Advertising language correction method and device based on large language model, equipment and storage medium

Abstract

The invention is suitable for the field of data processing, and provides an advertising language correction method and device based on a large language model, equipment and a storage medium. Multi-language advertising language data and target area culture rule data are obtained, emotion vectors are generated through analysis, and culture metaphor features are extracted through comparative learning; accurate capture of advertising language emotion and cultural semantics in a cross-cultural context is realized; a culture sensitivity index is calculated in combination with a culture and emotion mapping library to generate a compliant alternative scheme, and a target advertising language with culture adaptability and marketing conversion potential is screened through a conversion rate prediction model; the problems that in a traditional method, emotion misjudgment is prone to occurring in cross-culture advertising words, the culture compliance risk is high, and the marketing conversion effect is difficult to consider are effectively solved, the corrected target advertising words can meet the target area culture rule, audience emotion cognition and marketing conversion requirements at the same time, and the effectiveness of cross-culture advertising is improved.

Description

Technical Field

[0001] This application belongs to the field of data processing, and in particular relates to a method, apparatus, device and storage medium for revising advertising slogans based on a large language model. Background Technology

[0002] Cross-cultural advertising slogan modification refers to the process of adjusting the language, cultural elements, and emotional tone of advertising slogans in cross-regional (e.g., transnational, transcultural) scenarios. This is done to address potential issues such as cultural conflicts, emotional misjudgments, compliance risks, and reduced marketing effectiveness that may arise from language differences, cultural taboos, emotional perception biases, and differing regional compliance rules. Ultimately, the goal is to ensure that the advertising slogan conforms to the cultural customs and laws of the target region while also resonating with the emotional perceptions of the local audience, thus guaranteeing the effectiveness of the advertising campaign. Its core function is to solve the dual problems of language translation and cultural adaptation, avoiding situations where a literal translation is correct but culturally offensive.

[0003] Traditional cross-cultural advertising slogans rely on static emotional lexicons and manual extraction of cultural metaphors during the revision process. This can easily lead to a serious mismatch between the emotional tone of multilingual advertising slogans and the perception of the target audience. In addition, the reliance on manual extraction of cultural metaphors also carries the risk of missing taboo elements, which can easily cause advertising slogans to violate local cultural rules. It is evident that traditional advertising slogan revision has technical problems such as emotional misjudgment and high cultural compliance risks. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus, device and storage medium for correcting advertising slogans based on a large language model, which can solve the problems of emotional misjudgment and high cultural compliance risks in related technologies.

[0005] In a first aspect, embodiments of this application provide a method for revising advertising slogans based on a large language model, including: Acquire multilingual advertising slogans and cultural rules data for the target region; We analyze multilingual advertising data to generate sentiment vectors, and at the same time, we use a contrastive learning algorithm to align textual symbols and visual elements in the multilingual advertising data to extract cultural metaphor features. The system calls a pre-defined cultural and emotional mapping library, combines emotional vectors, cultural metaphor features, and cultural rule data to calculate a cultural sensitivity index. If the cultural sensitivity index is greater than a pre-defined threshold, multiple alternative advertising slogans are generated through a reinforcement learning algorithm. The predicted conversion rate data of various ad copy alternatives in the target area were evaluated by the conversion rate prediction model, and the target alternatives were selected as the corrected target ad copy based on the predicted conversion rate data. Output the target advertising slogan.

[0006] Secondly, embodiments of this application provide an advertising slogan correction device based on a large language model, the device comprising: The acquisition module is used to acquire multilingual advertising slogan data and cultural rule data for the target region. The parsing and extraction module is used to parse multilingual advertising data, generate sentiment vectors, and at the same time, align text symbols and visual elements in multilingual advertising data through a contrastive learning algorithm to extract cultural metaphor features. The generation module is used to call a preset cultural and emotional mapping library, combine emotional vectors, cultural metaphor features and cultural rule data to calculate the cultural sensitivity index. If the cultural sensitivity index is greater than a preset threshold, multiple advertising slogan alternatives are generated through reinforcement learning algorithm. The correction module is used to evaluate the predicted conversion rate data of multiple ad copy alternatives in the target area through the conversion rate prediction model, and select the optimal solution as the corrected target ad copy based on the predicted conversion rate data. The output module is used to output the target advertising slogan.

[0007] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described advertising slogan correction method based on a large language model.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described advertising slogan correction method based on a large language model.

[0009] Fifthly, embodiments of this application provide a computer program product that, when run on a terminal device, causes the terminal device to execute the aforementioned advertising slogan correction method based on a large language model.

[0010] The beneficial effects of this application embodiment compared with the prior art are as follows: This application embodiment first obtains multilingual advertising slogan data and target area cultural rule data, and then uses a pre-trained model to parse and generate sentiment vectors, replacing the traditional static sentiment lexicon. It can accurately capture the sentiment tendency in the cross-cultural context and avoid the disconnect from the cognition of the target area audience. At the same time, it uses a contrastive learning algorithm to align text symbols and visual elements and extract cultural metaphor features, replacing manual extraction, thereby reducing the risk of missing taboo elements. Subsequently, it combines the cultural and sentiment mapping library to calculate the cultural sensitivity index to generate compliant alternatives, and uses a conversion rate prediction model to screen the appropriate solutions, thus solving the technical problems of sentiment misjudgment and high cultural compliance risks in traditional advertising slogan correction. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram illustrating the implementation process of the advertising slogan correction method based on a large language model provided in this application embodiment.

[0013] Figure 2 This is a schematic diagram of the advertising slogan correction device based on a large language model provided in the embodiments of this application.

[0014] Figure 3 This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are protected by this application.

[0016] It should be noted that the terms "comprising," "including," and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application, are intended to cover non-exclusive inclusion. For example, a process, method, terminal, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. Terms such as "first" and "second" in the claims, specification, and accompanying drawings of this application, as well as relational terms, are used merely to distinguish one entity / operation / object from another entity / operation / object, and do not necessarily require or imply any such immediate relationship or order between these entities / operations / objects.

[0017] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0018] Cross-cultural advertising slogan modification refers to the process of adjusting the language, cultural elements, and emotional tone of advertising slogans in cross-regional (e.g., transnational, transcultural) scenarios. This is done to address potential issues such as cultural conflicts, emotional misjudgments, compliance risks, and reduced marketing effectiveness that may arise from language differences, cultural taboos, emotional perception biases, and differing regional compliance rules. Ultimately, the goal is to ensure that the advertising slogan conforms to the cultural customs and laws of the target region while also resonating with the emotional perceptions of the local audience, thus guaranteeing the effectiveness of the advertising campaign. Its core function is to solve the dual problems of language translation and cultural adaptation, avoiding situations where a literal translation is correct but culturally offensive.

[0019] Traditional cross-cultural advertising slogans rely on static emotional lexicons and manual extraction of cultural metaphors during the revision process. This can easily lead to a serious mismatch between the emotional tone of multilingual advertising slogans and the perception of the target audience. In addition, the reliance on manual extraction of cultural metaphors also carries the risk of missing taboo elements, which can easily cause advertising slogans to violate local cultural rules. It is evident that traditional advertising slogan revision has technical problems such as emotional misjudgment and high cultural compliance risks.

[0020] In view of this, this application provides an advertising slogan correction method based on a large language model. By first acquiring multilingual advertising slogan data and target region cultural rule data, and then using a pre-trained model to parse and generate sentiment vectors, it replaces the traditional static sentiment lexicon. This method can accurately capture the sentiment tendency in cross-cultural contexts and avoid disconnection from the audience's cognition in the target region. At the same time, it uses a contrastive learning algorithm to align text symbols and visual elements and extract cultural metaphor features, replacing manual extraction, thereby reducing the risk of missing taboo elements. Subsequently, it combines cultural and sentiment mapping libraries to calculate a cultural sensitivity index to generate compliant alternatives, and uses a conversion rate prediction model to select suitable solutions. This solves the technical problems of sentiment misjudgment and high cultural compliance risks in traditional advertising slogan correction.

[0021] To illustrate the technical solution of this application, specific embodiments are described below.

[0022] Figure 1 This illustration shows a flowchart of an advertising slogan correction method based on a large language model, provided in an embodiment of this application. This method can be applied to terminal devices. Terminal devices can be servers, service clusters, mobile phones, tablets, laptops, ultra-mobile personal computers (UMPCs), netbooks, etc.

[0023] Specifically, the above-mentioned advertising slogan correction method based on a large language model may include the following steps S101 to S105.

[0024] Step S101: Obtain multilingual advertising slogan data and cultural rule data for the target region.

[0025] Among them, terminal devices or advertising slogan correction devices based on large language models collect multilingual advertising slogan data to be corrected in batches or individually by connecting to online channels such as the brand's internal marketing content management system and the advertising material library of cross-border e-commerce platforms. This data must include the original text of the advertising slogan and the corresponding language identifier to distinguish the advertising content in different language versions. At the same time, it will also access the preset compliance database, the official policy interface of the advertising regulatory agency in the target area, or the rule dataset provided by the third-party cultural consulting service to obtain cultural rules data such as cultural taboos, advertising compliance constraints, and religious custom restrictions in the target area.

[0026] Optionally, if there are special cultural rules that are manually compiled, such as newly added holiday marketing restrictions or regional custom requirements in the target area, it is also possible to supplement and obtain this part of the data by importing local files or manually entering it through the user interface, so as to avoid deviations in subsequent corrections due to missing rules.

[0027] Step S102: Analyze the multilingual advertising data to generate sentiment vectors. At the same time, align the text symbols and visual elements in the multilingual advertising data using a contrastive learning algorithm to extract cultural metaphor features.

[0028] The process involves analyzing multilingual advertising slogan data to generate sentiment vectors, and simultaneously using a contrastive learning algorithm to align textual symbols and visual elements to extract cultural metaphor features. In the cultural metaphor feature extraction stage, a pre-defined visual element library (containing culturally relevant visual data from various countries, including taboo colors, religious symbols, and regional identifiers) is used to calculate the semantic similarity (e.g., cosine similarity) between core textual symbols in the advertising slogan (such as culturally significant terms like "halal" and "lucky bag") and elements in the visual element library. The visual elements with the strongest semantic associations are then selected. Finally, the feature vector of this visual element is concatenated and fused with the semantic feature vector of the textual symbol to generate cultural metaphor features representing cultural connotations (e.g., a 128-dimensional joint semantic vector).

[0029] Optionally, the multilingual advertising data is preprocessed to remove garbled characters, redundant spaces, and meaningless characters, and is classified and labeled according to language. The classified multilingual advertising data is then input into a pre-trained XLM-RoBERTa model, which extracts the sentiment features of the text through a cross-language semantic understanding module and outputs a 512-dimensional sentiment vector.

[0030] Specifically, in the sentiment vector generation stage, the multilingual advertising slogan data is first preprocessed to remove interfering information such as garbled characters, redundant spaces, and meaningless placeholder characters, and the advertising slogans in different languages ​​are classified according to language identifiers. Then, the classified advertising slogans are input into a pre-trained cross-language semantic understanding model. The cross-language semantic understanding model quantifies the sentiment tendency (positive, negative, neutral) and sentiment intensity in the text through the sentiment feature extraction module, and finally outputs a fixed-dimensional (e.g., 512-dimensional) sentiment vector, which can reflect the positioning of the advertising slogan in the cross-language sentiment space.

[0031] The preferred dimension for the sentiment vector is 512, which preserves the original sentiment characteristics (including positive and negative tendencies, sentiment intensity, and contextual relevance) while avoiding computational redundancy caused by excessive dimensionality. For example, in the sentiment characteristic analysis of the German word for "confidence" and the Japanese word for "confidence," a 512-dimensional vector can accurately distinguish the difference between the former's "recognition of personal ability" and the latter's "cautious expression under a hierarchical culture."

[0032] Optionally, if the advertising slogan is accompanied by visual materials such as advertising images or short video frames, the visual elements in the materials can be extracted through an image recognition model and then aligned with the text symbols to further improve the accuracy of cultural metaphor features.

[0033] Step S103: Call the preset cultural and emotional mapping library, combine the emotional vector, the cultural metaphor features and the cultural rule data to calculate the cultural sensitivity index. If the cultural sensitivity index is greater than the preset threshold, then generate multiple advertising slogan alternatives through reinforcement learning algorithm.

[0034] This involves calling a pre-defined cultural and emotional mapping library, combining emotional vectors, cultural metaphor features, and cultural rule data to calculate a cultural sensitivity index, and determining whether to generate alternative solutions based on whether the index exceeds a threshold, thereby achieving risk prediction and preliminary correction.

[0035] Specifically, when calculating the cultural sensitivity index, a weighted sum can be performed based on the matching degree between the emotional vector and cultural rule data, and the conflict degree between cultural metaphor features and the visual element library. The emotional vector matching degree is calculated by the cosine similarity between the emotional vector of the advertising slogan and the emotional acceptance threshold vector of the target region's audience. If the similarity is less than 0.5, it is considered an emotional conflict, and the corresponding weight coefficient w1 can be adjusted according to the emotional sensitivity of the target region. The cultural metaphor conflict degree is calculated by analyzing the overlap between cultural metaphor features and the taboo element vectors of the target region in the visual element library. If the overlap dimension exceeds a preset 100% threshold, it is considered a cultural conflict, and the corresponding weight coefficients w2 and w1 satisfy w1 + w2 = 1. For example, when the emotional vector matching degree of the advertising slogan for the Middle East region is 0.3 (emotional conflict exists), the cultural metaphor conflict degree is 0.8 (serious cultural conflict exists), and w1 = 0.4 and w2 = 0.6, the cultural sensitivity index = 0.3 × 0.4 + 0.8 × 0.6 = 0.6. If the preset threshold for this region is 0.6, a correction process can be triggered.

[0036] The mapping library pre-stores the relationships between sentiment vectors, cultural metaphor features, and cultural rules in different cultural scenarios, as well as sensitivity calculation logic (such as multi-dimensional weighted algorithms). Sentiment vectors, cultural metaphor features, and cultural rule data are input into the calculation module. By analyzing whether the sentiment aligns with the target region's acceptability, whether the cultural metaphors conflict, and whether the content violates rules, a quantified cultural sensitivity index (an indicator for assessing conflict and compliance risks) is obtained. Specifically, the cultural sensitivity index can be calculated by weighting and summing the matching degree between sentiment vectors and cultural rule data, and the conflict degree between cultural metaphor features and the visual element library. If the index exceeds a preset threshold (set according to the cultural sensitivity of the target region, e.g., 0.6 for highly sensitive regions), a reinforcement learning algorithm can be activated to generate multiple alternatives. The advertising text features, sentiment vectors, and cultural rule data are used as input to the reinforcement learning agent. The agent generates initial alternatives based on a pre-trained language generation model, and then filters them according to the criteria of "compliance with rules, emotional consistency, and semantic matching of the original core information" to obtain multiple alternatives that meet the basic requirements.

[0037] Optionally, if the target area has multiple sub-cultural groups, an index can be calculated for each sub-group and a suitable alternative can be generated.

[0038] Step S104: Evaluate the predicted conversion rate data of various advertising slogan alternatives in the target area using a conversion rate prediction model, and select the target alternative as the corrected target advertising slogan based on the predicted conversion rate data.

[0039] In the embodiments of this application, a conversion rate prediction model is used to evaluate the predicted conversion rate data of alternative solutions, and a target alternative solution is selected based on the data to ensure that the revised advertising copy is market-adaptable. First, each alternative solution is converted into a text feature vector (e.g., generated through a word embedding model), and combined with user characteristics of the target region (age distribution, consumption habits, historical advertising behavior, etc.) as model input. The conversion rate prediction model analyzes the degree of matching between the alternative solution and user characteristics, and outputs predicted conversion rate data (e.g., click-through rate, purchase conversion rate). Then, according to selection rules (e.g., selecting the one with the highest conversion rate, or selecting the one with the lowest sensitivity after the conversion rate reaches a target), the optimal solution is selected from the alternative solutions as the revised target advertising copy.

[0040] Optionally, if the conversion rates of multiple solutions are similar, auxiliary indicators such as semantic simplicity and propagation can be introduced for secondary screening.

[0041] Step S105: Output the target advertising slogan.

[0042] In the embodiments of this application, the target advertising slogan can be stored in the results database for use by the brand marketing system (e.g., for advertising in the target region); at the same time, it can be pushed to cross-border e-commerce advertising modules, multilingual advertising sections of brand websites, and other scenarios through interfaces.

[0043] Optionally, an evaluation report can be attached to the output, including information such as cultural sensitivity index, predicted conversion rate, and screening process.

[0044] The beneficial effects of this application embodiment compared with the prior art are as follows: This application embodiment obtains multilingual advertising slogan data and target region cultural rule data, and firstly achieves accurate capture of advertising slogan emotion and cultural semantics in cross-cultural context by parsing to generate emotion vectors and extracting cultural metaphor features through comparative learning; then, it calculates a cultural sensitivity index by combining cultural and emotion mapping libraries to generate compliant alternatives, and selects target advertising slogans that have both cultural adaptability and marketing conversion potential through a conversion rate prediction model. This effectively solves the problems of easy emotional misjudgment, high cultural compliance risk, and difficulty in taking into account marketing conversion effect in traditional methods. The corrected target advertising slogan can simultaneously meet the cultural rules of the target region, the audience's emotional cognition, and the marketing conversion needs, thus improving the effectiveness of cross-cultural advertising.

[0045] In some specific embodiments of this application, the generation of multiple advertising slogan alternatives through reinforcement learning algorithms may specifically include steps S401 to S402.

[0046] Step S401: The text features of the multilingual advertising data, the sentiment vector, and the cultural rule data are used as inputs to the reinforcement learning agent, wherein the agent generates an initial alternative based on a pre-trained XLM-RoBERTa model.

[0047] In the embodiments of this application, the reinforcement learning agent may include a state space, an action space, and a reward function. The state space is a fusion vector of multilingual advertising text feature vectors, sentiment vectors, and cultural rule data. The fusion method uses dimensional alignment and concatenation of the text feature vector (256 dimensions), sentiment vector (512 dimensions), and cultural rule binary vector (128 dimensions, 1 indicating compliance with the rule, 0 indicating violation of the rule) to form a 916-dimensional state vector. The action space consists of advertising expression adjustment actions that the agent can generate, including word replacement, sentence transformation, and cultural element supplementation, covering more than 1000 basic action combinations. The reward function uses multi-dimensional weighted calculation, specifically including a cultural rule compliance score (e.g., weight 0.4, maximum score 1 point, 0.2 points deducted for each rule violation), a sentiment consistency score (e.g., weight 0.3, calculating the similarity between the sentiment vector of the alternative and the sentiment vector of the original advertising text, 1 point for a similarity of 1), and a semantic relevance score (e.g., weight 0.3, calculating the matching degree between the alternative and the core semantics of the original advertising text, 1 point for a matching degree of 1).

[0048] The process begins by extracting text features from the multilingual advertising slogan data. Unstructured original slogans are converted into computable structured text feature vectors through word segmentation and word embedding techniques. Simultaneously, sentiment vectors (used to constrain the sentiment tendencies of alternative slogans) and cultural rule data from the target region (used to mitigate cultural taboos and compliance risks) generated in previous steps are retrieved. These three types of data are then integrated into a unified input data format according to preset feature fusion rules (e.g., concatenation after dimension alignment and superposition after weight allocation), and transmitted to the reinforcement learning agent. Subsequently, the reinforcement learning agent invokes a pre-trained XLM-RoBERTa model. Leveraging its cross-lingual semantic understanding capabilities, this model, combined with the sentiment constraints and cultural rule boundaries in the input data, generates multiple candidate texts that conform to the core semantic logic of the original advertising slogan, serving as initial advertising slogan alternatives.

[0049] Furthermore, the training of the XLM-RoBERTa model comprises three parts: first, a multilingual advertising text corpus (e.g., covering more than 30 languages, including more than 10 categories such as FMCG, home appliances, and clothing, totaling more than 500,000 texts, each text labeled with language identifiers and core semantic tags); second, manually annotated sentiment tags (e.g., annotated by 50 cross-cultural language experts, categorized by positive, neutral, and negative sentiment tendencies and sentiment intensity levels of 1-5, with annotation consistency verified by the Kappa coefficient, and a Kappa value greater than or equal to 0.85); and third, cultural rule-related texts (e.g., integrating advertising regulations from 50 countries, cultural taboo manuals, and historical conflict case analyses, totaling more than 2,000 rule texts, each rule labeled with its applicable region and violation risk level).

[0050] The training process uses the cross-entropy loss function to optimize the sentiment label prediction error, while introducing cultural rule constraint loss (e.g., increasing the loss value when the model generates text that violates cultural rules). The learning rate is set to 2e-5, the batch size is 32, the number of iterations is 100, and an early stopping mechanism is used (training stops when the accuracy on the validation set does not improve for 5 consecutive iterations). To adapt to minority language scenarios, data augmentation techniques are used (e.g., synonym replacement, sentence transformation, and context expansion for minority language texts, increasing the sample size by 3 times), and some parameters of the underlying model are frozen, training only the top semantic adaptation layer to improve accuracy in sentiment recognition of advertising texts in minority languages.

[0051] Optionally, the generation parameters of the XLM-RoBERTa model (such as the search width and maximum generation length of beamsearch) can be adjusted according to the language characteristics of the target region (such as minority languages ​​or languages ​​with complex grammar) to adapt to the expression habits of different languages; a semantic deviation threshold can also be set in advance for the agent so that the core information deviation between the initial alternative and the original advertising slogan does not exceed the preset range.

[0052] Step S402: The initial alternative is scored. If the score is lower than the preset reward threshold, the generation strategy of the agent is adjusted and a new solution is generated iteratively until multiple advertising slogan alternatives with satisfactory scores and semantic non-repetition are obtained.

[0053] The process begins by constructing a multi-dimensional scoring method. Each dimension must at least protect "cultural compliance" (determining whether the proposed solution violates cultural taboos and advertising regulations in the target region), "emotional consistency" (calculating the similarity between the proposed solution's emotional vector and the original advertising slogan's emotional vector), and "semantic relevance" (evaluating the degree of matching between the proposed solution and the core semantics of the original advertising slogan). A weighted summation algorithm is then used to calculate the comprehensive score for each initial alternative. If the comprehensive score of an initial alternative falls below a preset reward threshold, the generation strategy of the reinforcement learning agent is adjusted. This includes optimizing the agent's reward function (increasing the weight of cultural compliance and emotional consistency) and adjusting the output constraints of the XLM-RoBERTa model (e.g., adding a list of prohibited words and strengthening semantic relevance constraints). The agent is then controlled to regenerate a new advertising slogan alternative based on the adjusted strategy.

[0054] This iterative process continues until multiple solutions that meet the comprehensive score requirements are obtained. At the same time, semantic similarity algorithms (such as cosine similarity and Jaccard similarity coefficient) are used to deduplicatize the solutions that meet the requirements, and solutions with semantic repetition exceeding a preset threshold are eliminated. Finally, multiple advertising slogan alternatives that meet the requirements and have no semantic repetition are obtained.

[0055] Optionally, features of historically high-conversion ad copy in the target region can be introduced as a scoring reference to improve the suitability of the solution to the preferences of users in the target region; an upper limit on the number of iterations (e.g., 50 times) can also be set to avoid infinite loops in the iteration process due to extreme scenarios.

[0056] In the embodiments of this application, the terminal device, through reinforcement learning agent and pre-trained XLM-RoBERTa model, can ensure that the initial alternative has cross-linguistic semantic accuracy and basic emotional and cultural adaptability. Then, through multi-dimensional scoring iterative optimization and semantic deduplication, it effectively solves the problems of inconsistent quality, high semantic repetition rate, and difficulty in balancing emotional consistency and cultural compliance in the generation of traditional alternatives, and finally obtains high-quality and diversified advertising slogan alternatives.

[0057] In some specific embodiments of this application, before the step of using the text features of the multilingual advertising data, the sentiment vector, and the cultural rule data as input to the reinforcement learning agent, steps S501 to S502 may be included.

[0058] Step S501: Use multilingual advertising text corpus, manually annotated sentiment tags, and cultural rule-related text as the model training dataset.

[0059] The process involves three main steps: First, collecting multilingual advertising text corpora. This corpus can be derived from archives of historical multilingual advertising slogans from brands, publicly available advertising materials from cross-border e-commerce platforms, and industry marketing case databases. It must cover the main languages ​​relevant to the target region and include advertising slogans for different product categories (e.g., FMCG, home appliances, apparel) and marketing scenarios (e.g., holiday promotions, new product launches). Simultaneously, manually annotated sentiment tags are acquired. Professional annotators combine the semantic context of the advertising slogans and annotate them according to preset sentiment dimensions (e.g., positive, neutral, negative) and sentiment intensity levels. Cultural rule-related texts are also collected. These texts include advertising regulations, cultural taboos, religious customs, and other content related to the compliance of the advertising slogans in the target region, and are then linked to the corresponding language and cultural scenarios. These three types of data are then integrated and processed, removing duplicate text, garbled text, meaningless short sentences, and other invalid data. The dataset is then divided into training, validation, and test sets according to a preset ratio (e.g., 8:1:1) to ultimately form a structured model training dataset.

[0060] Optionally, data augmentation can be performed on the advertising text corpus in minority languages ​​by expanding the sample size through methods such as synonym replacement and sentence transformation, thereby solving the problem of scarce samples in minority languages.

[0061] Optionally, an annotation consistency verification method can be introduced, which involves cross-annotation by multiple people and calculation of the annotation consistency coefficient (e.g., Kappa coefficient) to ensure the annotation accuracy of sentiment tags.

[0062] Step S502: Input the training dataset into the initial XLM-RoBERTa model to fine-tune the initial XLM-RoBERTa model and obtain the XLM-RoBERTa model.

[0063] First, a pre-trained initial XLM-RoBERTa model is loaded. This model has basic cross-lingual semantic understanding capabilities, but it has not been optimized for advertising scenarios and cultural rules. Then, the training set data is processed according to the model input format. Multilingual advertising text and corresponding cultural rule-related text are used as input features, and manually labeled sentiment tags are used as prediction targets. Training parameters (such as learning rate, batch size, and number of iterations) are set, and the model is iteratively trained using gradient descent algorithm and cross-entropy loss function (used to optimize sentiment tag prediction error).

[0064] During training, the model performance is monitored in real time using the validation set. If the model's sentiment prediction accuracy on the validation set does not improve for several consecutive rounds, an early stopping mechanism is triggered to avoid overfitting. After training, the model's final performance is evaluated using the test set. If the test set accuracy reaches a preset standard (e.g., above 85%), the model is determined to be a finely tuned XLM-RoBERTa model. If it does not reach the standard, the training parameters are adjusted (e.g., reducing the learning rate or increasing the number of iterations) or the training dataset is supplemented and optimized (e.g., adding samples from languages ​​with high error rates), and the fine-tuning process is repeated.

[0065] Optionally, a parameter freezing strategy can be used to freeze the parameters of the underlying general semantic understanding layer of the initial model, and only train the top network layer related to the sentiment and cultural rules of the advertising slogan, thereby reducing the consumption of computing resources and speeding up the training process.

[0066] In the embodiments of this application, by using a dedicated training dataset containing multilingual advertising slogans, sentiment tags, and cultural rule-related texts, and by fine-tuning the initial XLM-RoBERTa model, the problem of insufficient accuracy in processing advertising slogan sentiment analysis and cultural compliance judgment caused by the lack of advertising slogan scene adaptability and cultural rule-related knowledge in the initial pre-trained model is effectively solved.

[0067] In some specific embodiments of this application, the step of evaluating the predicted conversion rate data of various advertising slogan alternatives in the target area through a conversion rate prediction model may specifically include steps S601 to S602.

[0068] Step S601: Convert the various advertising slogan alternatives into text feature vectors respectively, and combine them with the user features of the target region as input to the conversion rate prediction model.

[0069] In the embodiments of this application, for each advertising slogan alternative, a text feature extraction operation is first performed. The text content of the alternative is processed through a pre-trained text semantic coding model to convert unstructured text information into a fixed-dimensional structured text feature vector (e.g., 768-dimensional, 512-dimensional). This vector can quantitatively represent key information such as the semantic connotation and marketing style of the alternative. Simultaneously, user characteristic data of the target region is acquired. These characteristics include, but are not limited to, demographic characteristics (age stratification, gender ratio, and geographical distribution of users in the target region), consumer behavior characteristics (historical purchase categories, purchase frequency, and average order value), and advertising interaction characteristics (past click frequency, dwell time, and conversion rate of similar product advertisements). These user characteristics can be retrieved from the user behavior database of cross-border e-commerce platforms, brand-owned user management systems, or third-party user profiling service platforms.

[0070] The text feature vectors corresponding to each alternative are integrated with the user features of the target region according to the preset feature fusion rules. For example, vectors are concatenated after dimension alignment, weighted fusion is based on attention mechanism, or a feature cross layer is introduced to generate associated features, forming unified data that conforms to the input format of the conversion rate prediction model.

[0071] Optionally, user features can be preprocessed by using Min-Max normalization and Z-Score standardization to eliminate differences in the numerical range of different feature dimensions (such as "age" and "consumption frequency") and avoid interfering with model prediction.

[0072] Optionally, core user characteristics can be filtered based on the product category of the advertisement (e.g., FMCG, home appliances, luxury goods). For example, luxury goods advertisements focus on "the proportion of users with high average order value", while FMCG advertisements focus on "purchase frequency".

[0073] Step S602: Output the predicted conversion rate data of the advertising slogan alternative through the conversion rate prediction model. The predicted conversion rate data includes click conversion rate and purchase conversion rate.

[0074] In the embodiments of this application, a pre-trained conversion rate prediction model is loaded. This model has been trained using historical data (including text features of past multilingual advertising slogans, user features of corresponding regions, and click-through conversion rates and purchase conversion rates after actual deployment), and has the ability to associate text and user features with conversion behavior.

[0075] The integrated input data is input into the model in batches or individually. The model then uses internal prediction algorithms (such as deep learning networks, gradient boosting decision trees, logistic regression, etc.) to analyze the matching degree between the text features of alternative solutions and the characteristics of users in the target region. For example, it determines whether discount rhetoric matches the price-sensitive consumption characteristics of users in the target region, and whether quality promotion rhetoric aligns with the preferences of high-spending users, thereby outputting predicted conversion rate data. Click-through rate (CTR) is the probability that users in the target region will click on the ad after seeing it; purchase conversion rate is the probability that users will complete a purchase after clicking the ad.

[0076] The identifier of each alternative is associated with the corresponding two types of predicted conversion rate data and stored to obtain a structured evaluation result table.

[0077] The implementation method of this application effectively solves the problem that traditional evaluation methods only focus on the cultural compliance or semantic consistency of the alternative advertising slogans and ignore the impact of user preferences on conversion results by deeply combining the textual features of the alternative advertising slogans with the user features of the target area as input to the conversion rate prediction model.

[0078] In some specific embodiments of this application, the step of aligning text symbols and visual elements in the multilingual advertising data using a contrastive learning algorithm to extract cultural metaphor features may specifically include steps S701 and S703.

[0079] Step S701: Call a preset visual element library, which contains culturally relevant visual data from multiple countries, including taboo colors and cultural symbols.

[0080] The visual elements library is a pre-built and stored structured database containing culturally relevant visual data from multiple countries and regions around the world. Its core features include taboo colors (such as red, which is taboo in some parts of the Middle East, and white, which symbolizes differences in some East Asian traditions) and cultural symbols (such as religious symbols, regional totems, or festival logos).

[0081] Specifically, the data sources for the visual element library can include three categories: first, publicly available databases from cultural research institutions in various countries to obtain basic cultural symbols (such as religious totems and traditional clothing patterns); second, a database of cross-border advertising historical and cultural conflict cases (for example, compiling over 1,000 advertising cultural conflict incidents from the past 5 years and extracting the visual elements that triggered the conflicts); and third, local cultural materials from the target region (for example, extracting typical visual elements from local films and television works, traditional festival decorations, and official cultural promotional materials using image recognition models). All visual data is manually annotated (the annotation includes element name, region, cultural meaning, and taboo level) and then encoded into 256-dimensional feature vectors for storage, forming a structured library with element ID, region list, feature vector, taboo level, and update time.

[0082] In the embodiments of this application, these data can be collected and constructed through multiple channels. For example, basic visual taboo information can be obtained from public databases of cultural research institutions in various countries, visual elements in historical and cultural conflict cases of cross-border advertising can be manually sorted out, and typical symbols can be extracted from local cultural materials (such as film and television works, traditional clothing, and festival decorations) in the target region through image recognition technology. Through the library call interface, a subset of culturally relevant visual data specific to the target region corresponding to the current advertising slogan can be filtered out to avoid data from irrelevant regions interfering with subsequent calculations.

[0083] Optionally, a regular update mechanism can be set up for the visual element library, such as synchronizing the latest visual taboo information (e.g., new festival-specific visual symbols added in a certain region) from third-party cultural information platforms or official cultural institutions in the target area every month to ensure the timeliness of the data in the library.

[0084] Step S702: Calculate the semantic similarity between the core text symbols in the multilingual advertising data and each visual element in the visual element library.

[0085] In the embodiments of this application, the core symbols of the multilingual advertising text are first extracted. By combining the text keyword extraction algorithm with the cultural attribute screening rules, the core text symbols with cultural orientation are identified from the advertising text, and general words without cultural connotation are eliminated.

[0086] The extracted core text symbols are converted into semantic vectors. Specifically, a pre-trained cross-lingual semantic encoding model is used to encode the text symbols, generating fixed-dimensional text semantic vectors. Simultaneously, pre-encoded feature vectors of each visual element within the target region's visual data subset are retrieved from a visual element library. Using semantic similarity calculation algorithms (such as cosine similarity and Euclidean distance), the similarity value between the semantic vector of the core text symbol and the feature vector of each visual element is calculated one by one. This value is used to quantify the degree of cultural association between the text symbols and visual elements.

[0087] Optionally, the core text symbols can be processed for ambiguity. For example, when a symbol (such as "moon") has multiple cultural interpretations in the target area, semantic vectors of the symbol in different interpretation directions can be generated, and their similarity can be calculated with visual elements respectively.

[0088] Step S703: The feature vectors of the visual elements with the highest similarity are concatenated to generate a 128-dimensional joint semantic vector as the cultural metaphor feature.

[0089] In the embodiments of this application, the visual element feature vector with the highest value is selected from all similarity calculation results (if there are multiple vectors with the same and highest similarity, they can be filtered by the cultural priority weight of the visual elements, such as the priority of religious taboo visual elements is higher than that of ordinary regional symbols).

[0090] The semantic vectors of core text symbols and the feature vectors of selected visual elements are dimensionally adapted and concatenated. If their initial dimensions differ, a vector dimension adjustment algorithm is used to unify them to the same intermediate dimension. Then, the vectors are concatenated according to a preset concatenation order. Finally, feature normalization (e.g., L2 normalization) is used to compress the concatenated vector to 128 dimensions. This 128-dimensional joint semantic vector represents the cultural metaphor features that comprehensively reflect the cultural association between text symbols and visual elements. For example, experiments were conducted using over 1000 historical cultural conflict cases from 50 countries to test the optimal 128-dimensional cultural metaphor feature dimension. The 128-dimensional vector showed extremely high accuracy in identifying cultural conflicts, significantly improving upon the 64-dimensional vector, while significantly reducing computation time compared to the 256-dimensional vector. When identifying the association between the Arabic word "halal" and taboo colors in the visual element database, the 128-dimensional vector fully covers three core features: religious attributes, color taboos, and regional customs. In contrast, the 64-dimensional vector loses the dimension related to regional customs, leading to a higher misjudgment rate.

[0091] Optionally, an attention weighting mechanism can be introduced to assign different weights to each dimension of the concatenated vector based on the cultural association between the text symbols and the visual elements (e.g., the higher the association, the greater the weight).

[0092] The implementation method of this application calls a preset culturally relevant visual element library and combines semantic similarity calculation to achieve accurate alignment between text symbols and visual elements, generating a 128-dimensional joint semantic vector that integrates the cultural connotations of both. This effectively solves the problem that the extraction of traditional cultural metaphor features only focuses on text content and ignores the cultural association between text and visual elements, resulting in incomplete features and an inability to truly reflect the cross-cultural connotations of advertising slogans.

[0093] In some specific embodiments of this application, after the step of outputting the target advertising slogan, steps S801 and S802 may be included.

[0094] Step S801: Call a preset visual element library, which contains culturally relevant visual data from multiple countries, including taboo colors and cultural symbols.

[0095] In the implementation of this application, public opinion data is obtained by connecting to multiple data interfaces. Specifically, for user comments on the target advertising slogan, the advertising-related comment interfaces of cross-border e-commerce platforms (such as user reviews of related ads under product detail pages) and mainstream social media APIs of the target region (such as the comment section of TikTokShop in Southeast Asia and the user feedback interface of Souq in the Middle East) can be called to collect the text comments, sentiment ratings and interactive behaviors (such as likes and dislikes) posted by users, and extract key information related to the semantics and cultural expression of the advertising slogan; for media reports, news information aggregation platform interfaces (such as Google News and local news website APIs of the target region) can be used to filter reports containing the target advertising slogan and the corresponding brand or product, and extract the opinion bias (positive, negative, neutral) and cultural compliance-related expressions involved in the reports; for local advertising policy updates, the official interfaces of advertising regulatory agencies in the target region or third-party compliance service platforms can be accessed to obtain the latest policy texts such as advertising bans, expression restrictions, and industry standards.

[0096] After collection, the public opinion data is preprocessed to remove duplicate data, garbled text, and invalid information unrelated to the advertising slogan (such as comments that simply discuss product quality rather than the advertising slogan). The data format is standardized by text cleaning algorithms (such as removing special symbols and unifying character encoding), and the data is classified and labeled according to "user comments", "media reports", and "policy updates".

[0097] Optionally, a dynamic data collection frequency can be set, such as collecting data every 2 hours in the early stages of ad campaigning to quickly capture initial feedback, and then adjusting to collecting data once a day after the campaign stabilizes to balance timeliness and resource consumption.

[0098] Step S802: Dynamically update the mapping database based on the public opinion data.

[0099] In the implementation of this application, key update information in the public opinion data is analyzed. Specifically, if a large number of negative feedbacks about a certain text symbol appear in user comments (for example, the phrase "economical and affordable" in a target advertising slogan is interpreted as "cheap and inferior" in the target region), the association between the text symbol and the negative sentiment is extracted, and the strength of the association (e.g., the proportion of negative comments) is recorded. If media reports indicate that a certain type of cultural metaphor is controversial (e.g., the inappropriate use of symbols related to a specific religion), the sentiment tendency weight and risk level corresponding to that cultural symbol in the mapping library are updated. If local advertising policies add prohibited expressions (e.g., prohibiting the use of absolute terms such as "absolutely number one"), the prohibited expression and the corresponding compliance risk parameters are added to the cultural rules module of the mapping library.

[0100] The update operation is performed according to the structured format of the mapping library. Specifically, for adjustments to existing entries in the mapping library (such as the sentiment coefficient of a text symbol), a weighted fusion algorithm is used to combine historical rule data and new public opinion data to calculate the updated parameter value (e.g., the original sentiment coefficient was +0.6, adjusted to -0.4 after considering 70% negative feedback). For new entries (such as new policy bans), a new rule record is created in the mapping library, indicating the update time, data source, and scope of effect. After the update is completed, the mapping library is backed up and its integrity is verified. If a conflict occurs, a rollback mechanism is triggered to restore the previous version and re-parse the public opinion data.

[0101] Optionally, a manual review node can be set up. For updates involving high-risk cultural rules (such as religious taboos or politically sensitive expressions), they should first be pushed to professional reviewers for confirmation to avoid erroneous updates due to noise in public opinion data.

[0102] The implementation method of this application effectively solves the problem that the traditional mapping library, due to its static configuration and slow manual updates, cannot keep up with the speed of public opinion changes and policy updates in the target area by collecting advertising public opinion data in the target area in real time and dynamically updating the cultural and emotional mapping library.

[0103] Figure 3 This illustration shows a structural diagram of an advertising slogan correction device based on a large language model, according to an embodiment of this application. The aforementioned advertising slogan correction device 2 based on a large language model can be configured on a terminal device. Specifically, the aforementioned advertising slogan correction device 2 based on a large language model may include: Module 201 is used to acquire multilingual advertising slogan data and cultural rule data of the target area; The parsing and extraction module 202 is used to parse the multilingual advertising data, generate sentiment vectors, and simultaneously align the text symbols and visual elements in the multilingual advertising data through a contrastive learning algorithm to extract cultural metaphor features. The generation module 203 is used to call a preset cultural and emotional mapping library, combine the emotional vector, the cultural metaphor features and the cultural rule data to calculate the cultural sensitivity index. If the cultural sensitivity index is greater than a preset threshold, multiple advertising slogan alternatives are generated through reinforcement learning algorithm. The correction module 204 is used to evaluate the predicted conversion rate data of various advertising slogan alternatives in the target area through the conversion rate prediction model, and select the optimal solution as the corrected target advertising slogan based on the predicted conversion rate data. Output module 205 is used to output the target advertising slogan.

[0104] The beneficial effects of this application embodiment compared with the prior art are as follows: This application embodiment obtains multilingual advertising slogan data and target region cultural rule data, and firstly achieves accurate capture of advertising slogan emotion and cultural semantics in cross-cultural context by parsing to generate emotion vectors and extracting cultural metaphor features through comparative learning; then, it calculates a cultural sensitivity index by combining cultural and emotion mapping libraries to generate compliant alternatives, and selects target advertising slogans that have both cultural adaptability and marketing conversion potential through a conversion rate prediction model. This effectively solves the problems of easy emotional misjudgment, high cultural compliance risk, and difficulty in taking into account marketing conversion effect in traditional methods. The corrected target advertising slogan can simultaneously meet the cultural rules of the target region, the audience's emotional cognition, and the marketing conversion needs, thus improving the effectiveness of cross-cultural advertising.

[0105] In some embodiments of this application, the generation module 202 is further configured to: The text features of the multilingual advertising data, the sentiment vector, and the cultural rule data are used as inputs to a reinforcement learning agent, wherein the agent generates an initial alternative based on a pre-trained XLM-RoBERTa model. The initial alternative is scored. If the score is lower than the preset reward threshold, the generation strategy of the agent is adjusted and a new solution is generated iteratively until multiple advertising slogan alternatives with satisfactory scores and semantic non-repetition are obtained.

[0106] In some embodiments of this application, the above-mentioned advertising slogan correction device 2 based on a large language model further includes a training module for: The multilingual advertising text corpus, manually annotated sentiment tags, and cultural rule-related texts were used as the model training dataset. The training dataset is input into the initial XLM-RoBERTa model to fine-tune the initial XLM-RoBERTa model and obtain the XLM-RoBERTa model.

[0107] In some embodiments of this application, the parsing and extraction module 202 is further configured to: The multilingual advertising data is preprocessed to remove garbled characters, redundant spaces, and meaningless characters, and then categorized and labeled according to language. The classified multilingual advertising data is input into the pre-trained XLM-RoBERTa model. The model extracts the sentiment features of the text through the cross-language semantic understanding module and outputs the 512-dimensional sentiment vector.

[0108] In some embodiments of this application, the above-mentioned correction module 204 is further used for: The various advertising slogan alternatives are converted into text feature vectors, and combined with the user characteristics of the target region as input to the conversion rate prediction model; The conversion rate prediction model outputs the predicted conversion rate data for the advertising slogan alternative, including click conversion rate and purchase conversion rate.

[0109] In some embodiments of this application, the parsing and extraction module 202 is further configured to: Call a preset visual element library, which contains culturally relevant visual data from multiple countries, including taboo colors and cultural symbols; Calculate the semantic similarity between the core text symbols in the multilingual advertising data and each visual element in the visual element library; The feature vectors of the visual elements with the highest similarity are concatenated to generate a 128-dimensional joint semantic vector as the cultural metaphor feature.

[0110] In some embodiments of this application, the above-mentioned advertising slogan correction device 2 based on a large language model further includes a dynamic update module, used for: Obtain advertising sentiment data for the target region, including user comments on the target advertising slogan, media reports, and updates to local advertising policies; The mapping library is dynamically updated based on the public opinion data.

[0111] like Figure 3The diagram shown is a schematic of a terminal device provided in an embodiment of this application. The terminal device 3 may include: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301, such as an advertising slogan correction program based on a large language model. When the processor 301 executes the computer program 303, it implements the steps in the various advertising slogan correction embodiments based on large language models described above, for example... Figure 1 Steps S101 to S106 are shown.

[0112] A computer program can be divided into one or more modules / units. One or more modules / units are stored in memory 302 and executed by processor 301 to complete this application. One or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in a terminal device.

[0113] The terminal device may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of a terminal device and does not constitute a limitation on the terminal device. It may include more or fewer components than shown in the figure, or combine certain components, or different components. For example, a terminal device may also include input / output devices, network access devices, buses, etc.

[0114] The processor 301 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0115] The memory 302 can be an internal storage unit of the terminal device, such as the hard drive or RAM of the terminal device. The memory 302 can also be an external storage device of the terminal device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 302 can include both internal and external storage units of the terminal device. The memory 302 is used to store computer programs and other programs and data required by the terminal device. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0116] It should be noted that, for the sake of convenience and brevity, the structure of the terminal device described above can also be referred to the specific description of the structure in the method embodiment, which will not be repeated here.

[0117] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0118] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps in the above-described advertising slogan correction method based on a large language model.

[0119] This application provides a computer program product that, when run on a mobile terminal, enables the mobile terminal to implement the steps in the above-mentioned advertising slogan correction method based on a large language model.

[0120] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0121] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for various specific applications, but such implementations should not be considered beyond the scope of this application.

[0122] In the embodiments provided in this application, it should be understood that the disclosed apparatus / terminal devices and methods can be implemented in other ways. For example, the apparatus / terminal device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0123] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0125] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0126] The embodiments described above are merely illustrative of the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for revising advertising slogans based on a large language model, characterized in that, include: Acquire multilingual advertising slogans and cultural rules data for the target region; The multilingual advertising data is analyzed to generate sentiment vectors. At the same time, the text symbols and visual elements in the multilingual advertising data are aligned through a contrastive learning algorithm to extract cultural metaphor features. The system calls a preset cultural and emotional mapping library, combines the emotional vector, the cultural metaphor features, and the cultural rule data to calculate a cultural sensitivity index. If the cultural sensitivity index is greater than a preset threshold, multiple advertising slogan alternatives are generated through a reinforcement learning algorithm. The predicted conversion rate data of various ad copy alternatives in the target area are evaluated by a conversion rate prediction model, and the target alternative is selected as the corrected target ad copy based on the predicted conversion rate data. Output the target advertising slogan.

2. The advertising slogan correction method based on a large language model as described in claim 1, characterized in that, The method of generating multiple advertising slogan alternatives through reinforcement learning algorithms includes: The text features of the multilingual advertising data, the sentiment vector, and the cultural rule data are used as inputs to a reinforcement learning agent, wherein the agent generates an initial alternative based on a pre-trained XLM-RoBERTa model. The initial alternative is scored. If the score is lower than the preset reward threshold, the generation strategy of the agent is adjusted and a new solution is generated iteratively until multiple advertising slogan alternatives with satisfactory scores and semantic non-repetition are obtained.

3. The advertising slogan correction method based on a large language model as described in claim 2, characterized in that, Before the step of using the text features of the multilingual advertising data, the sentiment vector, and the cultural rule data as input to the reinforcement learning agent, the following steps are included: The multilingual advertising text corpus, manually annotated sentiment tags, and cultural rule-related texts were used as the model training dataset. The training dataset is input into the initial XLM-RoBERTa model to fine-tune the initial XLM-RoBERTa model and obtain the XLM-RoBERTa model.

4. The advertising slogan correction method based on a large language model as described in claim 1, characterized in that, The process of parsing the multilingual advertising data and generating sentiment vectors includes: The multilingual advertising data is preprocessed to remove garbled characters, redundant spaces, and meaningless characters, and then categorized and labeled according to language. The classified multilingual advertising data is input into the pre-trained XLM-RoBERTa model. The model extracts the sentiment features of the text through the cross-language semantic understanding module and outputs the 512-dimensional sentiment vector.

5. The advertising slogan correction method based on a large language model as described in claim 1, characterized in that, The evaluation of predicted conversion rate data for various advertising slogan alternatives in the target region using a conversion rate prediction model includes: The various advertising slogan alternatives are converted into text feature vectors, and combined with the user characteristics of the target region as input to the conversion rate prediction model; The conversion rate prediction model outputs the predicted conversion rate data for the advertising slogan alternative, including click conversion rate and purchase conversion rate.

6. The advertising slogan correction method based on a large language model as described in claim 1, characterized in that, The step of aligning textual symbols and visual elements in the multilingual advertising data using a contrastive learning algorithm to extract cultural metaphor features includes: Call a preset visual element library, which contains culturally relevant visual data from multiple countries, including taboo colors and cultural symbols; Calculate the semantic similarity between the core text symbols in the multilingual advertising data and each visual element in the visual element library; The feature vectors of the visual elements with the highest similarity are concatenated to generate a 128-dimensional joint semantic vector as the cultural metaphor feature.

7. The advertising slogan correction method based on a large language model as described in claim 1, characterized in that, After the step of outputting the target advertising slogan, the following steps are included: Obtain advertising sentiment data for the target region, including user comments on the target advertising slogan, media reports, and updates to local advertising policies; The mapping library is dynamically updated based on the public opinion data.

8. An advertising slogan correction device based on a large language model, characterized in that, The device includes: The acquisition module is used to acquire multilingual advertising slogan data and cultural rule data for the target region. The parsing and extraction module is used to parse the multilingual advertising data, generate sentiment vectors, and simultaneously align text symbols and visual elements in the multilingual advertising data through a contrastive learning algorithm to extract cultural metaphor features. The generation module is used to call a preset cultural and emotional mapping library, combine the emotional vector, the cultural metaphor features and the cultural rule data to calculate the cultural sensitivity index. If the cultural sensitivity index is greater than a preset threshold, multiple advertising slogan alternatives are generated through reinforcement learning algorithm. The correction module is used to evaluate the predicted conversion rate data of various advertising slogan alternatives in the target area through a conversion rate prediction model, and select the optimal solution as the corrected target advertising slogan based on the predicted conversion rate data. The output module is used to output the target advertising slogan.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the advertising slogan correction method based on any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the advertising slogan correction method based on a large language model as described in any one of claims 1 to 7.

Images