AIGC-based tourist attraction personalized propaganda copywriting automatic generation method and system

Abstract

The invention discloses an AIGC-based tourist attraction personalized publicity copywriting automatic generation method and system, and relates to the technical field of artificial intelligence and natural language processing. Through multi-modal data fusion and tourist state dynamic modeling, the problem that a traditional method cannot perceive tourist interests and emotion changes in real time is solved, and accurate input is provided for personalized generation. Deep integration and accurate calling of scenic region cultural knowledge are realized by utilizing fine-grained knowledge graph and graph neural network retrieval, and the defects that the existing scheme is empty in content and lacks cultural connotation are overcome; through reinforcement learning strategy control and a multi-task hierarchical generation model, adaptive precise regulation and control of copywriting styles, emotions and information are realized, culture accuracy and emotional resonance are ensured, content generation efficiency and personalized matching degree are greatly improved, and quality, suitability and propagation value of tourism propaganda content are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence and natural language processing, and particularly relates to a personalized tourism scenic spot propaganda script automatic generation method and system based on AIGC. BACKGROUND

[0002] As an important medium connecting scenic spots and tourists, tourism scenic spot propaganda scripts play a key role in stimulating tourism interest, spreading cultural value and improving tourist experience. The content quality of the tourism scenic spot propaganda scripts is directly related to the attractiveness and market competitiveness of the destination. However, the production of current tourism propaganda content generally faces systematic challenges. On the one hand, the traditional mode relying on manual creation is inefficient and costly, and it is difficult to meet the massive personalized content demand. On the other hand, although there are automatic schemes based on templates, these methods are often seriously homogeneous in content and lack deep adaptation to individual characteristics of tourists, real-time situations and cultural connotations.

[0003] In practical applications, the existing technology shows significant limitations. Most automatic generation systems only make static recommendations based on simple user tags or historical behaviors, and cannot dynamically perceive real-time state changes of tourists in the touring process, such as interest shifts caused by geographical position movement, emotional fluctuations caused by environmental factors (weather, season) and preference evolution reflected by social media behaviors. At the same time, these systems often fail to deeply integrate historical and cultural knowledge specific to scenic spots, resulting in a lack of cultural accuracy and emotional resonance in the generated content, and failing to reflect the unique value of the scenery, such as allusions and folk customs. SUMMARY

[0004] The purpose of the present application is to provide a personalized tourism scenic spot propaganda script automatic generation method and system based on AIGC, which can dynamically generate and optimize personalized propaganda scripts deeply adapted to individual characteristics of tourists, real-time situations and cultural backgrounds by real-time perception of tourist states through multi-modal data, combined with fine-grained scenic knowledge graphs and reinforcement learning strategies.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] The present application provides a personalized tourism scenic spot propaganda script automatic generation method based on AIGC, which comprises the following steps:

[0007] S1, real-time acquisition of heterogeneous data of tourists through Internet of Things sensors deployed in the scenic environment and user terminals, alignment and fusion of the heterogeneous data by using a multi-modal deep neural network to obtain quantitative representation of the tourist state;

[0008] S2, through the pre-constructed fine-grained scenic spot knowledge graph, combining the spatio-temporal coordinates and interest labels in the quantitative representation of the real-time tourist state, using a semantic retrieval algorithm based on a graph neural network, a personalized situational knowledge set is generated;

[0009] Among them, the pre-constructed fine-grained scenic spot knowledge graph contains scenic landscape entities, historical events, cultural elements and their spatio-temporal semantic relationships;

[0010] S3, through the reinforcement learning strategy network, the quantitative representation of the tourist state is analyzed, a mapping model from the tourist characteristics to the style control parameters of the text is established, and the optimal style control parameters are automatically generated;

[0011] S4, integrating the quantitative representation of the tourist state, the situational knowledge set and the optimal style control parameters, inputting a multi-task generation model based on the Transformer architecture, through the implementation of a three-layer progressive decoding strategy, outputting multiple versions of personalized candidate texts;

[0012] S5, through the content evaluation model, the multi-dimensional comprehensive score of the multiple versions of the candidate texts is carried out, combined with the format specification requirements of the target publishing channel, using an adaptive conversion algorithm for structure reorganization and format adaptation, generating the final content of different media;

[0013] S6, through the point monitoring module, the exposure, interaction and conversion behavior data of the user to the generated final content are collected in real time, using an attribution analysis model to associate the user behavior data with the corresponding generation control parameters and tourist characteristics, forming a traceable generation effect evaluation record;

[0014] S7, based on the constructed reinforcement learning model infrastructure, using the generation effect evaluation data as training samples, the mapping model of the generation control parameters is optimized by policy gradient, and the multi-task generation model is incrementally trained.

[0015] The application provides an AIGC-based personalized promotional text automatic generation system for scenic spots, which is used to realize the AIGC-based personalized promotional text automatic generation method for scenic spots, comprising:

[0016] The tourist feature perception module collects tourist positioning, environment, interaction and social data in real time through multi-source Internet of Things sensors, aligns and fuses them using a multi-modal deep neural network, and outputs a quantitative feature vector containing spatio-temporal preferences, interest intensity and emotional state;

[0017] The situational knowledge retrieval module, based on the pre-constructed fine-grained scenic spot knowledge graph, combining the real-time location, time and interest labels of the tourists, using a graph neural network semantic retrieval algorithm, outputs a personalized knowledge set highly matched with the current tourist situation;

[0018] The style parameter generation module maps the tourist feature vector into the style control parameters of the cultural concentration coefficient, the language style code, the emotional tone vector and the information density threshold through an Actor-Critic reinforcement learning network, and outputs an optimal parameter combination for guiding the generation of a script;

[0019] The content generation optimization module integrates the tourist features, situational knowledge and style parameters, and implements three-layer progressive decoding through a multi-task generation model based on a Transformer to output a set of multi-version personalized candidate scripts filtered in quality;

[0020] The multi-dimensional evaluation adaptation module adopts a three-dimensional evaluation model of cultural fit, personalized matching degree and propagation potential to score the candidate scripts in parallel, intelligently reconstructs the format in combination with the target channel specifications, and outputs the final content ready for the channel and the corresponding release tracking identifier;

[0021] The effect monitoring and attribution module collects user exposure, interaction and conversion behavior data in real time through a point embedding monitoring system, establishes a causal relationship between behavior and generated parameters by using a deep attribution analysis model, and outputs a structured evaluation record containing parameter impact analysis and optimization suggestions;

[0022] The closed-loop optimization and evolution module builds a reinforcement learning environment based on the effect evaluation data, optimizes the style parameter generation model online, and implements incremental training on the content generation model, and outputs the model parameters and performance indicators that are continuously iterated.

[0023] The beneficial effects of the present application are:

[0024] The present application acquires and fuses multi-source heterogeneous data such as the spatio-temporal trajectory and emotional state of tourists in real time through the deployment of Internet of Things sensors and multi-modal deep neural networks, constructs a dynamically updated quantitative representation of the state of tourists, effectively solves the problem that the traditional method in the background art cannot dynamically perceive the real-time state change of tourists (such as geographical position movement and environmental factor influence), and realizes the accurate capture of the real-time interest, emotion and behavior evolution of tourists, thereby laying an accurate data foundation for subsequent personalized content generation;

[0025] By introducing a fine-grained scenic spot knowledge graph and a graph neural network semantic retrieval algorithm, the present application can automatically retrieve and generate a personalized situational knowledge set that fuses cultural allusions and historical events according to the real-time state of tourists, overcoming the limitations of existing systems that are difficult to deeply integrate unique historical and cultural knowledge of scenic spots, leading to a lack of cultural accuracy and emotional resonance in the content, so that the generated promotional scripts not only dynamically adapt to the interests of tourists, but also deeply reflect the cultural value and unique connotation behind the landscape, significantly improving the accuracy and appeal of the content;

[0026] The three-layer progressive decoding strategy combining the reinforcement learning strategy network and the multi-task generation model is adopted, the adaptive mapping from the tourist features to the copywriting style control parameters is realized, and multiple versions of personalized candidate copywriting can be output, the bottleneck of low efficiency and serious content homogenization caused by the dependence on artificial or template method is broken, the automation degree and efficiency of content production are greatly improved, and the copywriting is ensured to be highly matched with the individual tourists and the real-time scene in terms of style, emotion and information density, so that the attractiveness and propagation effect of the propaganda content are comprehensively improved. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to better understand and implement, the technical solutions of the present application are described in detail below in combination with the drawings.

[0028] Figure 1 A flowchart of a method for automatically generating personalized copywriting for a tourist attraction based on AIGC is provided for Embodiment 1 of the present application.

[0029] Figure 2 A flowchart of step S2 in the method for automatically generating personalized copywriting for a tourist attraction based on AIGC is provided for Embodiment 1 of the present application.

[0030] Figure 3 A flowchart of step S4 in the method for automatically generating personalized copywriting for a tourist attraction based on AIGC is provided for Embodiment 1 of the present application.

[0031] Figure 4 A structural diagram of a system for automatically generating personalized copywriting for a tourist attraction based on AIGC is provided for Embodiment 2 of the present application. DETAILED DESCRIPTION

[0032] In order to better understand and implement, the technical solutions of the present application are described in detail below in combination with the drawings.

[0033] The terms used in the present application are merely for the purpose of describing specific embodiments and are not intended to limit the present application. The singular forms "a", "an" and "the" used in the present application are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein means and includes any or all possible combinations of one or more associated listed items.

[0034] The specific embodiments, features and effects of the present application are described in detail below in combination with the drawings and preferred embodiments.

[0035] Embodiment 1

[0036] Referring to Figures 1-3 The embodiment provides an AIGC-based personalized publicity copy generation method for a tourist attraction, AIGC is an artificial intelligence content generation technology, and the method comprises the following steps:

[0037] S1, through the Internet of Things sensors deployed in the environment of the scenic spot and the user terminal, real-time acquisition of the heterogeneous data of the tourist positioning track, environmental parameters, interactive behavior and social media data, alignment and fusion of the heterogeneous data by using a multi-modal deep neural network, extraction of a dynamic feature vector containing spatio-temporal preference, interest intensity and emotional state, and obtaining a quantitative representation of the tourist state;

[0038] Further, step S1 specifically comprises:

[0039] The spatio-temporal trajectory data and environmental parameter data of tourists are collected through the cameras, WiFi probes and environmental sensors deployed in the scenic spot; the interactive behavior data of tourists are collected through the mobile terminal application interface; the text, image and location check-in data published by tourists are collected through the authorized and open social media platform interface;

[0040] The collected heterogeneous data is subjected to cleaning, denoising and format standardization processing: the trajectory data is subjected to coordinate unification and path smoothing processing; the environmental parameter data is subjected to time synchronization and dimension normalization; the text data is subjected to word segmentation, stop word removal and sentiment polarity annotation; the image data is subjected to feature extraction and scene classification annotation;

[0041] The preprocessed multi-modal data is input into a multi-modal deep neural network based on an attention mechanism, the network contains a text encoder, an image encoder and a time series data encoder, and the alignment and weighted fusion of feature vectors of different modalities are realized through a cross-modal attention layer to generate a unified fusion feature representation;

[0042] Based on the unified fusion feature representation, the feature components in three dimensions of spatio-temporal preference intensity, interest label weight and emotional state value are respectively calculated through a fully connected neural network layer, and finally a dynamic feature vector with fixed dimensions is generated as the quantitative representation of the tourist state.

[0043] Among them, the spatio-temporal preference intensity is calculated by the length of stay and the visit frequency of tourists in different landscape areas; the interest label weight is determined by the interaction frequency and depth learning of tourists on various cultural theme contents; and the emotional state value is calculated by combining the sentiment analysis of the social media text of tourists with the environmental parameters.

[0044] Further, a unified fusion feature representation is generated, including: designing a three-way parallel processing architecture containing a text encoder, an image encoder and a time series data encoder for the special needs of personalized tour guide generation; the text encoder adopts a BERT model fine-tuned on a tourism corpus, which specifically extracts interest preferences, sentiment tendencies and cultural semantic features from tourist social media texts; the image encoder is based on the ResNet-50 architecture and adds a scene classification module, which focuses on identifying landscape types, seasonal features and activity scenes in tourist photos; the time series encoder adopts a bidirectional LSTM network, which specifically processes the spatiotemporal patterns of tourist mobile trajectories, stay sequences and interaction behaviors; through a cross-modal attention fusion mechanism, the system establishes a bidirectional alignment relationship between the three modalities of text, image and time series: the text-image attention layer calculates the relevance weight of semantic description and visual scene, the time-space attention layer associates the spatiotemporal patterns of behavior sequence and geographic coordinates, and the global multi-head attention layer comprehensively integrates all modal features for adaptive weighted fusion. This fusion mechanism specially introduces a tourism knowledge guided attention constraint to ensure that cultural related features obtain a higher weight in the fusion process; the final generated joint feature representation is converted through a fully connected layer, outputting a dynamic feature vector containing three dimensions of spatiotemporal preference intensity, cultural interest weight and real-time sentiment state, providing a precise and interpretable tourist state representation for subsequent personalized tour guide generation.

[0045] Specifically, by deploying a sensor network and using a specially designed multi-modal deep neural network, real-time fusion of tourist trajectory, environment, interaction and social media data is achieved, and a dynamic feature vector containing spatiotemporal preference, interest intensity and sentiment state is accurately extracted, directly solving the problem that existing systems in the background technology cannot dynamically perceive the real-time state evolution of tourists caused by changes in geographic location, environmental factors, etc. during the tour, achieving deep, accurate and quantifiable dynamic perception of individual characteristics and real-time context of tourists, laying a reliable data foundation for generating truly personalized tour guides, and overcoming the defects of traditional methods relying on static labels, resulting in superficial adaptation and rigid content.

[0046] S2, through the pre-constructed fine-grained scenic spot knowledge graph, combining the spatiotemporal coordinates and interest labels in the real-time generated quantitative representation of tourist state, using a semantic retrieval algorithm based on graph neural network, dynamically matching the landscape nodes, cultural allusions and activity information related to the current tourist context from the knowledge graph, generating a personalized context knowledge set;

[0047] Among them, the pre-constructed fine-grained scenic spot knowledge graph contains scenic landscape entities, historical events, cultural elements and their spatiotemporal semantic relationships;

[0048] Further, step S2 specifically includes:

[0049] S21, extract the real-time geographic position coordinates, time stamp and interest label set in the quantitative representation of the tourist state, and convert them into a knowledge graph query language, construct a composite query graph pattern combining spatio-temporal constraints and semantic interest constraints, the spatio-temporal constraints limit the time effectiveness and spatial proximity of the nodes, and the interest constraints are based on the semantic similarity between the tourist interest labels and the knowledge nodes;

[0050] S22, input the composite query into the pre-constructed fine-grained scenic spot knowledge graph, use a retrieval algorithm based on a graph attention network to calculate the matching degree of the query conditions and the nodes and relationships in the knowledge graph, traverse the nodes in the graph that meet the conditions, dynamically expand the retrieval range through a graph propagation mechanism, and finally extract a connected knowledge subgraph containing related entities, relationships and attributes;

[0051] S23, structure the knowledge subgraph obtained by retrieval, extract information segments such as landscape description, historical events and cultural elements, sort the knowledge segments according to the importance of the tourist interest labels, and adaptively adjust the knowledge content according to real-time environmental parameters (such as weather and season) to generate a structured personalized situational knowledge set.

[0052] Among them, the structured situational knowledge set is converted into a multi-modal representation form, including a text summary, a keyword set, an associated image identifier and a spatial coordinate sequence, forming a knowledge representation vector that can be directly input into a downstream generation model.

[0053] The pre-constructed fine-grained scenic spot knowledge graph includes: extracting entities, attributes and relationships from official materials, local records, academic literature and authoritative guide texts to establish a basic graph structure; based on user-generated content (UGC) and expert annotation data to complete knowledge and enhance semantic relationships; mapping entities and relationships to low-dimensional vector space through graph embedding algorithm, supporting efficient semantic similarity calculation and graph neural network operation.

[0054] The semantic retrieval algorithm based on graph neural network specifically uses a heterogeneous graph attention network, which can simultaneously process multiple node types (landscapes, events, characters) and relationship types (spatio-temporal relationships, semantic associations, subordinate relationships) in the knowledge graph, and realize accurate situational awareness retrieval.

[0055] Further, the specific implementation of constructing a composite query graph pattern and performing retrieval in step S2 includes:

[0056] Construction of composite query graph pattern: mapping the real-time geographic location coordinates of tourists into spatial constraint conditions, converting time stamps into time validity constraints, and converting the set of interest labels into semantic query vectors. Specifically, a "spatio-temporal-interest" two-dimensional constraint model is established: spatial constraints are limited by a geographic coordinate buffer algorithm to define a search radius (e.g., related nodes within a 500-meter range), time constraints are filtered by a time window function to screen seasonal landscapes and time-sensitive activities, interest constraints are assigned search weights to different interest labels by calculating semantic similarity, forming a weighted multi-constraint query graph pattern.

[0057] Matching degree calculation based on graph attention network: a heterogeneous attention network is used to calculate the matching degree of the composite query and the knowledge graph, including: first, the spatio-temporal constraints and semantic interest constraints in the composite query graph pattern are encoded into multi-dimensional query vectors, where the spatial constraints are converted into coordinate feature vectors through geographic coding, the temporal constraints are extracted through a time encoder, and the interest constraints are generated through a semantic encoder to generate interest embedding vectors; the algorithm uses a heterogeneous graph attention network architecture and designs a node type-aware attention mechanism to calculate attention weights for different node types such as landscape entities, historical events, and cultural elements: for landscape nodes, the spatial proximity weight is calculated, and a distance decay-based spatial attention function is used; for historical event nodes, the time correlation weight is calculated, and a time sequence attention module is used to analyze the time window matching degree; for cultural element nodes, the semantic similarity weight is emphasized, and the query interest vector and node semantic vector are compared through a cross-modal attention layer; in terms of relationship matching degree calculation, the algorithm uses a relationship-aware message passing mechanism to reason along the three core relationship paths of "spatio-temporal association-semantic association-subordinate association": the spatial relationship path calculates the coordinate distance and topological connection strength, the temporal relationship path analyzes the event time sequence logic and seasonal matching degree, and the semantic relationship path evaluates the concept correlation degree; finally, the multi-head attention aggregation layer integrates the node-level matching degree and the relationship-level matching degree to generate a comprehensive matching score for each candidate node, and sets an adaptive threshold to dynamically control the search range, achieving precise and efficient context-aware knowledge retrieval.

[0058] Specifically, by using the pre-constructed fine-grained scenic spot knowledge graph and the graph attention network algorithm, and combining the real-time state of tourists, the related landscape, history, and cultural information are dynamically retrieved and fused to generate a personalized context knowledge set, solving the core problem in the background technology that existing automatic generation systems cannot deeply integrate unique historical and cultural knowledge of scenic spots, resulting in a lack of cultural accuracy and emotional resonance in the promotional content, and achieving deep mining and precise adaptation of unique values such as scenic anecdotes and folk customs, thereby ensuring that the generated copy is not only highly relevant to the current spatio-temporal position and interests of tourists, but also conveys deep cultural connotations, effectively improving the attractiveness and value of the content.

[0059] S3, analyze the quantitative representation of the tourist state through the reinforcement learning strategy network, establish a mapping model from the tourist characteristics to the style control parameters, and automatically generate optimal style control parameters including cultural concentration coefficients, language style codes, emotional tone vectors, and information density thresholds to provide accurate guiding parameters for the AIGC generation process;

[0060] Further, step S3 specifically includes:

[0061] Extracting the preference feature vector, emotional state indicator, and cultural recognition parameter in the quantitative representation of the tourist state to form a standardized input feature, and simultaneously constructing a multi-dimensional parameter space.

[0062] The multi-dimensional parameter space includes: a continuous value cultural concentration coefficient for quantifying the density of cultural elements, a discrete language style code selected from a preset style word table, an emotional tone vector representing emotional tendency and intensity in a multi-dimensional emotional space, and an adjustable information density threshold for controlling the saturation of text information.

[0063] An Actor-Critic dual network architecture is constructed, in which the policy network (Actor) receives the standardized feature vector as the state input and generates the cultural concentration coefficient, language style code, and emotional tone vector through three parallel output layers. The value network (Critic) evaluates the expected return of state-action pairs, uses the tourist feature-parameter pairs of historical high-quality scripts for supervised pre-training to establish a basic mapping relationship and complete model initialization.

[0064] The trained model generates basic parameters based on real-time tourist features, dynamically adjusts them in combination with environmental factors such as time, weather, and holidays, and outputs optimal control parameters highly adapted to the current scene.

[0065] Numerical boundary checking, logical consistency checking, and abnormal pattern recognition are performed on the generated optimal control parameters. When abnormal output is detected, automatically switch to a backup parameter generation mode based on rules to ensure the stability of the AIGC generation process, and record abnormal cases for subsequent model optimization.

[0066] The generation of the cultural concentration coefficient uses a weighted fusion algorithm based on time decay. Specifically, according to the content interaction records of tourists on different cultural dimensions such as history, folk customs, and art, the interaction frequency and average interaction depth of each dimension within a preset time window are calculated, combined with preset dimension weight coefficients, and normalized by weighted summation to output the cultural concentration coefficient.

[0067] The generation of the emotional tone vector adopts a multi-source emotion fusion technology, specifically: performing emotion analysis on social media texts published by tourists to extract multi-dimensional emotional features; meanwhile, an environmental emotion analysis model is used to analyze the current weather, season and scene atmosphere to generate environmental emotion influencing factors; the text emotional features and environmental emotion factors are weighted and fused, and vector projection and standardization are performed in a preset emotion space to form the final emotional tone vector.

[0068] The determination of the information density threshold value adopts a multi-factor dynamic adjustment algorithm, specifically: matching the reading ability level according to the age characteristics of the tourists, determining the screen information bearing parameter according to the terminal device type, and calculating the environmental interference coefficient according to the real-time environmental complexity (such as the density of people flow, weather conditions); input the above three types of parameters into the adjustment function, dynamically calculate and output the information density threshold value that adapts to the current comprehensive conditions.

[0069] Specifically, an adaptive mapping model from the real-time state of the tourists to the copywriting style control parameter is constructed through a reinforcement learning strategy network, and the optimal control parameter can be dynamically generated by comprehensively considering environmental factors, which solves the core defects of the existing automatic methods in the background technology, i.e., the dependence on static templates or simple labels leads to content homogenization and lack of deep personalization, and realizes accurate quantification and dynamic regulation of the copywriting cultural density, language style, emotional tone and information density, so as to ensure that the generated promotional content can closely match the real-time preferences, emotional state and specific tour situation of each tourist, and significantly improve the individual adaptability and situational awareness of the copywriting.

[0070] S4, input the quantitative representation of the tourist state, the situational knowledge set and the optimal style control parameter into a multi-task generation model based on the Transformer architecture, and implement a three-layer progressive decoding strategy through the attention mechanism in the model to output multiple versions of personalized candidate copywriting;

[0071] Further, step S4 specifically includes:

[0072] S41, concatenate and position encode the quantitative representation of the tourist state, the graph embedding vector of the situational knowledge set and the optimal style control parameter vector to form a unified input sequence, calculate and establish the dynamic semantic association and weight distribution among the multi-source information through the multi-head self-attention mechanism in the Transformer model, and construct a joint representation that integrates personalized needs, domain knowledge and style instructions for the subsequent generation process;

[0073] S42, based on the joint representation, an initial script is generated that is semantically coherent and meets the preset constraint conditions using a three-layer progressive decoding strategy, wherein the first layer of the decoder generates a script framework containing themes, structures, and core information points based on the tourist characteristics; the second layer precisely injects relevant entities and attributes in the context knowledge set into the corresponding positions of the framework through a knowledge gating attention mechanism; the third layer iteratively optimizes the vocabulary, sentence structure, and rhetoric of the text according to the style control parameters;

[0074] S43, based on the joint representation and the initial script after layered decoding, a plurality of candidate script versions expressing diversity but consistent in content are generated by adjusting the temperature parameter of the decoding process and applying kernel sampling technology, and then basic quality filtering is performed on the generated multiple versions, including syntax correctness check, basic logic consistency verification, and preset constraint condition compliance check, to eliminate severely unqualified texts, and output a high-quality personalized candidate script set.

[0075] Among them, the preset constraint conditions specifically include: cultural accuracy constraint: requires historical facts, cultural allusions, etc. involved in the generated content to be consistent with the knowledge graph data; personalized adaptation constraint: requires the script content to be highly matched with the preferences and real-time context of the tourists; style consistency constraint: requires the language style and emotional tone of the generated text to be consistent with the target set by the style control parameters; information density constraint: requires the information content of the unit text to meet the set information density threshold range.

[0076] The knowledge gating attention mechanism realizes differential fusion weight control of different categories of knowledge fragments by introducing a knowledge type gating vector in cross-attention calculation.

[0077] Further, through the multi-head self-attention mechanism in the Transformer model, dynamic semantic association and weight distribution among multi-source information are calculated and established, specifically including: the tourist state representation vector, the context knowledge graph embedding vector and the style control parameter vector are spliced and added with position encoding to form a unified input sequence; then the query vector, the key vector and the value vector are calculated in parallel through the multi-head attention mechanism, wherein each attention head focuses on different dimensional semantic association, some attention heads focus on analyzing the matching relationship between tourist features and knowledge elements, some attention heads focus on the adaptation degree of style instructions and content framework, and some attention heads are specially used to process the corresponding relationship between space-time information and event description; through the soft attention weight matrix, the correlation scores between elements in the input sequence are dynamically calculated, so that related knowledge fragments can automatically obtain higher attention weight according to the tourist preference, and the style parameter can effectively guide the rhetorical direction of content generation; the finally formed attention distribution not only reflects the semantic association strength of multi-source information, but also retains the uniqueness of different source information through residual connection and layer normalization, providing comprehensive and accurate joint representation for subsequent decoding generation.

[0078] Specifically, through the multi-task generation model based on Transformer, the tourist state, the context knowledge and the style control parameter are integrated, and the three-layer progressive decoding is realized by using the attention mechanism, which solves the core problems of low efficiency and lack of depth and personalized adaptation of the generated content in the traditional artificial creation or template method in the background technology. Through knowledge gate attention and style parameter guidance, the system can ensure the cultural accuracy, personalized matching degree and style consistency of the content at the same time, realizes efficient and automatic generation of diversified high-quality candidate scripts under the premise of ensuring information accuracy and style adaptation, and solves the efficiency bottleneck and cultural connotation loss problem of traditional content production, thereby significantly improving the output efficiency and comprehensive quality of the propaganda content.

[0079] S5, through the content evaluation model, a multi-dimensional comprehensive score is given to the multi-version candidate script, an adaptive conversion algorithm is used to reorganize the structure and adapt the format of the candidate script according to the format specification requirements of the target release channel, and the final content of different media is generated;

[0080] Further, step S5 specifically includes:

[0081] The content evaluation model is used to evaluate the candidate script in parallel, including a cultural fit module, a personalized matching module and a propagation potential module, which respectively calculate the semantic similarity between the script and the context knowledge set, the adaptation degree of the script features and the tourist state vector, and the propagation potential of the script in the target channel, and output a three-dimensional score vector of each script;

[0082] According to the target channel characteristics and marketing needs, dynamic weight coefficients are configured for the three evaluation dimensions, the comprehensive score is calculated through a weighted fusion algorithm, and hierarchical screening is performed;

[0083] The hierarchical screening strategy includes: eliminating the copy with a cultural fit degree lower than a preset safety threshold, ranking the copies in descending order of comprehensive score, and finally selecting the top N copies as the preferred results to enter the next processing stage, wherein the value of N is dynamically adjusted according to the channel characteristics;

[0084] According to the technical specifications of different media channels, a conversion system combining rule engine and neural network is used to reconstruct the content of the candidate copy, and multiple rounds of quality verification are performed on the reconstructed content. Then the content and channel-specific metadata are standardized and packaged to generate the final content and corresponding publication tracking identifier that can be directly published through the platform API interface.

[0085] The semantic similarity between the copy and the situational knowledge set includes: using a pre-trained semantic encoding model to encode the cultural elements, historical anecdotes and other knowledge fragments in the candidate copy text and the situational knowledge set into high-dimensional semantic vectors respectively; the cosine similarity of the two vectors is calculated, and the entity overlap and semantic association are weighted and fused to finally obtain a semantic similarity score in the interval [0, 1].

[0086] The adaptation degree of the copy features and the tourist state vector includes: extracting the deep semantic features of the candidate copy through a text encoder to form a feature vector, and then performing similarity matching with the tourist state quantitative representation vector in a multi-dimensional feature space. The cosine similarity is used as the basic measure in the matching process, and the matching degree of the tourist interest label is weighted and corrected, and an individualized adaptation degree score is output.

[0087] The propagation potential of the copy in the target channel includes: constructing a propagation prediction model based on historical propagation data and channel characteristics. The model analyzes the emotional polarity, keyword popularity, sentence structure, readability index and other features of the copy, and combines the user portrait and behavior characteristics of the target channel to predict the expected sharing rate, forwarding rate and interaction depth of the copy in the channel, and finally outputs a standardized propagation potential score.

[0088] According to the target channel characteristics and marketing needs, dynamic weight coefficients are configured for the three evaluation dimensions, including: for official propaganda channels (such as official website, guide screen), give higher weight to cultural fit degree (such as 0.5), medium weight to individualized matching degree (such as 0.3), and lower weight to propagation potential (such as 0.2); for social media channels (such as Weibo, Douyin), appropriately increase the weights of propagation potential and individualized matching degree, and reduce the weight of cultural fit degree; for personalized recommendation scenarios (such as APP push), the weight of individualized matching degree is greatly increased, forming a dynamically adjustable weight configuration strategy.

[0089] A conversion system combining a rule engine and neural networks is used to reconstruct candidate text. Specifically, this includes: processing structured format requirements through the rule engine, such as automatically adjusting text length to adapt to the character limits of different channels, optimizing paragraph structure to conform to platform reading habits, processing specific format tags such as topic tags and hyperlinks, and injecting necessary metadata; and performing intelligent content reconstruction while preserving semantics through the neural network module, including attention-based summary generation, sentence style transfer optimization, and intelligent adaptation of multimodal elements, ensuring that the converted content fully complies with the technical specifications and user experience requirements of the target channel while maintaining the core semantics.

[0090] Specifically, by constructing a multi-dimensional content evaluation model and combining it with channel-aware adaptive conversion technology, the system performs refined screening and reconstruction of the generated multi-version candidate copy. This solves the problems of inconsistent content quality and difficulty in meeting the diverse format specifications of different media channels in existing automated generation solutions. By conducting quantitative evaluation and dynamic weighting from multiple perspectives such as cultural accuracy, personalization matching, and dissemination potential, and combining rule and intelligent reconstruction technology, the system ensures that the final published copy not only has cultural depth and personalized adaptability in content, but also perfectly matches the dissemination characteristics and technical requirements of different channels in form. This achieves a closed-loop quality process from automated generation to precise and channel-specific delivery, significantly improving the implementation effect and dissemination efficiency of promotional copy.

[0091] S6. By deploying a tracking module on the content display terminal, real-time data on user exposure, interaction and conversion behavior of the generated final content is collected. Attribution analysis model is used to correlate user behavior data with corresponding generation control parameters and visitor characteristics to form a traceable generation effect evaluation record.

[0092] Further, step S6 specifically includes:

[0093] By deploying standardized tracking components on the content display terminal, the system collects complete behavioral data sequences of users on the final content in real time, including structured raw data such as exposure trigger time, interaction action type, and conversion path nodes. The collected raw data is then cleaned by noise reduction, timestamp alignment, and format standardization to establish a unified set of behavioral data records.

[0094] Establish a reverse correlation mapping from user behavior to content generation process, bind behavioral data to final content through unique content identifiers, trace the correlation of style control parameter set and extracted visitor state feature vector, and adopt a feature alignment module based on attention mechanism to accurately align data from different sources in the time dimension to form a spatiotemporally consistent multi-source dataset.

[0095] The deep attribution analysis model based on the Transformer architecture is constructed, the influence weight of different generation parameters in the user behavior sequence is learned, the causal inference relationship from the control parameter to the behavior result is established, the multi-head attention mechanism is used to analyze the interaction effect between parameters, and the attribution analysis matrix containing the contribution degree of each style parameter is output, which quantitatively shows the influence degree of the cultural concentration coefficient, the emotional tone vector and other parameters on the user behavior;

[0096] The attribution analysis result is integrated with the original data to generate a structured evaluation record containing basic behavior indicators, parameter influence distribution, abnormal pattern recognition and optimization suggestions.

[0097] Among them, the attribution analysis model adopts a feature contribution degree calculation method based on Shapley value to ensure the fairness and interpretability of parameter influence evaluation; the evaluation record contains complete data traceability information to support complete path query from the final behavior to the original generation parameter.

[0098] Specifically, by deploying the buried point monitoring and the deep attribution analysis model, the reverse tracking and causal correlation analysis capability from user behavior data to content generation parameters is established, and the pain points of the existing automatic system in the background technology, such as the inability to dynamically optimize due to the lack of feedback loop and the difficulty in evaluating the content generation effect, are solved. By accurately associating and quantitatively attributing the exposure, interaction and conversion behavior of users with specific copy style parameters and tourist characteristics, the system first realizes the fine evaluation of the effect of personalized copy generation and the deep analysis of the influence factors, thereby providing data-driven decision basis for the continuous iteration and optimization of the generation model, forming a complete closed loop from content generation, release to effect evaluation and optimization, and effectively improving the ability of the system to adaptively evolve and the long-term effectiveness of the promotional content.

[0099] S7, based on the constructed reinforcement learning model infrastructure, taking the generation effect evaluation data as the training sample, performing policy gradient optimization on the mapping model of the generation control parameter, and periodically using the newly added data to perform incremental training on the multi-task generation model, forming a complete closed loop.

[0100] Further, step S7 specifically includes:

[0101] The attribution analysis result is used to construct a reinforcement learning environment, the initialized Actor-Critic network is used as an intelligent agent to be optimized, the causal relationship between the user interaction data and each generation parameter is calculated, a multi-level reward function is established, including a basic conversion rate reward, a content quality score reward and a propagation innovation degree reward, a proximal policy optimization algorithm is used to iteratively update the strategy network parameters, and online continuous optimization of the generation control parameter mapping strategy is realized;

[0102] Filter high-quality new samples from evaluation records, including content data with conversion rates higher than a threshold and their corresponding visitor features and generation parameters, use curriculum learning strategy to incrementally train multi-task generation model, integrate updated content of knowledge graph in training process, embed new cultural elements and historical anecdotes into generation model parameter space through knowledge adaptation layer to maintain timeliness and accuracy of content generation ability;

[0103] Establish a system performance monitoring index system, including generation quality stability index, personalized matching accuracy index and computing resource efficiency index, design an abnormality detection mechanism, automatically trigger a diagnosis process when performance abnormalities are detected, implement model rollback or retraining operations to ensure the coordinated evolution and stable operation of system modules.

[0104] Among them, the weights of each layer of the multi-level reward function are dynamically adjusted according to the training stage, focusing on basic conversion indicators at the initial stage, balancing quality and innovation at the middle stage, and focusing on long-term effect stability at the later stage; the incremental training adopts an elastic sample buffer management strategy; the performance monitoring establishes a real-time early warning and automatic recovery mechanism;

[0105] The multi-level reward function is represented as: R(t)=α(t)⋅Rc+β(t)⋅Rq +γ(t)⋅Ri;Wherein, Rc represents the basic conversion rate reward, reflecting the proportion of user conversion behavior; Rq represents the content quality score reward, coming from the content evaluation model output; Ri represents the propagation innovation degree reward, combining the propagation depth (level) and breadth (forwarding proportion); α(t), β(t), γ(t) are dynamic weights that change over time, used to adjust the importance of different reward components at the current stage.

[0106] The proximal policy optimization algorithm is used to iteratively update the policy network parameters, including: generating a policy distribution based on the current policy network parameters, collecting state-action-reward sequence data through importance sampling; calculating the probability ratio between the new and old policies, and using a clipping objective function to limit the policy update amplitude to ensure training stability; At the same time, an entropy regularization term is introduced to encourage policy exploration and prevent premature convergence to a local optimum; By constructing an advantage function estimator, the generalized advantage estimation algorithm is used to balance immediate rewards and long-term returns to optimize the gradient direction of the generated control parameter mapping strategy and continuously improve it online.

[0107] S7 step is based on the model framework established in S3, using actual data generated after running to complete iterative optimization of the model and closed-loop improvement of the overall performance of the system.

[0108] Specifically, by constructing an adaptive optimization closed loop based on reinforcement learning and incremental training, the user behavior feedback and content evaluation results are converted into the direct driving force for model iteration, solving the fundamental problem of system rigidity caused by the lack of data-driven continuous learning mechanism in the background technology, which cannot adapt to the long-term evolution of tourist preferences and market environment. By using actual interaction data to optimize and incrementally update the strategy network and the generation model online, the transformation from static rule execution to dynamic intelligent evolution is realized, ensuring that the copy generation strategy can continuously track and respond to changes in user preferences and knowledge updates, thereby continuously improving the individualization accuracy, cultural timeliness and overall dissemination efficiency of the content in the long run, and ultimately building an intelligent content generation ecosystem with self-evolution ability.

[0109] Embodiment 2

[0110] Please refer to Figure 4 The embodiment provides an AIGC-based personalized promotional copy automatic generation system for tourist attractions, which is used for realizing an AIGC-based personalized promotional copy automatic generation method for tourist attractions, and comprises the following steps of:

[0111] A tourist feature perception module, which collects tourist positioning, environment, interaction and social data in real time through multi-source Internet of Things sensors, aligns and fuses the data by using a multi-modal deep neural network, and directly outputs a quantitative feature vector containing spatiotemporal preferences, interest intensity and emotional state, thereby providing accurate tourist state representation for subsequent processing;

[0112] A situational knowledge retrieval module, which, based on a pre-constructed fine-grained tourist attraction knowledge graph, combines the real-time location, time and interest label of a tourist, adopts a graph neural network semantic retrieval algorithm, and outputs a personalized knowledge set highly matched with the current tourist situation, including structured information such as landscape description, historical anecdotes and cultural elements;

[0113] A style parameter generation module, which maps the tourist feature vector into cultural concentration coefficients, language style codes, emotional tone vectors and information density thresholds and other style control parameters by using an Actor-Critic reinforcement learning network, and outputs an optimal parameter combination for guiding copy generation;

[0114] A content generation optimization module, which integrates tourist features, situational knowledge and style parameters, implements three-layer progressive decoding by using a multi-task generation model based on a Transformer, and outputs a multi-version personalized candidate copy set filtered in quality, thereby ensuring the accuracy, adaptability and style consistency of the content;

[0115] A multi-dimensional evaluation adaptation module uses a three-dimensional evaluation model of cultural fit, personalized matching degree and communication potential to score candidate scripts in parallel, combines target channel specifications to perform intelligent format reconstruction, and outputs channel-ready final content and corresponding release tracking identifiers;

[0116] An effect monitoring attribution module collects user exposure, interaction and conversion behavior data in real time through a buried point monitoring system, establishes a causal relationship between behavior and generation parameters using a deep attribution analysis model, and outputs structured evaluation records containing parameter impact analysis and optimization suggestions;

[0117] A closed-loop optimization evolution module builds a reinforcement learning environment based on effect evaluation data, performs online optimization on the style parameter generation model, and implements incremental training on the content generation model, outputs continuously iterated model parameters and performance indicators, and realizes self-evolution and capability improvement of the system.

[0118] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any equivalent embodiments with equivalent changes and modifications made to the above embodiments according to the technical essence of the present application are still within the scope of the technical solution of the present application.

Claims

1. A method for automatically generating personalized promotional copy for tourist attractions based on AIGC, characterized by: Includes the following steps: S1. By deploying IoT sensors in the scenic area and user terminals, heterogeneous tourist data is acquired in real time. Multimodal deep neural networks are used to align and fuse the heterogeneous data to obtain a quantitative representation of the tourist status. S2. By using a pre-constructed fine-grained scenic area knowledge graph, combined with the spatiotemporal coordinates and interest tags in the quantitative representation of real-time tourist status, a semantic retrieval algorithm based on graph neural networks is adopted to generate a personalized set of contextual knowledge. The pre-constructed fine-grained scenic area knowledge graph includes scenic area landscape entities, historical events, cultural elements, and their spatiotemporal semantic relationships. S3. Analyze the quantitative representation of tourist status through reinforcement learning policy network, establish a mapping model from tourist characteristics to copywriting style control parameters, and automatically generate the optimal style control parameters. S4. Integrate the quantitative representation of tourist status, contextual knowledge set and optimal style control parameters, input the multi-task generation model based on the Transformer architecture, and output multiple versions of personalized candidate copy by implementing a three-layer progressive decoding strategy. S5. Through the content evaluation model, the candidate texts of multiple versions are comprehensively scored from multiple dimensions. Combined with the format specifications of the target publishing channels, the adaptive conversion algorithm is used to restructure and adapt the format to generate the final content for different media. S6. Through the tracking module, real-time data on user exposure, interaction and conversion behavior of the generated final content is collected. Attribution analysis model is used to correlate user behavior data with corresponding generation control parameters and visitor characteristics to form a traceable generation effect evaluation record. S7. Based on the constructed reinforcement learning model infrastructure, the generated performance evaluation data is used as training samples to optimize the policy gradient of the mapping model of the generation control parameters, while the multi-task generation model is incrementally trained.

2. The method for automatically generating personalized promotional copy for tourist attractions based on AIGC according to claim 1, characterized in that: Step S1 specifically includes: Tourist spatiotemporal trajectory data and environmental parameter data are collected through cameras, WiFi probes and environmental sensors deployed in the scenic area; tourist interaction behavior data is collected through mobile terminal application interfaces; and text, images and location check-in data posted by tourists are collected through authorized open social media platform interfaces. The collected heterogeneous data is cleaned, denoised, and standardized in format, including coordinate unification and path smoothing for trajectory data, time synchronization and dimensional normalization for environmental parameter data, word segmentation, stop word removal, and sentiment polarity labeling for text data; and feature extraction and scene classification labeling for image data. The preprocessed multimodal data is input into a multimodal deep neural network based on an attention mechanism. The network includes a text encoder, an image encoder, and a temporal data encoder. The alignment and weighted fusion of feature vectors from different modalities are achieved through a cross-modal attention layer to generate a unified fused feature representation. Based on a unified fusion feature representation, feature components of three dimensions—spatiotemporal preference intensity, interest tag weight, and emotional state value—are calculated separately through a fully connected neural network layer to generate a dynamic feature vector with fixed dimensions, which serves as a quantitative representation of tourist status.

3. The method for automatically generating personalized promotional copy for tourist attractions based on AIGC according to claim 1, characterized in that: Step S2 specifically includes: S21. Extract the real-time geographic location coordinates, timestamps, and interest tag set from the quantitative representation of tourist status, and transform them into a knowledge graph query language. Construct a composite query graph pattern that combines spatiotemporal constraints and semantic interest constraints. The spatiotemporal constraints limit the temporal validity and spatial proximity of nodes, and the interest constraints are based on the semantic similarity between tourist interest tags and knowledge nodes. S22. Input the composite query into the pre-constructed fine-grained scenic area knowledge graph, use a retrieval algorithm based on graph attention network to calculate the matching degree between the query conditions and the nodes and relationships in the knowledge graph, traverse the nodes in the graph that meet the conditions, dynamically expand the retrieval scope through the graph propagation mechanism, and finally extract the connected knowledge subgraph containing relevant entities, relationships and attributes. S23. Perform structured parsing on the retrieved knowledge subgraph, extract information fragments, sort the knowledge fragments by importance according to the weight of tourist interest tags, and adaptively adjust the knowledge content in combination with real-time environmental parameters to generate a structured personalized contextual knowledge set.

4. The method for automatically generating personalized promotional copy for tourist attractions based on AIGC according to claim 3, characterized in that: The pre-constructed fine-grained scenic area knowledge graph includes: extracting entities, attributes, and relationships from official scenic area materials, local chronicles, academic literature, and authoritative tour guide texts to establish a basic graph structure; performing knowledge completion and semantic relationship enhancement based on tourist-generated content and expert-annotated data; and mapping entities and relationships to a low-dimensional vector space through a graph embedding algorithm to perform semantic similarity calculation and graph neural network operations.

5. The method for automatically generating personalized promotional copy for tourist attractions based on AIGC according to claim 1, characterized in that: Step S3 specifically includes: Extract preference feature vectors, emotional state indicators, and cultural awareness parameters from the quantitative representation of tourist status to form standardized input features, and construct a multi-dimensional parameter space. The multidimensional parameter space includes: a continuous-valued cultural concentration coefficient for quantifying the density of cultural elements, a discrete language style code selected from a preset style vocabulary, an emotional tone vector representing emotional tendency and intensity in the multidimensional emotional space, and an adjustable information density threshold for controlling the saturation of text information. A dual-network architecture of Actor-Critic is constructed, in which Actor receives standardized feature vectors as state input and generates cultural concentration coefficient, language style code and sentiment tone vector through three parallel output layers; Critic evaluates the expected reward of state-action pairs and uses tourist feature-parameter pairs of historical high-quality copywriting for supervised pre-training to establish basic mapping relationships and complete model initialization. The trained model generates basic parameters based on real-time tourist characteristics, dynamically adjusts them in combination with environmental factors, and outputs optimal control parameters that are highly adapted to the current scene. The generated optimal control parameters are subjected to numerical boundary checks, logical consistency checks, and abnormal pattern recognition. When an abnormal output is detected, the system automatically switches to a rule-based backup parameter generation mode.

6. The method for automatically generating personalized promotional copy for tourist attractions based on AIGC according to claim 1, characterized in that: Step S4 specifically includes: S41. The tourist state quantification representation vector, the graph embedding vector of the contextual knowledge set, and the optimal style control parameter vector are concatenated and positionally encoded to form a unified input sequence. Through the multi-head self-attention mechanism in the Transformer model, the dynamic semantic association and weight distribution between multi-source information are calculated and established, so as to build a joint representation that integrates personalized needs, domain knowledge and style instructions for the subsequent generation process. S42. Based on joint representation, a three-layer progressive decoding strategy is adopted to generate semantically coherent initial copy that meets preset constraints. The first layer of the decoder generates a copy framework containing the theme, structure and core information points based on tourist characteristics. The second layer uses a knowledge gating attention mechanism to accurately inject relevant entities and attributes from the contextual knowledge set into the corresponding positions of the framework. The third layer iteratively optimizes the vocabulary, sentence structure and rhetoric of the text based on style control parameters. S43. Based on the joint representation and the initial text after hierarchical decoding, by adjusting the temperature parameters of the decoding process and applying kernel sampling technology, the model is driven to generate multiple candidate text versions with diverse expressions but consistent content. Then, basic quality filtering is performed on the generated multiple versions to output a high-quality personalized candidate text set.

7. The method for automatically generating personalized promotional copy for tourist attractions based on AIGC according to claim 1, characterized in that: Step S5 specifically includes: A content evaluation model is used to evaluate candidate copy in parallel, including a cultural fit module, a personalized matching module, and a dissemination potential module. The semantic similarity between the copy and the contextual knowledge set, the fit between the copy features and the tourist state vector, and the dissemination potential of the copy on the target channel are calculated respectively, and the three-dimensional score vector of each copy is output. Dynamic weighting coefficients are configured for the three evaluation dimensions based on the characteristics of the target channel and marketing needs. A comprehensive score is calculated through a weighted fusion algorithm, and tiered screening is performed. The tiered screening strategy includes: removing copywriting with cultural fit below a preset safety threshold, sorting by comprehensive score in descending order, and finally selecting the top N copywriting as the preferred result to enter the next processing stage, where the value of N is dynamically adjusted according to the characteristics of the channel. To address the technical specifications of different media channels, a transformation system combining a rule engine and a neural network is used to reconstruct candidate text. Multiple rounds of quality verification are then performed on the reconstructed content. Finally, the content and channel-specific metadata are standardized and encapsulated to generate the final content that can be directly published through the platform's API interface, along with the corresponding publication tracking identifier.

8. The method for automatically generating personalized promotional copy for tourist attractions based on AIGC according to claim 1, characterized in that: Step S6 specifically includes: By deploying standardized tracking components on the content display terminal, the system collects complete behavioral data sequences of users on the final content in real time. The collected raw data is then denoised, cleaned, timestamp aligned, and format standardized to establish a unified set of behavioral data records. Establish a reverse correlation mapping from user behavior to content generation process, bind behavioral data to final content through unique content identifiers, trace the correlation of style control parameter set and extracted visitor state feature vector, and adopt a feature alignment module based on attention mechanism to accurately align data from different sources in the time dimension to form a spatiotemporally consistent multi-source dataset. We construct a deep attribution analysis model based on the Transformer architecture. By learning the influence weights of different generation parameters in the user behavior sequence, we establish a causal inference relationship from control parameters to behavioral results. We use a multi-head attention mechanism to analyze the interaction effect between parameters and output the degree of influence of user behavior. The attribution analysis results are integrated with the raw data to generate a structured evaluation record that includes basic behavioral indicators, parameter influence distribution, anomaly pattern identification, and optimization suggestions.

9. The method for automatically generating personalized promotional copy for tourist attractions based on AIGC according to claim 1, characterized in that: Step S7 specifically includes: A reinforcement learning environment is constructed using the generated attribution analysis results. The initialized Actor-Critic network is used as the agent to be optimized. By calculating the causal relationship strength between user interaction data and each generation parameter, a multi-level reward function is established, including basic conversion rate reward, content quality score reward and dissemination innovation reward. The proximal policy optimization algorithm is used to iteratively update the policy network parameters to achieve online continuous optimization of the generation control parameter mapping policy. High-quality new samples are selected from the evaluation records, including content data with a conversion rate higher than the threshold and its corresponding tourist characteristics and generation parameters. A course learning strategy is used to incrementally train the multi-task generation model. During the training process, the updated content of the knowledge graph is integrated synchronously. New cultural elements and historical allusions are embedded into the parameter space of the generation model through the knowledge adaptation layer. Establish a system performance monitoring index system, including generation quality stability index, personalized matching accuracy index and computing resource efficiency index, design an anomaly detection mechanism, and automatically trigger the diagnostic process when performance anomalies are detected, and implement model rollback or retraining operations.

10. An AIGC-based automatic generation system for personalized promotional copy for tourist attractions, applied to the AIGC-based automatic generation method for personalized promotional copy for tourist attractions as described in any one of claims 1-9, characterized in that: include: The tourist feature perception module collects tourist location, environment, interaction and social data in real time through multi-source IoT sensors, and uses multimodal deep neural networks to align and fuse the data, outputting a quantitative feature vector containing spatiotemporal preferences, interest intensity and emotional state. The contextual knowledge retrieval module, based on a pre-built fine-grained scenic area knowledge graph, combines tourists' real-time location, time, and interest tags, and uses a graph neural network semantic retrieval algorithm to output a personalized knowledge set that is highly matched with the current tourist context. The style parameter generation module uses an Actor-Critic reinforcement learning network to map tourist feature vectors into style control parameters such as cultural concentration coefficient, language style code, emotional tone vector, and information density threshold, and outputs the optimal parameter combination to guide copywriting generation. The content generation optimization module integrates tourist characteristics, contextual knowledge, and style parameters. It implements three-layer progressive decoding through a Transformer-based multi-task generation model and outputs a set of personalized candidate texts with multiple versions that have undergone quality filtering. The multi-dimensional assessment and adaptation module uses a three-dimensional assessment model of cultural fit, personalization matching and dissemination potential to score candidate copy in parallel. It also intelligently reconstructs the format in conjunction with the target channel specifications and outputs the final content that is ready for the channel and the corresponding release tracking identifier. The effect monitoring and attribution module collects user exposure, interaction and conversion behavior data in real time through the tracking system, establishes causal relationships between behavior and generated parameters using a deep attribution analysis model, and outputs a structured evaluation record that includes parameter impact analysis and optimization suggestions. The closed-loop optimization and evolution module constructs a reinforcement learning environment based on the effect evaluation data, optimizes the style parameter generation model online, and performs incremental training on the content generation model, outputting continuously iterative model parameters and performance indicators.

Images