Travel service and system based on natural language intent recognition

Abstract

This application belongs to the field of travel services and discloses a travel service and system based on natural language intent recognition. The travel service method includes: receiving user travel requests and converting them into natural language text; extracting a domain dictionary and user historical behavior data using a pre-built domain knowledge base and historical request data; using a large language model for intent recognition and slot filling to determine the user intent type and slot information; when the confidence level of the user intent type is greater than a preset threshold and the slot information meets preset missing conditions, generating query information and performing a fare search accordingly, and finally returning the result. This solution automates the processing from natural language requests to fare searches, effectively integrates domain knowledge and user behavior data to improve the accuracy of intent recognition, reduces invalid searches through confidence level and missing condition judgments, and enhances the ability to process complex queries by combining a large language model, thereby significantly improving the efficiency and accuracy of fare searches in travel services and optimizing the user experience.

Description

Technical Field

[0001] This application belongs to the field of travel services, and specifically relates to a travel service, system, device and medium based on natural language intent recognition. Background Technology

[0002] With the continuous development of the global economy and the improvement of people's living standards, tourism and business travel have become increasingly complex. People's requirements for flight searches are no longer limited to simple price comparisons and flight queries. Users expect more personalized and accurate search results, as well as a more efficient and convenient search experience.

[0003] The current pain points of traditional freight rate search are: (1) Users need to learn complex search rules and have rigid input formats. (2) The system cannot understand users' colloquial and diverse expressions. (3) The system cannot handle vague, incomplete or queries with multiple constraints. (4) Search results do not meet the requirements, and users make repeated requests, which increases the pressure on the backend system. The above problems lead to low search efficiency, poor user experience, inability to meet personalized needs, waste of resources in the backend system, and the need to do a lot of useless work. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a travel service, system, device, and medium based on natural language intent recognition. This solution significantly improves the efficiency and accuracy of fare searches in travel services and optimizes the user experience.

[0005] The first aspect of this disclosure proposes a travel service based on natural language intent recognition, the travel service method comprising: Receive user travel requests and convert the user travel requests into natural language text; Obtain domain dictionaries and user history behavior data from a pre-built domain knowledge base and historical request data; Based on the domain dictionary and user history behavior data, the natural language text is subjected to intent recognition and slot filling through a pre-built large language model to determine the user intent type and slot information. When the confidence level of the user intent type is greater than a preset threshold and the slot information meets the preset missing conditions, query information is generated based on the slot information. Based on the query information, a freight rate search is performed, and the freight rate search results are returned.

[0006] According to a preferred embodiment of this disclosure, before performing intent recognition on the natural language text using the domain dictionary and user historical behavior data, the travel service method further includes: The natural language text is segmented and domain entity recognition is performed to determine the keywords of the natural language text.

[0007] According to a preferred embodiment of this disclosure, the step of determining user intent type and slot information by performing intent recognition and slot filling on the natural language text using a pre-built large language model based on the domain dictionary and user historical behavior data includes: Feature data in the natural language text is extracted using the domain dictionary and the user's historical behavior data, and then converted into vector feature data; Based on the intent classification model in the pre-built first large language model, the vector feature data is used to identify intent and determine the user intent type; Based on the slot filling model in the first large language model, the vector feature data is filled with slots according to the user intent type to generate slot information.

[0008] According to a preferred embodiment of this disclosure, the step of determining user intent type and slot information by performing intent recognition and slot filling on the natural language text using a pre-built large language model based on the domain dictionary and user historical behavior data includes: Feature data in the natural language text is extracted using the domain dictionary and the user's historical behavior data, and then converted into vector feature data; The vector feature data is input into a pre-built second language model for intent recognition and slot filling to generate user intent type and slot information.

[0009] According to a preferred embodiment of this disclosure, the travel service method further includes: When the confidence level of the user intent type is less than or equal to a preset threshold, candidate intent options are generated. Interact with the user through the candidate intent options, and update the user's travel request based on the interaction information provided by the user; The updated user travel request is used to re-identify the user's intent type, and it is determined whether the confidence level of the user's intent type is greater than a preset threshold.

[0010] According to a preferred embodiment of this disclosure, the travel service method further includes: When any slot in the slot information is missing information, the slot information does not meet the preset missing conditions, and a slot question is generated based on the context information in the user's travel request. The system interacts with users by addressing the slot issue and updates the user's travel request based on the user's feedback. The updated user travel request is used to fill the slots again, the slot information is obtained again, and it is determined whether the slot information meets the preset missing conditions.

[0011] According to a preferred embodiment of this disclosure, generating query information based on the slot information includes: Determine the backend format requirements for the fare search system; Based on the slot information, generate structured query information that conforms to the format requirements of the backend.

[0012] To address the aforementioned technical problems, a second aspect of this disclosure proposes a travel service system based on natural language intent recognition, the travel service system comprising: The user interaction module is used to receive user travel requests and convert the user travel requests into natural language text. The data acquisition module is used to acquire domain dictionaries and user history behavior data from a pre-built domain knowledge base and historical request data; The large model processing module is used to perform intent recognition and slot filling on the natural language text based on the domain dictionary and user historical behavior data through a pre-built large language model, and to determine the user intent type and slot information. The query information generation module is used to generate query information based on the slot information when the confidence level of the user intent type is greater than a preset threshold and the slot information meets the preset missing conditions. The backend service module is used to perform a fare search based on the query information and return the fare search results.

[0013] To address the aforementioned technical problems, a third aspect of this disclosure provides an electronic device, comprising: Processor; and A memory storing computer-executable instructions, which, when executed, cause the processor to perform the method described in any of the above embodiments.

[0014] To address the aforementioned technical problems, a fourth aspect of this disclosure provides a computer storage medium that stores one or more programs, which, when executed by a processor, implement the method described in any of the above embodiments.

[0015] Compared with existing technologies, this application has the following advantages: By receiving user travel requests and converting them into natural language text, it extracts domain dictionaries and user historical behavior data using a pre-built domain knowledge base and historical request data. It then uses a large language model for intent recognition and slot filling to determine the user intent type and slot information. When the confidence level of the user intent type is greater than a preset threshold and the slot information meets preset missing conditions, it generates query information and performs a fare search accordingly, finally returning the results. This solution automates the process from natural language requests to fare searches, effectively integrates domain knowledge and user behavior data to improve the accuracy of intent recognition, reduces invalid searches through confidence level and missing condition judgments, and enhances the ability to process complex queries by combining a large language model. This significantly improves the efficiency and accuracy of fare searches in travel services and optimizes the user experience.

[0016] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of a travel service process based on natural language intent recognition according to an embodiment of the present disclosure is shown; Figure 2 A schematic diagram of the architecture of a travel service implementation based on natural language intent recognition according to an embodiment of the present disclosure is shown; Figure 3 A second schematic diagram of a travel service process based on natural language intent recognition according to an embodiment of the present disclosure is shown; Figure 4 A third schematic diagram of a travel service process based on natural language intent recognition according to an embodiment of the present disclosure is shown; Figure 5 A fourth schematic diagram of a travel service process based on natural language intent recognition according to an embodiment of the present disclosure is shown; Figure 6 A fifth schematic diagram of a travel service process based on natural language intent recognition according to an embodiment of the present disclosure is shown; Figure 7A sixth schematic diagram of a travel service process based on natural language intent recognition according to an embodiment of the present disclosure is shown; Figure 8 A user interaction flowchart for intent recognition and slot filling according to an embodiment of the present disclosure is shown; Figure 9 A UML sequence diagram of fare search including user intent recognition is shown according to an embodiment of the present disclosure; Figure 10 A schematic diagram of the structure of a natural language intent recognition system for travel services according to an embodiment of the present disclosure is shown; Figure 11 A schematic diagram of an electronic device structure according to an embodiment of the present disclosure is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The same reference numerals in the accompanying drawings denote the same or similar elements, components, or parts, and therefore repeated descriptions of the same or similar elements, components, or parts may be omitted below. It should also be understood that although qualifiers such as first, second, third, etc., indicating numbers may be used herein to describe various devices, elements, components, or parts, these devices, elements, components, or parts should not be limited by these qualifiers. That is, these qualifiers are only used to distinguish one from another. For example, a first device may also be referred to as a second device, without departing from the essence of the technical solution of this disclosure. Furthermore, the terms "and / or" and "and / or" refer to all combinations including the first or more of the listed items.

[0021] Please see Figure 1 , Figure 1 This is one of the schematic diagrams of a travel service process based on natural language intent recognition provided in this disclosure, such as... Figure 1 As shown, the travel service methods include: S11. Receive the user's travel request and convert the user's travel request into natural language text.

[0022] In this embodiment, the system receives travel requests submitted by users in the form of voice, text, or structured forms by integrating multimodal interaction interfaces (such as a speech recognition module and a text input parser). It then utilizes semantic understanding, context analysis, and standardization conversion algorithms from Natural Language Processing (NLP) technology to transform the unstructured original requests into grammatically correct and semantically clear natural language text expressions, while preserving the integrity of key travel elements. This solution overcomes the limitations of user input methods and supports diverse interaction scenarios. Semantic standardization significantly improves the system's ability to understand colloquial and ambiguous expressions, reducing the error rate in intent recognition caused by differences in expression. This provides a high-quality data foundation for subsequent accurate intent classification and slot filling, thereby optimizing the overall processing efficiency and user experience of travel service requests.

[0023] In this embodiment, users express their travel requests via multiple entry points (including websites, apps, mini-programs, etc.) in the form of text or voice. These entry points convert the requests into natural language text.

[0024] S12. Obtain the domain dictionary and user history behavior data from the pre-built domain knowledge base and historical request data.

[0025] In this embodiment, a domain knowledge base covering travel-related terminology, business rules, and scenario logic is constructed. This base integrates historical user travel request data, and data extraction and cleaning techniques are used to extract a structured domain dictionary (such as entities and relationships like cities, flights, and cabin classes) from the knowledge base. Simultaneously, data analysis and mining techniques are employed to extract user behavior patterns (such as frequently used routes, preferred cabin classes, and query times) from historical request data, forming a feature dataset that can be used by subsequent models. The introduction of the domain dictionary significantly improves the system's ability to understand travel-related terminology and contextual semantics, reducing ambiguous resolution. The fusion of historical user behavior data enables the system to capture personalized demand patterns, providing context-dependent personalized prior knowledge for intent recognition and slot filling. This improves resolution accuracy while enhancing the precision of service recommendations and user stickiness.

[0026] S13. Based on the domain dictionary and user historical behavior data, perform intent recognition and slot filling on natural language text through a pre-built large language model to determine the user intent type and slot information.

[0027] In this embodiment, a pre-built domain dictionary (covering travel-related terminology, entity relationships, etc.) and user historical behavior data (such as frequently used routes, preferred cabin class, and other personalized features) are used as input. Through the semantic understanding capabilities of a large language model (such as a pre-trained model based on the Transformer architecture), the converted natural language text undergoes multi-dimensional analysis. Combining domain knowledge constraints and user behavior patterns, intent classification (such as "flight search" and "hotel booking") and slot extraction (such as departure location, date, and cabin class) are completed simultaneously. Finally, structured user intent types and corresponding slot information are output. The introduction of the domain dictionary significantly improves the model's accuracy in recognizing travel-related terminology, avoiding ambiguous interpretations. The fusion of user historical behavior data enables the model to perform contextual reasoning based on personalized preferences, enhancing its ability to handle complex queries (such as fuzzy time and implicit needs). The end-to-end learning mechanism of the large language model achieves joint optimization of intent recognition and slot filling, significantly reducing the risk of error accumulation in traditional phased models, thereby improving overall parsing efficiency and accuracy, and providing a reliable data foundation for subsequent personalized services.

[0028] In this embodiment, natural language text is segmented and domain entity recognition is performed to determine keywords. Rule-based or statistical model-based segmentation algorithms (such as maximum matching or Hidden Markov Models) are used to segment the natural language text into semantically or grammatically meaningful lexical units. Simultaneously, combined with an entity dictionary in a domain knowledge base (covering travel-related entities such as cities, flights, and hotels), Named Entity Recognition (NER) technology (such as BiLSTM-CRF model) is used to annotate domain entities in the text. Finally, high-frequency, core, or business-meaning words are extracted as keywords from the segmentation results and entity annotations. Segmentation transforms continuous text into discrete lexical units, providing a basic structure for subsequent analysis; domain entity recognition enhances the understanding of professional terminology and business scenarios by accurately annotating travel-related entities; the keywords extracted by combining these two methods effectively filter irrelevant information (such as function words and stop words), focusing on core user needs (such as departure point, destination, and time), significantly improving the accuracy and efficiency of intent recognition and slot filling, while reducing the computational complexity of subsequent model processing.

[0029] S14. When the confidence level of the user intent type is greater than the preset threshold and the slot information meets the preset missing conditions, generate query information based on the slot information.

[0030] In this embodiment, the confidence level of user intent types output by the large language model is filtered by a preset threshold, retaining only high-confidence intents. Simultaneously, preset missing conditions for slot information are defined (e.g., required slots are empty or optional slots are missing but can be inferred). When the intent confidence level meets the threshold and the slot status meets the conditions, a rule engine or template matching technology, combined with default values ​​from the domain knowledge base, historical behavior data, or contextual information, is used to intelligently fill in the missing slots or generate a query framework to be completed, ultimately outputting structured query information. This solution filters low-quality intents through confidence thresholds, reducing invalid searches; the combination of preset missing conditions and intelligent completion mechanisms avoids query failures due to incomplete slots and improves interaction efficiency by minimizing user input; the generated query information directly matches the requirements of the fare search interface, significantly shortening the link from intent parsing to result return, while maintaining query accuracy and personalization, ultimately optimizing user experience and system resource utilization.

[0031] In this embodiment, the specific generation of query information can be achieved by determining the backend format requirements of the fare search system; structured query information conforming to the backend format requirements is generated based on the slot information. According to the data interface specifications of the fare search system (such as field types, required fields, and format constraints), the fully filled slot information (such as city codes and date formats) is mapped and converted into structured query parameters (such as JSON or XML format) that conform to the backend requirements. This ensures seamless integration between the frontend request and the backend system, avoiding query failures due to format errors; structured data improves search efficiency, enabling accurate and rapid retrieval of fare information and optimizing the user's waiting experience.

[0032] S15. Perform a fare search based on the query information and return the fare search results.

[0033] In this embodiment, the structured query information generated in the preceding steps (including user intent type and complete / completed slot information, such as departure point, destination, date, cabin class, etc.) is used as input. By calling the fare calculation interface of airlines, ticketing platforms, or third-party GDS (Global Distribution System), and combining real-time inventory data with dynamic pricing rules, a multi-dimensional condition matching (such as lowest price, direct flight priority, time range, etc.) fare search algorithm is executed. Finally, the search results are filtered and sorted to select fare schemes that meet the user's needs, and returned to the user in a structured format (such as price, flight number, departure and arrival times, and refund / change rules). This solution achieves end-to-end closed-loop processing from the user's natural language request to accurate fare results. Real-time interface calls ensure the timeliness of price and inventory data, avoiding the "no tickets available" problem caused by data delays. The multi-dimensional search algorithm and personalized sorting mechanism (such as prioritizing direct flights based on user historical preferences) significantly improve the relevance of results and reduce user filtering costs. The structured return format supports rapid rendering and interaction (such as price calendars and cabin class comparisons), thereby optimizing user decision-making efficiency and the overall service experience.

[0034] like Figure 2 The diagram illustrates the architecture of a travel service based on natural language intent recognition. This solution adds an intent recognition engine. This engine is placed between the user interaction layer and the backend service layer. The engine primarily consists of an LLM module and a DSHOP-AGENT module. The LLM module is a large language model module, while the DSHOP AGENT module typically refers to a component used as a proxy or mediator in distributed systems or microservice architectures. They work together to convert natural language into structured language. Structured requests are used to call a traditional fare search engine, returning a large number of results. The DSHOP-AGENT then processes all the results before finally returning them to the requesting user.

[0035] Core Module Details: Module 1: LLM Module, a large language model supporting text cleaning, domain dictionary-enhanced analysis, intent classification, slot filling, multi-turn dialogue management, and other functions. It serves as the core of the language intent engine.

[0036] Module Two: The DSHOP-AGENT module acts as an "adapter" between the LLM module and the traditional fare search system. It provides the LLM model with a domain knowledge base (dictionary, rules), user historical behavior data, and auxiliary storage of contextual dialogues with the user. It transforms the identified content of each slot into a structured request conforming to the query interface format of the traditional fare search system and sends it to the system. It then processes and organizes the results returned by the traditional fare search system according to the user's specific requirements. Finally, it returns the results to the requesting user.

[0037] Module 3: Traditional Fare Search System. The traditional fare search system does not require major changes to its original functions.

[0038] In this embodiment, user travel requests are received and converted into natural language text. A pre-built domain knowledge base and historical request data are used to extract a domain dictionary and user historical behavior data. A large language model is then used for intent recognition and slot filling to determine the user intent type and slot information. When the confidence level of the user intent type is greater than a preset threshold and the slot information meets preset missing conditions, query information is generated and a fare search is performed accordingly, ultimately returning the result. This solution automates the process from natural language requests to fare searches, effectively integrates domain knowledge and user behavior data to improve the accuracy of intent recognition, reduces invalid searches through confidence level and missing condition judgments, and enhances the ability to process complex queries by combining a large language model. This significantly improves the efficiency and accuracy of fare searches in travel services and optimizes the user experience.

[0039] Please see Figure 3 , Figure 3 This is the second illustration of a travel service process based on natural language intent recognition provided in this public disclosure, such as... Figure 3 As shown, the travel service methods include: S21. Extract feature data from natural language text using domain dictionary and user history behavior data, and convert it into vector feature data.

[0040] In this embodiment, a pre-built domain dictionary (covering travel-related terminology, entity relationships, etc.) is used to perform keyword matching and semantic expansion on natural language text, extracting features such as domain-related entities and business attributes. Simultaneously, user historical behavior data (such as frequently used routes, preferred cabin classes, query times, etc.) is combined to mine personalized user characteristics. Text features and user behavior features are mapped into low-dimensional dense vectors using word embedding (Word2Vec, BERT, etc.) or graph embedding techniques, forming vector feature data that can be directly processed by machine learning models. The introduction of the domain dictionary significantly improves the accuracy of feature extraction and avoids interference from irrelevant information; the fusion of user historical behavior data allows vector features to contain personalized context, enhancing the model's ability to capture implicit user needs; vectorization transforms unstructured data into a unified format, reducing the complexity of subsequent model processing while preserving deep semantic and behavioral patterns, providing high-quality input for tasks such as intent recognition, slot filling, and fare recommendation, ultimately improving the overall service intelligence and user experience.

[0041] S22. Based on the intent classification model in the pre-built first language model, perform intent recognition on the vector feature data to determine the user intent type.

[0042] In this embodiment, the vector feature data generated in the preceding steps is input into the intent classification model within a pre-built first large language model (this model is typically based on the Transformer architecture, pre-trained and fine-tuned using massive corpora from the travel domain). The model utilizes its multi-layer self-attention mechanism to capture semantic relationships and contextual dependencies between vectors, and combines this with a Softmax classifier to perform multi-label or single-label classification of user intents, ultimately outputting high-confidence user intent types (such as flight searches, hotel bookings, car rental services, etc.). The intent classification model based on the large language model can deeply understand the semantics and behavioral patterns in vector features, significantly improving the accuracy of recognizing complex, ambiguous, or implicit intents. The pre-training and fine-tuning mechanism allows the model to quickly adapt to specific scenarios in the travel domain, reducing reliance on labeled data. The end-to-end vector processing avoids the separation between feature engineering and model training in traditional methods, reducing the risk of error accumulation, thus providing a reliable intent foundation for subsequent slot filling and personalized services, optimizing overall service response efficiency and user satisfaction.

[0043] S23. Based on the slot filling model in the first language model, fill the slots of the vector feature data according to the user intent type to generate slot information.

[0044] In this embodiment, the identified user intent type and corresponding vector feature data are input into the slot filling model in the first large language model (this model typically employs sequence labeling or a generative architecture, dynamically adjusting attention weights based on intent type). The model performs context-aware semantic parsing of the vector features, locates key information through conditional probability prediction or pointer networks, and fills the intent-related entities, attributes, and other elements in the vector into preset slots (such as departure location, date, and cabin class), ultimately generating structured slot information. The slot filling model based on the large language model can dynamically focus on the context in conjunction with user intent, significantly improving the filling accuracy of complex queries (such as implicit time and multi-entity associations); the generative architecture can handle out-of-vocabulary words and long-tail requirements, enhancing the model's generalization ability; the structured slot information directly matches subsequent service interfaces, reducing manual parsing costs, and providing data support for personalized recommendations (such as filling cabin class based on historical preferences), thereby optimizing service chain efficiency and user experience.

[0045] In this embodiment, a two-stage concatenated model of "intent recognition → slot filling" is used as the first major language model: The intent recognition model is based on a classifier of a major language model (such as BERT or RoBERTa), taking the vector features of the user request as input and outputting the intent type (such as flight search or hotel booking). The slot filling model is based on a conditional generation or sequence labeling model (such as BERT-CRF or T5), dynamically adjusting the filling strategy according to the intent type, and outputting structured slot information (such as departure location and date).

[0046] In this embodiment, specifically, the following solutions are included: Data collection, domain corpus: Collect travel domain dialogue data (such as flight ticket queries, hotel reservations), annotate intent types (such as flight_search) and slot labels (such as B-departure, I-date). User historical data: Integrate user historical query records (such as common departure locations, preferred cabin classes) for personalized feature enhancement. Domain dictionary: Construct a travel domain dictionary (such as city names, airline names) for feature expansion.

[0047] Data preprocessing, word segmentation and entity recognition: Use the domain dictionary to segment the text and annotate domain entities (such as Beijing → B-city). Vector encoding: Convert the text into vector features through models such as BERT (such as the [CLS] vector for intent classification and sub-word vectors for slot filling). Data augmentation: Perform synonym replacement and entity replacement on long-tail queries (such as "Shanghai" → "Hu") to generate more samples.

[0048] Intent recognition model, model selection: A classifier based on BERT (such as BERT-For-Sequence-Classification). Input: The vector features of the user request ([CLS] + sub-word vectors). Output: The probability distribution of intent types (such as flight_search: 0.95).

[0049] Slot filling model, model selection: Sequence labeling scheme: BERT-CRF (combining BERT features and CRF layer to constrain label transitions). Generative scheme: T5 or BART, directly generate slot sequences (such as departure location: Beijing date: 2024-01-01). Input: The original text vector (sub-word level). Intent type (converted into a vector through an embedding layer and concatenated with the text vector). Output: The slot label sequence (such as B-departure I-city O O B-date I-date).

[0050] Sequential logic, forward propagation: Input text → Intent recognition model → Output intent type. Concatenate the intent type embedding vector with the original text vector → Slot filling model → Output slot labels.

[0051] Model training and optimization, joint training: If resources are sufficient, end-to-end training can be performed (requiring shared BERT encoding layers). More commonly, staged training is used: first train the intent recognition model until convergence. Then, with the intent model parameters fixed, train the slot-filling model (input containing intent embeddings). Loss functions: Intent recognition: cross-entropy loss. Slot filling: CRF negative log-likelihood loss (or sequence cross-entropy loss). Optimization techniques: Domain adaptation: continue pre-training on a general BERT model. Few-shot learning: use cue learning or fine-tuning with a small amount of labeled data. Multi-task learning: share the BERT encoding layer and simultaneously optimize both intent and slot losses.

[0052] Please see Figure 4 , Figure 4 This is the third illustration of a travel service process based on natural language intent recognition provided in this public disclosure, such as... Figure 4 As shown, the travel service methods include: S31. Extract feature data from natural language text using domain dictionary and user history behavior data, and convert it into vector feature data.

[0053] In this embodiment, natural language text is segmented and entity-labeled based on a pre-built domain dictionary (covering professional terms, entities, and relationships in fields such as travel), extracting domain-related features (such as city names and business attributes). Simultaneously, personalized features are mined by combining user historical behavior data (such as frequently used query terms and preference categories). Text features and user behavior features are mapped into low-dimensional dense vectors using word embedding (Word2Vec, BERT) or graph embedding techniques, ultimately fusing them into unified vector feature data containing semantics and behavioral patterns. The introduction of the domain dictionary significantly improves the accuracy of feature extraction, filters irrelevant information, and strengthens domain semantic connections. The fusion of user historical behavior data imbues vector features with personalized context, enhancing the model's ability to capture implicit user needs. Vectorization unifies the processing format of unstructured data, reducing the complexity of subsequent model training while preserving deep semantics and behavioral patterns, providing high-quality input for tasks such as intent recognition and slot filling, ultimately improving service intelligence and user decision-making efficiency.

[0054] S32. Input the vector feature data into the pre-built second language model to perform intent recognition and slot filling, and generate user intent type and slot information.

[0055] In this embodiment, the vector feature data fused with semantics and user behavior patterns obtained from the preprocessing is input into a second large-scale language model (such as GPT, BERT variants, or a dedicated architecture) pre-trained on a large-scale domain corpus (such as travel and e-commerce dialogue data). Utilizing its multi-task learning capabilities or joint modeling mechanism, the model captures complex relationships between vectors through a self-attention mechanism, simultaneously completing intent classification (e.g., identifying the intent of "flight search") and slot sequence labeling (e.g., extracting entities such as "departure location - Beijing" and "date - 2024-10-01"). Finally, it outputs structured user intent types and slot information. Joint modeling based on a large language model fully leverages parameter sharing and context awareness capabilities, significantly improving the collaborative accuracy of intent recognition and slot filling (especially for fuzzy or long-tail queries). Pre-training and fine-tuning mechanisms enable the model to quickly adapt to vertical domains, reducing reliance on labeled data. End-to-end generation of structured results directly matches downstream service interfaces, reducing manual parsing costs and providing data support for personalized recommendations (e.g., filling in cabin class slots based on historical preferences), thereby optimizing service response efficiency and user experience.

[0056] In this embodiment, the solution uses a single Large Language Model (LLM) to simultaneously perform two tasks: intent classification and slot filling. Both tasks share underlying feature representations, thereby improving task relevance and recognition accuracy.

[0057] In this embodiment, a single model structure is specifically adopted: the first major language model uses a unified deep learning framework (such as the Transformer architecture) to handle both intent classification and slot filling tasks. The model input is the user's natural language text (such as "the cheapest flight from Beijing to Guangzhou next week"), and the output is a set of intent labels and slot key-value pairs. Shared underlying features: the intent classification and slot filling tasks share the context embedding vectors generated by the encoder, avoiding information fragmentation caused by independent modeling. For example, the model uses a self-attention mechanism to capture the semantic associations of words such as "Beijing," "Guangzhou," and "cheapest," providing global contextual support for both tasks.

[0058] In this embodiment, the training method employs multi-task learning: During the training phase, the model simultaneously optimizes the loss functions for intent classification and slot filling. For example: Intent classification loss: Cross-entropy loss function, used to predict intent labels (e.g., "query lowest price"). Slot filling loss: Conditional Random Field (CRF) or sequence labeling loss, used to predict slot sequences (e.g., {departure_city: "Beijing", arrival_city: "Guangzhou"}). Joint loss function: Weighted summation (e.g., Loss = α·Loss_intent + β·Loss_slot) balances the importance of the two tasks. Data augmentation and domain adaptation: Fine-tuning is performed using labeled data from the aviation travel domain (e.g., corpora containing professional terms such as "cabin class" and "round trip") to improve the model's ability to recognize domain terms. Simultaneously, data augmentation (e.g., synonym replacement, random entity masking) enhances the model's robustness.

[0059] In this embodiment, the inference process includes the following steps: Input preprocessing: The user query is segmented, corrected, and normalized to domain entities (e.g., "Capital Airport" is normalized to "PEK") to generate a text sequence that the model can process. Joint decoding: The model outputs the intent label and slot sequence in one go. For example: Intent classification result: "Query the lowest price" (confidence 0.95). Slot filling result: {departure_city: "Beijing", arrival_city: "Guangzhou", date: "2025-10-01", preference: "cheapest"}. Post-processing and clarification: If the model confidence is lower than the threshold (e.g., intent confidence < 0.8), a multi-turn dialogue is triggered to clarify missing or ambiguous slot information (e.g., "Which city do you want to depart from?"), and the user feedback is re-inputted to update the model's results.

[0060] Please see Figure 5 , Figure 5 This is the fourth illustration of a travel service process based on natural language intent recognition provided in this public disclosure, such as... Figure 5 As shown, the travel service methods include: S41. When the confidence level of the user intent type is less than or equal to the preset threshold, generate candidate intent options.

[0061] In this embodiment, during the intent recognition stage, after the model outputs the probability distribution of user intent types, it compares the highest confidence score with a preset threshold (e.g., 0.85). When the highest score is less than or equal to the threshold, the current intent recognition result is deemed unreliable. The model then selects the top N intents with higher confidence (e.g., Top-3) from the probability distribution as candidate options, or expands the candidate list from the intent tag library by combining semantic similarity calculations. Finally, the results are returned to the user or subsequent modules for secondary confirmation or interactive clarification. The threshold filtering mechanism effectively avoids the erroneous transmission of low-confidence intents, reducing service deviations caused by misjudgments in subsequent processes. Candidate intent options provide users or the system with multi-round interaction correction space, improving the robustness of parsing complex or fuzzy queries. The candidate generation strategy combined with confidence ranking optimizes the efficiency of manual review, while interactive clarification (e.g., "Do you want to check flight tickets or hotels?") enhances the user experience, ultimately achieving a dual improvement in intent recognition accuracy and service fault tolerance.

[0062] S42. Interact with the user through candidate intent options and update the user's travel request based on the interaction information provided by the user.

[0063] In this embodiment, after the system generates candidate intent options, it presents the options to the user through a multimodal interactive interface (such as voice prompts or pop-up selections). The user provides feedback on the selection result by clicking, confirming with voice, or inputting correction information. The system analyzes the user's feedback interaction information (such as selecting an intent or adding a new intent description), combines it with the original travel request text, dynamically updates the intent recognition result and context state, and re-triggers the slot filling or service routing process to ensure that subsequent processing aligns with the user's actual needs. The interaction clarification mechanism significantly improves the parsing accuracy of fuzzy or complex queries, avoiding service interruptions caused by intent misjudgment; dynamically updating the request context enhances the system's understanding of the coherence of multi-turn dialogues, supporting users to correct their needs midway (such as changing from "checking flight tickets" to "booking hotels"); the user's active participation in the intent confirmation process reduces the system's guessing cost, and at the same time, the model training data is optimized through real-time feedback (such as recording user correction behavior), forming a closed loop of "interaction-correction-learning" to continuously improve the level of service intelligence.

[0064] S43. Perform intent recognition again through the updated user travel request to obtain the user intent type again, and determine whether the confidence level of the user intent type is greater than the preset threshold.

[0065] In this embodiment, the updated travel request text (including a revised intent description or supplementary information) based on user interaction feedback is re-inputted into the intent recognition model for secondary inference, generating a new user intent type and corresponding confidence score. By comparing the new confidence score with a preset threshold (e.g., 0.9), the reliability of the current intent recognition result is determined. If it is still below the threshold, the next round of interaction clarification or manual processing may be triggered, forming a closed-loop optimization process of "request update - model inference - confidence verification". Secondary intent recognition uses user correction information to eliminate ambiguity or noise in the original request, significantly improving the parsing accuracy in complex scenarios. The dynamic confidence verification mechanism ensures that only highly reliable intents enter the subsequent process, avoiding service deviations caused by erroneous transmission. Closed-loop optimization gradually converges the user's true needs through multiple interactions, especially enhancing the fault tolerance of long-tail or fuzzy queries (e.g., "Book me a trip soon"), ultimately achieving a dual improvement in the accuracy of service response and the smoothness of user experience.

[0066] In this embodiment, if the confidence level of the user intent type is still less than the preset threshold, it means that the user intent has not been accurately identified. In this case, it is necessary to interact with the user again to improve the recognition accuracy.

[0067] Please see Figure 6 , Figure 6 This is the fifth illustration of a travel service process based on natural language intent recognition provided in this public disclosure, such as... Figure 6 As shown, the travel service methods include: S51. When any slot in the slot information is missing information, the slot information does not meet the preset missing conditions, and a slot question is generated based on the context information in the user's travel request.

[0068] In this embodiment, after completing intent recognition and slot filling, the system automatically verifies the completeness of slot information. If any required slot (such as departure location or date) is found to be missing information, the current slot status is determined to be inconsistent with the preset completeness condition (i.e., the range of allowed missing slots). Subsequently, based on the context of the user's original travel request (such as already filled slots, historical query records, and domain common sense), a targeted question (such as "What is your departure city?") is generated, guiding the user to supplement the missing information through the interactive interface to complete the slot data. The context-aware slot question generation mechanism significantly improves interaction efficiency and avoids user confusion caused by generalized questions (such as "Please supplement all missing information"). Inference based on already filled slots (such as knowing "destination - Beijing" and inferring "departure location") reduces the user's input burden while improving the accuracy of information completion. Dynamic verification and interactive closed loop ensure the integrity of slot data, providing reliable input for subsequent services (such as flight ticket queries), ultimately improving task completion rate and user satisfaction.

[0069] S52. Interact with users through slot-related questions and update user travel requests based on user feedback.

[0070] In this embodiment, based on the detected missing slots, a targeted question (such as "What is your departure date?") is generated and displayed to the user through an interactive interface. After the user provides feedback via voice, text, or clicking options, the system parses the interaction content, extracts valid data, fills it into the corresponding slot, and updates the context of the original travel request, forming a new request text containing complete slot information, providing an accurate basis for subsequent service processing. The interactive slot completion mechanism eliminates information ambiguity by proactively asking questions, significantly improving the parsing success rate of complex queries; dynamically updating the request context ensures the coherence of multi-turn dialogues, supporting users to gradually clarify their needs (such as selecting a destination first and then adding a date); user participation in the information correction process reduces the system's guessing cost, and the interaction data can be used to optimize the model (such as recording common missing slot types), forming a closed loop of "question-feedback-optimization," ultimately improving the accuracy of service response and the smoothness of user experience.

[0071] S53. Fill the slots again with the updated user travel requests, obtain the slot information again, and determine whether the slot information meets the preset missing conditions.

[0072] In this embodiment, based on the updated complete travel request text updated by user interaction feedback, the system re-invokes the slot filling model (such as sequence labeling or generative models) for secondary parsing to extract updated intent-related slot information (such as departure location, date, cabin class, etc.). Then, the system verifies the completeness of the new slot information using preset rules (such as non-empty required slots and valid data format). If all required slots are filled and the data is valid, the system is deemed to meet the missing data condition; otherwise, the next round of interaction clarification or exception handling is triggered. Secondary slot filling utilizes the user-corrected context information to eliminate ambiguity or omissions from the initial parsing, significantly improving the accuracy of slot data. The dynamic verification mechanism ensures that only complete and valid slot information enters subsequent service stages (such as flight booking), avoiding service failures due to missing information. Closed-loop optimization gradually converges the user's true needs through multiple interactions, especially enhancing the fault tolerance for complex or long-tail queries (such as "Book me a business class flight from Shanghai to Beijing next week"), ultimately achieving a dual improvement in service response reliability and user experience smoothness.

[0073] In combination with the above, as follows Figure 3 , 4 The embodiments shown in 5 and 6 are specifically provided in this solution as follows: (1) Preprocessing: This includes text segmentation, error correction, and domain entity recognition (e.g., normalizing "Capital Airport" to "PEK"). Contextual information, including user search history and the context of the current session (e.g., the results of the previous query), is used for disambiguation and completion.

[0074] (2) Model inference path: The system can select the joint model or serial model path according to the actual deployed model.

[0075] Joint model: This is a cutting-edge technology that uses a single model to complete both intent classification and slot filling tasks simultaneously. Both share underlying features, which can better capture the correlation between tasks and usually has higher accuracy.

[0076] Serial model: This is a classic method that first classifies the intent, and then uses the results as features to input into the slot filling model. It has a clear structure, but may suffer from error accumulation.

[0077] (3) Analysis of results during the process: • The intent classification result is a specific intent label (e.g., lowest price cheapest), and multiple intent labels can be preset, such as "full price mode", "low price mode", etc.

[0078] • Slot filling result: A set of key-value pairs (Slots), for example: {departure_city: "Beijing",arrival_city: "Shanghai", date: "2025-10-01", preference: "cheapest"}.

[0079] like Figure 7 As shown in Figure 6, a schematic diagram of a travel service process based on natural language intent recognition is presented. The process starts from the "start" node, first preprocessing the input data and converting it into a feature vector.

[0080] Model selection and preliminary processing: Determine whether to use a joint model or a single model. If different models are selected, processing will be performed separately using the intent classification model or the slot filling model; if a joint model is used, both the intent classification model and the slot filling model will be applied simultaneously. The model output is then parsed to obtain the intent classification results and slot filling results, and intent weight values ​​are calculated and slot information is checked.

[0081] Confidence Assessment and Subsequent Operations: The confidence level of the intent type is assessed. If the confidence threshold is not met, further checks are performed to determine if there are errors in slot filling. If there are errors, a structured query request is output, and the traditional fare search system is invoked before the process ends. If there are no errors in slot filling, the process directly proceeds to the interactive clarification module. If the confidence threshold is met, structured query information conforming to the format requirements of the fare search system's backend is generated.

[0082] End: The process has finally come to a close.

[0083] like Figure 8 As shown, the user interaction flowchart for intent recognition and slot filling in the above scheme is as follows: 1. Receiving and Analysis: Receive low-confidence results or missing slot information from the intent recognition module, and then analyze the root cause of the problem. 2. Problem Classification and Strategy Formulation: Classify problems into three categories: ambiguous intent and low confidence, missing key slots, and conflicting or ambiguous slot values, and formulate corresponding strategies. For ambiguous intent and low confidence, the strategy is to confirm the intent and generate candidate intent options; when key slots are missing, the strategy is to ask for missing information and generate questions based on the context; if there is a conflicting or ambiguous slot value, the strategy is to clarify the ambiguity and ask for conflicting values. 3. Question Generation and Return: Generate a natural language clarifying question, such as asking the user whether they want to check the lowest price or a specific flight, and then return the question to the user interface. 4. User Response Processing: Analyze whether the user response contains valid information. If the context is present, update the context information (e.g., supplement or modify the slot, confirm the intent), re-enter the updated information into the intent recognition module, and check again whether the intent and slot are clear. If clear, continue generating the final structured query; if unclear, return to the more advanced steps of problem analysis. If the user response is not valid, execute the default strategy or end the conversation.

[0084] To clearly illustrate the key processing flow of intent recognition, the following example is provided: Example 1 (Simple query): User input: "Check the cheapest flight from Beijing to Guangzhou next week".

[0085] Example 2 (Complex Query): User input: "My family and I are departing from Shanghai during the National Day holiday and want to go to an island. Our budget is under 5,000 yuan per person. Direct flights would be preferred."

[0086] Example 3 (Typical Interaction Example) Initial user input: "I want to book cheap tickets to the beach around National Day."

[0087]

[0088] like Figure 9 As shown, a UML sequence diagram for fare search including user intent recognition is presented: Step 1: Users express their travel requests through multiple entry points (including websites, apps, mini-programs, etc.) in the form of text or voice.

[0089] Step 2: Multiple entry points (including websites, apps, mini-programs, etc.) convert the request into natural language text to call DSHOP-AGENT.

[0090] Step 3: DSHOP-AGENT calls the LLM model to perform intent analysis.

[0091] Step 4: The LLM model performs text cleaning and identifies keywords.

[0092] Step 5: The LLM model obtains domain dictionary and user habits information from the DSHOP-AGENT to improve the accuracy of intent recognition.

[0093] Step 6: DSHOP-AGENT returns the requested information to the LLM model.

[0094] Step 7: The LLM model performs initial intent recognition.

[0095] Step 8: For cases where the semantic analysis is unclear, conduct a second interaction and return a question to the user (e.g., Which city would you like to depart from?). Step 9: The user answers the confirmation question that is returned.

[0096] Step 10: Perform intent recognition a second time. If recognition is successful, fill the slot.

[0097] Step 11: Return the data according to the agreed intent type and the corresponding slot.

[0098] Step 12: DSHOP-AGENT generates a formatted request that conforms to the format requirements of the domestic fare search system backend.

[0099] Step 13: DSHOP-AGENT initiates a fare search request to the backend of the domestic fare search system.

[0100] Step 14: The domestic freight rate search system returns the search results to DSHOP-AGENT.

[0101] Step 15: DSHOP-AGENT integrates the results returned from the backend according to the user's requirements.

[0102] Step 16: Return to the final result.

[0103] Please see Figure 10 , Figure 10 This disclosure provides a natural language intent recognition system for travel services. The recognition system includes: a user interaction module 11, a data acquisition module 12, a large model processing module 13, a query information generation module 14, and a backend service module 15.

[0104] In this embodiment, the user interaction module 11 is used to receive user travel requests and convert the user travel requests into natural language text.

[0105] In this embodiment, the data acquisition module 12 is used to acquire domain dictionary and user historical behavior data from a pre-built domain knowledge base and historical request data.

[0106] In this embodiment, the large model processing module 13 is used to perform intent recognition and slot filling on natural language text based on the domain dictionary and user historical behavior data, and to determine the user intent type and slot information.

[0107] In this embodiment, the query information generation module 14 is used to generate query information based on slot information when the confidence level of the user intent type is greater than a preset threshold and the slot information meets the preset missing conditions.

[0108] In this embodiment, the backend service module 15 is used to perform a fare search based on the query information and return the fare search results.

[0109] In this embodiment, the recognition system further includes a data preprocessing module, used to perform word segmentation and domain entity recognition on natural language text to determine the keywords of natural language text.

[0110] In this embodiment, the large model processing module 13 is specifically used to extract feature data from natural language text through domain dictionary and user historical behavior data, and convert it into vector feature data; based on the intent classification model in the pre-built first large language model, it performs intent recognition on the vector feature data to determine the user intent type; based on the slot filling model in the first large language model, it fills the slots of the vector feature data according to the user intent type to generate slot information.

[0111] In this embodiment, the large model processing module 13 is specifically used to extract feature data from natural language text through domain dictionary and user historical behavior data, and convert it into vector feature data; input the vector feature data into the pre-built second large language model, perform intent recognition and slot filling, and generate user intent type and slot information.

[0112] In this embodiment, the identification system further includes: a secondary interaction module, specifically used to generate candidate intent options when the confidence level of the user intent type is less than or equal to a preset threshold; interact with the user through the candidate intent options, update the user's travel request based on the user's feedback interaction information; and perform intent identification again through the updated user travel request to obtain the user intent type again, and determine whether the confidence level of the user intent type is greater than the preset threshold.

[0113] In this embodiment, the secondary interaction module is specifically used to generate a slot question based on the context information in the user's travel request when any slot in the slot information is missing information and the slot information does not meet the preset missing conditions; to interact with the user through the slot question and update the user's travel request based on the user's feedback interaction information; to fill the slot again through the updated user travel request, obtain the slot information again, and determine whether the slot information meets the preset missing conditions.

[0114] In this embodiment, the query information generation module 14 is specifically used to determine the backend format requirements of the fare search system; and to generate structured query information that conforms to the backend format requirements based on the slot information.

[0115] like Figure 11 As shown, this embodiment of the present disclosure provides an electronic device, including a processor 1110, a communication interface 1120, a memory 1130, and a communication bus 1140, wherein the processor 1110, the communication interface 1120, and the memory 1130 communicate with each other through the communication bus 1140. Memory 1130 is used to store computer programs; When the processor 1110 executes the program stored in the memory 1130, it implements any of the above methods.

[0116] The electronic device provided in this embodiment of the present disclosure includes a processor 1110 that receives a user's travel request by executing a program stored in a memory 1130 and converts the user's travel request into natural language text; it obtains a domain dictionary and user historical behavior data from a pre-built domain knowledge base and historical request data; based on the domain dictionary and user historical behavior data, it performs intent recognition and slot filling on the natural language text using a pre-built large language model to determine the user's intent type and slot information; when the confidence level of the user's intent type is greater than a preset threshold and the slot information meets a preset missing condition, it generates query information based on the slot information; it performs a fare search based on the query information and returns the fare search results.

[0117] The communication bus 1140 mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, and a component bus, etc. For ease of illustration, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus.

[0118] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.

[0119] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1130 may also be at least one storage device located remotely from the aforementioned processor 1110.

[0120] The processor 1110 mentioned above can be a general-purpose processor 1110, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0121] This disclosure provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors 1110 to implement the methods of any of the above embodiments.

[0122] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this disclosure is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0123] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A travel service method based on natural language intent recognition, characterized in that, The travel service methods include: Receive user travel requests and convert the user travel requests into natural language text; Obtain domain dictionaries and user history behavior data from a pre-built domain knowledge base and historical request data; Based on the domain dictionary and user history behavior data, the natural language text is subjected to intent recognition and slot filling through a pre-built large language model to determine the user intent type and slot information. When the confidence level of the user intent type is greater than a preset threshold and the slot information meets the preset missing conditions, query information is generated based on the slot information. Based on the query information, a freight rate search is performed, and the freight rate search results are returned.

2. The travel service method according to claim 1, characterized in that, Before performing intent recognition on the natural language text using the domain dictionary and user historical behavior data, the travel service method further includes: The natural language text is segmented and domain entity recognition is performed to determine the keywords of the natural language text.

3. The travel service method according to claim 1, characterized in that, The process of determining user intent type and slot information by performing intent recognition and slot filling on the natural language text based on the domain dictionary and user historical behavior data through a pre-built large language model includes: Feature data in the natural language text is extracted using the domain dictionary and the user's historical behavior data, and then converted into vector feature data; Based on the intent classification model in the pre-built first large language model, the vector feature data is used to identify intent and determine the user intent type; Based on the slot filling model in the first large language model, the vector feature data is filled with slots according to the user intent type to generate slot information.

4. The travel service method according to claim 1, characterized in that, The process of determining user intent type and slot information by performing intent recognition and slot filling on the natural language text based on the domain dictionary and user historical behavior data through a pre-built large language model includes: Feature data in the natural language text is extracted using the domain dictionary and the user's historical behavior data, and then converted into vector feature data; The vector feature data is input into a pre-built second language model for intent recognition and slot filling to generate user intent type and slot information.

5. The travel service method according to claim 1, characterized in that, The travel service methods also include: When the confidence level of the user intent type is less than or equal to a preset threshold, candidate intent options are generated. Interact with the user through the candidate intent options, and update the user's travel request based on the interaction information provided by the user; The updated user travel request is used to re-identify the user's intent type, and it is determined whether the confidence level of the user's intent type is greater than a preset threshold.

6. The travel service method according to claim 1, characterized in that, The travel service methods also include: When any slot in the slot information is missing information, the slot information does not meet the preset missing conditions, and a slot question is generated based on the context information in the user's travel request. The system interacts with users by addressing the slot issue and updates the user's travel request based on the user's feedback. The updated user travel request is used to fill the slots again, the slot information is obtained again, and it is determined whether the slot information meets the preset missing conditions.

7. The travel service method according to any one of claims 1 to 6, characterized in that, The step of generating query information based on the slot information includes: Determine the backend format requirements for the fare search system; Based on the slot information, generate structured query information that conforms to the format requirements of the backend.

8. A travel service system based on natural language intent recognition, characterized in that, The travel service system includes: The user interaction module is used to receive user travel requests and convert the user travel requests into natural language text. The data acquisition module is used to acquire domain dictionaries and user history behavior data from a pre-built domain knowledge base and historical request data; The large model processing module is used to perform intent recognition and slot filling on the natural language text based on the domain dictionary and user historical behavior data through a pre-built large language model, and to determine the user intent type and slot information. The query information generation module is used to generate query information based on the slot information when the confidence level of the user intent type is greater than a preset threshold and the slot information meets the preset missing conditions. The backend service module is used to perform a fare search based on the query information and return the fare search results.

9. An electronic device, characterized in that, include: processor; as well as A memory storing computer-executable instructions, which, when executed, cause the processor to perform the method according to any one of claims 1-7.

10. A computer storage medium, characterized in that, in, The computer storage medium stores one or more programs that, when executed by a processor, implement the method of any one of claims 1-7.

Images