Method and device for generating content distribution scheme based on large model

Abstract

The invention relates to a method for generating a content distribution scheme based on a large model, and the method comprises the steps: receiving a natural language instruction related to the content distribution scheme, enabling a pre-trained large model to analyze the natural language instruction to extract control parameters, and enabling the control parameters to at least comprise a scheme theme and resource scheduling limitation; analyzing semantic features of the scheme theme, and establishing a semantic mapping relationship between the semantic features and the attribute tags of the candidate objects to screen and obtain a target object set matched with the scheme theme; analyzing behavior interest characteristics of each object in the target object set, and constructing content distribution data for the target object set in combination with a scheme theme and resource scheduling limitation; and generating a content distribution scheme based on the target object set and the content distribution data, and executing a pushing task according to the content distribution scheme. According to the method and the device, a large-model-based content matching and distribution mechanism is used for replacing a tedious manual decision, so that human resources are effectively released, and the task execution efficiency is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus, electronic device, and computer program product for generating a content distribution scheme based on a large model. Background Technology

[0002] With the rapid development of mobile internet and big data technologies, various internet platforms have accumulated massive amounts of information resources. To accurately deliver suitable information to target users, content distribution technology has emerged. Through efficient content distribution, platforms can push information in a targeted manner based on user characteristics. This not only improves the efficiency of information dissemination but also becomes a key link connecting platform content with user needs, and is widely used in scenarios such as news pushes and event notifications.

[0003] However, existing content distribution technologies still suffer from insufficient automation in practical applications. The process of building a content distribution solution involves crucial steps such as material preparation and distribution strategy formulation, which typically rely on human experience for parameter configuration and process operation. This highly manual approach not only results in a long generation cycle for distribution solutions but also fails to meet the current application scenarios' demands for high-frequency and timely content distribution. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus, electronic device, and computer program product for generating a content distribution scheme based on a large model, used to solve at least one technical problem.

[0005] This application provides a method for generating a content distribution scheme based on a large model, comprising: receiving a natural language instruction regarding the content distribution scheme; parsing the natural language instruction using a pre-trained large model to extract control parameters, the control parameters including at least: scheme topic and resource scheduling constraints; parsing the semantic features of the scheme topic and establishing a semantic mapping relationship between the semantic features and the attribute tags of candidate objects to filter and obtain a set of target objects that are compatible with the scheme topic; analyzing the behavioral and interest features of each object in the set of target objects and constructing content distribution data for the set of target objects in combination with the scheme topic and resource scheduling constraints; generating a content distribution scheme based on the set of target objects and the content distribution data, and executing a push task according to the content distribution scheme.

[0006] The generation method described above, which establishes a semantic mapping relationship between the semantic features and the attribute tags of candidate objects to filter and obtain a set of target objects that are compatible with the theme of the scheme, includes: extracting multi-dimensional topic tags from the semantic features, wherein the multi-dimensional topic tags include one or more of industry tags, regional tags, and identity tags; calculating the semantic relevance between the multi-dimensional topic tags and the attribute tags of candidate objects, and determining the objects whose semantic relevance meets a preset matching threshold as the first set of target objects.

[0007] The generation method described above further includes: predicting the click probability of each object on the topic of the scheme based on the historical behavior data of each object in the first target object set; sorting the objects in the first target object set based on the click probability, and determining the objects with a click probability higher than a preset probability threshold as the second target object set, and using the second target object set as the final target object set.

[0008] The generation method described above analyzes the behavioral interest characteristics of each object in the target object set, and combines the scheme theme and resource scheduling constraints to construct content distribution data for the target object set. This includes: determining the behavioral preference type of each object based on its historical behavior logs, whereby the behavioral preference type includes a preference for industry information and a preference for service entities; and, under the premise of satisfying the resource scheduling constraints, performing differentiated content processing based on the behavioral preference type and scheme theme of each object to construct corresponding content distribution data.

[0009] As described above, the differentiated content processing based on the behavioral preference type and solution theme of each object to construct corresponding content distribution data includes: if the object's behavioral preference type is industry information preference and its historical behavior logs are identified as being concentrated in specific industry tags, then industry dynamic copy is generated based on the solution theme and specific industry tags, and the industry dynamic copy is used as content distribution data; if the object's behavioral preference type is service entity preference and its historical behavior logs are identified as being concentrated in specific business attributes, then target service entities that match the solution theme and specific business attributes are retrieved, and the target service entities are encapsulated as distribution landing pages as content distribution data.

[0010] As described above, the step of executing the push task according to the content distribution scheme includes: extracting the historical active logs of each object in the target object set, and performing time-domain analysis on the historical active logs to extract the active time period characteristics of each object; determining the high-frequency active time window of each object based on the active time period characteristics, and mapping the trigger time of the push task for each object to the high-frequency active time window.

[0011] The generation method described above further includes: during the execution of the push task, collecting the actual feedback indicators of the task and calculating the deviation value between the actual feedback indicators and the estimated indicators; if the deviation value exceeds a preset abnormal threshold, then executing a dynamic optimization process for the content distribution scheme; the dynamic optimization process includes at least one of the following: extracting high-response feature vectors based on objects that generate positive interactive behavior in the executed tasks, performing similarity expansion on the target object set based on the high-response feature vectors, and generating supplementary push tasks for the expanded target object set; identifying silent objects that generate exposure behavior but do not generate deep interactive behavior in the executed tasks, regenerating secondary content distribution data that differs from the original content distribution data for the silent objects, and pushing it to the silent objects.

[0012] According to another aspect of this application, a device for generating a content distribution scheme based on a large model is proposed, comprising: a receiving module for receiving natural language instructions regarding the content distribution scheme, wherein a pre-trained large model parses the natural language instructions to extract control parameters, the control parameters including at least: scheme topic and resource scheduling constraints; a filtering module for parsing the semantic features of the scheme topic and establishing a semantic mapping relationship between the semantic features and the attribute tags of candidate objects to filter and obtain a set of target objects that are compatible with the scheme topic; a content distribution construction module for analyzing the behavioral and interest features of each object in the set of target objects and, in conjunction with the scheme topic and resource scheduling constraints, constructing content distribution data for the set of target objects; and a content distribution scheme generation module for generating a content distribution scheme based on the set of target objects and the content distribution data, and executing a push task according to the content distribution scheme.

[0013] According to another aspect of this application, an electronic device is proposed, including a processor and a memory, wherein a set of computer program instructions is stored on the memory, and the processor executes the set of computer program instructions on the memory to implement the generation method of the content distribution scheme based on the large model as described above.

[0014] According to another aspect of this application, a computer program product is proposed, which includes a set of computer program instructions that, when executed by a processor, implement the generation method of the content distribution scheme based on the large model as described above.

[0015] This application, through an automated content matching and distribution mechanism, effectively frees up the human resources of business personnel, freeing them from the need to invest significant effort in audience screening, content organization, and distribution strategy formulation, thus significantly improving the efficiency of daily distribution tasks. Simultaneously, it eliminates the need for repetitive development of customized tags or specific content for different scenarios, greatly reducing R&D and maintenance costs and saving substantial technical investment. Attached Figure Description

[0016] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for generating a content distribution scheme based on a large model according to an embodiment of this application.

[0017] Figure 2 This is a flowchart of the method for filtering the target object set in step S120.

[0018] Figure 3 This is a flowchart of the method for constructing content distribution data in step S130.

[0019] Figure 4 This is a schematic diagram illustrating the process of developing a campus recruitment marketing plan based on a large model, according to one embodiment of this application.

[0020] Figure 5 This is a schematic diagram of a content distribution data push interface according to an embodiment of this application.

[0021] Figure 6 This is a schematic diagram of the generation device for a content distribution scheme based on a large model according to an embodiment of this application.

[0022] Figure 7 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In the following detailed description, reference can be made to the accompanying drawings, which form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Specific embodiments of the present application are described in sufficient detail below to enable those skilled in the art to implement the technical solutions of the present application. It should be understood that other embodiments may also be utilized, or structural, logical, or electrical changes may be made to the embodiments of the present application.

[0025] The large-scale model in this application refers to an artificial intelligence model with a massive number of parameters, based on a deep learning architecture (such as deep neural networks, multi-layer neural networks, or convolutional neural networks). This type of model, through pre-training on massive amounts of multi-dimensional data, can learn the high-dimensional feature representations and complex logical relationships inherent in the data, thereby possessing powerful generalization and information processing capabilities. In the technical solution of this application, the large-scale model can be used for deep feature extraction and semantic understanding of massive amounts of material, user behavior, and tag data in content distribution scenarios.

[0026] The content distribution solution described in this application refers to a pre-built task configuration executable by computing devices to achieve targeted delivery of specific content to a target user group. This content distribution solution constrains the distribution recipients, conditions, timing, and methods of the content, enabling the content to reach and be delivered according to the intended strategy. The content distribution solution can be applied to marketing campaigns, advertising promotions, or other information dissemination scenarios.

[0027] Figure 1 This is a flowchart illustrating a method for generating a content distribution scheme based on a large model according to an embodiment of this application. Figure 1 As shown, the method includes: S110, Receive natural language instructions about content distribution scheme, and a pre-trained large model parses the natural language instructions to extract control parameters, the control parameters including at least: scheme topic and resource scheduling constraints; S120, parse the semantic features of the scheme theme, and establish a semantic mapping relationship between the semantic features and the attribute tags of the candidate objects, so as to filter out a set of target objects that are compatible with the scheme theme; S130, Analyze the behavioral and interest characteristics of each object in the target object set, and combine the scheme theme and resource scheduling constraints to construct content distribution data for the target object set; S140, Generate a content distribution scheme based on the target object set and content distribution data, and execute the push task according to the content distribution scheme.

[0028] In step S110, the natural language instruction refers to the unstructured business requirements input by the staff, such as: "Starting an online campus recruitment campaign on September 1st, primarily targeting campus recruiters and recent graduates, with a goal of achieving 100,000 daily unique visitors (UV) on the campaign page." The pre-trained large-scale model parses this natural language instruction to extract control parameters. Taking the campus recruitment scenario as an example, the solution theme is parsed as "job and company recommendations for recent graduates"; resource scheduling constraints cover the time frame (September to October), budget cost (e.g., 3 cents / message with a daily budget of 1000 yuan), and reach frequency (e.g., push notifications limited to 2 times per person per day). Through the intent recognition of the large-scale model, complex business objectives can be automatically broken down into underlying executable constraints, significantly improving the efficiency of solution initialization.

[0029] In step S120, candidate objects refer to all users stored within the platform, and attribute tags can be static dimensions of the users (such as education level, major, graduation time). In this embodiment, by extracting semantic features from the theme "campus recruitment," the mapping logic automatically associates users with attribute tags such as "2026 graduates," "interns," and "current students," thereby quickly identifying a set of target objects highly relevant to this recruitment activity from a user database of tens of millions. This semantic-based matching method, rather than a single hard rule, improves the generalization and accuracy of the target audience coverage.

[0030] In step S130, behavioral interest characteristics refer to the user's dynamic interaction records within the platform, such as recently searching for 'computer vision algorithm position' or tending to click on positions marked "HR online." In constructing content distribution data, this application embodiment intelligently combines content from different channels while considering resource scheduling constraints. For example, for highly active users, push-format job card distribution data is constructed; for less active but matching users, SMS notification content is constructed within budget. By coupling user interests, goal achievement paths, and cost constraints in a multi-dimensional way, it ensures that the generated distribution data both conforms to user preferences and does not exceed resource capacity.

[0031] In step S140, the generated content distribution plan is a content combination strategy that includes audience segments, delivery channels, creative content, and execution timing. Executing push tasks according to this plan ensures that content most aligned with the target audience's interests reaches users at the most appropriate time, while meeting preset resource constraints.

[0032] This application, through an automated content matching and distribution mechanism, effectively frees up the human resources of business personnel, freeing them from the need to invest significant effort in audience screening, content organization, and distribution strategy formulation, thus significantly improving the execution efficiency of daily distribution tasks. Simultaneously, this application eliminates the need for repetitive development of customized tags or specific content for different scenarios, greatly reducing system development and maintenance costs and saving substantial technical investment.

[0033] Figure 2 This is the flowchart for step S120, which involves filtering the set of target objects. Figure 2 As shown, the method includes: S121, extract multi-dimensional topic tags from the semantic features, the multi-dimensional topic tags including one or more of industry tags, regional tags and identity tags; S122, calculate the semantic relevance between the multidimensional topic tags and the attribute tags of the candidate objects, and determine the objects whose semantic relevance meets the preset matching threshold as the first target object set; S123, Based on the historical behavior data of each object in the first target object set, predict the click probability of each object on the topic of the solution; S124, sort the objects in the first target object set based on the click probability, and determine the objects with a click probability higher than a preset probability threshold as the second target object set, and use the second target object set as the final target object set.

[0034] In step S121, the subject tags constitute the core dimensions describing the business scenario, typically encompassing industry tags, regional tags, and identity tags. For example, when the solution topic involves "a major internet company's campus recruitment activities in Shanghai," semantic analysis can map it to the industry tag "Internet / Artificial Intelligence," the regional tag "Shanghai / East China," and the identity tag "recent graduate / student at this university." Through this tagging process, unstructured natural language instructions are transformed into structured feature vectors that can be processed by computers, laying the data foundation for subsequent user matching.

[0035] In step S122, by calculating the semantic relevance between the aforementioned multi-dimensional topic tags and the attribute tags of each object in the candidate object library, the degree of fit between the candidate objects and the current solution topic can be quickly measured. This process can utilize vector space distance calculation or semantic alignment algorithms to classify objects whose relevance meets a preset matching threshold into the first target object set. This step achieves accurate audience recall, significantly reduces the data scale of subsequent processing, thereby reducing the ineffective consumption of computing resources, while ensuring the basic relevance of the selected audience in terms of business attributes.

[0036] In step S123, to further enhance the accuracy of target object screening, this application introduces a deep prediction mechanism based on historical behavioral data. By analyzing the interaction logs of objects within a historical time window, such as clickstream, browsing duration, and historical response rate, a pre-defined prediction algorithm is used to accurately predict the probability of each object clicking on the current topic. This moves from static attribute matching to dynamic behavioral prediction, effectively identifying users who, although their tags match, are inactive or have decreased interest in this type of content, thus focusing on truly active, high-potential users with conversion intent.

[0037] In step S124, the first target object set is sorted and truncated based on the predicted click probabilities. By setting a preset probability threshold, objects with click probabilities higher than the threshold are filtered out and identified as the second target object set, which is then used as the final target object set for subsequent push tasks. This probability-based optimization strategy ensures that limited push resources (such as SMS budget and push channel bandwidth) are concentrated on allocating to the group with the highest conversion probability, significantly improving the overall conversion rate and resource utilization efficiency of the content distribution solution while maintaining distribution accuracy.

[0038] Figure 3 This is a flowchart illustrating the method for constructing content distribution data in step S130. (Example) Figure 3 As shown, the method includes: S131, based on the historical behavior logs of each object in the target object set, determine the behavior preference type of each object, the behavior preference type including preference for industry information type and preference for service entity type; S132, Under the premise of satisfying the resource scheduling constraints, differentiated content processing is performed based on the behavioral preference type and scheme theme of each object to construct corresponding content distribution data.

[0039] In step S131, historical interaction records can be clustered to identify high-frequency content features and determine preference types. Significant differences exist among different groups: for example, white-collar workers tend to focus on industry reports or company prospects, reflecting a demand for unstructured information, thus they are labeled as preferring industry information; while blue-collar workers focus on certification training or job postings, reflecting a strong demand for service entities, thus they are labeled as preferring service entities. This step not only identifies user intent but also transforms abstract profiles into algorithmically executable strategy tags, laying a data foundation for subsequent precise distribution.

[0040] In step S132, under the premise of meeting resource constraints, content distribution data is dynamically generated based on the scheme theme and the aforementioned determined preference types. This content distribution data includes one or more of the following: job postings / resumes, copywriting, industry consulting, graphics and text, promotional activities, and landing pages. This differentiated approach effectively avoids mismatches between content and audience needs (such as pushing obscure analytics to blue-collar workers), achieving precise matching under limited resource constraints, thereby significantly improving the final conversion rate.

[0041] This application leverages in-depth analysis of the target audience's historical interaction logs to accurately identify their behavioral preference types, dynamically constructing differentiated distribution data under the same theme. This tailored strategy effectively achieves personalized content distribution.

[0042] According to one embodiment of this application, differentiated content processing is performed based on the behavioral preference type and theme of each object to construct corresponding content distribution data, including: If the object's behavioral preference type is industry information preference, and its historical behavioral logs are identified as being concentrated in specific industry tags, then the big model is used to generate industry dynamic copy based on the solution theme and specific industry tags, and the industry dynamic copy is used as content distribution data. For users who prefer industry-specific information, the processing logic focuses on generating creative content from unstructured text. When historical behavior logs show that the user frequently follows specific industry tags (such as internet, finance, etc.), the text generation capability is invoked. Centered on the theme of the campaign, it integrates industry dynamics or trend analysis related to that specific industry tag to generate highly readable industry-related copy. This approach leverages users' thirst for industry knowledge, transforming what would otherwise be rigid marketing promotion into valuable information delivery, thereby reducing user resistance and effectively increasing click-through rates.

[0043] If the object's behavior preference type is a preference service entity type, and its historical behavior logs are identified as being concentrated on specific business attributes, then the target service entity that matches the solution theme and specific business attributes is retrieved, and the target service entity is encapsulated as a distribution landing page as content distribution data.

[0044] For objects identified as having preferred service entities, the processing logic shifts to retrieval of structured data and page encapsulation. When historical behavior logs show that the object has recently been consistently browsing specific business attributes (such as a specific type of job or a specific salary range) or specific interaction patterns (such as liking online HR communication), a precise search is performed in the business database to filter out target service entities that highly match the current solution theme and that specific business attribute. These entities can be specific job cards (including job title, salary, and requirements) or real-time service status (such as the current online live broadcast status of HR).

[0045] Building upon this, the crucial step of encapsulating the target service entity into a landing page is executed. This encapsulation process essentially involves dynamically injecting discrete service entity data into a visual page template. Combining this with a standard page structure, the processing logic first generates a header visual and slogan based on the solution theme, establishing the page's tone. Then, the retrieved service entities (such as job listings or HR-focused entry points) are rendered as core interactive components. For example, if a user prefers "HR online," their avatar, online status, and communication button are encapsulated as interactive cards; if they prefer specific industry positions, relevant job information is aggregated into a list component. Finally, these components are assembled into a directly interactive page (such as an HTML5 or mini-program page). This effectively shortens the conversion path, allowing users to directly access their chosen aggregation page without needing a secondary search, thus significantly improving traffic acquisition efficiency and business conversion rates.

[0046] According to one embodiment of this application, performing a push task according to the content distribution scheme includes: Extract the historical active logs of each object in the target object set, and perform time-domain analysis on the historical active logs to extract the active time period characteristics of each object; The historical activity logs for each object meticulously record the timestamps of user interactions such as logins, clicks, and browsing over a past period. By performing time-domain analysis on this massive amount of discrete-time data, statistically significant active period characteristics can be extracted from seemingly random time points. This time-domain analysis not only focuses on single points in time but also emphasizes uncovering periodic patterns in users across daily, weekly, or monthly dimensions, thereby building time-dimensional user profiles for subsequent precise outreach.

[0047] Based on the active time period characteristics, a high-frequency active time window is determined for each object, and the trigger time of the push task for each object is mapped to the high-frequency active time window.

[0048] The above steps translate a user's potential activity patterns into specific task execution instructions. For example, if historical data analysis shows that a user habitually opens the application to browse on Monday mornings, the processing logic will lock that "Monday morning" time period as the user's high-frequency activity window. Subsequently, when executing a specific push notification task, the push notification trigger time for that user will be dynamically mapped to this high-frequency activity window.

[0049] This timing-based, differentiated triggering strategy not only smooths out peak and trough server traffic, but more importantly, it ensures that information is delivered when users are most likely to be online or most focused, thereby effectively increasing the probability of message opening and click conversion, and achieving individualized optimization of push efficiency.

[0050] To overcome the prediction bias that may occur when static pushes encounter complex and ever-changing real-time user behavior, this application introduces an automatic iteration and dynamic optimization mechanism based on feedback data to ensure that the distribution effect is always maintained at the expected level.

[0051] According to one embodiment of this application, during the execution of a push task, the actual feedback indicators of the task are collected, and the deviation between the actual feedback indicators and the estimated indicators is calculated. If the deviation value exceeds a preset abnormal threshold, a dynamic optimization process for the content distribution scheme is executed. During the execution of push notification tasks, the processing logic collects actual feedback metrics in real time according to preset time granularity (e.g., hourly iterations) or task progress nodes (e.g., every 100,000 push notifications completed). These metrics include, but are not limited to, real-time click-through rate, conversion rate, or page visit depth. Subsequently, the health of the current task is quantified by calculating the deviation between the actual feedback metrics and the model's predicted metrics.

[0052] If the calculated deviation exceeds the preset abnormal threshold, for example, if the overall real-time effect is significantly lower than the model's predicted effect, this usually means that there is a mismatch between the current audience positioning or content strategy and the actual market response. At this time, the processing logic will automatically trigger a dynamic optimization process for the content distribution plan, and adjust the remaining tasks or additional tasks in a targeted manner.

[0053] The dynamic optimization process includes at least one of the following: High-response feature vectors are extracted from objects that generate positive interactive behavior in executed tasks. The target object set is then expanded based on the high-response feature vectors, and supplementary push tasks are generated for the expanded target object set. The process involves filtering out objects that generate positive interactive behaviors (such as high-frequency clicks) from executed tasks and extracting feature vectors from these high-response objects. Based on these feature vectors, the clustering or similarity matching capabilities of a large model are used to expand the target object set into a larger candidate pool, essentially finding users similar to the high-response user profiles. Supplementary push tasks are then generated for this expanded target object set. This technical solution utilizes validated successful samples to calibrate the audience model, thereby enabling rapid scaling of high-quality traffic acquisition.

[0054] Identify silent objects that have generated exposure behavior but not deep interaction behavior in the executed tasks, regenerate secondary content distribution data that differs from the original content distribution data for the silent objects, and push it to the silent objects.

[0055] The lack of deep interaction suggests users have initial interest in the topic, but the original content failed to trigger further action. Therefore, for this group of inactive users, secondary content distribution data that differs significantly from the original data in copywriting style, visual elements, or entry point is generated and pushed to them for reactivation. In this way, differentiated content stimulation solves the content mismatch problem, effectively winning back potentially lost users and improving overall traffic retention and conversion depth.

[0056] The implementation methods and advantages of the embodiments of this application have been described above through multiple examples. The specific processing procedures of the embodiments of this application are described in detail below, using a specific example of an online campus recruitment marketing plan.

[0057] Figure 4 This is a schematic diagram illustrating the process of developing a campus recruitment marketing plan based on a large model, according to one embodiment of this application.

[0058] Before executing a marketing plan, the large model needs to be pre-trained to meet specific functionalities. To ensure that the model can generate structured solutions that conform to business logic, the model building and training process in this application mainly includes three stages: infrastructure construction, pre-training in a general language, and fine-tuning of instructions in the vertical domain.

[0059] At the model architecture level, the large model in this application can adopt a Transformer-based deep neural network architecture, whose core module includes a multi-layered stacked multi-head self-attention mechanism and a feedforward neural network. By processing the temporal features of the input sequence through a positional encoding module, the model is able to handle long text dependencies, which is crucial for understanding complex marketing instructions containing multiple constraints (such as "budget of 1000 yuan" and "lasting for 60 days").

[0060] In the training step that integrates the model with a specific marketing domain, a dedicated instruction dataset containing massive amounts of historical marketing logs was constructed. Specifically, the training process employed a supervised fine-tuning strategy to establish an intrinsic correlation between input features and output schemes. The input data was configured as a "structured Prompt," which included a description of the historical campaign's objectives (e.g., "increase UV"), resource constraints (e.g., "SMS budget"), target audience attribute tags (e.g., "C9 universities"), and the target audience's historical behavioral characteristics; the output data corresponded to the "preferred marketing strategies" and / or "high-click copy" that actually generated high conversion rates in that historical scenario.

[0061] In terms of training parameters and target settings, the model updates its parameters by minimizing the cross-entropy loss function between the predicted output and the actual high-optimal solution. To enhance the model's ability to predict performance, a regression prediction head is introduced at the end of the network, using the actual click-through rate of historical tasks as a supervisory signal for joint training. Through this setup, the model not only learns how to generate fluent marketing copy (text generation capability), but also learns how to deduce the optimal "channel combination" and "sending frequency" based on a quantitative target such as "100,000 daily UV". Ultimately, the trained model can accurately map unstructured natural language requirements into structured marketing solutions that conform to preset definitions, achieving a deep integration of the algorithm and specific business scenarios.

[0062] like Figure 4 As shown, the pre-trained large model receives user input in natural language text: "On XX month XX day, an online campus recruitment event will be launched. The theme of the event is to provide a customized online recruitment session for campus recruiters and fresh graduates, mainly focusing on job and company recommendations; it is expected to last for 30 days. The event goal is to achieve an average daily UV of XX million for the event page, with an estimated cost of XX yuan; the core target audience is fresh graduates. Please help develop a marketing outreach plan. Optional outreach channels include: push notifications (10 million messages per day, 2 messages per person), SMS (3 cents / message); email (1 million messages per day, 1 message per person), WeChat (500,000 messages per day, 1 message per person)." The large model uses NLP capabilities to extract entities and identify intents from the above text, transforming the vague business requirements into structured control parameters (including core objectives, budget boundaries, and available channel pools) as the input benchmark for subsequent inference.

[0063] Following this, the process enters the deep reasoning and strategy generation stage (i.e., the "thinking" stage in the diagram). The processing logic in the background uses a large model to perform joint analysis of control parameters and historical data. Regarding the matching of resources and objectives, the gap between the theoretical conversion rate of currently available channels (Push, SMS, email, etc.) and the target UV is calculated, and an initial model for resource allocation weights is planned. In audience analysis, semantic alignment is performed. For example, the generalized concept of "recent graduates" in the instruction is mapped to user tags in the database to identify specific target groups such as "C9 universities" and "985 liberal arts students". At the same time, the response preferences of recent graduates in the database to different channels are analyzed (such as "preferring push notifications" and "preferring WeChat"). In terms of content strategy, user preferences are differentiated based on historical behavior. It was identified that some users prefer "industry news," while others prefer "job postings."

[0064] Based on the above reasoning, the processing logic outputs an initial list of marketing plans (as shown in the table in the figure). This list structurally displays a combination of strategies for different audience segments: for example, for the "C9 University" audience (estimated at XXX million), match "consulting" content, predict a click-through rate of XX%, and suggest reaching them via "WeChat" at 10:00 AM on "XX day"; for the "985 Liberal Arts" audience, match "recruitment" content, and suggest reaching them via "Push" channels.

[0065] In addition, users can provide feedback commands for fine-tuning. As shown in the lower right corner of the image: "More job recommendations are needed," the larger model will fine-tune the strategy based on this feedback, increasing the weight of "job / service entity" content.

[0066] Finally, the confirmed plan is instantiated into specific content distribution data, as shown in the table in the figure. Users can click "Confirm All Execution" to send the specific content distribution data to the target users.

[0067] refer to Figure 5 Taking the "C9 universities" demographic as an example, this paper introduces the process of generating and pushing content distribution data. Figure 5 It is known that the target group is "C9 universities", which includes approximately XX0,000 high-potential users with top academic backgrounds, and the execution period of the task is strictly anchored within the time window of "October 1 to October 30".

[0068] Next, the core content construction and rendering phase begins. Given that the "C9 University" group typically possesses high academic qualifications, strong research capabilities, and a high level of interest in high-end positions, the generative logic underwent deep customization when constructing the email content. Visually, image generation capabilities or material matching algorithms were used to render a dynamic banner in the email header containing the slogan "You have one new job opportunity to view," utilizing vibrant orange tones and cartoon elements to reduce reading pressure and attract visual focus. Regarding the selection of core service entities, the processing logic performed precise business attribute matching, retrieving data for highly specialized positions such as "xxxx Co., Ltd. - Strategic Planning and Industrial Policy Research (PhD Required)" and "Chip System Architect," which are highly relevant to the "C9" tag.

[0069] Next, these discrete job data were packaged into structured interactive cards. The cards clearly displayed the job title, salary range (e.g., 40,000-60,000 RMB) and attractive tags such as "state-owned enterprise", and were equipped with interactive buttons for "communicate now" and "view job responsibilities", thus directly building a mini job landing page in the email body.

[0070] Corresponding to the method embodiments of this application, this application also provides a generation apparatus for a content distribution scheme based on a large model, such as... Figure 6 As shown, the generating apparatus includes: The receiving module 110 is used to receive natural language instructions about the content distribution scheme. The pre-trained large model parses the natural language instructions to extract control parameters, which include at least: scheme topic and resource scheduling constraints. The filtering module 120 is used to parse the semantic features of the scheme theme and establish a semantic mapping relationship between the semantic features and the attribute tags of the candidate objects, so as to filter out a set of target objects that are compatible with the scheme theme; The content distribution construction module 130 is used to analyze the behavioral and interest characteristics of each object in the target object set, and, in combination with the scheme theme and resource scheduling constraints, construct content distribution data for the target object set. The content distribution scheme generation module 140 is used to generate a content distribution scheme based on the target object set and content distribution data, and to execute the push task according to the content distribution scheme.

[0071] Figure 7This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. The electronic device can be implemented as a server or other various terminal devices, such as desktop personal computers, tablet computers, laptop computers, mobile phones, etc., including a processor 601 and a memory 602. The memory 602 stores a program instruction set. When the processor 601 executes the program instruction set in the memory 602, it implements any of the aforementioned methods for generating content distribution schemes based on large models.

[0072] Specifically, the processor 601 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of the present invention.

[0073] Memory 602 may include mass storage for data or instructions. For example, and not limitingly, memory 602 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 602 may include removable or non-removable (or fixed) media. Where appropriate, memory 602 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 602 is non-volatile solid-state memory.

[0074] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the generation method of the large-model-based content distribution scheme provided by this invention.

[0075] In one example, the electronic device may also include a communication interface 603 and a bus 604. The processor 601, memory 602, and communication interface 603 are connected via bus 604 and communicate with each other. Communication interface 603 is primarily used to enable communication between modules, devices, units, and / or equipment in the embodiments of the present invention. Bus 604 includes hardware, software, or both, coupling components of the online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) Interconnect, Industry Standard Architecture (ISA) bus, Infinite Bandwidth Interconnect, Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 604 may include one or more buses. Although specific buses are described and illustrated in the embodiments of the present invention, the present invention is contemplated by any suitable bus or interconnect.

[0076] The present invention also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement any of the large-model-based content distribution scheme generation methods described in the foregoing embodiments. The computer-readable storage medium can be any medium that can tangibly contain or store computer-executable instructions for use by or in conjunction with an instruction execution system, apparatus, or device. The storage medium can be a transient computer-readable storage medium or a non-transitory computer-readable storage medium. Non-transitory computer-readable storage media may include, but are not limited to, magnetic storage devices, optical storage devices, and / or semiconductor storage devices. Examples of such storage devices include, for example, magnetic disks, optical discs based on CD, DVD, or Blu-ray technology, and persistent solid-state storage such as flash memory and solid-state drives.

[0077] This invention also provides a computer program product, comprising a set of computer program instructions, which, when executed by a processor, implements any of the content distribution scheme generation methods based on a large model in the foregoing embodiments. The computer program product includes, but is not limited to, application installation packages published on websites and app stores, application plugins, and mini-programs that can run within certain applications.

[0078] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0079] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the scope of the invention. Therefore, all equivalent technical solutions should also fall within the scope of the invention.

Claims

1. A method for generating a content distribution scheme based on a large model, characterized in that, include: The system receives natural language instructions regarding a content distribution scheme. A pre-trained large model parses the natural language instructions to extract control parameters, which include at least: scheme topic and resource scheduling constraints. The semantic features of the proposed solution theme are analyzed, and a semantic mapping relationship is established between the semantic features and the attribute tags of the candidate objects to filter and obtain a set of target objects that are compatible with the proposed solution theme; Analyze the behavioral and interest characteristics of each object in the target object set, and combine the scheme theme and resource scheduling constraints to construct content distribution data for the target object set; A content distribution scheme is generated based on the target object set and content distribution data, and the push task is executed according to the content distribution scheme.

2. The generation method according to claim 1, characterized in that, Establishing a semantic mapping relationship between the semantic features and the attribute tags of the candidate objects to filter out a set of target objects that are compatible with the theme of the scheme includes: Multidimensional topic tags are extracted from the semantic features, and the multidimensional topic tags include one or more of the following: industry tags, regional tags, and identity tags; Calculate the semantic relevance between the multidimensional topic tags and the attribute tags of the candidate objects, and determine the objects whose semantic relevance meets the preset matching threshold as the first target object set.

3. The generation method according to claim 2, characterized in that, Further includes: Based on the historical behavior data of each object in the first target object set, predict the click probability of each object on the topic of the solution; Based on the click probability, the objects in the first target object set are sorted, and the objects with a click probability higher than a preset probability threshold are determined as the second target object set, and the second target object set is used as the final target object set.

4. The generation method according to claim 1, characterized in that, Analyze the behavioral and interest characteristics of each object in the target object set, and combine this with the scheme theme and resource scheduling constraints to construct content distribution data for the target object set, including: Based on the historical behavior logs of each object in the target object set, the behavior preference type of each object is determined, including preference for industry information and preference for service entities. Under the premise of meeting the resource scheduling constraints, differentiated content processing is performed based on the behavioral preference type and scheme theme of each object to construct corresponding content distribution data.

5. The generation method according to claim 4, characterized in that, The differentiated content processing based on the behavioral preference type and theme of each object to construct corresponding content distribution data includes: If the object's behavioral preference type is industry information preference, and its historical behavior logs are identified as being concentrated in specific industry tags, then industry dynamic copy is generated based on the solution theme and specific industry tags, and the industry dynamic copy is used as content distribution data; If the object's behavior preference type is a preference service entity type, and its historical behavior logs are identified as being concentrated on specific business attributes, then the target service entity that matches the solution theme and specific business attributes is retrieved, and the target service entity is encapsulated as a distribution landing page as content distribution data.

6. The generation method according to claim 1, characterized in that, The step of executing the push task according to the content distribution scheme includes: Extract the historical active logs of each object in the target object set, and perform time-domain analysis on the historical active logs to extract the active time period characteristics of each object; Based on the active time period characteristics, a high-frequency active time window is determined for each object, and the trigger time of the push task for each object is mapped to the high-frequency active time window.

7. The generation method according to claim 1, characterized in that, Further includes: During the execution of the push task, the actual feedback indicators of the task are collected, and the deviation between the actual feedback indicators and the predicted indicators is calculated. If the deviation value exceeds a preset abnormal threshold, a dynamic optimization process for the content distribution scheme is executed; the dynamic optimization process includes at least one of the following: High-response feature vectors are extracted from objects that generate positive interactive behavior in executed tasks. The target object set is then expanded based on the high-response feature vectors, and supplementary push tasks are generated for the expanded target object set. Identify silent objects that have generated exposure behavior but not deep interaction behavior in the executed tasks, regenerate secondary content distribution data that differs from the original content distribution data for the silent objects, and push it to the silent objects.

8. A generation device for a content distribution scheme based on a large model, characterized in that, include: A receiving module is used to receive natural language instructions about a content distribution scheme. A pre-trained large model parses the natural language instructions to extract control parameters, which include at least: scheme topic and resource scheduling constraints. The filtering module is used to parse the semantic features of the scheme theme and establish a semantic mapping relationship between the semantic features and the attribute tags of the candidate objects, so as to filter out a set of target objects that are compatible with the scheme theme; The content distribution construction module is used to analyze the behavioral and interest characteristics of each object in the target object set, and combine the scheme theme and resource scheduling constraints to construct content distribution data for the target object set; The content distribution scheme generation module is used to generate a content distribution scheme based on the target object set and content distribution data, and to execute the push task according to the content distribution scheme.

9. An electronic device, characterized in that, This includes a processor and a memory, the memory storing a set of computer program instructions, which, when executed by the processor, implement the generation method of the content distribution scheme based on any one of claims 1-7.

10. A computer program product, characterized in that, It includes a set of computer program instructions that, when executed by a processor, implement the generation method of the content distribution scheme based on any one of claims 1-7.

Images