Dynamic advertisement putting method and system oriented to large language model

Abstract

The invention relates to a dynamic advertisement putting method and system oriented to a large language model, and belongs to the technical field of generative artificial intelligence and calculation advertisement crossing. According to the method, the semantic splitting problem of advertisements and natural language streams is solved through a three-level collaborative architecture, firstly, a semantic depth alignment module is constructed, and advertisement and natural language vectors are projected to a unified semantic space by using dual-channel adversarial training and orthogonal mapping constraints; a dynamic context adaptation module is constructed, dominant intentions are analyzed based on a self-attention mechanism, and scene advertisements are generated by fusing space-time-equipment signals; and finally, constructing an anti-fragility regulation and control module, and dynamically balancing the advertisement density and the user experience through an exponential attenuation model and a reinforcement learning fusing mechanism. The method breaks through the limitation of a static advertisement template, realizes the adaptive fusion of the advertisement content and the generative interaction scene, and is suitable for scenes such as a multi-language dialogue system and an intelligent search engine.

Description

Technical Field

[0001] The present invention relates to a dynamic advertising delivery method and system for a large language model, belonging to the technical field of intersection of generative artificial intelligence and computational advertising. Background Art

[0002] Against the backdrop of rapid advancements in artificial intelligence (AI) technology, intelligent interactive systems driven by large language models (LLMs) are gradually replacing traditional search engines. This transformation has had a profound impact on the digital advertising ecosystem. Traditional digital advertising technologies, such as keyword matching and user behavior prediction, have significant flaws in generative scenarios driven by large language models:

[0003] (1) Semantic fragmentation: Static advertising templates cannot adapt to dynamic conversation flows, resulting in commercial information being disconnected from the context;

[0004] (2) Distribution shift: The ad vector representation deviates from the LLM semantic space, causing information filtering attenuation;

[0005] (3) Lack of adaptive regulation: The lack of a dynamic intensity adjustment mechanism can easily lead to a deterioration of user experience due to advertising density overload.

[0006] Some improved solutions, such as knowledge graph-based ad insertion methods, struggle to capture dynamic intent, and fixed template-based ad generation technologies are unable to adapt to multimodal scenarios. They also suffer from issues such as coarse contextual awareness and insufficient semantic alignment. To address these issues, the present invention proposes an ad delivery method with a three-level collaborative architecture that achieves semantic consistency, scenario awareness, and automatic system adjustment. This approach provides an innovative path for intelligent interactive systems that balances user acceptance and commercial conversion efficiency, fostering a sustainable and evolving ad delivery ecosystem. Summary of the Invention

[0007] The purpose of the present invention is to provide a dynamic advertising delivery method and system for large language models, aiming to solve technical problems such as semantic fragmentation, insufficient scene adaptation and lack of ecological regulation in generative interaction scenarios in the existing technology.

[0008] To achieve the above objectives, the technical solution of the present invention is: a dynamic advertising delivery method and system for large language models. This method achieves seamless integration of advertising content and natural language generation streams through a three-level collaborative architecture of deep semantic alignment, dynamic context adaptation, and anti-fragile regulation. The specific steps are as follows:

[0009] Step 1: Build a deep semantic alignment module to achieve deep semantic alignment between advertisements and natural language, bridge the gap between advertisements and natural language expression, and obtain advertisement content that conforms to natural language expression;

[0010] Step 2: Build a dynamic context adaptation module to enhance the relevance of the ad content, the user's current intent, and the interaction scenario through intelligent context analysis, solving the problem of commercial information being disconnected from the context and achieving content that seamlessly integrates with the ad.

[0011] Step 3: Build an anti-fragile control module and adopt a content delivery strategy that dynamically adjusts and seamlessly integrates with advertising to achieve coordinated optimization of commercial value and content quality.

[0012] The step 1 is specifically as follows:

[0013] Building the ad encoder E a With the natural language encoder E n , semantic space alignment is achieved through a dual-channel generative adversarial network;

[0014] The ad encoder uses a gated recurrent unit network and takes as input a decommercialized ad text sequence, wherein the decommercialization process includes replacing explicit promotional words with regular expressions and verifying grammatical compliance with a sequence labeling model.

[0015] The natural language encoder is adjusted based on the pre-trained language model, and the output layer is connected to the orthogonal mapping matrix ,satisfy Constrain the vector space orthogonality, where for The transposed matrix of

[0016] In adversarial training, the generator optimizes the domain discrimination difference optimization objective through the gradient reversal layer. ,in is the total loss, α and β are weight coefficients, , To combat losses:

[0017]

[0018]

[0019] in, is the discriminator, For the ad encoder, convert the ad data Mapped to the semantic space V, For natural language encoder, natural language Mapping to semantic space , To get the expectation of the advertising data distribution, The goal of the discriminator is to maximize this loss, while the goal of the ad encoder is to minimize it.

[0020] The step 2 is specifically as follows:

[0021] Explicit intent parsing: Extract the top-K entities from the large language model self-attention matrix, and then expand the associated entity set through the knowledge graph:

[0022]

[0023] in, Represents a collection of related entities, is the i-th entity, represents the similarity score between the i-th entity and the core entity in the knowledge graph, and θ is the threshold;

[0024] Implicit scene modeling: Constructing a spatiotemporal-device joint feature vector = (geolocation ⊕ timestamp ⊕ device type;

[0025] Knowledge-driven ad generation: Triggered when the activation frequency of the associated entity exceeds the threshold, an ad card containing competitor comparisons and multimodal adaptation is generated.

[0026] The step three is specifically as follows:

[0027] The ad weight of the dynamic decay algorithm is adjusted according to the exponential model, and the formula is:

[0028]

[0029] in, is the adjusted weight, W0 is the initial weight, λ is the step size, N1 is the number of repeated exposures of the advertisement, CRT t is the click rate of the current window period, CRT avg is the historical average click-through rate;

[0030] Monitor user response time and ad ratio in real time. If the limit is exceeded, temporarily adjust the lambda value or suspend ad insertion.

[0031] Circuit breaker mechanism: When the negative feedback rate of exposure exceeds the threshold for several consecutive times, the delivery is stopped and the reinforcement learning strategy optimization module is started. The reinforcement learning reward is calculated based on the click-through rate, context matching degree and sentiment score. The reward function is:

[0032]

[0033] Among them, a, b, and c are the corresponding weight coefficients respectively. The optimized advertisements will be re-launched after manual review.

[0034] Furthermore, the natural language encoder adopts a multilingual pre-training model to support cross-language semantic alignment.

[0035] Furthermore, the advertising card is a knowledge-enhanced advertising card. When the knowledge-enhanced advertising card is generated, a multilingual knowledge graph interface is called to adapt to the grammatical structure and cultural characteristics of the target language.

[0036] Furthermore, the knowledge enhancement advertising card includes:

[0037] Competitive product comparison module: obtains functional parameters and user reviews from public data sources and generates visual charts highlighting advantages;

[0038] Multimodal Adaptor: Dynamically selects rich media formats based on device signals, with constraints including file size, loading time, ad card size, screen ratio, and click hotspot size.

[0039] Furthermore, the reinforcement learning strategy optimization module of the circuit breaker mechanism includes:

[0040] Input layer: advertisement text vector, user historical behavior characteristics, and device scenario labels;

[0041] Policy network: A 3-layer fully connected network whose output actions are instructions for generating ad content variants.

[0042] Manual review and annotation dimensions include information accuracy, compliance, and cultural adaptability.

[0043] To achieve the above objectives, the present invention further provides a dynamic advertising delivery system for a large language model, comprising:

[0044] Semantic Deep Alignment Module: performs two-channel adversarial training and cross-lingual vector projection;

[0045] Dynamic context adaptation module: integrates self-attention parser, knowledge graph interface and multimodal renderer;

[0046] Anti-fragile control module: deploys dynamic attenuation algorithm, circuit breaker decision tree and reinforcement learning strategy network.

[0047] The innovation of the present invention is:

[0048] Semantic deep alignment: Semantic deep alignment technology is used to bridge the gap between advertising and natural language expression, allowing commercial information to be naturally integrated into the generated content flow, avoiding the semantic gap caused by traditional rigid implantation.

[0049] Dynamic context adaptation: Based on a multimodal dynamic perception framework, it realizes intelligent scene analysis and personalized adaptation, breaking through the limitations of static delivery mode and significantly enhancing the relevance of advertisements to users' current intentions and interaction contexts.

[0050] Anti-fragile regulation: The anti-fragile regulation mechanism establishes a self-balancing system for business ecology and language generation. Through dynamic weight adjustment and intelligent circuit breaker strategies, it effectively prevents advertising density overload from interfering with user experience, and achieves coordinated optimization of commercial value and content quality.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] This approach, through a three-level collaborative architecture, significantly improves the semantic consistency between ads and context, enhances multi-scenario adaptability, and establishes a dynamic control mechanism to prevent user experience degradation. It effectively eliminates the interference of advertising content on language generation fluency, significantly improves the fit between commercial information and context, and is compatible with mainstream large language model architectures. It significantly enhances the technical robustness and applicability of large language model commercialization, providing an innovative path for intelligent interactive systems that balances user acceptance and commercial conversion efficiency. It can be applied to scenarios such as conversational AI systems, intelligent search engines, and commercial information integration on multimodal content platforms, achieving precise and flexible advertising delivery results and forming a sustainable and evolving advertising delivery ecosystem. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the steps of the present invention;

[0054] Figure 2 It is a flow chart of dual-channel adversarial training of the present invention;

[0055] Figure 3 It is a dynamic attenuation schematic diagram of the present invention;

[0056] Figure 4 It is a diagram of the fuse mechanism of the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0058] like Figure 1 As shown, a dynamic advertising delivery method and system for a large language model specifically includes the following steps:

[0059] Step 1: Build a semantic deep alignment module to achieve deep semantic alignment between advertisements and natural language, and obtain advertisement content that conforms to natural language expression.

[0060] Specifically, we constructed the advertising dataset D1 by extracting 100,000 pieces of advertising text from the original materials provided by advertisers. We then used regular expressions to filter out explicit promotional terms (such as "limited-time sale" and "buy one get one free") and replaced them with neutral descriptions. We then verified the semantic consistency of the replaced text with the original ad, ensuring that the core information was retained after the replacement. For example, we reconstructed "RTX 4080 limited-time sale" into "high-performance graphics card," annotating them with structured semantic tags (such as "Bluetooth 5.0 protocol"), sentiment keywords (such as "stable" and "efficient"), and scenario-specific tags (such as "mobile device compatible"). These tags were stored in JSON format and mapped one-to-one with the ad text for subsequent model training.

[0061] Specifically, the D2 natural dialogue dataset was constructed by integrating open-source community corpora (e.g., Reddit and Zhihu Q&A) covering 20 vertical sectors, including technology, e-commerce, and travel. Generative models (e.g., DeepSeek) were used to expand the data to 1 million conversations, ensuring semantic diversity. For example, a user question, "How do I solve my Bluetooth pairing problem?" was expanded into a multi-turn conversation containing technical details (e.g., "Check the Bluetooth protocol version on my phone"). A noise injection strategy was introduced to randomly insert typos or colloquial expressions (e.g., "What should I do if my Bluetooth connection fails?") to improve model robustness.

[0062] Specifically, a dual-channel encoder is trained, such as Figure 2 As shown;

[0063] For Ad Encoder E a :

[0064] The network structure is as follows: the input layer has 768 dimensions and the maximum sequence length is 256; the hidden layer has 3 layers of GRU, each layer has a hidden state dimension of 1024, and a dropout rate of 0.2 to prevent overfitting; the output layer has a fully connected layer that projects the hidden state to 768 dimensions.

[0065] Input: De-commercialized advertising text sequence (e.g., "high-performance graphics card" instead of "RTX 4080 limited-time sale").

[0066] For the natural language encoder E n :

[0067] The natural language encoder adopts a multilingual pre-training model to support cross-language semantic alignment. Specifically, based on the fine-tuning of the BERT-Large pre-training model, the output layer is connected to the orthogonal mapping matrix W proj Positive samples are semantically similar conversation pairs (e.g., “Bluetooth headset pairing problem” and “device protocol compatibility”), and negative samples are randomly sampled from conversations in different fields (e.g., “graphics card performance comparison” is sampled from the technology category as a negative sample for the travel category). The temperature coefficient is 0.07.

[0068] Specifically, the adversarial training process is:

[0069] Discriminator D pre-training: fixed E a With E n Parameters, use cross entropy loss to optimize the domain classification ability of D, and output domain classification probability (0: advertising, 1: natural language);

[0070] Freeze the D parameter and back-propagate the domain classification loss through the Gradient Reversal Layer;

[0071] Optimization goal: minimize the weighted sum of semantic similarity loss and adversarial loss ,in is the total loss, α and β are weight coefficients, , To combat the loss, the present invention sets α=0.7 and β=0.3.

[0072] Step 2: Build a dynamic context adaptation module to enhance the relevance of the advertisement content, the user's current intention, and the interaction scenario through intelligent context analysis, thereby obtaining content that is seamlessly integrated with the advertisement.

[0073] First, perform explicit intent analysis:

[0074] Self-attention matrix analysis: Extract the self-attention weight matrix A∈R from the 6th and 12th layers of the large language model n×n ;

[0075] Calculate the entity attention score from the number of entities detected in the current conversation round, and extract the top-K entities based on the attention score;

[0076] Knowledge graph expansion: call the knowledge graph to expand synonymous entities to the candidate set ,in, Represents a collection of related entities, is the i-th entity, Represents the similarity score between the i-th entity and the core entity in the knowledge graph, θ is the threshold, and in this invention =0.7.

[0077] Then, implicit scene modeling:

[0078] Constructing joint feature vectors = (Geographic location ⊕ timestamp ⊕ device type, where geographic location encoding uses the Geohash algorithm (6-digit precision, ±0.6km error) to convert the user location into a string (for example, "Xihu District, Hangzhou City" is encoded as "WTW3SP9X"); regional features (such as "commercial area" and "residential area") are obtained through the reverse geocoding API and encoded into a 16-dimensional vector.

[0079] Decompose the timestamp into hour (t_h∈[0,23]), day of week (t_d∈[0,6]) and holiday flag (t_holy∈{0,1}), normalize them and concatenate them into a 4-dimensional vector;

[0080] The device type includes the device type code, such as Mobile_Android, iOS, and PC, which is one-hot encoded into an 8-dimensional vector. The screen resolution is discretized into standard sizes (such as "1080×1920" and "1440×2560") and encoded into a 6-dimensional vector.

[0081] Joint vector generation: The above features are mapped into scene vectors through a 3-layer MLP (256→128→64 dimensions).

[0082] Finally, generate knowledge-enhanced ads:

[0083] Count the number of occurrences of related entities per thousand characters, with a threshold of f e ≥2×10⁻³ chars⁻¹ (i.e., triggered at least once for every 500 characters). When the entity activation frequency is greater than the threshold, the knowledge graph API is called to generate an advertising card. For example, the entity chain is: Bluetooth protocol → compatible accessories → promotional SKU. Based on the user's historical interaction rate (for example, when the click-through rate CTR is greater than 5%, the threshold is reduced to 1.5×10⁻³).

[0084] Real-time competitor data acquisition: Obtain data from mainstream e-commerce platforms, including price, user ratings (≥ 4 stars), and functional parameters (using regular expressions to extract numerical features).

[0085] Multimodal rendering constraints: For example, on mobile devices: the width of the ad card occupies 90% of the screen, the height is dynamically adjusted (20%-25%), the text size is ≥12pt, the button spacing is ≥8pt, the click hotspot is ≥48×48 pixels, the anti-mistouch delay is 300ms, and the sliding trigger threshold is ≥50px; rich media scene embedding uses WebGL to render the interactive model (.glb format), the model size is limited to 5MB, and the loading time is ≤1.5 seconds; audio ads: when the device audio output is detected to be activated, a 5-second brand prompt tone is inserted (sampling rate 44.1kHz, bit rate 128kbps), and the volume automatically adapts to 70% of the system setting.

[0086] Step 3: Build an anti-fragile control module and adopt a content delivery strategy that dynamically adjusts and seamlessly integrates with advertising to achieve coordinated optimization of commercial value and content quality.

[0087] Specifically, the advertising weight of the dynamic decay algorithm is adjusted according to the exponential model, and the formula is:

[0088]

[0089] in, is the adjusted weight, W0 is the initial weight, λ is the step size, N1 is the number of repeated exposures of the advertisement, CRT t is the click rate of the current window period, CRT avg is the historical average click rate, and the dynamic decay is shown as follows Figure 3 As shown;

[0090] Real-time monitoring of user response time (≤3 seconds) and ad ratio (≤15%). If these limits are exceeded, the lambda value will be temporarily adjusted or ad insertion will be suspended.

[0091] The initial weight of the present invention is W0=1, λ=0.2, the grid search range is 0.1-0.3, and the step size is 0.05.

[0092] Circuit breaker mechanism: When the negative feedback rate of exposure exceeds the threshold for several consecutive times, the trigger stops the delivery and starts the reinforcement learning strategy optimization module. The reinforcement learning reward is calculated by the click rate, context matching and sentiment score. The circuit breaker mechanism is as follows: Figure 4 shown.

[0093] Furthermore, the negative feedback rate of the window period (the last 10 ad requests) is calculated in real time. :

[0094]

[0095] Negative feedback user behaviors include "closing ads", "complaining", and "staying time <1 second". In the present invention, when the negative feedback rate of exposure is ≥30% for three consecutive times, the ad delivery is stopped and the abnormal ID is recorded.

[0096] Furthermore, the reinforcement learning strategy network architecture is:

[0097] State space: ad text vector (256 dimensions) + user behavior features (128 dimensions);

[0098] Action Space:

[0099] Dimension 1: Adjustment of advertising tone (promotional → technical comparison);

[0100] Dimension 2: content density adjustment (information volume ±20%);

[0101] Reward function:

[0102]

[0103] Wherein, a, b, and c are corresponding weight coefficients respectively. In the present invention, a=0.6, b=0.3, and c=0.1.

[0104] The optimized ads will be put on the market again after manual review.

[0105] Specifically, manual review includes audit marking dimensions such as "information accuracy" (such as whether product parameters are correct) and "compliance" (such as no false advertising).

[0106] Furthermore, the reinforcement learning strategy optimization module of the circuit breaker mechanism includes:

[0107] Input layer: advertisement text vector, user historical behavior characteristics, and device scenario labels;

[0108] Policy network: A 3-layer fully connected network whose output actions are instructions for generating ad content variants.

[0109] Manual review and annotation dimensions include information accuracy, compliance, and cultural adaptability.

[0110] The present invention also provides a dynamic advertising delivery system for a large language model, comprising:

[0111] Semantic Deep Alignment Module: performs two-channel adversarial training and cross-lingual vector projection;

[0112] Dynamic context adaptation module: integrates self-attention parser, knowledge graph interface and multimodal renderer;

[0113] Anti-fragile control module: deploys dynamic attenuation algorithm, circuit breaker decision tree and reinforcement learning strategy network.

[0114] Example 1: Dynamic advertising delivery in social media interaction scenarios

[0115] Scenario description:

[0116] A user posts on a social media platform: "I want to go camping with friends this weekend. Can you recommend some lightweight tents and cooking utensils?" The system parses the text in real time and triggers the dynamic ad delivery process.

[0117] System processing flow:

[0118] Semantic Depth Alignment:

[0119] Ad text reconstruction: The original ad, "XX brand tents, 50% off for a limited time!", has been de-commercialized and replaced with "XX brand ultralight tent (weighing only 1.2kg, windproof level 5)"

[0120] Vector Alignment: Ad Encoder E a The processed text is mapped to the same semantic space as user dynamics (natural language), with a cosine similarity of 0.89.

[0121] Dynamic context adaptation:

[0122] Explicit intent analysis: Extract core entities "tent", "cooking utensils", and "camping" from the text, and expand related entities (such as "windproof tent" and "portable gas stove") through the knowledge graph.

[0123] Implicit scene modeling:

[0124] Geographic location: The user's IP address is resolved to "xx city", the Geohash code is "WTW3SJ", and it is associated with a nearby camping site (such as "xxx wetland").

[0125] Device type: Mobile (iOS), screen resolution 1125×2436.

[0126] Joint eigenvector Driver selects the "Image + Geolocation Tag" ad template.

[0127] Knowledge-enhanced ad generation:

[0128] Data comparison: Shows the weight (1.2kg vs. 8kg) and price (￥399 vs. ￥499) of "XX tent" and its competitors.

[0129] Multimodal adaptation: Embedded 3D tent expansion animation (.glb format, 3MB), supporting users' 360-degree viewing.

[0130] Anti-fragile regulation:

[0131] Dynamic decay: The initial weight of the ad is W0 = 0.7. If the user browses for 3 seconds without clicking, the weight decays to W according to λ = 0.2. t =0.56.

[0132] Circuit breaker mechanism: If a user replies "This is not what I need..." twice after two consecutive exposures, the ad weight drops from 0.45 to 0.2, triggering a circuit breaker and replacing it with pure recommendation content: "Here is another user's camping equipment list →"

[0133] Example 2: Advertisement Adaptation for Multilingual Intelligent Customer Service Scenario

[0134] Scenario description:

[0135] A user asks the AI customer service representative in French: "My laptop is slow, what should I do?" The system needs to generate ads in French and adapt to multilingual scenarios.

[0136] System processing flow:

[0137] Semantic Depth Alignment:

[0138] Multi-language processing: The original advertisement "XX SSD 500GB special price 299 yuan!" was reconstructed after de-commercialization to "SSD500GB (read and write speed 550MB / s, compatible with laptops)".

[0139] Cross-lingual vector alignment: Natural Language Encoder E n Supports multilingual BERT, mapping French queries and ad text into a unified semantic space with a similarity of 0.85.

[0140] Dynamic context adaptation:

[0141] Explicit intent parsing: Extract the entity laptop and slow speed, and expand to related entities such as "SSD" and "memory upgrade".

[0142] Implicit scene modeling:

[0143] Geographic location: The user IP address is located in "XX City" and the timestamp is 15:00 on a weekday (normalized t / 24=0.625).

[0144] Device signal: PC (Windows 11), screen resolution 1920×1080.

[0145] Joint eigenvector Trigger the "Technical Solutions + Promotional Information" ad template.

[0146] Knowledge-enhanced ad generation:

[0147] Data comparison: Compare the read and write speeds (550MB / s vs. 450MB / s) and warranty period (5 years vs. 3 years) of "XX SSD" with those of competing products.

[0148] Multimodal adaptation: Embedded SSD installation tutorial video (30 seconds, 720p resolution), supports French subtitles.

[0149] Anti-fragile regulation:

[0150] Dynamic decay: The initial weight W0=0.6. After the user clicks the ad, the weight increases to 0.65. If the user does not click but stays for 5 seconds, the weight is fine-tuned to W. t =0.55 (λ=0.15).

[0151] The above describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A dynamic advertising delivery method for a large language model, characterized in that: The following steps are involved: Step 1: Build a semantic deep alignment module to achieve deep alignment of advertisements with natural language semantics, and obtain advertisement content that conforms to natural language expression; Step 2: Build a dynamic context adaptation module to enhance the relevance of the ad content, the user's current intent, and the interaction scenario through intelligent context analysis, resulting in content that seamlessly integrates with the ad; Step 3: Build an anti-fragile control module and adopt a content delivery strategy that dynamically adjusts and seamlessly integrates with advertising to achieve coordinated optimization of commercial value and content quality.

2. A dynamic advertising delivery method for a large language model according to claim 1, characterized in that: The step 1 is specifically as follows: Building the ad encoder E a With the natural language encoder E n , semantic space alignment is achieved through a dual-channel generative adversarial network; The ad encoder uses a gated recurrent unit network and takes as input a decommercialized ad text sequence, wherein the decommercialization process includes replacing explicit promotional words with regular expressions and verifying grammatical compliance with a sequence labeling model. The natural language encoder is adjusted based on the pre-trained language model, and the output layer is connected to the orthogonal mapping matrix ,satisfy Constrain the vector space orthogonality, where for The transposed matrix of In adversarial training, the generator optimizes the domain discrimination difference optimization objective through the gradient reversal layer. ,in is the total loss, α and β are weight coefficients, , To combat losses: ； ； in, is the discriminator, For the ad encoder, convert the ad data Mapped to the semantic space V, For natural language encoder, natural language Mapping to semantic space , To get the expectation of the advertising data distribution, To take the expectation of the natural data distribution.

3. The method for dynamic advertising delivery based on a large language model according to claim 1, wherein: The step 2 is specifically as follows: Explicit intent parsing: Extract the top-K entities from the large language model self-attention matrix, and then expand the associated entity set through the knowledge graph: ； in, Represents a collection of related entities, is the i-th entity, represents the similarity score between the i-th entity and the core entity in the knowledge graph, and θ is the threshold; Implicit scene modeling: Constructing a spatiotemporal-device joint feature vector = (geolocation ⊕ timestamp ⊕ device type; Knowledge-driven ad generation: Triggered when the activation frequency of the associated entity exceeds the threshold, an ad card containing competitor comparisons and multimodal adaptation is generated.

4. The method for dynamic advertising delivery based on a large language model according to claim 1, wherein: The step three is specifically as follows: The ad weight of the dynamic decay algorithm is adjusted according to the exponential model, and the formula is: ； in, is the adjusted weight, W0 is the initial weight, λ is the step size, N1 is the number of repeated exposures of the advertisement, CRT t is the click rate of the current window period, CRT avg is the historical average click-through rate; Monitor user response time and ad ratio in real time. If the limit is exceeded, temporarily adjust the lambda value or suspend ad insertion. Circuit breaker mechanism: When the negative feedback rate of exposure exceeds the threshold for several consecutive times, the delivery is stopped and the reinforcement learning strategy optimization module is started. The reinforcement learning reward is calculated based on the click-through rate, context matching degree and sentiment score. The reward function is:

5. Among them, a, b, and c are the corresponding weight coefficients. The optimized ads will be re-launched after manual review.

6. The method for dynamic advertising delivery based on a large language model according to claim 2, wherein: The natural language encoder adopts a multilingual pre-training model and supports cross-language semantic alignment.

7. The method for dynamic advertising delivery based on a large language model according to claim 3, characterized in that: The advertising card is a knowledge-enhanced advertising card. When the knowledge-enhanced advertising card is generated, a multilingual knowledge graph interface is called to adapt to the grammatical structure and cultural characteristics of the target language.

8. The method for dynamic advertising delivery based on a large language model according to claim 6, characterized in that: The knowledge enhancement advertising card includes: Competitive product comparison module: obtains functional parameters and user reviews from public data sources and generates visual charts highlighting advantages; Multimodal Adaptor: Dynamically selects rich media formats based on device signals, with constraints including file size, loading time, ad card size, screen ratio, and click hotspot size.

9. The method for dynamic advertising delivery based on a large language model according to claim 4, wherein: The reinforcement learning strategy optimization module of the circuit breaker mechanism includes: Input layer: advertisement text vector, user historical behavior characteristics, and device scenario labels; Policy network: A 3-layer fully connected network whose output actions are instructions for generating ad content variants. Manual review and annotation dimensions include information accuracy, compliance, and cultural adaptability.

10. A system for implementing the large language model-oriented dynamic advertising delivery method according to claim 1, characterized in that: include: Semantic Deep Alignment Module: performs two-channel adversarial training and cross-lingual vector projection; Dynamic context adaptation module: integrates self-attention parser, knowledge graph interface and multimodal renderer; Anti-fragile control module: deploys dynamic attenuation algorithm, circuit breaker decision tree and reinforcement learning strategy network.

Images