Intelligent advertisement promotion method based on large language model

Abstract

The invention discloses a large language model advertisement generation method based on retrieval enhancement generation model and conversion feedback optimization. The method comprises the following steps: S1, collecting advertisement input data; s2, generating a user behavior vector; s3, constructing a retrieval query vector and inputting the retrieval query vector into a retrieval module; s4, obtaining document content matched with the query vector from an advertisement knowledge base; s5, fusing the document content and the multi-source context to generate an advertisement context embedding vector; s6, inputting the context embedding vector into an advertisement generation model to generate a candidate advertisement text; s7, performing conversion rate prediction on the candidate advertisement texts; and S8, constructing a preference sample pair based on user feedback, and optimizing the advertisement generation model by adopting a direct preference optimization algorithm. According to the method, context modeling, semantic information enhancement and behavior-driven generation strategy optimization in the advertisement content generation process are realized, the context correlation and the putting conversion effect of the advertisement generation result are improved, and the method is suitable for a multi-platform intelligent advertisement generation scene.

Description

Technical Field

[0001] The present invention relates to the field of intelligent advertising technology, and in particular to an intelligent advertising promotion method based on a large language model. Background Art

[0002] With the rapid development of natural language processing and generative AI technologies, large language models have demonstrated powerful capabilities in text generation tasks and are widely used in a variety of fields, including content creation, question-answering systems, intelligent customer service, and advertising copy generation. In advertising technology scenarios, businesses and platforms are increasingly demanding efficient, accurate, and controllable ad content generation, prompting a gradual transition in ad generation technology from traditional template filling and static scripting to intelligent generation based on language models.

[0003] Existing ad generation methods primarily fall into two categories: rule- or template-based systems that construct ad copy using fixed fields and predefined statements; and large language model-based text generation systems that typically rely on prompts to drive copy generation. These models typically take product descriptions, keywords, or phrases as input and generate the ad copy in a single pass. While these methods have achieved some progress in automation, they still face significant deficiencies in contextual adaptability, user behavior alignment, and platform context adaptation.

[0004] Current ad generation systems generally lack a unified encoding mechanism for structured input data. Product information, user behavior sequences, platform environment characteristics, and ad placement targets are often processed separately or directly concatenated, lacking a unified modeling structure. This input processing approach makes it difficult to establish stable semantic representations, making it difficult for language models to accurately understand the semantic focus of advertising context, and resulting in a lack of personalization and structure in the generated content.

[0005] While introducing contextual search enhancement mechanisms (i.e., RAG structures) before ad content generation can improve content relevance to a certain extent, most existing systems only support static prompt text retrieval or keyword-level queries, failing to construct query vectors driven by the multidimensional features of the ad context. The query construction process typically fails to incorporate user behavior representations and platform environmental factors, nor does it design structured search expressions tailored to the advertising task. This results in a poor match between the search content and the advertising objective, which in turn affects the quality of the generated content.

[0006] When it comes to processing user feedback, many current ad generation systems still employ a static generation and fixed delivery strategy. Once generated, the model doesn't adjust its generation path based on actual user behavior. While some systems support fine-tuning training, most rely on manual scoring, click probability prediction, or predefined conversion probability estimation models for supervision, lacking a true optimization process built on a closed-loop feedback loop. Existing methods are unable to extract preference signals from user behaviors such as clicks, stays, and conversions, nor do they translate this behavioral data into model training signals for optimization.

[0007] Furthermore, most existing generative model training methods rely solely on a single language loss objective (such as cross-entropy), ignoring the multiple requirements that ad copy faces in real-world delivery scenarios, such as content quality, contextual relevance, user behavior guidance, and conversion rate optimization. This training strategy can easily result in models generating fluent content but weak conversion performance, failing to meet the business demands of actual advertising campaigns.

[0008] In recent years, some advanced methods have attempted to introduce reinforcement learning or human feedback training mechanisms (such as RLHF), such as using algorithms such as PPO to optimize model generation strategies. However, these methods generally suffer from high training costs and the need to build reward models or manual scoring systems. In advertising scenarios, users' actual click or conversion behaviors can be considered natural preference signals, making them more suitable for preference comparison learning paths that do not require environment modeling, such as direct preference optimization (DPO). However, DPO has not yet been systematically integrated and applied in ad generation tasks, especially in the context of combining RAG structures with multidimensional context modeling, and its potential has not been fully explored.

[0009] In summary, current ad generation technology still faces technical gaps in areas such as input data structure modeling, contextual retrieval control strategies, feedback closed-loop optimization paths, and multi-objective modeling. The industry urgently needs a systematic ad generation approach that integrates structured ad context, multi-source retrieval control, user behavior preference responses, and updateable generation strategies to meet the personalized, dynamic, and conversion-oriented demands of modern advertising.

[0010] Therefore, how to provide an intelligent advertising promotion method based on a large language model is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0011] One purpose of the present invention is to propose an intelligent advertising promotion method based on a large language model, aiming to improve the contextual relevance, personality adaptability and conversion responsiveness of advertising content. The method generates a retrieval query vector by constructing an input structure containing product information, user behavior, platform context and advertising goals, obtains matching documents through a semantic retrieval module, and fuses it with multi-source context to form an advertising context embedding vector, which is input into a large language model to generate candidate advertising text. Further, preference sample pairs are constructed based on user behavior feedback, and a direct preference optimization algorithm is used to train the generation strategy. This method realizes dynamic retrieval enhancement, semantic structure control and closed-loop optimization based on conversion feedback of advertising content, thereby improving advertising generation effect and conversion efficiency.

[0012] According to an embodiment of the present invention, a method for intelligent advertising promotion based on a large language model includes the following steps:

[0013] S1. Collect advertising input data;

[0014] S2. Generate a user behavior vector based on the advertisement input data;

[0015] S3. Concatenate the user behavior vector and the advertisement input data to construct a search query vector, and input the vector into a search module based on a vector matching structure. The search module calculates semantic similarity and outputs multiple related document contents.

[0016] S4. Obtaining document content matching the search query vector from an advertising knowledge base through the search module;

[0017] S5. Concatenate the document content with the user behavior vector, the advertisement input data, and the platform context data to generate an advertisement context embedding vector;

[0018] S6. Inputting the advertisement context embedding vector into an advertisement generation model to generate candidate advertisement texts;

[0019] S7, inputting the candidate advertisement text into a conversion rate prediction module constructed based on the feature interaction structure to obtain a conversion probability score corresponding to the candidate advertisement text;

[0020] S8. Deliver the candidate advertisement text to a target platform, record user feedback data associated with the candidate advertisement text, and optimize and train the advertisement generation model based on the user feedback data.

[0021] Optionally, the S1 specifically includes:

[0022] S11. Collect product information, user behavior data, platform context data, and advertising target parameters;

[0023] S12: Encode the product information, user behavior data, platform context data, and advertising target parameters into product representation vectors x p , user behavior representation vector x u , platform context representation vector x c and the advertising target representation vector x t ;

[0024] S13. Construct an advertisement input vector X. The input vector is represented by the following formula:

[0025] X=Concat(x p ,x u ,x c ,x t )+α·TimeEmbed(x c )+β·WeightNorm(x t );

[0026] Where Concat(·) represents the vector concatenation function, TimeEmbed(·) represents the temporal position embedding function of the platform context vector, WeightNorm(·) represents the weighted normalization function of the advertising target vector, and α and β are scale adjustment parameters.

[0027] S14: Submit the advertisement input vector X to the user behavior vector generation module and the search query vector construction module as advertisement input data.

[0028] Optionally, the S2 specifically includes:

[0029] S21. Receive the user behavior representation part x in the advertisement input vector X. u ;

[0030] S22. Mark and map the behavior types in the behavior record, encoding each behavior type into a fixed-dimensional behavior type vector;

[0031] S23. Calculate a behavior time factor based on the timestamp information of each behavior record, where the behavior time factor is used to represent the distance between the time when the behavior occurred and the current time;

[0032] S24, combining the behavior type vector and its corresponding time factor to generate a behavior segment representation with a time-weighted scale;

[0033] S25. Inputting all behavior segment representations into a behavior aggregation module, wherein the behavior aggregation module includes one or more layers of sequential modeling networks for constructing a behavior temporal context structure;

[0034] S26. Extracting a final user behavior vector u from the output of the behavior aggregation module, where the user behavior vector is a vector representation of a fixed dimension.

[0035] Optionally, the S3 specifically includes:

[0036] S31. Concatenate the product representation vector, the user behavior vector, the platform context representation vector, and the advertising target representation vector to generate an initial query structure vector;

[0037] S32, performing multi-layer linear mapping conversion on the initial vector of the query structure to compress the dimension and form a unified representation structure;

[0038] S33, introduce the platform current time and display scene fields, and perform position encoding on the linear mapping output through the position embedding function;

[0039] S34, setting a weight parameter group according to the source category of each input vector, applying an attention fusion strategy to the above-mentioned encoded representation, and generating a multi-source structure weighted representation;

[0040] S35. Use the attention fusion result as the retrieval query vector, perform similarity calculation with the advertising knowledge base index through the vector matching mechanism, and determine the retrieval document set.

[0041] Optionally, the S4 specifically includes:

[0042] S41, mapping the search query vector to a fixed semantic space, and performing a semantic comparison with a pre-generated document vector index structure in the advertising knowledge base;

[0043] S42, using the cosine similarity function as a matching evaluation criterion to calculate the similarity score between the query vector and each document index vector;

[0044] S43, setting a search matching threshold or a fixed return quantity parameter, selecting the top k documents with the highest similarity scores from the advertising knowledge base to form a preliminary search result set;

[0045] S44. Applying a domain filtering strategy to the preliminary search result set to filter out document entries that do not meet the requirements based on advertising category tags, platform delivery restrictions, or delivery time windows;

[0046] S45. Extracting structural information fields from the retained document entries, including document summaries, keyword tags, original content segments, and matching score values;

[0047] S46: Encode the structure information field into a unified vector format, and generate a document content representation set.

[0048] Optionally, the S5 specifically includes:

[0049] S51, decomposing the multiple document contents output by the retrieval module, extracting the document summary field, the keyword tag field, and the matching score field;

[0050] S52, embedding and encoding the fields, and concatenating them to generate document structure representation vectors, and combining all document structure representations to form a document representation matrix;

[0051] S53: Interact the user behavior vector, the platform context vector, and the advertising target vector with the document representation matrix, respectively, using a multi-head attention mechanism to form multiple context-related fusion representations;

[0052] S54, applying component normalization and structure compression strategies to the fusion representation, and outputting a document-behavior fusion representation, a document-platform fusion representation, and a document-target fusion representation;

[0053] S55, splicing the three types of fused representations with the original document structure representation, and performing structural integration to form a fused structure representation;

[0054] S56. Unifying the fused structure representation through a fusion function, wherein the fusion function includes a splicing control structure, an attention enhancement structure, and a component weight gating structure, for adjusting the combined ratio of user behavior, platform context, and advertising target in the final embedded representation;

[0055] S57: Generate an advertising context embedding vector and submit it to the advertising generation model.

[0056] Optionally, the S6 specifically includes:

[0057] S61, inputting the advertisement context embedding vector into the encoding layer of the advertisement generation model, wherein the encoding layer is used to establish semantic associations between embedding components;

[0058] S62: Generate path control parameters based on the user behavior component, platform context component, and advertising target component included in the advertising context embedding vector, wherein the control parameters include a behavior adjustment weight, a platform style label, and a target template number;

[0059] S63: calling a corresponding decoding path according to the control parameter, wherein the decoding path corresponds to a set of style-specific text generation submodules, and the modules include differentiated language structure control strategies;

[0060] S64. During the decoding process, an advertisement text sequence is generated in an autoregressive manner. The advertisement context embedding vector is introduced as a conditional input at each decoding moment. The consistency of the association between the text and the input embedding is maintained through the attention mechanism.

[0061] S65. Setting a field generation sequence structure for the advertisement text, where the fields include a brand description field, a product element field, a scene semantic field, and an action instruction field, and controlling the paragraph structure output of the text decoding module according to the field sequence;

[0062] S66: Generate a complete candidate advertisement text, and output the generation control path record, decoding state sequence, and attention weight distribution corresponding to the advertisement text as part of the advertisement generation output data structure.

[0063] Optionally, the S7 specifically includes:

[0064] S71. Receive candidate advertisement texts generated by an advertisement generation model;

[0065] S72. Encode the candidate advertisement text to extract a semantic vector representation of the text. The encoding process uses a text embedding network that is consistent with the advertisement generation model structure to ensure consistent input representation dimensions.

[0066] S73: Concatenate the text embedding representation with the user behavior vector, the platform context vector, and the advertising target representation vector to construct a prediction input vector;

[0067] S74: Input the predicted input vector into a conversion rate prediction model, where the model includes at least one layer of feature interaction network and one layer of nonlinear mapping network, for calculating a conversion probability value;

[0068] S75. Setting a learnable weight group in the conversion rate prediction model, corresponding to the advertisement content semantic component, the user behavior component, the platform environment component, and the goal-oriented component, respectively, to adjust the degree of influence of different sources on the prediction result;

[0069] S76: Output the conversion rate prediction value, and bind the value with the candidate advertisement text to generate a structured output result, wherein the structured output result includes the advertisement text body, the prediction score and its corresponding vector mapping and scoring weight structure.

[0070] Optionally, the S8 specifically includes:

[0071] S81. Collect user feedback data after the candidate ad text is released online, where the user feedback data includes a conversion flag indicating whether the user has completed a conversion behavior after the ad is displayed;

[0072] S82. Constructing advertising copy preference sample pairs based on the conversion mark, wherein the preference sample pairs include a set of positive sample copies and a set of negative sample copies, wherein the positive sample copies correspond to texts with conversions and the negative sample copies correspond to texts without conversions;

[0073] S83, semantically encode the positive and negative sample texts to generate positive text representation and negative text representation, and jointly represent them with the advertising context embedding vector to form a training sample structure;

[0074] S84. Construct a preference probability objective function, and construct an optimization objective based on the assumption that positive samples should be preferred in the preference pair, wherein the objective function is used to maximize the difference between the probability of generating positive samples and minimize the probability of generating negative samples;

[0075] S85. During the training process, freeze the parameter structure of the basic language model and only update the submodules related to the generation control path, including the fusion layer parameters and the decoding control weights;

[0076] S86. The advertisement generation model after the optimization training is completed replaces the original model structure.

[0077] The beneficial effects of the present invention are:

[0078] (1) The structured modeling of multi-source advertising context is realized, which integrates product information, user behavior, platform environment and advertising goals, constructs advertising context embedding vectors, and improves the semantic integrity and context expression ability of the generated input.

[0079] (2) Introducing a retrieval enhancement generation structure and constructing a behavior-driven, platform-guided query vector to call knowledge documents that are highly relevant to the context before generating advertising content, thus achieving semantically targeted retrieval and content pre-enhancement in advertising generation.

[0080] (3) A preference comparison training path based on user feedback was established, and a direct preference optimization algorithm was used to update the strategy of the generation model, forming a closed-loop optimization mechanism that does not require manual labeling, thereby improving the behavioral response and conversion performance of advertising copy in the delivery scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0082] Figure 1 This is an overall process structure diagram of an intelligent advertising promotion method based on a large language model proposed by the present invention. DETAILED DESCRIPTION

[0083] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0084] refer to Figure 1, an intelligent advertising promotion method based on a large language model, comprising the following steps:

[0085] S1. Collect advertising input data;

[0086] S2. Generate a user behavior vector based on the advertisement input data;

[0087] S3. Concatenate the user behavior vector and the advertisement input data to construct a search query vector, and input the vector into a search module based on a vector matching structure. The search module calculates semantic similarity and outputs multiple related document contents.

[0088] S4. Obtaining document content matching the search query vector from an advertising knowledge base through the search module;

[0089] S5. Concatenate the document content with the user behavior vector, the advertisement input data, and the platform context data to generate an advertisement context embedding vector;

[0090] S6. Inputting the advertisement context embedding vector into an advertisement generation model to generate candidate advertisement texts;

[0091] S7, inputting the candidate advertisement text into a conversion rate prediction module constructed based on the feature interaction structure to obtain a conversion probability score corresponding to the candidate advertisement text;

[0092] S8. Deliver the candidate advertisement text to a target platform, record user feedback data associated with the candidate advertisement text, and optimize and train the advertisement generation model based on the user feedback data.

[0093] In this embodiment, S1 specifically includes:

[0094] S11. Collect product information, user behavior data, platform context data, and advertising target parameters;

[0095] S12: Encode the product information, user behavior data, platform context data, and advertising target parameters into product representation vectors x p , user behavior representation vector x u , platform context representation vector x c and the advertising target representation vector x t ;

[0096] S13. Construct an advertisement input vector X. The input vector is represented by the following formula:

[0097] X=Concat(x p ,x u ,x c ,x t )+α·TimeEmbed(xc )+β·WeightNorm(x t );

[0098] Where Concat(·) represents the vector concatenation function, TimeEmbed(·) represents the temporal position embedding function of the platform context vector, WeightNorm(·) represents the weighted normalization function of the advertising target vector, and α and β are scale adjustment parameters.

[0099] S14: Submit the advertisement input vector X to the user behavior vector generation module and the search query vector construction module as advertisement input data.

[0100] This implementation captures all key contextual factors involved in ad generation through the collection of ad input data. By encoding the raw field content, unifying the vector dimensions, and organizing it into a structured representation, it facilitates the precise expression of user intent, platform preferences, and content goals in subsequent tasks. The context adjustment factor is used to adjust the fusion ratio of various input components in different ad contexts, improving subsequent generation quality.

[0101] In this embodiment, S2 specifically includes:

[0102] S21. Receive the user behavior representation part x in the advertisement input vector X. u ;

[0103] S22. Mark and map the behavior types in the behavior record, encoding each behavior type into a fixed-dimensional behavior type vector;

[0104] S23. Calculate a behavior time factor based on the timestamp information of each behavior record, where the behavior time factor is used to represent the distance between the time when the behavior occurred and the current time;

[0105] S24, combining the behavior type vector and its corresponding time factor to generate a behavior segment representation with a time-weighted scale;

[0106] S25. Inputting all behavior segment representations into a behavior aggregation module, wherein the behavior aggregation module includes one or more layers of sequential modeling networks for constructing a behavior temporal context structure;

[0107] S26. Extracting a final user behavior vector u from the output of the behavior aggregation module, where the user behavior vector is a vector representation of a fixed dimension.

[0108] This implementation organizes user behavior data into a time series structure and introduces a time factor to weight the importance of behaviors, thereby achieving a focused representation of users' recent preferences. The behavior aggregation module can encode behavior sequences using structures such as bidirectional GRU, Transformer, or sliding attention mechanisms, effectively integrating short-term and long-term behavioral features to improve the accuracy and contextual awareness of user behavior representation. The resulting user behavior vector serves not only as a generation input but also in subsequent retrieval path construction and conversion prediction tasks.

[0109] In this embodiment, S3 specifically includes:

[0110] S31. Concatenate the product representation vector, the user behavior vector, the platform context representation vector, and the advertising target representation vector to generate an initial query structure vector;

[0111] S32, performing multi-layer linear mapping conversion on the initial vector of the query structure to compress the dimension and form a unified representation structure;

[0112] S33, introduce the platform current time and display scene fields, and perform position encoding on the linear mapping output through the position embedding function;

[0113] S34, setting a weight parameter group according to the source category of each input vector, applying an attention fusion strategy to the above-mentioned encoded representation, and generating a multi-source structure weighted representation;

[0114] S35. Use the attention fusion result as the retrieval query vector, perform similarity calculation with the advertising knowledge base index through the vector matching mechanism, and determine the retrieval document set.

[0115] This implementation method realizes semantic query construction based on multi-dimensional advertising context by introducing structured fusion strategy, multi-layer semantic mapping and position embedding mechanism, improves the context adaptability and matching accuracy of document retrieval, and provides more directional and relevant document support for advertising content generation.

[0116] In this embodiment, the S4 specifically includes:

[0117] S41, mapping the search query vector to a fixed semantic space, and performing a semantic comparison with a pre-generated document vector index structure in the advertising knowledge base;

[0118] S42, using the cosine similarity function as a matching evaluation criterion to calculate the similarity score between the query vector and each document index vector;

[0119] S43, setting a search matching threshold or a fixed return quantity parameter, selecting the top k documents with the highest similarity scores from the advertising knowledge base to form a preliminary search result set;

[0120] S44. Applying a domain filtering strategy to the preliminary search result set to filter out document entries that do not meet the requirements based on advertising category tags, platform delivery restrictions, or delivery time windows;

[0121] S45. Extracting structural information fields from the retained document entries, including document summaries, keyword tags, original content segments, and matching score values;

[0122] S46: Encode the structure information field into a unified vector format, and generate a document content representation set.

[0123] This implementation method achieves accurate document content extraction for advertising context tasks by constructing a document retrieval path based on semantic space alignment and similarity evaluation, and combining domain constraint filtering and structure encoding processing, providing a highly relevant information foundation for the context embedding construction and text generation stages.

[0124] In this embodiment, the S5 specifically includes:

[0125] S51, decomposing the multiple document contents output by the retrieval module, extracting the document summary field, the keyword tag field, and the matching score field;

[0126] S52, embedding and encoding the fields, and concatenating them to generate document structure representation vectors, and combining all document structure representations to form a document representation matrix;

[0127] S53: Interact the user behavior vector, the platform context vector, and the advertising target vector with the document representation matrix, respectively, using a multi-head attention mechanism to form multiple context-related fusion representations;

[0128] S54, applying component normalization and structure compression strategies to the fusion representation, and outputting a document-behavior fusion representation, a document-platform fusion representation, and a document-target fusion representation;

[0129] S55, splicing the three types of fused representations with the original document structure representation, and performing structural integration to form a fused structure representation;

[0130] S56. Unifying the fused structure representation through a fusion function, wherein the fusion function includes a splicing control structure, an attention enhancement structure, and a component weight gating structure, for adjusting the combined ratio of user behavior, platform context, and advertising target in the final embedded representation;

[0131] S57: Generate an advertising context embedding vector and submit it to the advertising generation model.

[0132] This implementation method constructs an advertising context embedding vector with structural constraints and contextual expression capabilities through the fusion calculation of document content and multi-source contextual information, combined with the attention mechanism and gating control method. It provides a stable, controllable and expressive input basis for the generation model, and helps to improve the consistency and context adaptability of advertising content generation.

[0133] In this embodiment, S6 specifically includes:

[0134] S61, inputting the advertisement context embedding vector into the encoding layer of the advertisement generation model, wherein the encoding layer is used to establish semantic associations between embedding components;

[0135] S62: Generate path control parameters based on the user behavior component, platform context component, and advertising target component included in the advertising context embedding vector, wherein the control parameters include a behavior adjustment weight, a platform style label, and a target template number;

[0136] S63: calling a corresponding decoding path according to the control parameter, wherein the decoding path corresponds to a set of style-specific text generation submodules, and the modules include differentiated language structure control strategies;

[0137] S64. During the decoding process, an advertisement text sequence is generated in an autoregressive manner. The advertisement context embedding vector is introduced as a conditional input at each decoding moment. The consistency of the association between the text and the input embedding is maintained through the attention mechanism.

[0138] S65. Setting a field generation sequence structure for the advertisement text, where the fields include a brand description field, a product element field, a scene semantic field, and an action instruction field, and controlling the paragraph structure output of the text decoding module according to the field sequence;

[0139] S66: Generate a complete candidate advertisement text, and output the generation control path record, decoding state sequence, and attention weight distribution corresponding to the advertisement text as part of the advertisement generation output data structure.

[0140] This implementation method ensures the controllability of the advertising copy in terms of semantic style, expression order and generation process by constructing an advertising generation decoding process with a multi-path and field-controlled structure, and combining it with embedded context-guided generation, thereby improving the structural consistency and expression stability of the generated content in language expression and delivery adaptation.

[0141] In this embodiment, the S7 specifically includes:

[0142] S71. Receive candidate advertisement texts generated by an advertisement generation model;

[0143] S72. Encode the candidate advertisement text to extract a semantic vector representation of the text. The encoding process uses a text embedding network that is consistent with the advertisement generation model structure to ensure consistent input representation dimensions.

[0144] S73: Concatenate the text embedding representation with the user behavior vector, the platform context vector, and the advertising target representation vector to construct a prediction input vector;

[0145] S74: Input the predicted input vector into a conversion rate prediction model, where the model includes at least one layer of feature interaction network and one layer of nonlinear mapping network, for calculating a conversion probability value;

[0146] S75. Setting a learnable weight group in the conversion rate prediction model, corresponding to the advertisement content semantic component, the user behavior component, the platform environment component, and the goal-oriented component, respectively, to adjust the degree of influence of different sources on the prediction result;

[0147] S76: Output the conversion rate prediction value, and bind the value with the candidate advertisement text to generate a structured output result, wherein the structured output result includes the advertisement text body, the prediction score and its corresponding vector mapping and scoring weight structure.

[0148] This implementation introduces a conversion rate prediction module to perform semantic evaluation and behavioral response prediction on advertising copy, forming a unified judgment basis by integrating the fusion representation of user behavior, platform context and advertising goals with the generated content, achieving a structured evaluation of copy effectiveness, and providing basic scoring input for subsequent feedback training modules.

[0149] In this embodiment, S8 specifically includes:

[0150] S81. Collect user feedback data after the candidate ad text is released online, where the user feedback data includes a conversion flag indicating whether the user has completed a conversion behavior after the ad is displayed;

[0151] S82. Constructing advertising copy preference sample pairs based on the conversion mark, wherein the preference sample pairs include a set of positive sample copies and a set of negative sample copies, wherein the positive sample copies correspond to texts with conversions and the negative sample copies correspond to texts without conversions;

[0152] S83, semantically encode the positive and negative sample texts to generate positive text representation and negative text representation, and jointly represent them with the advertising context embedding vector to form a training sample structure;

[0153] S84. Construct a preference probability objective function, and construct an optimization objective based on the assumption that positive samples should be preferred in the preference pair, wherein the objective function is used to maximize the difference between the probability of generating positive samples and minimize the probability of generating negative samples;

[0154] S85. During the training process, freeze the parameter structure of the basic language model and only update the submodules related to the generation control path, including the fusion layer parameters and the decoding control weights;

[0155] S86. The advertisement generation model after the optimization training is completed replaces the original model structure.

[0156] This implementation converts user conversion feedback into preference training sample pairs, and uses the direct preference optimization algorithm (DPO) to update the structure of the advertising generation model, forming a closed-loop training path that does not require external reward modeling. It achieves behavior-driven adaptive improvement of the generation strategy and effectively enhances the responsive coupling between advertising content and the actual delivery environment.

[0157] Example 1:

[0158] To verify the feasibility of this invention, we applied it to a large, comprehensive e-commerce platform with over 20 million daily active users. This platform features a rich variety of ad placements, including homepage banners, activity page carousels, detail page recommendations, and search-linked cards. Over 100,000 ad copy is delivered daily, covering over 50 categories of goods. The original system, which combined manual templates with operational rules for copy management, suffered from low generation efficiency, a single content style, and an inability to quickly adapt to changing user interests.

[0159] In this context, the advertisement generation method based on retrieval enhancement generation model and user feedback optimization proposed in the present invention has been fully deployed on the platform. The overall structure of the system consists of an input vector construction module, a semantic retrieval module, an advertisement generation model, a conversion rate prediction module and a feedback optimization module. First, the platform extracts basic commodity fields from the commodity database, including category labels, brand information, keywords, pricing information, etc.; user behavior data includes browsing, collecting, adding to cart, purchasing and other behaviors. The system uses a sliding window structure to extract high-frequency behaviors in the past 15 days, and uses a time decay mechanism to construct a user behavior vector; platform environment data includes display location, channel source, access time, device identification, etc.; advertising delivery targets are composed of the population targeting, budget limit, delivery cycle, etc. set in the delivery plan.

[0160] The aforementioned input data is structured and normalized, then concatenated into a unified advertising context representation vector, which is used to generate the search query representation. The retrieval module is built based on cosine similarity and a vector index structure. The platform's in-house knowledge base contains over 48 million documents, including historical advertising copy, extracted product reviews, brand corpus, and real user reviews. Each document is preprocessed into summary, keyword, and trending comment fields, and semantic encoding is completed offline.

[0161] Before copywriting is generated, the system retrieves the document content most similar to the ad input based on the current context. It then integrates multi-source contextual fusion with user behavior, platform environment, and ad objectives to ultimately form an ad context embedding vector. This embedding vector is then fed into a large language model custom-trained on the platform. The model employs an autoregressive architecture, integrating a field control module with a template guidance strategy, to output a multi-field structured ad copy, including a brand opening statement, product highlights, scenario placement, and behavioral guidance segments.

[0162] The generated copy is submitted to the platform's advertising engine for actual delivery, and user behaviors such as clicks, stays, add-to-cart, and conversions are collected in real time. Based on conversion results, the system automatically constructs preference sample pairs, using conversion-producing copy as positive samples and non-conversion or low-performing copy as negative samples. Using the Direct Preference Optimization (DPO) algorithm, the copy generation model is optimized through preference learning. Only parameters related to the copy structure are fine-tuned, ensuring that the model generation strategy continuously aligns with user response preferences, achieving a closed-loop optimization.

[0163] To validate the effectiveness of our method, the platform conducted a 10-day, large-scale experiment comparing three system solutions: a control group used a traditional template system without contextual modeling and policy feedback; solution one employed our RAG search enhancement module but without DPO; and solution two fully implemented our method. Each of the three experimental groups received equal traffic resources, tested 4,800 SKUs, and generated over 18 million impressions. The key experimental results are as follows:

[0164] Table 1 Comparison of experimental data of advertisement generation system

[0165]

[0166] As can be seen from the table, Solution 2, which fully applies the method of the present invention, is significantly better than the other two groups in all indicators. The click-through rate increased by more than 115% compared with the control group, the conversion rate increased by more than 230%, and the user's stay time on the product page was significantly extended, indicating that the generated copy is more attractive and has the ability to guide purchases. At the same time, the average user rating of the copy reached 4.6 points, far exceeding the manual template copy, proving that the generated results have high readability and marketing performance. Due to the increase in click-through rate and conversion rate, the unit click delivery cost decreased by more than 28% compared with the baseline group, resulting in actual delivery budget savings.

[0167] Compared with the traditional prompt generation method, the core advantages of the method of the present invention lie in three points: first, through input structure standardization and multi-source context modeling, it ensures that user preferences and display environment information are fully obtained before copy generation, thereby improving the quality of semantic input; second, through the semantic retrieval enhancement module, highly relevant document content is introduced to achieve the transformation from "language-driven generation" to "knowledge-supported generation"; third, through preference sample construction and DPO training mechanism, real user behavior feedback is directly used to optimize the generation strategy, construct a self-learning copy generation model, and avoid the problem of pure language model drift that deviates from business goals.

[0168] At present, the platform has deployed the method of the present invention to three major traffic entrances, namely homepage focus advertising, detail page recommendations, and brand venue targeted copywriting. It generates more than 140,000 copywritings per day, with the response delay controlled within 800 milliseconds. The system has been running stably for more than 90 days, supporting high-concurrency, low-cost, and real-time response copywriting tasks.

[0169] Comprehensive experimental verification and platform application feedback results show that the advertisement generation method provided by the present invention has clear business conversion gains, cost advantages and user experience improvement effects. It is suitable for many scenarios such as large-scale delivery, precise matching of user groups and brand semantic consistency control, and has broad industry application value and technology promotion prospects.

[0170] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent advertising promotion method based on a large language model, characterized in that: The steps include: S1. Collect advertising input data; S2. Generate a user behavior vector based on the advertisement input data; S3. Concatenate the user behavior vector and the advertisement input data to construct a search query vector, and input the vector into a search module based on a vector matching structure. The search module calculates semantic similarity and outputs multiple related document contents. S4. Obtaining document content matching the search query vector from an advertising knowledge base through the search module; S5. Concatenate the document content with the user behavior vector, the advertisement input data, and the platform context data to generate an advertisement context embedding vector; S6. Inputting the advertisement context embedding vector into an advertisement generation model to generate candidate advertisement texts; S7, inputting the candidate advertisement text into a conversion rate prediction module constructed based on the feature interaction structure to obtain a conversion probability score corresponding to the candidate advertisement text; S8. Deliver the candidate advertisement text to a target platform, record user feedback data associated with the candidate advertisement text, and optimize and train the advertisement generation model based on the user feedback data.

2. The intelligent advertising promotion method based on a large language model according to claim 1, characterized in that: Said S1 specifically includes: S11. Collect product information, user behavior data, platform context data, and advertising target parameters; S12: Encode the product information, user behavior data, platform context data, and advertising target parameters into product representation vectors x p , user behavior representation vector x u , platform context representation vector x c and the advertising target representation vector x t ; S13. Construct an advertisement input vector X. The input vector is represented by the following formula: X=Concat(x p ,x u ,x c ,x t )+α·TimeEmbed(x c )+β·WeightNorm(x t ); Where Concat(·) represents the vector concatenation function, TimeEmbed(·) represents the temporal position embedding function of the platform context vector, WeightNorm(·) represents the weighted normalization function of the advertising target vector, and α and β are scale adjustment parameters. S14: Submit the advertisement input vector X to the user behavior vector generation module and the search query vector construction module as advertisement input data.

3. The intelligent advertising promotion method based on a large language model according to claim 1, characterized in that: The S2 specifically includes: S21. Receive the user behavior representation part x in the advertisement input vector X. u ; S22. Mark and map the behavior types in the behavior record, encoding each behavior type into a fixed-dimensional behavior type vector; S23. Calculate a behavior time factor based on the timestamp information of each behavior record, where the behavior time factor is used to represent the distance between the time when the behavior occurred and the current time; S24, combining the behavior type vector and its corresponding time factor to generate a behavior segment representation with a time-weighted scale; S25. Inputting all behavior segment representations into a behavior aggregation module, wherein the behavior aggregation module includes one or more layers of sequential modeling networks for constructing a behavior temporal context structure; S26. Extracting a final user behavior vector u from the output of the behavior aggregation module, where the user behavior vector is a vector representation of a fixed dimension.

4. The intelligent advertising promotion method based on a large language model according to claim 1, characterized in that: The S3 specifically includes: S31. Concatenate the product representation vector, the user behavior vector, the platform context representation vector, and the advertising target representation vector to generate an initial query structure vector; S32, performing multi-layer linear mapping conversion on the initial vector of the query structure to compress the dimension and form a unified representation structure; S33, introduce the platform current time and display scene fields, and perform position encoding on the linear mapping output through the position embedding function; S34, setting a weight parameter group according to the source category of each input vector, applying an attention fusion strategy to the above-mentioned encoded representation, and generating a multi-source structure weighted representation; S35. Use the attention fusion result as the retrieval query vector, perform similarity calculation with the advertising knowledge base index through the vector matching mechanism, and determine the retrieval document set.

5. The intelligent advertising promotion method based on a large language model according to claim 1, characterized in that: The S4 specifically includes: S41, mapping the search query vector to a fixed semantic space, and performing a semantic comparison with a pre-generated document vector index structure in the advertising knowledge base; S42, using the cosine similarity function as a matching evaluation criterion to calculate the similarity score between the query vector and each document index vector; S43, setting a search matching threshold or a fixed return quantity parameter, selecting the top k documents with the highest similarity scores from the advertising knowledge base to form a preliminary search result set; S44. Applying a domain filtering strategy to the preliminary search result set to filter out document entries that do not meet the requirements based on advertising category tags, platform delivery restrictions, or delivery time windows; S45. Extracting structural information fields from the retained document entries, including document summaries, keyword tags, original content segments, and matching score values; S46: Encode the structure information field into a unified vector format, and generate a document content representation set.

6. The intelligent advertising promotion method based on a large language model according to claim 1, characterized in that: The S5 specifically includes: S51, decomposing the multiple document contents output by the retrieval module, extracting the document summary field, the keyword tag field, and the matching score field; S52, embedding and encoding the fields, and concatenating them to generate document structure representation vectors, and combining all document structure representations to form a document representation matrix; S53: Interact the user behavior vector, the platform context vector, and the advertising target vector with the document representation matrix, respectively, using a multi-head attention mechanism to form multiple context-related fusion representations; S54, applying component normalization and structure compression strategies to the fusion representation, and outputting a document-behavior fusion representation, a document-platform fusion representation, and a document-target fusion representation; S55, splicing the three types of fused representations with the original document structure representation, and performing structural integration to form a fused structure representation; S56. Unifying the fused structure representation through a fusion function, wherein the fusion function includes a splicing control structure, an attention enhancement structure, and a component weight gating structure, for adjusting the combined ratio of user behavior, platform context, and advertising target in the final embedded representation; S57: Generate an advertising context embedding vector and submit it to the advertising generation model.

7. The intelligent advertising promotion method based on a large language model according to claim 1, characterized in that: The S6 specifically includes: S61, inputting the advertisement context embedding vector into the encoding layer of the advertisement generation model, wherein the encoding layer is used to establish semantic associations between embedding components; S62: Generate path control parameters based on the user behavior component, platform context component, and advertising target component included in the advertising context embedding vector, wherein the control parameters include a behavior adjustment weight, a platform style label, and a target template number; S63: calling a corresponding decoding path according to the control parameter, wherein the decoding path corresponds to a set of style-specific text generation submodules, and the modules include differentiated language structure control strategies; S64. During the decoding process, an advertisement text sequence is generated in an autoregressive manner. The advertisement context embedding vector is introduced as a conditional input at each decoding moment. The consistency of the association between the text and the input embedding is maintained through the attention mechanism. S65. Setting a field generation sequence structure for the advertisement text, where the fields include a brand description field, a product element field, a scene semantic field, and an action instruction field, and controlling the paragraph structure output of the text decoding module according to the field sequence; S66: Generate a complete candidate advertisement text, and output the generation control path record, decoding state sequence, and attention weight distribution corresponding to the advertisement text as part of the advertisement generation output data structure.

8. The intelligent advertising promotion method based on a large language model according to claim 1, characterized in that: The S7 specifically includes: S71. Receive candidate advertisement texts generated by an advertisement generation model; S72. Encode the candidate advertisement text to extract a semantic vector representation of the text. The encoding process uses a text embedding network that is consistent with the advertisement generation model structure to ensure consistent input representation dimensions. S73: Concatenate the text embedding representation with the user behavior vector, the platform context vector, and the advertising target representation vector to construct a prediction input vector; S74: Input the predicted input vector into a conversion rate prediction model, where the model includes at least one layer of feature interaction network and one layer of nonlinear mapping network, for calculating a conversion probability value; S75. Setting a learnable weight group in the conversion rate prediction model, corresponding to the advertisement content semantic component, the user behavior component, the platform environment component, and the goal-oriented component, respectively, to adjust the degree of influence of different sources on the prediction result; S76: Output the conversion rate prediction value, and bind the value with the candidate advertisement text to generate a structured output result, wherein the structured output result includes the advertisement text body, the prediction score and its corresponding vector mapping and scoring weight structure.

9. The intelligent advertising promotion method based on a large language model according to claim 1, characterized in that: The S8 specifically includes: S81. Collect user feedback data after the candidate ad text is released online, where the user feedback data includes a conversion flag indicating whether the user has completed a conversion behavior after the ad is displayed; S82. Constructing advertising copy preference sample pairs based on the conversion mark, wherein the preference sample pairs include a set of positive sample copies and a set of negative sample copies, wherein the positive sample copies correspond to texts with conversions and the negative sample copies correspond to texts without conversions; S83, semantically encode the positive and negative sample texts to generate positive text representation and negative text representation, and jointly represent them with the advertising context embedding vector to form a training sample structure; S84. Construct a preference probability objective function, and construct an optimization objective based on the assumption that positive samples should be preferred in the preference pair, wherein the objective function is used to maximize the difference between the probability of generating positive samples and minimize the probability of generating negative samples; S85. During the training process, freeze the parameter structure of the basic language model and only update the submodules related to the generation control path, including the fusion layer parameters and the decoding control weights; S86. The advertisement generation model after the optimization training is completed replaces the original model structure.

Images