Scenic spot escort service method and system based on generative AI and knowledge base linkage

Abstract

The present invention discloses a scenic spot companion service method and system based on the linkage of generative AI and knowledge base, which specifically relates to the field of artificial intelligence and smart tourism technology, and is used to solve the problem that the generative AI model in the existing scenic spot guide system forgets core knowledge and reduces the accuracy of answers due to dynamic knowledge updating; it extracts keywords of tourists' questions in real time and retrieves the knowledge base to generate related data; dynamically weights the priority of related data based on the density of tourists' gathering; generates initial answers through real-time parameter adjustment; integrates spatiotemporal clustering and knowledge graph structural entropy to generate weight factors; corrects answer coverage defects and evaluates the results; outputs the final answer after targeted additional core knowledge iterative training; ensures that generative AI stably retains the core knowledge of the scenic spot during the dynamic learning process, and accurately responds to real-time scene needs, significantly improving the reliability, accuracy and personalization of the guide service.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and smart tourism technology, and more specifically, to a scenic area tour escort service method and system based on the linkage of generative AI and a knowledge base. Background Art

[0002] Currently, smart tour guide systems in scenic spots generally use generative artificial intelligence technology to provide question-and-answer services for tourists, and realize information query functions by connecting to the scenic spot knowledge base. Existing technologies usually adopt static knowledge injection methods, that is, the basic data of the scenic spot is imported into the AI model for pre-training before the system is deployed to form a fixed knowledge response system. During the service operation, some systems try to supplement real-time information by dynamically updating the knowledge base content, such as adding introductions to new attractions or temporary event notifications, so as to improve the timeliness and completeness of information services.

[0003] However, this dynamic update mechanism has defects in actual application: when the generative artificial intelligence model performs real-time learning on new knowledge, the global adjustment characteristics of the neural network parameters will lead to systematic forgetting of the core knowledge of the scenic spot that has been mastered, causing the AI model to have a lower accuracy when answering tourists' basic and common sense questions, such as confusing the historical age of the scenic spot, incorrectly describing architectural features, etc., which seriously affects the reliability of the guided tour service and user experience. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide a scenic spot tour service method and system based on the linkage of generative AI and knowledge base to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The scenic area tour service method based on the linkage between generative AI and knowledge base includes the following steps:

[0007] S1. Real-time acquisition of question information input by tourists through mobile terminals, extraction of keyword sets in the question information and retrieval of matching related data from the scenic spot knowledge base;

[0008] S2. Calculate the tourist concentration density of the scenic spots corresponding to the associated data based on the tourist distribution thermal data to adjust the priority weight of the real-time updated knowledge in the associated data and generate a dynamically weighted associated data set;

[0009] S3. Input the associated data set and question information into the pre-trained generative AI model, complete real-time adaptive training through dynamic parameter adjustment, and generate initial answer information;

[0010] S4. Generate spatiotemporal weight factors based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph;

[0011] S5. Use the spatiotemporal weight factor to modify the core knowledge coverage of the initial answer information and generate a spatiotemporal constraint coverage evaluation result;

[0012] S6. When the coverage evaluation result does not meet the threshold, the core knowledge data of the corresponding keyword set in the scenic area knowledge base is called, appended to the associated data, and the generative AI model is retrained to output the final answer information to the mobile terminal.

[0013] In a preferred embodiment, the question information input by tourists through mobile terminals is obtained in real time, a keyword set in the question information is extracted, and matching related data is retrieved from the scenic spot knowledge base, including:

[0014] Collect tourists' question information in real time through the input interface of the mobile terminal;

[0015] Perform word segmentation on the question information and filter out stop words to generate an initial keyword set;

[0016] Filter entity nouns that match the scenic spot knowledge base from the initial keyword set to form a final keyword set;

[0017] Match the coordinate data, historical event data, and cultural explanation data of the corresponding scenic spots in the scenic area knowledge base according to the final keyword set;

[0018] Merge the matched coordinate data, historical event data, and cultural interpretation data into related data.

[0019] In a preferred embodiment, the scenic spot knowledge base includes structured data stored according to geographical distribution, historical background and cultural attributes.

[0020] In a preferred embodiment, the tourist concentration density of the scenic spot corresponding to the associated data is calculated based on the tourist distribution thermal data to adjust the priority weight of the real-time updated knowledge in the associated data, and generate a dynamically weighted associated data set, including:

[0021] Obtain the real-time tourist location coordinates of the corresponding scenic spots based on tourist distribution thermal data;

[0022] The geographical area of each scenic spot is divided into grid cells of a preset size, and the number of real-time tourist location coordinates in each grid cell is counted as the tourist gathering density;

[0023] According to the density of tourist gatherings and the matching degree of content tags of real-time updated knowledge in the scenic area knowledge base, the priority weight coefficient of real-time updated knowledge is set;

[0024] The real-time updated knowledge is dynamically weighted based on the priority weight coefficient to generate a dynamically weighted associated data set containing weighted values.

[0025] In a preferred embodiment, the associated data set and question information are input into a pre-trained generative AI model, and real-time adaptive training is completed and initial answer information is generated through dynamic parameter adjustment, including:

[0026] Vectorize and concatenate the weighted values in the dynamically weighted associated data set with the question information to generate a fusion input vector;

[0027] Dynamically adjust the attention weight distribution ratio of the generative AI model based on the distribution characteristics of the weighted values in the fusion input vector;

[0028] Perform gradient descent optimization on the decoder layer of the generative AI model based on the adjusted attention weight distribution ratio to update the model parameters;

[0029] The updated generative AI model is used to infer and calculate the question information to generate initial answer information containing the core knowledge features of the scenic spot.

[0030] In a preferred embodiment, a spatiotemporal weight factor is generated based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph, including:

[0031] The continuous positioning data of tourists' mobile terminals are segmented in time and space, and trajectory segments with a density higher than a preset threshold are identified as spatiotemporal hotspots;

[0032] Construct a logical dependency graph between events, people, and buildings in the scenic area knowledge base and calculate the intermediary index of each node;

[0033] The spatiotemporal correlation strength is calculated based on the overlapping area between the geographical boundaries of the spatiotemporal hotspot area and the associated areas of high-betweenness nodes in the logical dependency graph;

[0034] The topological structure disorder index of the logical dependency graph and the spatiotemporal correlation strength are normalized and weighted and fused to generate the spatiotemporal weight factor.

[0035] Normalize the spatiotemporal weight factors and output a set of normalized spatiotemporal weight factors.

[0036] In a preferred embodiment, the core knowledge coverage of the initial answer information is modified using the spatiotemporal weight factor to generate a spatiotemporal constrained coverage evaluation result, including:

[0037] Extract semantic features that match the core knowledge labels of the scenic spot knowledge base from the initial answer information to generate a core knowledge feature set;

[0038] Prioritizing the core knowledge feature set based on the normalized spatiotemporal weight factor set to generate a core knowledge feature priority sequence;

[0039] According to the comparison between the priority sequence of core knowledge features and the semantic coverage of the initial answer information, the feature segments with missing coverage are identified;

[0040] Dynamically correlate the missing coverage fragments with the spatiotemporal weight factors to generate spatiotemporal constrained coverage evaluation results.

[0041] In a preferred embodiment, when the coverage evaluation result does not meet the threshold, the core knowledge data of the corresponding keyword set in the scenic area knowledge base is called, appended to the associated data, and the generative AI model is retrained to output the final answer information to the mobile terminal, including:

[0042] When the priority and weight value of the defect mark in the coverage assessment result meet the preset conditions, a knowledge append trigger instruction is generated;

[0043] According to the knowledge addition trigger instruction, the core knowledge data corresponding to the coverage-missing feature fragments are retrieved from the scenic area knowledge base to generate a knowledge addition set;

[0044] Dynamically appending the knowledge append set to the dynamically weighted linked data set to form an enhanced linked data set;

[0045] Incrementally train the decoder layer of the generative AI model based on the enhanced set of associated data to update the model parameters;

[0046] The updated generative AI model performs semantic fusion calculations on the initial answer information to generate the final answer information and output it to the mobile terminal.

[0047] On the other hand, the present invention provides a scenic area tour service system based on generative AI and knowledge base linkage, including the following modules:

[0048] The question retrieval module is used to obtain the question information input by tourists through mobile terminals in real time, extract the keyword set in the question information and retrieve the matching related data from the scenic spot knowledge base;

[0049] The thermal weighting module is used to calculate the tourist concentration density of the scenic spots corresponding to the associated data based on the tourist distribution thermal data to adjust the priority weight of the real-time updated knowledge in the associated data and generate a dynamically weighted associated data set;

[0050] The model adaptation module is used to input the associated data set and question information into the pre-trained generative AI model, complete real-time adaptation training through dynamic parameter adjustment, and generate initial answer information;

[0051] The spatiotemporal weight module is used to generate spatiotemporal weight factors based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph;

[0052] The coverage correction module is used to correct the core knowledge coverage of the initial answer information using the spatiotemporal weight factor and generate a coverage evaluation result with spatiotemporal constraints;

[0053] The result output module is used to call the core knowledge data of the corresponding keyword set in the scenic spot knowledge base when the coverage evaluation result does not meet the threshold, append it to the associated data, retrain the generative AI model, and then output the final answer information to the mobile terminal.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. Through the dynamic weighted association data and spatiotemporal coupling weight mechanism, the generative AI model effectively maintains the long-term memory ability of the core knowledge of the scenic spot; in the process of real-time knowledge updating, a priority weight dynamic allocation strategy is adopted to associate the density of tourist gatherings with the criticality of knowledge, ensuring that the core knowledge of high-density areas is stably strengthened during model training; at the same time, the spatiotemporal weight factor is used to quantify the knowledge coverage defects, and the core knowledge addition and local parameter adjustment are triggered in a targeted manner to block the interference coverage of existing knowledge by global training; so that when the model absorbs real-time information, it can not only quickly adapt to dynamic scenarios, but also avoid large-scale forgetting of historical data, significantly improving the consistency of the answer content in basic historical facts, geographical information and other dimensions, and ensuring the accuracy and credibility of the information obtained by tourists.

[0056] 2. Through knowledge coverage evaluation and closed-loop iteration mechanisms, self-optimization and personalized adaptation of service quality are achieved; the spatiotemporal weight factors generated based on tourist behavior trajectories and knowledge topological characteristics can accurately identify insufficient coverage problems in answers caused by the dynamic nature of the scene. Combined with targeted retrieval and incremental training of the knowledge base, a continuous improvement cycle of "evaluation-compensation-verification" is formed; this not only strengthens the generative AI's ability to understand complex knowledge relationships, but also reduces resource consumption through lightweight training strategies, allowing the system to stably output guide content that meets tourists' spatiotemporal needs and interest preferences in high-concurrency scenarios, taking into account both service efficiency and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of the scenic area tour service method based on the linkage between generative AI and knowledge base of the present invention;

[0058] Figure 2 This is a structural diagram of the scenic area tour service system based on the linkage between generative AI and knowledge base in the present invention. DETAILED DESCRIPTION

[0059] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0060] Example 1: Figure 1 The present invention provides a scenic area tour service method based on generative AI and knowledge base linkage, which includes the following steps:

[0061] S1. Real-time acquisition of question information input by tourists through mobile terminals, extraction of keyword sets in the question information and retrieval of matching related data from the scenic spot knowledge base;

[0062] S2. Calculate the tourist concentration density of the scenic spots corresponding to the associated data based on the tourist distribution thermal data to adjust the priority weight of the real-time updated knowledge in the associated data and generate a dynamically weighted associated data set;

[0063] S3. Input the associated data set and question information into the pre-trained generative AI model, complete real-time adaptive training through dynamic parameter adjustment, and generate initial answer information;

[0064] S4. Generate spatiotemporal weight factors based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph;

[0065] S5. Use the spatiotemporal weight factor to modify the core knowledge coverage of the initial answer information and generate a spatiotemporal constraint coverage evaluation result;

[0066] S6. When the coverage evaluation result does not meet the threshold, the core knowledge data of the corresponding keyword set in the scenic area knowledge base is called, appended to the associated data, and the generative AI model is retrained to output the final answer information to the mobile terminal.

[0067] S1. Real-time acquisition of question information input by tourists through mobile terminals, extraction of keyword sets in the question information, and retrieval of matching related data from the scenic spot knowledge base. The specific implementation is as follows:

[0068] Visitors' questions are collected in real time through the mobile terminal's input interface. The mobile terminal's input interface includes voice input and text input. The voice input function uses the terminal's built-in microphone to collect visitors' voice questions and converts them into text using voice recognition technology. The text input function receives text questions entered by visitors through the terminal's touch screen or keyboard. Voice recognition technology converts voice signals into text and outputs the questions in text format.

[0069] Perform word segmentation on the question information and filter out stop words to generate an initial keyword set. The word segmentation process uses the maximum forward matching method based on a dictionary, dividing the text question information according to the scenic area-specific dictionary, which contains a vocabulary set of scenic spot names, historical figure names, and cultural terms. For example, when a tourist asks "What is the history of this scenic area?", the words obtained after word segmentation are "this scenic area", "history", "have", and "which", etc. Filtering out stop words removes虚词 and auxiliary words without actual meaning through a predefined stop word list, and the stop word list includes common words such as "of", "already", "right", "have", and "which", etc. After filtering, an initial keyword set is generated, such as "this scenic area" and "history".

[0070] Screen out the entity nouns in the initial keyword set that match the scenic area knowledge base to form a final keyword set. The screening process is achieved by matching each word in the initial keyword set with the entity nouns stored in the scenic area knowledge base classified by geographical distribution, historical background, and cultural attributes. If a word exists in the entity noun set of any classification, the word is retained. For example, "this scenic area" matches the scenic spot name set in the geographical distribution classification, and "history" matches the event type set in the historical background classification, and both are retained as the final keyword set. Unmatched words such as "which" are excluded. The entity noun set in the geographical distribution classification includes scenic spot names and coordinate data, the historical background classification includes historical events and chronological data, and the cultural attribute classification includes dialect words and cultural explanation labels.

[0071] Match the corresponding scenic spot coordinate data, historical event data, and cultural explanation data in the scenic area knowledge base according to the final keyword set. The matching process retrieves data from different classifications respectively: according to the geographical distribution type words in the final keyword set, such as "this scenic area", retrieve the longitude and latitude coordinate data of this scenic spot stored in the scenic area knowledge base; according to the historical background type words, such as "history", retrieve the historical event description data that meets the time range in the scenic area knowledge base, and the time range is automatically set according to the implicit time information in the tourist's question. For example, when the time is not specified, all historical events are retrieved by default; according to the cultural attribute type words, calculate their semantic similarity with the cultural explanation labels, and screen out the associated cultural explanation data with a similarity higher than the set value. The semantic similarity is achieved through the distance measurement in the semantic vector space.

[0072] Merge the matched coordinate data, historical event data, and cultural explanation data into associated data. The merging operation includes data format unification and redundancy processing: convert the coordinate data into geographical coordinate format, arrange the historical event data in chronological order and remove duplicate descriptions, and sort the cultural explanation data by similarity scores and retain the entry with the highest score. The associated data is used as the input for subsequent steps for the training of the generative model and answer generation.

[0073] S2. Calculate the tourist density of the scenic spot corresponding to the associated data based on the tourist distribution thermal data to adjust the priority weight of the real-time updated knowledge in the associated data and generate a dynamically weighted associated data set. The specific implementation is as follows:

[0074] Based on visitor distribution thermal data, a real-time visitor location coordinate set for each scenic spot is obtained. This visitor distribution thermal data is collected in real time through an application programming interface (API) integrated with an electronic map service. It includes the geographic coordinates of the visitor's mobile terminal, timestamp information, and a unique device identifier. Geographic coordinates are obtained using satellite positioning systems or cellular network base station positioning technology. Satellite positioning systems utilize the Global Positioning System (GPS) or the Beidou Navigation System (BDS). Base station positioning technology triangulates positioning by measuring the difference in signal strength between the terminal and at least three base stations. Coordinate data is stored in the internationally accepted decimal latitude and longitude format, for example, "116.28°E, 39.99°N." The real-time visitor location coordinate set is divided into regions based on the predefined geographic boundaries of the scenic spot in the scenic spot knowledge base. The geographic boundaries of the scenic spot are formed by enclosing the coordinate data of the scenic spot's polygon vertices stored in the scenic spot knowledge base. The polygon vertex coordinate data is collected on-site using surveying and mapping equipment and stored in the geographic information database. For example, the geographical boundary of a rectangular scenic spot is composed of four vertex coordinates (116.28°E, 39.99°N), (116.30°E, 39.99°N), (116.30°E, 40.01°N), and (116.28°E, 40.01°N) connected in sequence. The real-time tourist location coordinate set only retains the coordinate points that fall within the polygon.

[0075] Each scenic spot's geographic area is divided into grid cells of a preset size, and the number of real-time visitor location coordinates within each grid cell is counted as the visitor density. The preset grid cell size is dynamically adjusted based on the area of the scenic spot. The adjustment rule is as follows: for scenic spots with an area greater than or equal to 1 square kilometer, the grid cell side length is 100 meters; for scenic spots with an area less than 1 square kilometer and greater than or equal to 0.1 square kilometer, the grid cell side length is 50 meters; and for scenic spots with an area less than 0.1 square kilometer, the grid cell side length is 20 meters. Grid division is achieved using a plane rectangular coordinate system projection transformation technique. The specific steps include: converting the longitude and latitude coordinates of the scenic spot's geographic boundary into coordinate points in a plane rectangular coordinate system using the Gauss-Krüger projection algorithm, with the plane coordinate system origin set to the minimum longitude and latitude values of the scenic spot's polygon vertices; generating equidistant grid cells in the plane coordinate system based on the preset grid side lengths, with each grid cell defined by the plane coordinates of its lower left and upper right corners. When calculating the visitor density, the real-time visitor location coordinates are converted to plane coordinates using the same projection algorithm, and the grid cell to which they belong is determined. For example, in the plane coordinate system of a scenic spot, the coordinates of the lower left corner of a grid cell are (100, 200) and the coordinates of the upper right corner are (150, 250). Then the plane coordinate point (120, 220) belongs to this grid cell, and the number of all coordinate points in this grid cell is the tourist gathering density.

[0076] The priority weight coefficient for real-time updated knowledge is determined based on the degree of match between visitor density and the content tags of real-time updated knowledge in the scenic area's knowledge base. The content tags of real-time updated knowledge are categorized into three types: emergency notification, event change, and facility status. Specifically, emergency notification tags are associated with safety guidance, temporary closures, and disaster warnings; event change tags are associated with performance schedule adjustments, ticketing policy changes, and event cancellations; and facility status tags are associated with equipment maintenance, route closures, and service suspensions. The matching rule is: if the scenic area associated with the content tag of the real-time updated knowledge matches the scenic area of a grid cell where the visitor density exceeds a preset threshold, the area is considered a match. The preset threshold is dynamically set at 20% of the scenic area's maximum capacity, which is stored in the scenic area knowledge base. For example, if the maximum capacity of a scenic area is 5,000 and the preset threshold is 1,000, a match is triggered when the visitor density within a grid cell exceeds 1,000. The weighting formula is designed differently based on the knowledge type: Emergency Notification Knowledge Weighting = Visitor Density / Preset Threshold × 2; Event Change Knowledge Weighting = Visitor Density / Preset Threshold × 1.5; Facility Status Knowledge Weighting = Visitor Density / Preset Threshold × 1. For example, if the visitor density for a grid is 1,200 and the preset threshold is 1,000, the associated Emergency Notification Knowledge Weighting is 1,200 / 1,000 × 2 = 2.4, rounded to two decimal places.

[0077] A dynamic weighted calculation is performed on real-time updated knowledge based on priority weight coefficients to generate a dynamically weighted linked data set containing weighted values. This dynamic weighting calculation is achieved through the following steps: extracting all real-time updated knowledge items from the scenic area knowledge base, filtering out eligible knowledge items based on content tag matching; combining the filtered knowledge items with their weight coefficients by appending a weight value field after the knowledge content field, with the fields separated by a "|" separator. For example, the real-time updated knowledge "Attraction B East Gate Temporarily Closed" is weighted to "Attraction B East Gate Temporarily Closed | Weight 3.2." After generating the dynamically weighted linked data set, it is merged with the linked data retrieved in step S1 based on data identifiers generated from the knowledge content's keyword hash values. The merge rule includes: if the same data identifier exists in both sets, its weight value is accumulated; if it exists only in one set, it is retained. For example, if the dynamically weighted linked data set contains "show time adjustment | weight 2.5" and the linked data from step S1 contains "show time adjustment | weight 1.0," the merged result is "show time adjustment | weight 3.5." The final dynamically weighted associated data set is sorted in descending order of weight value to ensure that high-priority knowledge is called first in subsequent model training.

[0078] S3. Input the associated data set and question information into the pre-trained generative AI model, complete real-time adaptive training through dynamic parameter adjustment, and generate initial answer information. The specific implementation is as follows:

[0079] The weighted values in the dynamically weighted associated data set are vectorized and concatenated with the question information to generate a fused input vector. The dynamically weighted associated data set contains weighted entries of real-time updated knowledge and core scenic area knowledge. The weighted values are stored in floating-point format and are generated through the dynamic weighting calculation in step S2. The question information text is converted into a semantic vector using a pre-trained word vector model. The word vector model is implemented using the Word2Vec model trained on public corpus, mapping each word to a 300-dimensional real number vector. The concatenation process of the weighted value and the semantic vector is as follows: the weight value is expanded to a constant vector of the same dimension as the semantic vector. For example, a weight of 3.5 is expanded to a 300-dimensional vector [3.5, 3.5, …, 3.5], and then added to the semantic vector element by element. For example, if the semantic vector is [0.2, -0.5, 1.3, …, 0.7], the expanded weight vector is [3.5, 3.5, …, 3.5], and the fused vector is [3.7, 3.0, 4.8, …, 4.2]. The fused input vector is used as the input of the generative AI model for subsequent attention weight adjustment.

[0080] The attention weight distribution ratio of the generative AI model is dynamically adjusted based on the distribution characteristics of the weighted values in the fused input vector. The attention weight distribution ratio reflects the model's attention intensity for different input positions. The adjustment process includes the following steps: calculating the mean square error of all weighted values in the fused input vector. If the mean square error exceeds a preset threshold (e.g., 1.5), the weight distribution difference is determined to be significant; for positions with weighted values above the mean, the attention weight is proportionally increased. For example, a fused input vector contains three weighted values of 3.5, 1.2, and 0.8, with a mean of 1.83 and a mean square error of 1.2. If it is below the threshold, no adjustment is made. If the weighted values are 5.0, 0.5, and 0.3, with a mean square error of 2.3, the attention weight for the high-weight position (5.0) is increased by 20% if it exceeds the threshold. The adjusted attention weight distribution ratio is stored in the form of a probability matrix, where the rows correspond to the decoding time steps, the columns correspond to the input positions, and the element values represent the attention probability.

[0081] Based on the adjusted attention weight distribution, the decoder layer of the generative AI model is optimized using gradient descent to update the model parameters. The gradient descent optimization loss function consists of two parts: the first is the answer generation loss, which calculates the cross entropy between the generated text and the standard answer; the second is the attention consistency loss, which calculates the mean squared error between the adjusted attention weights and the initial weights. During the parameter update process, the learning rate is dynamically adjusted based on the weight values: when the weight values are above the mean, a smaller learning rate (e.g., 0.001) is used to prevent overfitting; when the weight values are below the mean, a larger learning rate (e.g., 0.005) is used to accelerate knowledge absorption. For example, a weight value of 3.5 corresponds to a learning rate of 0.001, and a weight value of 0.8 corresponds to a learning rate of 0.005. The number of optimization iterations is set based on the dimensionality of the fused input vector. The default number of iterations for a 300-dimensional vector is 50. Convergence of the loss function is verified after each iteration. Training is terminated early if the loss decreases by less than 1% for five consecutive iterations.

[0082] The updated generative AI model infers the question information and generates an initial response that incorporates the core knowledge features of the scenic spot. The inference process consists of two phases: encoding and decoding. The encoding phase feeds the question text into the model encoder to generate a context vector representation. The decoding phase gradually generates the response text based on the context vector and the adjusted attention weight distribution. During the generation process, the model prioritizes semantic segments corresponding to high-weighted associated data. For example, when the weight of "show time adjustment" is 3.5, the decoder assigns a higher probability to generating words related to "time." The initial response information is then output after text post-processing (such as punctuation correction and grammar verification). For example, for the question "Has today's show time changed?", the response generated is "The current show time has been adjusted to 2:00 PM. Please refer to the scenic spot announcement for details." The generative AI model's parameter updates and inference calculations share the same computational graph structure, ensuring logical consistency between training and inference.

[0083] S4. Generate spatiotemporal weight factors based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph. The specific implementation is as follows:

[0084] The continuous positioning data from tourists' mobile terminals is segmented in time and space, and trajectory segments with a density above a preset threshold are identified as spatiotemporal hotspots. Continuous positioning data is collected in real time via the mobile terminal's global positioning system and includes longitude, latitude, and timestamp information with timestamp accuracy down to the second. Spatiotemporal segmentation is achieved through the collaborative processing of sliding time windows and geographic grid division: the time window length is set to 15 minutes, and the geographic grid side length is 50 meters. The time window and grid side length parameters are dynamically adjusted based on the area of the scenic spot and the characteristics of tourist flow. For example, scenic spots larger than 10 square kilometers use a 30-minute time window and a 100-meter grid side length.

[0085] Within each time window, the number of location data points in each grid is counted. If the number of data points in a grid exceeds a preset threshold, the temporal and spatial range of that grid is identified as a spatiotemporal hotspot. The preset threshold is calculated as follows: Single-grid threshold = Maximum capacity of the scenic area / Total number of grids × 5%. The maximum capacity of the scenic area is derived from the scenic area management department's operational planning documents, and the total number of grids is calculated by dividing the scenic area's geographic area by the grid area. For example, if a scenic area has a total area of 5 square kilometers and a grid side length of 50 meters, the total number of grids is (5,000 meters / 50 meters) × (5,000 meters / 50 meters) = 10,000. If the maximum capacity is 20,000 people, the single-grid threshold is 20,000 / 10,000 × 5% = 0.1 person, rounded to 1 person in practice. If the number of location data points in a grid exceeds one within a time window, it is marked as a hotspot.

[0086] A logical dependency graph was constructed between events, people, and buildings in the scenic area knowledge base, and the betweenness index of each node was calculated. The logical dependency graph was constructed as follows: events, people, and buildings were used as nodes, and directed edges were established if there was a direct causal relationship or spatial subordination between two nodes. Causal relationships were defined as clear decision-making connections within historical events, and the edge weight was the frequency of mentions of the decision in historical documents. Spatial subordination was defined as the inclusion relationship between a building and a geographical area, and the edge weight was the certainty score of the inclusion relationship (ranging from 0 to 1).

[0087] The betweenness index is calculated using the Brandes algorithm. The specific steps are: traversing the shortest paths between all node pairs and counting the number of times each node appears in these paths. The calculation formula is: node betweenness index = (number of shortest paths passing through the node) / (total number of shortest paths between all node pairs in the entire graph) × 100%. Nodes with a betweenness index above 10% are considered high betweenness nodes. For example, if a node appears in 300 shortest paths, and there are 3,000 shortest paths in the entire graph, its betweenness index is 10%, which meets the threshold.

[0088] The strength of spatiotemporal associations is calculated based on the overlap between the geographic boundaries of spatiotemporal hotspots and the associated regions of high-betweenness nodes in the logical dependency graph. The geographic boundaries of spatiotemporal hotspots are determined by the vertex coordinates of the circumscribed polygon of the hotspot grid, which are calculated using a convex hull algorithm. The associated regions of high-betweenness nodes are predefined node influence spheres in the knowledge base. This influence sphere data is derived from historical activity records or geographic radiation models. For example, the influence sphere of a historical event is defined as a circular area with a radius of 500 meters centered at the location of the event.

[0089] The overlapping area is calculated using polygon Boolean operations in a plane coordinate system: the spatiotemporal hotspot polygon and the node-associated area polygon are transformed into the same plane coordinate system using a Gauss-Krüger projection, and the intersection area is calculated. The spatiotemporal correlation strength is calculated as follows: Spatiotemporal Correlation Strength = Intersection Area / (Spatiotemporal Hotspot Area + Node-Associated Area - Intersection Area). The result is linearly normalized to a value between 0 and 1. For example, if a spatiotemporal hotspot has an area of 1200 square meters and a node-associated area is a circular area with a radius of 30 meters (approximately 2827 square meters), the intersection area is 800 square meters. The spatiotemporal correlation strength is 800 / (1200 + 2827 - 800) = 800 / 3227 ≈ 0.248, which is a normalized value of 0.248.

[0090] The topological structure disorder index of the logical dependency graph and the spatiotemporal correlation strength are normalized and then weighted fused to generate the spatiotemporal weight factor.

[0091] The topological disorder index analyzes the topological characteristics of a logical dependency graph using a spectral clustering algorithm. The calculation process includes the following steps: First, an adjacency matrix is constructed for the logical dependency graph. The rows and columns of the adjacency matrix correspond to nodes in the graph, and the matrix elements represent the edge weights between nodes. Edge weights are determined based on the strength of the dependencies between nodes. For example, the causal weight between a historical event and a decision-maker is the ratio of the number of related mentions in historical documents to the total number of documents, and the spatial subordination weight is the certainty score of the geographic inclusion relationship. Next, all eigenvalues of the adjacency matrix are calculated. Eigenvalues represent the structural characteristics of the graph along different dimensions. After calculation, the eigenvalues are sorted in descending order of magnitude, and the top k eigenvalues are selected for subsequent analysis. The value of k is the square root of the total number of nodes in the graph, rounded down. For example, if the graph contains 100 nodes, k = 10. These top k eigenvalues are then normalized. This involves calculating the proportion of each eigenvalue to the sum of the k eigenvalues to obtain a normalized eigenvalue sequence. The entropy value is calculated based on the normalized eigenvalue sequence. The physical meaning of the entropy value is the degree of chaos in the eigenvalue distribution. It is calculated by summing the product of each normalized eigenvalue and its natural logarithm and taking the negative. Specifically, each normalized eigenvalue is multiplied by the natural logarithm of the value, all the product results are added up and the negative is taken to finally obtain the entropy value.

[0092] The entropy value ranges from 0 to the maximum value of the natural logarithm. The higher the value, the more uniform the distribution of the eigenvalues and the more chaotic the graph topology. For example, the eigenvalues of a completely randomly connected graph are evenly distributed, and the entropy value is close to the theoretical maximum value; the eigenvalues of a graph with a clear hierarchy and centralized structure are concentrated in a few dimensions, and the entropy value is close to 0.

[0093] To facilitate subsequent calculations, the entropy values are linearly scaled to a standard range of 0 to 1. The scaling formula is: Normalized Entropy = (Original Entropy - Minimum Theoretical Entropy) / (Maximum Theoretical Entropy - Minimum Theoretical Entropy), where the minimum theoretical entropy corresponds to a completely centralized structure (dominated by a single eigenvalue), and the maximum theoretical entropy corresponds to a completely uniform distribution. The resulting normalized entropy is an indicator of topological disorder, which is positively correlated with the structural complexity of the graph and is used to quantify the degree of topological disorder in knowledge dependencies. For example, when multiple nodes in a historical event graph have cross-dependencies, the entropy value is significantly higher than when the graph only contains linear causal chains, indicating a more complex structure and more dispersed knowledge connections.

[0094] Normalization uses a maximum-minimum scaling method to map entropy values to a range of 0 to 1. The weighted fusion formula is: spatiotemporal weight factor = normalized topological structure disorder index × α + normalized spatiotemporal correlation strength × β, where α and β are weighting coefficients, and α + β = 1. Initial values are set to α = 0.6 and β = 0.4. In actual operation, dynamic adjustments are made based on the effectiveness of scenic area knowledge coverage. The adjustment rule is: if the core knowledge coverage assessment result is lower than expected, increase α to strengthen the influence of topological disorder. For example, if the normalized disorder index of a region is 0.8 and the normalized spatiotemporal correlation strength is 0.5, the spatiotemporal weight factor = 0.8 × 0.6 + 0.5 × 0.4 = 0.68.

[0095] The spatiotemporal weight factors are normalized, outputting a set of normalized spatiotemporal weight factors. This normalization uses the Z-score method, calculated as follows: Normalized value = (original value - mean) / standard deviation. The mean and standard deviation are calculated in real time based on all spatiotemporal weight factors for the current scenic area. For example, if a scenic area has 100 spatiotemporal weight factors with a mean of 0.5 and a standard deviation of 0.15, and the original value of a weight factor is 0.68, then the normalized value = (0.68 - 0.5) / 0.15 ≈ 1.2. The set of normalized spatiotemporal weight factors is sorted in descending order by normalized value, with entries with high normalized values prioritized for core knowledge coverage correction in subsequent steps. Normalization also includes outlier filtering: if the absolute value of the normalized value of a weight factor is greater than 3, it is considered an outlier and replaced with the mean of the adjacent entries to prevent extreme values from interfering with model training. For example, if a weight factor's normalized value is 3.5, it is replaced with the mean of the two adjacent entries (1.2 and 1.0) in the sorted order, which is 1.1.

[0096] Step S4 solves the core knowledge coverage deviation problem caused by the lack of knowledge relevance in dynamic learning of generative AI models by coupling the spatiotemporal characteristics of tourist behavior with the complexity of the knowledge topological structure. Compared with the existing technology that only relies on the knowledge update frequency or static label weighting, this method innovatively introduces the fusion weight of spatiotemporal correlation strength and topological chaos, so that the model can synchronously perceive the physical space aggregation state and the evolution of the knowledge network structure during training, blocking the interference coverage of high-chaos knowledge on core knowledge; through the joint quantification of spatiotemporal and knowledge dual-domain features, it breaks through the limitations of traditional single-dimensional weight distribution, while ensuring real-time knowledge injection, it directionally strengthens the stability of core knowledge in the generation process, thereby significantly reducing the knowledge forgetting rate and improving the answer accuracy under the same hardware conditions.

[0097] S5. Use the spatiotemporal weight factor to modify the core knowledge coverage of the initial answer information and generate a spatiotemporal constraint coverage evaluation result. The specific implementation is as follows:

[0098] Semantic features that match the core knowledge tags of the scenic area knowledge base are extracted from the initial answer information to generate a core knowledge feature set. Core knowledge tags are predefined knowledge classification identifiers in the scenic area knowledge base, including historical event tags, cultural term tags, and geographic coordinate tags. Semantic feature extraction is achieved through a pre-trained word vector model. The word vector model uses the Word2Vec model trained on public corpus. After segmenting the text in the initial answer information into vocabulary units, the semantic similarity between each vocabulary unit and the core knowledge tag is calculated. The semantic similarity is calculated using the cosine similarity algorithm, and vocabulary units with similarity above the set threshold are selected as core knowledge features.

[0099] The core knowledge feature set is prioritized based on the normalized spatiotemporal weight factor set to generate a core knowledge feature priority sequence. The normalized spatiotemporal weight factor set is derived from the standardized weight value generated in step S4, and each weight factor is associated with a specific spatiotemporal hotspot area and a knowledge dependency relationship. The priority sorting rule is: match each feature in the core knowledge feature set with the weight value of the corresponding feature in the normalized spatiotemporal weight factor set, sort from high to low by weight value, and generate a priority sequence. If a feature does not match a weight value in the normalized spatiotemporal weight factor set, it is assigned the lowest priority by default and placed at the end of the sequence.

[0100] By comparing the core knowledge feature priority sequence with the semantic coverage of the initial answer information, feature segments with missing coverage are identified. The semantic coverage is determined by the core knowledge features that already exist in the initial answer information. The specific method is: the initial answer information is converted into a semantic vector through a word vector model, and the similarity is matched with each feature vector in the core knowledge feature set. If the similarity is higher than the threshold, it is determined to be covered. The rule for determining missing coverage is: in the core knowledge feature priority sequence, if the top N features are not covered in the initial answer information, it is determined to be missing coverage. The value of N is dynamically adjusted according to the length of the priority sequence. The rule is that when the sequence length is greater than or equal to 5, N=3, and when it is less than 5, N=1.

[0101] Dynamically relevance-matching is performed on feature segments with missing coverage and spatiotemporal weighting factors to generate a coverage assessment result with spatiotemporal constraints. The dynamic relevance matching rule is as follows: traverse the feature segments with missing coverage and retrieve their associated normalized spatiotemporal weighting factors. If the weighting factor is above a preset threshold, a coverage defect flag is generated and the weight value is recorded. If the weighting factor is below the threshold, the defect is marked as a low-priority defect. The coverage assessment result is output as a set of defect flags, with high-priority defects receiving priority compensation in subsequent processing.

[0102] S6. When the coverage evaluation result does not meet the threshold, the core knowledge data of the corresponding keyword set in the scenic area knowledge base is called, appended to the associated data, and the generative AI model is retrained to output the final answer information to the mobile terminal. The specific implementation is as follows:

[0103] When the priority and weight value of the defect mark in the coverage assessment result meet the preset conditions, a knowledge append trigger instruction is generated. The setting rule of the preset conditions is: the defect mark in the coverage assessment result must meet the priority of "high priority" and the associated normalized spatiotemporal weight factor value is greater than the preset threshold, or the weight value ranks in the top 20% of all defect marks. The priority classification is set according to the scenic area knowledge coverage strategy. "High priority" corresponds to core knowledge loopholes that need to be repaired immediately, and "low priority" corresponds to minor knowledge defects that can be processed later. The preset threshold is 0.5, and the top 20% of the weight ranking is a mark with a weight value ≥0.6. For marks that meet the trigger conditions, two knowledge append trigger instructions are generated. The trigger instruction contains knowledge tags, weight values and priority information to guide subsequent knowledge retrieval.

[0104] Based on the knowledge append trigger, core knowledge data corresponding to the missing coverage feature fragments is retrieved from the scenic area knowledge base to generate a knowledge append set. The retrieval process is achieved through a combination of exact matching of knowledge tags and semantic expansion matching: Exact matching prioritizes searching for knowledge entries that are completely consistent with the trigger tag within the "Historical Events," "Geographic Coordinates," and "Cultural Terms" categories of the scenic area knowledge base. If no exact match is found, the search scope is expanded through semantic similarity calculation, screening for related knowledge with a similarity above a threshold (e.g., 0.7) and adding them to the knowledge append set. Entries in the knowledge append set are sorted by knowledge type and authority score. The authority score is calculated based on the literature level of the data source (e.g., first-level historical materials, second-level research papers) and the expert review status (e.g., reviewed / unreviewed), ensuring that highly authoritative entries are retained first.

[0105] The knowledge append set is dynamically appended to the dynamically weighted linked data set to form an enhanced linked data set. The dynamic append operation includes data deduplication, weight fusion, and timeliness verification: If an entry in the knowledge append set already exists in the dynamically weighted linked data set, the weights of the two are compared, the larger value is retained, and the timeliness label (e.g., "latest version") is updated. If the entry is new, an initial weight value (e.g., 1.0) is assigned and the timeliness label "newly added" is added. The enhanced linked data set is stored using a dual index of timestamps and weight values to ensure that high-weight, high-timeliness data is prioritized during training.

[0106] The decoder layer of the generative AI model is incrementally trained based on the enhanced associated data set, and the model parameters are updated. Incremental training adopts a local parameter optimization strategy. The specific steps are as follows: the parameters of the encoder layer of the generative AI model are fixed, and only the fully connected network parameters of the decoder layer are back-propagated and optimized. The training data consists of the knowledge content and its weight values in the enhanced associated data set. The loss function is designed to be the sum of a weighted cross-entropy loss and a regularization term. The weighted cross-entropy loss adjusts the loss contribution based on the weight values of the knowledge items, and the regularization term is used to prevent overfitting of the decoder layer. The number of training iterations is dynamically adjusted based on the size of the enhanced associated data set. The rule is: 10 iterations for less than 50 data items, 20 iterations for 50-100 data items, and 30 iterations for more than 100 data items. The convergence of the loss function is verified after each iteration. If the loss decreases by less than 1% for 5 consecutive iterations, training is terminated early.

[0107] The updated generative AI model performs semantic fusion calculations on the initial response information to generate the final response information, which is then output to the mobile terminal. This semantic fusion calculation is implemented by combining an attention mechanism with contextual splicing: the initial response information is concatenated with high-weighted knowledge text from the enhanced contextual data as an input sequence, and a joint semantic representation is generated through the encoder. During decoding, the model's attention layer dynamically assigns attention weight to the newly added knowledge to generate the fused response text. After grammatical correction, redundancy filtering, and sensitive word detection, the final response information is pushed to the visitor's device via the mobile terminal's network interface. The push format is JSON structured data, including text content, knowledge source annotation, and timeliness tags.

[0108] The solution of this embodiment solves the common problem of insufficient core knowledge coverage of generative AI models in scenic spot knowledge services through multi-dimensional feature coupling and dynamic iteration mechanism. Existing technologies usually rely on knowledge update frequency or static label weighting, and do not consider the dynamic correlation between the physical space aggregation state and the knowledge topological structure, resulting in highly chaotic knowledge interfering with the expression of core knowledge. By introducing a spatiotemporal weight factor as a dynamic regulator of knowledge coverage, the spatiotemporal clustering results of tourist behavior trajectories are fused with the spectral clustering structure entropy of the knowledge graph, so that the weight factor reflects both the spatial clustering intensity and the complexity of the knowledge network, forming a dual-domain feature coupling mechanism. The coverage evaluation results trigger targeted knowledge appending and decoder incremental training, realizing the closed-loop optimization of "evaluation-append-training-generation", significantly reducing computing resource consumption. The spatiotemporal dynamics and knowledge topological evolution are incorporated into a unified weight framework, and through priority-driven local parameter optimization, the interference of irrelevant knowledge on the generation process is blocked, thereby improving the answer accuracy and knowledge stability under the same hardware conditions.

[0109] Notably, the collaborative architecture of a cloud data center and WeChat mini-programs enables intelligent scenic spot navigation services. The cloud data center integrates a structured scenic spot knowledge base, where data is categorized and stored by geographic coordinates, historical events, and cultural attributes. For example, geographic coordinate data includes the latitude and longitude of scenic spots and their coverage radius, while historical event data is linked to a timeline and a person-to-person relationship graph. When the WeChat mini-program obtains a visitor's real-time location through an electronic map interface, it triggers a keyword search mechanism within the knowledge base. Entity nouns are extracted from the visitor's question text, matched to structured data of corresponding dimensions in the knowledge base, and the associated content is dynamically fed into a pre-trained generative AI model for real-time fine-tuning. This fine-tuning utilizes a lightweight gradient update strategy, adjusting only the decoder layer parameters to adapt to the scenic spot's knowledge characteristics, generating accurate responses that incorporate geographic coordinates and historical allusions. Furthermore, a hybrid recommendation algorithm (collaborative filtering + knowledge graph path recommendation) generates personalized guided tours based on visitor location, length of stay, and historical interaction data. The cloud architecture utilizes a vertically sharded database to isolate data across multiple scenic spots. Redis caches hotspot knowledge data to reduce retrieval latency. Intelligent routing middleware dynamically allocates computing resources, and containerized virtual scheduling supports high-concurrency access. Deeply couple the static knowledge base, dynamic AI training and scenario-based recommendation system to form a closed loop of "data-model-service".

[0110] Example 2: Figure 2 The present invention provides a structural diagram of a scenic area tour companion service system based on generative AI and knowledge base linkage. The scenic area tour companion service system based on generative AI and knowledge base linkage includes the following modules:

[0111] The question retrieval module is used to obtain the question information input by tourists through mobile terminals in real time, extract the keyword set in the question information and retrieve the matching related data from the scenic spot knowledge base;

[0112] The thermal weighting module is used to calculate the tourist concentration density of the scenic spots corresponding to the associated data based on the tourist distribution thermal data to adjust the priority weight of the real-time updated knowledge in the associated data and generate a dynamically weighted associated data set;

[0113] The model adaptation module is used to input the associated data set and question information into the pre-trained generative AI model, complete real-time adaptation training through dynamic parameter adjustment, and generate initial answer information;

[0114] The spatiotemporal weight module is used to generate spatiotemporal weight factors based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph;

[0115] The coverage correction module is used to correct the core knowledge coverage of the initial answer information using the spatiotemporal weight factor and generate a coverage evaluation result with spatiotemporal constraints;

[0116] The result output module is used to call the core knowledge data of the corresponding keyword set in the scenic spot knowledge base when the coverage evaluation result does not meet the threshold, append it to the associated data, retrain the generative AI model, and then output the final answer information to the mobile terminal.

[0117] The calculations involved in the embodiments are all dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to actual conditions.

[0118] It should be noted that the present invention can be deployed on the device itself to implement embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0119] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0120] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0122] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0123] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0124] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0125] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0126] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A scenic area tour service method based on generative AI and knowledge base linkage, characterized by: The steps include: S1. Real-time acquisition of question information input by tourists through mobile terminals, extraction of keyword sets in the question information and retrieval of matching related data from the scenic spot knowledge base; S2. Calculate the tourist concentration density of the scenic spots corresponding to the associated data based on the tourist distribution thermal data to adjust the priority weight of the real-time updated knowledge in the associated data and generate a dynamically weighted associated data set; S3. Input the associated data set and question information into the pre-trained generative AI model, complete real-time adaptive training through dynamic parameter adjustment, and generate initial answer information; S4. Generate spatiotemporal weight factors based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph; S5. Use the spatiotemporal weight factors to modify the core knowledge coverage of the initial answer information and generate a spatiotemporal constraint coverage evaluation result, including: Extract semantic features that match the core knowledge labels of the scenic spot knowledge base from the initial answer information to generate a core knowledge feature set; Prioritizing the core knowledge feature set based on the normalized spatiotemporal weight factor set to generate a core knowledge feature priority sequence; According to the comparison between the priority sequence of core knowledge features and the semantic coverage of the initial answer information, the feature segments with missing coverage are identified; Dynamically match the coverage-missing feature fragments with spatiotemporal weight factors to generate spatiotemporal constrained coverage evaluation results; S6. When the coverage evaluation result does not meet the threshold, the core knowledge data of the corresponding keyword set in the scenic area knowledge base is called, appended to the associated data, and the generative AI model is retrained to output the final answer information to the mobile terminal.

2. The scenic area tour service method based on generative AI and knowledge base linkage according to claim 1 is characterized in that: Acquire the question information input by tourists through mobile terminals in real time, extract the keyword set in the question information and retrieve the matching related data from the scenic spot knowledge base, including: Collect tourists' question information in real time through the input interface of the mobile terminal; Perform word segmentation on the question information and filter out stop words to generate an initial keyword set; Filter entity nouns that match the scenic spot knowledge base from the initial keyword set to form a final keyword set; Match the coordinate data, historical event data, and cultural explanation data of the corresponding scenic spots in the scenic area knowledge base according to the final keyword set; Merge the matched coordinate data, historical event data, and cultural interpretation data into related data.

3. The scenic area tour service method based on generative AI and knowledge base linkage according to claim 2 is characterized in that: The scenic spot knowledge base contains structured data stored according to geographical distribution, historical background and cultural attributes.

4. The scenic area tour service method based on generative AI and knowledge base linkage according to claim 1 is characterized in that: Based on the tourist distribution thermal data, the tourist concentration density of the corresponding scenic spots in the associated data is calculated to adjust the priority weight of the real-time updated knowledge in the associated data, and generate a dynamically weighted associated data set, including: Obtain the real-time tourist location coordinates of the corresponding scenic spots based on tourist distribution thermal data; The geographical area of each scenic spot is divided into grid cells of a preset size, and the number of real-time tourist location coordinates in each grid cell is counted as the tourist gathering density; According to the density of tourist gatherings and the matching degree of content tags of real-time updated knowledge in the scenic area knowledge base, the priority weight coefficient of real-time updated knowledge is set; The real-time updated knowledge is dynamically weighted based on the priority weight coefficient to generate a dynamically weighted associated data set containing weighted values.

5. The scenic area tour service method based on generative AI and knowledge base linkage according to claim 1 is characterized in that: The associated data set and question information are input into the pre-trained generative AI model. Through dynamic parameter adjustment, real-time adaptive training is completed and initial answer information is generated, including: Vectorize and concatenate the weighted values in the dynamically weighted associated data set with the question information to generate a fusion input vector; Dynamically adjust the attention weight distribution ratio of the generative AI model based on the distribution characteristics of the weighted values in the fusion input vector; Perform gradient descent optimization on the decoder layer of the generative AI model based on the adjusted attention weight distribution ratio to update the model parameters; The updated generative AI model is used to infer and calculate the question information to generate initial answer information containing the core knowledge features of the scenic spot.

6. The scenic area tour service method based on generative AI and knowledge base linkage according to claim 1 is characterized in that: Based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph, spatiotemporal weight factors are generated, including: The continuous positioning data of tourists' mobile terminals are segmented in time and space, and trajectory segments with a density higher than a preset threshold are identified as spatiotemporal hotspots; Construct a logical dependency graph between events, people, and buildings in the scenic area knowledge base and calculate the intermediary index of each node; The spatiotemporal correlation strength is calculated based on the overlapping area between the geographical boundaries of the spatiotemporal hotspot area and the associated areas of high-betweenness nodes in the logical dependency graph; The topological structure disorder index of the logical dependency graph and the spatiotemporal correlation strength are normalized and weighted and fused to generate the spatiotemporal weight factor. Normalize the spatiotemporal weight factors and output a set of normalized spatiotemporal weight factors.

7. The scenic area tour service method based on generative AI and knowledge base linkage according to claim 1 is characterized in that: When the coverage evaluation result does not meet the threshold, the core knowledge data of the corresponding keyword set in the scenic area knowledge base is called, appended to the associated data, and the generative AI model is retrained to output the final answer information to the mobile terminal, including: When the priority and weight value of the defect mark in the coverage assessment result meet the preset conditions, a knowledge append trigger instruction is generated; According to the knowledge addition trigger instruction, the core knowledge data corresponding to the coverage-missing feature fragments are retrieved from the scenic area knowledge base to generate a knowledge addition set; Dynamically appending the knowledge append set to the dynamically weighted linked data set to form an enhanced linked data set; Incrementally train the decoder layer of the generative AI model based on the enhanced set of associated data to update the model parameters; The updated generative AI model performs semantic fusion calculations on the initial answer information to generate the final answer information and output it to the mobile terminal.

8. A scenic area tour companion service system based on generative AI and knowledge base linkage, used to implement the scenic area tour companion service method based on generative AI and knowledge base linkage according to any one of claims 1 to 7, characterized in that: Includes the following modules: The question retrieval module is used to obtain the question information input by tourists through mobile terminals in real time, extract the keyword set in the question information and retrieve the matching related data from the scenic spot knowledge base; The thermal weighting module is used to calculate the tourist concentration density of the scenic spots corresponding to the associated data based on the tourist distribution thermal data to adjust the priority weight of the real-time updated knowledge in the associated data and generate a dynamically weighted associated data set; The model adaptation module is used to input the associated data set and question information into the pre-trained generative AI model, complete real-time adaptation training through dynamic parameter adjustment, and generate initial answer information; The spatiotemporal weight module is used to generate spatiotemporal weight factors based on the spatiotemporal clustering results of tourist behavior trajectories and the spectral clustering structure entropy of the dynamic knowledge graph; The coverage correction module is used to correct the core knowledge coverage of the initial answer information using the spatiotemporal weight factor and generate a coverage evaluation result with spatiotemporal constraints; The result output module is used to call the core knowledge data of the corresponding keyword set in the scenic spot knowledge base when the coverage evaluation result does not meet the threshold, append it to the associated data, retrain the generative AI model, and then output the final answer information to the mobile terminal.

Images