning model based on the low-rank matrix parameters corresponding to the backbone model parameters of the preset layer to obtain the second fine-tuning model.

[0129] During the second fine-tuning of the first fine-tuned model, the parameter unfreezing strategy was to only unfreeze the parameters of the last 12 layers of the Qwen3-0.6B backbone model, i.e., the bottom preset layer number K=12, while keeping the parameters of the remaining 12 layers frozen. A top-down learning rate decay was introduced during the learning rate configuration process, with the top layer (layer 24) having a learning rate η. top =2e-4, decay factor δ=0.9, learning rate η of layer l l =η top δ 24-l For example, in the 23rd layer, η = (2e-4) × 0.9 = 1.8e-4.

[0130] In other words, during the second fine-tuning of the first fine-tuned model, layered LoRA is unfrozen for fine-tuning, and high-level semantic enhancement is performed. Building upon stage one, the sensitivity of the model's higher levels to travel semantic structures is further activated, improving the recognition accuracy of complex slots (such as date, city, and mode of transportation). Layered strategy: Although LoRA itself does not "unfreeze" the original weights, "layered adaptation" can be achieved by enabling LoRA only in the last K layers, or by setting different learning rates for different layers of LoRA. This solution adopts the latter: all LL layers retain the LoRA module; only the LoRA parameters of the last K layers (e.g., K=12, corresponding to the high-level semantic layers of Qwen3-0.6B) are updated with a high learning rate, while the LoRA parameters of the remaining layers continue to be trained but with a significantly reduced learning rate (or frozen). Assume the Transformer backbone has L layers, and the parameters of the last K layers (e.g., K=14) are unfrozen. The optimization target has been updated to:

[0131] At this point, the gradient propagates to... The remaining layers remain frozen. To prevent high-level LoRA updates from causing training instability, a top-down learning rate decay is introduced:

[0132] in For the top learning rate, As the attenuation factor, This is the current layer number.

[0133] The training process, based on the model parameters after the first fine-tuning, continues to use the same optimizer and loss function to train on 8 samples, focusing on optimizing the upper layer's ability to understand complex travel semantics, resulting in the second fine-tuned model. At this point, the model can accurately identify slots in complex scenarios such as typos, ambiguous times, and implicit modes of transportation.

[0134] It should be noted that this strategy enables higher-level LoRA to achieve greater adaptability, while keeping mid-to-low-level LoRA stable, thereby enhancing task specificity without compromising general language capabilities. The instruction fine-tuning paradigm continues, but a high learning rate is applied only to specified LoRA layers. Gradients are not backpropagated. Therefore, there is no risk of gradient explosion; learning rate decay is mainly used to control the intensity of task adaptation and avoid overfitting at higher levels.

[0135] Traditional slot extraction methods perform well when handling standard instructions, but tend to fail when dealing with colloquial expressions, implicit intentions, and long-tailed place names. This application introduces an optimization mechanism for boundary scenarios during adaptive hierarchical fine-tuning. The boundary scenario data includes the following categories: Fuzzy time expression: Users can input non-standard time formats, such as "the Wednesday after next" or "the first day after National Day".

[0136] Deep semantic description: The user input implies the mode of transportation, such as "flying from Guangzhou to Beijing today and returning the day after tomorrow".

[0137] Negation and Correction Intent: Users may include self-correction in their statements, such as, "Book me a ticket to Shanghai, oh no, I mean Hangzhou." The model needs to accurately extract "Hangzhou" instead of "Shanghai."

[0138] Mixed itineraries (transportation): A single sentence contains multiple itineraries, for example: "First, I book the high-speed train from Harbin to Beijing tomorrow morning, and then I book the flight to Guangzhou tonight."

[0139] Typos: There are typos in the user input, which need to be corrected, such as "Help me initiate a business trip application from Guangzhou to Harbin."

[0140] Special description: If the input is extremely simple, or even uses special symbols, such as "Guangzhou -> Shenzhen, going today, high-speed rail".

[0141] S304, the first training sample pair is input into the second fine-tuning model for training, and the target slot value sequence corresponding to the preset slot value sequence is output as the first output data.

[0142] All training sample pairs are input into the second fine-tuned model. The model semantically encodes user instructions through forward propagation and outputs the target slot value sequence based on the fine-tuned low-rank matrix parameters and high-level semantic parameters. In the input sample pair, the user instruction is "Help me initiate a business trip application from Guangzhou to Harbin, departing next Tuesday"; the model output, the first output data, i.e., the target slot value sequence, is "Guangzhou, Harbin, next Tuesday, none, none, business trip," which is completely consistent with the first supervised data, and the loss function value approaches 0.

[0143] S305: Select first verification data with a first preset ratio from the target slot value dataset, input the first verification data into the trained second fine-tuning model, and output the slot value sequence prediction result.

[0144] The validation data is selected by randomly choosing 10% of the samples from the target slot value dataset as the first validation data, covering all scenarios including standard expressions, multiple trips, less popular cities, and special dates. The prediction execution step is to input the first validation data into the trained second fine-tuning model and output the predicted slot value sequence. For example, the validation input is: "Go to Beijing on the first day after National Day, return the following Wednesday, bidding"; the predicted output is: "Beijing, none, the first day after National Day, the following Wednesday, none, bidding"; the true label is: "Beijing, none, the first day after National Day, the following Wednesday, none, bidding", and the prediction result matches the true label perfectly.

[0145] S306. Based on the slot value sequence prediction results, the trained second fine-tuning model is optimized to obtain the slot extraction model.

[0146] Based on the prediction results from the validation data, the model performance metrics were calculated. For prediction biases in a few less popular cities such as "Shuozhou" and "Yibin," the hyperparameters were adjusted, and the model was trained again for two epochs. This improved the accuracy on the validation set, ultimately yielding the slot extraction model.

[0147] As shown above, the adaptive hierarchical LoRA fine-tuning method for generative large models, while freezing the Qwen3-0.6B backbone parameters, efficiently achieves travel slot extraction through two-stage progressive low-rank adaptation. Through two-stage fine-tuning and a learning rate decay mechanism, the model accurately captures semantic features in the travel domain, significantly improving its ability to recognize complex scenarios such as colloquial expressions, misspellings, ambiguous times, and implicit needs, thereby enhancing the model's slot extraction accuracy and adapting to complex travel scenarios. The model is based on a lightweight architecture, coupled with a slot value sequence output structure, significantly improving inference speed to ensure model inference efficiency and meet real-time interaction requirements. Furthermore, updating only a portion of the model parameters significantly saves GPU memory and computing power compared to full parameter fine-tuning, requiring no high-end hardware support, greatly reducing training costs, and improving the feasibility of industrial application. Furthermore, the class balancing weighting strategy introduced during fine-tuning enhances the model's attention to long-tail samples such as less popular cities, special dates, and multiple itineraries. In the long-tail sample test set, the slot extraction accuracy is significantly higher than the model without this strategy, ensuring stable extraction performance under complex travel needs and improving model robustness. In the long-tail case optimization scheme, to address the class imbalance problem caused by the scarcity of long-tail data samples such as less popular cities and uncommon abbreviations, this application introduces a class balancing weighting strategy (or Focal Loss mechanism) in adaptive hierarchical fine-tuning. The system statistically analyzes the frequency of each slot category in the training set. For long-tail slots with extremely low occurrence frequencies (such as specific event codes and less popular sites), a higher penalty weight is assigned when calculating the loss function. During model training, for "difficult samples" (usually long-tail cases) with low prediction probabilities, the Focal Loss mechanism amplifies their gradient contribution. This strategy allows the adaptive fine-tuning process to focus more attention on these difficult-to-identify long-tail samples, rather than being dominated by a large number of common, simple samples, thereby significantly improving the recall rate in long-tail scenarios.

[0148] Experimental results show that the proposed method achieves a slot extraction accuracy of 92.7% on long-tailed samples, significantly higher than the 89.1% of full-parameter fine-tuning and the 84.6% of single-stage LoRA fine-tuning. The two-stage adaptive fine-tuning strategy proposed in this invention achieves high-precision domain adaptation with extremely low parameter update volume through a progressive optimization of "lightweight adaptation → local unfreezing." This perfectly meets the core requirements of travel intelligence agents for "high accuracy + low computational overhead," demonstrating significant technological innovation and industrial application value.

[0149] Further, please refer to Figure 4 , Figure 4 This is another schematic flowchart illustrating the travel slot extraction method shown in the embodiments of this application, which includes the following steps: S401, the target slot value dataset is used as the second input data, and the binary classification labels corresponding to the target slot value dataset generated based on the target slot value sequence are used as the second supervision data to form a second training sample pair of the second input data and the second supervision data, and the second training sample pair is input into the second model.

[0150] In this embodiment, the slot existence discrimination model training process involves sample pair construction, three-stage network inference, and verification optimization. Based on a high-quality target slot value dataset, a lightweight and efficient slot existence discrimination model specifically designed for business travel scenarios is trained. This model enables the slot existence discrimination model to accurately determine whether each preset slot exists in the user input, providing a reliable probabilistic basis for subsequent parallel verification.

[0151] It should be noted that the training sample pairs are constructed using the semantic features of user commands in the target slot value dataset as the second input data, and binary labels generated based on the slot value sequence (1 for presence and 0 for absence) as the second supervision data. This forms a one-to-one correspondence between the input features and the supervision labels, ensuring that the model learns the mapping relationship between travel semantics and slot existence. The second model architecture design adopts a Transformer-based approach. The Encoder's lightweight architecture adds three fully connected layers and two layer normalization modules to the original architecture, balancing feature extraction capability and inference efficiency. In the inference process, the input layer maps semantic features to a high-dimensional space, the feature extraction layer enhances key semantic information, and the scoring layer outputs a dimension vector matching the preset number of slots, which is then mapped to the existence probability in the (0,1) interval using the Sigmoid function. Supervised training is performed using mean squared error as the loss function, and the existence probability output by the trained model closely matches the binary classification supervision label. Validation data is selected from the target slot value dataset, such as 10% of the second preset proportion, to evaluate the model's generalization ability. Model optimization iterations adjust the model weights and hyperparameters based on the predicted existence probability values ​​of the validation data to ensure that the model meets the criteria on the independent validation set, ultimately resulting in a high-precision, lightweight slot existence discrimination model.

[0152] In some embodiments, the entire training process of the slot existence discrimination model is combined, wherein the preset slot labels include departure point, destination, departure time, end time, mode of transportation, and reason for business trip; the second model is optimized based on the TransformerEncoder architecture; the second preset ratio is 10%; and the preset probability threshold is 0.5.

[0153] Specifically, the target slot value dataset contains optimized high-quality samples, and one of the samples is selected to construct a training pair. The second input data is the semantic features of the user command "By the way, are there any economy class cabins left for Shanghai to Beijing tomorrow? For business trips" and is the vector representation of the [CLS] label in the Transformer Encoder; the second supervision data, namely the binary classification label, is generated based on the standard slot value sequence "Shanghai, Beijing, tomorrow, none, airplane, business trip", where "departure location = exists (1), destination = exists (1), departure time = exists (1), end time = does not exist (0), mode of transportation = exists (1), reason for business trip = exists (1)", and the label sequence is [1,1,1,0,1,1]; the second training sample pair is the [CLS] vector feature, [1,1,1,0,1,1].

[0154] A lightweight post-processing verification mechanism based on a "slot existence discriminator" performs probability calibration and logical filtering on the main task output during the model inference stage, significantly reducing the false positive error rate and improving system robustness. The core processing flow is as follows: the slot existence discriminator outputs the existence of each slot based on the user's text. The results of the slot existence discriminator and the slot extraction model are input into the verification module for consistency checks (e.g., if a slot is extracted, but the slot existence discriminator outputs that the slot does not exist → it may be a misidentification). To improve the efficiency of slot existence detection, the slot existence discriminator is trained using a small model. It is based on the Trasformer Encoder architecture, with three fully connected layers and two normalization layers added. The specific structure is as follows: Figure 5 Embedding represents the [CLS] marker in the Encoder structure, Fully Connected Layer represents a fully connected layer, and LayerNorm represents layer normalization.

[0155] S402, after the second training sample pair is processed sequentially by the input layer, feature extraction layer and scoring layer in the second model, the output is a dimension vector that matches the number of preset slot labels, and the dimension vector is mapped by a function to obtain the second output data corresponding to the existence probability of each preset slot label.

[0156] For the input layer, firstly, the [CLS] marker in the Encoder structure represents the semantic expression of the user input. Then, it passes through a fully connected layer with an output dimension of 1024, used to map the original features to a higher-dimensional space to better capture complex semantic information. The input layer processing involves feeding the second input data into the first fully connected layer, mapping the output dimension to 1024, and enhancing the feature expression through the GELU activation function and layer normalization, as shown in the following formula:

[0157] in, This represents the output of the first fully connected layer, with a dimension of 1024. W1 represents the input layer weights, b1 represents the input layer bias, GELU represents the activation function performing a non-linear transformation on the output to improve the model's generalization ability, and LayerNorm represents layer normalization.

[0158] The feature extraction layer passes the normalized features to the second fully connected layer. This layer has both 1024 input and output dimensions, as shown in the following formula:

[0159] in b1 represents the output of the feature extraction layer, with a dimension of 1024; W2 represents the weight of the feature extraction layer, and b2 represents the bias of the feature extraction layer.

[0160] The scoring layer, after the normalization layer, allows features to enter the third fully connected layer. This layer has an input dimension of 1024 and an output dimension of 5, used to determine the existence of the five slots. Each dimension of this 5-dimensional vector corresponds to a logit for a binary classification task. A sigmoid function is applied to each logit, compressing it to the range (0, 1), which serves as the probability that each class is positive, i.e., the score for each of the five slots. The formula is as follows: x5 = Sigmoid(W3x) 1024 +b3) Here, x5 represents the output of the scoring layer (dimension 5), W3 represents the weights of the scoring layer, and b3 represents the bias of the scoring layer. The sigmoid activation function is used to generate the probability of each slot's existence. Among the five outputs of the scoring layer, if the value of a slot is greater than or equal to 0.5, it indicates that the corresponding slot exists. Otherwise, it indicates that the slot does not exist.

[0161] S403: Select second verification data with a second preset ratio from the target slot value dataset, input the second verification data into the trained second model, and output the existence probability prediction value of each preset slot label.

[0162] All second training samples are input into the second model, with the mean squared error as the loss function. The optimization objective is to minimize the difference between the existence probability of the model output and the squared error of the second supervision data. For example, the initial model output existence probability is [0.92, 0.88, 0.95, 0.32, 0.85, 0.90]; the second supervision data is [1, 1, 1, 0, 1, 1]. The loss is calculated by using the mean squared error function to calculate the difference between the two. The model weights are updated by backpropagation. After iterative training with 15 samples, the model output probability converges to [0.98, 0.99, 0.97, 0.05, 0.96, 0.94], which closely matches the supervision labels.

[0163] The validation data is selected by randomly choosing 10% of the samples from the target slot value dataset as the second validation data, covering all scenarios such as colloquial omissions, multiple trips, less popular cities, and typos. The prediction step involves inputting the semantic features of the second validation data into the trained second model, which outputs the predicted existence probability value for each slot. For example, if the validation input instruction is "Help me initiate a business trip application to Beijing now, departing tomorrow", the model outputs the predicted existence probability values: [0.08, 0.99, 0.96, 0.21, 0.35, 0.92]; the true binary classification label is: [0, 1, 1, 0, 0, 1]. The predicted values ​​deviate very little from the true labels, meeting the accuracy requirements.

[0164] S404. Based on the predicted existence probability, the trained second model is optimized to obtain the slot existence discrimination model.

[0165] Based on the prediction results from the validation data, the performance metrics of the computational model have met the core evaluation metrics. For a few ambiguous expressions such as "going to Chengdu after National Day," where there is a probability prediction bias in the "end time," the learning rate of the third fully connected layer was adjusted, and the model was retrained with three more samples to finally obtain the slot existence discrimination model.

[0166] As shown above, the trained slot existence discrimination model can accurately identify the existence status of each slot in the user input, significantly improving the accuracy of identifying erroneous slots and effectively intercepting false positive errors. This results in a substantial improvement in the overall slot extraction accuracy compared to the basic model, significantly reducing the risk of subsequent business logic failures and achieving accurate slot existence judgment while intercepting false positive errors. The slot existence discrimination model is trained on a multi-source heterogeneous, data-augmented target slot value dataset, making it adaptable to complex scenarios such as colloquial omissions, typos, ambiguous expressions, and multi-run scenarios, thus enhancing the model's robustness.

[0167] Based on experimental data, the method for extracting travel slots is described below. This method includes the following steps: Step 1: Constructing a high-quality dataset. Adjust the structure of the output in the fine-tuning data and collect data from three types of sources: real user data, model-generated prompt words, and rule-generated data. Then, use an intelligent quality scoring filtering method to filter out low-quality data.

[0168] Step 2: Automatic verification of slot extraction results. By adjusting the network structure of the small model, the probability of each slot existing is obtained to determine the accuracy of the travel slot extraction model and improve its robustness.

[0169] Step 3: Design and training of the travel slot extraction model. The travel slot extraction model is adaptively fine-tuned hierarchically using the constructed high-quality dataset. This step offers better generalization ability and efficiency compared to traditional full-parameter fine-tuning.

[0170] Step 4: Result Verification. The model automatically verifies the slot extraction results. If the verification result indicates that the slot information extraction is incorrect, a larger and more capable model will automatically correct it using prompt word engineering. If the slot information is correct, the slot information will be automatically filled into the system.

[0171] The following table compares the accuracy, response time, and other data of different slot extraction methods:

[0172] Test Dataset Description: This experiment is based on a test set containing 4,000 user inputs across various travel scenarios. This dataset covers 50+ cities, 200+ date expressions, 10+ modes of transportation, 50+ reasons for business trips, and various sentence structures used by users. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio to ensure even distribution across scenarios. The testing environment uses an NVIDIA P100 GPU, and the model is deployed privately. Analysis of experimental results: The method of the present invention achieved an accuracy of 95.8% on the test set, which is 8.2 percentage points higher than the LLM direct extraction method (87.6%) and 13.5 percentage points higher than the traditional BiLSTM+CRF method (82.3%). From the experimental data, although the traditional method has the fastest response speed, it has the lowest accuracy and is applicable to low-complexity scenarios. Although the LLM direct extraction method can adapt to more complex scenarios, the response time is too slow and needs to be improved in the travel scenario. The accuracy also has room for improvement. Overall, the method of the present invention is the best in terms of accuracy and applicable scenarios, and has a greater advantage in terms of latency than directly using LLM to extract slots. The accuracy improvement of the method of the present invention mainly comes from three aspects: (1) the high-quality dataset makes the model more accurate in understanding the semantics of the travel domain; (2) the slot existence discriminator effectively intercepts 15.7% of the "ghost slot" errors; (3) the adaptive hierarchical fine-tuning strategy improves the model's ability to recognize boundary scenarios by 22.3%.

[0173] In summary, see Figure 6 The present invention provides steps including the following: (1) Get user input and receive text information input by the user.

[0174] (2) Concurrently call the slot existence discriminator model and the slot extraction model to obtain the slot extraction results (specific slot information) and slot information detection results (whether each slot exists). ① Construct a high-quality fine-tuning dataset. The data sources mainly include: data generated using code, data generated by the large model using prompt words, and real user data. ② Use the constructed data to perform LoRA fine-tuning on the Qwen3-0.6B large model.

[0175] (3) Slot verification: The slot existence discriminator outputs whether each slot exists. Combined with the actual slot extraction results, it verifies whether the slot extraction is complete and whether there are any extraction errors.

[0176] (4) Determine if slot extraction passes. ① If the slot extraction model extracts incorrect slot information (e.g., it should extract 5 slots, but the fine-tuning model only extracts 4), then use the large model to correct the slots using prompt word engineering as a fallback. ② If the slot information extracted by the fine-tuning model is correct.

[0177] In intelligent dialogue systems for business travel, the slot extraction module generally suffers from three core problems: slow inference speed, low recognition accuracy, and high resource consumption. Especially when facing complex scenarios such as users' colloquial expressions and mixed domain terminology, traditional models or general-purpose language models suffer from low semantic understanding efficiency due to high training data noise, structural mismatch with vertical domains, and high parameter redundancy, making it difficult to meet the dual requirements of real-time interaction and high-precision extraction. Specifically, the method in this embodiment aims to overcome the following key technical bottlenecks: Optimize the dataset: Obtain high-quality training data for the travel industry by adjusting the data structure of the model's output, collecting data from multiple sources, scoring data quality, and augmenting data.

[0178] Improve the efficiency of travel slot extraction: By adaptively and hierarchically fine-tuning a small but large language model, such as using Qwen3-0.6B, leverage its high-efficiency response capabilities to quickly extract slot information; Enhance the verification capability of slot extraction results: By training the slot existence discriminator, the extracted slot results are automatically verified, thereby enhancing the reliability and improving the accuracy of the slot extraction results; When the fine-tuned model fails to extract slot information correctly, it automatically switches to the large language model to perform contextual understanding and extract slot information, thereby improving the recognition success rate in complex semantic scenarios. By dynamically switching between different model strategies, a more intelligent and smoother slot extraction process is achieved, improving the efficiency and accuracy of the overall slot extraction environment.

[0179] This invention provides a method to improve the speed and accuracy of slot extraction in a travel agent, achieving technological breakthroughs in three aspects: Data structure optimization: abandoning traditional JSON output and adopting a "slot value sequence" format, reducing redundant tokens by approximately 40% and improving inference speed by over 30%; Parallel verification architecture: introducing a "slot existence discriminator" for the first time in the travel domain, enabling real-time verification in parallel with the main model, reducing the false recognition rate by 50%; Adaptive hierarchical fine-tuning: proposing a "two-stage LoRA + high-level unfreezing" strategy, achieving accuracy comparable to or even better than full parameter fine-tuning on travel slot tasks with only 0.1% parameter updates. The key technical implementations of this invention mainly include the following aspects: 1. A method for constructing a high-quality slot extraction model training dataset for business travel scenarios: A high-quality, high-coverage, and low-noise corpus is constructed through the strategy of "multi-source heterogeneous data fusion + intelligent quality scoring filtering + dynamic sample augmentation". Specifically, it includes: (1) Optimizing the model output structure, retaining only slot information, reducing redundant tokens, and improving inference efficiency. In constructing a large slot extraction model fine-tuning dataset, the common output structure is in JSON format, such as: {"departure date": "next Tuesday afternoon", "departure location": "Guangzhou", "destination": "Beijing", "reason for business trip": "meeting", ...}. However, this output format will output redundant tokens, which seriously affects the output efficiency of the model. Therefore, we optimized the output format of the fine-tuning data, outputting only slot information and avoiding irrelevant information as much as possible, thereby optimizing the output efficiency of the model. The output format is as follows: "Shanghai, Beijing, tomorrow, next Friday, airplane, meeting". The above slots represent "departure location, destination, departure time, end time, mode of transportation, reason for business trip". In this way, the model outputs as few tokens as possible that are irrelevant to the slot information, thereby improving the efficiency of the model. (2) Integrate three types of data sources - real user data, large model generated data and rule template generated data to cover diverse expressions and boundary scenarios; (3) Train a dedicated quality scoring model to automatically score and filter samples, and combine dynamic sampling and semantic deduplication to screen high-quality data; (4) Enhance sample diversity and model robustness through synonym replacement, sentence transformation and adversarial perturbation. This method aims to provide a high-quality data foundation for slot extraction model fine-tuning and improve model performance and generalization ability.

[0180] 2. An automatic verification method for slot extraction results: Construct a lightweight "slot existence discriminator" that receives the semantic feature vector output by the language model and outputs the existence probability of each slot (∈[0,1]); This discriminator is deployed in parallel with the main slot label extractor, and the consistency of their outputs is verified during the inference phase: If a slot is extracted by the main model but its existence probability is <0.5, it is determined to be a "ghost slot", triggering a filtering, clarification, or downgrade processing mechanism; This method does not require additional labeling, and the labels are automatically generated by the main task, which can effectively intercept false positive errors and greatly improve system reliability and user experience.

[0181] 3. An Adaptive Hierarchical Fine-Tuning Method for Travel Slot Extraction: Addressing the overfitting risk due to data scarcity and high computational cost of full parameter updates when directly fine-tuning general-purpose language models for travel slot extraction tasks, this method proposes a two-stage adaptive fine-tuning strategy: "lightweight adaptation followed by hierarchical fine-tuning." The first stage freezes all original parameters of the Qwen3-0.6B backbone model, training only the low-rank adaptation module (LoRA) to quickly capture key semantic features of the travel domain with minimal resource overhead. The second stage, while retaining the LoRA, fine-tunes the LoRA parameters of the model's higher layers (the last K layers) and introduces a top-down layer learning rate decay mechanism. This enhances the model's ability to recognize complex slots (such as fuzzy dates, city pairs, and cabin class terminology) while effectively maintaining training stability. This approach achieves high-precision understanding and structured extraction of travel semantic structures while significantly reducing memory usage and overfitting risk. It combines technological innovation with engineering feasibility, making it particularly suitable for efficient fine-tuning of large-scale models in vertical domains under resource constraints.

[0182] Corresponding to the aforementioned application function implementation method embodiments, this application also provides a travel cabin extraction device, electronic device, and corresponding embodiments.

[0183] Figure 7 This is a schematic diagram of the travel cabin extraction device shown in an embodiment of this application. See also... Figure 7 The travel slot extraction device 700 can be applied to electronic devices such as smartphones, tablets, and laptops. Specifically, the travel slot extraction device 700 may include an acquisition module 701, a first processing module 702, a second processing module 703, a verification module 704, and a generation module 705, as detailed below: The system comprises: an acquisition module 701 for acquiring initial travel information input by the user; a first processing module 702 for concurrently calling a slot extraction model and a slot existence discrimination model, extracting slot values ​​from the initial travel information according to the target slot value sequence based on the slot extraction model, and obtaining slot value extraction results; a second processing module 703 for determining the existence probability of each preset slot tag from the initial travel information based on the slot existence discrimination model, and obtaining slot tag detection results, wherein the existence probability of valid slot tags included in the slot tag detection results is greater than or equal to a preset probability threshold; a verification module 704 for performing semantic and logical verification on each slot value in the slot value extraction results based on the initial travel information if all slot tags corresponding to the slot values ​​included in the slot value extraction results are valid slot tags, and each valid slot tag in the slot tag detection results is covered by slot tags; and a generation module 705 for generating target travel information based on the slot value extraction results if the slot value verification results indicate that all slot values ​​in the slot value extraction results are normal.

[0184] In some embodiments, the travel slot extraction device 700 may further include a model training module. Before the step of parallel invocation of the slot extraction model and the slot existence discrimination model based on the first processing module 702, the model training module may be used to: construct a preset model output structure and obtain an initial slot value dataset based on the preset model output structure, wherein the initial slot value dataset includes real user data, and / or preset model prompt word generation data, and / or code rule generation data; perform quality scoring on each slot value data in the initial slot value dataset, and filter the initial slot value dataset based on the scoring results; perform deduplication processing on the filtered dataset based on semantic similarity; perform data augmentation processing on the deduplicated dataset to obtain a target slot value dataset; and train the model based on the target slot value dataset to obtain the slot extraction model and the slot existence discrimination model.

[0185] Optionally, the model training module may include a first model training submodule, which can be used to: take the target slot value dataset as the first input data and a preset slot value sequence corresponding to the target slot value dataset as the first supervision data to form a first training sample pair of the first input data and the first supervision data; freeze the backbone model parameters in the first model and perform the first fine-tuning of the first model based on the first target parameters to obtain a first fine-tuned model, wherein the first target parameters include the low-rank matrix parameters and slot label embedding parameters corresponding to each layer of the backbone model parameters; and fine-tune the backbone model located at the bottom preset number of layers in the first fine-tuned model. The model parameters are unfrozen, while the parameters of the backbone model (except for the preset number of layers) remain frozen. The first fine-tuning model is then fine-tuned a second time based on the low-rank matrix parameters corresponding to the preset number of backbone model parameters, resulting in the second fine-tuning model. The first training sample is input into the second fine-tuning model for training, and the output target slot value sequence corresponding to the preset slot value sequence is used as the first output data. First verification data of a first preset proportion is selected from the target slot value dataset and input into the trained second fine-tuning model, outputting the slot value sequence prediction result. The trained second fine-tuning model is then optimized based on the slot value sequence prediction result to obtain the slot extraction model.

[0186] Optionally, the model training module may further include a second model training submodule, which can be used to: take the target slot value dataset as the second input data, and the binary classification labels corresponding to the target slot value dataset generated based on the target slot value sequence as the second supervision data, forming a second training sample pair of the second input data and the second supervision data, and input the second training sample pair into the second model; after the second training sample pair is processed sequentially by the input layer, feature extraction layer and scoring layer in the second model, the output is a dimension vector matching the preset number of slot labels, and the dimension vector is mapped by a function to obtain the second output data corresponding to the existence probability of each preset slot label; select a second preset proportion of second verification data from the target slot value dataset, and input the second verification data into the trained second model, outputting the existence probability prediction value of each preset slot label; optimize the trained second model based on the existence probability prediction value to obtain a slot existence discrimination model.

[0187] In this embodiment, the travel slot extraction device 700 improves the efficiency of travel slot extraction by using a parallel call to the slot extraction model and the slot existence discrimination model, and improves the accuracy of slot extraction by judging the consistency between the slot value extraction result and the slot label detection result.

[0188] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0189] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.

[0190] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for extracting travel slots, characterized in that, include: Obtain the initial travel information input by the user; The slot extraction model and the slot existence determination model are invoked in parallel. Based on the slot extraction model, the initial travel information is extracted according to the target slot value sequence to obtain the slot value extraction result. Based on the slot existence discrimination model, the existence probability of each preset slot tag is determined from the initial travel information to obtain the slot tag detection result, wherein the existence probability of the valid slot tags included in the slot tag detection result is greater than or equal to the preset probability threshold. If all the slot labels corresponding to the slot values ​​included in the slot value extraction result are valid slot labels, and each valid slot label in the slot label detection result is covered by the slot label, then semantic and logical verification is performed on each slot value in the slot value extraction result based on the initial travel information. If the slot value verification result indicates that all the slot values ​​in the slot value extraction result are normal, then the target travel information is generated based on the slot value extraction result.

2. The travel cabin extraction method according to claim 1, characterized in that, The method further includes: If the slot value extraction result includes invalid slot labels corresponding to the slot values, and / or the slot labels do not cover all the valid slot labels, then the initial travel information is re-extracted based on the preset model using prompt word engineering to obtain the first corrected slot value extraction result. Wherein, the first slot labels corresponding to the first slot values ​​included in the first corrected slot value extraction result are all valid slot labels, and each valid slot label in the slot label detection result is covered by the first slot label, and the probability of the existence of the invalid slot label is less than the preset probability threshold. Based on the initial travel information, semantic and logical verification is performed on each of the first slot values ​​in the first modified slot value extraction result.

3. The travel cabin extraction method according to claim 1, characterized in that, The method further includes: If the slot value verification result indicates that there is an abnormal slot value in the slot value extraction result, the initial travel information is re-extracted based on the preset model using prompt word engineering to obtain a second corrected slot value extraction result, wherein the second slot value included in the second corrected slot value extraction result matches the initial travel information. The target travel information is generated based on the extraction result of the second corrected slot value.

4. The travel cabin extraction method according to claim 1, characterized in that, Before the parallel invocation of the slot extraction model and the slot existence determination model, the method further includes: Construct a preset model output structure, and obtain an initial slot value dataset based on the preset model output structure, wherein the initial slot value dataset includes real user data, and / or preset model prompt word generation data, and / or code rule generation data; The initial slot value dataset is scored for the quality of each slot value, and the initial slot value dataset is filtered based on the scoring results. Deduplication is performed on the filtered dataset based on semantic similarity. Data augmentation is performed on the deduplicated dataset to obtain the target slot value dataset. The model is trained based on the target slot value dataset to obtain the slot extraction model and the slot existence discrimination model.

5. The travel cabin extraction method according to claim 4, characterized in that, The process of training the model based on the target slot value dataset to obtain the slot extraction model includes: The target slot value dataset is used as the first input data, and the preset slot value sequence corresponding to the target slot value dataset is used as the first supervision data to form a first training sample pair of the first input data and the first supervision data. The backbone model parameters in the first model are frozen, and the first model is fine-tuned for the first time based on the first target parameters to obtain the first fine-tuned model. The first target parameters include the low-rank matrix parameters and slot label embedding parameters corresponding to each layer of the backbone model parameters. Unfreeze the backbone model parameters located at the bottom preset number of layers in the first fine-tuning model, keep the backbone model parameters other than the preset number of layers frozen, and perform a second fine-tuning on the first fine-tuning model based on the low-rank matrix parameters corresponding to the backbone model parameters of the preset number of layers to obtain the second fine-tuning model; The first training sample is input into the second fine-tuning model for training, and the target slot value sequence corresponding to the preset slot value sequence is output as the first output data. Select a first set of verification data from the target slot value dataset, and input the first verification data into the trained second fine-tuning model to output the slot value sequence prediction result; The second fine-tuning model after training is optimized based on the predicted slot value sequence to obtain the slot extraction model.

6. The travel cabin extraction method according to claim 4, characterized in that, The process of training the model based on the target slot value dataset to obtain the slot existence discrimination model includes: The target slot value dataset is used as the second input data, and the binary classification labels corresponding to the target slot value dataset generated based on the target slot value sequence are used as the second supervision data to form a second training sample pair of the second input data and the second supervision data, and the second training sample pair is input into the second model; After the second training sample pair is processed sequentially through the input layer, feature extraction layer and scoring layer in the second model, the output is a dimension vector that matches the number of preset slot labels. The dimension vector is then mapped by a function to obtain the second output data corresponding to the existence probability of each preset slot label. Select a second verification data with a second preset ratio from the target slot value dataset, and input the second verification data into the trained second model to output the existence probability prediction value of each preset slot label; Based on the predicted existence probability, the trained second model is optimized to obtain the slot existence discrimination model.

7. A travel cabin space extraction device, characterized in that, include: The acquisition module is used to acquire the initial travel information input by the user; The first processing module is used to call the slot extraction model and the slot existence discrimination model in parallel, and extract slot values ​​from the initial travel information according to the target slot value sequence based on the slot extraction model to obtain the slot value extraction result. The second processing module is used to determine the existence probability of each preset slot tag from the initial travel information based on the slot existence discrimination model, and obtain the slot tag detection result, wherein the existence probability of the valid slot tags included in the slot tag detection result is greater than or equal to the preset probability threshold. The verification module is used to perform semantic and logical verification on each slot value in the slot value extraction result based on the initial travel information if all the slot labels corresponding to the slot values ​​included in the slot value extraction result are valid slot labels, and each valid slot label in the slot label detection result is covered by the slot label. The generation module is used to generate target travel information based on the slot value extraction result if the slot value verification result indicates that all the slot values ​​in the slot value extraction result are normal.

8. The travel cabin extraction device according to claim 7, characterized in that, The verification module can also be used to re-extract slot values ​​from the initial travel information based on the preset model using prompt word engineering if the slot value extraction result includes invalid slot labels and / or the slot labels do not cover all valid slot labels, thereby obtaining a first corrected slot value extraction result; wherein, the first slot labels corresponding to the first slot values ​​included in the first corrected slot value extraction result are all valid slot labels, and each valid slot label in the slot label detection result is covered by the first slot label, and the probability of the existence of invalid slot labels is less than the preset probability threshold; and semantic and logical verification is performed on each first slot value in the first corrected slot value extraction result based on the initial travel information.

9. A computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes computer instructions that, when executed by a processor, implement the method described in any one of claims 1 to 6.

Images