Tourist fuzzy intention analysis method and system based on short text expansion

Abstract

The application discloses a tourist fuzzy intention analysis method and system based on short text expansion, and relates to the field of smart tourism semantic understanding.The method comprises the following steps: modeling the tourism entity extracted from the short text in combination with the pre-constructed tourism knowledge graph and the context information such as the current position of the tourist, the query time and the historical preference, screening the target expansion entity through adjacent entity expansion and comprehensive correlation degree scoring, performing semantic completion on the original query and generating a rich semantic representation vector; and introducing a fuzzy intention reasoning mechanism oriented to the tourism scene on the rich semantic representation vector, and replacing the single category label with a multi-intention membership degree distribution. Thus, the primary and secondary intention relationship and the intention uncertainty can be described, and the intention constraint that is explainable and adjustable is provided for the downstream search, recommendation and travel planning.

Description

Technical Field

[0001] This application relates to the field of semantic understanding technology in smart tourism, and in particular to a method and system for parsing fuzzy tourist intentions based on short text expansion. Background Technology

[0002] In smart tourism scenarios, applications such as search, recommendation, Q&A, and customer service increasingly rely on short texts entered by tourists on their mobile phones or web pages, such as "How to plan a three-day trip in Chongqing?" or "Where to go with kids this weekend?" These queries are often short and colloquial, containing key information such as destination and activities, but also full of omissions and ambiguities, making it difficult for the system to accurately understand the tourist's true itinerary type, activity preferences, or consumption demands. Furthermore, queries submitted by tourists on different platforms and at different times may be independent, making each short text message itself extremely sparse in information.

[0003] Currently, one approach directly inputs short texts into general classification models or pre-trained language models, using vectorization and classifiers to identify intent. However, these methods typically lack specific knowledge support from the tourism sector, making it difficult to accurately distinguish fine-grained purposes such as "cultural check-in," "family experience," and "summer vacation," and also struggling to grasp the impact of spatiotemporal factors like holidays and peak seasons on intent. Another approach attempts to introduce knowledge graphs or external documents to expand or enhance the features of short texts, but these often remain at the level of general corpora or static graphs, failing to deeply integrate with contextual information such as tourists' current geographical location, travel time, and historical preferences.

[0004] Furthermore, tourist intentions themselves are characterized by obvious ambiguity and multiple overlapping intentions. For example, "going to Xi'an to visit historical sites and eat snacks" involves both sightseeing and food. Traditional hard classification only outputs a single category, which cannot express the combination relationship of "main image + secondary intention" or quantify the uncertainty of the system's judgment of each intention. This makes recommendations and itinerary planning either too simplistic or overly reliant on empirical rules, making it difficult to form interpretable and adjustable intention analysis results. Summary of the Invention

[0005] This application provides a method, system, storage medium, computer program product, and electronic device for resolving fuzzy tourist intent based on short text expansion, in order to at least solve the problems of low accuracy in recognizing tourist short text intent in smart tourism scenarios and the difficulty in comprehensively characterizing multiple intents and contextual needs in current related technologies.

[0006] In a first aspect, embodiments of this application provide a method for parsing fuzzy tourist intent based on short text expansion. The method includes: acquiring a short text query input by a tourist and associated user context information; the user context information includes at least one of the following: the tourist's current location, query time, and historical preference records; performing word segmentation, part-of-speech tagging, and named entity recognition on the short text query to obtain a set of tourism entities, and determining graph nodes corresponding to the set of tourism entities in a pre-constructed tourism knowledge graph; based on the adjacent entities associated with the set of tourism entities in the tourism knowledge graph, and in conjunction with the user context information, generating a graph node corresponding to the tourist's current location, query time, and historical preference records; and generating a graph node corresponding to the tourist's current location, query time, and historical preference records. The system first obtains a semantically relevant candidate entity set for expanding a short text query; then performs a comprehensive relevance score on each candidate entity in the candidate entity set and selects a target expansion entity set based on the score results; finally, it inputs the short text query and the target expansion entity set into an expansion fusion model, performs semantic expansion on the short text query based on the tourism knowledge graph, and generates a rich semantic representation vector to represent the short text query; and performs fuzzy intent reasoning on the rich semantic representation vector based on a preset set of tourism intent categories, calculates the intent membership degree of each tourism intent category, and forms a fuzzy intent distribution of tourists for the short text query.

[0007] Secondly, embodiments of this application provide a tourist fuzzy intent parsing system based on short text expansion. The system includes: a short text acquisition unit, used to acquire a short text query input by a tourist and associated user context information; the user context information includes at least one of the following: the tourist's current location, query time, and historical preference records; a semantic parsing unit, used to perform word segmentation, part-of-speech tagging, and named entity recognition on the short text query to obtain a set of tourism entities, and to determine the graph nodes corresponding to the set of tourism entities in a pre-constructed tourism knowledge graph; and a candidate expansion generation unit, used to generate a candidate expansion based on the adjacent entities associated with the set of tourism entities in the tourism knowledge graph, combined with the user context information. The system comprises: a semantically relevant candidate entity set for expanding a short text query; a relevance scoring unit, used to perform a comprehensive relevance score on each candidate entity in the candidate entity set and select a target expanded entity set based on the score results; an expansion fusion unit, used to input the short text query and the target expanded entity set into the expansion fusion model, and perform semantic expansion on the short text query based on the tourism knowledge graph to generate a rich semantic representation vector for representing the short text query; and a fuzzy intent reasoning unit, used to perform fuzzy intent reasoning on the rich semantic representation vector based on a preset tourism intent category set, calculate the intent membership degree of each tourism intent category, and form a fuzzy intent distribution of tourists for the short text query.

[0008] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the visitor fuzzy intent parsing method based on short text expansion according to any embodiment of this application.

[0009] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the visitor fuzzy intent parsing method based on short text expansion according to any embodiment of this application.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the visitor fuzzy intent parsing method based on short text expansion according to any embodiment of this application.

[0011] The method and system for parsing fuzzy tourist intent based on short text expansion provided in this application can achieve at least the following technical effects: (1) By jointly modeling short text queries with tourism knowledge graphs and user context information, the expressive power of short texts is significantly enhanced at the semantic level. Entity recognition is performed on short texts and mapped to nodes in the tourism knowledge graph. Then, adjacent entities are combined to generate expansion candidates. The candidate entities are scored and filtered based on the context of the tourist's current location, query time, and historical preferences. Semantic elements such as tourism themes, attraction types, and activity formats that are highly relevant to the current travel scenario are explicitly introduced into the representation. Subsequently, the expansion fusion model jointly encodes the short text and the target expansion entity to obtain a rich semantic representation vector that can simultaneously characterize text semantics, domain knowledge, and contextual constraints. Thus, targeted semantic completion and refinement are achieved for queries that are originally short in length and semantically ambiguous, making the feature carriers on which intent parsing relies more comprehensive and more in line with actual travel needs.

[0012] (2) Based on the rich semantic representation vector, a fuzzy intent reasoning mechanism for tourism scenarios is further introduced, which expands the single intent judgment into a membership degree distribution expression on a preset set of tourism intent categories. By calculating and comparing the membership degree of each intent category, the strength relationship between the dominant intent and the secondary intent can be reflected at the same time, and a continuous and quantifiable expression of uncertainty can be given for cases where the semantic boundaries are unclear. This allows tourists' multiple demands such as "sightseeing + food" and "family + leisure" to be fully represented in a unified framework, providing more granular intent-side constraints for tourism product recommendation and itinerary planning, and also making the intent results adjustable and combinable, which is convenient for flexibly controlling the scope and focus of recommendations according to business strategies.

[0013] This technical solution first utilizes a tourism knowledge graph and user context to semantically expand and fuse short texts. Then, through fuzzy intent reasoning, it maps rich semantic information into a multi-dimensional intent membership distribution, forming a computable and adjustable fuzzy intent distribution derived from sparse short texts. Thus, at a macro level, it effectively solves the problem of accurately, comprehensively, and meticulously expressing tourists' short text intents in smart tourism scenarios, providing more reliable and accurate intent recognition results for downstream search, recommendation, and itinerary planning. Attached Figure Description

[0014] 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.

[0015] Figure 1 A flowchart illustrating an example of a tourist fuzzy intent parsing method based on short text expansion according to an embodiment of this application is shown. Figure 2 A flowchart illustrating an example of constructing a tourism knowledge graph according to an embodiment of this application is shown. Figure 3 A flowchart illustrating an example of generating a tourist fuzzy intent distribution for a short text query through fuzzy intent reasoning according to an embodiment of this application is shown. Figure 4 A flowchart illustrating an example of building an intent combination pattern library is provided. Figure 5 A flowchart illustrating an example of a tourist fuzzy intent parsing method based on short text expansion according to an embodiment of this application is shown. Figure 6 A schematic diagram showing the performance comparison results of each comparison method on the test set is provided. Figure 7 This is an example illustration of the effect of the membership degree distribution of intent combinations corresponding to different queries; Figure 8 A structural block diagram of an example of a visitor fuzzy intent parsing system based on short text expansion according to an embodiment of this application is shown. Detailed Implementation

[0016] 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.

[0017] Currently, in the semantic representation and expansion of short texts, some studies have proposed using external documents to supplement information in short texts. Some works use deep memory networks or neural models with memory modules to retrieve long documents related to short texts from large-scale document sets, select valuable sentences and segments, and fuse the external text with the original short text representation. Such methods can significantly alleviate the feature sparsity problem in general short text classification tasks, but the expansion corpora are mostly from general domains such as news and encyclopedias, lacking specialized modeling of tourism-related knowledge such as scenic spot names, local cuisine, and festival activities; at the same time, the expansion process mainly relies on document content similarity, which is basically decoupled from contextual information such as the user's current location, travel time, and historical behavior, and does not take into account the uncertainty and multiple intent superposition characteristics of tourists' intentions.

[0018] Another type of research attempts to enhance short text representation by introducing knowledge graphs and graph neural networks. This involves constructing hierarchical heterogeneous graphs from the words, parts of speech, and entities appearing in the short text, and then propagating semantic information across the graph structure through graph convolution or graph attention networks, thus compensating to some extent for the lack of syntactic and semantic cues. Some works also combine topic models with graph structures, using topic distribution to adjust the importance of different nodes in the graph. These methods enhance the structured representation capabilities of short texts, but they typically rely on pre-constructed and relatively fixed graph structures, making it difficult to integrate continuously added attractions, routes, and activities in tourism scenarios in a timely manner. Furthermore, they often employ modeling methods oriented towards a single language and static corpora, failing to incorporate factors such as individual tourist differences and time sensitivity, and lacking modeling for "fuzzy intentions" and "complex intentions."

[0019] In the area of ​​smart tourism intent recognition, some studies have built tourist intent prediction models based on social media text or online reviews. These models use sentence vector encoding in conjunction with traditional classifiers to determine whether users have visited a particular destination or whether they intend to revisit it. These methods demonstrate that machine learning can extract potential travel intentions from natural language text, achieving high accuracy in binary or minority-category scenarios such as visit / not visit and revisit / not revisit. However, their intent space is typically simplified to a single label, making it difficult to distinguish between complex motivations such as "sightseeing + food" or "family + shopping." Furthermore, they do not introduce the concepts of fuzzy sets or membership degrees to characterize the relative strength and uncertainty between different intents. Therefore, they struggle to support the refined requirements of personalized recommendations and itinerary planning that require "primary intent + secondary intent."

[0020] As tourism information becomes increasingly structured, tourism knowledge graphs are gradually being used to organize various entities such as attractions, restaurants, transportation methods, and activity themes, as well as semantic relationships between entities such as subordination, proximity, similarity, and adaptation. Some studies have constructed small- to medium-scale tourism knowledge graphs based on ontology and applied them to tourism question answering or information retrieval to support finer-grained query expressions and answer generation. However, these knowledge graphs are mostly used for static queries and knowledge retrieval, and are not deeply coupled with the short text expansion process. They lack mechanisms to use graph structures to guide short queries to expand to entities related to destinations, activity types, and topic tags; furthermore, they fail to map users' current geographical location, travel time, and historical preferences to contextual constraints on the graph, making it difficult to filter and weight the expansion results according to specific scenarios.

[0021] Current research attempts to combine retrieval-enhanced generative frameworks with large-scale language models for travel planning. By retrieving online travel documents and itinerary examples, richer spatiotemporal constraints and real-world knowledge are provided to large models when generating routes, thereby improving the rationality of automated itinerary generation. These methods mainly focus on the generation quality of "how to plan routes" and "how to arrange itineraries." For the intent parsing stage when users input short queries, they often use relatively simple classifications or rules, and a multi-intent, interpretable intent parsing model for fuzzy queries has not yet been developed.

[0022] In summary, while current technologies have made some progress in short text expansion, knowledge graph modeling, and tourism intent recognition, they still generally suffer from the following shortcomings: First, short text expansion relies heavily on general corpora or static graph structures, lacking entity and topic modeling specific to the tourism sector, making it difficult to accurately complete key information such as destination and activity type. Second, the expansion and intent recognition processes rarely incorporate contextual constraints such as the tourist's current location, query time, and historical preferences, failing to provide different interpretations of the same sentence in different scenarios. Third, there is a lack of systematic fuzzy modeling and combinatorial reasoning mechanisms for tourists' ambiguous and overlapping intents, relying mainly on single-label hard classification, making it difficult to output primary and secondary intents and uncertainty levels. These problems collectively result in smart tourism systems struggling to obtain stable, fine-grained, and interpretable intent parsing results when faced with short text queries in real-world scenarios.

[0023] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0024] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0025] Figure 1 A flowchart illustrating an example of a tourist fuzzy intent parsing method based on short text expansion according to an embodiment of this application is shown.

[0026] Regarding the execution subject of the method in this application embodiment, it can be any controller or processor with computing or processing capabilities, such as a tourism semantic management platform controller. It performs semantic expansion driven by knowledge graph and user context around short text queries. By introducing target expansion entities closely coupled with the scene on the input side, a unified rich semantic vector is formed on the representation side, and multi-intent parsing results are given in the form of membership degree on the output side. This allows tourists' ambiguous intentions to be characterized in a fine-grained, composable and uncertain manner within a unified representation framework, thereby providing more refined and interpretable intent parsing capabilities for smart tourism applications.

[0027] In some examples, it may be integrated into an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device may be diverse.

[0028] like Figure 1As shown, in step S110, the short text query input by the tourist and the associated user context information are obtained; the user context information includes at least one of the following: the tourist's current location, query time, and historical preference records.

[0029] In some implementations, the platform receives short text queries entered by tourists on mobile devices or web pages, such as "Where to go for a three-day trip in Chongqing?" or "What to do with kids this weekend?". Simultaneously, it extracts user context information related to the query from user profiles and the current conversation environment. The tourist's current location can be obtained through GPS positioning, base station positioning, or information from the selected departure city, and converted into standard geocoding or administrative region identifiers. The query time can be refined into specific dates, weekday information, and time tags such as whether it falls on a statutory holiday, summer vacation, or winter vacation. Historical preference records can be derived from tourists' browsing history, favorites, order records, and historical reviews on the travel platform, and summarized into structured features such as preferred destination type (city / nature / theme park, etc.), preferred activity type (sightseeing / leisure / outdoor / family, etc.), and price range preference.

[0030] By simultaneously supplementing contextual information such as the user's current location, query time, and historical preferences during the short text query collection stage, semantic parsing no longer relies solely on a single text and can incorporate constraints and preference features closely related to the travel scenario into a unified data structure.

[0031] In step S120, the short text query is segmented, part-of-speech tagging is performed, and named entity recognition is performed to obtain a set of tourism entities, and the graph nodes corresponding to the set of tourism entities are determined in the pre-constructed tourism knowledge graph.

[0032] In some implementations, word segmentation and part-of-speech tagging preferably adopt a combination of a general word segmentation model and a tourism-specific dictionary, adding proprietary terms such as scenic spot names, city names, route names, event names, and festival names to the domain lexicon to reduce missegmentation and mislabeling; the named entity recognition part can use sequence labeling models such as BiLSTM-CRF and Transformer-CRF to classify and recognize entities such as "place name entities", "scenic spot entities", "food entities", "festival event entities", and "mode of transportation entities".

[0033] Furthermore, the identified set of tourism entities is matched with a pre-constructed tourism knowledge graph. Matching can employ string matching strategies based on entity aliases, pinyin, and thesaurus, or it can combine similarity matching based on vector representation and entity type constraints. This determines a set of graph nodes in the knowledge graph corresponding to the tourism entities in the short text. Thus, the mapping from unstructured query text to structured knowledge entities is achieved, obtaining a standardized set of tourism entities and their anchored positions in the knowledge graph.

[0034] In step S130, based on the neighboring entities associated with the tourism entity set in the tourism knowledge graph and combined with user context information, a set of expanded candidate entities semantically related to the short text query is generated.

[0035] In some implementations, the system takes the nodes of the above-mentioned tourism entity set in the knowledge graph as the center, retrieves the adjacent entities that have predefined relationships with them, such as entity nodes that have relationships with the current city or scenic spot such as "suitable people", "suitable season", "typical activities", "popular activities", "representative food", "nearby attractions", "mode of transportation", etc., and limits the relationship type and hop number of the adjacency according to business needs (for example, giving priority to one-hop or two-hop neighbors) to form a preliminary set of candidate entities for expansion.

[0036] When generating expanded candidate entities, user context information is also incorporated: for example, when the query time is during the summer vacation and the current location is a southern city, relevant entities such as "summer retreat", "seaside", and "water activities" are prioritized in the adjacent entities; when historical preference records show that the user has selected "family travel" products multiple times, entities such as "family amusement park", "edutainment experience", and "child-friendly hotel" are prioritized, thus introducing filtering and amplification of spatiotemporal conditions and personalized preferences at the candidate stage.

[0037] In step S140, a comprehensive relevance score is performed on each candidate entity in the candidate entity expansion set, and the target expansion entity set is selected based on the score results.

[0038] In some implementations, the comprehensive relevance score of the expanded candidate entity can be composed of multiple weighted sub-scores, such as: the similarity score between the candidate entity and the original short text in the semantic space, the distance and relationship weight score between the candidate entity and the identified tourism entity in the graph structure, the spatial distance between the candidate entity and the user's current location and the seasonality score with the query time, and the matching score with the user's historical preference tags, etc. In addition, the weight coefficients of each sub-score can be determined by offline training, or fused using a learned scoring model, which should not be restricted here.

[0039] Furthermore, based on the comprehensive relevance score, entities with high scores that meet a certain threshold or rank in the top K are selected as the target expanded entity set. This effectively suppresses noisy entities that are irrelevant or weakly related to the current query, ensuring that the remaining expanded entities are semantically related to the user's current question and highly consistent with geographical location, time, and personal preferences.

[0040] In step S150, the short text query and the target expanded entity set are input into the expansion fusion model, and the short text query is semantically expanded based on the tourism knowledge graph to generate a rich semantic representation vector for representing the short text query.

[0041] In some implementations, the original short text query and the target expanded entity set are jointly input into the expansion fusion model, which may include a text encoding subnetwork and an entity encoding subnetwork, and jointly model using the structural and attribute information of entities in the tourism knowledge graph.

[0042] In some simplified examples, the short text segmentation sequence can be encoded together with the name, type label, and graph adjacency summary of each target expanded entity. Multi-head attention interaction can be performed through a Transformer-like model, so that the keywords in the short text can focus on the features of the expanded entity that are highly related to them. At the same time, entity vectors pre-trained in the knowledge graph can be introduced for each entity, so that the graph structure is integrated into the overall representation.

[0043] Here, the output of the expansion fusion model is a semantically rich representation vector that characterizes the short text query. This vector carries the semantic information of the original short text, the domain knowledge of the expanded entity in the knowledge graph, and the spatiotemporal and preference constraints reflected in the user context in a low-dimensional space, providing information-rich and structurally stable feature inputs for subsequent intent reasoning.

[0044] In step S160, fuzzy intent reasoning is performed on the rich semantic representation vector based on a preset set of tourism intent categories, and the intent membership degree of each tourism intent category is calculated to form a fuzzy intent distribution of tourists for short text queries.

[0045] In some implementations, fuzzy intent reasoning is performed on the aforementioned semantically rich representation vectors based on a preset set of tourism intent categories (such as sightseeing, cultural experiences, family trips, leisure vacations, outdoor hiking, food exploration, nighttime entertainment, etc.).

[0046] In some simplified examples, a multi-label classification network can be used, which uses activation functions such as Sigmoid to output the membership value corresponding to each intent category. The obtained membership value is between 0 and 1, which is used to represent the degree of matching between the short text query and each intent category and the uncertainty of the matching relationship. On this basis, a rule-based fuzzy inference unit is introduced to adjust the membership of samples containing typical entity combinations or contextual patterns.

[0047] Through this fuzzy intent reasoning process, a single semantically rich representation vector is mapped to a multidimensional distribution of fuzzy tourist intents. This allows the system to simultaneously characterize the relative importance of dominant and secondary intents under the same query, while preserving continuity and adjustability in scenarios with fuzzy semantic boundaries.

[0048] It should be understood that the types and distribution of intents in a fuzzy intent distribution can be diverse. In some extreme cases, the fuzzy intent distribution contains only a single or a small number of intent types with high membership, which can express, to some extent, the tourist's clear play preferences. In other extreme cases, the fuzzy intent distribution contains a large number of intent types, or the intent types are evenly distributed, which can express, to some extent, the tourist's diverse play preferences. These advantages are not available in single-label intent recognition systems.

[0049] In the embodiments of this application, the fuzzy intent distribution can be stored as a structured result or directly used to drive diverse downstream applications, such as route recommendation, product ranking, and intelligent question answering. This allows short text-based tourism intent parsing to be semantically rich, context-sensitive, and capable of expressing multiple intents while maintaining model computational efficiency.

[0050] Figure 2 A flowchart illustrating an example of constructing a tourism knowledge graph according to an embodiment of this application is shown.

[0051] like Figure 2 As shown, in step S210, multi-source tourism data is collected, and the multi-source tourism data is cleaned and deduplicated to obtain the original tourism dataset.

[0052] Here, multi-source tourism data includes at least one of the following: tourism platform query logs, travelogues and reviews, scenic spot and activity information released by scenic area management departments, and structured scenic spot data provided by online tourism platforms.

[0053] In some implementations, the system accesses multiple data sources, including query log databases from online travel platforms, travelogue and review content libraries, publicly available data interfaces from scenic area management departments, and structured attraction data interfaces provided by online travel platforms. Separate acquisition adapters can be designed for raw data from different sources. After acquisition, the multi-source data undergoes unified encoding format conversion, field alignment, and timestamp standardization, and is cleaned using denoising and deduplication strategies. Specifically, quality control can be performed simultaneously based on a preset rule base and statistical features. For example, empty queries, records containing only stop words or emoticons are filtered out, as are obvious advertising statements, bot spam content, and records where the language detection indicates a language other than the target language. This significantly reduces the interference of noisy data on subsequent entity extraction and relation mining.

[0054] In step S220, tourism entities including attractions, activities, modes of transportation, destination cities and festivals are extracted from the original tourism dataset, and entity relationships used to represent spatial subordination, thematic association and temporal constraint are extracted to form a set of tourism entities and a set of entity relationships.

[0055] In some implementations, the system uses a combination of natural language processing and rule mining to extract tourism entities and entity relationships from the original tourism dataset. Specifically, it first performs word segmentation, part-of-speech tagging, and named entity recognition on unstructured text such as query text, travelogue content, and scenic spot introductions. This is combined with a custom dictionary for the tourism domain to identify candidate entity phrases such as scenic spot names, activity types, modes of transportation, destination cities, and festivals. For scenic spot entries, activity lists, and transportation information in structured data, their field values ​​are directly read as candidate entities. Subsequently, a unified entity normalization module unifies aliases with different spellings but referring to the same scenic spot or activity, such as merging "Hongyadong Pedestrian Street" and "Hongyadong Scenic Area" into the same scenic spot entity. For ambiguous expressions such as "Old Street" and "Ancient City," the city or region to which they belong is determined by combining contextual geographical location or co-occurring scenic spot information.

[0056] In entity relationship extraction, spatial subordination, thematic association, and temporal constraint relationships can be identified based on rule templates and statistical learning methods. For example, by identifying expressions such as "located in," "situated in," and "belonging to a certain scenic area," as well as administrative division information in structured fields, spatial subordination relationships can be established between destination cities and attractions, and between scenic areas and sub-attractions. By analyzing phrases such as "suitable for families," "suitable for couples," "suitable for hiking," and "beautiful night view" in travelogues and reviews, combined with a predefined thematic tag system, thematic association relationships can be established between attractions and activities and thematic tags such as "family travel," "food experience," and "night tour." By parsing the start and end times of activities, holiday names, and time expressions such as "every year on [month] [day]" and "during National Day," temporal constraint relationships can be established between festival events and corresponding attractions or activities. After this process, a set of tourism entities covering attractions, activities, modes of transportation, destination cities, and festival events can be formed, as well as a set of entity relationships used to characterize spatial, thematic, and temporal semantics.

[0057] In step S230, tourism entities are used as graph nodes and entity relationships are used as connecting edges to construct a basic tourism knowledge graph that represents semantic connections in the tourism field. Weights are assigned to each entity node and relationship edge based on the co-occurrence frequency of the entity or relationship in historical queries.

[0058] Specifically, each attraction, activity, mode of transportation, destination city, and festival is modeled as a node object with a unique identifier and attribute fields, including name, category, geographic coordinates, set of thematic tags, and opening hours. Simultaneously, different types of edges are defined for spatial subordination, thematic association, and temporal constraints. Metadata such as the starting entity, ending entity, relationship type, and relationship source are recorded in the graph. This creates a heterogeneous graph structure with multiple types of nodes and edges, allowing the semantic relationships between different entities to be explicitly presented within the graph structure.

[0059] To reflect the importance of different entities and entity relationships in real tourist behavior, the system assigns weights to each entity node and relationship edge based on the co-occurrence frequency in historical queries. For example, the algorithm can count the number of times entity pairs such as "city-attraction," "attraction-activity," and "attraction-theme tag" co-occur in short text queries within a certain time window, and use the co-occurrence frequency or normalized co-occurrence probability as edge weights. For core attractions, popular activities, and high-profile festivals that are frequently mentioned in a large number of queries, higher node weights can be assigned to represent their representativeness in the tourism scenario.

[0060] Therefore, after weighting, the basic tourism knowledge graph not only depicts "what entities exist and what relationships exist", but also "the strength differences of these relationships in real tourism behavior". Thus, when expanding short texts and calculating entity associations in the future, the graph can be expanded along high-value relationships based on weights, thereby improving the accuracy and practicality of semantic association mining.

[0061] In step S240, based on the tourist interaction behavior data recorded in the original tourism dataset, the topic tags, destination cities, and travel time periods used to represent tourist travel preferences are extracted, and the topic tags, destination cities, and travel time periods are abstracted into context entities.

[0062] Here, we construct context-dimensional knowledge entities using tourist interaction data centrally recorded in the original tourism dataset. More specifically, we mine thematic tags representing tourist travel preferences from user query logs, click records, order records, and itinerary reviews, combining product categories, tag systems, and text clustering / classification models, such as "family travel," "in-depth cultural travel," "food exploration," "summer retreat," and "self-driving tours to surrounding areas." We extract actual destination cities or city combinations from order information and itinerary planning results and standardize them into unified city entities. We extract travel dates from the travel time field and divide travel time periods accordingly, such as "weekend / weekday," "mini-holiday / long-holiday," "summer and winter vacations," "off-season / peak season," and more granular seasonal tags such as "spring, summer, autumn, and winter."

[0063] The aforementioned topic tags, destination cities, and travel time periods are all abstracted as contextual entities and introduced into the graph as a type of node parallel to traditional tourism entities. This clearly depicts the three contextual dimensions closely related to tourist behavior in the graph: "user preference themes," "travel time characteristics," and "destination selection," achieving abstraction and structuring from individual behavior records to general contextual characterization.

[0064] In step S250, contextual association edges are established between each context entity and the tourism entities in the tourism entity set that have a co-occurrence relationship with them in historical behavior data, so as to form a context-enhanced structure on the basis of the basic tourism knowledge graph to characterize the strength of entity association under different context scenarios, thereby obtaining a context-enhanced tourism knowledge graph that supports short text expansion and fuzzy intent reasoning.

[0065] In some implementations, the system explicitly embeds statistical associations between context entities and tourism entities into the basic tourism knowledge graph based on co-occurrence information recorded in historical behavioral data. Specifically, for each context entity (e.g., "family travel," "National Day holiday," "Chongqing surrounding areas," "summer travel," etc.), the system counts the frequency and distribution of its co-occurrence with various tourism entities such as attractions, activities, cities, and modes of transportation in historical queries, clicks, or orders. It calculates conditions such as conditional probability, mutual information, or weighted co-occurrence frequency between context entities and tourism entities, and sets a significance threshold, retaining only entity pairs with a certain level of association as context association edges. While establishing association edges, the above statistical indicators are normalized and used as edge weights to characterize the strength of actual selection or attention of each tourism entity in a specific contextual scenario. Finally, a context-enhanced structure is superimposed on the basic tourism knowledge graph, meaning that the same graph contains both general semantic relationships between tourism entities and scenario-specific associations between context entities and tourism entities.

[0066] Based on this context-enhanced tourism knowledge graph, during short text expansion and fuzzy intent reasoning, when the user's current context (such as current location, query time, and preferred topics) is known, the system can use context-related edges to perform biased selection and weighted diffusion of candidate entities, so that the expansion results automatically focus on attractions and activities that match the current scenario, thereby achieving context-sensitive semantic expansion and intent parsing capabilities.

[0067] Regarding the implementation details of generating the expanded candidate entity set in step S130, in some examples of the embodiments of this application, for each entity in the tourism entity set, based on the synonym relationship, alias relationship and tag information in the tourism knowledge graph, entities that semantically match the entity are retrieved and recorded as the first type of expanded candidate entities, which are used to cover different ways of expressing tourists.

[0068] Specifically, during the construction of the tourism knowledge graph, multiple name nodes (such as official names, common abbreviations, foreign names, dialect names, etc.) are maintained for the same real scenic spot or destination, and explicit connections are made through "synonym relations" and "alias relations". At the same time, semantic tags are assigned to entities based on frequently used tags in historical text data (such as "popular tourist spot", "beautiful night view", "old street", etc.). When generating expanded candidate entities, the system takes a certain entity in the current tourism entity set as the starting point, searches for entity nodes directly or indirectly connected to it along the "synonym relations" and "alias relations" edges, and combines entity type constraints and tag similarity (such as requiring the entity type to be consistent or highly similar and to share some topic tags) to filter out semantically equivalent or highly similar entities, and classify them as the first type of expanded candidate entities corresponding to that entity. For example, for the scenic spot entity "Hongyadong", it can be expanded into alias or tagged entities such as "Hongyadong Scenic Area", "Hongyadong Night View", and "Hongyadong Folk Custom Area". Through the above processing, without changing the user's original intent, multiple expressions that tourists may use in natural language are explicitly covered, improving compatibility and recall capabilities for colloquial queries and non-standard expressions.

[0069] Based on the hierarchical and semantic relationships in the tourism knowledge graph, starting from the topic tags in the tourism entity set, the entities related to the topic tags are expanded along at least one preset expansion relationship to obtain more granular scenic spot entities or specific activity entities of the destination. The entities are recorded as second-class expansion candidate entities to refine the vague purpose in the short text.

[0070] Specifically, the system first identifies entity nodes in the tourism entity set that belong to types such as "theme tags," "tourism types," and "activity categories," such as "family travel," "cultural attractions," and "summer resorts." Then, starting from these theme tags, it performs a constrained traversal of the graph along at least one preset extension relationship. The extension relationship can include: "hierarchical subordination relationships" (such as "cultural attractions" → "museums," "historical sites" → specific scenic spot entities) used to represent the subordinate relationship between higher-level categories and lower-level instances; "theme-related relationships" (such as "family travel" ↔ "zoo," "science museum," "family amusement park," etc.) used to represent the thematic proximity or common combinations; and "activity-related relationships" (such as "river view pier" ↔ "night cruise," "light show performance," etc.) used to represent the linkage between scenic spots and activities.

[0071] Preferably, the system can also limit the maximum number of hops and the number of candidates at each level to prevent excessive graph traversal and prioritize retaining entities with higher co-occurrence frequency and similarity scores to the current topic tag, thereby generating a set of scenic spot and activity entities that can concretize the fuzzy topic. For example, when extracting the topic tag "Internet celebrity check-in" from the short text "Three-day tour of Chongqing", the system can expand to specific scenic spot entities such as "Hongyadong", "Liziba", and "Yangtze River Cableway" through hierarchical and topic-related relationships.

[0072] The first type of expansion candidate entities and the second type of expansion candidate entities are merged and deduplicated to form an initial expansion candidate entity set.

[0073] Since the same real object may appear multiple times in the candidate list in the graph through synonyms, aliases and different extension paths, the system merges multiple candidates pointing to the same entity into one by combining the entity's unique identifier (such as entity ID), standardized name and entity type, thus avoiding the duplicate counting of the same entity.

[0074] Then, based on the visitor's current location, query time, and historical preference records in the user context information, each expansion candidate entity in the initial expansion candidate entity set is filtered out. Expansion candidate entities that are spatially outside the preset range of the visitor's current location, are inaccessible at the query time, or are obviously mismatched with the historical preference topics are removed. The expansion candidate entity set is then determined based on the filtered and retained expansion candidate entities.

[0075] Here, the system further filters the initial set of candidate entities for expansion based on user context information to obtain a set of candidate entities that highly match the current query. Specifically, regarding spatial constraints, the system uses the latitude and longitude or city code of the tourist's current location, combined with the geographical coordinates of candidate entities (such as attractions and event locations), to calculate the spatial distance or estimated travel time between them. Different accessibility thresholds are set for different travel types. For example, for a "weekend day trip" scenario, it can be limited to a two-hour one-way drive, while for "long-distance holiday travel," it can be extended to cross-province areas, and entities that are too far away to meet the preset range are eliminated. Regarding time constraints, the system compares the query time with the operating hours, opening season, and start and end times of festivals and events of candidate entities. If a candidate is a seasonal attraction (such as a ski resort or flower sea scenic area) or a limited-time event (such as a lantern festival or music festival), and the current time is not within its valid window, the candidate entity is marked as inaccessible and eliminated.

[0076] Furthermore, in terms of preference matching, the system reads thematic preferences summarized from tourists' historical preference records (such as multiple selections of products like "family," "food," and "outdoors"), calculates the similarity between the thematic tags attached to candidate entities and the preference vector, and eliminates entities that are clearly mismatched with the user's preferred themes, or downweights entities with low similarity. For example, for tourists who prefer "family," unsuitable entity options such as "high-intensity hiking" and "extreme sports" are filtered out. Through a comprehensive screening mechanism that considers spatial distance, temporal accessibility, and historical preferences, the retained set of expanded candidate entities is not only highly semantically relevant to the short text query, but also conforms to the geographical and temporal constraints of the current travel scenario and is consistent with individual tourist preferences. This makes the subsequent semantic expansion results naturally possess scenario adaptability and personalized characteristics, improving the overall effectiveness of intent parsing and recommendation planning.

[0077] Regarding the implementation details of filtering the target expanded entity set in step S140, in some examples of the embodiments of this application, the tourism entity set is denoted as... For the first candidate entity in the expanded candidate entity set... One expanded candidate entity ,based on Text representation and short text query The semantic similarity between the text representations is calculated to obtain a semantic similarity score. Semantic similarity score is used to reflect How close it is to short text queries in the semantic space.

[0078] Specifically, a text encoder based on a pre-trained language model can be used to translate scenic spots or event entities. The standard name, alias, and core tag are concatenated and encoded into a vector representation for querying. Similarly, it is encoded as a vector representation, and then cosine similarity or inner product is used as... The value of is used to characterize the closeness between the two in the semantic space: when the keywords appearing in the query are highly consistent with the candidate entity name or tag (e.g., the query contains "night view" and "river cruise", while the candidate entity is "Jialing River night cruise"), It achieves a large value; when the two are semantically unrelated or only weakly correlated, the value decreases significantly.

[0079] Based on the visitor's current location, query time, and historical preference records in the user context information, the candidate entities are expanded. Calculate the normalized value of geographic distance Time difference normalized value and the normalized value of preference difference To construct expanded candidate entities Contextual differences .

[0080] Equation (1) In the formula, The weight coefficients are non-negative and satisfy the following conditions: ; Used for characterization The relative distance between the corresponding scenic spot location and the tourist's current location. Used for characterization The degree of time mismatch between the available visiting time range of a tourist attraction and the time of query. Used for characterization The degree of difference between the theme tags of the corresponding attractions or activities and the set of preference tags in the historical preference records.

[0081] Specifically, based on the visitor's current location and physical location The geographical coordinates of the corresponding scenic spots are used to calculate the geographical distance, and then the result is obtained through methods such as maximum distance normalization. This value is set to the range [0,1], with larger values ​​indicating greater spatial distance. Then, the query time is compared with the visitable time range and opening season of the attraction or activity to determine the degree of time mismatch. For example, the similarity is close to 0 if the visit is entirely during business hours, and close to 1 if it is outside the visiting hours or the season is not suitable. Then, the similarity between the entity's set of topic tags and the user's historical preference tag set is calculated, using methods such as 1-Jaccard similarity or 1-cosine similarity, and the result is used as the normalized value for preference differences. .

[0082] In equation (1), nonnegative weighting coefficients are used. The three types of differences are weighted and summed to form a comprehensive contextual difference degree. The greater the deviation in a certain aspect, the higher its contribution to the overall degree of difference.

[0083] Contextual differences Transformed into a context matching score using a monotonically decreasing exponential mapping function. The context matching score is calculated according to the following formula: Equation (2) in, This is a context sensitivity coefficient used to control the discriminatory power between "almost" and "very much" in rating; when expanding candidate entities... When the location, time accessibility, and preferred topic are all similar to the user's contextual information, Smaller values ​​and It is close to 1; however, when any factor deviates significantly... Enlarge and cause The number of candidates is reduced exponentially, thereby automatically suppressing candidate entities that do not match the current travel scenario in the scoring.

[0084] Based on tourism knowledge graph Corresponding graph nodes and tourism entity set Calculate the weighted shortest path distance between the corresponding nodes of each entity in the graph. With tourism entities Minimum graph distance between .

[0085] Equation (3) In the formula, This indicates that in the tourism knowledge graph, from Corresponding graph nodes to entities The weighted shortest path distance for the corresponding graph node.

[0086] Based on minimum graph distance Calculate structural proximity score .

[0087] Equation (4) in, For structural sensitivity coefficient, when When an entity in the graph structure is unreachable from any entity in the tourism entity set, A penalty value greater than the preset upper limit is set to adjust the corresponding structural proximity score. It is close to zero.

[0088] Regarding structural relevance, in equation (3), for candidate entities For each corresponding node in the graph, the system calculates its connection to the set of tourism entities. Each entity node Weighted shortest path distance The weights are derived from the importance or co-occurrence frequency of edges in the graph; the minimum value among all paths is taken as the minimum graph distance. ,reflect The degree of closest association between an entity and an already identified entity on the graph structure.

[0089] Subsequently, in equation (4), the minimum graph distance is mapped to a structural proximity score, so that the graph distance is smaller (i.e., the candidate entity can be connected to the scenic spot or theme involved in the current query through a small number of high-weight relationships). The closer it is to 1, the greater the distance, even to the point of being unreachable. Approaching 0. Through this design, the system can prioritize retaining entities in the knowledge graph that are in the same semantic cluster or on the same typical combination path as the city, attraction, or topic involved in the current query, thereby improving the rationality of the expanded results from a structural perspective.

[0090] semantic similarity scoring Structural proximity score Context matching score Each entity is normalized according to a preset method, and then linearly weighted and fused to obtain expanded candidate entities. Comprehensive correlation score : Equation (5) In the formula, , and The weight coefficients are non-negative and satisfy the following conditions: ; Based on the comprehensive relevance score of each expanded candidate entity The candidate entities in the candidate entity set are sorted, and at least one candidate entity with a comprehensive relevance score not lower than the score threshold and ranked first is selected according to the preset scoring threshold and the sorting results to form the target candidate entity set.

[0091] In some implementations, min-max normalization or standardization based on historical statistical distribution can be used. Then, the three scores are linearly weighted and fused according to equation (5) to obtain the comprehensive relevance score for each candidate entity to be expanded, where... , and This is used to balance the importance of semantic matching, graph structure relevance, and context adaptation. For example, in scenarios that prioritize precise semantic matching, the appropriate improvement can be made. The value of ; in local tourism scenarios that emphasize scene fit, it can improve The system ultimately scores each candidate entity based on its overall relevance. The candidate entities for expansion are sorted, and at least one candidate entity with a score not lower than the threshold and ranked at the top is selected as the target expansion entity set based on a preset scoring threshold. This effectively filters out noisy entities that are semantically deviated, structurally isolated, or context-mismatched while ensuring rich recall, thus providing high-quality expansion entity input for subsequent short text semantic expansion and fuzzy intent reasoning.

[0092] Regarding the implementation details of generating the semantically rich representation vector in step S150, in some examples of embodiments of this application, the short text query is input into the query encoder to obtain the initial vector representation of the short text query. The user context information is then input into the context encoder to obtain the context vector representation. .

[0093] For each target expanded entity in the target expanded entity set Based on tourism knowledge graph acquisition Embedding vectors of corresponding graph nodes and their relationships with The embedding vectors of adjacent adjacent entity nodes, and for The embedding vector of an entity is aggregated with the embedding vectors of its neighboring entity nodes to form an expanded entity feature vector that contains graph structure semantics. The expanded set of entity feature vectors is denoted as .

[0094] In some implementations, to fully utilize the correlation between the short text content and user context information during the expansion stage, the short text query is first input into a query encoder (e.g., a text encoding network based on bidirectional LSTM, Transformer, or a pre-trained language model) to obtain an initial vector representation of the short text query. This vector is used to characterize semantic information such as destination and activity description in the query. Simultaneously, the visitor's current location (latitude and longitude or city code), query time (date, holiday markers, season markers, etc.), and historical preference records (preference topic tags, price range preferences, etc.) are encoded into structured feature vectors and input into a context encoder (e.g., a multilayer perceptron or a dedicated context encoding network) to obtain a context vector representation. This is used to centrally represent the current travel scenario. Based on this, each target expanded entity in the target expanded entity set is then... Based on tourism knowledge graph acquisition The embedding vectors of corresponding graph nodes and their adjacent entity nodes can be aggregated onto the central node using methods such as weighted summation, attention aggregation, or graph convolution, thereby forming an expanded entity feature vector that contains graph structural semantics. The expanded set of entity feature vectors is denoted as .

[0095] Initial vector representation based on short text query Context vector representation and the feature vectors of each extended entity Calculate the corresponding attention weights The expanded entity feature vectors are then weighted and summed to obtain the expanded feature vectors. .

[0096] Equation (6) Equation (7) In the formula, This represents a vector concatenation operation. This is a trainable attention parameter vector. for The transpose of ; It is a traversal index used to traverse each entity in the target expanded entity collection; The sum of importance scores for each target expanded entity under the current query and context conditions is used to normalize the attention weights into a probability distribution.

[0097] As shown in equation (6), the system will each With query vector and context vectors By splicing them together, we can obtain and with the trainable attention parameter vector The inner product is calculated by performing an exponential operation and then dividing by the normalization factor obtained by summing over all candidate entities, to obtain the i-th... Attention weights of each expanded entity This weight reflects the target expanded entity under the combined constraints of the current query semantics and context. The relative importance of all The sum of these values ​​is 1, forming a probability distribution.

[0098] Subsequently, as shown in equation (7), for all expanded entity feature vectors According to attention weight Perform a weighted summation to obtain the expanded feature vector. This allows multiple target expanded entities to be compressed into a single vector representation in the semantic space, and automatically highlights the expanded entities that are most critical to the current query and context, while weakening entities with weak semantic association or potential noise.

[0099] Construct a gated fusion unit for the initial vector representation of short text queries. With expanding feature vectors Perform weighted fusion to generate semantically rich representation vectors. .

[0100] Equation (8) Equation (9) In the formula, This represents an element-wise non-linear activation function. For the gated weight matrix, For bias vectors, This is the gating coefficient vector. This represents the element-wise multiplication of vectors of the same dimension.

[0101] In this embodiment, a gated fusion unit is constructed to perform weighted fusion of the above-mentioned multi-source features. As shown in equation (8), the gated unit first concatenates the vectors of the three features. Multiply by the gate weight matrix Add the bias vector Then, through an element-wise nonlinear activation function (Preferably Sigmoid, so that the output is in the [0,1] interval), thus obtaining the gating coefficient vector. , and , Having the same dimensions, the value of each dimension indicates whether the expanded feature or the initial query feature is preferred on that dimension. Then, as shown in equation (9), The system uses the vector obtained by inverting each element of the gating coefficients to perform element-wise weighted fusion of the two types of features.

[0102] Among them, by adjusting the parameters during model training. , and Learning enables the gated fusion unit to learn from the expanded feature vector. With context vector The correlation adaptive adjustment of the gating coefficient vector The value of .

[0103] More specifically, on the one hand, when the target expanded entity set contains noisy expanded information or expanded information that does not match the current context, the gating coefficient vector The value of tends to decrease, making the semantically rich representation vector More inclined to preserve the initial vector representation This avoids the interference of incorrect expansion on the results.

[0104] On the other hand, when the target expanded entity set can complete the omitted semantics in the short text query and match the user context information, the gating coefficient vector As the value of increases, the semantically rich representation vector becomes more... More inclined to enhance and expand feature vectors . contributions.

[0105] Thus, while maintaining the semantic gain brought by expansion, the model's sensitivity to expansion quality and context matching is improved through an explicit gating mechanism. This enables the generated semantically rich representation vectors to reflect the original intent of the short text and to make full use of the tourism knowledge graph and context expansion information, providing stable and reliable vector input for subsequent fuzzy intent reasoning.

[0106] Figure 3 A flowchart illustrating an example of generating a fuzzy intent distribution of a visitor for a short text query through fuzzy intent reasoning, according to an embodiment of this application, is shown.

[0107] like Figure 3 As shown, in step S310, a set of tourism intention categories is predefined to represent different travel purposes. Examples include sightseeing, cultural experiences, family trips, and culinary explorations, with categories for each travel intent. Configure the corresponding intent classification parameter vector and bias parameters Used to transform rich semantic representation vectors Mapping to the intent score space yields intent scores for each tourism intent category. .

[0108] Equation (10) In the reasoning process of equation (10), the semantically rich representation vector is... Input a linear mapping layer and calculate an intent score for each intent category. .in, This can be understood as corresponding to the first in the vector space. The "direction" of class intent. This is the bias term, learned through training on labeled samples, such that when... The score of a class is higher when it is closer to a certain type of intent feature. Larger. Thus, the semantically rich vector is projected onto a larger... The intention score space of the dimension.

[0109] In step S320, the basic intent membership degree of each tourism intent category is obtained through normalization operation.

[0110] Equation (11) In the formula, Index representing categories of travel intent Indicates the total number of categories of travel intentions; This indicates that the short text query belongs to the first... Category of travel intentions The basic degree of membership, satisfying and ; The sum of intent scores obtained by exponentially transforming the semantically rich representation vectors across all preset travel intent categories is used to normalize the individual intent scores.

[0111] In equation (11), the system assigns the aforementioned intent score vector. The basic intent membership degree of each tourism intent category is obtained by normalization using Softmax. In this formula, the numerator exponentially amplifies the score of a single category, while the denominator, by summing the exponential values ​​of the scores of all categories, normalizes the result. Between 0 and 1, and all categories The sum of them equals 1. Therefore, This can be interpreted as short text queries belonging to the first category. Category of travel intentions The more concentrated the distribution of basic membership degree or "probabilistic weights," the more confident the system is in a single intention; the more uniform the distribution, the more blurred the boundaries between multiple intentions.

[0112] In step S330, based on the basic intent membership distribution Calculate fuzzy entropy This is used to measure the degree of uncertainty in intent for short text queries.

[0113] Equation (12) Among them, for The category of travel intent is ignored in the calculation to avoid errors. Numerical problems.

[0114] To further measure the degree of uncertainty of the overall intent of the current query, the system uses equation (12) to measure the basic intent membership distribution. Calculate fuzzy entropy based on In equation (12), the entropy is minimized when the membership degree of a certain category is close to 1 and the membership degrees of the other categories are close to 0, indicating the lowest uncertainty; the entropy is maximized when the membership degrees of all categories are close to a uniform distribution, indicating that the system cannot clearly determine which category is dominant.

[0115] In step S340, based on the basic intent membership distribution Determine the maximum intention membership degree and the second largest intention membership degree and fuzzy entropy With the preset fuzzy threshold Compare and use the difference The difference threshold with respect to the preset value Compare them.

[0116] In step S351, when the fuzzy entropy Not greater than the fuzzy threshold And the difference Not less than the difference threshold At that time, short text queries are identified as queries with clear intent, and the corresponding maximum intent membership degree is assigned. The travel intent category is determined as the target intent category, and the target intent category is used as the tourist intent output for the short text query.

[0117] It should be noted that the difference It reflects the dominance of the principal graph (i.e., the first largest membership degree) relative to the secondary intention (i.e., the second largest membership degree); at the same time, it includes fuzzy entropy. With the preset fuzzy threshold Compare the values. When the fuzzy entropy is not greater than the threshold... (Overall uncertainty is low) and Not less than the threshold When the difference between the main intent and the secondary intent is sufficiently large, the system classifies the short text query as a "clear intent query." In this case, the corresponding maximum membership degree can be directly assigned. Categories of travel intentions The target intent category is identified and output as the tourist intent for the current query, which drives subsequent itinerary planning or resource recommendations based on explicit intent.

[0118] In step S353, when the fuzzy entropy Greater than the fuzzy threshold Or difference Less than the difference threshold At that time, short text queries are identified as fuzzy intent queries, and the membership degree distribution of basic intents is determined based on the intent combination pattern library. Perform complex intent reasoning to output a fuzzy intent distribution of tourists for short text queries.

[0119] Here, the basic intent membership distribution exhibits a clear multi-peak or nearly uniform pattern, leading the system to identify this short text query as a "fuzzy intent query." Instead of simply selecting a single largest category, the system then considers a pre-built intent combination pattern library... Perform complex intent reasoning. The intent combination pattern library can store common intent combinations such as "sightseeing + food", "culture + night tour", and "family + leisure vacation" and their weight structures. Based on the combination of multiple high-value categories in the current membership distribution, the system matches or infers the most suitable complex intent pattern and performs reasoning within that pattern framework. By renormalizing or adjusting the weights, a final distribution of fuzzy visitor intentions is formed. This approach preserves the real user demands that are multifaceted and highly uncertain, while also organizing them into a structured and interpretable combination of intentions.

[0120] In some examples of embodiments of this application, the intent combination pattern library contains multiple cluster center vectors, and each cluster center vector has a corresponding intent combination pattern.

[0121] In some embodiments, the intent combination pattern library is obtained by fuzzy clustering of rich semantic representation vectors of a large number of historical short text queries. Each cluster center vector in the library corresponds to a common intent combination pattern, such as "sightseeing + food", "culture + night tour", "family + leisure vacation", etc.

[0122] When reasoning about the current short text query, the system uses rich semantic representation vectors. The short text query's cluster membership degree to each cluster center is determined by the distance in the representation space between the query and the cluster center vectors in the intent combination pattern library. Equation (13) In the formula, and For cluster center index, and These are the corresponding cluster center vectors; This represents the total number of cluster centers. This is a fuzzy index used to control the "degree of fuzziness" of membership; when... As the distance to a center approaches 1, its advantage is amplified, resulting in... Closer to hard clustering; when When the membership is large, multiple centers can simultaneously obtain a high degree of membership, which better reflects the situation of multiple intentions superimposed. Indicates the first The cluster membership degree corresponding to each cluster center satisfies .

[0123] Since equation (13) applies to each of Both are calculated using relative distance ratios, naturally satisfying... In other words, the membership degrees of all cluster centers constitute a probability distribution. Through this fuzzy clustering method, the system can determine in the continuous vector space which typical intent combinations the current query is closer to, and allows it to "partially belong" to multiple patterns simultaneously.

[0124] Based on the membership degree distribution of basic intent Clustering membership degree and the intent combination feature vectors corresponding to each cluster center in the intent combination pattern library. The intent membership degree of the short text query in each tourism intent category is combined and modified to obtain a composite intent membership degree distribution.

[0125] Equation (14) In the formula, This is the fusion coefficient, used to balance the influence of the current individual basic intent distribution and historical intent combination patterns in short text queries; Indicates the first In the intention combination pattern corresponding to the cluster center, the first... The degree of membership of a composite intent category for each tourism intent; Indicates the first In the intention combination pattern corresponding to the cluster center, the first... The typical membership degree of a tourism intent category is used to characterize the common intent combinations represented by the cluster center.

[0126] In equation (14), when When it approaches 1, This is mainly due to the model's judgment of single intent. The decision was made to make only minor pattern adjustments; when When smaller, More towards "Closer" means moving closer to the typical membership distribution of common intent combinations most similar to the current query. Among these, It can be learned offline by statistical averaging or weighting of historical samples belonging to the cluster across each intent category, thus characterizing "the proportion of each intent category in this type of combination pattern".

[0127] Through equation (14), the basic membership degree Based on this, the "proportion of multiple intents typically presented in similar historical queries" is introduced for smoothing and correction: on the one hand, when the basic prediction is already relatively clear, The settings ensure that the combination results are not overly skewed by historical patterns; on the other hand, when the basic predictions are somewhat vague but the membership of certain cluster centers is... At high levels, the system can utilize The provided pattern prior will The distribution of intentions has been adjusted to better reflect real-world travel habits, thereby improving the stability and rationality of intention reasoning in ambiguous scenarios.

[0128] Based on the membership degree distribution of composite intentions Output the fuzzy intent distribution of tourists for short text queries; wherein, when there are at least two composite intent categories whose membership degree is not lower than a preset membership degree threshold, the tourist intent category with the highest composite intent membership degree is determined as the primary intent category, and one or more tourist intent categories whose composite intent membership degree is at the second highest level and not lower than the membership degree threshold are determined as secondary intent categories, so as to explicitly characterize the composite intent pattern corresponding to the short text query.

[0129] Specifically, the system first checks whether there are at least two travel intent categories whose combined membership degree is not lower than a preset membership degree threshold (e.g., 0.3 or 0.4). If so, the intent category with the highest combined membership degree is determined as the primary intent category, and one or more intent categories with membership degrees at the second highest level but still not lower than the threshold are determined as secondary intent categories, thus explicitly characterizing the "primary intent + secondary intent" pattern. If only one category exceeds the threshold, the query can also be considered to have a single primary intent. Through this threshold- and ranking-based interpretation method, on the one hand, it ensures that the output intent distribution still retains complete fuzzy information, which can be used by downstream modules as needed; on the other hand, it provides a clear primary and secondary structure in terms of business, making it easier for the itinerary planning module to arrange the core route according to the primary intent map and insert auxiliary activities such as food, night tours, and family activities according to secondary intents, so that the final recommendation results are more in line with the real multi-intent travel needs.

[0130] Figure 4 A flowchart illustrating an example of building an intent combination pattern library is shown.

[0131] like Figure 4 As shown, in step S410, the rich semantic representation vector corresponding to each historical short text query in the historical tourist short text query set is obtained. and the distribution of basic intent membership degree across various tourism intent categories. .in This represents the sample index for historical short text queries.

[0132] Here, the system reads samples one by one from the historical short text query set of tourists, and stores the rich semantic representation vector of each historical short text query obtained according to the aforementioned process of this application, as well as the basic intent membership distribution on each tourist intent category. Thus, all historical samples are mapped to a unified vector space and a unified intent membership space.

[0133] In step S420, the semantically rich representation vector In the constructed representation space, fuzzy clustering algorithm is used to perform clustering analysis on short text queries of historical tourists to obtain multiple cluster centers. and the corresponding cluster membership degree Fuzzy clustering algorithms optimize cluster centers and cluster membership by minimizing an objective function. This is the cluster center index.

[0134] Here, the system employs a fuzzy clustering algorithm (e.g., fuzzy C-means algorithm) to perform clustering analysis in a representation space composed of semantically rich representation vectors of historical samples, in order to automatically discover common query patterns. Specifically, the objective function is defined as follows:

[0135] Equation (15) In the formula, Describe the objective function. Indicates the first Historical short text query for the first The cluster membership degree of each cluster center satisfies .

[0136] In equation (15), the objective function is essentially the sum of the weighted squared distances of all sample-cluster center pairs, where the weights are... The higher the membership degree of a sample to a center, the greater the contribution of its distance to that center to the objective function. This is achieved through iterative updates. and Minimizing equation (15) allows semantically similar historical queries to cluster in the vector space to the same or similar cluster centers, thus forming several representative "travel intention combinations" clusters, each cluster using a cluster center. and a set of membership degrees To depict.

[0137] In step S430, for each cluster center Clustering membership based on samples within a cluster and the basic intent membership degree of each historical short text query in each tourism intent category. Calculate the cluster centers The corresponding intention combination feature vector .

[0138] Here, for each cluster center The system utilizes the cluster membership degree of samples within a cluster. And the basic intent membership of these samples across various tourism intent categories. Calculate the intention combination feature vector corresponding to the cluster center. .

[0139] Equation (16) In equation (16), the molecule belongs to the cluster. Historical samples in The weighted sum of membership degrees across each tourism intent category, with weights of . Therefore, for samples with higher membership to the cluster center, their intended membership is... The greater the influence, the better; the denominator is the sum of the weighted membership degrees of all samples to that cluster center, used for normalization. The resulting... This can be understood as "in the first" In the intention combination pattern, the first The "typical membership degree" of each tourism intent category is the average proportion of samples within that cluster across all intent categories. Therefore, each cluster center has not only a semantic center vector. It is also associated with a set of feature vectors reflecting the combined structure of multiple intentions. This solidifies the common intent combinations of historical queries into computable pattern priors.

[0140] In step S440, based on each cluster center vector and its corresponding intention combination feature vector A library of intent combination patterns is constructed to provide pattern-level prior constraints for correcting the membership of each tourism intent category when performing composite intent reasoning for fuzzy intent queries.

[0141] In some implementations, the system is based on all cluster center vectors and its corresponding intention combination feature vector Build an intent combination pattern library and combine each pair It can be viewed as a retrieveable intent combination pattern unit. During online inference, for new fuzzy intent queries, its rich semantic representation vector can be used in conjunction with various... The distance between them is used to calculate the cluster membership degree, and then combined with each The basic intent membership distribution is modified at the pattern level (as in the operation of Equation (14) above), thereby introducing common intent combination patterns from historical large-scale data while maintaining individual prediction results.

[0142] In this embodiment of the application, by pre-constructing such an intent combination pattern library, stable pattern-level prior constraints are provided in the composite intent reasoning stage, so that the system can still output a fuzzy intent distribution with reasonable structure and clear hierarchy when facing queries with multiple intent superposition, sparse samples or fuzzy semantics.

[0143] Figure 5 A flowchart illustrating an example of a visitor fuzzy intent parsing method based on short text expansion according to an embodiment of this application is shown.

[0144] like Figure 5 As shown, the system first receives the input short tourist query and combines it with user context information. After word segmentation and noise reduction by the preprocessing module, the query is mapped to the tourism knowledge graph for domain knowledge retrieval. Subsequently, the adaptive expansion module generates and merges expanded entities under the constraints of the knowledge graph and user context to form a rich semantic representation of the query. Finally, the fuzzy intent reasoning module calculates the membership distribution of each intent category based on the rich semantic representation and user context, and outputs the predicted intent for the short text query.

[0145] To verify the effectiveness of the proposed method, a dataset containing 50,000 travel queries was constructed, with 30,000 used for training, 10,000 for validation, and 10,000 for testing. Each query includes text, geolocation, timestamp, and user preference information. Eight intent categories (sightseeing, dining, shopping, cultural experiences, nature exploration, leisure vacations, business travel, and general) were manually labeled. Evaluation metrics included accuracy, F1 score, macro-average recall, and membership entropy statistics.

[0146] The comparison method selected is as follows: STLR (Sentence Transformer + Logistic Regression): Train a logistic regression classifier after encoding short texts using a sentence transformer; Graph-LDA: A method that fuses hierarchical heterogeneous graphs with LDA; Fine-tuned BERT: Using BERT to fine-tune intent classification on travel data; Proposed Method: The short text expansion + fuzzy intent reasoning model proposed in this paper.

[0147] Figure 6 A schematic diagram showing the performance comparison results of each comparison method on the test set is provided.

[0148] like Figure 6 As shown, the accuracy of STLR and Graph-LDA methods based on sentence transformers is 85% and 88%, respectively, while the BERT fine-tuning method achieves 90%. The method proposed in this paper, after introducing expansion and fuzzy inference, improves the accuracy to 94% and the F1 score to 0.93. This demonstrates that integrating tourism knowledge graphs and context into the expansion process can significantly improve intent parsing performance.

[0149] Figure 7 This is a schematic diagram illustrating an example of the membership degree distribution of intent combinations corresponding to different queries.

[0150] Specifically, four queries are randomly selected and their membership degrees for four different intents are displayed. Figure 7The horizontal axis represents four categories of travel intents (sightseeing, dining, shopping, and culture), and the vertical axis represents four example queries Q1-Q4. The color intensity and the numerical values ​​in the cells indicate the degree of membership of each query to its corresponding intent. It can be seen that the model can assign reasonable membership degrees to each query across multiple intents: for example, Q1 has a membership degree of 0.70 for sightseeing and a certain membership degree of 0.40 for culture; Q2 has the highest membership degree for dining at 0.60; Q3 is primarily biased towards shopping (membership degree 0.70); while Q4 has relatively high membership degrees for both sightseeing and culture (0.50 and 0.60 respectively). Through this membership degree matrix representation, the system can explicitly characterize the relative strength of multiple intents within a single query, enabling the system to provide diversified suggestions based on uncertainties or combined intent identification results.

[0151] This paper addresses the issues of sparse and ambiguous short text queries and intents in smart tourism scenarios. It proposes a method for parsing fuzzy intents of tourists based on short text expansion. This method constructs a tourism knowledge graph, uses a context-aware graph expansion network to generate candidate expansions, and leverages fuzzy sets and fuzzy clustering to infer tourist intents. Experimental results show that this method achieves optimal inference performance across multiple key metrics, demonstrating significant advantages in complex intent parsing tasks.

[0152] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0153] Figure 8 A structural block diagram of an example of a visitor fuzzy intent parsing system based on short text expansion according to an embodiment of this application is shown.

[0154] like Figure 8 As shown, the tourist fuzzy intent parsing system 800 based on short text expansion includes a short text acquisition unit 810, a semantic parsing unit 820, a candidate expansion generation unit 830, a relevance scoring unit 840, an expansion fusion unit 850, and a fuzzy intent reasoning unit 860.

[0155] The short text acquisition unit 810 is used to acquire the short text query input by the tourist and the associated user context information; the user context information includes at least one of the following: the tourist's current location, query time, and historical preference records.

[0156] The semantic parsing unit 820 is used to perform word segmentation, part-of-speech tagging, and named entity recognition on the short text query to obtain a set of tourism entities, and to determine the graph nodes corresponding to the set of tourism entities in the pre-constructed tourism knowledge graph.

[0157] The candidate expansion generation unit 830 is used to generate a set of expansion candidate entities that are semantically related to the short text query, based on the neighboring entities associated with the tourism entity set in the tourism knowledge graph and in combination with the user context information.

[0158] The relevance scoring unit 840 is used to perform a comprehensive relevance score on each of the expansion candidate entities in the expansion candidate entity set, and to select the target expansion entity set based on the scoring results.

[0159] The expansion and fusion unit 850 is used to input the short text query and the target expansion entity set into the expansion and fusion model, and to perform semantic expansion on the short text query based on the tourism knowledge graph to generate a rich semantic representation vector for representing the short text query.

[0160] The fuzzy intent reasoning unit 860 is used to perform fuzzy intent reasoning on the rich semantic representation vector based on a preset set of tourism intent categories, and calculate the intent membership degree of each tourism intent category to form a fuzzy intent distribution of tourists for the short text query.

[0161] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. The execution instructions can be read and executed by an electronic device (including but not limited to a computer, server, or network device) to perform the steps of any of the above-described methods for parsing fuzzy intent of visitors based on short text expansion.

[0162] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform the steps of any of the above-described methods for resolving visitor fuzzy intent based on short text expansion.

[0163] In some embodiments, this application also provides an electronic device comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform steps of a visitor fuzzy intent parsing method based on short text expansion.

[0164] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0165] The electronic devices in this application can exist in various forms, including but not limited to: mobile communication devices, ultra-mobile personal computer devices, portable entertainment devices, or other airborne electronic devices with data interaction functions.

[0166] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0167] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. 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. Such 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 method for resolving fuzzy tourist intent based on short text expansion, characterized in that, The method includes: Obtain the short text query input by the tourist and the associated user context information; the user context information includes at least one of the following: the tourist's current location, query time, and historical preference records; The short text query is segmented, part-of-speech tagging is performed, and named entity recognition is performed to obtain a set of tourism entities, and the graph nodes corresponding to the set of tourism entities are determined in the pre-constructed tourism knowledge graph. Based on the adjacent entities associated with the tourism entity set in the tourism knowledge graph, and combined with the user context information, a set of expanded candidate entities semantically related to the short text query is generated. A comprehensive relevance score is performed on each candidate entity in the candidate entity set for expansion, and a target set of entities for expansion is selected based on the score results; The short text query and the target expanded entity set are input into the expansion fusion model, and the short text query is semantically expanded based on the tourism knowledge graph to generate a rich semantic representation vector to represent the short text query. Based on a preset set of tourism intent categories, fuzzy intent reasoning is performed on the rich semantic representation vector to calculate the intent membership degree of each tourism intent category, so as to form a fuzzy intent distribution of tourists for the short text query.

2. The method according to claim 1, characterized in that, The construction of the tourism knowledge graph includes: Collect multi-source tourism data, clean and deduplicate the multi-source tourism data to obtain the original tourism dataset; the multi-source tourism data includes at least one of the following: tourism platform query logs, travelogues and reviews, scenic spot and activity information released by scenic area management departments, and structured scenic spot data provided by online tourism platforms; From the original tourism dataset, tourism entities containing attractions, activities, modes of transportation, destination cities, and festivals are extracted, and entity relationships used to represent spatial subordination, thematic association, and temporal constraints are extracted to form a set of tourism entities and a set of entity relationships. By using the tourism entities as graph nodes and the entity relationships as connecting edges, a basic tourism knowledge graph is constructed to represent semantic connections in the tourism field. Weights are assigned to each entity node and relationship edge based on the co-occurrence frequency of the entity or relationship in historical queries. Based on the tourist interaction behavior data recorded in the original tourism dataset, the topic tags, destination cities and travel time periods used to represent tourist travel preferences are extracted, and the topic tags, destination cities and travel time periods are abstracted into context entities. Each context entity is connected to a tourism entity in the set of tourism entities that have a co-occurrence relationship with it in the historical behavior data. This creates a context-enhanced structure on top of the basic tourism knowledge graph to characterize the strength of entity associations in different contextual scenarios, thereby obtaining a context-enhanced tourism knowledge graph that supports short text expansion and fuzzy intent reasoning.

3. The method according to claim 1 or 2, characterized in that, The step of generating a set of expanded candidate entities semantically related to the short text query based on the adjacent entities associated with the tourism entity set in the tourism knowledge graph and in combination with the user context information includes: For each entity in the tourism entity set, based on the synonym relations, alias relations and tag information in the tourism knowledge graph, entities that semantically match the entity are retrieved and recorded as the first type of expansion candidate entities, which are used to cover different ways of tourists to express themselves. Based on the hierarchical and semantic relationships in the tourism knowledge graph, starting from the topic tags in the tourism entity set, entities related to the topic tags are expanded along at least one preset expansion relationship to obtain more granular destination attractions or specific activity entities. These entities are recorded as second-class expansion candidate entities to refine the ambiguous purpose in the short text. The expansion relationship includes any one of the following: hierarchical subordinate relationship to represent hierarchical subordination, topic-related relationship to represent topic proximity, and activity association relationship to represent the interactive connection between attractions and activities. The first type of expansion candidate entities and the second type of expansion candidate entities are merged and deduplicated to form an initial expansion candidate entity set; Based on the visitor's current location, query time, and historical preference records in the user context information, each expansion candidate entity in the initial expansion candidate entity set is filtered out. Expansion candidate entities that are spatially outside the visitor's current location, inaccessible at the query time, or clearly mismatched with historical preference topics are removed. The expansion candidate entity set is then determined based on the filtered and retained expansion candidate entities.

4. The method according to claim 1, characterized in that, The step of performing a comprehensive relevance score on each candidate entity in the candidate entity set for expansion, and selecting the target entity set for expansion based on the score results, includes: Let the set of tourism entities be denoted as For the first in the set of candidate entities to be expanded One expanded candidate entity ,based on The text representation and the short text query The semantic similarity between the text representations is calculated to obtain a semantic similarity score. The semantic similarity score is used to reflect The degree of proximity to the short text query in the semantic space; Based on the visitor's current location, query time, and historical preference records in the user context information, the candidate entities are expanded. Calculate the normalized value of geographic distance Time difference normalized value and the normalized value of preference difference To construct expanded candidate entities Contextual differences : ， In the formula, The weight coefficients are non-negative and satisfy the following conditions: ; Used for characterization The relative distance between the corresponding scenic spot location and the tourist's current location. Used for characterization The degree of time mismatch between the available visiting time range of the corresponding scenic spot and the query time. Used for characterization The degree of difference between the thematic tags of the corresponding attractions or activities and the set of preference tags in the historical preference records; Contextual differences Transformed into a context matching score using a monotonically decreasing exponential mapping function. The context matching score is calculated according to the following formula: ， in, This is the context sensitivity coefficient; when expanding candidate entities When the user's spatial location, temporal accessibility, and preferred topic are all similar to the user's context information, Smaller values ​​and It is close to 1; however, when any factor deviates significantly... Enlarge and cause Decrease exponentially; Based on the tourism knowledge graph Corresponding graph nodes and the set of tourism entities Calculate the weighted shortest path distance between the corresponding nodes of each entity in the graph. With the aforementioned set of tourism entities Minimum graph distance between : ， In the formula, In the tourism knowledge graph, it is indicated that from Corresponding graph nodes to entities The weighted shortest path distance corresponding to the nodes in the graph; According to the minimum graph distance Calculate structural proximity score ; ， in, For structural sensitivity coefficient, when When the entity in the graph structure is unreachable from any entity in the set of tourism entities, A penalty value greater than the preset upper limit is set to adjust the corresponding structural proximity score. Approaching zero; The semantic similarity score The structural proximity score and the context matching score Each entity is normalized according to a preset method, and then linearly weighted and fused to obtain expanded candidate entities. Comprehensive correlation score : ， In the formula, , and The weight coefficients are non-negative and satisfy the following conditions: ; Based on the comprehensive relevance score of each expanded candidate entity The candidate entities in the candidate entity set are sorted, and at least one candidate entity with a comprehensive relevance score not lower than the score threshold and ranked higher is selected according to the preset scoring threshold and the sorting result to form the target candidate entity set.

5. The method according to claim 1, characterized in that, The step of inputting the short text query and the target expanded entity set into the expansion fusion model, and semantically expanding the short text query based on the tourism knowledge graph to generate a rich semantic representation vector for representing the short text query includes: The short text query is input into the query encoder to obtain the initial vector representation of the short text query. The user context information is then input into the context encoder to obtain a context vector representation. ; For each target rewriting entity in the target rewriting entity set Based on the tourism knowledge graph, obtain Embedding vectors of corresponding graph nodes and their relationships with The embedding vectors of adjacent adjacent entity nodes, and for The embedding vector of an entity is aggregated with the embedding vectors of its neighboring entity nodes to form an expanded entity feature vector that contains graph structure semantics. The set of extended entity feature vectors is denoted as ; Initial vector representation based on the short text query The context vector representation and the feature vectors of each extended entity Calculate the corresponding attention weights The expanded entity feature vector is then weighted and summed to obtain the expanded feature vector. : ， ， In the formula, This represents a vector concatenation operation. This is a trainable attention parameter vector. for The transpose of ; It is a traversal index used to traverse each entity in the target expanded entity collection; The sum of importance scores for each target expanded entity under the current query and context conditions is used to normalize the attention weights into a probability distribution; Construct a gated fusion unit to initialize the vector representation of the short text query. With the expanded feature vector Perform weighted fusion to generate semantically rich representation vectors. : ， ， In the formula, This represents an element-wise non-linear activation function. For the gated weight matrix, For bias vectors, This is the gating coefficient vector. This represents the element-wise multiplication of vectors of the same dimension. Among them, by adjusting the parameters during model training , and Learning is performed so that the gated fusion unit can learn from the expanded feature vector. With the context vector The correlation adaptively adjusts the gate coefficient vector The possible values ​​of ; When the target rewriting entity set contains noisy rewriting or rewriting information that does not match the current context, the gating coefficient vector... The value of tends to decrease, making the semantically rich representation vector It is more inclined to retain the initial vector representation. ;as well as When the target expanded entity set can complete the omitted semantics in the short text query and matches the user context information, the gating coefficient vector As the value of increases, the semantically rich representation vector... It is more inclined to enhance the expanded feature vector. . contributions.

6. The method according to claim 1, characterized in that, The process of performing fuzzy intent reasoning on the rich semantic representation vector based on a preset set of tourism intent categories, calculating the intent membership degree of each tourism intent category, and forming a fuzzy intent distribution of tourists for the short text query includes: A predefined set of travel intention categories to represent different travel purposes And for each category of travel intent Configure the corresponding intent classification parameter vector and bias parameters Used to transform rich semantic representation vectors Mapping to the intent score space yields intent scores for each tourism intent category. : ， The basic intent membership degree of each tourism intent category is obtained through normalization operation: ， In the formula, Index representing categories of travel intent This indicates the total number of categories of travel intentions; This indicates that the short text query belongs to the first... Category of travel intentions The basic degree of membership, satisfying and ; The sum of the intent scores of the semantically rich representation vectors on all preset tourism intent categories is used to normalize the individual intent scores. Based on the membership degree distribution of basic intent Calculate fuzzy entropy This is used to measure the degree of uncertainty in the intent of the short text query. ， Among them, for The category of travel intention is ignored in the calculation; Based on the aforementioned basic intent membership distribution Determine the maximum intention membership degree and the second largest intention membership degree and the fuzzy entropy With the preset fuzzy threshold Compare, The difference threshold with respect to the preset value Compare; Wherein, when the fuzzy entropy Not greater than the fuzzy threshold And the difference Not less than the difference threshold At that time, the short text query is identified as a query with a clear intent, and the membership degree corresponding to the maximum intent is assigned. The travel intent category is determined as the target intent category, and the target intent category is output as the traveler intent of the short text query; When the fuzzy entropy Greater than the fuzzy threshold Or difference Less than the difference threshold At that time, the short text query is identified as a fuzzy intent query, and the membership degree distribution of the basic intent is determined according to the intent combination pattern library. Perform complex intent reasoning to output a fuzzy intent distribution of tourists for the short text query.

7. The method according to claim 6, characterized in that, The intent combination pattern library contains multiple cluster center vectors, and each cluster center vector has a corresponding intent combination pattern. Wherein, the basic intent membership degree distribution according to the intent combination pattern library Perform complex intent reasoning to output a fuzzy intent distribution of tourists for the short text query, including: Based on the rich semantic representation vector The short text query's cluster membership degree to each cluster center is determined by its distance in the representation space from the vectors of each cluster center in the intent combination pattern library. ， In the formula, and For cluster center index, and These are the corresponding cluster center vectors; The total number of cluster centers. For fuzzy index; Indicates the first The cluster membership degree corresponding to each cluster center satisfies ; Based on the aforementioned basic intent membership distribution Clustering membership degree and the intent combination feature vectors corresponding to each cluster center in the intent combination pattern library. The intent membership degrees of the short text query across various tourism intent categories are combined and adjusted to obtain a composite intent membership degree distribution: ， In the formula, This is the fusion coefficient, used to balance the influence of the current individual basic intent distribution and historical intent combination patterns in short text queries; Indicates the first In the intention combination pattern corresponding to the cluster center, the first... The degree of membership of a composite intent category for each tourism intent; Indicates the first In the intention combination pattern corresponding to the cluster center, the first... The typical membership degree of each tourism intent category is used to characterize the common intent combination relationships represented by the cluster center; Based on the membership degree distribution of the composite intent Output the fuzzy intent distribution of tourists for the short text query; wherein, when there are at least two tourist intent categories whose composite intent membership degree is not lower than a preset membership degree threshold, the tourist intent category with the largest composite intent membership degree is determined as the primary intent category, and one or more tourist intent categories whose composite intent membership degree is at the second highest level and not lower than the membership degree threshold are determined as secondary intent categories, so as to explicitly characterize the composite intent pattern corresponding to the short text query.

8. The method according to claim 6 or 7, characterized in that, The construction of the intent combination pattern library includes: Obtain the rich semantic representation vector corresponding to each historical short text query in the historical visitor short text query set. and the distribution of basic intent membership degree across various tourism intent categories. ,in This represents a sample index for historical short text queries; In the semantically rich representation vector In the constructed representation space, fuzzy clustering algorithm is used to perform cluster analysis on the historical tourist short text queries to obtain multiple cluster centers. and the corresponding cluster membership degree ,in, As the cluster center index, the fuzzy clustering algorithm optimizes the cluster centers and cluster membership by minimizing the objective function: , In the formula, Describe the objective function. Indicates the first Historical short text query for the first The cluster membership degree of each cluster center satisfies ; For each cluster center Clustering membership based on samples within a cluster and the basic intent membership degree of each historical short text query in each tourism intent category. Calculate the cluster centers The corresponding intention combination feature vector : ， Based on the cluster center vectors and its corresponding intention combination feature vector A library of intent combination patterns is constructed to provide pattern-level prior constraints for correcting the membership of each tourism intent category when performing composite intent reasoning for fuzzy intent queries.

9. A tourist fuzzy intent parsing system based on short text expansion, characterized in that, The system includes: The short text acquisition unit is used to acquire the short text query input by the tourist and the associated user context information; the user context information includes at least one of the following: the tourist's current location, query time, and historical preference records; The semantic parsing unit is used to perform word segmentation, part-of-speech tagging, and named entity recognition on the short text query to obtain a set of tourism entities, and to determine the graph nodes corresponding to the set of tourism entities in the pre-constructed tourism knowledge graph. The candidate expansion generation unit is used to generate a set of expansion candidate entities that are semantically related to the short text query, based on the adjacent entities associated with the tourism entity set in the tourism knowledge graph and in combination with the user context information. The relevance scoring unit is used to perform a comprehensive relevance score on each candidate entity in the candidate entity set for expansion, and to select the target entity set for expansion based on the score results; The expansion and fusion unit is used to input the short text query and the target expansion entity set into the expansion and fusion model, and to perform semantic expansion on the short text query based on the tourism knowledge graph to generate a rich semantic representation vector for representing the short text query. The fuzzy intent reasoning unit is used to perform fuzzy intent reasoning on the rich semantic representation vector based on a preset set of tourism intent categories, and calculate the intent membership degree of each tourism intent category to form a fuzzy intent distribution of tourists for the short text query.

Images