An inbound tourism multi-source big data integration and intelligent analysis method

Abstract

The application discloses a kind of inbound tour multi-source big data integration and intelligent analysis method, the method is through to the acquisition of multi-source data, semantic analysis, fusion analysis and service generation etc. Process is cooperatively designed, realizes the closed loop processing of multi-source data integration and intelligent analysis, each processing process is sequentially linked according to the technical path of " knowledge construction-real-time perception-fusion decision-service generation", and through the introduction feedback update mechanism, model parameters or fusion strategy is dynamically adjusted, so that the robustness and accuracy of multi-source data fusion analysis in complex cross-cultural scene are improved while guaranteeing the real-time of system.The method of the application is a set of inbound tourism scene design, deep fusion cultural semantics and real-time response capability big data integration and intelligent analysis scheme, can promote inbound tourism data analysis to the evolution of depth intelligent, real-time personalized direction.

Description

Technical Field

[0001] This invention belongs to the field of data processing and artificial intelligence technology, specifically relating to a method for integrating and intelligently analyzing multi-source big data of inbound tourism, which is particularly suitable for tourist behavior analysis and service recommendation in inbound tourism scenarios. Background Technology

[0002] Against the backdrop of global digital transformation in the tourism industry and the upgrading of cultural and tourism consumption, leveraging data intelligence to improve the service quality, management efficiency, and marketing precision of inbound tourism has become a core issue for industry development. Inbound tourist behavior is characterized by cross-cultural aspects, short itineraries, diverse scenarios, and heterogeneous data sources. Traditional analytical methods, such as those relying on single data sources or offline questionnaires, are no longer sufficient to meet the demands for real-time insights, precise services, and efficient management. Therefore, integrating multi-source, heterogeneous, and multimodal data and achieving intelligent analysis constitutes the main direction of current smart tourism technology evolution.

[0003] To achieve a precise understanding of tourists, especially inbound tourists, technological practices have evolved from early single-point information management to the current systematic integration and mining of multimodal data. The core of this process lies in imitating the way humans use multiple senses to perceive the world, integrating data from different sources and formats through technological means to construct a comprehensive and dynamic tourist profile. These data modalities mainly include: (1) textual and semantic data, such as multilingual online reviews, travelogues, and search keywords, which contain tourists' emotional tendencies, cultural interests, and consumption concerns; (2) spatiotemporal trajectory data, such as mobile phone signaling, GPS positioning, and Wi-Fi probe data, which objectively record tourists' movement patterns, hotspot areas, and stay patterns; (3) consumption transaction data, such as cross-border payment records and OTA (Online Travel Agent) platform orders, which directly reflect tourists' consumption capacity and preference structure; and (4) environmental and scene data, such as real-time visitor flow monitoring videos of scenic spots and environmental information such as temperature, humidity, and congestion collected by IoT sensors, which provide conditions for enhancing the tourist experience.

[0004] Currently, the technological systems supporting this integration process primarily rely on cutting-edge advancements in data science and artificial intelligence. Natural language processing (NLP) technology has been used to extract themes, sentiments, and fine-grained preferences from massive amounts of unstructured multilingual text; spatiotemporal data mining algorithms can identify meaningful access patterns, travel routes, and clustering events from continuous trajectory points. To achieve unified modeling and correlation analysis of these multi-source, heterogeneous, and multimodal data, multimodal fusion and knowledge graph technologies become crucial. By constructing a tourism knowledge graph with nodes representing "tourists-attractions-activities-consumption" and utilizing techniques such as graph neural networks for relational reasoning, deep semantic connections can be established between data from different modalities, thus going beyond simple data juxtaposition to achieve information complementarity and enhancement.

[0005] In practical industrial applications, some existing analytical systems have begun to explore this path. For example, Chinese patent technology with publication number CN118656596B provides a multi-source heterogeneous scene data fusion analysis management system and method. This solution discloses a highly general data fusion analysis management system, the core of which lies in a fusion analysis method: first, feature extraction and classification of multi-source heterogeneous data are performed to construct a "common information set" and a "dissimilar information set"; second, a unified spatiotemporal reference system is determined to perform spatiotemporal alignment and fusion of the data; finally, decision analysis is performed based on the fusion results and confidence levels are assessed. This system and its method represent a mainstream technical approach to processing cross-domain, heterogeneous data to achieve comprehensive decision-making, and its systematic fusion framework provides a methodological foundation for data integration in multiple fields, including tourism. However, such general-purpose data fusion systems still exhibit significant mismatches and limitations when dealing with the vertical scenario of inbound tourism. Most existing systems focus on spatiotemporal alignment and structured processing at the data level, failing to deeply address the three core challenges in this scenario: 1. Lack of cross-linguistic and cross-cultural semantic understanding; existing systems typically lack the ability to perform deep semantic modeling in multilingual and cross-cultural contexts, making it difficult to identify the impact of cultural differences on tourist behavior and intentions, resulting in biased analysis of tourist preferences and satisfaction.

[0006] 2. Insufficient real-time dynamic response capability; the existing architecture is mostly based on batch processing mode, which cannot meet the high dynamic needs of inbound tourists with tight itineraries and rapid shifts in points of interest, resulting in excessive delays from data collection to service feedback.

[0007] 3. Weak scenario adaptability; when integrating multimodal data, general frameworks often ignore the unique data quality fluctuations, semantic conflicts and cultural background differences of inbound tourism, resulting in poor robustness and low credibility of the integration results.

[0008] Despite progress in general data processing and multimodal fusion, existing systems have not yet achieved truly cross-language, cross-cultural, and real-time dynamic intelligent analysis capabilities in the inbound tourism sector. How to deeply integrate these general frameworks with the unique needs of the specific scenario of inbound tourism to achieve deeper and more scenario-adaptive intelligent analysis has become a crucial stage in technological development that urgently needs breakthroughs.

[0009] Currently, fusion analysis technologies for multi-source heterogeneous and multimodal data have spawned a series of systems built on general frameworks. While these systems provide a basic path for multi-source data integration, they reveal the following fundamental limitations in terms of system design philosophy, core architecture, and key capabilities when dealing with the highly dynamic, culturally specific, and data heterogeneous vertical scenario of "inbound tourism": 1. Superficial semantic understanding of scenarios lacks quantitative cultural logic support. Existing systems typically focus on the spatiotemporal alignment of multi-source data, and their analytical depth is limited by general NLP (Natural Language Processing) models. For inbound tourism scenarios, simply achieving data-level alignment is insufficient to reveal the cross-cultural intentions behind behavior. Because existing system architectures lack calculable and reasonable quantitative cultural dimension models, the systems cannot logically link tourists' textual intentions with deep cultural preferences. Furthermore, traditional models often employ unidirectional feature extraction and lack effective cross-attention mechanisms to integrate domain knowledge, resulting in analytical results that remain at the statistical description level and fail to address the true semantic motivations of tourists.

[0010] 2. Batch processing architecture suffers from high latency and lacks state machine management for dynamic behavior. From an implementation perspective, existing systems generally rely on batch computation or simple pipeline task scheduling. Inbound tourists have tight schedules, and their points of interest shift rapidly with the environment, requiring extremely high system response speeds. However, existing technologies lack event-driven stream processing engines and in-memory state management mechanisms. Because they cannot capture the state machine transition from "quickly moving through" to "stopping to observe" in real time, and traditional rule matching is inefficient, the system struggles to complete a closed-loop response to new data in sub-second increments, failing to support truly "accompanying" real-time services.

[0011] 3. The robustness of multi-source fusion mechanisms is insufficient, neglecting the interference of spatiotemporal offsets. In real-world deployment environments, multi-source data is often accompanied by GPS drift and semantic noise. Existing fusion methods typically assume that the information sources are statically reliable, lacking online feature evaluation of data quality. More importantly, existing fusion algorithms do not consider the impact of spatiotemporal distance on the attenuation of evidence confidence, leading to outdated or contradictory evidence from the far field severely interfering with the fusion decision results. This significantly reduces the reliability of the system's output conclusions in complex noisy environments.

[0012] 4. Limited cross-cultural adaptation methods hinder deep intervention at the model's underlying level. The core challenge of inbound tourism is cultural diversity. Existing general systems often handle cultural differences at a superficial level of "translation" or "label filtering." This "one-size-fits-all" approach typically only patches the output and fails to delve into the feature distribution at the model's core. Lacking techniques like parameter generation networks and conditional layer normalization, the system cannot adjust the model's neuron activation state in real-time based on tourists' fine-grained, dynamic cultural profiles. This results in significant cultural inconsistencies and semantic gaps in the generated recommendations and service information.

[0013] In summary, the limitations of existing systems in dealing with inbound tourism scenarios stem from system-level design flaws: a lack of quantitative knowledge-driven approaches at the semantic level, a lack of dynamic state machine management at the architectural level, a lack of spatiotemporal robustness at the integration level, and a failure to achieve parameter-level cultural adaptation at the cognitive level. These shortcomings collectively constitute the core bottleneck for the current technological upgrade towards deep intelligence and real-time personalized services. Summary of the Invention

[0014] To address the issues of insufficient semantic understanding, limited real-time response capabilities, and poor cross-cultural adaptability in multi-source heterogeneous data analysis for inbound tourism scenarios, this invention provides a method for multi-source big data integration and intelligent analysis in inbound tourism. This method achieves closed-loop processing of multi-source data integration and intelligent analysis by collaboratively designing the processes of multi-source data collection, semantic parsing, fusion analysis, and service generation. Each processing step is sequentially connected according to the technical path of "knowledge construction - real-time perception - fusion decision-making - service generation," and a feedback update mechanism is introduced to dynamically adjust model parameters or fusion strategies. This ensures the real-time performance of the system while improving the robustness and accuracy of multi-source data fusion analysis in complex cross-cultural scenarios.

[0015] A method for integrating and intelligently analyzing multi-source big data on inbound tourism includes the following steps: (1) Collect multi-source heterogeneous data of inbound tourists and convert data from different data sources into a unified event stream with timestamps. The event stream includes text semantic events, spatiotemporal behavioral events and consumption behavior events; (2) Call the semantic parsing model to process the semantic events of the text, and constrain the parsing results based on the domain knowledge model during the processing, and map the unstructured text into a structured semantic feature vector as the input for subsequent state evaluation judgment and fusion processing; (3) Based on the cultural feature mapping model, the semantic feature vector is converted into at least one dynamic cultural feature vector with continuous values, which serves as a condition parameter for distinguishing different semantic feature vector processing methods and participates in the subsequent decision generation process; (4) For different event streams, calculate the confidence weight associated with each data source based on the event's timeliness parameter, historical consistency parameter and integrity index, so as to reflect the reliability of the corresponding data source in the current state. (5) Based on the confidence weight, the improved DS (Dempster-Shaffer) evidence theory inference engine is invoked to perform uncertainty fusion processing on the state judgment results from different data sources; during the fusion process, a spatiotemporal sensitivity correction factor is introduced based on the time difference and spatial displacement of the event to compensate for the attenuation of the support of each piece of evidence; by performing joint normalization processing on the support and conflict of different state judgment results, a comprehensive judgment result of the tourist's current state with high trust is generated; (6) The comprehensive judgment result and the dynamic cultural feature vector are used as input. The scaling factor and translation bias generated by the PGN (parameter generation network) based on the dynamic mapping of cultural features are dynamically modulated through the conditional layer normalization method to generate service decision information that matches the background of tourists in terms of cultural connotation and expression style.

[0016] Furthermore, in step (1), the CEP (Complex Event Processing) engine is used to perform single-point access and time-series alignment of heterogeneous real-time data streams, including GPS trajectories and mobile payments. A sliding window mechanism is introduced, and by performing pattern matching on the continuous data streams within the window, the scattered raw data signals are abstracted in real time into standardized event objects with clear business semantics. At the same time, a lightweight, in-memory, atomically updatable visitor context object is maintained for each active visitor. The data structure of this object not only includes user ID, location, and intent vector, but also integrates the state bits of a finite state machine. When a new event arrives and updates the visitor context object, a lightweight rule engine based on the Rete algorithm is used to instantly trigger the corresponding condition judgment and action execution through a topology matching mechanism, without the need for frequent database polling.

[0017] Furthermore, in step (2), the domain knowledge model uses OWL (Web Ontology Language) to uniformly model domain entities during the knowledge modeling stage. Based on the Hofstede cultural dimension theoretical framework, it introduces quantitative attribute vectors that reflect cultural differences for entities. It calls the relation extraction model based on the Transformer architecture to deeply mine multilingual travelogues and uses a regression prediction model to transform cultural tendencies in unstructured texts into continuous values ​​between 0 and 1, thereby transforming abstract cultural features into computer-computable logical attributes. The "entity-relationship-attribute" triple data after the above extraction and quantification are stored in the graph database in the form of RDF (Resource Description Framework), forming a cross-cultural knowledge graph that supports deep reasoning.

[0018] Furthermore, in step (2), the semantic parsing model is based on XLM-RoBERTa (cross-language RoBERTa model) and integrates a knowledge retrieval module to perform deep processing on the text generated by tourists. In the processing, text entities are first identified and linked to the knowledge graph. Then, the contextual representation of the text is used as the query, and the quantized attribute vector retrieved from the knowledge graph is used as the key and value. A structured semantic feature vector containing intent and cultural features is generated through a cross-attention mechanism.

[0019] Furthermore, the cultural feature mapping model in step (3) is based on the generation mechanism of dynamic cultural portrait and parameterized condition modulation. In the portrait construction stage, a dynamic cultural feature vector with continuous values ​​is generated for each tourist in real time through an encoder network based on the Transformer architecture. This encoding process goes beyond static nationality labels and deeply integrates multi-source behavioral signals, including the embedded representation of real-time behavioral event sequences, structured semantic feature vectors containing intentions, quantitative attribute vectors reflecting cultural differences, and implicit feedback signals generated by user interaction with the system through a self-attention mechanism. The final output dynamic cultural feature vector constitutes a digital portrait of tourist cultural preferences.

[0020] Furthermore, in step (4), for each data modality of the event stream, a lightweight confidence assessment model is first deployed to extract quality feature indicators in real time. For physical sensing modalities (such as GPS), the number of visible satellites, HDOP (horizontal precision factor), and signal-to-noise ratio are extracted. For text semantic modalities, the prediction entropy of the language model and the Softmax probability distribution of the parsing results are extracted. Then, these quality feature indicators are input into a multilayer perceptron network to calculate and generate real-time confidence weights between 0 and 1 online, which are used to reflect the reliability of the corresponding data source in the current state.

[0021] Furthermore, in step (5), a modified BPA (Basic Probability Assignment) value is used during the fusion of state judgment results from different data sources using the DS evidence theory inference engine to ensure that "strongly spatiotemporally correlated" data that is closer to the current time and location receives higher decision weights during the fusion process; the expression for the modified BPA value is: Where: for any data source i A BPA function is constructed by combining confidence weights and support bias. and Data sources i any subset A BPA values ​​before and after correction α and βThese are the preset time decay coefficient and spatial weighting coefficient, respectively. and These represent the time offset and spatial location deviation of the data, respectively.

[0022] Furthermore, in step (6), the intermediate layer parameters of the service generation model are dynamically modulated through conditional layer normalization, that is, for any layer of input features in the model decoder architecture... x The following expression is used to perform conditional normalization: in: Input features x The result after conditional normalization. This indicates element-wise multiplication. and Representing the input features respectively x The mean and standard deviation, As a dynamic cultural feature vector, This is the spatiotemporal context vector at the current decoding moment. and They represent and The scaling factor and translation bias are input into the PGN (Parameter Generation Network) and generated in real time.

[0023] A computer device includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the above-mentioned method for integrating and intelligently analyzing multi-source big data of inbound tourism.

[0024] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for integrating and intelligently analyzing multi-source big data on inbound tourism.

[0025] Based on the above technical solution, the present invention can solve the following technical problems: 1. The computability and logical reasoning issues of cross-cultural intentions. Addressing the problem that existing systems' analysis of tourist behavior remains at a superficial statistical level and fails to delve into deep motivations, this invention constructs a domain knowledge model that integrates quantitative cultural dimension attributes. This model transforms ambiguous cultural symbols, tourism intentions, and consumption behaviors into structured semantic tuples that are computationally computable and reasonable. It utilizes cross-attention mechanisms to achieve a deep fusion of textual features and prior cultural knowledge, thereby solving the cognitive leap from describing "what happened" to explaining "what the intention is."

[0026] 2. Sub-second response and state management issues for perception and decision-making in highly dynamic scenarios. To address the mismatch between the response latency of batch processing architectures and the high-frequency movement characteristics of inbound tourists, this invention constructs an event-driven real-time perception architecture. By introducing a stream processing engine and in-memory context objects, it achieves real-time capture of tourist behavior state machine transitions. Simultaneously, by optimizing the rule matching mechanism, it solves the problem of real-time triggering of business rules under large-scale concurrent event streams, thereby compressing the closed-loop response time of "perception-decision-execution" to the sub-second level.

[0027] 3. Conflict Resolution and Spatiotemporal Robustness of Multi-Source Fusion in Complex Noise Environments. To address the decision distortion caused by fluctuations in the quality of multi-source data and contradictory evidence, this invention establishes an uncertainty fusion mechanism with quality awareness. This mechanism not only requires online evaluation of the original reliability of each information source, such as GPS accuracy and semantic prediction entropy, but more importantly, it addresses the impact of spatiotemporal offsets on the fusion weights. By introducing a spatiotemporal sensitivity correction factor, the contribution of different pieces of evidence in the fusion process is dynamically adjusted, thereby outputting a comprehensive judgment result with high reliability and robustness even in the presence of noise and information conflicts.

[0028] 4. Deep Adaptation and Low-Level Parameter Modulation in Cross-Cultural Service Generation. Addressing the limitation of existing cultural adaptation methods, which are confined to surface-level label filtering and fail to bridge deep cultural gaps, this invention achieves deep cultural adaptive generation at the model level. By constructing dynamic cultural profile vectors and combining them with real-time spatiotemporal context, a parameter generation network is used to modulate the intermediate layer parameters of the generation model in real time. This approach aims to solve the problem of changing the feature distribution from the model's underlying layers, ensuring that the generated service content naturally aligns with tourists' cultural backgrounds in terms of expression style and logical emphasis, thus achieving true deep cultural alignment.

[0029] By overcoming the aforementioned interrelated technical challenges, this invention aims to advance inbound tourism data analysis technology from a general, lagging data processing stage to a deep, agile, and reliable stage of intelligent cognition and service decision-making in inbound tourism scenarios. Compared with existing technologies, it offers the following beneficial technical effects: 1. This invention achieves a cognitive leap from superficial semantic matching to quantitative cultural intent reasoning. Existing technologies focus on spatiotemporal alignment, lacking in-depth analysis of user intent in cross-cultural contexts. This invention constructs a knowledge graph incorporating Hofstede's six-dimensional quantitative attributes and combines it with a cross-attention mechanism to transform abstract cultural features into computable spatiotemporal context vectors. This improvement enables the system not only to identify where tourists "went" but also to accurately discern deep motivations such as "pursuing uniqueness" by calculating the association weights between textual representations and cultural dimension vectors. This provides a solid cognitive foundation for achieving accurate "semantic alignment" in a multicultural context and significantly improves the accuracy of intent recognition.

[0030] 2. An architectural transformation from high-latency batch processing to low-latency real-time response has been achieved. Addressing the highly dynamic nature of inbound tourism scenarios, this invention abandons the traditional "batch processing" architecture and introduces a sliding window mechanism and a lightweight rule engine based on the Rete algorithm. By maintaining a memory-state context object integrating limited state information at edge nodes, the system achieves an extremely short closed loop of "perception-state transition-topology matching-execution." This architectural transformation improves response speed to sub-second levels, solving the problem of delayed recommendation information caused by processing latency, and truly achieving a service leap from "post-event analysis" to "real-time support during the event."

[0031] 3. Significantly enhanced decision robustness in complex environments. Existing fusion methods are vulnerable to data quality fluctuations (such as GPS drift and text noise). This invention outputs confidence scores online through a multilayer perceptron regression network and innovatively introduces a spatiotemporal sensitivity correction factor into the DS evidence theory. This mechanism can automatically identify and reduce the weight of outdated or far-field interference data, ensuring that decisions are always based on the most relevant and reliable information sources. Through this explicit modeling and dynamic weighting of uncertainty, the system can maintain a high level of decision credibility and stability even in real-world scenarios with contradictory data and severe noise interference.

[0032] 4. This invention achieves deep cultural adaptation from surface labels to the underlying parameters of the model. Existing systems often rely on static labels for simple content filtering, which is insufficient to handle fine-grained cultural differences. This invention utilizes a parameter generation network to receive dynamic cultural vectors and spatiotemporal context vectors, and directly intervenes in the intermediate layer parameters of the generation model through conditional layer normalization technology. This "parameter modulation" rather than "text replacement" technique enables the generation model to automatically adjust its expression style, logical weights, and interaction strategies according to the cultural background of tourists. This not only improves the cultural relevance of recommended content and user satisfaction, but also solves the semantic gap problem in cross-cultural communication at the underlying technical level, demonstrating significant commercial potential and social benefits. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the multi-source big data integration and intelligent analysis method for inbound tourism of the present invention. Detailed Implementation

[0034] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] like Figure 1 As shown in the figure, this embodiment provides a method for integrating and intelligently analyzing multi-source big data on inbound tourism, including the following steps: (1) Event-driven data acquisition and preprocessing.

[0036] This invention collects multi-source heterogeneous data of inbound tourists through an event-driven data access method, and converts data from different data sources into a unified event stream format with timestamps. The event stream includes text semantic events, spatiotemporal behavioral events, and consumption behavior events. During the event generation process, format mapping, time synchronization, and abnormal noise suppression are performed on the original data.

[0037] The system's real-time processing capability is built upon a layered data processing architecture, with the core objective of transforming high-throughput raw data streams into low-latency, high-semantic-value business events. Addressing the highly dynamic nature of inbound tourist behavior, this invention constructs an event-driven real-time perception system on edge computing nodes. Through deep collaboration between the stream processing engine and memory-based state management, it achieves sub-second response times from raw data acquisition to business command triggering.

[0038] At the data access layer, the system no longer employs traditional batch processing logic. Instead, it utilizes a Complex Event Processing (CEP) engine to perform single-point access and time-series alignment of heterogeneous real-time data streams such as GPS trajectories and mobile payments. Its core technical approach lies in introducing a sliding window mechanism. By performing pattern matching on continuous data streams within a window, it abstracts fragmented, low-value raw signals into standardized "business event" objects with clear business semantics in real time. For example, the system can use sliding window analysis to determine whether tourists have stayed at a specific cultural landmark for longer than a preset threshold, thereby instantly generating high-order event signals carrying time, location, and subject context, and publishing them to a message queue for further decision-making.

[0039] To support real-time, accompanying services, the system maintains a lightweight, atomically updatable, in-memory Visitor Context Object (TCO) for each active visitor. This object's data structure not only includes the user ID, location, and intent vector, but also, crucially, integrates the state bits of a finite state machine. This design allows the system to record the dynamic transitions of visitor behavior states, such as the state transition from "quickly moving through" to "staying to observe," thereby accurately depicting their current stage of the visit.

[0040] When a new event arrives and updates the context object, the embedded lightweight rule engine based on the Rete algorithm starts working simultaneously. This invention pre-encodes complex business logic, such as cross-cultural push trigger conditions, into production rules and constructs them as an efficient directed graph topology network. Whenever the memory context is updated, the system asserts it as a "fact" in the rule engine's working memory, instantly triggering the corresponding condition judgment and action execution through a topology matching mechanism, without frequent database polling. This closed-loop mechanism of "perception-state transition-topology matching-execution" greatly reduces end-to-end latency, providing tourists with a truly real-time personalized service experience.

[0041] This implementation uses the CEP engine to achieve event-driven data transformation. The CEP engine predefines various atomic event patterns, such as a tourist entering / leaving a specific geofence or completing a transaction. Taking the processing of tourist location trajectories as an example, the system predefines a state machine outside the geofence. Inside the fence The engine continuously monitors real-time GPS coordinate streams in two basic states. and through spatial predicates The real-time determination of the positional relationship between coordinate points and geofences, and the specific processing logic and execution process are as follows: State transition determination: When the system detects the current coordinates Outside the fence (status is) ), and the coordinates of the next immediate time step. Once the predicate determines that the area is within the fence (i.e., the predicate returns true), the system will automatically drive the state machine from... Towards A state transition occurs.

[0042] Structured event instantiation: The system will synchronously trigger the structured event object at the moment a state transition occurs. The instantiation process of the object encapsulates the event type, subject identifier, millisecond-level timestamp, precise coordinates and geofence identifier, transforming the originally isolated coordinate point into a semantic object with clear business meaning, and then publishes it to a unified distributed event message bus such as Kafka.

[0043] This design ensures that the system can accurately extract key dynamic information with business decision-making value from high-frequency, low-value raw signals. In actual deployment, the geofence can be set to a circle or polygon according to the scale of the scenic area, such as 50 meters to 500 meters. Combined with millisecond-level timestamp verification, it ensures the consistency of event generation time and the immediacy of response.

[0044] To achieve sub-second response times, the system uses a managed state backend based on the Apache Flink framework to maintain a context object in memory for each active visitor that supports atomic updates. This object serves as a real-time "digital mirror" of the tourist within the system, dynamically recording their current geographical location, multi-dimensional intent probability distribution, and recent behavioral event sequence. For massive concurrent event streams, the system's processing is divided into the following two key stages: Atomic update phase based on keyed state: whenever a new business event is generated. When the data flows into the processing layer, the system will synchronously call the state update function. This process fully utilizes Flink's keyed state mechanism, ensuring that all events of a specific visitor are serialized within the same operator instance by hashing the visitor identifier. This design avoids lock contention in traditional concurrency control, thereby achieving lock-free atomic merging of context information.

[0045] The logic instant matching stage based on the Rete algorithm: After the context is updated, the system uses the embedded Rete algorithm rule engine to make logical judgments. The system pre-constructs complex cross-cultural business logic (for example, if the "cultural authenticity" score in the tourist intent vector is >0.8 and the current location is within the "intangible cultural heritage block", then in-depth explanation push is triggered) into an efficient directed cyclic filtering network.

[0046] Specifically, context object Any attribute change will be asserted as a new "fact" in the working memory of the Rete network; thanks to the node sharing and state caching characteristics of the Rete algorithm in the pattern matching process, the system can instantly complete the condition matching and trigger the corresponding action function without traversing redundant business rules, generating highly timely service instructions and sending them to the execution link.

[0047] (2) Knowledge-enhanced semantic event parsing.

[0048] This invention processes text semantic events by calling a semantic parsing model, and constrains the parsing results based on a domain knowledge model during the parsing process, mapping unstructured text into structured semantic feature vectors, which serve as inputs for subsequent state evaluation and fusion processing.

[0049] As a structured knowledge foundation for the system, the construction process of a domain knowledge graph deeply integrates formal ontology modeling, quantitative cultural dimension theory, and multi-source heterogeneous data mining technology, providing logical support for subsequent semantic parsing and service generation. To address the superficial semantic understanding problem in inbound tourism scenarios caused by the lack of a computable and reasonable domain knowledge model, this invention first constructs a domain knowledge graph that integrates cultural dimensions, and then uses this as a basis to drive a fine-grained semantic parsing process.

[0050] In the knowledge modeling phase, the system does not simply pile up information, but adopts ontology engineering methods and uses web ontology language to uniformly model entities such as "attractions," "cultural concepts," and "tourist intentions." Its core improvement lies in introducing quantitative attribute vectors that reflect cultural differences to the entities. (That is, based on the Hofstede six-dimensional model). During the construction process, the system utilizes a relation extraction model based on the Transformer architecture to deeply mine multilingual travelogues, and combines this with a regression prediction model to transform cultural tendencies in unstructured text into continuous values ​​between 0 and 1, thereby converting abstract cultural features into computer-computable logical attributes. Finally, these triplet data are stored in a graph database in the form of a resource description framework, forming a cross-cultural semantic network that supports deep reasoning.

[0051] After acquiring the aforementioned knowledge support, the system performs deep processing on the visitor-generated text using a parsing model that integrates a knowledge retrieval module based on a cross-language model-robust optimization BERT pre-training method. The core technology of this parsing process lies in the application of a cross-attention mechanism: the system first identifies text entities and links them to a knowledge graph, then uses the contextual representation of the text as a query, and uses the retrieved... The attention weights are calculated using the attribute vectors as keys and values, respectively, by the following formula: in: Q This represents the contextual representation vector of the tourist's input text after being encoded by a multilingual pre-trained model. It carries the tourist's current immediate intent, such as semantic features in search terms or reviews. K It represents the relevant cultural entities and their feature vectors retrieved from the domain knowledge graph. It serves as a matching benchmark to measure the strength of the association between the input text and specific cultural background knowledge. V This represents the quantified cultural dimension attribute value corresponding to the key (i.e., the aforementioned). Vectors are the substantial cultural knowledge information that is ultimately injected into the model and used to enhance semantic understanding. The dimension of the input vector is represented by dividing by Scaling is performed to prevent the dot product from becoming too large when the vector dimension is high, thereby ensuring the stability of the gradient of the Softmax function and improving the accuracy of semantic fusion.

[0052] In this way, the model not only matches at the semantic level, but also deeply integrates text features with quantified cultural background at the underlying level. The final output is no longer isolated keywords, but structured semantic tuples containing intent and cultural features. For example, when the system analyzes tourists' preference for "niche teahouses", it can automatically associate it with the cultural intent of "pursuing uniqueness" and give the corresponding confidence level. This process provides a culturally deep underlying input for subsequent real-time perception and adaptive decision-making.

[0053] The system's construction begins with formal ontology modeling using a web ontology language to describe logic, aiming to clearly define the hierarchical structure and complex logical relationships of core entities such as "tourists," "attractions," "cultural concepts," and "tourism activities." A key improvement in this implementation is the introduction of computable, quantifiable cultural attribute vectors for these entities. Based on theoretical frameworks such as Hofstede's cultural dimensions, the system defines a series of continuously valued numerical attributes for relevant entity classes. For example, the "tourism activity" class defines the data attribute `individualism_index` to quantify the individualistic tendencies embodied in the activity; the "attraction" class defines `uncertainty_avoidance_index` to characterize its attractiveness to tourists seeking predictable itineraries. This quantitative design transforms abstract cultural differences into dense vectors that can be processed by computers. .

[0054] After completing the top-level design of the ontology structure, the system enters the knowledge graph filling stage. This implementation method constructs an integrated pipeline of "multimodal extraction - cross-cultural quantification - graph construction" to realize the transformation from raw heterogeneous data to deep semantic knowledge. The specific implementation process is as follows: Collaborative extraction of heterogeneous data: The system utilizes extraction, transformation, and loading processes to extract basic entity attributes (such as geographical coordinates and business hours) from structured databases of hotels and scenic spots. Building upon this, for massive amounts of unstructured corpora such as multilingual travelogues and online reviews, the system invokes a relation extraction model based on the Transformer architecture. This model employs a remote supervision mechanism, using pre-set seed triples to perform automated annotation and alignment in the text corpus, thereby identifying and extracting implicit semantic relationships such as "originating market - preference tendencies - cultural activities."

[0055] Quantitative regression based on extracted entities to perform cultural dimensions: For the abstract entities obtained in the preceding steps, in order to meet the needs of cross-cultural computing, the system needs to numerically model their cultural attributes. For implicit attributes such as social distance preferences that cannot be directly observed, the system constructs a feature mapping regression model, extracts the average word vector of the entity in the multilingual context as the input feature, calculates it through a non-linear mapping layer, and outputs a continuous confidence score between 0 and 1. This quantification process uses mean squared error as the loss function and uses the Adam optimizer to fine-tune it on the labeled dataset to ensure that each newly extracted knowledge item carries a computable cultural metric label.

[0056] Execution of RDF structured storage and persistent reasoning: The "entity-relationship-attribute" dataset, after the above extraction and quantization processes, is uniformly transformed into standard RDF triples and persistently stored in the Neo4j graph database. The storage method utilizes the adjacency matrix characteristics of the graph structure to achieve millisecond-level cross-node path queries. Simultaneously, semantic reasoning is performed using the ontology logic layer to discover potential tourist demand associations through known cultural metric attributes, thereby constructing a dynamic cross-cultural knowledge network with self-evolving capabilities.

[0057] This invention aims to perform deep semantic deconstruction of multilingual text data. Its core lies in constructing a neural semantic parsing model that integrates external cultural knowledge perception capabilities, achieving a mapping from literal meaning to deeper intent. The model uses the cross-lingual pre-trained model XLM-RoBERTa as the basic feature extraction layer. To effectively incorporate cultural context constraints during the encoding process, this implementation integrates a knowledge-aware attention fusion layer into the encoder architecture. The specific implementation process is as follows: Model Entity Alignment and Semantic Linking: The model performs named entity recognition on the input text, and the recognized text fragments are precisely aligned with specific entity nodes in the aforementioned domain knowledge graph through entity linking technology.

[0058] Cross-modal feature retrieval: Based on the alignment results, the system dynamically retrieves the quantified cultural dimension vectors of entity nodes from the knowledge graph. and its associated topological features.

[0059] Dual-path information fusion computation: Utilizing the cross-attention mechanism, the text's own contextual representation sequence is used as the query vector, and the retrieved external knowledge vector is used as the key vector and value vector. By calculating the affinity weights of the two, the external cultural prior knowledge is weighted and fused into the text representation to generate the final, knowledge-enhanced high-order semantic representation vector.

[0060] Based on the knowledge-enhanced semantic representation, the model performs joint task modeling through successive structured output heads, generating semantic tuples containing behavior, modifiers, and cultural motivations. A specific parsing example is as follows: For the text input "finding a quiet, non-touristy temple for meditation", the model, while recognizing the literal intent, combines the cultural associations of "temple, meditation, non-touristy" in the knowledge graph and outputs the following structured tuple: Subject - User, Predicate - Seeking, Object - Temple Activity, Modifier - {Atmosphere: Quiet, Type: Non-touristy}, Cultural Intent - [Spiritual pursuit, preference for solitude].

[0061] Through this mechanism, the system can not only capture the surface-level behavioral descriptions of tourists, but also reveal the abstract intentions and cultural psychological motivations hidden behind the textual context, providing a data benchmark for subsequent accurate decision-making.

[0062] (3) Cultural feature vector mapping.

[0063] Based on a pre-defined cultural feature mapping model, semantic feature vectors are converted into dynamic cultural feature vectors with at least one continuous value. These cultural feature vectors serve as conditional parameters to distinguish different semantic feature vector processing methods and participate in subsequent fusion and generation processes.

[0064] To completely eliminate the semantic gap and service barriers caused by ignoring cultural differences, this invention constructs a generation mechanism based on dynamic cultural profiling and parameterized condition modulation, realizing deep cultural adaptation of service content at the underlying model level.

[0065] During the profile construction phase, the system uses an encoder network based on the Transformer architecture to generate dynamic, continuously valued cultural feature vectors for each tourist in real time. This encoding process goes beyond static nationality labels, deeply fusing multi-source behavioral signals through a self-attention mechanism: including the basic cultural dimension of the customer origin mapped from the knowledge graph, real-time behavioral event embeddings, and the intent vector obtained from the aforementioned semantic parsing, ultimately outputting a low-dimensional dense vector. This constitutes a digital portrait of tourists' cultural preferences.

[0066] This step aims to achieve deep modeling of tourists' personalized characteristics and precise adaptation of service output based on the fusion intent and quality evidence produced in the previous stages. In order to transform scattered behavioral signals into structured knowledge that can guide the generative model, the system first performs the task of constructing a dynamic cultural profile. Specifically, this task is implemented through a deep neural network based on a Transformer encoder, the core function of which is to map heterogeneous behavioral signals to the cultural representation dimension of a higher-order semantic space.

[0067] In this representation construction process, the system first performs feature alignment on multi-source signals to form an input vector that integrates the following four feature dimensions: an embedded representation of real-time behavioral event sequences, a structured intent vector output by the semantic parsing module, and a baseline cultural dimension vector based on tourist origin locations. Implicit feedback signals generated by user interaction with the system.

[0068] The Transformer encoder, through its internal self-attention mechanism, performs deep fusion and abstract extraction on the input vector, ultimately outputting a low-dimensional, dense, dynamic cultural feature vector. This vector, serving as a digital portrait of tourists' real-time cultural preferences, will be continuously updated and stored over time, providing crucial conditional input for subsequent service adaptation.

[0069] (4) Multi-source data quality perception and confidence calculation.

[0070] For different event streams and their corresponding data sources, confidence weights associated with each data source are calculated based on the event's timeliness parameters, historical consistency parameters, and integrity indicators. These weights reflect the reliability of the corresponding information source in its current state.

[0071] To address the issue of unreliable fusion decisions due to data quality fluctuations in real-world tourism environments, this invention significantly improves the system's robustness under complex noise interference by fusing online confidence assessment with improved evidence-based decision-making. The system first constructs a lightweight online assessment model to quantify the confidence of input data in real time. This model extracts multidimensional quality features based on the physical characteristics of different modalities of data. These features are input into a multilayer perceptron regression network, dynamically outputting confidence scores. Furthermore, the model possesses adaptive evolution capabilities, enabling online fine-tuning based on feedback from downstream services regarding recommendation accuracy.

[0072] To address decision-making biases caused by data quality fluctuations and information conflicts from multiple sources in real-world tourism scenarios, this implementation employs a robust fusion strategy based on uncertainty modeling at the decision-making level. The system first establishes a confidence assessment layer to perceive the reliability of different information sources in real time. For each data modality, the system deploys a lightweight confidence assessment model to extract quality feature indicators in real time. For physical sensing modes (such as GPS): extract the number of visible satellites, HDOP, and signal-to-noise ratio to measure positioning accuracy.

[0073] For text semantic modalities: extract the prediction entropy of the multilingual model and the Softmax probability distribution of the parsing results to evaluate the determinism of semantic expression.

[0074] These quality characteristics are fed into a shallow multilayer perceptron network to calculate and generate real-time confidence scores between 0 and 1 online. The model has online learning capabilities and can adaptively fine-tune the evaluation parameters based on user feedback such as click-through rate and dwell time after the decision.

[0075] (5) Multi-source fusion decision based on confidence constraints.

[0076] Based on the confidence weight, the improved DS evidence theory inference engine is invoked to perform uncertainty fusion processing on the state judgment results from different data sources. During the fusion process, a spatiotemporal sensitivity correction factor is introduced based on the time difference and spatial displacement of the event to compensate for the attenuation of the support of each piece of evidence. By jointly normalizing the support and conflict of different state judgment results, a comprehensive judgment result of the tourist's current state with high confidence is generated.

[0077] At the decision-making level, this implementation method utilizes the DS evidence theory framework to quantify and handle information uncertainty. The specific implementation steps are as follows: Constructing basic probability assignment: Defining the identification framework ,in H This represents the target hypothesis to be verified (such as high tourist engagement with current cultural attractions), for each data source. i Based on its confidence level With supportive tendencies Constructing the basic probability assignment function .

[0078] Introducing spatiotemporal sensitivity correction: To reflect the timeliness of tourism scenarios, the system... Make corrections, that is ,in Indicates data source i Basic probability assignment values ​​after spatiotemporal decay correction and These represent the spatiotemporal offset between the time the evidence was generated and the current decision-making moment. and These are preset time decay coefficients and spatial offset weight constants, used to adjust the sensitivity of spatiotemporal information to confidence levels, ensuring that the weights of outdated or far-field interference data are effectively decayed. Through exponential decay correction, the system ensures that "strongly spatiotemporally correlated" data that is closer to the current time and location receives higher decision weights during the fusion process, effectively solving the interference caused by outdated or far-field data in inbound tourism scenarios.

[0079] Enforcing the evidence combination rule: Applying Dempster's combination rule to synthesize all the corrected independent data, the basic probability assignments after synthesis are determined. satisfy: in: Used to represent a subset of the identification framework from all independent sources of evidence. A The degree of common support; Indicates the first j Each independent data source is any subset selected within its identification framework; This represents a constraint condition where the intersection of subset sequences selected from all data sources is performed, and the result of the operation is exactly equal to the subset sequence. A All possible combinations; This indicates that each independent source of evidence corresponds to its respective subset. The product of the assigned probability mass values; K This represents the normalization factor, whose value is equal to the sum of the probabilities of all evidence combinations whose intersection is an empty set. This represents the compatibility between different data sources and serves as a normalization term to ensure that the sum of the probabilities of the synthesized data is 1. This is achieved through a normalization factor. K Addressing conflicting evidence can effectively prevent decision-making distortions caused by logical contradictions from different sources.

[0080] After the synthetic calculation is completed, the system outputs the hypothesis. H Trust level and the probability quality that characterizes overall uncertainty The system ultimately executes its decision-making logic based on a dual-threshold mechanism: when and When the system determines that the current intent is credible, it accepts the service decision suggestion and executes the recommended delivery. Conversely, when highly conflicting evidence leads to... When the preset threshold is exceeded, the system determines that the decision is unreliable and will trigger a compensation mechanism that delays execution or requests the user to provide supplementary information, thereby ensuring the high robustness and credibility of the output results at the system architecture level.

[0081] This fusion paradigm based on uncertainty modeling ensures that the system can maintain a high level of decision credibility and robustness even under non-ideal conditions such as GPS offset or semantic ambiguity.

[0082] (6) Conditional service decision generation.

[0083] The comprehensive judgment result and dynamic cultural feature vector are used as input. The scaling factor and translation bias generated by the parameter generation network based on the dynamic mapping of cultural features are used. The intermediate layer parameters of the generation model are dynamically modulated through conditional layer normalization, thereby generating service decision information that matches the tourist's background in terms of cultural connotation and expression style. This information includes at least one or more of the following: recommendation information, itinerary planning information, and prompt information.

[0084] In the final service generation stage, such as recommendation, itinerary planning, or copywriting generation, this invention employs conditional parameter modulation technology to intervene in the generative neural network in real time. Its core technical approach is to replace the standard normalization layer with a conditional normalization layer (CLN). The system will generate feature vectors... The process propagates to the service generation layer, where conditional normalization technology enables low-level intervention in the decoding logic of the generative model. To achieve this deep adaptation intervention, the system integrates a Parameter Generation Network (PGN) into the decoder architecture of the generative model. The core operating mechanism of this PGN lies in its modification of the scaling parameters in traditional models. and bias parameters Instead of addressing the fixed limitations, two independent feedforward neural networks are used, along with the aforementioned dynamic cultural vectors. Perform real-time mapping calculations for the independent variables.

[0085] Specifically, PGN is used to dynamically control the internal feature distribution of the generative model based on tourist characteristics, for the input features of a certain layer of the decoder. x Its conditional normalization process follows the following formula: in: and Representing the input features respectively x The mean and standard deviation, The spatiotemporal context vector representing the current decoding moment (such as current weather, geofencing status, and the time tourists have spent at the attraction). and This is by PGN and The input is a dynamically generated function.

[0086] Through this mechanism, the model is able to adapt to different cultural vectors during the forward propagation process. By modulating its internal feature distribution in real time, the generated content automatically undergoes adaptive evolution in terms of information filtering, narrative style, and value ranking, closely matching the user's cultural background, thereby eliminating the semantic gap in cross-cultural scenarios.

[0087] (7) Cross-module collaborative operation mechanism and data interaction.

[0088] This implementation method organically integrates the aforementioned core technology modules through an event-driven architecture and a microservice paradigm, constructing a real-time collaborative operation system with high throughput and low latency. The specific data interaction and module collaboration process is as follows: Data aggregation and event instantiation stage: Raw multi-source heterogeneous data (such as GPS tracks, text streams, transaction records, etc.) are accessed by the system through a unified adaptation layer. Then, the complex event processing engine converts the raw data stream into standardized business events in real time according to the preset finite state machine logic. These events serve as the "first driving force" for system operation and are published to distributed event message buses such as Kafka.

[0089] Asynchronous Parallel Computing and Decision Fusion Phase: The aforementioned business events serve as driving signals, triggering the downstream computing cluster through asynchronous subscription. In this phase, the stateless semantic parsing service module and the multimodal quality assessment service module execute in parallel. The former utilizes knowledge graphs to extract higher-order intents, while the latter uses DS evidence theory to assess the credibility of each modality. This asynchronous parallel mechanism significantly improves the system's response speed under the impact of massive amounts of data.

[0090] State evolution and dynamic profile update stage: The processing results of each computing unit are pushed to the state storage layer in real time, and then the system calls the context update function to update the tourist context object and dynamic cultural feature vector in the centralized storage. The atomic updates ensure that the visitor's "digital mirror image" can evolve dynamically in milliseconds following their real-time behavior.

[0091] Request-driven service adaptation and generation phase: When the terminal receives a user's proactive service request, the service gateway performs context probing, automatically aggregating the current request parameters with the latest cultural feature vector in the storage layer. Based on this, the system calls the aforementioned conditional generation model based on PGN and CLN to produce personalized content highly aligned with the current cultural profile, and feeds it back to the user terminal via the API link.

[0092] System governance and reliability assurance: Service discovery, traffic distribution, link tracing and circuit breaking of the entire system are uniformly managed through the service mesh. Through this control plane, the system can achieve fine-grained governance of multi-source concurrent flows, ensuring that the system still has high availability, sub-second latency and strong observability in cross-regional and large-scale deployment scenarios.

[0093] In this embodiment, the above steps (1) to (7) are performed in real time or near real time to meet the immediate service needs of inbound tourists in dynamic scenarios.

[0094] The following section uses a specific inbound tourism test scenario to illustrate the entire operational mechanism of the technical solution of this invention.

[0095] Target audience: An American tourist with a highly individualistic cultural background (User ID: US001).

[0096] Environmental awareness: When tourists arrive in Shanghai, the system collects GPS coordinate streams, multilingual comment texts, and sensor status data in real time through their mobile terminals.

[0097] This embodiment performs cross-cultural adaptive recommendation according to the following steps: Event-driven transition: When a visitor enters the "Bund" pre-defined geofence, the CEP engine detects a change in the state machine. Migrate to Then, the business event Event_Enter_Geofence(US001, Bund, T1) is instantiated and published to the event bus.

[0098] Context Update: After receiving the above events, the context management module updates the context object atomically. The real-time location attribute in the data is used to mark the visitor's current spatial status as "Bund Activity".

[0099] Fine-grained semantic parsing: A tourist posted a voice comment through the terminal: "I want to find a traditional tea house to experience local culture". The semantic parsing module then linked to the "tea house" entity in the knowledge graph with cultural_authenticity_index=0.9, and parsed out the structured tuple {subject: US001, predicate: seek, object: traditional tea house, cultural intention: [cultural authenticity]}, with a semantic confidence score of 0.88.

[0100] Uncertainty Decision Fusion: The system collects evidence from the following three dimensions: Location evidence (sensor modality): Located in the core area of ​​the Bund, supporting the "cultural experience" intent, confidence level. ; Semantic evidence (textual modality): The above parsing results support the intent of "cultural experience," with a confidence level of [insert confidence level here]. ; Historical behavioral evidence: No consumption was found on that day, so it is considered neutral evidence.

[0101] Fusion computation: The improved DS evidence theory is applied for synthesis to calculate the target hypothesis. H Trustworthiness (for a strong cultural experience) Characterizing the uncertainty of the overall conflict Because the double threshold judgment condition is met ( The decision-makers determined that the intention was valid.

[0102] Dynamic profile representation: The dynamic cultural profile module aggregates the source region characteristics and real-time intent characteristics of US001, and generates dynamic cultural vectors through the Transformer encoder. In this vector, the individualism dimension score appears to be high (0.82).

[0103] Adaptive content generation: When tourists approach the "City God Temple" area, which features a traditional teahouse, a service generation command is triggered. The model decoder then calls the parameter generation network to generate content based on... The text, with its strong individualistic tendencies, employs a conditional layer normalization modulation generation strategy, focusing on the distribution of words related to "uniqueness, seclusion, and in-depth experience," resulting in the following recommended content: "Discover a hidden tea house in the old town for an authentic and personal cultural immersion."

[0104] Through the implementation of the above-described entire process, the system completes a closed-loop processing step spanning "geographical perception, knowledge reasoning, intent fusion, profile modulation, and content generation" within sub-seconds after a tourist's request is generated. Experiments demonstrate that the recommended content output by this solution highly aligns with tourists' psychological motivations in terms of cultural connotation, significantly improving the accuracy of information acquisition and service experience satisfaction for inbound tourists in heterogeneous cultural environments.

[0105] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A method for integrating and intelligently analyzing multi-source big data on inbound tourism, characterized in that, Includes the following steps: (1) Collect multi-source heterogeneous data of inbound tourists and convert data from different data sources into a unified event stream format with timestamps; (2) Call the semantic parsing model to process the semantic events of the text, and constrain the parsing results based on the domain knowledge model during the processing, so as to map the unstructured text into a structured semantic feature vector; In the knowledge modeling stage, the domain knowledge model uses OWL to uniformly model domain entities. Based on the Hofstede cultural dimension theoretical framework, it introduces quantitative attribute vectors reflecting cultural differences for entities. It calls a relation extraction model based on the Transformer architecture to deeply mine multilingual travelogues, and uses a regression prediction model to transform cultural tendencies in unstructured texts into continuous values ​​between 0 and 1, thereby transforming abstract cultural features into computer-computable logical attributes. After the above extraction and quantification, the "entity-relationship-attribute" triple data is stored in the graph database in RDF format, forming a cross-cultural knowledge graph that supports deep reasoning. (3) Based on the cultural feature mapping model, the semantic feature vector is converted into a dynamic cultural feature vector with continuously taking values; The cultural feature mapping model is based on the generation mechanism of dynamic cultural profile and parameterized conditional modulation. In the profile construction stage, an encoder network based on the Transformer architecture generates a dynamic cultural feature vector with continuously taking values ​​for each tourist in real time. This encoding process goes beyond static nationality labels and deeply integrates multi-source behavioral signals, including the embedded representation of real-time behavioral event sequences, structured semantic feature vectors containing intent, quantitative attribute vectors reflecting cultural differences, and implicit feedback signals generated by user interaction with the system through a self-attention mechanism. The final output dynamic cultural feature vector constitutes a digital portrait of tourist cultural preferences. (4) For different event streams, calculate the confidence weight associated with each data source based on the event's timeliness parameter, historical consistency parameter, and integrity index. (5) Based on the confidence weight, the improved DS evidence theory inference engine is invoked to perform uncertainty fusion processing on the state judgment results from different data sources, and a comprehensive judgment result of the tourist's current state is generated; (6) The comprehensive judgment result and the dynamic cultural feature vector are used as input. The scaling factor and translation bias generated by PGN based on the dynamic mapping of cultural features are used to dynamically modulate the intermediate layer parameters of the service generation model through the conditional layer normalization method, thereby generating service decision information that matches the tourist background in terms of cultural connotation and expression style.

2. The method for integrating and intelligently analyzing multi-source big data of inbound tourism according to claim 1, characterized in that: In step (1), the CEP engine is used to perform single-point access and time-series alignment of heterogeneous real-time data streams, including GPS trajectories and mobile payments. A sliding window mechanism is introduced, which performs pattern matching on the continuous data streams within the window to abstract the scattered raw data signals into standardized event objects with clear business semantics in real time. At the same time, a lightweight, in-memory, atomically updatable visitor context object is maintained for each active visitor. The data structure of this object not only includes user ID, location, and intent vector, but also integrates the state bits of a finite state machine. When a new event arrives and updates the visitor context object, a lightweight rule engine based on the Rete algorithm is used to trigger the corresponding condition judgment and action execution instantaneously through the topology matching mechanism, without the need for frequent database polling.

3. The method for integrating and intelligently analyzing multi-source big data on inbound tourism according to claim 1, characterized in that: In step (2), the semantic parsing model uses XLM-RoBERTa as its base and integrates a knowledge retrieval module to perform deep processing on the text generated by tourists. During the processing, text entities are first identified and linked to the knowledge graph. Then, the contextual representation of the text is used as the query, and the quantified attribute vector retrieved from the knowledge graph is used as the key and value. A structured semantic feature vector containing intent and cultural characteristics is generated through a cross-attention mechanism.

4. The method for integrating and intelligently analyzing multi-source big data of inbound tourism according to claim 1, characterized in that: In step (4), for each data modality of the event stream, a lightweight confidence assessment model is first deployed to extract quality feature indicators in real time. For the physical sensing modality, the number of visible satellites, HDOP, and signal-to-noise ratio are extracted. For the text semantic modality, the prediction entropy of the language model and the Softmax probability distribution of the parsing results are extracted. Then, these quality feature indicators are input into a multilayer perceptron network to calculate and generate real-time confidence weights between 0 and 1 online, which are used to reflect the reliability of the corresponding data source in the current state.

5. The method for integrating and intelligently analyzing multi-source big data on inbound tourism according to claim 1, characterized in that: In step (5), the modified BPA value is used in the process of fusing the state judgment results from different data sources using the DS evidence theory inference engine to ensure that the "strong spatiotemporal correlation" data that is closer to the current time and the current position obtains a higher decision weight in the fusion process. The revised BPA value expression is: Where: for any data source i A BPA function is constructed by combining confidence weights and support bias. and Data sources i any subset A BPA values ​​before and after correction α and β These are the preset time decay coefficient and spatial weighting coefficient, respectively. and These represent the time offset and spatial location deviation of the data, respectively.

6. The method for integrating and intelligently analyzing multi-source big data of inbound tourism according to claim 1, characterized in that: In step (6), the intermediate layer parameters of the service generation model are dynamically modulated through conditional layer normalization, that is, for any layer of input features in the model decoder architecture... x The following expression is used to perform conditional normalization: in: Input features x The result after conditional normalization. This indicates element-wise multiplication. and Representing the input features respectively x The mean and standard deviation, As a dynamic cultural feature vector, This is the spatiotemporal context vector at the current decoding moment. and They represent and Input the scaling factor and translation bias generated in real time by PGN.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor is used to execute the computer program to implement the inbound tourism multi-source big data integration and intelligent analysis method as described in any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the method for multi-source big data integration and intelligent analysis of inbound tourism as described in any one of claims 1 to 6.

Images