Methods for constructing a tourism point of interest database, storage media, and computer equipment

Abstract

This application discloses a method, storage medium, and computer equipment for constructing a tourism point of interest (POI) database. The method includes: inputting multi-source heterogeneous data into a preset multimodal model to generate POI data; wherein the preset multimodal model uses a preset feature value system for standardized feature value generation, the preset feature value system including general feature categories and multiple dimensions of specific feature categories, and the POI data includes general feature values ​​output according to the general feature categories and specific feature values ​​output according to the specific feature categories of the corresponding dimensions of the POI data; storing the general feature values ​​corresponding to the same POI data in a general feature master table, and associating the corresponding specific feature values ​​and the unique identifiers of the POIs in the general feature values ​​with a specific feature sub-table of the corresponding specific feature category, thus obtaining the tourism POI database. This application can achieve intelligent and efficient extraction and structured storage of tourism POI data, effectively improving the construction efficiency and data quality of the tourism POI database.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method for constructing a tourist points of interest database, a storage medium, and a computer device. Background Technology

[0002] With the digital development of the tourism industry, the construction of a Point of Interest (POI) database is crucial for itinerary planning and service management. Currently, the construction of POI databases typically relies on manual input or simple rule-based matching methods. Specifically, traditional solutions mainly involve manually reviewing documents such as itineraries and supplier information to extract and input relevant information, or processing data in specific formats using preset matching rules. Although neural network technology has mature applications in general fields such as image recognition and natural language processing, a systematic application solution has not yet been developed for the specific scenario of building POI databases, and traditional data processing methods are still the primary approach.

[0003] The existing methods for building point-of-interest (POI) databases based on manual input or simple rule matching have several drawbacks: First, the efficiency of POI data extraction is low, making it difficult to process unstructured data such as image-based itineraries and text-based service agreements in batches, thus failing to meet the needs of massive data processing. Second, the information structuring is poor, making it difficult to accurately segment and correlate the unique characteristics (such as dining flavor, accommodation level, shopping advantages, etc.) across all dimensions of "food, transportation, accommodation, sightseeing, shopping, and entertainment" in the tourism scenario. Third, the dynamic update capability is weak, making it difficult to quickly synchronize when POI data (such as price, business status, opening hours, etc.) changes, resulting in insufficient database timeliness. In addition, the cross-data source integration is poor, making it difficult to achieve unified modeling for data of different formats (such as documents, images, web pages, etc.) and different sources (such as travel agency itineraries, online travel platform data, supplier feedback, etc.), failing to meet the core requirements of the tourism scenario for the accuracy, completeness, and timeliness of POI data. Summary of the Invention

[0004] This application mainly provides a method for constructing a tourism point of interest database, a storage medium, and a computer device, aiming to solve the problem of low construction efficiency of tourism point of interest databases in the prior art.

[0005] To address the aforementioned technical problems, this application adopts the following technical solution: A method for constructing a tourism point of interest (POI) database. This method includes: inputting multi-source heterogeneous data into a preset multimodal model to generate POI data; wherein the preset multimodal model uses a preset feature value system for standardized feature value generation, the preset feature value system including general feature categories and multiple dimensions of specific feature categories, the POI data including general feature values ​​and specific feature values, the general feature values ​​being output according to the general feature categories, and the specific feature values ​​being output according to the specific feature categories of the corresponding dimensions of the POI data; storing the general feature values ​​corresponding to the same POI data as data entries in a general feature master table, and associating and storing the specific feature values ​​corresponding to the same POI data and the unique POI identifiers in the general feature values ​​as data entries in a specific feature sub-table corresponding to the specific feature categories, thus obtaining the tourism POI database.

[0006] In some embodiments, the method for constructing a tourism point of interest database further includes: real-time monitoring of data source changes, incrementally generating changed feature values ​​through the preset multimodal model; using a contrastive learning algorithm to compare the changed feature values ​​with the corresponding feature values ​​in the tourism point of interest database, identifying and updating the feature values ​​that need to be changed; and reporting the corresponding feature values ​​when the change magnitude of a feature value reaches a preset update threshold.

[0007] In some embodiments, the method for constructing a tourism point of interest database further includes: when the point of interest data contains border crossing information or port information, splitting the point of interest data into pre-border and post-border sections using the preset multimodal model.

[0008] In some embodiments, the preset multimodal model outputs the confidence scores corresponding to each feature value of the point of interest data, as well as the overall confidence score of the point of interest data; wherein, when the data source explicitly records the feature values, the confidence score is 1; when the feature values ​​of the point of interest are obtained through context inference, the confidence score is a preset inference confidence score; and the overall confidence score is the average of the confidence scores of all feature values ​​in the point of interest data.

[0009] In some embodiments, the method for constructing a tourism point of interest database further includes: when feature values ​​of the same point of interest data extracted from multiple sources conflict, performing weighted voting based on the confidence level output by the preset multimodal model; and marking feature values ​​whose confidence levels after weighted voting are within a preset high confidence range as feature values ​​not within the preset high confidence range as pending verification.

[0010] In some embodiments, after inputting multi-source heterogeneous data into a preset multimodal model to generate point-of-interest (POI) data, the method further includes: determining whether the POI data has missing feature values, and marking missing feature values ​​as to be supplemented; based on common sense about tourism scenarios, determining whether the feature values ​​of the POI data are reasonable through the preset multimodal model, and marking unreasonable feature values ​​as abnormal.

[0011] In some embodiments, before inputting multi-source heterogeneous data into a preset multimodal model to generate corresponding point-of-interest (POI) data, the method further includes: constructing the preset feature value system; wherein, the feature values ​​of the general feature category include: unique identifier of the POI, POI name, data source identifier, city, price, and information confidence level; the dimensions of the specific feature category include dining dimension, accommodation dimension, transportation dimension, sightseeing dimension, shopping dimension, and entertainment dimension; the feature values ​​of the dining dimension include: dining time, food type, and flavor label; the feature values ​​of the accommodation dimension include: accommodation level and room type; the feature values ​​of the transportation dimension include: departure point and destination, vehicle type, additional fees, transportation type, and waypoints; the feature values ​​of the sightseeing dimension include: sightseeing name and fee type; the feature values ​​of the shopping dimension include: shopping name, shopping features, and fee type; the feature values ​​of the entertainment dimension include: entertainment name and fee type.

[0012] In some embodiments, the multi-source heterogeneous data includes at least image data and text data; the preset multimodal model includes at least an image processing sub-model and a text processing sub-model; the image processing sub-model includes a convolutional neural network module and a deep neural network module based on a self-attention mechanism; the convolutional neural network module is used to extract features from the image data and identify text regions in the image; the deep neural network module based on a self-attention mechanism is used to perform text understanding and information extraction on the identified text regions and output the interest point data; the text processing sub-model includes a bidirectional long short-term memory network module and a conditional random field module; the bidirectional long short-term memory network module is used to bidirectionally encode the text data to generate a context representation sequence; the conditional random field module is used to perform named entity recognition on the context representation sequence and output the interest point data.

[0013] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a storage medium storing program data, characterized in that the program data, when executed by a processor, implements the steps of the above-described method for constructing a tourism point of interest database.

[0014] To solve the above-mentioned technical problems, another technical solution adopted in this application is to provide a computer device, which includes a processor and a memory connected to each other. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the above-described method for constructing a tourist interest database.

[0015] The beneficial effects of this application are as follows: Unlike existing technologies, this application discloses a method for constructing a tourism point-of-interest (POI) database, a storage medium, and computer equipment. This application uses a pre-defined multimodal model to extract features and standardize multi-source heterogeneous data, efficiently integrating data resources from different sources and formats, thus solving the problem of difficult unstructured data processing in traditional methods. By constructing a unified, hierarchical pre-defined feature value system, it achieves systematic identification and structured output of multi-dimensional tourism feature values, significantly improving the completeness and accuracy of POI data. By constructing an associated storage structure of a general feature master table and dedicated feature sub-tables, it achieves accurate splitting and structured storage of POI data in tourism scenarios, satisfying the unified management needs of all-dimensional feature values ​​while supporting flexible querying and dynamic updates based on business dimensions, thus achieving efficient and flexible storage of POI data. Through the above-mentioned method of constructing a tourism POI database, intelligent and efficient extraction and structured storage of tourism POI data can be achieved, effectively improving the construction efficiency and data quality of the tourism POI database, and contributing to the upgrading of intelligent management and service levels in the tourism industry. Attached Figure Description

[0016] 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating an embodiment of the method for constructing a tourist points of interest database provided in this application; Figure 2 This is a schematic diagram of the structure of an embodiment of the storage medium provided in this application; Figure 3 This is a schematic diagram of the structure of an embodiment of the computer device provided in this application. Detailed Implementation

[0017] 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 a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] The terms "first," "second," and "third" used in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] This application provides a method for constructing a tourism point of interest database, see reference. Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the method for constructing a tourism point of interest database provided in this application. The method for constructing a tourism point of interest database includes: Step 10: Input multi-source heterogeneous data into a preset multimodal model to generate point of interest (POI) data; wherein, the preset multimodal model uses a preset feature value system to standardize the generation of feature values. The preset feature value system includes general feature categories and multiple dimension-specific feature categories. The POI data includes general feature values ​​and specific feature values. General feature values ​​are output according to the general feature category, and specific feature values ​​are output according to the specific feature category of the corresponding dimension of the POI data. Step 20: Store the general feature values ​​corresponding to the same point of interest data as data entries in the general feature master table, and associate and store the specific feature values ​​corresponding to the same point of interest data and the unique identifier of the point of interest in the general feature values ​​as data entries in the specific feature sub-table of the corresponding specific feature category, thus obtaining the tourism point of interest database.

[0021] In this embodiment, multi-source heterogeneous data refers to a collection of raw data from multiple different data sources with various data structure forms. Specifically, "multi-source" means that the data comes from multiple different data providers or data collection channels, including but not limited to: itinerary documents provided by travel agencies (such as itinerary schedules in PDF format, service agreements in Word format, etc.), web page data from online travel agencies (OTAs), structured or semi-structured data fed back by suppliers, image materials uploaded by users (such as photos of scenic spots, restaurant menus, scanned copies of hotel brochures, etc.), and various text materials (such as travel guides, scenic spot introductions, service descriptions, etc.). "Heterogeneous" means that these data have significant differences in format, structure, and expression: from the perspective of data format, it includes structured data (such as database tables, Excel spreadsheets), semi-structured data (such as JSON, XML format data), and unstructured data (such as plain text, images, scanned copies, etc.); from the perspective of content expression, information on the same point of interest may be described in different ways, with different field names, different data granularities, and even information conflicts or inconsistencies may exist in different data sources. For example, the name of the same restaurant may be recorded as its full name, abbreviation, or alias in different data sources; price information may differ due to timeliness; and business hours may be inaccurate due to untimely data updates. This complexity of multi-source heterogeneous data is the fundamental reason for the inefficiency of traditional manual processing methods. This application, through its intelligent processing capabilities based on a pre-set multimodal model, can effectively address this complexity and achieve unified data integration and standardized output.

[0022] In this embodiment, the pre-trained multimodal model refers to a deep learning model that is pre-trained and capable of processing multiple modalities of data simultaneously. This model possesses multimodal information fusion processing capabilities, employing appropriate processing mechanisms for different types of input data (such as images and text), and ultimately outputting structured point-of-interest (POI) data. The core advantage of the pre-trained multimodal model lies in its ability to bridge the gap between different data modalities, effectively integrating visual information from images and semantic information from text, thereby achieving comprehensive and accurate extraction of tourism POI data. For example, for an image of a restaurant menu, the pre-trained multimodal model can not only recognize the text content (such as dish names and prices) in the image, but also understand the overall layout structure of the image, determining which information belongs to the restaurant name, which belongs to the dish description, and which belongs to the price label, and outputting the structured data according to the requirements of the pre-trained feature value system. The pre-trained multimodal model typically consists of multiple sub-models, each responsible for processing a specific type of data. The outputs of each sub-model are integrated by a fusion module to ultimately generate complete POI data. This modular design gives the pre-defined multimodal model excellent scalability. When new data types need to be supported, only the corresponding sub-model needs to be added, without requiring large-scale adjustments to the overall architecture. For example, in addition to text and images, the ability to process modal data such as speech, video, and odor can be introduced. Simply add a speech processing sub-model or a video processing sub-model to the existing model, and achieve comprehensive analysis of multimodal information through the fusion module.

[0023] In this embodiment, because the preset multimodal model can simultaneously process complex and diverse multi-source heterogeneous data and possesses deep semantic understanding capabilities, inputting multi-source heterogeneous data into the preset multimodal model for processing yields interest point data that is standardized according to a preset feature value system. This interest point data is characterized by unified structure, complete information, and accurate semantics. Specifically, the preset multimodal model automatically identifies the type of input data during processing and calls the corresponding processing sub-model for feature extraction. For example, for image data, the model first performs optical character recognition (OCR) to extract text information from the image, then combines the visual features of the image for semantic understanding, and finally outputs structured interest point data; for text data, the model performs natural language understanding to identify named entities, relationships, and attributes, and maps and outputs them according to the preset feature value system. Regardless of the original format of the input data, after processing by the preset multimodal model, the output interest point data follows a unified data structure specification, facilitating subsequent storage, querying, and management operations. This standardized output method effectively solves the problems of inconsistent data formats, chaotic field naming, and difficulty in associating information caused by different data sources in traditional methods.

[0024] In this embodiment, to ensure that the obtained point-of-interest (POI) data has a unified structural standard and complete feature coverage, a pre-defined feature value system is specifically adopted to standardize and constrain the POI data. This pre-defined feature value system is a feature classification framework specifically designed for tourism scenarios. Its design concept is a two-layer architecture of "general features + specific features," which not only ensures the comparability and correlation between different types of POIs but also fully reflects the personalized characteristics of various POIs. The construction process of the pre-defined feature value system needs to comprehensively consider the business characteristics of the tourism industry, the actual needs of data management, and common scenarios such as user queries. In specific applications, a large language model can be used as the core component of the pre-defined multimodal model. Through carefully designed prompts combined with workflow or the use of an agent, the model can be guided to output according to the pre-defined feature value system, clearly specifying the names, meanings, value ranges, and hierarchical relationships between various features, enabling the model to accurately understand and execute the feature extraction task. At the same time, traditional rule engines or template matching methods can also be used. Feature extraction rules and mapping relationship tables can be predefined to map fields in the original data to corresponding features in the pre-defined feature value system. Regardless of the technical implementation method used, the preset feature value system serves as a "data standard," ensuring that the interest point data output from different data sources and processing methods has a consistent structure and semantics.

[0025] In this embodiment, the preset feature value system includes general feature categories and multiple dimensions of exclusive feature categories to achieve comprehensive coverage and accurate characterization of tourism interest point data. The general feature categories refer to the set of basic features possessed by all types of tourism interest points. These features serve as the identity identifiers and basic attributes of interest points, applicable to various tourism scenarios such as food, transportation, accommodation, sightseeing, shopping, and entertainment. The purpose of setting general feature categories is to establish a unified index and association foundation between different types of interest points, enabling users to retrieve various interest point data through a unified query portal. Exclusive feature categories are personalized feature sets designed for specific types of tourism interest points. Their dimensional divisions are consistent with the full-process classification method of the tourism industry (e.g., including dimensions such as catering, accommodation, transportation, sightseeing, shopping, and entertainment). The feature values ​​under each dimension are attribute information unique to that type of interest point, enabling a more accurate description of the service content, quality characteristics, and user focus of the interest point. For example, specific features for the dining dimension include dining time (such as breakfast, lunch, dinner, and late-night snack), food type (such as Chinese, Western, Japanese, and hot pot), and flavor profile (such as Sichuan, Cantonese, and Shanghainese cuisine). These features help users quickly filter restaurants that match their taste preferences and dining schedules. Similarly, specific features for the accommodation dimension include accommodation level (such as five-star, four-star, and budget) and room type (such as standard room, double room, and suite). These features help users assess accommodation conditions and choose suitable room types. Through this "general + specific" feature classification approach, the pre-set feature value system ensures both data consistency and takes into account the personalized characteristics of various points of interest, achieving a comprehensive and accurate portrayal of tourism interest data.

[0026] In this embodiment, under the constraints and guidance of the preset feature value system, the preset multimodal model will classify and output the point of interest data extracted from multi-source heterogeneous data into general feature values ​​and specific feature values. General feature values ​​are output according to general feature categories. Each point of interest data will contain a complete set of general feature values, such as a unique identifier (used to uniquely identify the point of interest), the name of the point of interest (the official name of the point of interest), the data source identifier (identifying which data source the data originates from), the city (the name of the city where the point of interest is located), the price (price information or price range of the point of interest), and the information confidence level (indicating the credibility of the extracted information). Specific feature values ​​are output according to the specific feature dimensions corresponding to the point of interest data. Different dimensions of point of interest will output different types of specific feature values. For example, when the preset multimodal model identifies a point of interest as belonging to the catering type, it will output feature values ​​such as dining time, food type, and flavor label according to the specific feature categories of the catering dimension; when it identifies a point of interest as belonging to the accommodation type, it will output feature values ​​such as accommodation level and room type according to the specific feature categories of the accommodation dimension. In addition, when outputting feature values, the pre-defined multimodal model will also output the confidence information of each feature value, which is used to identify the credibility of the feature value and provide a basis for subsequent data quality assessment and conflict resolution.

[0027] In this embodiment, for example, for a series of input information such as "Restaurant A / Hong Kong Style Meal / Hong Kong / 20 yuan, Day Trip / Wong Tai Sin → Victoria Harbour Cruise / Hong Kong Tour Bus (45-65 seats) / 0 yuan" retrieved from a certain platform, the preset multimodal model will perform standardized processing according to the preset feature value system. For the point of interest "Chinese food", the model extracts general feature values ​​such as the unique identifier CY26000001, the name of the point of interest "Restaurant A", the data source identifier "a certain platform", the city "Hong Kong", the price "20 yuan", and the information confidence level (derived from a comprehensive evaluation based on the reliability of the data source and the model processing results, such as 0.97). At the same time, according to the specific feature categories of the catering dimension, specific feature values ​​are extracted such as the dining time "lunch" (which can be determined based on dining habits and system time, and may also be marked as "unknown" or "to be supplemented" under some processing methods), the type of food "Chinese food" (which can be determined based on whether Hong Kong-style food belongs to Chinese food or the term "Restaurant A" is Chinese restaurant, and may also be marked as "unknown" or "to be supplemented" under some processing methods), and the flavor label "Hong Kong-style food". For the point of interest "day trip," the model first extracts general feature values, such as the unique identifier of the point of interest, the name "day trip," the data source identifier, and the city "Hong Kong." Then, based on the specific content of the tour itinerary, it extracts specific feature values ​​according to the transportation dimension, such as departure and destination "Wong Tai Sin → Victoria Harbour cruise," mode of transportation "Hong Kong tour bus," vehicle type "45-65 seats," and additional fees "0 yuan." This method of outputting specific feature values ​​by dimension avoids the data redundancy and low query efficiency caused by mixing all types of feature values ​​together, while ensuring that each type of point of interest receives a sufficient feature description.

[0028] In this embodiment, after generating standardized point-of-interest (POI) data, this data can be stored in a database to construct a complete tourism POI database. Specifically, this application adopts a "master table + sub-table" relational storage architecture for data storage. This architecture design fully considers the characteristics and query requirements of tourism POI data. The general feature master table stores the general feature values ​​of all POIs, with each record corresponding to one POI, using the unique identifier of the POI as the primary key. The design of the general feature master table allows users to retrieve all types of POIs through a unified query entry, facilitating cross-type comprehensive queries and statistical analysis. The exclusive feature sub-tables are divided according to the dimensions of exclusive feature categories, with each dimension corresponding to an exclusive feature sub-table, such as a catering feature sub-table, an accommodation feature sub-table, and a transportation feature sub-table. Each exclusive feature sub-table stores the exclusive feature values ​​of the POIs corresponding to the dimension and is associated with the general feature master table through the unique identifier of the POI. The advantages of this design are twofold: First, it avoids the data sparsity problem caused by storing all feature values ​​in a single wide table (because different types of points of interest have different specific features, storing them in one table would result in a large number of null values). Second, it facilitates specialized queries based on business dimensions. For example, if a user only wants to query restaurant-related points of interest, they can directly search the restaurant feature sub-table without scanning other types of feature data, thus improving query efficiency. Furthermore, this architecture facilitates subsequent dynamic expansion. When it's necessary to add new types of points of interest or new specific features, simply add the corresponding specific feature sub-table or add fields to existing sub-tables; this will not affect the structure of the general feature main table or the data in other specific feature sub-tables.

[0029] In the above embodiments, intelligent processing of multi-source heterogeneous data through a preset multimodal model achieves unified integration and standardized output of data from different sources and in different formats, solving the problems of low data processing efficiency and high labor costs in traditional methods. By constructing a preset feature value system of "general features + specific features," a comprehensive and accurate characterization of tourism point of interest data is achieved, ensuring comparability between different types of points of interest while fully reflecting the personalized characteristics of each type. By adopting a "master table + sub-table" associated storage architecture, efficient storage and flexible querying of tourism point of interest data are achieved, facilitating cross-type comprehensive retrieval and supporting specialized queries based on business dimensions. The above technical solutions work together to construct a complete method for building a tourism point of interest database, effectively improving the construction efficiency and data quality of the tourism point of interest database.

[0030] Optionally, in some embodiments, the multi-source heterogeneous data includes at least image data and text data; the preset multimodal model includes at least an image processing sub-model and a text processing sub-model; the image processing sub-model includes a convolutional neural network module and a deep neural network module based on a self-attention mechanism; the convolutional neural network module is used to extract features from the image data and identify text regions in the image; the deep neural network module based on a self-attention mechanism is used to perform text understanding and information extraction on the identified text regions and output the interest point data; the text processing sub-model includes a bidirectional long short-term memory network module and a conditional random field module; the bidirectional long short-term memory network module is used to bidirectionally encode the text data to generate a context representation sequence; the conditional random field module is used to perform named entity recognition on the context representation sequence and output the interest point data.

[0031] In this optional embodiment, a preset multimodal model structure and functional implementation method are provided. Specifically, the multi-source heterogeneous data includes at least image data and text data, which are the two most common data types in tourism scenarios. Image data includes, but is not limited to: itineraries provided by travel agencies, promotional images of attractions, restaurant menu photos, hotel facility images, scanned copies of transportation tickets, etc.; text data includes, but is not limited to: electronic itinerary documents, service agreement texts, travel guide articles, attraction introductions, user reviews, etc. For these two different types of data, the preset multimodal model is configured with corresponding processing sub-models.

[0032] In this optional embodiment, the image processing sub-model adopts a hybrid processing architecture of "Convolutional Neural Network (CNN) + Self-Attention mechanism": The convolutional neural network (CNN) module, as the first stage, is responsible for low-level feature extraction from image data. Its core function is to identify text regions within images. Through multiple convolutional and pooling operations, the CNN effectively extracts edge, texture, and region features from images, thereby locating regions that may contain text. In some implementations, the CNN module may first preprocess the input image (e.g., size normalization, grayscale conversion, noise reduction), then extract local features through convolutional layers, reduce feature dimensionality and enhance translation invariance through pooling layers, and finally output the candidate bounding box positions and confidence scores for the text regions. The output of this stage is the coordinate information of each text region in the image and a cropped image fragment.

[0033] Deep neural network modules based on self-attention mechanisms (such as large language models based on the Transformer architecture) serve as the second stage, responsible for deep semantic understanding and information extraction from the identified text regions. Self-attention mechanisms capture long-distance dependencies by calculating the correlations between elements in the sequence, thus better understanding the semantic content of the text regions. In some implementations, a deep neural network module based on self-attention mechanisms may first perform Optical Character Recognition (OCR) on the text regions, converting the text in the image into a processable text sequence; then, it encodes the text sequence using a self-attention mechanism, capturing the semantic relationships between words in the text; finally, it outputs structured interest point data through a decoding layer. Compared to traditional sequence processing models, self-attention mechanisms can better handle long-distance dependencies in text, such as understanding the modifier relationships between "five-star," "city center," and "hotel" in the expression "a five-star hotel located in the city center," thereby more accurately extracting the attribute information of interest points.

[0034] In this optional embodiment, the text processing sub-model adopts a sequence labeling architecture of "bidirectional long short-term memory network + conditional random field": The Bidirectional Long Short-Term Memory (Bi-LSTM) module is responsible for bidirectional encoding of textual data, generating contextual representation sequences. Long Short-Term Memory (LSTM) is a special type of recurrent neural network that effectively solves the vanishing gradient problem in traditional recurrent neural networks, making it suitable for processing long sequences of data. By simultaneously considering both forward and backward information of the sequence, Bi-LSTM can more comprehensively capture the contextual semantics of each position. For example, when processing the sentence "This Sichuan restaurant's signature dish is Mapo Tofu," Bi-LSTM, when encoding the word "Sichuan restaurant," can not only obtain the forward contextual information of "this restaurant" but also the backward contextual information of "signature dish," thus more accurately understanding the semantic role of "Sichuan restaurant."

[0035] The Conditional Random Field (CRF) module is responsible for performing Named Entity Recognition (NER) on the context representation sequence output by the bidirectional Long Short-Term Memory (LSTM) network module, outputting structured point-of-interest (POI) data. CRF is a discriminative probabilistic graphical model, particularly suitable for sequence labeling tasks. Unlike traditional classification models, CRF considers the constraints between adjacent labels when predicting the label at each position, learning the transition probabilities between labels to ensure the rationality of the output label sequence and avoid outputting unreasonable label sequences. Through the combination of the bidirectional LSM network and CRF, the text processing sub-model can accurately identify key information such as the name, location, price, and time of POIs in the text and output them in a structured manner according to a pre-defined feature value system.

[0036] Optionally, in some embodiments, before inputting multi-source heterogeneous data into a preset multimodal model to generate corresponding point-of-interest (POI) data, the method further includes: constructing the preset feature value system; wherein, the feature values ​​of the general feature category include: the unique identifier of the POI, the name of the POI, the data source identifier, the city, the price, and the information confidence level; the dimensions of the specific feature category include catering, accommodation, transportation, sightseeing, shopping, and entertainment; the feature values ​​of the catering dimension include: dining time, food type, and flavor label; the feature values ​​of the accommodation dimension include: accommodation level and room type; the feature values ​​of the transportation dimension include: departure point and destination, vehicle type, additional fees, transportation type, and waypoints; the feature values ​​of the sightseeing dimension include: sightseeing name and fee type; the feature values ​​of the shopping dimension include: shopping name, shopping features, and fee type; and the feature values ​​of the entertainment dimension include: entertainment name and fee type.

[0037] This optional embodiment specifically describes the construction process of the preset feature value system before generating point-of-interest (POI) data and the definition of the feature values ​​under this system. It should be understood that the construction of the preset feature value system is the foundational work for building a tourism POI database. Its design quality directly affects the efficiency of subsequent data processing and the convenience of data application. In this optional embodiment, the construction of the preset feature value system follows the principles of comprehensiveness (the feature value system can cover the core attribute information of various POIs in the tourism scenario), hierarchy (the feature value system has a clear hierarchical structure, facilitating data management and user understanding), scalability (the feature value system reserves expansion space, facilitating the addition of new feature types later), and practicality (the feature value system closely matches the actual query needs of users, avoiding setting too many redundant features) to ensure the accuracy and completeness of the feature definitions.

[0038] In this optional embodiment, the general feature category defines the basic features that all types of points of interest should possess. Among them, the unique identifier of a point of interest is a code used to uniquely identify each point of interest, usually in the form of a UUID or an auto-incrementing ID, ensuring that there are no duplicate records in the database (for example, in some embodiments, it may use the format of the specific dimension category corresponding to the point of interest data + the last two digits of the year + the sequence number, such as CY26123456 representing the 123456th restaurant-related point of interest in 2026). The name of the point of interest is its official or commonly used name, such as restaurant name, hotel name, attraction name, etc., and is the main basis for users to identify and retrieve points of interest. The data source identifier records which data provider the point of interest data comes from, facilitating data traceability and quality tracking. The city recorded records the name of the city where the point of interest is located, which is an important basis for users to filter by region. The price recorded the price information or price range of the point of interest, such as the average cost per person in a restaurant, the room rate in a hotel, and the ticket price of an attraction, which is an important reference for users to make consumption decisions. Information confidence is a numerical indicator used to represent the credibility of extracted information. Its value typically ranges from 0 to 1, with higher values ​​indicating more credible information. These six feature values ​​of the general feature category form the basic framework of the point of interest data, providing a unified association basis for subsequent specific features.

[0039] In this optional embodiment, the exclusive feature category includes at least six dimensions, categorized according to the traditional classification methods of the tourism industry, wherein: The dining dimension corresponds to the "food" element in the tourism scenario. Its characteristic values ​​include: dining time (e.g., breakfast, lunch, dinner, late-night snack, afternoon tea, etc.), used to identify the restaurant's main operating hours; food type (e.g., Chinese food, Western food, Japanese food, Korean food, Southeast Asian food, hot pot, barbecue, fast food, etc.), used to identify the main types of dishes offered by the restaurant; and flavor profile (e.g., Sichuan cuisine, Cantonese cuisine, Hunan cuisine, etc.), used to identify the restaurant's culinary style. These three characteristic values ​​help users quickly filter out restaurants that match their taste preferences and dining time arrangements.

[0040] The accommodation dimension corresponds to the "stay" element in the tourism scenario. Its characteristic values ​​include: accommodation level (such as five-star, four-star, three-star, budget, homestay, youth hostel, etc.), used to identify the service level of accommodation facilities; and room type (such as standard room, double room, twin room, suite, family room, etc.), used to identify the main room types offered by accommodation facilities. These two characteristic values ​​help users evaluate accommodation conditions and choose a suitable room type.

[0041] The transportation dimension corresponds to the "travel" element in a tourism scenario. Its characteristic values ​​include: origin and destination (e.g., "Shenzhen-Hong Kong", "Airport-Residential Area"), used to identify the start and end points of the travel itinerary; vehicle type (e.g., airplane, high-speed rail, bullet train, bus, sedan, minivan, etc.), used to identify the type of transportation; additional fees (e.g., fuel surcharge, airport construction fee, toll fees, etc.), used to identify extra costs in addition to the base fare; transportation type (e.g., flight, train, car, ferry, etc.), used to identify the major category of transportation; and transit points (e.g., stopover cities, transfer stations, etc.), used to identify intermediate stops in the travel itinerary. These five characteristic values ​​can comprehensively describe all the key information of the travel itinerary.

[0042] The tour dimension corresponds to the "tour" element in the tourism scenario. Its characteristic values ​​include: tour name (such as attraction name, scenic area name, etc.), which is used to identify the name of the tour destination; and fee type (such as free, paid, package ticket, etc.), which is used to identify the cost of the tour.

[0043] The shopping dimension corresponds to the "purchase" element in the tourism scenario. Its characteristic values ​​include: shopping name (such as mall name, store name, etc.), used to identify the name of the shopping venue; shopping features (such as local specialties, duty-free goods, luxury goods, etc.), used to identify the main types of goods in the shopping venue; and fee type (such as free admission, admission fee, etc.), used to identify the fees for entering the shopping venue.

[0044] The entertainment dimension corresponds to the "entertainment" element in the tourism scenario. Its characteristic values ​​include: entertainment name (such as performance name, event name, etc.), used to identify the name of the entertainment project; and cost type (such as free, paid, charge per project, etc.), used to identify the cost of the entertainment project.

[0045] This optional embodiment, through the above-mentioned six dimensions of exclusive feature definition and general feature definition, preset feature value system can comprehensively cover the information needs of the six major elements of "food, transportation, accommodation, sightseeing, shopping and entertainment" in the tourism scenario, accurately reflect the business characteristics of various points of interest, support refined data management and query, and provide a complete data standard framework for the construction of the tourism point of interest database.

[0046] Optionally, in some embodiments, the method for constructing the tourism point of interest database further includes: real-time monitoring of data source changes, incrementally generating changed feature values ​​through the preset multimodal model; using a contrastive learning algorithm to compare the changed feature values ​​with the corresponding feature values ​​in the tourism point of interest database, identifying and updating the feature values ​​that need to be changed; and reporting the corresponding feature values ​​when the change magnitude of a feature value reaches a preset update threshold.

[0047] In this optional embodiment, a dynamic update mechanism based on incremental learning is provided to avoid the problem of inaccurate information caused by the long update cycle of traditional point of interest databases in tourism scenarios with high requirements for timeliness and dynamism.

[0048] In this optional embodiment, real-time monitoring of data source changes refers to the system continuously monitoring the status changes of each data source, such as suppliers updating prices, attractions adjusting opening hours, and hotels changing room types. Monitoring of data source changes can be achieved in two ways: one is an active fetching method, where data is automatically fetched periodically by configuring the data source address to detect any new or changed information; the other is a passive receiving method, where point-of-interest data is output according to a preset workflow and changes are detected when the data source provider submits data.

[0049] In this optional embodiment, when a change in the data source is detected, an incremental processing flow can be triggered. This means that only the changed portion of the data is processed, rather than reprocessing all the data. This significantly improves update efficiency and reduces system load. Incremental generation of modified feature values ​​refers to extracting modified feature values ​​by analyzing and processing the changed data using a preset multimodal model. Since changed data typically only involves some features of some points of interest, incremental processing can quickly complete the generation and updating of feature values.

[0050] In this optional embodiment, the use of a contrastive learning algorithm for feature value comparison is the core component of the dynamic update mechanism. Contrastive learning is a machine learning method whose core idea is to represent features by learning the similarities and differences between samples. In this application, the contrastive learning algorithm is used to compare the differences between the changed feature values ​​and the existing feature values ​​in the tourism point of interest database. Specifically, the algorithm encodes the changed feature values ​​into vector representations and then calculates the similarity with the vector representations of the corresponding feature values ​​in the database. Similarity calculation can employ metrics such as cosine similarity, Euclidean distance, and Manhattan distance. Through similarity calculation, the system can identify which feature values ​​have undergone substantial changes and need to be updated; and which feature values ​​have changes within an acceptable range and can remain unchanged. This update strategy based on contrastive learning can effectively avoid unnecessary update operations triggered by minor fluctuations in the data source (such as format adjustments, changes in expression, etc.), improving the stability and efficiency of change processing.

[0051] In this optional embodiment, in addition to the direct comparative learning method described above, the incremental update process can also adopt a simplified strategy based on the type of feature value. That is, when the specific feature values ​​are inconsistent, the strategy of directly adding new data entries is adopted, while only comparison and change processing are performed on changes to general feature values, in order to further reduce computational complexity and improve update efficiency. For example, when the food type feature value of the catering dimension changes, since this feature value usually has a clear classification standard, the strategy of adding new data entries can be directly adopted, and the new food type feature value can be added as a new record to the point of interest database. For changes to general feature values ​​such as price or information confidence, a comparative learning algorithm is used for comparison, and the update operation is performed only when the difference exceeds a preset threshold. In addition, after the feature value is changed, its historical version can also be recorded to facilitate subsequent data auditing and rollback operations. The historical version record can also include key information such as change time, change content, and change operator, to ensure that every data update is traceable.

[0052] In this optional embodiment, reporting when a feature value change reaches a preset update threshold is a risk control step in the dynamic update mechanism. The preset update threshold is a pre-defined value used to determine whether a feature value change warrants manual intervention. For example, the change range for a price feature value can be set to 10%. When the price change exceeds 10%, the feature value can be reported, prompting management to conduct manual verification. This preset update threshold can be dynamically adjusted based on the business importance and change frequency of different feature values. Large changes in feature values ​​usually indicate data extraction errors, abnormal data sources, or significant changes in points of interest, while small changes may be normal business fluctuations. Therefore, by setting a reasonable preset update threshold, the management costs caused by frequent updates can be effectively controlled while ensuring data timeliness.

[0053] In this optional embodiment, the reporting methods for feature values ​​include, but are not limited to, system log recording, email notifications, SMS reminders, and work order generation. Any method that enables relevant management personnel to promptly obtain information about feature value changes and take appropriate action based on the actual situation is acceptable. Furthermore, in addition to notifying relevant management personnel, the changed feature values ​​can also be promptly synchronized to associated itinerary planning systems (i.e., various tourism applications or platforms that reference the corresponding point-of-interest (POI) data) to ensure users receive the latest and most accurate POI data, avoiding untimely or erroneous itinerary planning that could harm the interests of users or the platform. Through the aforementioned reporting and synchronization of feature values, data quality risks can be effectively controlled while ensuring data timeliness, preventing erroneous data from entering the POI database, and further enhancing the practicality and reliability of the tourism POI database in real-world application scenarios.

[0054] Optionally, in some embodiments, the method for constructing the tourism point of interest database further includes: when the point of interest data contains border crossing information or port information, splitting the point of interest data into pre-border and post-border sections using the preset multimodal model.

[0055] This optional embodiment describes in detail a special data processing mechanism for cross-border tourism scenarios. In cross-border tourism, itineraries typically involve customs clearance at border crossings or ports of entry, and the itinerary arrangements, service content, and cost standards before and after these crossings may differ significantly. For example, in a tour from Shenzhen to Hong Kong, tourists need to clear customs at the Shenzhen port before entering Hong Kong. The transportation, dining arrangements, and tour guide services before and after customs clearance belong to different service providers and resource types. Storing the entire itinerary as a single point of interest would result in overly coarse data granularity, an inability to accurately reflect the business characteristics of cross-border trips, and difficulties in refined resource management and cost accounting. Therefore, this application designs a splitting mechanism based on border or port information, splitting point-of-interest data containing border or port information into two independent point-of-interest data sets: one before and one after customs clearance, to achieve precise splitting and refined management of cross-border trips.

[0056] In this optional embodiment, the identification of border crossing or port information is a prerequisite for the splitting mechanism. When processing point-of-interest (POI) data, the preset multimodal model first determines whether it includes cross-border routes based on source data such as departure point, destination, and waypoints. The source data must explicitly specify border crossing or port information. Specifically, the model performs semantic analysis on the departure point, destination, and waypoint fields of transportation-related POI data to identify whether they contain keywords such as "port," "border crossing," "customs," or "border," or whether the departure point and destination belong to different administrative divisions or countries / regions. For example, when the departure point is "Shenzhen" and the destination is "Hong Kong," the model can identify the trip as a cross-border route; when the waypoint field contains information such as "Luohu Port," "Huanggang Port," or "Futian Port," the model can determine the specific border crossing. After identifying cross-border routes, the model will further determine the checkpoints or ports of entry and split the original point of interest data according to the checkpoint location, forming two independent point of interest data records: "departure point → checkpoint" and "checkpoint → destination". At the same time, it can add segment markers to each record to indicate whether it belongs to the "before checkpoint" or "after checkpoint" travel segment, so as to carry out targeted resource management and service arrangements in the future.

[0057] This optional embodiment, through its cross-border scenario-adapted splitting processing method, effectively solves the problem of inaccurate identification in general large models under cross-border scenarios, ensuring that cross-border travel data can be correctly split and stored according to business rules. The split points of interest can correspond to different service providers, pricing standards, and resource types, supporting travel agencies in refined itinerary planning and resource allocation. Simultaneously, this splitting processing method facilitates subsequent data statistics and analysis, such as statistically analyzing customs clearance traffic at various ports and analyzing the travel preferences of cross-border tourists, providing data support for optimizing tourism business decisions. Furthermore, by adding segmentation markers to the points of interest data, the complete cross-border itinerary can be clearly presented during data querying and display, improving data operability and traceability.

[0058] Optionally, in some embodiments, the preset multimodal model outputs the confidence level corresponding to each feature value of the interest point data, as well as the overall confidence level of the interest point data; wherein, when the data source explicitly records the feature value, the confidence level is 1; when the interest point feature value is obtained through context inference, the confidence level is a preset inference confidence level; and the overall confidence level is the average of the confidence levels of all feature values ​​in the interest point data.

[0059] In this optional embodiment, the preset multimodal model outputs not only the interest point data but also the confidence score corresponding to each feature value and the overall confidence score of the interest point data. The confidence score is an indicator similar to a statistical feature value, used to evaluate the reliability of the feature value data. Due to the inherent uncertainty of the data source, the data source may not contain all feature value information; some feature values ​​need to be processed through inference or default conditions (specified through prompt words). Therefore, a confidence score mechanism is introduced to quantify the reliability of the feature values.

[0060] In this optional embodiment, each feature value corresponds to a separate confidence level. Specifically, when the data source explicitly records the data corresponding to the feature value, the confidence level is 1, indicating that the feature value comes directly from the original data and has the highest credibility. For example, if the data source explicitly records "had a Hong Kong-style meal in Hong Kong for lunch, priced at 20 yuan," then the confidence levels of the feature values ​​such as "Chinese food," "Hong Kong-style meal," "Hong Kong," and "20 yuan" are all 1. When the feature value is obtained through context inference, the confidence level is a preset inference confidence level (e.g., 0.93, which can be set according to the model's inference accuracy on a specific training set or the business's requirements for data quality), indicating that the feature value is inferred from context information and has a slightly lower credibility than directly obtained feature values. For example, if the data source only records "had roast goose rice in Shenzhen," and the model infers the flavor label feature value of "Cantonese cuisine" based on "roast goose rice," then the confidence level of "Cantonese cuisine" is 0.93.

[0061] In this optional embodiment, the overall confidence score is the average of the confidence scores of all feature values ​​in the point of interest (POI) data, used to represent the overall credibility of the POI data. Specifically, the formula for calculating the overall confidence score is: Overall Confidence Score = Σ(Confidence Score of Each Feature Value) / Number of Feature Values. For example, if a POI contains 6 feature values, where the confidence score of 3 feature values ​​is 1 and the confidence score of 3 feature values ​​is 0.93, then the overall confidence score is (3 × 1 + 3 × 0.93) / 6 = 0.965. The overall confidence score can intuitively reflect the overall quality level of the POI data, facilitating a quick assessment of the data's credibility. Simultaneously, the overall confidence score can also serve as a basis for data filtering, such as setting an overall confidence score threshold to only show POI data with an overall confidence score higher than the threshold, or prioritizing the display of data with higher overall confidence scores. The introduction of the confidence score mechanism makes the data quality of the tourism POI database more transparent and controllable, providing a more reliable basis for data analysis and application.

[0062] Optionally, in some embodiments, the method for constructing the tourism point of interest database further includes: when feature values ​​of the same point of interest data extracted from multiple sources conflict, performing weighted voting based on the confidence level output by the preset multimodal model; and marking feature values ​​whose confidence level after weighted voting is within a preset high confidence level range as feature values ​​that are not within the preset high confidence level range as pending verification.

[0063] In this optional embodiment, since multi-source heterogeneous data originates from different data providers, the same feature value of the same point of interest may be inconsistent in different data sources, leading to data conflict issues. For example, the average cost per person at the same restaurant might be marked as 80 yuan in data source A and 100 yuan in data source B, resulting in a conflict in price feature values. Improper handling of data conflicts can lead to incorrect information being stored in the point of interest database, affecting user experience and decision-making quality. Therefore, this application designs a confidence-based weighted voting mechanism to resolve the data conflict problem.

[0064] In this embodiment, the core idea of ​​the weighted voting mechanism is to perform weighted calculations based on the confidence levels of feature values ​​provided by each data source, and select the feature value with the highest confidence level after weighting as the final result. Specifically, when multiple different values ​​are detected for the same feature value of the same point of interest, the confidence level corresponding to each value can be obtained, and weights can be added based on factors such as the timeliness of the data and the reliability of the corresponding data source (e.g., assigning higher weights to point of interest data from highly reliable data sources with strong timeliness). This weighted voting can also be executed by a preset multimodal model to achieve automated conflict resolution. The preset multimodal model will perform a comprehensive analysis based on the feature values ​​and their confidence levels of each data source, and output the final feature value after weighted voting and its corresponding weighted confidence level.

[0065] In this optional embodiment, the preset high confidence range is a threshold interval used to determine whether a feature value is sufficiently trustworthy. When the confidence level after weighted voting is within the preset high confidence range (e.g., confidence level ≥ 0.8), the feature value is directly adopted, considered to have sufficient credibility, and requires no manual intervention. When the confidence level after weighted voting is not within the preset high confidence range (e.g., confidence level < 0.8), the feature value is marked as "pending verification," triggering a manual review process for confirmation and processing. This conflict handling mechanism maximizes the automation of data conflict resolution while ensuring data quality, reducing the workload of manual intervention and improving the efficiency of building the point of interest database. Simultaneously, quantifying the credibility of feature values ​​through the confidence mechanism makes the data processing process more transparent and traceable, facilitating subsequent data quality management and problem investigation.

[0066] Optionally, in some embodiments, after inputting multi-source heterogeneous data into a preset multimodal model to generate point-of-interest (POI) data, the method further includes: determining whether the POI data has missing feature values, and marking the missing feature values ​​as to be supplemented; based on common sense about tourism scenarios, determining whether the feature values ​​of the POI data are reasonable through the preset multimodal model, and marking unreasonable feature values ​​as abnormal.

[0067] This optional embodiment provides a mechanism for verifying the integrity and rationality of the generated Point of Interest (POI) data. The integrity verification aims to detect whether any required feature values ​​are missing in the PIO data, ensuring that each PIO record contains the core information required for business operations. Specifically, a preset multimodal model can be used to scan the PIO data, checking each feature value against a list of required feature values ​​corresponding to the PIO type. For example, required feature values ​​for dining-related PIOs include dining time, food type, city, and price; required feature values ​​for transportation-related PIOs include departure and destination, transportation type, and vehicle type. If a required feature value is found to be empty or not extracted, it is marked as "to be supplemented," and the reason for the absence is recorded. This missing feature value can then be supplemented manually or through other data sources or semantic completion mechanisms. The integrity verification ensures the integrity of the PIO data, preventing the loss of key feature values ​​from affecting subsequent business applications such as itinerary planning and cost calculation.

[0068] In this optional embodiment, the rationality check aims to detect whether the feature values ​​of the point of interest data conform to the business logic and common sense rules of the tourism scenario, and to identify possible abnormal data or erroneous information. Specifically, a tourism common sense knowledge base (such as the correspondence between cities and attractions, the rationality of price ranges, the rationality of travel time, the rationality of geographical location, etc.) can be used to perform semantic understanding and logical judgment on the feature values ​​through a preset multimodal model. For example, if the departure point of a transportation-related point of interest data is "Hong Kong", the destination is "Beijing", and the transportation type is "bus", the model will judge the data as unreasonable based on geographical common sense, because the distance from Hong Kong to Beijing is more than 2,000 kilometers, and it is impossible to travel directly by bus; as another example, if the ticket price of an attraction is "5,000 yuan", which is far beyond the normal price range of similar attractions (usually 50-500 yuan), the model will judge the price as abnormal; as yet another example, if the dining time of a catering-related point of interest is marked as "3 am", which does not conform to the operating hours, it will also be judged as abnormal. If an unreasonable feature value is detected, the feature value is marked as "abnormal" and a review process is triggered, which is confirmed and corrected manually or by algorithm. Reasonableness checks can effectively identify outliers and errors in the data, improve the data quality of the point of interest database, and prevent erroneous data from negatively impacting business decisions.

[0069] This optional embodiment employs a dual quality control mechanism of integrity and rationality checks. This mechanism effectively screens Points of Interest (POI) data before it is added to the database, ensuring the completeness and accuracy of the stored data. Integrity checks guarantee data completeness, preventing the loss of critical information; rationality checks ensure data correctness, preventing the inclusion of abnormal data. These two mechanisms work together to form a comprehensive data quality assurance system, effectively improving the data quality of the tourism POI database and providing a reliable data foundation for subsequent business applications such as itinerary planning, resource allocation, and cost calculation. Furthermore, the "to be supplemented" and "abnormal" labeling mechanism makes the data processing process more transparent and traceable, facilitating subsequent data quality management and problem troubleshooting.

[0070] Optionally, in some embodiments, the preset multimodal model is further used to standardize feature values ​​based on preset feature value standardization rules; wherein, the preset feature value standardization rules include format unification rules and content specification rules.

[0071] In this optional embodiment, the process of eigenvalue standardization for ensuring the consistency and comparability of stored data is elaborated in detail. The data of tourist interest points comes from multi-source heterogeneous channels, and there may be differences in the expression forms and contents of their eigenvalues when they are initially generated. Without unified standardization processing, it will directly affect the effects of data aggregation, statistics, analysis and application. For example, the price corresponding to the same interest point may be expressed as "￥1000", "RMB1000", "1000 yuan", "one thousand yuan" or "0.1 ten thousand yuan" in different source data. Without standardizing the price unit, it will cause difficulties in subsequent cost accounting or price comparison. Another example is that the name of the same city may be expressed as "Beijing", "the capital", "Beijing" or "Beijing Municipality", etc. Without standardizing the city name, it will affect the accuracy of regional screening and attribution statistics.

[0072] In this optional embodiment, the preset eigenvalue standardization rules mainly include two categories: format unification rules and content specification rules. The format unification rules aim to eliminate the differences in different expression forms of the same eigenvalue and convert it into a unified format. For example, for the price eigenvalue, the unit is uniformly converted to "yuan" and other currency symbols or explanatory texts are removed. The content specification rules aim to standardize the specific content of the eigenvalue to make it conform to a predetermined standard or classification system. For example, for the city name eigenvalue, the standard naming of national administrative divisions is uniformly adopted (such as "Beijing Municipality" instead of "Beijing", "Shanghai Municipality" instead of "Shanghai", and the full name is used for provincial capital cities, etc.) to avoid ambiguity.

[0073] In this optional embodiment, the preset multimodal model will process the extracted eigenvalues one by one according to the preset eigenvalue standardization rules. For the format unification rules, the model will identify the current format of the eigenvalue and apply the corresponding conversion rules to convert it into a unified format. For the content specification rules, the model can use a predefined dictionary or classification system to match and replace the eigenvalue. Through this standardization processing, the eigenvalues in the tourist interest point data will have a consistent expression form and content, thus improving the consistency and comparability of the data. This consistency not only helps with subsequent data aggregation, statistics and analysis work, but also enhances the user experience when querying and browsing interest point data. At the same time, the standardization processing also provides strong support for the automation and intelligence of tourism services. For example, in scenarios such as itinerary planning, resource allocation and cost accounting, the system can more accurately understand and process interest point data, thus providing more accurate and personalized services. In addition, the preset eigenvalue standardization rules can be flexibly adjusted and extended according to actual business needs to adapt to the changing market environment and user requirements.

[0074] In the above embodiments and optional embodiments of this application, feature extraction and standardization of multi-source heterogeneous data are performed using a preset multimodal model, which can efficiently integrate data resources from different sources and in different formats, solving the problem of difficult unstructured data processing in traditional methods; by constructing a unified and hierarchical preset feature value system, systematic identification and structured output of multi-dimensional tourism feature values ​​are achieved, significantly improving the completeness and accuracy of point-of-interest (POI) data; by constructing an associated storage structure of a general feature master table and a dedicated feature sub-table, accurate splitting and structured storage of POI data in tourism scenarios are achieved, which not only meets the unified management requirements of all-dimensional feature values, but also supports flexible querying and dynamic updates according to business dimensions; through a dynamic update mechanism combined with incremental learning, timely updates of POI data are achieved, effectively solving the problems of traditional tourism This application addresses the issue of delayed updates to the Points of Interest (POI) database. Through cross-border scenario adaptation, it achieves precise segmentation of cross-border itineraries, ensuring that cross-border itinerary data is correctly segmented and stored according to business rules. A confidence output mechanism enables quantitative assessment of the credibility of feature values, making the data processing process more transparent and traceable. A conflict resolution mechanism automates the resolution of multi-source data conflicts, minimizing human intervention while ensuring data quality. An integrity verification mechanism ensures the completeness of PIO data, preventing the loss of key information. A rationality verification mechanism ensures the correctness of PIO data, preventing abnormal data from being included in the database. A feature value standardization mechanism unifies and standardizes the expression and semantic content of multi-source heterogeneous data, significantly improving the consistency, comparability, and aggregation of the included data. Through the above methods, this application enables intelligent and efficient extraction and structured storage of tourism PIO data, effectively improving the construction efficiency and data quality of the tourism PIO database. In some specific tests, the extraction efficiency of point of interest data is more than 80% higher than that of manual extraction. It supports batch processing of tens of thousands of data points, with a data standardization rate of ≥95%. The update delay of point of interest data and corresponding feature values ​​is ≤2 hours, which effectively solves the problem of lagging updates in traditional tourism point of interest databases.

[0075] See Figure 2 , Figure 2 This is a schematic diagram of an embodiment of the storage medium provided in this application.

[0076] The storage medium 30 stores program data 31, which, when executed by the processor, implements, as follows: Figure 1 The method for constructing a database of tourist points of interest is described.

[0077] The program data 31 is stored in a storage medium 30 and includes several instructions for causing a network device (such as a router, personal computer, or server) or processor to execute all or part of the steps of the methods described in the various embodiments of this application.

[0078] Optionally, the storage medium 30 can be any medium capable of storing program data 31, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), disk, or optical disc.

[0079] See Figure 3 , Figure 3 This is a schematic diagram of the structure of an embodiment of the computer device provided in this application.

[0080] The computer device 40 includes a processor 42 and a memory 41 interconnected. The memory 41 stores a computer program, and when the processor 42 executes the computer program, it implements, for example, Figure 1 The described method for constructing a tourist points of interest database. The memory 41 may include the storage medium 30, or it may be another separately developed memory.

[0081] Unlike existing technologies, this application discloses a method for constructing a tourism point of interest (POI) database, a storage medium, and computer equipment. By using a pre-defined multimodal model to extract features and standardize multi-source heterogeneous data, it efficiently integrates data resources from different sources and formats, solving the problem of difficult unstructured data processing in traditional methods. By constructing a unified, hierarchical pre-defined feature value system, it achieves systematic identification and structured output of multi-dimensional tourism feature values, significantly improving the completeness and accuracy of POI data. By constructing an associative storage structure of a general feature master table and dedicated feature sub-tables, it achieves accurate splitting and structured storage of POI data in tourism scenarios, satisfying the unified management needs of all-dimensional feature values ​​while supporting flexible querying and dynamic updates based on business dimensions, thus achieving efficient and flexible storage of POI data. Through the above-described method of constructing a tourism POI database, intelligent and efficient extraction and structured storage of tourism POI data can be achieved, effectively improving the construction efficiency and data quality of the database, and contributing to the upgrading of intelligent management and service levels in the tourism industry.

[0082] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the storage medium embodiments and computer device embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0083] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method of constructing a tourist point of interest library, characterized by, include: Multi-source heterogeneous data is input into a preset multimodal model to generate point of interest (POI) data. The preset multimodal model uses a preset feature value system to standardize the generation of feature values. The preset feature value system includes general feature categories and multiple dimension-specific feature categories. The POI data includes general feature values ​​and specific feature values. The general feature values ​​are output according to the general feature categories, and the specific feature values ​​are output according to the specific feature categories of the corresponding dimensions of the POI data. The common feature value corresponding to the same point of interest data is stored as a data entry in the common feature master table, and the specific feature value corresponding to the same point of interest data and the unique identifier of the point of interest in the common feature value are associated and stored as data entries in the specific feature sub-table corresponding to the specific feature category, thus obtaining the tourism point of interest database.

2. The tourist point of interest library construction method of claim 1, wherein, The method for constructing the tourism point of interest database also includes: Real-time monitoring of data source changes, and incremental generation of changed feature values ​​through the preset multimodal model; A contrastive learning algorithm is used to compare the changed feature values ​​with the corresponding feature values ​​in the tourism interest database, identify and update the feature values ​​that need to be changed; When the change in a feature value reaches the preset update threshold, the corresponding feature value is reported.

3. The tourist point of interest library construction method of claim 1, wherein, The method for constructing the tourism point of interest database also includes: When the point of interest data contains border crossing information or port information, the point of interest data is split into before and after the border crossing using the preset multimodal model.

4. The tourist point of interest library construction method of claim 1, wherein, The preset multimodal model outputs the confidence scores corresponding to each feature value of the interest point data, as well as the overall confidence score of the interest point data. Specifically, when the data source explicitly records the feature value, the confidence level is 1; when the feature value of the point of interest is obtained through context inference, the confidence level is a preset inference confidence level; and the overall confidence level is the average of the confidence levels of all feature values ​​in the point of interest data.

5. The tourist point of interest library construction method of claim 1, wherein, The method for constructing the tourism point of interest database also includes: When feature values ​​of the same point of interest extracted from multiple sources conflict, a weighted vote is performed based on the confidence level output by the preset multimodal model. Feature values ​​whose confidence level falls within a preset high confidence range after weighted voting are marked as pending verification if they do not fall within the preset high confidence range.

6. The tourist point of interest library construction method of claim 1, wherein, After inputting multi-source heterogeneous data into a preset multimodal model to generate point-of-interest data, the method further includes: Determine whether the point of interest data has missing feature values, and mark the missing feature values ​​as to be supplemented; Based on common sense about tourism scenarios, the system uses a preset multimodal model to determine whether the feature values ​​of the point of interest data are reasonable, and labels unreasonable feature values ​​as abnormal.

7. The tourist point of interest library construction method of claim 1, wherein, Before inputting multi-source heterogeneous data into a preset multimodal model to generate corresponding interest point data, the method further includes: constructing the preset feature value system; The feature values ​​of the general feature category include: the unique identifier of the point of interest, the name of the point of interest, the data source identifier, the city, the price, and the information confidence level; The dimensions of the exclusive feature categories include dining, accommodation, transportation, sightseeing, shopping, and entertainment. The feature values ​​for the dining dimension include: dining time, food type, and flavor profile. The feature values ​​for the accommodation dimension include: accommodation level and room type. The feature values ​​for the transportation dimension include: departure and destination, vehicle type, additional fees, transportation type, and stops along the way. The feature values ​​for the sightseeing dimension include: sightseeing name and fee type. The feature values ​​for the shopping dimension include: shopping name, shopping features, and fee type. The feature values ​​for the entertainment dimension include: entertainment name and fee type.

8. The tourist point of interest library construction method of claim 1, wherein, The multi-source heterogeneous data includes at least image data and text data; the preset multimodal model includes at least an image processing sub-model and a text processing sub-model. The image processing sub-model includes a convolutional neural network module and a deep neural network module based on a self-attention mechanism; the convolutional neural network module is used to extract features from image data and identify text regions in the image; the deep neural network module based on a self-attention mechanism is used to perform text understanding and information extraction on the identified text regions and output the interest point data. The text processing sub-model includes a bidirectional long short-term memory network module and a conditional random field module; the bidirectional long short-term memory network module is used to bidirectionally encode text data to generate a context representation sequence; the conditional random field module is used to perform named entity recognition on the context representation sequence and output the point of interest data.

9. A storage medium having stored thereon program data, characterized in that When the program data is executed by the processor, the steps of the method for constructing a tourism point of interest database as described in any one of claims 1 to 8 are implemented.

10. A computer device, comprising: It includes an interconnected processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the tourist points of interest database construction method as described in any one of claims 1 to 8.

Images