Content generation control method, apparatus, device, and storage medium

Abstract

The application discloses a content generation control method and device, equipment and storage medium. The scheme comprises: in response to a recommendation request submitted to a recommendation engine, monitoring the state of a candidate pool of the recommendation engine; the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool to generate a recommendation result after sorting. When it is monitored that the state of the candidate pool is that there is gap data, a targeted generation instruction is generated based on the semantic description corresponding to the gap data; the gap data is used to represent the state that the candidate pool cannot provide a sufficient number of recommendable content for the recommendation request; the semantic description is used to indicate the content attribute involved in the gap data; the targeted generation instruction is used to indicate the task description of generating content corresponding to the gap data to fill the gap. A content generation model is called to generate target content for filling the gap data according to the targeted generation instruction, and the target content is written into the candidate pool for the recommendation engine to include it in the sorting when processing subsequent recommendation requests.

Description

Technical Field

[0001] This application relates to the field of content generation technology, and in particular to a content generation control method, apparatus, device and storage medium. Background Technology

[0002] The core function of intelligent content platforms (such as online learning systems, news recommendation apps, and short video distribution platforms) is to use recommendation engines to filter and rank personalized content from a massive content library that matches the user's current interests, abilities, or context, thereby achieving precise distribution. The operational quality of these platforms highly depends on the richness and coverage of their content library—only when the candidate pool contains a sufficient quantity, diversity, and appropriate difficulty level of content can the recommendation engine effectively complete the ranking and push, thus continuously satisfying the user experience. However, with the expansion of the user base and dynamic changes in demand, the content library often experiences supply gaps in specific categories, difficulty ranges, or types, posing a serious challenge to the platform's automated operation capabilities.

[0003] In the current industry-standard architecture, recommendation systems and content generation systems operate independently: when processing user requests, recommendation systems only sort existing content, unable to detect gaps in the candidate pool such as insufficient quantity, difficulty gaps, or decreased diversity in real time, and even less able to proactively trigger content completion; while content generation systems rely on scheduled tasks or manual instructions for batch production, severely disconnected from the immediate needs of the recommendation side. A few linked offline statistical solutions (such as planning the next generation direction based on historical popularity data) are inherently lagging and unable to cope with the personalized gaps emerging in real-time recommendation requests. Therefore, existing technologies lack a closed-loop mechanism that can respond to recommendation requests, monitor the candidate pool status in real time, automatically drive content generation based on semantic descriptions of gaps, and immediately inject the generated results into the candidate pool. This is precisely the problem that this invention aims to solve. Summary of the Invention

[0004] This application proposes a content generation control method, apparatus, device, and storage medium, aiming to solve the problem that existing recommendation systems and content generation systems are disconnected and cannot automatically fill in gaps in real-time content. Accordingly, the technical solution of this application is as follows: In a first aspect, embodiments of this application provide a content generation control method, including: In response to a recommendation request submitted to the recommendation engine, the status of the candidate pool of the recommendation engine is monitored; wherein, the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool, sort it, and generate recommendation results; When the candidate pool is detected to have gap data, a targeted generation instruction is generated based on the semantic description corresponding to the gap data; wherein, the gap data is used to characterize the state that the candidate pool cannot provide a sufficient number of recommendable contents for the recommendation request; the semantic description is used to indicate the content attributes involved in the gap data; and the targeted generation instruction is used to instruct the generation of a task description to fill the gap data. The content generation model is invoked to generate target content to fill the data gaps according to the directed generation instructions; The target content is written into the candidate pool of the recommendation engine so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests.

[0005] Secondly, embodiments of this application provide a content generation control device, including: The candidate pool monitoring module monitors the status of the candidate pool of the recommendation engine in response to the recommendation request submitted by the recommendation engine; wherein, the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool, sort it, and generate recommendation results; The instruction generation module, when detecting that the candidate pool has gap data, generates a targeted generation instruction based on the semantic description corresponding to the gap data; wherein, the gap data is used to characterize the state that the candidate pool cannot provide a sufficient number of recommendable contents for the recommendation request; the semantic description is used to indicate the content attributes involved in the gap data; and the targeted generation instruction is used to instruct the generation of a task description to fill the gap data. The instruction execution module calls the content generation model to generate target content to fill the gap data according to the directed generation instruction; The content filling module writes the target content into the candidate pool of the recommendation engine, so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests.

[0006] Thirdly, embodiments of this application provide an electronic device, including a processor; and a memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the method described in the first aspect.

[0007] Fourthly, embodiments of this application provide a computer-readable storage medium for storing computer-executable instructions that, when executed by a processor, implement the method described in the first aspect.

[0008] In a typical intelligent content platform architecture, the recommendation engine is responsible for selecting content from a candidate pool, ranking it, and generating recommendations. This candidate pool stores all content available for recommendation. The content generation system, on the other hand, operates independently, producing content in batches according to preset cycles or manual instructions. The lack of a linkage mechanism triggered by recommendation requests means that when processing user requests, if the candidate pool has gaps such as insufficient difficulty levels or a lack of content categories, the recommendation engine can only passively reduce recommendation quality. Furthermore, the content generation system cannot detect these real-time gaps and implement targeted improvements. To address this technical problem, this application proposes the following solution: In response to a recommendation request submitted by the recommendation engine, the status of the candidate pool is monitored in real time. When a gap in the candidate pool is detected, this gap data represents the state that the candidate pool cannot provide a sufficient number of recommendable contents for the current recommendation request. A targeted generation instruction is generated based on the semantic description corresponding to the gap data. The semantic description indicates the content attributes involved in the gap, including content category, difficulty level, and content type. The targeted generation instruction indicates the specific task description needed to fill the gap. Subsequently, the content generation model is invoked to generate target content according to the targeted generation instruction, and this target content is written into the candidate pool so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests. In the above steps, gap detection is embedded in the real-time processing chain of recommendation requests, enabling the discovery of content gaps to be precisely triggered with each recommendation request. This captures the personalized and instantaneous changes in users' content needs, overcoming the lag of offline statistics or manual inspection. Targeted generation instructions are generated based on semantic descriptions, ensuring that the attributes of the generated content perfectly match the gaps, avoiding the problem of content being out of sync with demand in traditional batch production. The generated content is written into a candidate pool, significantly shortening the delay from gap detection to content availability, achieving a real-time closed loop of perception, decision-making, and execution. Therefore, this solution enables the recommendation engine to proactively trigger content completion, while providing real-time and semantic demand input to the content generation system. It effectively solves the architectural defects of passive quality degradation on the recommendation side and blind production on the generation side, significantly improving the timeliness and matching accuracy of content supply, reducing the cost of manual operational intervention, and ensuring the stable quality of recommendation results under dynamic changes in the content library. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the content generation control method according to an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of the application architecture of the content generation control method in an embodiment of this application.

[0012] Figure 3 This is a schematic diagram of the structure of the content generation control device for embodiments of this application.

[0013] Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0015] One embodiment of this application provides a content generation control method. Figure 1 This is a flowchart illustrating the content generation control method, including: S101, in response to a recommendation request submitted to the recommendation engine, monitor the status of the candidate pool of the recommendation engine; wherein, the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool, sort it, and generate recommendation results.

[0016] This embodiment first requires establishing a real-time monitoring relationship between recommendation requests and the candidate pool status. In a conventional intelligent content platform architecture, the recommendation engine is an independently operating distribution module. Its core function is to receive recommendation requests from users, select content from the candidate pool, and generate recommendation results according to a preset sorting strategy, returning them to the user. The candidate pool serves as the content supply source for the recommendation engine, storing all content data available for recommendation. However, in existing architectures, the recommendation engine passively retrieves existing content from the candidate pool and lacks the ability to perceive the content supply status of the candidate pool, let alone feed this status back to any other module. This means that when the number of recommendable content in a certain difficulty range or content category in the candidate pool is insufficient, the recommendation engine can only sort based on the limited content, leading to a decrease in the coverage or quality of the recommendation results, and the system as a whole lacks awareness and countermeasures for this. Therefore, this embodiment binds the monitoring of the candidate pool status to the processing link of recommendation requests in the first step of the method, providing a technical foundation for the timely discovery of content gaps.

[0017] Specifically, this embodiment responds to recommendation requests submitted by the recommendation engine by performing real-time monitoring of the candidate pool status. The recommendation request, as referred to here, is generated by a platform user when initiating a content retrieval action, such as a user entering an online learning application requesting recommended practice questions, opening a news aggregation application requesting recommended news articles, or entering a short video application requesting a refresh of the recommendation list. Once submitted to the recommendation engine, this request triggers the subsequent content selection and sorting process. Recommendation requests typically carry user identifiers, scenario identifiers, and optional filtering conditions (such as content category preferences, difficulty range constraints, etc.). The recommendation engine uses this information to match content from the candidate pool that meets the conditions. This embodiment synchronously binds the monitoring action with the reception of recommendation requests; that is, every time a recommendation request arrives at the recommendation engine, this embodiment performs a real-time status check on the candidate pool, rather than using periodic polling or offline batch analysis. The advantage of this synchronous triggering mechanism is that the monitoring overhead only occurs when there is an actual recommendation demand, avoiding the waste of computational resources caused by idle polling; at the same time, the monitoring results directly reflect the actual supply situation faced by the current request, without any perception lag caused by time differences.

[0018] The status monitoring of the candidate pool includes at least the following three dimensions of evaluation. The first dimension is whether the total number of content available for recommendation in the candidate pool meets the minimum number of candidates required for this recommendation request. When performing ranking operations, the recommendation engine typically needs a sufficiently large candidate set to ensure the diversity and accuracy of the ranking. If the total amount of content matching the conditions is lower than this threshold, the quality of the ranking results cannot be guaranteed. The second dimension is whether there are gaps in the distribution of content within a specific difficulty range in the candidate pool. In scenarios oriented towards learning or ability matching, the recommendation engine needs to push content that is appropriate to the user's current ability level. If the candidate pool lacks content reserves within the user's ability range, even if the overall number of candidates in the pool is sufficient, the user will not receive effective recommendation results. The third dimension is whether there is sufficient recommendable content under a specific content category in the candidate pool. Content category is the basic dimension for topic matching in recommendation systems. If the content under a certain category is insufficient to support the filtering requirements of the recommendation request, it will lead to a lack of recommendation results under that category. These three dimensions each reflect different aspects of the supply status of the candidate pool and can trigger gap detection independently, or multiple dimensions of gap reporting can be triggered simultaneously in the same recommendation request.

[0019] For example, in the usage scenario of an online learning platform, a user with an ability score of 3.5 requests recommended practice questions from the recommendation engine. The recommendation engine needs to filter content from the candidate pool that matches the user's current ability level, usually searching within a range near the user's ability score, such as the difficulty range [3.3, 3.7]. If there are only 2 items in the candidate pool within this difficulty range, while the recommendation engine requires 10 items to function normally, this monitoring will simultaneously identify gaps in two dimensions: first, the total number of items in the candidate pool that match the request conditions is insufficient, which falls under the category of "insufficient content in the candidate pool"; second, the gap is concentrated within a specific difficulty range, which falls under the category of "content gaps exist in a specific difficulty range in the candidate pool". By triggering the monitoring action synchronously with the recommendation request, this embodiment ensures that the content supply status of the candidate pool can be grasped in real time during each recommendation service process. Any situation that may affect the recommendation quality due to insufficient content supply can be detected in time during the current request processing stage, thereby providing a precise triggering time and gap description for subsequent targeted generation.

[0020] S102, when it is detected that the candidate pool has gap data, a targeted generation instruction is generated based on the semantic description corresponding to the gap data; wherein, the gap data is used to characterize the state that the candidate pool cannot provide a sufficient number of recommendable contents for the recommendation request; the semantic description is used to indicate the content attributes involved in the gap data; and the targeted generation instruction is used to indicate the task description for generating the content corresponding to fill the gap data.

[0021] Building upon the real-time monitoring achieved in the previous step, this embodiment further addresses the challenge of transforming the detected undersupply in the candidate pool into usable information for subsequent content production. A simple binary judgment of "a gap exists" is insufficient to drive any meaningful generation behavior, as different types of gaps correspond to drastically different content production needs. Specifically, an absolute shortage of content in the candidate pool, a lack of content within specific difficulty ranges, and a decrease in the diversity of recommended content types all require entirely different generation directions and content. Without a precise characterization of the gap's nature, subsequent content production will lack a clear objective, and the generated content may be misaligned with actual needs. Therefore, this embodiment, upon detecting a gap, further transforms the original monitoring results into structured gap data and generates targeted generation instructions based on the semantic description of the gap data, thereby providing precise task input for the content production process.

[0022] Gap data is a parameterized representation of the insufficient supply in the candidate pool, serving to accurately present the information of "what is missing" in the form of structured data. In this embodiment, the state represented by gap data includes at least one of the following situations: insufficient content in the candidate pool, i.e., the total number of content in the candidate pool that matches the filtering conditions of the current recommendation request is lower than the minimum number of candidates required for the recommendation engine to perform normal ranking; content gaps exist in specific difficulty ranges in the candidate pool, i.e., although the overall number of candidates in the candidate pool may be sufficient, there is a lack of sufficient content to distribute within the user's ability range corresponding to the current recommendation request; decreased diversity of content types in the recommendation results generated based on the candidate pool, i.e., the content type distribution in the candidate pool is too homogeneous, causing the recommendation engine to be unable to provide users with diverse recommendation results. The above three gap situations can occur independently or simultaneously. For example, on an online learning platform, when a user with an ability score of 3.5 requests recommendations for practice questions, the recommendation engine filters within the difficulty range [3.3, 3.7]. If there are only 2 items in the candidate pool for this range, while normal sorting requires at least 10 candidates, then the gap data simultaneously represents two states: "insufficient content in the candidate pool" and "a gap in content within a specific difficulty range in the candidate pool".

[0023] After acquiring the gap data, this embodiment further generates targeted generation instructions based on the semantic description corresponding to the gap data. The semantic description is used to indicate the content attributes involved in the gap data in a semantic manner, including at least one of the following: the content category corresponding to the gap data, used to indicate the knowledge domain or topic category of the gap, such as "mathematics" or "Chinese"; the difficulty range corresponding to the gap data, used to indicate the ability value range of the gap, such as "3.3 to 3.7"; the content type corresponding to the gap data, used to indicate the specific question type or format of the missing content, such as "multiple choice" or "fill in the blank"; the priority corresponding to the gap data, used to indicate the urgency of responding to the gap when it is detected, which can be determined based on the user's waiting state corresponding to the recommendation request that triggered the gap. The targeted generation instructions are task descriptions further generated based on the above semantic descriptions, used to instruct the content generation model to perform specific tasks. These targeted generation instructions differ from traditional manual instructions or scheduled task instructions; their content is entirely driven by the semantic information in the gap data. Taking the aforementioned online learning platform scenario as an example, when the gap data indicates that "the amount of content in category A within the difficulty range of 3.3 to 3.7 is insufficient," the targeted generation instruction generated based on the semantic description of the gap data in this embodiment can be expressed as "generate several multiple-choice questions with a difficulty range of 3.3 to 3.7 for category A." By transforming the gap data into a semantic description and further generating targeted generation instructions, this embodiment achieves the key transformation from "discovering insufficient supply" to "clarifying production goals."

[0024] S103, invoke the content generation model to generate target content to fill the gap data according to the targeted generation instructions.

[0025] In the previous step, this embodiment has already clarified "what content needs to be generated" through targeted generation instructions. However, these instructions remain at the task description level and cannot be directly stored in the candidate pool as content data. Therefore, after obtaining the targeted generation instructions, this embodiment further calls the content generation model to perform content generation operations according to the instructions, thereby transforming the abstract generation task into specific target content and making the instructions truly usable content resources.

[0026] Specifically, this embodiment invokes a content generation model, using the targeted generation instruction as its task input. The content generation model then performs the content generation operation, outputting target content to fill in the gap data. The content generation model can employ a large language model based on a transformer architecture or a multimodal generation model with content creation capabilities, capable of automatically generating content data that meets the requirements based on a textual task description. The targeted generation instruction serves as input prompts for the content generation model, providing clear generation goals and constraints, including key parameters such as content classification, difficulty level, and content type, ensuring that the target content output by the model is consistent in attributes with the missing content indicated by the gap data. In one implementation, the recommendation request processed in this embodiment is initiated by a target user; that is, the user object served by this recommendation request is a clearly defined target user, and the content gaps in the candidate pool are detected under the filtering conditions of this target user. In this case, when invoking the content generation model to generate target content, this embodiment further obtains the long-term preference profile of the target user. Long-term preference profiles refer to data extracted from the target user's historical interaction behavior to characterize the target user's content preference features. Examples include whether the user prefers text / image or video content, multiple-choice or fill-in-the-blank questions, and preferred language style or expression. These preference features can be statistically analyzed and mined from the target user's past clicks, browsing, completions, ratings, and other interactive behaviors. In this embodiment, after obtaining the long-term preference profile, the user preference features represented by the long-term preference profile, together with the targeted generation instruction, serve as generation constraints. This instructs the content generation model to further generate target content adapted to the target user's personalized preferences, provided that the classification, difficulty, and type requirements specified in the targeted generation instruction are met. For example, if the targeted generation instruction requires "generating several multiple-choice questions with a difficulty range of 3.3 to 3.7 for category A," and the target user's long-term preference profile indicates that the user prefers questions combining text and images and prefers concise and clear question stems, then this embodiment instructs the content generation model to incorporate text and image elements when generating questions and to organize the question stem language in a concise style. In this way, the generated target content not only accurately corresponds to the gap status of the candidate pool in terms of content classification, difficulty range and content type, but also matches the personalized preferences of the target users who triggered the gap detection in terms of content style and expression. This makes the generated content not only fill the structural supply shortage in the candidate pool, but also have higher user acceptance and recommendation conversion potential.

[0027] S104. Write the target content into the candidate pool of the recommendation engine so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests.

[0028] It should be noted that if the target content is stored in an offline content library and awaits a batch update cycle, the time delay between gap detection and gap filling will weaken the timeliness of the real-time closed-loop mechanism, and may even result in the gap disappearing with the end of the current recommendation request before the filling content is ready. Therefore, in this embodiment, after the target content is generated, it is immediately written into the candidate pool of the recommendation engine, so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests. "Immediate writing" means that the target content, after passing a preset quality check, is directly added to the data storage structure of the candidate pool, without needing a batch synchronization process of the offline content library or a manual review and queuing process. After writing, the total amount of content in the candidate pool increases immediately, and the target content is in the same recommendable state as the existing content in the pool. When processing subsequent recommendation requests, the target content can be overridden by the selection logic of the recommendation engine and participate in the ranking and distribution. For example, after generating target content in response to a user's recommendation request and immediately writing it into the candidate pool, if the user initiates another recommendation request within a preset time window, the recommendation engine can include the target content in the ranking; if the user does not initiate another request, the target content remains in the candidate pool and can be used by other users matching the same content attribute conditions. Thus, the time delay from gap detection to content availability is compressed into the time window required for content generation and immediate writing.

[0029] To ensure the basic usability and quality of the target content written into the candidate pool, this embodiment performs a quality assessment on the target content before it is immediately written into the recommendation engine's candidate pool. The quality assessment includes at least one of the following: content validity verification, format conformity check, and semantic consistency verification with the gap data. Content validity verification checks whether the generated content is complete and free of obvious errors or logical contradictions, such as whether the generated questions contain complete stems and options, and whether the answers are correct, preventing invalid content from entering the candidate pool. Format conformity check checks whether the generated content conforms to the platform's predefined content structure standards, such as whether content fields are complete and whether the data format can be correctly parsed by the recommendation engine, preventing errors when the recommendation engine reads the candidate pool due to format abnormalities. Semantic consistency verification with the gap data checks whether the generated content is consistent with the gap data that triggered the generation in terms of key attributes such as content category, difficulty level, and content type. Only when the target content passes all or part of the above quality assessment checks is it allowed to be written into the candidate pool; if it fails the quality assessment, the target content is not written into the candidate pool, and regeneration or recording of anomalies may be triggered. By introducing a quality assessment mechanism, this embodiment effectively avoids low-quality or undesirable output from polluting the candidate pool, thus ensuring the bottom line of candidate content quality for the recommendation engine.

[0030] Furthermore, this embodiment introduces a priority scheduling mechanism for processing gap data. Gaps data with different semantic descriptions are assigned corresponding priorities. Since the recommendation engine continuously processes recommendation requests from a large number of users, multiple gap data with different semantic descriptions may be detected within the same time period, resulting in the generation of multiple targeted generation instructions. However, the concurrent processing capability of the content generation model is limited. When multiple targeted generation instructions are pending completion, the execution order of each instruction needs to be determined. Therefore, this embodiment determines the execution order according to the priority of the gap data to which each targeted generation instruction belongs, with higher-priority targeted generation instructions being scheduled for execution first. Further, this embodiment also monitors the perception frequency of gap data with the same semantic description within a preset time window, i.e., counting the number of times gap data with completely identical semantic descriptions are detected within a specified duration (e.g., one hour), and adjusts the priority of the corresponding gap data based on this perception frequency. The priority of any target semantic description is positively correlated with its corresponding perception frequency; that is, the higher the frequency of perception of gap data with the same semantic description within the preset time window, the greater the increase in the priority of that gap data. For example, a gap with the semantic description "Content Category = Category A, Difficulty Range = 3.3 to 3.7" was detected five times in one hour by recommendation requests from different users, while another gap with the semantic description "Content Category = Category B, Difficulty Range = 4.0 to 4.5" was detected only once in the same time window. In this case, the former was perceived significantly more frequently than the latter, and this embodiment prioritizes the former. The rationale for this design is that the fact that the same gap is perceived multiple times in a short period of time means that the gap affects not only a single user's single recommendation experience, but also the common supply shortage faced by multiple users in the same content segment. Its urgency in filling and its impact are greater, and therefore it should receive a higher processing priority.

[0031] Furthermore, this embodiment establishes a content value accumulation mechanism from temporary supplementation to long-term accumulation. The candidate pool is mainly used to support the real-time distribution of the recommendation engine, and its storage resources are usually limited, making it impossible to retain all the target content generated in real time indefinitely. If all generated content resides in the candidate pool for a long time, it will lead to the expansion of the candidate pool content, which will reduce the ranking efficiency of the recommendation engine. Therefore, after the target content is included in the recommendation results by the recommendation engine, this embodiment further obtains user feedback signals for the target content. User feedback signals refer to the behavioral response data of users to the content after the target content is recommended to users and generates exposure and interaction, such as whether the user clicked or viewed the content, whether they completed the task contained in the content (such as completing a practice question), and whether they gave a rating or evaluation of the content. Based on these user feedback signals, this embodiment determines whether the target content meets the preset persistence conditions. The preset persistence conditions may include, but are not limited to: the click-through rate of the content after it is recommended exceeds a preset threshold, the user completion rate exceeds a preset threshold, the user rating is higher than a preset score, or a combination of the above indicators. If the persistence condition is met, it means that the target content not only fills the immediate gap in the candidate pool but also possesses a quality level that is continuously accepted and recognized by users. In this embodiment, the target content is persistently stored in a knowledge base. This knowledge base is independent of the candidate pool and is used to store verified high-quality content for a long time, serving as a reference corpus and retrieval enhancement source for the content generation model. In subsequent generation tasks, the content generation model can retrieve relevant high-quality samples from the knowledge base as reference examples or contextual enhancement information during generation, thereby improving the quality and stability of subsequent content generation. Through this mechanism, this embodiment achieves a functional leap from "immediately generated temporary content" to "verified and accumulated long-term content".

[0032] The system module deployment and execution logic of the method in this embodiment will be described below.

[0033] Figure 2This example illustrates the system architecture upon which this embodiment is based. This architecture is divided into a user layer, a recommendation engine layer, a content generation agent layer, and a content management layer, from the user side to the content generation side. The recommendation engine layer deploys a real-time recommendation engine and a gap detector. The real-time recommendation engine is responsible for retrieving and ranking content, i.e., selecting content from the candidate pool (i.e., the content library in the attached figure) and generating recommendation results after ranking according to a preset strategy. The gap detector is embedded within the recommendation engine layer and deeply coupled with the recommendation request processing chain. It is configured to automatically sense the recommendation result generation process corresponding to each recommendation request submitted to the recommendation engine. When the recommendation engine encounters a situation such as insufficient candidate content, a significant decrease in the diversity of ranking results, or an empty recommended content range in a specific capability interval while processing the request, the gap detector immediately triggers a gap report, outputting the insufficient supply status of the candidate pool as gap data to a waiting queue storing multiple targeted generation instructions (i.e., the gap instruction queue in the attached figure). By embedding the gap detector into the recommendation engine, this solution makes gap perception capability a part of the recommendation engine's own functionality, avoiding the additional communication overhead and timing delays caused by introducing an external independent monitoring module.

[0034] In the content generation agent layer, there are orchestration agents and generation agents (i.e., content generation models). The orchestration agent, as the core execution entity for semantic parsing and task scheduling, communicates with the waiting queue in the recommendation engine layer. Gap data reported by the gap detector enters the waiting queue containing multiple targeted generation instructions according to the detection sequence. This queue has a priority sorting function, which can dynamically sort the gap data in the queue based on the priority attributes carried in the gap data. The orchestration agent retrieves gap data from this queue sequentially and is responsible for parsing the semantic descriptions carried by the gap data. The parsed content includes attributes such as the content category, difficulty range, content type, and priority corresponding to the gap, and generates targeted generation instructions based on the parsing results. These targeted generation instructions are no longer general keywords or rough directional prompts, but rather a complete task description containing the generation quantity, content attribute constraints, and task objectives. The orchestration agent sends the generated targeted generation instructions to the generation agent, which then executes the content generation operation according to the instructions. The generation agent has the ability to understand gap semantics and generate corresponding content. It can use a large language model based on a transformer architecture or a multimodal generation model with content creation capabilities to output target content for filling gap data.

[0035] Within the content management layer, a content library is deployed, which serves as the candidate pool upon which the recommendation engine relies. The target content output by the generation agent is evaluated for quality and then immediately written back to the content library, ensuring that newly generated content immediately enters the recommendation engine's recommended range for subsequent recommendation requests to be included in the ranking. Through the collaborative work of the gap detector, waiting queue, orchestration agent, and generation agent, this embodiment achieves a complete chain from recommendation request-driven gap perception, to priority-driven task scheduling, to semantic parsing-driven targeted generation, and finally to real-time writing-driven closed-loop completion. The perception, decision-making, and execution stages are modularly divided at the architectural level, while maintaining end-to-end real-time linkage capabilities.

[0036] In summary, the method in this embodiment monitors the candidate pool status in real time in response to recommendation requests. When gap data is detected, it generates targeted generation instructions based on the semantic description of the gap, and then calls the content generation model to generate target content and writes it into the candidate pool immediately for inclusion in the ranking of subsequent recommendation requests. In the above process, gap detection is embedded in the processing chain of each recommendation request, enabling the discovery of content gaps to be triggered accurately with each recommendation request, overcoming the lag of offline statistics or manual inspection. Generating targeted generation instructions based on the semantic description of the gap ensures that the attributes of the generated content completely match the gap, avoiding the problem of content being out of touch with demand in traditional batch production. Writing the generated content into the candidate pool immediately significantly shortens the delay from gap detection to content availability, achieving a real-time closed loop of perception, decision-making, and execution. In addition, the quality assessment mechanism ensures the bottom line of content quality in the candidate pool, the priority scheduling mechanism ensures that limited generation resources are tilted towards high-impact gaps, and the user feedback-driven persistence mechanism accumulates long-term reusable knowledge assets for the content generation system. Therefore, this solution enables the recommendation engine to proactively trigger content completion, while providing real-time and semantic demand input to the content generation system. It effectively solves the architectural defects of passive quality degradation on the recommendation side and blind production on the generation side, significantly improves the timeliness and matching accuracy of content supply, reduces the cost of manual operation and intervention, and ensures the stable quality of recommendation results under dynamic changes in the content library.

[0037] In addition, corresponding Figure 1 As shown, another embodiment of this application also provides a content generation control device. Figure 3 This is a schematic diagram of the content generation control device, including: The candidate pool monitoring module 310 monitors the status of the candidate pool of the recommendation engine in response to a recommendation request submitted by the recommendation engine; wherein the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool, sort it, and generate recommendation results.

[0038] The instruction generation module 320, when detecting that the candidate pool has gap data, generates a targeted generation instruction based on the semantic description corresponding to the gap data; wherein, the gap data is used to characterize the state that the candidate pool cannot provide a sufficient number of recommendable contents for the recommendation request; the semantic description is used to indicate the content attributes involved in the gap data; and the targeted generation instruction is used to instruct the generation of a task description to fill the gap data.

[0039] The instruction execution module 330 calls the content generation model to generate target content to fill the gap data according to the directional generation instruction.

[0040] The content filling module 340 writes the target content into the candidate pool of the recommendation engine in real time, so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests.

[0041] Optionally, before writing the target content into the candidate pool of the recommendation engine in real time, the content filling module 340 is further configured to: perform a quality assessment on the target content; the quality assessment includes at least one of content validity verification, format standardization check, and semantic consistency check with the gap data; wherein, the target content is allowed to be written into the candidate pool when it passes the quality assessment.

[0042] Optionally, gap data with different semantic descriptions are each assigned a corresponding priority. When there are multiple directional generation instructions to be completed, the instruction execution module 330 determines the execution order for each directional generation instruction according to the priority of the gap data to which it belongs. The instruction execution module 330 is also used to: monitor the perception frequency of gap data with the same semantic description within a preset time window; and adjust the priority of the corresponding gap data based on the perception frequency. Wherein, the priority of any target semantic description is positively correlated with its corresponding perception frequency.

[0043] Optionally, the content attributes indicated by the semantic description include at least one of the following: the content category of the content corresponding to the gap data; the difficulty level of the content corresponding to the gap data; the content type of the content corresponding to the gap data; and the priority of the gap data.

[0044] Optionally, the apparatus in this embodiment further includes: The knowledge supplementation module is used to: after the target content is included in the recommendation results by the recommendation engine, obtain user feedback signals for the target content; determine whether the target content meets the preset persistence conditions based on the user feedback signals; if it does, persist the target content to the knowledge base; wherein the knowledge base serves as the reference corpus and / or retrieval enhancement source for the content generation model.

[0045] Optionally, the recommendation request processed by the recommendation engine is initiated by the target user; the instruction execution module 330 calls the content generation model to generate target content according to the targeted generation instruction, including: obtaining the long-term preference profile of the target user, wherein the long-term preference profile refers to data extracted based on the target user's historical interaction behavior to characterize the content preference features of the target user; calling the content generation model, using the long-term preference profile as a constraint, to generate target content adapted to the personalized preferences of the target user according to the targeted generation instruction.

[0046] Optionally, the state represented by the gap data includes at least one of the following: insufficient content in the candidate pool; content gaps in a specific difficulty range in the candidate pool; and decreased content type diversity in the recommendation results generated based on the candidate pool.

[0047] It should be noted that the content generation control device in this embodiment can be used as... Figure 2 The execution body of the method shown is therefore able to achieve... Figure 2 The steps and functions of the method shown are illustrated.

[0048] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Please refer to it. Figure 4 At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.

[0049] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0050] Memory is used to store computer programs. Specifically, a computer program may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides the computer program to the processor.

[0051] Specifically, the processor reads the corresponding computer program from non-volatile memory into memory and then runs it, forming the above-mentioned logical structure. Figure 3 The content generation control device is shown. Correspondingly, the processor executes the program stored in memory and specifically performs the following operations: In response to a recommendation request submitted to the recommendation engine, the status of the candidate pool of the recommendation engine is monitored; wherein the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool, sort it, and generate recommendation results.

[0052] When the candidate pool is detected to have gap data, a targeted generation instruction is generated based on the semantic description corresponding to the gap data; wherein, the gap data is used to characterize the state that the candidate pool cannot provide a sufficient number of recommendable contents for the recommendation request; the semantic description is used to indicate the content attributes involved in the gap data; and the targeted generation instruction is used to instruct the generation of a task description to fill the gap data.

[0053] The content generation model is invoked to generate target content to fill the data gaps according to the directed generation instructions.

[0054] The target content is written into the candidate pool of the recommendation engine in real time, so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests.

[0055] The above is as described in this instruction manual. Figure 1The content generation control method disclosed in the illustrated embodiments can be applied to a processor and implemented by the processor. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the processor's hardware or by instructions in software form. The processor mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0056] Of course, in addition to software implementation, the electronic device described in this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0057] Furthermore, embodiments of this application also propose a computer-readable storage medium that stores one or more computer programs, the one or more computer programs including instructions.

[0058] When the aforementioned instructions are executed by a portable electronic device that includes multiple applications, they enable the portable electronic device to perform... Figure 1 The steps in the method shown include: In response to a recommendation request submitted to the recommendation engine, the status of the candidate pool of the recommendation engine is monitored; wherein the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool, sort it, and generate recommendation results.

[0059] When the candidate pool is detected to have gap data, a targeted generation instruction is generated based on the semantic description corresponding to the gap data; wherein, the gap data is used to characterize the state that the candidate pool cannot provide a sufficient number of recommendable contents for the recommendation request; the semantic description is used to indicate the content attributes involved in the gap data; and the targeted generation instruction is used to instruct the generation of a task description to fill the gap data.

[0060] The content generation model is invoked to generate target content to fill the data gaps according to the directed generation instructions.

[0061] The target content is written into the candidate pool of the recommendation engine in real time, so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests.

[0062] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0063] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0064] The above are merely embodiments of this specification and are not intended to limit the scope of this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification. Furthermore, all other embodiments obtained by those skilled in the art without inventive effort should fall within the protection scope of this document.

Claims

1. A content generation control method, characterized in that, include: In response to a recommendation request submitted to the recommendation engine, the status of the candidate pool of the recommendation engine is monitored; wherein, the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool, sort it, and generate recommendation results; When the candidate pool is detected to have gap data, a targeted generation instruction is generated based on the semantic description corresponding to the gap data; wherein, the gap data is used to characterize the state that the candidate pool cannot provide a sufficient number of recommendable contents for the recommendation request; the semantic description is used to indicate the content attributes involved in the gap data; and the targeted generation instruction is used to instruct the generation of a task description to fill the gap data. The content generation model is invoked to generate target content to fill the data gaps according to the directed generation instructions; The target content is written into the candidate pool of the recommendation engine so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests.

2. The method according to claim 1, characterized in that, Before writing the target content into the candidate pool of the recommendation engine, the method further includes: A quality assessment is performed on the target content; the quality assessment includes at least one of content validity verification, format conformity check, and semantic consistency check with the gap data; wherein, the target content is allowed to be written into the candidate pool when it passes the quality assessment.

3. The method according to claim 1, characterized in that, The gap data with different semantic descriptions are each assigned a corresponding priority. When there are multiple directional generation instructions to be completed, the execution order of each directional generation instruction is determined according to the priority of its corresponding gap data. The method further includes: Monitor the frequency of perception of gap data with the same semantic description within a preset time window; The priority of the corresponding gap data is adjusted based on the perceived frequency; wherein, the priority of any target semantic description is positively correlated with its corresponding perceived frequency.

4. The method according to claim 3, characterized in that, The content attributes indicated by the semantic description include at least one of the following: The content classification of the gap data; The difficulty level of the content corresponding to the gap data; The content type corresponding to the gap data; The priority corresponding to the gap data.

5. The method according to claim 1, characterized in that, Also includes: Once the target content is included in the recommendation results by the recommendation engine, user feedback signals for the target content are obtained. Based on the user feedback signal, determine whether the target content meets the preset persistence conditions; If the conditions are met, the target content is persistently stored in the knowledge base; wherein the knowledge base serves as the reference corpus and / or retrieval enhancement source for the content generation model.

6. The method according to claim 1, characterized in that, The recommendation engine processes recommendation requests initiated by the target user. The content generation model is invoked to generate target content according to the targeted generation instructions, including: Obtain the long-term preference profile of the target user, wherein the long-term preference profile refers to data extracted based on the target user's historical interaction behavior to characterize the content preference features of the target user; The content generation model is invoked, and the long-term preference profile is used as a constraint to generate target content that matches the personalized preferences of the target user according to the targeted generation instructions.

7. The method according to any one of claims 1 to 6, characterized in that, The state represented by the gap data includes at least one of the following: The number of candidates in the candidate pool is insufficient; There are blank spaces in a specific difficulty range of the candidate pool; The diversity of content types in the recommendation results generated based on the candidate pool decreases.

8. A content generation control device, characterized in that, include: The candidate pool monitoring module monitors the status of the candidate pool of the recommendation engine in response to the recommendation request submitted by the recommendation engine; wherein, the candidate pool is used to store content that can be recommended, and the recommendation engine is used to select content from the candidate pool, sort it, and generate recommendation results; The instruction generation module, when detecting that the candidate pool has gap data, generates a targeted generation instruction based on the semantic description corresponding to the gap data; wherein, the gap data is used to characterize the state that the candidate pool cannot provide a sufficient number of recommendable contents for the recommendation request; the semantic description is used to indicate the content attributes involved in the gap data; and the targeted generation instruction is used to instruct the generation of a task description to fill the gap data. The instruction execution module calls the content generation model to generate target content to fill the gap data according to the directional generation instruction; The content filling module writes the target content into the candidate pool of the recommendation engine, so that the recommendation engine can include the target content in the ranking when processing subsequent recommendation requests.

9. An electronic device comprising a processor; and a memory configured to store computer-executable instructions, characterized in that, When executed, the computer-executable instructions cause the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium for storing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a processor, implement the method as described in any one of claims 1-7.

Images