Advertisement creativity automatic generation system and method based on natural language generation

Abstract

The invention discloses an automatic advertisement creativity generation method and system based on natural language generation. According to the invention, through establishing a keyword and context construction module, a brand semantic modeling and control module, an industry trend extraction module, an advertisement generation engine, a diversity control module and a putting adaptation and post-processing module, automatic generation of advertisement originality is realized. The system firstly receives a keyword and a mood instruction input by an advertiser, and constructs an initial semantic context; and then, coding a brand sample text by utilizing a pre-training model, and generating a brand semantic center vector as a prompt embedding so as to guide the generated content to accord with brand tonality. Meanwhile, through a keyword reservation supervision mechanism, it is ensured that core keywords appear naturally and at high frequency in the process of generating the copywriting. In the generation process, the system also uses a sampling strategy through a diversity control module and uses a semantic vector to remove duplication, and finally outputs diversified and non-redundant high-quality advertisement copywriting.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence and digital marketing, and in particular to an automated method and system for generating advertising copy and creative scripts based on natural language processing (NLP) and natural language generation (NLG) technologies. Background Technology Advertising creativity plays a crucial role in marketing campaigns, directly impacting the effectiveness of advertising. Traditional advertising copywriting and creative scripts typically rely on human copywriters to manually compose content based on specific marketing objectives, target audience attributes, and brand messaging. This manual creation model has significant limitations, including long production cycles, low output efficiency, and difficulty in quickly generating diverse content versions for different audiences, platforms, and application scenarios.

[0002] While some advertising platforms on the market have integrated intelligent keyword recommendation packages or template generation functions, these tools have improved efficiency to some extent. However, they are still essentially limited by static material libraries and lack dynamic generation capabilities and semantic diversity. This means they cannot flexibly adjust and innovate based on real-time market trends, user feedback, or brand context.

[0003] In recent years, pre-trained language models (such as GPT, T5, and BART) have made breakthroughs in the field of natural language generation, demonstrating extremely high text generation capabilities in tasks such as copywriting and summarization. For example, CN10317934A discloses a method and system for personalized recommendation of advertising information based on generative AI, which uses AI for pre-training and calls the T5 model for advertising language generation. However, how to effectively integrate these powerful language models with the specific requirements of the advertising context (such as brand intent, industry trends, and platform tone) to ensure the commercial relevance and accuracy of the generated content remains a challenge for current technology, and mature and efficient solutions are lacking.

[0004] Therefore, the industry urgently needs an automated advertising creative generation method that can integrate pre-trained language models with advertising scenario context control mechanisms to achieve large-scale, diversified, and context-consistent advertising creative content production, thereby overcoming the shortcomings of existing methods. Summary of the Invention

[0005] This invention proposes an automated method for generating advertising creatives based on natural language processing. Its core lies in constructing two key inventive points to achieve precise control over the generation process and diversified output.

[0006] According to a first aspect of the embodiments of this specification, an automated advertising creative generation system based on natural language generation is provided, characterized in that it includes: The keyword and context construction module is used to receive keywords and tone instructions input by advertisers and construct the initial semantic context; The brand semantic modeling and control module is used to encode brand sample text using a pre-trained model, generate a brand semantic center vector representing the brand tone, and embed the vector as a cue into the advertising generation engine to guide the generated content to conform to the brand tone; and to implement a keyword retention supervision mechanism to ensure that the keywords appear naturally and frequently in the generated copy. This mechanism includes: an auxiliary loss function, used during the model training or inference phase to calculate a penalty term by evaluating the difference between the occurrence of keywords in the generated copy and the preset target, and weighted summing it with the loss of the main language model; and a bundle search scoring adjustment, used during the text decoding phase to prioritize increasing the score of the generation path containing the keywords. The industry trend extraction module is used to analyze the current season, industry hotspots, and popular terms and styles on advertising platforms to extract tone prompts; The advertising generation engine, based on a pre-trained language model (such as T5), is used to receive brand control vectors from the brand semantic modeling and control module, tone prompts from the industry trend extraction module (30), and core keywords from the keyword and context construction module, and to generate a draft of the advertising copy. The diversity control module is used to iterate through the initial copy output by the ad generation engine, and uses a semantic vector deduplication mechanism to remove semantically similar copy, ensuring the diversity of the final output content. The ad delivery adaptation and post-processing module is used to structure the generated copy to adapt it to the advertising platform and to perform reverse optimization of model parameters based on the ad delivery performance feedback data.

[0007] Another aspect of the present invention is to provide an automated advertising creative generation system based on natural language processing, characterized in that it includes: The keyword and context construction module is used to receive keywords and tone instructions input by advertisers and construct the initial semantic context; The brand semantic modeling and control module is used to encode brand sample text using a pre-trained model, generate a brand semantic center vector representing the brand tone, and embed the vector as a cue into the advertising generation engine to guide the generated content to conform to the brand tone. It is also used to implement a keyword retention supervision mechanism to ensure that the keywords appear naturally and frequently in the generated copy. The industry trend extraction module is used to analyze the current season, industry hotspots, and popular terms and styles on advertising platforms to extract tone prompts. The advertising generation engine, based on a pre-trained language model (such as T5), receives brand control vectors from the brand semantic modeling and control module, tone prompts from the industry trend extraction module, and core keywords from the keyword and context construction module, and generates a draft of the advertising copy. The diversity control module is used to iterate and generate multiple rounds of copy drafts output by the advertising generation engine. The iterative generation process uses Top-k sampling and core sampling strategies to explore different expressions and styles. At the same time, it uses a semantic vector deduplication mechanism to calculate the semantic similarity of all generated copy (e.g., using cosine similarity) to identify and remove semantically similar copy, so as to ensure the quality and non-redundancy of the final output content. The ad delivery adaptation and post-processing module is used to structure the generated copy to adapt it to the advertising platform and to perform reverse optimization of model parameters based on the ad delivery performance feedback data.

[0008] Preferably, in the brand semantic modeling and control module, the pre-trained model is BERT or RoBERTa.

[0009] Preferably, the brand semantic modeling and control module also includes a keyword retention supervision mechanism to ensure that the keywords appear naturally and frequently in the generated copy.

[0010] Preferably, the keyword retention supervision mechanism includes the auxiliary loss function, which is used to calculate a penalty term by evaluating the difference between the occurrence of keywords in the generated copy and the preset target during the model training or inference phase, and then weighted and summed with the loss of the main language model.

[0011] Preferably, in the diversity control module, the semantic vector deduplication mechanism is used to calculate the semantic similarity of all generated texts (e.g., using cosine similarity) to identify and remove semantically identical texts, so as to ensure the quality and non-redundancy of the final output content.

[0012] Preferably, before the final output, the diversity control module further filters and categorizes the generated text according to preset tone instructions. Another aspect of the present invention is to provide a method for automatically generating advertising creatives based on natural language processing, characterized by comprising the following steps: S1. System Input Reception and Context Construction: Receives keywords and tone instructions input by advertisers and constructs an initial semantic context; S2. Brand semantic modeling and control vector generation: The brand sample text is encoded using a pre-trained model to generate a brand semantic center vector that represents the brand tone. The vector is then embedded into the advertising generation engine as a cue to guide the generated content to conform to the brand tone. S3. Industry Trends and Platform Style Extraction: Analyze the current season, industry hotspots, and popular terms and styles of advertising platforms to extract tone prompts; S4. Initial draft generation of the advertising copy: Based on a pre-trained language model (such as T5), the engine receives the brand semantic center vector, the intonation prompts, and the core keywords, and generates an initial draft of the advertising copy. S5. Diversity Control and Semantic Deduplication: The initial draft of the advertising copy is generated through multiple iterations, and a semantic vector deduplication mechanism is used to remove highly similar copy in terms of semantics, so as to ensure the diversity of the final output content. S6. Adaptation and Feedback: The generated copy is structured to adapt to the advertising platform; and the parameters of the pre-trained language model are optimized in reverse based on the advertising effect feedback data.

[0013] Preferably, step S2 further includes implementing a keyword retention monitoring mechanism to ensure that the keywords appear naturally and frequently in the generated copy. This mechanism includes: The auxiliary loss function is used during the model training or inference phase to calculate a penalty term by evaluating the difference between the occurrence of keywords in the generated copy and the preset target, and then weighted and summed with the loss of the main language model. The bundle search scoring adjustment is used to prioritize increasing the score of generation paths containing the keywords during the text decoding stage.

[0014] Preferably, in step S5, the multi-round iterative generation process uses Top-k sampling and core sampling strategies to explore different expression methods and styles, and the semantic vector deduplication mechanism is used to calculate the semantic similarity of all generated texts, identify and remove texts that are highly similar in meaning, so as to ensure the quality and non-redundancy of the final output content.

[0015] This invention constructs a brand semantic control vector. By fine-tuning a pre-trained language model, this vector effectively guides the generated results, ensuring that they not only retain a specific brand tone but also that the frequency and placement of core keywords align with the brand's message. This overcomes the challenge that general-purpose language models may deviate from brand style and core information when generating content.

[0016] Secondly, this invention also designs an ad generation diversity control mechanism. This mechanism combines multi-round sampling techniques (such as Top-k and Nucleus sampling) with industry knowledge templates, enabling the generation of multiple versions of ad creatives with different semantics and styles under the same marketing objective. This greatly enriches the creative library, meets the personalized needs of different platforms and audiences, and solves the problem of content homogenization.

[0017] Through the synergistic effect of the aforementioned "brand semantic control vector" and "advertising generation diversity regulation mechanism", this invention can significantly improve the efficiency of advertising creative production and adapt to the needs of rapidly changing advertising scenarios. Attached Figure Description To more clearly illustrate the technical solutions in the embodiments or related technologies of this specification, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a diagram illustrating the overall architecture of the advertising creative generation system of the present invention. Figure 2 This is a flowchart illustrating the functional implementation of the brand semantic modeling and control module of the present invention. Figure 3 This is a flowchart illustrating the implementation of the diversity control and semantic deduplication functions of the present invention. Figure 4 This is a flowchart of the advertising creative generation process of the present invention; Detailed Implementation The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that the terms "comprising" and "having," and any variations thereof, in the embodiments and drawings of this specification are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, 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.

[0020] The following examples are combined Figure 1 The technical solution of this invention will be described in detail to demonstrate the system's functional modules and their collaborative mechanisms. This automated advertising innovation generation system based on natural language processing includes the following modules: Keyword and Context Construction Module 10: Responsible for receiving user input and constructing the initial semantic context.

[0021] Brand Semantic Modeling and Control Module 20: Enables semantic control and generation of brand tone and core keywords.

[0022] Industry Trend Extraction Module 30: Dynamically analyze and extract industry hotspots and platform styles.

[0023] Ad Generation Engine 40: Core text generation based on pre-trained language models.

[0024] Diversity Control Module 50: Ensures diversity and style variation in generated content.

[0025] Delivery Adaptation and Post-Processing Module 60: Responsible for the structured processing of generated content and optimization of effect feedback.

[0026] The following is a detailed description of each module: Keyword and Context Construction Module 10 This module serves as the system's entry point for receiving user intent. Its primary function is to receive key information input by advertisers, such as the product name, specific promotional goals (e.g., "customer acquisition" or "conversion"), and target audience attributes (e.g., age, gender, interests). The module intelligently extracts core keywords from this input and further utilizes the system's historical advertising corpus and user feedback word frequency data to construct an initial, rich semantic context. This context lays the foundation for subsequent ad creative generation, ensuring that the generated content is closely related to the advertiser's initial needs. Brand Semantic Modeling and Control Module 20 This module is key to achieving "semantic control generation" of advertising content. It aims to guide the language model to generate advertising content highly consistent with the brand's tone, avoiding brand style deviations that may occur when generating content using a general language model. The steps for implementing this module are as follows (e.g., Figure 4 (as shown) S21. Building a Brand Semantic Corpus: The system first collects historical advertising materials for a specified brand, including graphic ads, short video scripts, brand slogans, etc., to build a dedicated brand semantic training set. This corpus is the foundation for learning the brand's unique style and core values.

[0027] S22. Obtaining the Brand Tone Vector: Using pre-trained models such as BERT or RoBERTa, the brand sample text is encoded. These models can bidirectionally understand the text context, capturing complex relationships between words, thereby generating high-dimensional embedding vectors. The system averages these embedding vectors to obtain a "brand semantic center vector" that represents the overall style and semantic features of the brand. This vector is the mathematical representation of the brand tone in vector space.

[0028] The pre-trained model BERT (Bidirectional Encoder Representations from Transformers), developed by Google, uses the encoder part of the Transformer architecture to generate word representations by bidirectionally understanding the text context. During pre-training, BERT learns through two tasks: Masked Language Modeling (MLM): randomly masking a portion of the words in the input text and then having the model predict what those masked words are; and Next Sentence Prediction (NSP): taking two sentences as input and having the model determine whether they are consecutive in the original text.

[0029] RoBERTa (Robustly Optimized BERT Approach) is a model developed by Facebook AI that optimizes BERT. RoBERTa improves upon BERT's training process and typically outperforms it on multiple NLP tasks. Some key optimizations of RoBERTa include: removing the next-sentence prediction task: research found this task could introduce noise, so RoBERTa focuses solely on masking the language model training. Dynamic masking: Unlike BERT's static masking, RoBERTa changes the masking pattern in each training epoch, which helps the model generalize better. Larger training data and longer training time: RoBERTa uses a much larger training dataset than BERT and trains for a longer time with larger batch sizes.

[0030] In terms of applications, models such as BERT and RoBERTa can encode text, capture complex relationships between words, and generate high-dimensional embedding vectors. These vectors can serve as mathematical representations of information such as brand tone and industry trends, and can be injected into generative models to guide the generation of text that meets specific requirements, achieving "semantic-controlled generation."

[0031] S23. Control Vector Injection: During ad generation, this brand semantic center vector is incorporated as a "Prompt Embedding" into the encoder or decoder of a pre-trained generative model (such as GPT or BART). For Transformer-type models, this vector can be injected into the query vector of the input token embedding layer or attention mechanism through concatenation or weighting. This injection method allows the generative model to continuously reference the brand's semantic information when generating each word, thus ensuring brand consistency in the generated results.

[0032] S24. Introduction of a "Keyword Retention Monitoring Mechanism": To ensure that the core keywords specified by the brand (such as "low sugar," "eco-friendly packaging," and "safe ingredients") appear naturally and frequently in the generated text, this module introduces a keyword retention monitoring mechanism. This is a soft constraint method. Specific technical methods include: S241. Adding an auxiliary loss function: During model training or inference, an auxiliary loss function is added to encourage the generation of text to contain specified keywords. This loss function penalizes generated results that fail to contain keywords or whose keywords appear too infrequently, thus prompting the model to actively incorporate these key information during generation. An example of such an auxiliary loss function is the Bag-of-Keywords (BoK) loss. It calculates the loss by evaluating the difference between the occurrence of keywords in the generated text and the preset target, and then weights and sums this loss with the loss of the main language model. This ensures both text quality and effective keyword integration, achieving precise control over the generated content.

[0033] Specifically, the loss function is calculated as follows: in, The i-th keyword assigned to the brand; The sequence of advertising copy generated for the model; For indicator functions, when the keyword The value is 1 if the keyword appears in the generated text, and 0 otherwise; λ is the weighting coefficient (e.g., 0.3-0.5). This loss function is combined with the model's original cross-entropy loss in a weighted manner, thereby improving keyword hit rate while maintaining language fluency.

[0034] For example, for each keyword Check if it appears in the generated text. If it appears: then Then this item is No penalty; if it does not occur: Then this item is This will result in punishment.

[0035] This loss function is very useful in advertising copy generation, especially for ensuring that keywords specified by the user or brand must appear in the output content.

[0036] For example, if you specify the keyword as: "Moisturizing" "Oil control" Sensitive skin If any one of these keywords is missing from the model output, it will be penalized by this loss function, thus making the optimization direction during training more biased towards keyword coverage.

[0037] The advantage of using the Bag-of-Keywords (BoK) loss function is that it can be directly embedded into the training of existing text generation models such as T5, and it is suitable for "soft constraint" keyword control, which does not force keywords but encourages natural integration.

[0038] S242. Beam Search Adjustment: In natural language generation tasks, "Beam Search" is a commonly used decoding strategy. It achieves a balance between generation quality and computational efficiency by retaining multiple most likely candidate sequences (controlled by the "beam width" parameter) at each step, rather than simply selecting the most likely single word (like a greedy search). In the text decoding stage, the Beam Search algorithm is adjusted to prioritize improving the score of the generation path containing the specified keyword.

[0039] Beam Search explores the generation space by preserving a fixed number ("beam width") of optimal sequences at each step. By adjusting the scoring mechanism, it can effectively guide the model to generate sequences that contain keywords.

[0040] S25. Scalable to Multi-Brand Mixed Scenarios: This module is also scalable, supporting the handling of multi-brand mixed marketing scenarios. By linearly combining the semantic vectors of multiple brands, a composite control vector can be generated, thereby achieving style control in co-branded products or multi-brand marketing campaigns, ensuring that the generated content simultaneously conforms to the tone requirements of multiple brands.

[0041] This invention, through the aforementioned brand semantic modeling and control mechanism, can effectively guide the language model to generate advertising content that aligns with the brand's tone and contains core keywords. This "semantic control generation" capability is a key feature distinguishing it from general language model generation. For example, when generating advertising slogans for a brand that emphasizes "eco-friendly packaging," even without explicit instructions, the system can be guided by the brand's semantic vector to tend to generate copy containing concepts such as "environmentally friendly" or "sustainable." This capability ensures that the generated content not only meets the requirements on the surface but also maintains consistency with the brand's values ​​at a deeper semantic level.

[0042] To more intuitively demonstrate the impact of brand semantic control vectors on ad generation content, Table 1 below provides a comparison example before and after applying this mechanism: Table 1. Impact of Brand Semantic Control Vector on Ad Generated Content As the table above shows, by introducing brand semantic control vectors and keyword retention supervision mechanisms, the generated content can more accurately retain the brand tone and ensure the appearance of core keywords. This transforms the generated content from generalized descriptions into creative copy with clear brand recognition, greatly enhancing the commercial value and effectiveness of advertising content. This capability makes advertising creative generation no longer a simple continuation of text, but an intelligent creation deeply integrated with brand strategy.

[0043] Industry Trend Extraction Module 30 This module ensures the real-time nature and market relevance of generated content. It analyzes recent trending keywords, popular topics, and specific styles from mainstream social media platforms within the target industry (e.g., beauty, baby products, consumer electronics). Through web crawling technology, the system can capture the latest industry-related data in real time and perform in-depth analysis of this massive dataset using trend models (e.g., TextRank and TF-IDF) to extract high-frequency phrases, popular expressions, and tone templates specific to certain platforms (e.g., humor on TikTok, professionalism on Zhihu).

[0044] The industry trend extraction module 30 is used to construct content guidelines adapted to the current industry environment. Data sources include publicly available short video platforms (such as Douyin and Kuaishou), search engine trending lists, and social media comments, obtained through API crawling and keyword-based web scraping. After text processing using jieba Chinese word segmentation, high-frequency words are weighted and ranked using the TF-IDF algorithm, or the TextRank model is used to calculate the graph structure of summary sentences or topic words. Finally, several hot words and their combinations are selected as fragments of the platform's current popular discourse. These fragments are introduced into the T5 model input using soft prompts, specifically by embedding them at the beginning of the sequence through setting virtual tokens or additional input fields, to improve the industry relevance and discourse system adaptability of the generated content.

[0045] The extracted industry trend information is provided to the ad generation engine as tone prompts. This ability to dynamically inject external knowledge allows the system to continuously update its perception of market trends, ensuring that the generated ad creatives not only meet brand requirements but also keep up with current trends and can quickly respond to public opinion and hot topics. For example, when generating ads for sunscreen in the summer, the module will identify popular seasonal terms such as "refreshing on the street" and "don't get tanned," and combine them with frequently used phrases and humorous styles favored by users on online platforms to generate more attractive copy. This dynamic adaptability significantly enhances the utility and commercial value of this invention in practical applications, making it an intelligent system capable of continuous learning and evolution.

[0046] Ad generation engine 40 The ad generation engine 40 is responsible for transforming all input information into specific ad copy or creative scripts. This module is built on current advanced large-scale pre-trained language models (such as T5).

[0047] Its inputs include: brand control vectors from the brand semantic modeling and control module 20, ensuring the content aligns with the brand tone; industry templates and tone prompts from the industry trend extraction module 30, ensuring the content's real-time performance and platform adaptability; and core keyword prompts from the keyword and context construction module 10. By integrating these multi-dimensional inputs, the ad generation engine can output complete ad copy or script paragraphs, supporting text output of up to 512 characters, and possessing multilingual (e.g., Chinese / English) generation capabilities to meet the needs of global marketing.

[0048] Diversity control module 50 This diversity control module 50 aims to ensure that each generated ad creative has rich semantic diversity and style variation to meet the refined needs of different delivery scenarios and user preferences. The steps for implementing this module are as follows (e.g., Figure 3 (as shown) S51, using a Top-k + Nucleus core sampling multi-round generation mechanism: This module employs advanced text decoding strategies, including Top-k sampling and Nucleus sampling (core sampling) for multiple rounds of content generation.

[0049] Top-k sampling restricts the model to select from the K most probable candidate words when generating the next word, thus ensuring the coherence and relevance of the generated text while maintaining a certain degree of randomness. A smaller K value makes the output more deterministic, while a larger K value increases diversity but may introduce incoherence.

[0050] Nucleus sampling is a more flexible strategy that dynamically selects the smallest subset of words whose cumulative probability exceeds a threshold P for sampling. This method can flexibly adjust the size of the candidate word pool according to the probability distribution of the current context, thereby generating more natural and less repetitive text while maintaining fluency. It effectively avoids the repetitive and unnatural text problems that are easily generated by traditional decoding strategies (such as maximum likelihood decoding and Beam Search).

[0051] By combining these two sampling strategies, this invention can finely control the randomness and diversity of generated content, effectively avoiding repetition and patterning of output content.

[0052] S52. Introduce a semantic vector deduplication mechanism: To ensure true semantic diversity in the generated content, this module introduces a semantic vector deduplication mechanism. This mechanism calculates the cosine similarity between the semantic vectors of different generated texts (e.g., the embedding vectors obtained after encoding text using models such as BERT / RoBERTa).

[0053] Cosine similarity deduplication is a measure of directional similarity between two non-zero vectors, commonly used in text similarity detection. If the cosine similarity between the semantic vectors of two texts is higher than a preset threshold (e.g., 0.9), they are considered highly similar or semantically repetitive.16 By setting a reasonable threshold, this module can effectively identify and remove semantically repetitive expressions, ensuring that the multiple versions of the final output are truly semantically distinct. This two-pronged approach ensures that the "30 versions of stylized content output" are genuinely different creative content, rather than merely superficial grammatical differences.

[0054] S53. Specify the generated style: Users or the system can specify the specific style of the generated content, such as "humorous style", "inspirational style", "question style", etc., to match the specific needs of different platforms (such as the humor of Douyin, the professionalism of Zhihu) or marketing activities.

[0055] S54. Automatic Selection: This module supports triggering the generation process by combining multiple prompts with different sampling parameters, and can automatically evaluate and select the best version to further optimize creative output.

[0056] This intelligent diversity management approach not only introduces variability at the generation level through sampling strategies, but also ensures the effectiveness of variation at the semantic level through deduplication mechanisms. This provides advertisers with truly meaningful creative options, significantly improving the usability and choice of ad creatives.

[0057] To more clearly demonstrate the effectiveness of the diversity control module, Table 2 below compares the content diversity under different generation strategies: As can be seen from the table above, by combining Top-k and Nucleus sampling with semantic deduplication, this invention can generate truly diverse and stylistically varied advertising creatives, avoiding the repetition and homogenization problems of traditional generation methods.

[0058] 60 deployment adaptation and post-processing modules This module is a crucial link in achieving a closed loop from ad creative generation to actual deployment. Its main function is to structure the generated ad copy and scripts, enabling them to automatically adapt to the delivery formats of various advertising platforms. This includes automatically matching suitable cover images, generating relevant tags based on the copy content, and adding effective CTAs (Call-To-Actions) to ensure that the generated content can be directly used for commercial deployment.

[0059] Furthermore, this module integrates a crucial feedback loop mechanism. By combining A / B test results data from actual ad campaigns (such as click-through rate, engagement rate, and conversion rate), the system can perform reverse optimization of the ad generation engine parameters. This continuous feedback learning mechanism ensures that the system can constantly adjust and improve itself based on market performance, thereby maximizing ad campaign effectiveness. This allows the invention to transcend a simple content generator, becoming an end-to-end intelligent marketing assistant capable of learning from and continuously optimizing from real-world data. This "reverse optimization" capability, through iterative model improvements driven by real-time performance data, significantly enhances the long-term value and market competitiveness of the invention.

[0060] This invention will elaborate on its workflow and technical advantages through a specific application scenario.

[0061] Application Example: Creative Generation for Brand A's New Sunscreen Advertisement Scene description: Brand A plans to advertise its newly launched sunscreen on a short video platform. The target audience for this marketing campaign is young women, and the brand hopes the advertising copy will be delivered in a humorous tone to better attract its target audience.

[0062] Input parameters: Advertisers input the following key parameters into the system: Enter keywords: "lightweight", "non-greasy", "high SPF sun protection".

[0063] Control your tone: Young woman, humorous tone.

[0064] System-generated results: After processing by the system of this invention, the system successfully generated more than thirty diverse texts that met the requirements, including but not limited to the following examples: "Invisible coverage, no fear of sunburn in summer! High SPF sun protection! Feel refreshed on the street!"

[0065] "A savior for oily skin! Have you tried a sunscreen that's lighter than not wearing makeup?"

[0066] "Sisters, don't get a tan! It can even safely film vlogs in the park."

[0067] The background technical solution program flow of the above embodiments is as follows: After receiving the above input, the background technical solution program flow of the system of the present invention is as follows: Step S1: The system receives input and constructs context.

[0068] The system first receives keywords ("lightweight," "non-greasy," "high SPF") and tone control instructions (young women, humorous tone) from brand A. Then, the system calls the keyword and context construction module 10 to extract and construct initial contextual semantics based on this input information and the system's internal historical advertising corpus. For example, the system will identify the product category as "sunscreen," the target audience as "young women," and preliminarily determine the tone of the advertisement. This step serves as a bridge between user intent and system-processable data, ensuring the accuracy of subsequent processing.

[0069] Step S2: Brand semantic modeling and control vector generation.

[0070] The system then activates the brand semantic modeling module 20. This module loads brand A's past marketing copy, product introductions, brand slogans, and other materials to build a unique semantic corpus for brand A. Using the BERT model to encode this corpus, the system generates a "brand semantic center vector" that represents brand A's unique style and core values.

[0071] Taking the BERT model as an example, encoding keywords and tone control instructions such as "lightweight," "non-greasy," and "high SPF" using the BERT model can be done in the following steps: 1. Token Segmentation: First, the input keywords and tone commands are segmented into token sequences that the BERT model can understand. BERT typically uses the WordPiece tokenizer. Keywords such as "lightweight", "non-greasy", and "high-SPF sun protection" will be segmented into corresponding tokens, for example, ['light', 'thin'], ['non', 'greasy', 'oily'], ['high', 'times', 'protection', 'sun']. Tone commands such as "young woman" and "humorous tone" will also be segmented, for example, ['year', 'light', 'woman', 'sex'], ['humorous', 'silent', 'mouth', 'kiss'].

[0072] 2. Input Formatting: The BERT model requires a specific input format, typically including special markers such as [CLS] and [SEP]. These markers are used to indicate the beginning and end of sentences.

[0073] For example, you can concatenate all keywords and tone instructions into a sentence and then wrap it with [CLS] and [SEP].

[0074] [CLS]+Lightweight+[SEP]+Non-greasy+[SEP]+High SPF+[SEP]+Young women+[SEP]+Humorous tone+[SEP].

[0075] 3. Token Embedding: Converts the formatted token sequence into initial embedding vectors. Each token (such as light, thin, high, multiple, etc.) corresponds to a pre-trained embedding vector, which contains the basic semantic information of the word.

[0076] 4. Positional Encoding: Positional encoding is added to the token embedding vector to distinguish the position of each token in the sequence. This is crucial for understanding word order.

[0077] 5. Transformer Encoder Processing: The sequence of embedding vectors with positional information is input into BERT's Transformer encoder. The encoder learns the contextual information of each token throughout the sequence bidirectionally through a multi-layer self-attention mechanism.

[0078] For example, when processing the word "light and thin", the model will not only consider "light", but also "thin" and subsequent words such as "non-greasy" and "high SPF sun protection", thus generating a more context-dependent vector representation.

[0079] 6. Generating Control Vectors: After processing by the encoder, BERT generates a high-dimensional embedding vector for each token. Typically, the final embedding vector corresponding to [CLS] is considered a comprehensive representation of the entire input sequence.

[0080] This [CLS] vector can be used as a control vector, which combines keywords such as "lightweight", "non-greasy" and "high SPF" with tone information such as "young woman" and "humorous tone".

[0081] This control vector represents the direction and style of ad creative generation, and can then be injected as "PromptEmbedding" into generation models such as T5 to guide them in generating ad copy that meets the requirements.

[0082] The aforementioned vector, used as a prompt embedding, is injected into the T5 model used by the ad generation engine 40 to guide the model in generating content that aligns with the brand's tone. Simultaneously, a keyword retention supervision mechanism is activated, using an auxiliary loss function and Beam Search scoring adjustments to ensure that core keywords such as "high SPF" and "lightweight" appear naturally and frequently in the generated copy. This process concretizes the abstract concept of "brand tone" into an operable numerical vector and proactively guides the language model to output text that meets the brand's requirements.

[0083] When the T5 model begins decoding and generating ad copy, it first receives keywords ("lightweight," "non-greasy," "high SPF") and tone instructions ("young woman," "humorous tone") from the advertiser. Beam Search selects the k highest-scoring candidate sequences from the initial token probability distribution provided by the model. In each subsequent generation step, the system evaluates each candidate sequence and adjusts its score based on multiple factors. Each candidate sequence has a base score calculated by the T5 model itself, typically the sum of the logarithms of its token sequence probabilities. To ensure the appearance of core keywords such as "high SPF" and "lightweight," the system awards bonus points to candidate sequences containing these keywords. For example, if a candidate sequence subsequently generates the token "sunscreen," the system boosts its score, making it more likely to be selected in the next iteration. To ensure the style of "young woman, humorous tone," the system also awards bonus points to words or phrases that fit these tones. For example, when candidate sequences contain words like "oily skin savior" or "sisters," their scores are boosted. After each scoring adjustment, Beam Search reorders all candidate sequences based on the final comprehensive score (base score + various bonuses) and selects the k highest-scoring sequences for the next iteration. Through this dynamic adjustment, even sequences containing keywords or matching the tone have slightly lower probabilities in the model's original predictions, but they can still achieve higher total scores through scoring adjustments and are thus prioritized by Beam Search. Finally, the process continues until a preset sentence terminator is generated or the maximum generation length is reached. Ultimately, the sequence with the highest score is selected as the final advertising copy. This allows the system to generate copy like "Oily skin savior! Have you tried a sunscreen lighter than no makeup?" and "Sisters, don't get tanned! You can even shoot a vlog in the park with peace of mind!"—copy that contains keywords and meets specific tone requirements.

[0084] The T5 model is a pre-trained model with an Encoder-Decoder architecture. In this example, all control information (keywords, tone instructions, brand tone vectors) needs to be integrated into the model's input. Input keywords and instructions: First, the keywords ("lightweight", "non-greasy", "high SPF sun protection") and tone instructions ("young woman", "humorous tone") provided by the advertiser are concatenated into a unified input sequence, which is used as the input to the encoder of the T5 model.

[0085] Brand semantic control vector: The "brand semantic center vector" you mentioned earlier is a high-dimensional vector representation that represents the overall tone of the brand. This vector can be injected into the T5 model as a "Prompt Embedding," and there are several specific ways to do so: Encoder Injection: This vector is concatenated or weighted with the Token Embedding of the encoder input, so that the encoder takes the brand tone into account when processing the input.

[0086] Decoder injection: Use this vector as the initial hidden state of the decoder, or fuse it with the current hidden state at each decoding step to continuously guide the generation process.

[0087] Multi-task learning framework: To better control the generation, the T5 model can be designed as a multi-task learning framework, with one task being to generate advertising copy and another task being to ensure that the copy contains specific keywords, so that the model learns how to take multiple objectives into account during the training phase.

[0088] 2. Introduce a "keyword retention monitoring mechanism" During the training and inference process of the T5 model, the following two mechanisms are used to ensure the natural and high-frequency occurrence of core keywords: a) Auxiliary Loss Function Objective: To add an auxiliary loss function to the traditional language model loss function (such as cross-entropy loss) during the training of the T5 model.

[0089] Working principle: After each step of the generation process, check whether the generated text contains the specified keywords ("high SPF sun protection", "lightweight").

[0090] If the generated text does not contain these keywords, or the frequency of the keywords does not meet the requirements, the auxiliary loss function will generate a penalty term.

[0091] This penalty term will propagate back through the main loss function, prompting the model to adjust its parameters so that it is more inclined to use these keywords in subsequent generation.

[0092] Advantages: This method solves the problem from the source of model training, enabling the model to "internalize" the requirements of keywords, and the generated copy will incorporate keywords more naturally.

[0093] b. Adjustments based on Beam Search scoring Beam Search is a decoding strategy that balances generation quality and computational efficiency by retaining multiple most likely candidate sequences at each step. In the example above, the specific steps for score adjustment combining the auxiliary loss function and Beam Search are as follows: Initialization: Starting with the T5 model decoder, the first token is generated. Beam Search will retain the k candidate sequences with the highest scores (k is the beam width) based on the probabilities given by the model.

[0094] Iterative generation: In each subsequent step, for each candidate sequence, the model generates the next possible word and produces a new candidate sequence.

[0095] Score Adjustment: In this process, the score of each candidate sequence can be adjusted by combining the auxiliary loss function you provide.

[0096] Specifically, the set of keywords is {"lightweight", "non-greasy", "high SPF"}. Therefore, M=3, ω1 = "lightweight", ω2 = "non-greasy", ω3 = "high SPF".

[0097] During the T5 model training or inference phase, the system will check the generated text. Does Y contain these three keywords?

[0098] If the copy Y contains the keyword "lightweight", then the value of the indicator function II ("lightweight" ∈ Y) is 1, and the corresponding penalty term is 1−1=0.

[0099] If the copy Y lacks the keyword "not greasy", then the value of the indicator function II ("not greasy" ∈ Y) is 0, and the corresponding penalty term is 1−0=1, which will produce a penalty.

[0100] This loss function is weighted and summed with the original cross-entropy loss of the T5 model, thereby prompting the model to actively incorporate all specified keywords during training, otherwise it will be penalized.

[0101] Beam Search scoring adjustments This mechanism provides real-time control over the generated text during the model decoding stage, prioritizing the selection of generation paths that contain specified keywords.

[0102] Basis for score adjustment: Keywords: The advertiser's keywords are "lightweight", "non-greasy", and "high SPF sun protection".

[0103] Tone control: The advertiser inputs tone commands such as "young woman" and "humorous tone".

[0104] Adjustment process: The T5 model generates multiple possible candidate token sequences at each decoding step.

[0105] Beam Search calculates a base score for each sequence (usually the logarithm of its probability).

[0106] The system was adjusted based on this: Keyword Bonus: If the next term in a candidate sequence is a keyword (e.g., "sunscreen"), or can form a keyword (e.g., "high SPF" followed by "sunscreen"), then the sequence will receive an additional weighted boost in score.

[0107] Bonus points for tone: Additionally, a candidate sequence will receive extra points if its vocabulary choices or sentence structure match the style of "young women" and "humorous tone." For example, the score will be boosted if words like "sisters" or "oily skin savior" appear in the candidate sequence.

[0108] Through this adjustment, even if some sequences containing keywords or specific styles have a slightly lower probability in the model's original predictions, they can still get a higher total score and thus be prioritized by Beam Search, ultimately generating copy like "Oily skin savior! Have you tried a sunscreen that's lighter than no makeup?"

[0109] Step S3: Extracting industry trends and platform styles.

[0110] The industry trend extraction module 30 begins its work, analyzing current seasonal (e.g., summer) and popular terms in the beauty industry (e.g., "refreshing for the streets," "can't get tanned"), as well as frequently used terms, popular memes, and the unique humorous tone of short videos on a particular short video platform (e.g., Douyin). This dynamic industry trend and platform style information is extracted and used as tone prompts, provided to the ad generation engine 40 to ensure that the generated content is both in line with industry trends and fits the characteristics of the short video platform. This step demonstrates the system's dynamic adaptability to real-time external information, guaranteeing the market relevance and attractiveness of the generated content.

[0111] Step S4: Initial draft generated by the ad generation engine.

[0112] The ad generation engine 40 receives the brand control vector from step two, the industry trend and platform tone prompt from step three, and the keyword prompt from step one. Based on these combined inputs, the T5 model begins to generate multiple initial drafts of ad copy. These drafts have already begun to integrate the tone of brand A, current trends in the beauty industry, and humorous elements from short video platforms. At this stage, the core generation capabilities of the large language model are demonstrated, but its output is subject to multiple constraints.

[0113] The T5 model is a pre-trained text generation model based on the Transformer architecture, undertaking the core task of text creation in the advertising generation engine. Essentially, it's a text-to-text transfer model (Text-to-Text Transformer) capable of generating natural, fluent, and content-appropriate copy based on multimodal or multidimensional input conditions. In this system, the T5 model's generation relies not only on a single prompt but also integrates constrained inputs from multiple sources, including brand control vectors (reflecting brand tone and word preferences), industry trend corpora (reflecting current content popularity and changes in consumer psychology), and platform tone cues (such as the style and rhythm of popular phrases on short video platforms). This multi-source input is internally mapped to a unified contextual representation through embedding and fusion mechanisms, thereby guiding the T5 model to better align with brand tone and communication context in the output stage, achieving customized advertising creation. The T5 model (Text-to-Text Transfer Transformer) is a pre-trained generative model based on the Transformer architecture. Its core design lies in unifying all natural language processing tasks into a "text-to-text" format, that is, encoding the input information as a text prompt and outputting the target text. During the pre-training stage, the model learns language structure and semantic representation through a large-scale general corpus. In the specific application stage, it can be guided to generate text using task-specific prompts and control vectors.

[0114] In this embodiment, the T5 model used by the ad generation engine 40 can be deployed through publicly available pre-trained weights, such as the T5-base or T5-large model based on the HuggingFace open-source platform. This model is built in Python and integrated into the Transformers library, loading the model structure and weights through a standard interface. In its implementation, the system embeds the brand control vector and platform tone prompts, then concatenates them with the original text sequence constructed from the keyword prompt to form a composite input sequence, which is then input into the Encoder part of the T5 model for contextual encoding. During the Decoder decoding process, targeted control of the generated copy content is achieved through the setting of guidance tokens and the repeated injection mechanism of control signals. This system also supports further fine-tuning of the T5 model, with training corpora including the brand's own marketing materials, industry sample data, and user comments to improve the model's adaptability and generation quality. Step S5: Diversity control and semantic deduplication.

[0115] The diversity control module 50 iterates through multiple rounds of copy generation from the initial draft output by the ad generation engine 40. It employs Top-k and Nucleus sampling strategies to explore different expressions and styles while ensuring semantic coherence, thus generating more diverse versions. Simultaneously, the semantic vector deduplication module calculates semantic similarity (e.g., using cosine similarity) for all generated copy, identifying and removing highly similar copy to ensure that the final output of over thirty pieces of copy has substantial differences in content. The system further filters and categorizes the generated results based on a preset "humorous tone" style instruction. This step is crucial for achieving meaningful diversity, ensuring the quality and non-redundancy of the creative output. Specifically, Top-k sampling involves restricting the model to select the next lexical unit from the K most probable candidate words. A smaller K value results in more conservative generated content, while a larger K value increases randomness and diversity. For styles like "humorous tone," which require more creativity and uncertainty, the system might use a larger K value.

[0116] Nucleus sampling (Top-p sampling): This method dynamically selects the smallest subset of words whose cumulative probability exceeds a threshold P for sampling. This strategy allows for more flexible control over the diversity of generated words and avoids the repetitive and unnatural problems that easily occur in traditional methods.

[0117] By combining these two strategies, the system can perform multiple rounds of generation, exploring different expressions and styles, for example: Example 1: It tends to highlight the direct effects of "lightweight" and "high SPF sun protection", generating messages such as "Invisible with a single swipe, no fear of sunburn in summer! High SPF sun protection! Feel refreshed on the street!"

[0118] Example 2: Preferring to use a "humorous tone" combined with the characteristic of "non-greasy", generating "Oily skin savior! Have you tried a sunscreen that's lighter than no makeup?".

[0119] Example 3: Combining the context of "young women" and using humorous expressions, generate "Sisters, don't get tanned! It's even safe to film a vlog in the park."

[0120] After generating a large number of initial drafts of copy in multiple rounds, the system will activate a semantic vector deduplication mechanism to ensure that the final output copy is truly different in semantics.

[0121] Encoding: The system uses models such as BERT or RoBERTa to encode all generated text, converting it into high-dimensional semantic vectors.

[0122] Calculate similarity: Then, the semantic similarity between texts is measured by calculating the cosine similarity between these vectors.

[0123] Deduplication: If the cosine similarity between the semantic vectors of two texts is higher than a preset threshold (e.g., 0.9), the system will consider them to be highly similar or repetitive in semantics and retain only one of them.

[0124] For example, if the system generates the following two texts: "No need to fear the sun in summer! This high-SPF sunscreen is refreshing and non-greasy!" "This refreshing, non-greasy, high-SPF sunscreen will keep you safe from the sun this summer!" Step S6: Adaptation and Feedback on Results.

[0125] Although these two pieces of copy differ slightly in grammar, their semantic vectors have a very high cosine similarity. The semantic deduplication mechanism will identify them as highly similar and remove one of them to ensure that the final output copy has substantial differences in content.

[0126] Finally, the system further filters and categorizes the generated results based on preset style instructions such as "humorous tone," providing advertisers with high-quality copy that matches specific styles. This method ensures that the final output creative content is diverse, non-redundant, and precisely meets the advertiser's marketing needs.

[0127] The ad delivery adaptation module 60 structures the selected high-quality copy, automatically matching sunscreen-related tags (such as "#HighSPFSunscreen", "#RefreshingOuting") and action prompts (such as "Buy Now"), and converting them into a format directly usable on short video platforms. After the copy is delivered to the short video platform, the system continuously collects and analyzes A / B testing results data (e.g., click-through rate and engagement rate of the copy). This feedback data is used to reverse-optimize the parameters of the ad generation engine, forming a closed-loop learning system that continuously improves the quality of future ad creatives and their actual delivery effectiveness. This final step emphasizes the practicality and continuous improvement capabilities of the invention, making it a complete intelligent marketing solution from generation to optimization and learning.

[0128] The automated advertising creative generation method proposed in the above examples of the present invention demonstrates significant technical effects in practical applications, specifically in the following aspects: Compared to traditional manual creation methods, this invention can increase the efficiency of generating advertising copy and creative scripts by more than 5 times, greatly shortening the content production cycle and enabling advertisers to respond to market changes more quickly.

[0129] The system supports generating up to 30 versions of stylized content output, covering the styles of various platforms, such as the lively tone of short videos, the professional tone of Zhihu, and the product copywriting style of Tmall, meeting the refined needs of different channels and target audiences.

[0130] A / B testing has verified that the ad creatives generated using this invention have increased the click-through rate (CTR) and engagement rate by an average of 17% to 28%, directly demonstrating the high quality and attractiveness of the generated content.

[0131] The system supports content generation and modification within minutes, enabling advertising creatives to quickly adapt to public opinion hotspots and breaking events, ensuring the real-time nature and timeliness of marketing content.

[0132] This invention provides a method and system for automatically generating advertising creatives based on natural language processing, successfully overcoming the limitations of traditional template-based advertising creation. By deeply integrating pre-trained language models with advertising context modeling technology, this invention demonstrates significant technical advantages such as real-time generation, diversity control, and brand consistency.

[0133] Furthermore, this method and system fundamentally change the way advertising creatives are produced, shifting from a static, rule-driven system to a dynamic, AI-driven, and context-aware generation model. This paradigm shift frees advertising creative production from limitations imposed by human efficiency and creative bottlenecks, enabling large-scale, high-quality, and personalized rapid output. This invention can be widely applied in various online marketing scenarios, significantly improving the efficiency of advertising creative production and the effectiveness of advertising campaigns. Its demonstrated efficiency improvements, content diversity, and optimized campaign performance in practical applications prove its enormous potential and commercial value in the field of digital marketing.

[0134] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0135] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and 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 the present invention.

Claims

1. An automated advertising creative generation system based on natural language processing, characterized in that, include: The keyword and context construction module (10) is used to receive keywords and tone instructions input by advertisers and construct the initial semantic context; The brand semantic modeling and control module (20) is used to encode the brand sample text using a pre-trained model, generate a brand semantic center vector representing the brand tone, and embed the vector as a prompt into the advertising generation engine to guide the generated content to conform to the brand tone. And a keyword retention supervision mechanism is implemented to ensure that the keywords appear naturally and frequently in the generated text. This mechanism includes: an auxiliary loss function, which calculates a penalty term by evaluating the difference between the occurrence of keywords in the generated text and a preset target during the model training or inference phase, and performs a weighted summation with the loss of the main language model; and a bundle search scoring adjustment, which prioritizes increasing the score of the generation path containing the keywords during the text decoding phase. The industry trend extraction module (30) is used to analyze the current season, industry hotspots, and popular terms and styles of the advertising platform, and extract tone prompts. The advertising generation engine (40), based on a pre-trained language model (such as T5), is used to receive brand control vectors from the brand semantic modeling and control module (20), tone prompts from the industry trend extraction module (30), and core keywords from the keyword and context construction module (10), and generate a draft of the advertising copy. The diversity control module (50) is used to generate multiple rounds of copy drafts output by the advertising generation engine (40), and to remove highly similar copy in semantics by using the semantic vector deduplication mechanism to ensure the diversity of the final output content. The delivery adaptation and post-processing module (60) is used to structure the generated copy to adapt to the advertising platform and to perform reverse optimization of the model parameters based on the delivery effect feedback data.

2. An automated advertising creative generation system based on natural language processing, characterized in that, include: The keyword and context construction module (10) is used to receive keywords and tone instructions input by advertisers and construct the initial semantic context; The brand semantic modeling and control module (20) is used to encode the brand sample text using a pre-trained model, generate a brand semantic center vector representing the brand tone, and embed the vector as a prompt into the advertising generation engine to guide the generated content to conform to the brand tone, and to implement a keyword retention supervision mechanism to ensure that the keywords appear naturally and frequently in the generated copy. The industry trend extraction module (30) is used to analyze the current season, industry hotspots, and popular terms and styles of the advertising platform, and extract tone prompts. The advertising generation engine (40), based on a pre-trained language model (such as T5), is used to receive brand control vectors from the brand semantic modeling and control module (20), tone prompts from the industry trend extraction module (30), and core keywords from the keyword and context construction module (10), and generate a draft of the advertising copy. The diversity control module (50) is used to perform multiple rounds of iterative generation of the initial copy output by the advertising generation engine (40). The iterative generation process uses Top-k sampling and core sampling strategies to explore different expression methods and styles. At the same time, it uses a semantic vector deduplication mechanism to calculate the semantic similarity of all generated copy (e.g., using cosine similarity) to identify and remove semantically similar copy, so as to ensure the quality and non-redundancy of the final output content. The delivery adaptation and post-processing module (60) is used to structure the generated copy to adapt to the advertising platform and to perform reverse optimization of the model parameters based on the delivery effect feedback data.

3. The system according to claim 1 or 2, characterized in that, In the brand semantic modeling and control module (20), the pre-trained model is BERT or RoBERTa.

4. The system according to claim 1 or 2, characterized in that, The brand semantic modeling and control module (20) also includes a keyword retention supervision mechanism to ensure that the keywords appear naturally and frequently in the generated copy.

5. The system according to claim 4, characterized in that, The keyword retention supervision mechanism includes the auxiliary loss function, which is used to calculate a penalty term during the model training or inference phase by evaluating the difference between the occurrence of keywords in the generated copy and the preset target, and then weighted and summed with the loss of the main language model.

6. The system according to claim 1 or 5, characterized in that, In the diversity control module (50), the semantic vector deduplication mechanism is used to perform semantic similarity calculation on all generated texts (e.g., using cosine similarity) to identify and remove semantically identical texts, so as to ensure the quality and non-redundancy of the final output content.

7. The system according to claim 1, characterized in that, Before the final output, the diversity control module (50) also filters and classifies the generated text according to preset tone instructions.

8. A method for automatically generating advertising creatives based on natural language processing, characterized in that, Includes the following steps: S1. System Input Reception and Context Construction: Receives keywords and tone instructions input by advertisers and constructs an initial semantic context; S2. Brand semantic modeling and control vector generation: The brand sample text is encoded using a pre-trained model to generate a brand semantic center vector that represents the brand tone. The vector is then embedded into the advertising generation engine as a cue to guide the generated content to conform to the brand tone. S3. Industry Trends and Platform Style Extraction: Analyze the current season, industry hotspots, and popular terms and styles of advertising platforms to extract tone prompts; S4. Initial draft generation of the advertising copy: Based on a pre-trained language model (such as T5), the engine receives the brand semantic center vector, the intonation prompts, and the core keywords, and generates an initial draft of the advertising copy. S5. Diversity Control and Semantic Deduplication: The initial draft of the advertising copy is generated through multiple iterations, and a semantic vector deduplication mechanism is used to remove highly similar copy in terms of semantics, so as to ensure the diversity of the final output content. S6. Adaptation and Feedback: The generated copy is structured to adapt to the advertising platform; and the parameters of the pre-trained language model are optimized in reverse based on the advertising effect feedback data.

9. The method according to claim 8, characterized in that, Step S2 further includes implementing a keyword retention monitoring mechanism to ensure that the keywords appear naturally and frequently in the generated copy. This mechanism includes: The auxiliary loss function is used during the model training or inference phase to calculate a penalty term by evaluating the difference between the occurrence of keywords in the generated copy and the preset target, and then weighted and summed with the loss of the main language model. The bundle search scoring adjustment is used to prioritize increasing the score of generation paths containing the keywords during the text decoding stage.

10. The method according to claim 8, characterized in that, In step S5, the multi-round iterative generation process uses Top-k sampling and core sampling strategies to explore different expression methods and styles. The semantic vector deduplication mechanism is used to calculate the semantic similarity of all generated texts, identify and remove texts that are highly similar in meaning, so as to ensure the quality and non-redundancy of the final output content.

Images