Method and system for generating corpus for dynamic scene-based delivery based on large model

Abstract

The application discloses a large model-based dynamic scene-based delivery corpus generation method and system, and belongs to the technical field of computers. Through the technical scheme provided by the embodiment of the application, the user dynamic scene can be captured in real time to generate adaptive corpus, and response lag caused by static image can be avoided; the strategy engine is used to drive corpus correction, so that the accuracy and flexibility of conversion induction are improved; the internalization learning of platform rules is realized by using a large language model, and the repeated adaptation problem of cross-platform delivery is solved. Therefore, the generated delivery corpus can closely match the real-time needs of users, and the conversion effect of different scenes and platforms is maximized under the compliance premise.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and system for generating dynamic, scenario-based delivery corpora based on large models. Background Technology

[0002] In the field of digital marketing, the technology of automatically generating personalized advertising corpora has become a key means to improve user conversion rates.

[0003] Related technologies typically generate standardized promotional text based on matching rules between static user profiles and product information, or adapt the language to relevant scenarios using pre-set templates combined with simple rules. For example, some systems generate fixed-structured recommendation scripts by analyzing basic user attributes (such as gender and age) and product keywords, or call preset templates to fill dynamic parameters on specific advertising channels (such as e-commerce homepages). While such methods can achieve basic personalization, they rely on manually preset corpora or rule engines, limiting their flexibility and scenario adaptability.

[0004] Therefore, there is an urgent need for a method to generate corpora that can be deployed, in order to solve the core problems of lagging scenario response, lack of strategy guidance and insufficient platform adaptation in related technologies. Summary of the Invention

[0005] This application provides a method and system for generating dynamic, scenario-based delivery corpus based on a large model. It has the advantages of improving scenario targeting, enhancing strategy guidance, and optimizing platform adaptability. The technical solution is as follows:

[0006] On the one hand, a method for generating dynamic, contextualized delivery corpus based on a large model is provided, the method comprising:

[0007] In response to the request to generate corpus for the target object, the system determines the target delivery scenario, target delivery strategy, and target product corresponding to the target object, and obtains the object information and delivery reference information of the target object. The target delivery scenario is the scenario when the target object triggers the request to generate corpus. Based on the target delivery scenario, object information, and target product, the system generates the initial corpus corresponding to the target object. Based on the delivery reference information and target delivery strategy of the target object, the initial corpus is modified to obtain the first modified corpus. Through the target large language model, the first modified corpus is modified based on the delivery platform rules carried in the corpus generation request to obtain the target corpus corresponding to the target object. The target corpus is used to attract the target object to click on the product display control corresponding to the target product.

[0008] Furthermore, this application also proposes, in response to a request to generate corpus for delivery to a target object, to determine the target delivery scenario, delivery requirement description information, and delivery stage of the target object corresponding to the request to generate corpus for delivery to the target object, wherein the delivery stage is the stage in which corpus is delivered to the target object; and to determine the target delivery strategy corresponding to the target object based on the delivery requirement description information and the delivery stage.

[0009] Furthermore, this application proposes, in response to a request to generate corpus data for the target audience, obtaining the delivery channel description information, delivery method description information, delivery requirement description information, and the target audience's corpus delivery records corresponding to the corpus generation request; determining the target delivery scenario based on the delivery channel description information, delivery method description information, and delivery requirement description information; determining the delivery stage based on the corpus delivery records; determining the expected delivery effect and delivery requirements based on the delivery requirement description information; determining a reference delivery strategy from multiple candidate delivery strategies based on the delivery stage and expected delivery effect, wherein the reference delivery strategy is a candidate delivery strategy that matches the delivery stage and expected delivery effect; and adjusting the reference delivery strategy using the delivery requirements to obtain the target delivery strategy corresponding to the target audience.

[0010] Furthermore, this application proposes that the target information includes historical behavioral data, a set of interest preference tags, and demographic attributes. Time-series pattern mining is performed on the historical behavioral data to generate behavioral characteristics of the target object. These behavioral characteristics include behavioral frequency distribution, content theme preference, and conversion path sensitivity. The application also proposes determining the tag association strength corresponding to each interest preference tag in the set of interest preference tags. The association strength is determined based on tag co-occurrence frequency and timeliness weight. Tag co-occurrence frequency refers to the frequency with which each interest preference tag in the set of interest preference tags appears together with other interest preference tags within a preset time window. The timeliness weight is used to represent the degree of effectiveness decay of each interest preference tag in the set of interest preference tags over time. Based on the behavioral characteristics, the set of interest preference tags, tag association strength, and demographic attributes, a target object profile is determined. Finally, based on the product information of the target product, the target delivery scenario, and the target object profile, initial delivery corpus is generated.

[0011] Furthermore, this application proposes to identify dynamic behavioral patterns and key behavioral sequences in historical behavioral data. Dynamic behavioral patterns refer to a set of behavioral patterns with temporal evolution characteristics and scenario relevance. Key behavioral sequences refer to continuous behavioral segments strongly correlated with the expected delivery effect corresponding to the delivery corpus generation request. Based on dynamic behavioral patterns and key behavioral sequences, the application determines behavioral frequency distribution, content theme preferences, and conversion path sensitivity. Conversion path sensitivity is determined based on the conversion rate of the key behavioral sequences. An interest topology network is constructed based on multiple interest preference tags whose tag association strength meets preset conditions in the interest preference tag set. Baseline features of the target object are determined based on behavioral characteristics and demographic attributes. An object profile is generated based on the interest topology network and baseline features. A semantic generation framework is determined according to the target delivery scenario. Product information and the object profile are injected into the semantic generation framework to obtain corpus generation information. Initial delivery corpus is generated based on the corpus generation information.

[0012] Furthermore, this application proposes to use benchmark features to perform matching in an interest topology network to obtain multiple target nodes in the interest topology network. Target nodes are nodes whose matching degree with the benchmark features is greater than or equal to a preset matching degree. Based on the node information of multiple target nodes and the benchmark features, an object profile is generated. The corpus generation information is parsed to obtain core interest descriptions, conversion guidance instructions, and scene modification elements. The core interest description is a structured field of the target object's core interest preferences, the conversion guidance instruction is a command parameter used to guide the target object to perform conversion behavior, and the scene modification elements are a set of semantic modifiers adapted to the target delivery scenario. According to the corpus generation rules of the target delivery scenario, the core interest description, conversion guidance instructions, and scene modification elements are combined into a semantic structure tree. The semantic structure tree is converted into the initial delivery corpus through a pre-trained scene adaptation generator.

[0013] Furthermore, this application proposes to generate reference correction information based on competitor corpora and historical campaign performance in the campaign reference information, wherein the competitor corpora are campaign corpora used by competitors of the target product; determine strategy correction parameters based on the target campaign strategy, including style enhancement coefficient and conversion induction intensity; input the reference correction information and strategy correction parameters into the strategy correction engine to generate a multi-dimensional correction rule set; and use the strategy correction engine to semantically reconstruct the initial campaign corpora according to the multi-dimensional correction rule set to output the first corrected campaign corpora.

[0014] Furthermore, this application proposes the following steps: extracting the semantic feature distribution and sentiment parameters of competitor corpora; analyzing the conversion decay trend and user feedback characteristics in historical campaign performance; integrating the semantic feature distribution, sentiment parameters, conversion decay trend, and user feedback characteristics to obtain reference correction information; determining the style guidance identifier and conversion priority parameter in the target campaign strategy; determining the style enhancement coefficient based on the style guidance identifier; determining the conversion induction intensity based on the conversion priority parameter and the conversion behavior characteristics of the target object, whereby the conversion behavior characteristics are determined based on historical behavior data included in the object information; deconstructing the initial campaign corpus into an initial semantic unit set through a strategy correction engine; reorganizing the core semantic units and modifying semantic units in the initial semantic unit set based on a multi-dimensional correction rule set to obtain the target semantic unit set; and optimizing the grammatical coherence of the target semantic unit set to obtain the first corrected campaign corpus.

[0015] Furthermore, this application proposes the following steps: performing cross-platform semantic alignment on competitor corpora to generate unified semantic codes; extracting high-frequency semantic clusters and corresponding sentiment weights from the unified semantic codes based on an attention mechanism; mapping the distribution features of high-frequency semantic clusters with sentiment weights to obtain semantic feature distribution and sentiment parameters; analyzing style enhancement rules and conversion induction rules in a multi-dimensional correction rule set; reconstructing modifier semantic units based on style enhancement rules to inject style-adaptive modifier components; reconstructing core semantic units based on conversion induction rules; fusing the reconstructed modifier semantic units and core semantic units to obtain a target semantic unit set; constructing a grammatical dependency graph of the target semantic unit set; detecting and repairing dependency conflicts in the grammatical dependency graph using a grammatical rule engine; and converting the repaired target semantic unit set into the first corrected delivery corpus.

[0016] Furthermore, this application proposes to parse the platform rules to generate a platform compliance constraint set and platform expression features, including the distribution of platform idioms and user interaction pattern features; to filter the first modified delivery corpus for compliance based on the platform compliance constraint set using a target large language model to obtain a second modified delivery corpus; to input the second modified delivery corpus and platform expression features into the target large language model, and to encode and iteratively decode the second modified delivery corpus and platform expression features using an attention mechanism based on the target large language model to obtain the target delivery corpus.

[0017] On the one hand, a dynamic, contextualized delivery corpus generation system based on a large model is provided, the system comprising:

[0018] The determination module is used to respond to the request to generate delivery corpus for the target object, determine the target delivery scenario, target delivery strategy and target product corresponding to the target object, and obtain the object information and delivery reference information of the target object, wherein the target delivery scenario is the scenario when the target object triggers the request to generate delivery corpus;

[0019] The generation module is used to generate the initial delivery corpus corresponding to the target object based on the target delivery scenario of the target object, the object information of the target object, and the target product.

[0020] The correction module is used to correct the initial delivery corpus based on the delivery reference information of the target object and the target delivery strategy to obtain the first corrected delivery corpus;

[0021] The correction module is further configured to correct the first corrected delivery corpus based on the delivery platform rules carried in the delivery corpus generation request using the target large language model, thereby obtaining the target delivery corpus corresponding to the target object. The target delivery corpus is used to attract the target object to click on the product display control corresponding to the target product.

[0022] On one hand, a computer device is provided, the computer device including one or more processors and one or more memories, the one or more memories storing at least one computer program, the computer program being loaded and executed by the one or more processors to implement the dynamic scene-based delivery corpus generation method based on large models.

[0023] On the one hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, the computer program being loaded and executed by a processor to implement the dynamic scene-based delivery corpus generation method based on a large model.

[0024] On the one hand, a computer program product or computer program is provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to execute the above-mentioned dynamic scenario-based delivery corpus generation method based on a large model.

[0025] The technical solution provided in this application can capture dynamic user scenarios in real time to generate adaptive corpora, avoiding response delays caused by static profiles; it drives corpus correction through a strategy engine, improving the accuracy and flexibility of conversion inducement; and it utilizes a large language model to internalize platform rules, solving the problem of repetitive adaptation in cross-platform advertising. This allows the generated advertising corpora to closely match users' real-time needs, maximizing conversion effects across different scenarios and platforms while ensuring compliance. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram of the implementation environment of a dynamic scenario-based delivery corpus generation method based on a large model provided in an embodiment of this application;

[0028] Figure 2 This is a flowchart of a method for generating dynamic, contextualized delivery corpus based on a large model, provided in an embodiment of this application.

[0029] Figure 3 This is a flowchart illustrating the target delivery strategy provided in an embodiment of this application;

[0030] Figure 4 This is a flowchart of generating the initial delivery corpus provided in an embodiment of this application;

[0031] Figure 5 This is a flowchart of determining the first modified delivery corpus provided in an embodiment of this application;

[0032] Figure 6 This is a flowchart illustrating the process of determining the target delivery corpus provided in an embodiment of this application;

[0033] Figure 7 This is a schematic diagram of the structure of a dynamic, scenario-based delivery corpus generation system based on a large model, provided in an embodiment of this application.

[0034] Figure 8 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0036] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0037] Artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain better results.

[0038] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge sub-models to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instruction-based learning.

[0039] LLM (Large Language Model): LLM is an artificial intelligence model (such as GPT, BERT, PaLM, etc.) trained on massive amounts of text data, utilizing deep learning architectures (especially Transformer) to understand and generate human-like text. In related technologies, LLM is widely used in tasks such as text summarization, translation, question answering, and content creation. In this application, LLM specifically refers to the "target large language model" in the claims, which is customized for use in the generation and optimization process of dynamic advertising corpora.

[0040] Advertising corpus: In the advertising and marketing field, "advertising corpus" usually refers to advertising copy, material descriptions, or promotional information prepared to be displayed to target users on specific channels (such as feed ads, search engine ads, and social media posts). Its core objective is to attract user attention and drive clicks or conversions.

[0041] Strategy Adjustment Engine: In related technologies, rule engines or simple A / B testing tools may be used to adjust ad copy. These tools typically perform limited keyword replacements or version selections based on preset rules or historical data, lacking deep semantic understanding and strategy-driven dynamic reconstruction capabilities. In this application, the strategy adjustment engine is a core, rule-driven intelligent processing module.

[0042] Semantic Reconstruction: In the field of Natural Language Processing (NLP), "semantic reconstruction" broadly refers to the process of changing the structure of a sentence or text while preserving (or specifically altering) its meaning, such as paraphrasing. In ad optimization, this may manifest as simple synonym replacement or sentence structure adjustment. In this application, semantic reconstruction specifically refers to a highly structured and policy-driven deep corpus transformation process performed by a policy correction engine.

[0043] User profiling: "User profiling" is a general concept in marketing and the internet industry, referring to a virtual user model constructed by collecting and analyzing user data (such as demographics, behavior, and interests) to represent the characteristics of a user group or individual. Traditional profiling may be based on static tags or simple statistics. In this application, user profiling refers to a highly refined and structured feature model dynamically constructed for a target object.

[0044] Advertising platform: "Advertising platform" broadly refers to digital channels or systems used for publishing and displaying advertisements, such as social media platforms, search engines, news feed platforms, programmatic advertising exchanges, etc. Each platform has its unique user base, content format, algorithm rules, and review policies. In this application, advertising platform refers to the external online system or channel that ultimately hosts and displays the target advertising corpus.

[0045] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0046] Figure 1 This is a schematic diagram illustrating the implementation environment of a dynamic, contextualized delivery corpus generation method based on a large model, as provided in this application embodiment. See also... Figure 1 The implementation environment may include terminal 110 and server 140.

[0047] Terminal 110 is connected to server 140 via a wireless or wired network. Optionally, terminal 110 can be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. Terminal 110 has an application installed and running that supports the generation of dynamic, contextualized delivery corpora based on large models.

[0048] Server 140 is a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. Server 140 can provide background services for applications running on terminal 110.

[0049] In related technologies, the digital marketing field has long relied on static user profiles and pre-set templates to generate standardized promotional texts. Traditional methods match fixed-structured scripts by analyzing basic user attributes and product keywords, or fill dynamic parameters using templates on specific channels. For example, e-commerce platform homepage pushes recommendation texts based on user age and browsing history. Such methods are limited by manually preset rules and limited scenario adaptation, resulting in content that cannot respond to dynamic user behavior scenarios in real time. Strategy adjustments rely on human experience, and rules need to be repeatedly adapted when deploying across platforms. When users are at different behavioral nodes or on different pages, statically generated texts struggle to capture scene changes, the lack of strategy guidance leads to rigid conversion inducement logic, and differences in platform rules raise content compliance risks.

[0050] To address the aforementioned issues, the inventors identified three major shortcomings in related technologies: delayed scenario response, lack of strategy guidance, and insufficient platform adaptation. By analyzing user behavior data streams, they recognized the dynamic nature of the delivery scenario and the real-time value of strategy guidance. Further research into the evolution patterns of competitor corpora and platform rule characteristics led to the proposal of a technical approach that combines scenario capture, strategy correction, and platform optimization. Based on the multi-round iteration capabilities of a large language model, a complete process from initial generation to dual correction was constructed, achieving dynamic coupling between corpus generation and the delivery environment.

[0051] Therefore, this application proposes a method for generating dynamic, contextualized delivery corpora based on a large model, see [link to relevant documentation]. Figure 2 Taking the server as the executing entity as an example, the method includes the following steps.

[0052] 201. In response to the request to generate delivery corpus for the target object, determine the target delivery scenario, target delivery strategy and target product corresponding to the target object, and obtain the object information and delivery reference information of the target object. The target delivery scenario is the scenario when the target object triggers the request to generate delivery corpus.

[0053] 202. Based on the target audience's target delivery scenario, target audience information, and target product, generate the initial delivery corpus corresponding to the target audience;

[0054] 203. Based on the target audience's referencing information and the target audience's delivery strategy, the initial delivery corpus is revised to obtain the first revised delivery corpus;

[0055] 204. Using the target large language model, the first modified delivery corpus is modified based on the delivery platform rules carried in the delivery corpus generation request to obtain the target delivery corpus corresponding to the target object. The target delivery corpus is used to attract the target object to click on the product display control corresponding to the target product.

[0056] The target delivery scenario refers to the real-time interactive environment in which the user is when the request is triggered. This can be determined by analyzing the user's current page visit, behavioral trajectory, and contextual information, used to capture the user's immediate needs. Object information includes dynamically updated behavioral characteristics and preference tags. This can be generated by using time-series pattern mining algorithms to process historical behavioral data and combining it with tag association strength calculations, supporting the construction of accurate user profiles. The target delivery strategy is a guiding rule based on the delivery stage and demand generation. This can be achieved by parsing the delivery demand description information and stage characteristics through a strategy engine, guiding adjustments to corpus style and conversion logic. Delivery reference information includes the evolution patterns of competitor corpora. This can be achieved by extracting high-frequency semantic features through cross-platform semantic alignment and sentiment analysis, supporting corpus optimization. The target large language model refers to a generated model fine-tuned according to platform rules. This can be achieved by using an attention mechanism to integrate platform expression features, ensuring that the output content conforms to specifications.

[0057] Specifically, when a user triggers a campaign request on a product details page of a social media platform, the system first analyzes the current page elements and dwell time to determine the target campaign scenario as a deep browsing scenario. It then uses behavioral data from the past 30 days to identify user sensitivity to discounts through time-series pattern mining, and constructs an interest topology network by combining this with real-time preference tags. Based on the core selling points and scenario characteristics of the target product, initial corpus highlighting limited-time offers is generated. Subsequently, the system analyzes the high-conversion rhetoric characteristics of competitors in similar scenarios, and reconstructs the persuasive logic strength of the corpus based on the conversion priority required by the current strategy. Finally, the revised corpus is input into a large language model that has learned the interaction patterns of this social media platform, optimizing it into expressions that conform to the platform's emoji usage habits, generating target campaign corpus containing countdown elements and emoji embellishments.

[0058] Compared to related technologies, traditional methods still push generic scripts based on homepage browsing history when users enter the product comparison stage, while this solution can instantly recognize scene changes and generate comparative advantage scripts. Existing strategy adjustments require manual rule setting, while this solution automatically matches the delivery stage to generate differentiated incentive strengths through a strategy engine. When delivering across platforms, traditional methods require redesigning the script structure, while this solution automatically adapts to the platform's expression characteristics through a large language model, maintaining consistency in core information while adjusting the expression style.

[0059] Through the above technical solutions, this application can capture dynamic user scenarios in real time to generate adaptive corpora, avoiding response delays caused by static profiles; drive corpus correction through a strategy engine to improve the accuracy and flexibility of conversion inducements; and utilize a large language model to internalize platform rules, solving the problem of repetitive adaptation in cross-platform advertising. This enables the generated advertising corpora to closely match users' real-time needs, maximizing conversion effects across different scenarios and platforms while ensuring compliance.

[0060] See Figure 3 Taking the server as the executing entity as an example, this application further proposes the following technical solutions.

[0061] 301. In response to the request to generate delivery corpus for the target audience, determine the target delivery scenario, delivery requirement description information, and delivery stage of the target audience corresponding to the delivery corpus generation request;

[0062] 302. Based on the description of the delivery requirements and the delivery stage, determine the target delivery strategy corresponding to the target audience.

[0063] The ad placement requirement description refers to the ad placement objectives, budget constraints, and time requirements described in natural language. Specifically, this can be achieved by using natural language processing techniques to extract ad placement target keywords, budget threshold parameters, and time window parameters, capturing the core needs and real-time business requirements of the current ad placement task. The ad placement stage refers to the phase in which the ad placement is delivered to the target audience, i.e., the ad placement lifecycle nodes divided based on historical ad placement records. This can be achieved by using time series clustering algorithms to divide the target audience's click behavior and conversion behavior into stages, representing the target audience's state characteristics in the current ad placement cycle. The target ad placement strategy refers to the set of dynamically adjusted ad placement generation rules based on the ad placement stage. This can be achieved by combining strategy template library matching with dynamic parameter injection, guiding style selection, conversion induction intensity control, and scenario adaptation rule application during ad placement generation.

[0064] Specifically, when the system receives a request to generate a delivery corpus, it first extracts the delivery requirement description information embedded in the request through a semantic parsing module. For example, it uses named entity recognition technology to extract budget parameters, time parameters, and target keywords. Simultaneously, based on the target audience's historical interaction records, it uses behavioral sequence modeling to identify the current delivery stage, such as the initial reach stage, interest cultivation stage, or repeat purchase guidance stage. Then, it maps the real-time business objectives in the delivery requirement description information to the strategy templates corresponding to the delivery stage. For example, in the repeat purchase guidance stage, it prioritizes matching strategy templates containing loyalty incentive rules, and dynamically injects the budget constraint parameters from the delivery requirement description information into the template parameter adjustment module to generate a target delivery strategy that adapts to the current stage and meets real-time business needs.

[0065] Compared to related technologies, traditional methods generate delivery strategies based solely on static rules or fixed templates, failing to dynamically adjust strategy parameters according to changes in the target audience's delivery stage. For example, related technologies still employ the same incentive strategies used in the initial reach stage during the repeat purchase stage, leading to decreased conversion efficiency. This application, however, establishes a dynamic mapping mechanism between delivery stages and strategy templates, enabling the strategy generation process to inherit historical delivery experience and adapt to real-time business needs, effectively solving the problem of strategy-stage disconnect.

[0066] Through the above technical solution, this application achieves dynamic adaptation of the entire lifecycle of the delivery strategy and the delivery stage, avoiding the decline in conversion rate caused by the rigidity of the strategy. For example, when the target audience is in the interest cultivation stage, the system automatically matches a strategy template containing multi-round interactive guidance rules and adjusts the frequency of content presentation according to real-time delivery needs, thereby improving the response speed and adaptation accuracy of the strategy to changes in user behavior.

[0067] This application further proposes a method for responding to a request to generate corpus data for a target audience, which involves obtaining the following information: descriptions of the delivery channel, delivery method, and delivery requirements corresponding to the request; determining the target delivery scenario based on the delivery channel, delivery method, and delivery requirements; determining the delivery stage based on the corpus delivery records; determining the desired delivery effect and delivery requirements based on the delivery requirements; determining a reference delivery strategy from multiple candidate delivery strategies based on the delivery stage and desired delivery effect; and adjusting the reference delivery strategy according to the delivery requirements to obtain the target delivery strategy for the target audience.

[0068] The delivery channel description information refers to the platform or media type when the target audience triggers the delivery request. This can be implemented using channel identifiers and interaction interface parameters to characterize the channel characteristics and interaction methods of the delivery scenario, ensuring that the delivery strategy is compatible with channel characteristics. The delivery method description information refers to the content display format and user reach method specified in the delivery request. This can be implemented using interaction mode encoding and display template identifiers to capture the interaction characteristics of the delivery scenario and guide strategy generation to match user reach preferences. The corpus delivery record refers to the corpus content and feedback data received by the target audience during historical delivery periods. This can be implemented using time-series behavior logs and conversion event databases. By analyzing historical delivery frequency, response rate, and conversion path, the current delivery stage can be dynamically identified. The expected delivery effect refers to the core conversion goal set in the delivery requirement description information. This can be implemented using conversion type identifiers and effect intensity parameters to establish effect matching constraints during candidate strategy selection. The candidate delivery strategy refers to the validated delivery strategies stored in a pre-established strategy library. This can be implemented by associating a strategy template library with effect evaluation indicators. By matching the delivery stage with the expected effect, the basic strategy framework can be quickly located. The delivery requirements refer to the personalized execution constraints specified in the delivery demand description information. Specifically, they can be implemented using compliance rule sets and style preference parameters, which are used to optimize the parameters and inject rules into the reference strategy to achieve customized adjustments to the strategy.

[0069] Specifically, by acquiring descriptions of the delivery channel, delivery method, and delivery requirements, the system can fully extract the channel attributes, interaction characteristics, and target demands of the current delivery scenario, avoiding the strategy mismatch problem caused by the one-sided extraction of scenario features in traditional methods. Time-series analysis of the delivery records can identify different delivery stages, such as initial contact, secondary engagement, or deep conversion, providing a dynamic basis for strategy selection. During strategy generation, a joint matching mechanism between expected delivery effects and delivery stages can filter reference strategies from the candidate strategy library that both conform to stage characteristics and effect goals, forming a basic strategy framework. Further, by analyzing the personalized requirements in the delivery demands, the parameters of the reference strategies are adjusted and rules are optimized, ultimately generating a target delivery strategy that balances general effectiveness with scenario specificity. This dual mechanism of phased selection and dynamic optimization effectively resolves the contradiction between static rules and dynamic requirements during strategy generation.

[0070] Compared to related technologies, existing campaign strategy generation methods typically select strategies based on fixed rules or single-stage characteristics, lacking the ability to dynamically respond to the evolving characteristics and changing needs of the campaign stage. For example, traditional systems may continue to use the strategy template from the initial reach stage even after users have entered the deep conversion stage, resulting in insufficient conversion incentives. This solution, however, determines the dynamic campaign stage by analyzing campaign records in real time and combines a stage-effect matching mechanism from the candidate strategy library. It can automatically switch the strategy benchmark based on the user's current stage and dynamically optimize strategy parameters through needs analysis, significantly improving the scenario adaptability and effectiveness of strategy generation.

[0071] Through the above technical solutions, this application can dynamically adjust the strategy baseline framework according to the real-time identified campaign stage, avoiding the decline in conversion efficiency caused by strategy lag; through the demand analysis and strategy optimization mechanism, personalized execution requirements are incorporated while ensuring the effectiveness of the strategy, resolving the contradiction between rigid strategies and changing demands in traditional strategy generation methods; the scenario recognition mechanism based on channel and interaction characteristics ensures that the generated campaign strategy is accurately matched with the interaction characteristics of the specific campaign scenario, improving user acceptance and conversion inducement effect during strategy execution.

[0072] This application further proposes a technical solution for generating initial delivery corpus based on the target delivery scenario, object information, and target product. (See [link to relevant documentation]). Figure 4 Taking the server as the executing entity as an example, the following steps are included.

[0073] 401. Perform time-series pattern mining on historical behavioral data to generate behavioral characteristics of the target object;

[0074] 402. Determine the tag association strength corresponding to each interest preference tag in the interest preference tag set;

[0075] 403. Based on behavioral characteristics, interest preference tag sets, tag association strength, and demographic attributes, determine the object profile of the target object;

[0076] 404. Generate initial delivery corpus based on product information of the target product, target delivery scenario and target profile.

[0077] The object information includes historical behavioral data, a set of interest preference tags, and demographic attributes. Behavioral characteristics include behavioral frequency distribution, content theme preference, and conversion path sensitivity. Association strength is determined based on tag co-occurrence frequency and time-sensitivity weight. Tag co-occurrence frequency refers to the frequency with which each interest preference tag in the set appears together with other interest preference tags within a preset time window. Time-sensitivity weight represents the degree of effectiveness decay of each interest preference tag over time. Time-series pattern mining refers to the dynamic analysis of time series patterns in user historical behavioral data, specifically implemented using time-series clustering algorithms and Hidden Markov Models, to capture the temporal evolution characteristics of user behavior frequency distribution and the real-time sensitivity of conversion paths. Tag association strength refers to the dynamic association degree between interest preference tags, specifically calculated using co-occurrence frequency statistics combined with a time decay function, used to eliminate time-sensitivity interference from historical tags and reflect the structural characteristics of the interest network. Time-sensitivity weight refers to the decay coefficient of the effectiveness of interest preference tags over time, specifically implemented using an exponential decay model, used to dynamically adjust the contribution of tags in profile construction. User profiling refers to a multi-dimensional representation of users that integrates dynamic behavioral characteristics, time-sensitive interest networks, and demographic attributes. Specifically, it can be constructed using graph neural networks and feature fusion algorithms to accurately depict users' real-time interests and behavioral patterns.

[0078] Specifically, time series analysis is used to extract behavioral frequency distribution, content theme preferences, and conversion path sensitivity from historical behavioral data, forming dynamic behavioral characteristics. Co-occurrence frequency statistics and time-sensitivity weight calculations are performed on interest preference tags to construct a time-sensitive interest association network. The dynamic behavioral characteristics, interest network, and demographic attributes are input into the profile generation model, outputting a multi-dimensional object profile. Based on the attribute characteristics of the target product and the semantic framework of the delivery scenario, the object profile is mapped to the semantic structure of the initial delivery corpus. Natural language generation technology is used to output promotional content adapted to real-time user characteristics. Specifically, time series pattern mining addresses the lag problem of traditional static statistical methods by capturing the periodic and trend changes in user behavior; tag association strength calculation eliminates noise interference from historical data by dynamically adjusting the association weights of interest tags; and the time-sensitivity weight mechanism automatically updates tag validity through an exponential decay model, ensuring the real-time accuracy of the interest network.

[0079] Compared to related technologies, which typically employ static tag libraries and fixed time windows to statistically analyze user behavior, resulting in weak interest correlation and lagging updates to behavior patterns, this solution dynamically captures the evolutionary patterns of user behavior through time-series pattern mining. It then adjusts the effectiveness of interest tags using time-sensitivity weights, enabling real-time updates to user profiles. Traditional methods cannot effectively handle the time-sensitivity decay of interest tags; this solution introduces a time decay function to dynamically calculate tag weights, improving the representation accuracy of the interest network. Furthermore, object profiles in related technologies are often discrete feature sets; this solution constructs a multi-dimensional fusion profile model to enhance the semantic relevance of user representations.

[0080] Through the above technical solutions, this application can dynamically capture the temporal evolution characteristics of user behavior patterns, solving the problem of lagging updates in traditional static profiles. By calculating the association strength and timeliness weight of interest tags, the accuracy and real-time performance of the interest network representation are improved. Based on a multi-dimensional fusion-based profile generation mechanism, the matching degree between the delivery corpus and the user's real-time interests is enhanced. Utilizing a dual mechanism of time-series analysis and timeliness weight adjustment, it is ensured that the generated initial corpus not only conforms to the user's current behavioral characteristics but also adapts to the real-time needs of the target delivery scenario.

[0081] This application further proposes a technical solution to perform time-series pattern mining on historical behavioral data to generate behavioral features of target objects, determine object profiles based on behavioral features, interest preference tag sets, tag association strength and demographic attributes, and generate initial delivery corpus according to the target delivery scenario.

[0082] Among them, dynamic behavior patterns refer to a set of behavioral patterns with temporal evolution characteristics and scenario relevance. Specifically, a sliding window algorithm combined with a hidden Markov model can be used to identify the periodic changes and scenario dependence of user behavior over time, capturing the dynamic patterns of user behavior changes with time and scenario. Key behavior sequences refer to continuous behavioral segments that are strongly correlated with the expected delivery effect corresponding to the delivery corpus generation request. Specifically, association rule mining algorithms can be used to extract continuous action combinations that are statistically significant with conversion behavior from users' historical behavior, identifying key trigger points in the user conversion path. Tag association strength is determined based on tag co-occurrence frequency and time-sensitivity weight. Specifically, a co-occurrence matrix calculation combined with an exponential decay function can be used to quantify the synergistic effect and time sensitivity between interest tags, addressing the problem that traditional tag weight calculation ignores interest relevance and time-sensitivity decay. The interest topology network is constructed from multiple interest preference tags whose tag association strength meets preset conditions. Specifically, a graph neural network can be used to cluster highly correlated tags into subgraph structures, representing the hierarchical relationship between user interests. The semantic generation framework is determined based on the target delivery scenario. Specifically, a template selection mechanism based on scenario classification can be adopted. For example, price-sensitive verbal structures can be selected for e-commerce promotion scenarios to ensure the semantic adaptability of the initial corpus to the delivery scenario.

[0083] Specifically, user historical behavior data is decomposed into dynamic behavior patterns and key behavior sequences through time-series pattern mining. Dynamic behavior patterns reveal the evolution of user behavior over time, while key behavior sequences identify behavioral segments strongly correlated with conversion. The association strength of interest preference tags is calculated using co-occurrence frequency and timeliness weights to select highly correlated tags and construct an interest topology network. Combined with baseline features formed by behavioral characteristics and demographic attributes, an object profile integrating dynamic behavior and timeliness interests is generated. The target product information and object profile are injected into a semantic generation framework adapted to the target scenario. For example, in a social media advertising scenario, the framework prioritizes interactive dialogue templates, maps core nodes in the user interest topology network to corpus keywords, adjusts the intensity of persuasive dialogue based on the conversion sensitivity of key behavior sequences, and finally outputs the initial delivery corpus through a pre-trained scenario adaptation generator.

[0084] Compared to related technologies, existing solutions rely on static user profiles and fixed templates to generate corpora, failing to capture the temporal changes in user behavior and the time-dependent decay of interest preferences. For example, traditional methods only match product keywords with basic user attributes to generate scripts, ignoring the impact of recent changes in user behavior patterns on the conversion path. This solution addresses the issues of delayed behavior capture and interest decay by mining dynamic behavior patterns and calculating time-dependent weights; it also achieves deep matching between corpus content and delivery scenarios by constructing an interest topology network and a scenario-adaptive semantic generation framework.

[0085] Through the above technical solutions, this application can improve the accuracy of dynamic capture of user behavior patterns, such as accurately identifying changes in users' shopping preferences at different times; enhance the timeliness of interest preference tags, such as reducing the interference of expired interest tags on profile construction; and improve the semantic adaptability of object profiles and delivery scenarios, such as automatically strengthening the generation weight of limited-time offer messages in holiday marketing scenarios, thereby improving the accuracy and conversion inducement effect of dynamic scenario-based delivery corpus.

[0086] This application further proposes a method for generating object profiles based on interest topology networks and benchmark features, including using benchmark features to match multiple target nodes in the interest topology network, generating object profiles based on the node information of the target nodes and benchmark features; when generating the initial delivery corpus, core interest descriptions, conversion guidance instructions and scene modification elements are obtained by parsing the corpus generation information, and combined into a semantic structure tree and converted into the initial delivery corpus.

[0087] The baseline features are feature vectors generated through historical behavioral data and demographic attribute analysis. Specifically, time-series pattern mining algorithms can be used to extract behavior frequency distribution, content theme preferences, and conversion path sensitivity, combined with demographic attributes to generate a multi-dimensional feature set to characterize user dynamic behavior patterns and static attributes. The interest topology network is a graph structure composed of interest preference tags and their relationships. Specifically, tag co-occurrence frequency can be used to calculate node connection strength, and graph embedding algorithms can be used to construct the network topology, reflecting the dynamic relationships between interest tags. Target nodes are network nodes whose matching degree with the baseline features exceeds a threshold. Specifically, cosine similarity can be used to calculate the matching degree between the baseline feature vector and the node feature vector, filtering out the most correlated interest tag clusters. The semantic structure tree is a structured semantic representation organized according to scene rules. Specifically, context-free grammars can be used to define generation rules, with core interest descriptions as root nodes and transformation guidance indicators and scene modification elements as child nodes to construct a tree structure. The scene adaptation generator is a model that converts structured semantics into natural language. Specifically, a Transformer-based sequence-to-sequence model can be used, with an attention mechanism to map semantic elements to natural language sentences.

[0088] Specifically, in generating user profiles, user behavior features are first extracted through time-series analysis and combined with demographic attributes to form a baseline feature vector. This vector is input into an interest topology network for similarity matching, filtering out target node clusters with high correlation strength. The label association information carried by the network nodes is fused with the baseline features to generate dynamically updated user profiles. In the corpus generation stage, product information and dynamic profiles are parsed into structured elements, core interest descriptions extract core user preferences, conversion guidance indicators set behavioral guidance parameters, and scene-adaptive modifiers are added to scene-specific elements. These elements are combined into a semantic structure tree according to scene grammar rules and converted into natural language text that conforms to scene expression habits through a pre-trained generator.

[0089] Compared to related technologies, traditional methods rely on static tag matching to generate fixed dialogue templates, making it difficult to capture the dynamic correlation of user interests. This solution constructs an interest topology network to achieve dynamic tag association analysis, and combines this with baseline feature matching to filter highly relevant interest clusters, enabling the generated user profile to reflect real-time interest evolution. In the corpus generation stage, related technologies use linear template filling, resulting in rigid semantic structures. In contrast, this solution uses a semantic structure tree to achieve hierarchical organization of elements, and combines scene rules to guide the generation process, ensuring that semantic expression both aligns with user interests and adapts to platform interaction features.

[0090] Through the above technical solutions, this application solves the problem of insufficient adaptation between the delivery corpus and dynamic scenarios caused by the static user profile, and realizes the real-time update of the profile by dynamically matching interest network nodes; at the same time, it overcomes the defect of low matching degree between semantic structure and scene rules, and ensures that the corpus content accurately expresses user needs and conforms to the platform interaction specifications through the structured semantic tree generation mechanism.

[0091] This application further proposes the following technical solutions, see [link / reference] Figure 5 Taking the server as the executing entity as an example, the following steps are included.

[0092] 501. Based on competitor data and historical campaign performance in the campaign reference information, generate reference correction information;

[0093] 502. Determine strategy adjustment parameters based on the target delivery strategy. Strategy adjustment parameters include style enhancement coefficient and conversion induction intensity.

[0094] 503. Input the reference correction information and strategy correction parameters into the strategy correction engine to generate a multi-dimensional correction rule set;

[0095] 504. Using the strategy correction engine, the initial delivery corpus is semantically reconstructed based on a multi-dimensional correction rule set, and the first corrected delivery corpus is output.

[0096] Among them, competitor corpus refers to the advertising data used by competitors of the target product. Specifically, it can be obtained through cross-platform data collection interfaces, which can be used to analyze market trends and competitor strategies. Historical advertising performance refers to conversion rate and click-through rate metrics in the historical advertising data of the target audience or similar users. Specifically, it can be extracted through event tracking logs and user behavior tracking systems to identify effective advertising patterns. Style enhancement coefficient is a quantitative parameter used to control the intensity of the expressive style of the corpus. Specifically, it can be generated by feature mapping of style-oriented identifiers in the target advertising strategy through natural language processing models. Conversion induction intensity is a priority parameter used to adjust the conversion guidance instructions in the corpus. Specifically, it can be calculated and generated through a relationship model between conversion behavior features and advertising stages. Strategy correction engine is a computational module that performs multi-dimensional rule fusion and semantic reconstruction. Specifically, it can be implemented using a hybrid architecture of rule engine and neural network model to transform abstract strategies into executable corpus correction operations.

[0097] Specifically, this technical solution generates reference correction information through cross-analysis of competitor corpora and historical campaign performance. For example, it extracts the semantic feature distribution and sentiment parameters of competitor corpora, combining them with conversion decay trends and user feedback characteristics from historical campaign performance to form a basis for correction based on market trends and user preferences. Simultaneously, it determines style enhancement coefficients and conversion induction intensity based on the target campaign strategy. For instance, it adjusts the enhancement level of the corpus's expressive style according to style-oriented identifiers in the strategy, and generates induction intensity parameters by matching the conversion behavior characteristics of the target audience with conversion priority parameters. A strategy correction engine integrates the reference correction information and strategy correction parameters to generate a multi-dimensional correction rule set. For example, it logically combines style enhancement rules and conversion induction rules to form a composite correction logic that considers both market trends and strategic objectives. Finally, it deconstructs and reorganizes the initial campaign corpus into semantic units. For instance, it rearranges and combines core semantic units and modifying semantic units according to correction rules, and then optimizes them through grammatical dependency relationships to generate corrected corpora that conform to the strategy orientation.

[0098] Compared to related technologies, existing methods typically rely solely on preset rules or single-dimensional user profiles for corpus adjustments, lacking a comprehensive consideration of competitor data and dynamic strategies. For example, traditional systems match keywords based only on static user attributes, failing to respond to real-time market trend changes, and the separation of strategy parameters from the corpus generation process leads to strategy execution deviations. This solution constructs a dual correction mechanism based on competitor data and campaign strategies, achieving synergy between horizontal market comparison and vertical strategy guidance during corpus generation, thus resolving the problem of insufficient strategy matching caused by static rules.

[0099] Through the above technical solution, this application can improve the dynamic adaptability of the target corpus to market trends and targeting strategies. By analyzing competitor corpus data to avoid inefficient targeting patterns and learning from successful experiences, and by combining strategy parameters to transform abstract targeting goals into actionable corpus correction rules, the generated corpus not only conforms to the characteristics of the current market environment but also accurately matches the preset targeting strategy, thereby effectively increasing the probability of triggering conversion behavior of the target audience.

[0100] This application further proposes a technical solution to generate reference correction information based on competitor corpus and historical campaign performance in the campaign reference information, determine strategy correction parameters, and perform semantic reconstruction through a strategy correction engine.

[0101] The semantic feature distribution refers to the distribution characteristics of high-frequency semantic clusters in the competitor's corpus. This can be achieved through cross-platform semantic alignment and attention mechanisms, used to capture the expressive patterns of successful industry cases. The sentiment tendency parameter refers to the implicit sentiment tendency weights in the competitor's corpus, which can be quantified using sentiment analysis models, used to identify user preference trends. The conversion decay trend refers to the pattern of conversion rate decline over time in historical campaign performance, which can be analyzed using time series analysis algorithms, used to identify corpus failure points. User feedback features refer to the user evaluation data features of historical campaign corpus, which can be mined using natural language processing techniques, used to reveal semantic expression pain points. Style guidance identifiers refer to the preset corpus style type markings in the campaign strategy, which can be parsed using strategy configuration files, used to ensure that the corpus expression is consistent with the brand tone. The conversion priority parameter refers to the conversion target weight value set in the campaign strategy, which can be parsed using strategy engine parameters, used to control the induction intensity. Conversion behavior features refer to the conversion path sensitivity exhibited by the target audience in historical behavior, which can be calculated using time series pattern mining algorithms, used to personalize the adjustment of induction intensity. A grammatical dependency graph is a graph of grammatical structural relationships between semantic units. It can be constructed using dependency parsing tools and is used to detect and repair grammatical structural breaks.

[0102] Specifically, competitor corpora undergo cross-platform semantic alignment to generate unified semantic encoding. High-frequency semantic clusters and their sentiment weights are extracted using an attention mechanism, forming semantic feature distribution and sentiment parameters. Historical campaign performance data is analyzed over time to identify conversion decay trends. Combined with feature mining results from user feedback text, reference correction information is constructed. Style-oriented identifiers in the target campaign strategy are converted into style enhancement coefficients. Conversion priority parameters are combined with the conversion path sensitivity of the target audience to calculate conversion induction intensity. The strategy correction engine deconstructs the initial campaign corpus into a set of independently adjustable semantic units. Based on style enhancement rules, it reconstructs and modifies these semantic units to inject style-adaptive components, and adjusts the expressive strength of core semantic units based on conversion induction rules. The reconstructed semantic unit set detects dependency conflicts through a grammatical dependency graph and is then repaired by a grammatical rule engine to generate grammatically coherent corrected corpus.

[0103] Compared to related technologies, traditional methods rely on manually preset templates to adjust the style of the corpus, failing to dynamically adapt to the expression patterns of competitors and historical performance data. Related technologies lack correlation analysis between conversion behavior characteristics and strategy parameters, resulting in a mismatch between the induction intensity and the actual user conversion path. Existing semantic reconstruction processes lack fine-grained unit deconstruction capabilities, making it difficult to achieve coordinated adjustment between preserving core information and optimizing modifying components.

[0104] Through the aforementioned technical solutions, this application achieves dynamic optimization of the delivery corpus in terms of style expression and conversion induction, resolving the issue of insufficient corpus appeal caused by a lack of strategic guidance. The conflict detection and repair mechanism using a grammatical dependency graph ensures the platform compliance and fluency of the semantically reconstructed corpus. The fusion analysis of competitor semantic features and historical performance data enhances the corpus's adaptability to industry trends and operational pain points.

[0105] This application further proposes extracting semantic feature distribution and sentiment parameters from competitor corpora, including cross-platform semantic alignment of competitor corpora to generate unified semantic encoding, extracting high-frequency semantic clusters and sentiment weights based on attention mechanisms, and mapping the distribution features of high-frequency semantic clusters with sentiment weights to obtain semantic feature distribution and sentiment parameters; reorganizing core semantic units and modifying semantic units in the initial semantic unit set based on a multi-dimensional correction rule set, including parsing style enhancement rules and transformation induction rules, reconstructing modifying semantic units and injecting style-adaptive modifying components, and fusing the reconstructed core semantic units to obtain the target semantic unit set; optimizing the grammatical coherence of the target semantic unit set, including constructing a grammatical dependency graph, detecting and repairing dependency conflicts, and converting it into the first corrected delivery corpus.

[0106] Cross-platform semantic alignment refers to converting competitor corpora from different advertising platforms into a unified encoding format. This can be achieved using a multimodal semantic encoder, eliminating feature extraction biases caused by differences in corpus structure across platforms. High-frequency semantic clusters refer to recurring core semantic segments in competitor corpora. This can be achieved using multi-head attention weight distribution in attention mechanisms, extracting and focusing on the marketing priorities of competitor corpora. Sentiment weights refer to the implicit emotional guidance strength of semantic segments, which can be achieved using the probability distribution output by a sentiment classification model. Sentiment weights identify the emotional inducement strategies of competitor corpora. Style-adaptive modifiers refer to corpus modifiers adapted to the target advertising strategy, which can be achieved using a set of modifiers generated by a style transfer model. Injecting style-adaptive modifiers enhances the scenario adaptability of the corpus's expression. A grammatical dependency graph refers to the grammatical structural relationship network between semantic units, which can be constructed using a dependency parser. The grammatical dependency graph detects and repairs grammatical logical conflicts between semantic units.

[0107] Specifically, the competitor corpus is first input into a cross-platform semantic alignment module, where a multimodal semantic encoder converts heterogeneous corpora from different platforms into a unified semantic code. Under the attention mechanism, the encoded semantic features are decomposed into multiple semantic clusters, with high-frequency semantic clusters and their corresponding sentiment weights automatically filtered through a multi-head attention layer. The mapping between semantic feature distribution and sentiment parameters is completed through a cross-attention mechanism, forming a competitor analysis benchmark that includes semantic topic distribution and sentiment orientation strength. Under the multi-dimensional correction rule set, the initial semantic units are split into core semantic units and modifier semantic units. Style enhancement rules drive the modifier semantic units to add style elements that fit the target audience, such as injecting popular internet slang modifiers for younger users. Conversion induction rules drive the core semantic units to reconstruct the expression of product selling points, such as converting price advantage statements into limited-time offer prompts. After the reconstructed semantic units are constructed using a grammatical dependency graph, grammatical conflicts such as subject-verb agreement and modifier positions are detected through dependency arcs. A dependency tree structure adjustment algorithm automatically corrects grammatical errors, ultimately generating the first corrected targeting corpus that conforms to language norms.

[0108] In some specific implementations, cross-platform semantic alignment can be achieved using a Transformer-based encoder, for example, uniformly encoding short copy from e-commerce platforms and long text from social media into a 768-dimensional vector. High-frequency semantic cluster extraction can employ a Top-K attention weight selection mechanism, for example, selecting the top 10% of semantic segments by attention weight as high-frequency semantic clusters. Style-adaptive modifier injection can utilize a contrastive learning-based style transfer model, for example, generating modifiers containing specific popular phrases based on the "trendy style" identifier in the advertising strategy. Grammatical dependency conflict resolution can employ a graph-based neural network model, for example, detecting and adjusting logically contradictory dependency arcs using a graph convolutional network.

[0109] Compared to related technologies, existing competitor corpus analysis methods typically extract keyword frequency or sentiment polarity directly, failing to address feature extraction biases caused by cross-platform corpus structural differences and lacking correlation analysis between semantic distribution and sentiment parameters. This solution extracts high-frequency semantic clusters through cross-platform semantic alignment and attention mechanisms, effectively eliminating platform-specific interference and focusing on core marketing elements. Existing semantic reconstruction methods often use single rules to adjust the corpus, making it difficult to collaboratively optimize style adaptation and conversion guidance. This solution separates core semantic units and modifying semantic units, applying conversion-inducing rules and style-enhancing rules respectively, achieving a balance between the completeness of marketing information and the flexibility of expression. Existing grammar correction methods rely on fixed template matching, unable to handle complex semantic structures. This solution uses dynamic conflict detection and repair based on grammatical dependency graphs, significantly improving the linguistic standardization of the generated corpus.

[0110] Through the above technical solutions, this application solves the problem of feature extraction bias caused by cross-platform structural differences in competitor corpus analysis, improves the accuracy of competitor strategy identification through semantic alignment and attention mechanisms, optimizes the collaborative mechanism of style adaptation and conversion guidance in the semantic reconstruction process, achieves dual improvement in marketing effectiveness and expression form through the separation of core and modification units, and improves the grammatical coherence of corpus generation by dynamically repairing and eliminating grammatical errors that are difficult to cover by manual rules through dependency graphs, thereby improving the compliance and attractiveness of the delivered corpus.

[0111] This application further proposes the following technical solutions, see [link / reference] Figure 6 Taking the server as the executing entity as an example, the following steps are included.

[0112] 601. Analyze the rules of the advertising platform to generate a set of platform compliance constraints and platform expression characteristics. The platform expression characteristics include the distribution of platform idiomatic expressions and user interaction pattern characteristics.

[0113] 602. Using the target large language model, the first revised delivery corpus is filtered for compliance based on the platform's compliance constraint set to obtain the second revised delivery corpus;

[0114] 603. Input the second revised delivery corpus and platform expression features into the target large language model. The target large language model encodes the second revised delivery corpus and platform expression features and performs multiple rounds of iterative decoding based on the attention mechanism to obtain the target delivery corpus.

[0115] The platform compliance constraint set refers to the set of mandatory content specifications extracted from the platform's rules. This can be achieved by using natural language processing technology to analyze keyword filtering lists and sensitive topic restrictions in the platform's rule documents, ensuring that the delivered corpus does not violate the platform's content security policy. The platform's expressive features refer to the language usage habits and interaction patterns of platform users. This can be achieved by analyzing the distribution of high-frequency words, sentence structure features, and user comment interaction patterns in the platform's historical corpus, guiding the generation of corpus that aligns with the expressive preferences of platform users. Multi-round iterative decoding refers to the encoding-decoding process based on an attention mechanism, repeatedly adjusting the corpus structure. This can be achieved by using the sequence generation capabilities of a large language model to semantically reconstruct intermediate corpora, progressively optimizing the match between the corpus and the platform's expressive features.

[0116] Specifically, the platform rule parsing phase breaks down the platform rules into two dimensions: compliance constraints and expressive features. The compliance filtering phase uses a large language model to replace keywords or remove sensitive content from the first revised delivery corpus; for example, when prohibited words are detected, they are automatically replaced with compliant synonyms. During multi-round iterative decoding, the distribution of platform idioms is encoded as attention weights, guiding the model to prioritize the use of commonly used platform vocabulary. User interaction pattern features are adjusted through the decoder's feedback mechanism to optimize the density and position of interactive guidance statements in the corpus. The encoding phase establishes a mapping between corpus fragments and platform features, while the decoding phase gradually optimizes the platform adaptability of the corpus through multi-round self-attention calculations; for example, increasing the proportion of promotional countdown statements on e-commerce platforms and enhancing the density of topic-based interactive questions on social platforms.

[0117] Compared to related technologies, traditional methods rely solely on keyword matching for platform compliance checks, failing to dynamically adapt to the stylistic differences across platforms. Template-based methods in related technologies struggle to capture user interaction characteristics, leading to discrepancies between the generated corpus and platform user behavior patterns. This solution leverages the semantic understanding capabilities of a large language model to deeply integrate platform-specific language expression patterns while ensuring compliance, thus resolving the adaptability issue during cross-platform deployment.

[0118] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0119] Through the above technical solution, this application can simultaneously meet the dual requirements of platform content security regulations and user interaction characteristics. The generated delivery corpus avoids being blocked due to violations of platform rules and can enhance the interactive effect by using expressions that conform to the habits of users on the target platform. In e-commerce platform scenarios, it can automatically generate corpus containing limited-time discount prompts and conforming to the platform's promotional language specifications; in social media scenarios, it can generate promotional text with topic tag interaction mechanisms and conforming to community content management rules, thereby improving the click-through rate of the target audience.

[0120] Figure 7 This is a schematic diagram of the structure of a dynamic, contextualized delivery corpus generation system based on a large model, provided in an embodiment of this application. See also... Figure 7 The system includes:

[0121] The determination module 701 is used to respond to the request to generate delivery corpus for the target object, determine the target delivery scenario, target delivery strategy and target product corresponding to the target object, and obtain the object information and delivery reference information of the target object. The target delivery scenario is the scenario when the target object triggers the request to generate delivery corpus.

[0122] The generation module 702 is used to generate the initial delivery corpus corresponding to the target object based on the target delivery scenario of the target object, the object information of the target object, and the target product.

[0123] The correction module 703 is used to correct the initial delivery corpus based on the delivery reference information of the target object and the target delivery strategy to obtain the first corrected delivery corpus;

[0124] The correction module 703 is also used to correct the first corrected delivery corpus based on the delivery platform rules carried in the delivery corpus generation request through the target large language model, so as to obtain the target delivery corpus corresponding to the target object. The target delivery corpus is used to attract the target object to click on the product display control corresponding to the target product.

[0125] It should be noted that the above-described embodiments of the dynamic scenario-based delivery corpus generation system based on a large model are only illustrative of the division of the functional modules described above. In practical applications, the functions described above can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the dynamic scenario-based delivery corpus generation system based on a large model and the method embodiment based on a large model belong to the same concept, and their specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0126] The technical solution provided in this application can capture dynamic user scenarios in real time to generate adaptive corpora, avoiding response delays caused by static profiles; it drives corpus correction through a strategy engine, improving the accuracy and flexibility of conversion inducement; and it utilizes a large language model to internalize platform rules, solving the problem of repetitive adaptation in cross-platform advertising. This allows the generated advertising corpora to closely match users' real-time needs, maximizing conversion effects across different scenarios and platforms while ensuring compliance.

[0127] Figure 8 This is a schematic diagram of a server structure provided in an embodiment of this application. The server 800 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 801 and one or more memories 802. The one or more memories 802 store at least one computer program, which is loaded and executed by the one or more processors 801 to implement the methods provided in the various method embodiments described above. Of course, the server 800 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server 800 may also include other components for implementing device functions, which will not be elaborated upon here.

[0128] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a computer program that can be executed by a processor to perform the dynamic scene-based delivery corpus generation method based on a large model in the above embodiments. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0129] In an exemplary embodiment, a computer program product or computer program is also provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to perform the above-described method for generating dynamic contextualized delivery corpus based on a large model.

[0130] In some embodiments, the computer program involved in the present application embodiments may be deployed and executed on a computer device, or executed on multiple computer devices located in one location, or executed on multiple computer devices distributed in multiple locations and interconnected through a communication network. Multiple computer devices distributed in multiple locations and interconnected through a communication network may constitute a blockchain system.

[0131] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0132] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating dynamic, contextualized delivery corpus based on a large model, characterized in that, The method includes: In response to a request to generate corpus data for a target object, the system acquires the following information: corpus generation request description information, delivery method description information, delivery requirement description information, and corpus delivery records for the target object. Based on the delivery method description information, delivery requirement description information, and delivery requirement description information, a target delivery scenario is determined, whereby the target object triggers the corpus generation request. Based on the corpus delivery records, the delivery stage for the target object is determined, whereby the delivery stage is the stage in which corpus data is delivered to the target object. Based on the delivery requirement description information, the desired delivery effect and delivery requirements are determined. Based on the delivery stage and the desired delivery effect, a reference delivery strategy is determined from multiple candidate delivery strategies, whereby the reference delivery strategy is a candidate delivery strategy that matches the delivery stage and the desired delivery effect. The reference delivery strategy is adjusted using the delivery requirements to obtain the target delivery strategy corresponding to the target object. Identify the target product corresponding to the target object, and obtain the object information and deployment reference information of the target object; Based on the target delivery scenario of the target object, the object information of the target object, and the target product, generate the initial delivery corpus corresponding to the target object; Based on the target object's delivery reference information and the target delivery strategy, the initial delivery corpus is modified to obtain the first modified delivery corpus; By using the target large language model, the first modified delivery corpus is modified based on the delivery platform rules carried in the delivery corpus generation request to obtain the target delivery corpus corresponding to the target object. The target delivery corpus is used to attract the target object to click on the product display control corresponding to the target product.

2. The method according to claim 1, characterized in that, The object information includes historical behavioral data, a set of interest preference tags, and demographic attributes. The process of generating initial delivery corpus corresponding to the target object based on the target delivery scenario, the object information of the target object, and the target product includes: Time-series pattern mining is performed on the historical behavior data to generate behavioral features of the target object, including behavior frequency distribution, content theme preference, and conversion path sensitivity. The association strength of each interest preference tag in the interest preference tag set is determined. The association strength is determined based on the tag co-occurrence frequency and the timeliness weight. The tag co-occurrence frequency refers to the frequency with which each interest preference tag in the interest preference tag set appears together with other interest preference tags within a preset time window. The timeliness weight is used to represent the degree of effectiveness decay of each interest preference tag in the interest preference tag set over time. Based on the behavioral characteristics, the set of interest preference tags, the tag association strength, and the demographic attributes, a profile of the target object is determined. The initial delivery corpus is generated based on the product information of the target product, the target delivery scenario, and the object profile.

3. The method according to claim 2, characterized in that, The step of performing time-series pattern mining on the historical behavior data to generate behavioral features of the target object includes: Identify dynamic behavior patterns and key behavior sequences in the historical behavior data. The dynamic behavior patterns refer to a set of behavioral patterns with time evolution characteristics and scene relevance. The key behavior sequences refer to continuous behavioral segments that are strongly correlated with the expected delivery effect corresponding to the delivery corpus generation request. Based on the dynamic behavior patterns and the key behavior sequences, determine the behavior frequency distribution, the content theme preference, and the conversion path sensitivity. The conversion path sensitivity is determined based on the conversion rate of the key behavior sequences. The process of determining the object profile of the target object based on the behavioral characteristics, the set of interest preference tags, the tag association strength, and the demographic attributes includes: Based on multiple interest preference tags in the interest preference tag set whose tag association strength meets preset conditions, an interest topology network is constructed; based on the behavioral characteristics and the demographic attributes, the baseline characteristics of the target object are determined; based on the interest topology network and the baseline characteristics, the object profile is generated. The process of generating the initial delivery corpus based on the product information of the target product, the target delivery scenario, and the object profile includes: A semantic generation framework is determined based on the target delivery scenario; the product information and the object profile are injected into the semantic generation framework to obtain corpus generation information; and the initial delivery corpus is generated based on the corpus generation information.

4. The method according to claim 3, characterized in that, The process of generating the object profile based on the interest topology network and the baseline features includes: The reference features are used to perform matching in the interest topology network to obtain multiple target nodes in the interest topology network. The target nodes are nodes whose matching degree with the reference features is greater than or equal to a preset matching degree. Based on the node information of the multiple target nodes and the reference features, the object profile is generated. The step of generating the initial delivery corpus based on the corpus information includes: The corpus generation information is parsed to obtain a core interest description, a conversion inducement instruction, and scene modification elements. The core interest description is a structured field representing the core interest preferences of the target object. The conversion inducement instruction is a set of command parameters used to guide the target object to perform conversion behavior. The scene modification elements are a set of semantic modifiers adapted to the target delivery scenario. According to the corpus generation rules of the target delivery scenario, the core interest description, the conversion inducement instruction, and the scene modification elements are combined into a semantic structure tree. The semantic structure tree is then converted into the initial delivery corpus using a pre-trained scene adaptation generator.

5. The method according to claim 1, characterized in that, The process of modifying the initial delivery corpus based on the delivery reference information of the target object and the target delivery strategy to obtain a first modified delivery corpus includes: Based on the competitor corpus and historical campaign performance in the campaign reference information, reference correction information is generated, wherein the competitor corpus is the campaign corpus used by the competitors of the target product; Based on the target delivery strategy, strategy adjustment parameters are determined, including style enhancement coefficient and conversion induction intensity. The reference correction information and policy correction parameters are input into the policy correction engine to generate a multi-dimensional correction rule set; The strategy correction engine performs semantic reconstruction on the initial delivery corpus according to the multi-dimensional correction rule set, and outputs the first corrected delivery corpus.

6. The method according to claim 5, characterized in that, The step of generating reference correction information based on competitor corpus and historical campaign performance in the campaign reference information includes: Extract the semantic feature distribution and sentiment parameters of the competitor's corpus; analyze the conversion decay trend and user feedback features in the historical campaign performance; and fuse the semantic feature distribution, sentiment parameters, conversion decay trend, and user feedback features to obtain the reference correction information. The strategy adjustment parameters are determined based on the target delivery strategy. These parameters include a style enhancement coefficient and a conversion induction intensity, comprising: The style guidance identifier and conversion priority parameter in the target delivery strategy are determined; the style reinforcement coefficient is determined based on the style guidance identifier; the conversion induction intensity is determined based on the conversion priority parameter and the conversion behavior characteristics of the target object, wherein the conversion behavior characteristics are determined based on the historical behavior data included in the object information; The step of using the strategy correction engine to semantically reconstruct the initial delivery corpus according to the multi-dimensional correction rule set, and outputting the first corrected delivery corpus, includes: The strategy correction engine deconstructs the initial delivery corpus into an initial semantic unit set; based on the multi-dimensional correction rule set, the core semantic units and modifying semantic units in the initial semantic unit set are reorganized to obtain a target semantic unit set; the target semantic unit set is then optimized for grammatical coherence to obtain the first corrected delivery corpus.

7. The method according to claim 6, characterized in that, The extraction of semantic feature distribution and sentiment parameters from the competitor's corpus includes: Cross-platform semantic alignment is performed on the competitor's corpus to generate a unified semantic code; high-frequency semantic clusters and corresponding sentiment tendency weights are extracted from the unified semantic code based on an attention mechanism; the distribution features of the high-frequency semantic clusters are associated and mapped with the sentiment tendency weights to obtain the semantic feature distribution and sentiment tendency parameters. The core semantic units and modifying semantic units in the initial semantic unit set are reorganized based on the multi-dimensional correction rule set to obtain the target semantic unit set, including: The style enhancement rules and transformation induction rules in the multi-dimensional correction rule set are analyzed; the modification semantic units are reconstructed based on the style enhancement rules to inject style-adaptive modification components; the core semantic units are reconstructed based on the transformation induction rules; and the reconstructed modification semantic units and core semantic units are merged to obtain the target semantic unit set. The step of optimizing the syntactic coherence of the target semantic unit set to obtain the first modified delivery corpus includes: Construct a grammatical dependency graph of the target semantic unit set; detect and repair dependency conflicts in the grammatical dependency graph using a grammar rule engine; and convert the repaired target semantic unit set into the first corrected delivery corpus.

8. A dynamic, contextualized delivery corpus generation system based on a large model, characterized in that, The system includes: A determination module is used to respond to a request to generate corpus data for a target object, and to obtain the delivery channel description information, delivery method description information, delivery requirement description information, and corpus delivery records of the target object corresponding to the corpus data generation request; based on the delivery channel description information, the delivery method description information, and the delivery requirement description information, to determine a target delivery scenario, wherein the target delivery scenario is the scenario in which the target object triggers the corpus data generation request; based on the corpus delivery records, to determine the delivery stage of the target object, wherein the delivery stage is the stage in which corpus data is delivered to the target object; based on the delivery requirement description information, to determine the expected delivery effect and delivery requirements; based on the delivery stage and the expected delivery effect, to determine a reference delivery strategy from multiple candidate delivery strategies, wherein the reference delivery strategy is a candidate delivery strategy that matches the delivery stage and the expected delivery effect; to adjust the reference delivery strategy according to the delivery requirements to obtain the target delivery strategy corresponding to the target object; to determine the target product corresponding to the target object, and to obtain the object information and delivery reference information of the target object; The generation module is used to generate the initial delivery corpus corresponding to the target object based on the target delivery scenario of the target object, the object information of the target object, and the target product. The correction module is used to correct the initial delivery corpus based on the delivery reference information of the target object and the target delivery strategy to obtain the first corrected delivery corpus; The correction module is further configured to correct the first corrected delivery corpus based on the delivery platform rules carried in the delivery corpus generation request using the target large language model, thereby obtaining the target delivery corpus corresponding to the target object. The target delivery corpus is used to attract the target object to click on the product display control corresponding to the target product.

Images