Topic recommendation method and apparatus

Abstract

Embodiments of the present application provide a topic recommendation method and device, computer equipment, computer readable storage medium and computer program product, relating to artificial intelligence and data processing technology. The method is applied to a server and includes: receiving a topic recommendation request containing an object identifier of a target content creator; obtaining tool description information of a plurality of data tools and constructing a task processing state; calling a target model to process the task processing state; when the model processing result contains tool call information, calling a target data tool, writing tool results into the task processing state and calling the target model again; and when the model processing result contains candidate topic generation information, generating and outputting candidate topic data based on the candidate topic generation information and accumulated tool results. This technical solution can continuously accumulate multi-source tool results in the same task processing state and provide them for subsequent model processing reuse, improving the continuity of cross-round data processing.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing technology, and in particular to a topic recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product. Background Technology

[0002] In content generation scenarios, servers can generate candidate topics based on topic information, historical content data, or trending data provided by the target object. Related processing typically requires acquiring different types of data from multiple data sources and providing this data to the generation model. When the results from different data sources exist in different processing stages, cross-round model processing may encounter problems such as discontinuous data states, repeated construction of acquired data, or some data not being utilized by subsequent processing. Furthermore, the generated results in free text format may be difficult for subsequent data processing programs to directly read. Therefore, improving the reusability and consistency of multi-source data call results in continuous model processing has become one of the technical problems that needs to be addressed in related technologies.

[0003] It should be noted that the above content is not necessarily prior art, nor is it intended to limit the scope of patent protection of this application. Summary of the Invention

[0004] This application provides a method, apparatus, computer device, computer-readable storage medium, and computer program product for recommending topics, in order to solve or alleviate one or more of the technical problems mentioned above.

[0005] One aspect of this application provides a topic recommendation method, applied to a server, the method comprising: Receive a topic recommendation request, wherein the topic recommendation request includes the object identifier of the target content creator; Obtain tool description information corresponding to each of the multiple data tools, wherein the multiple data tools are used to obtain content creator association data and hot topic data of the target content creator; The task processing status is constructed based on the topic recommendation request and the tool description information; The target model is invoked to process the task processing state, and the model processing result is obtained; If the model processing result includes tool call information, the target data tool is called according to the tool call information to obtain the tool result. The tool result is written into the task processing status, and the target model is called again based on the updated task processing status. If the model processing result includes candidate topic generation information, candidate topic data is generated based on the candidate topic generation information and the tool results accumulated in the task processing state. Output the candidate topic data.

[0006] Another aspect of this application provides a topic recommendation method, applied to a content creator's terminal, the method comprising: Obtain the object identifier of the target content creator; A topic recommendation request is generated based on the object identifier, and the topic recommendation request is sent to the topic recommendation server. Receive structured candidate topic data returned by the topic recommendation server. The structured candidate topic data includes multiple candidate topic entries. Each candidate topic entry includes a candidate topic title, recommendation basis, content generation angle, and / or exclusive content adaptation information. Read the candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information corresponding to each candidate topic item from the structured candidate topic data, so as to generate candidate topic display information corresponding to each candidate topic item; The control displays the candidate topic information.

[0007] Another aspect of this application provides a topic recommendation device applied to a server, the device comprising: The receiving module is used to receive topic recommendation requests, wherein the topic recommendation request includes the object identifier of the target content creator; The module is used to obtain the tool description information corresponding to each of the multiple data tools, and to construct the task processing status based on the topic recommendation request and the tool description information. The multiple data tools are used to obtain the content creator association data and hot topic data of the target content creator. The processing module is used to call the target model to process the task processing status and obtain the model processing result; The calling module is used to call the target data tool according to the tool calling information when the model processing result includes tool calling information, obtain the tool result, and write the tool result into the task processing status; The processing module is also used to invoke the target model again based on the updated task processing status; The generation module is used to generate candidate topic data based on the candidate topic generation information and the tool results accumulated in the task processing state, when the model processing result includes candidate topic generation information. The output module is used to output the candidate topic data.

[0008] Another aspect of this application provides a topic recommendation device, applied to a content creator's terminal, the device comprising: The acquisition module is used to obtain the object identifier of the target content creator; The sending module is used to generate a topic recommendation request based on the object identifier and send the topic recommendation request to the topic recommendation server. The receiving module is used to receive structured candidate topic data returned by the topic recommendation server. The structured candidate topic data includes multiple candidate topic entries, and each candidate topic entry includes a candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information. The processing module is used to read the candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information corresponding to each candidate topic item from the structured candidate topic data, so as to generate candidate topic display information corresponding to each candidate topic item; The control module is used to control the display of the candidate topic information.

[0009] Another aspect of this application provides a computer device, including: At least one processor; and A memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.

[0010] Another aspect of this application provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method described above.

[0011] Another aspect of this application provides a computer program product including a computer program that, when executed by a processor, implements the method described above.

[0012] The embodiments of this application employing the above technical solution may include the following advantages: by constructing a task processing state on the server side, writing the tool results obtained from each round of tool calls into the same task processing state, and calling the target model again based on the updated task processing state, the acquired multi-source data can be retained and reused during continuous model processing; when the model processing results include candidate topic generation information, candidate topic data is generated by combining the accumulated tool results. Therefore, data state breaks and repeated construction of acquired data in cross-round processing can be reduced, improving the continuity of multi-source data processing and the stability of the candidate topic data generation process. Attached Figure Description

[0013] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of the claims. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0014] Figure 1 The diagram illustrates the operating environment of the topic selection and recommendation method according to Embodiment 1 of this application; Figure 2 A flowchart illustrating the topic selection and recommendation method according to Embodiment 1 of this application is shown schematically. Figure 3 This schematically illustrates a closed-loop flowchart of topic selection recommendation and evaluation according to Embodiment 1 of this application; Figure 4 This schematically illustrates a state transition diagram for reliable execution of an asynchronous task according to Embodiment 1 of this application; Figure 5 A flowchart illustrating the topic selection and recommendation method according to Embodiment 2 of this application is shown schematically. Figure 6 The client graphical interface according to Embodiment 2 of this application is illustrated schematically; Figure 7 A block diagram of a topic recommendation device according to Embodiment 3 of this application is shown schematically; Figure 8 A block diagram of a topic selection recommendation device according to Embodiment 4 of this application is shown schematically; Figure 9 A schematic diagram of the hardware architecture of a computer device according to Embodiment 5 of this application is shown. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0016] It should be noted that the descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0017] It should be noted that the object identification data, content data, interaction data, status data, and other related data involved in the embodiments of this application can be data obtained and processed under the premise of complying with laws, regulations, and relevant normative requirements. Related data can be data that has undergone authorization, anonymization, de-identification, or aggregation processing. The processes of acquiring, storing, transmitting, using, and deleting related data can employ methods such as access control, encrypted transmission, access control, logging, data anonymization, or minimal necessary processing to reduce security risks during data processing.

[0018] In the description of this application, it should be understood that the numerical labels preceding the steps do not indicate the absolute order of execution, but are only used for the convenience of describing this application and distinguishing different processing steps, and therefore should not be construed as a limitation of this application. Where the technical logic allows, some steps may be executed in parallel, interleaved, or in different execution orders.

[0019] First, some of the terms used in the embodiments of this application will be explained.

[0020] A target content creator can refer to an entity that generates, edits, or publishes content within a content delivery system. Each target content creator can correspond to a single content creation account.

[0021] An object identifier can refer to data used to distinguish a target content creator in a data processing system. In some embodiments, the object identifier can be a positive integer identifier corresponding to a content creation account. The object identifier can also take other forms of identification that can uniquely or unambiguously identify the target content creator in a topic recommendation server.

[0022] Content creator-related data refers to data related to the content creation, publication, interaction, or service status of a target content creator. This data can include content creator characteristic data, historical content data, recent content data, statistics on followed users, dedicated content service status, and content publication restrictions.

[0023] A dedicated content service can refer to a content service used to control the access scope of partial content, content fragments, supplementary materials, or interactive resources. For example, a dedicated content service may correspond to a content service that requires a target access object to meet a preset service status before accessing the full version of the content. The specific access rules of a dedicated content service do not constitute a limitation of this application. In some embodiments, a dedicated content service can control the accessibility of a target content resource through content access status. For example, a target content resource may include a public content portion and a dedicated content portion controlled by the access status. When the service status corresponding to the target access object meets the preset access conditions, the full version of the target content resource can be provided to the target access object; when the service status does not meet the preset access conditions, only the public content portion, content summary, preview fragment, or prompt data indicating the existence of dedicated content can be provided. A dedicated content service can also be used to control the access scope of supplementary content resources, supplementary material resources, or interactive resources. The above content forms are only used to illustrate possible implementations of access control and are not intended to limit the inclusion of specific business rights in a dedicated content service.

[0024] Trending data refers to data used to characterize the level of attention, dissemination, or interaction surrounding a content topic within a target timeframe. Trending data can originate from content rankings, search data, comment data, bullet screen data, online community data, industry news data, or manually maintained data.

[0025] Data tools can refer to program modules, service interfaces, or processing nodes that can obtain target data, perform data queries, or provide data processing results based on tool call parameters. Data tools can be local tools within a server or remote tools invoked over a network.

[0026] Tool description information may include tool identifier, tool function description, input parameter structure, output data structure or calling constraints, which are used to enable the target model to identify the data tools that can be called and how to call them.

[0027] Processing rules refer to rule information used to constrain the processing objectives, data acquisition order, output content, or output format of the target model. The main processing rule can contain the complete rule content of the current task, while other available processing rules can be provided in the form of summary information.

[0028] Task processing status refers to the state data continuously saved during a topic recommendation process. Task processing status can include request information, processing rules, tool description information, model messages, tool invocation information, tool results, processing rounds, and candidate topic generation information, etc.

[0029] The target model can be a model capable of generating information based on input messages or by calling information from a tool or by generating candidate topics. The target model can be a large language model, a generative model adjusted according to instructions, or a generative model that supports the use of other tools. This application does not limit the specific network structure, training method, or model parameters of the target model.

[0030] Tool invocation information may include the tool identifier of the target data tool and the tool invocation parameters. Tool invocation information may also include invocation batch, invocation order, or multiple tool invocation tasks.

[0031] Candidate topic data can refer to data used to represent at least one potential content topic and its associated descriptions. Candidate topic data can be stored in the form of objects, arrays, lists, tables, or other structured data formats.

[0032] Task processing status checkpoints can refer to snapshots of task processing status saved during task execution, used for runtime record association, troubleshooting, or task status recovery.

[0033] Processing link records can refer to data that associates and records task processing nodes, node inputs and outputs, tool calls, model processing, and exception information.

[0034] Processing volume usage data can refer to statistical information on the amount of data processed during model processing or tool processing, such as the number of input characters, the number of output characters, the number of model tokens used, or the number of calls.

[0035] Recommended running records can refer to the input data, output data, status data, and link data corresponding to a complete topic selection recommendation running process or one of the processing steps.

[0036] An evaluation instance can refer to a task object created based on an evaluation template and evaluation configuration data, used to evaluate the recommended running process or target processing steps.

[0037] The compare-and-swap operation can refer to an atomic update operation that updates the target data to the target state if the current value of the target data satisfies the preset prerequisite state.

[0038] Secondly, to facilitate understanding of the technical solutions provided in the embodiments of this application by those skilled in the art, the relevant technologies are described below: With the development of large language model technology, artificial intelligence-assisted content creation has gradually become an important capability of content platforms. For example, based on the creator's active input of the topic being conceived, artificial intelligence tools can intelligently generate outlines, cover titles, and reference videos by integrating the creator's historical creations and platform trends; by identifying the creator's main submission categories and mining the hot topics and videos within the corresponding categories, the platform can provide the creator with inspiration for topic selection; and general large language models can directly generate corresponding candidate topics. However, the inventors have learned that existing topic recommendation technologies still have the following shortcomings: First, they lack personalized recommendation strategies. Target content creators in different content categories and with different target audiences have significant differences in content positioning, creative direction, and operational needs. For example, science and technology creators and humanities and history creators typically have different topic selection directions and recommendation strategies. Existing topic recommendation technologies cannot provide differentiated topic recommendations based on the characteristics of different content creators, and cannot meet personalized needs. Second, existing topic recommendation technologies cannot meet the needs of creation scenarios for exclusive content services. Unlike regular content, which primarily aims to increase viewership and audience engagement, the selection of topics for personalized content services requires comprehensive consideration of factors such as content depth, the audience's willingness to subscribe to personalized content services, and the relationship with regular content. For target content creators, the difficulty in selecting topics is one of the main pain points in the creation process of personalized content services, and existing topic recommendation technologies are insufficient to provide effective support for them. Third, general AI recommendations lack product strategy constraints, resulting in poor controllability of candidate topic data. General AI typically generates candidate topic data directly based on audience input, which can easily lead to problems such as unclear recommendation sources, unstable analysis processes, and a lack of data support for recommendations, resulting in low rationality and credibility of candidate topic data. In addition, because general AI lacks an understanding of platform community rules and operational strategies, it may also generate risky candidate topic data that does not meet the requirements of the platform community ecosystem, thereby affecting the quality and usability of candidate topic data.

[0039] To address this, this application provides a topic recommendation method. This method automatically acquires multi-source data, including target content creator feature data, historical content data, information on followed objects, the activation status of dedicated content services, and trending data. It then combines a target model with tool calls to iteratively complete multi-source data acquisition, candidate topic generation, and structured processing, enabling personalized topic recommendations for different content categories, styles, and creation stages. Furthermore, the method incorporates dedicated content service adaptability into the topic recommendation process. It analyzes factors such as the activation status of the target content creator's dedicated content service, content depth, and the potential for expansion of the dedicated content service, providing strategy interpretation and adaptability scores for each candidate topic. This achieves topic recommendations tailored to dedicated content service creation scenarios, improving the matching degree between candidate topic data and the needs of dedicated content creation services. Furthermore, a tool-calling loop ensures the reasoning process is based on solid evidence, and structured extraction and data validation guarantee the standardized format of the recommendation results, facilitating subsequent processing. An asynchronous, reliable execution state machine ensures the reliability of task execution. Simultaneously, a dual-channel evaluation loop combining content security assessment and backtesting of content data from target content creators is constructed to verify the content security of the recommendation results and match them with the actual creation trajectory of content creators. This achieves secure interception, ranking correction, and continuous optimization of candidate topic data, thereby improving the security, stability, and traceability of candidate topic data. See below for details.

[0040] Finally, for ease of understanding, an exemplary operating environment is provided below.

[0041] like Figure 1 As shown, the operating environment may include a topic recommendation server 100, a content creator terminal 200, a data and tool service 300, a model service 400, a task scheduling and state storage system 500, an evaluation system 600, and a network 700.

[0042] The topic recommendation server 100 can communicate with the content creator terminal 200, data and tool service 300, model service 400, task scheduling and state storage system 500 and evaluation system 600 via network 700.

[0043] The content creator terminal 200 can be used to send topic recommendation requests, query asynchronous task status, and receive candidate topic data. The content creator terminal 200 can be a smartphone, tablet, personal computer, workstation, or other electronic device capable of running applications.

[0044] The topic recommendation server 100 is the main body for executing the method of this application. It can be used to receive topic recommendation requests, construct task processing status, call model service 400, call data and tool service 300, generate candidate topic data, and perform structured processing, sorting and output control.

[0045] Data and Tools Service 300 may include account data services, content data services, follower data services, personalized content status data services, trending data services, and local processing tools. Different data tools within Data and Tools Service 300 can access the topic recommendation server 100 through a unified tool call interface.

[0046] Model service 400 can be used to perform target model inference or evaluation model inference. Model service 400 can be deployed on the same computer device as topic recommendation server 100, or it can be deployed on a separate server or server cluster.

[0047] The task scheduling and state storage system 500 can be used to store asynchronous task records, publish task messages, save task processing status checkpoints, and maintain an index of tasks to be published. The task scheduling and state storage system 500 can be implemented using at least one of a task queue, a key-value store, a relational database, an object storage system, or other data storage systems.

[0048] The evaluation system 600 can be used to create evaluation instances, extract evaluation variables, call evaluation models, verify evaluation outputs, and record evaluation results. The evaluation system 600 can be deployed as a standalone service or implemented as a functional module within the topic recommendation server 100.

[0049] It should be noted that the above-mentioned devices, services, modules and their quantities are only examples, and can be added, removed, split, merged or replaced according to actual needs in different embodiments.

[0050] The technical solution of this application will be described below through multiple embodiments, using the topic recommendation server 100 as the execution entity.

[0051] Example 1 Figure 2 A flowchart illustrating the topic selection and recommendation method according to Embodiment 1 of this application is shown schematically.

[0052] like Figure 2 As shown, the topic selection recommendation method may include steps S200 to S212, wherein: Step S200: Receive a topic recommendation request, wherein the topic recommendation request includes the object identifier of the target content creator.

[0053] Step S202: Obtain tool description information corresponding to each of the multiple data tools, wherein the multiple data tools are used to obtain content creator association data and hot topic data of the target content creator.

[0054] Step S204: Construct a task processing status based on the topic recommendation request and the tool description information.

[0055] Step S206: Call the target model to process the task processing state and obtain the model processing result.

[0056] Step S208: If the model processing result includes tool call information, the target data tool is called according to the tool call information to obtain the tool result. The tool result is written into the task processing status, and the target model is called again based on the updated task processing status.

[0057] Step S210: If the model processing result includes candidate topic generation information, generate candidate topic data based on the candidate topic generation information and the tool results accumulated in the task processing state.

[0058] Step S212: Output the candidate topic data.

[0059] In this embodiment, the topic recommendation server does not directly generate topics based solely on object identifiers or single text inputs. Instead, it first constructs a continuously updated task processing state, and then the target model determines whether to invoke the data tool based on this task processing state. The tool results returned by the data tool are written into the same task processing state and participate in the next round of model processing. Thus, the target model can gradually acquire content creator-related data and trending data during the processing, and determine whether to continue invoking the tool or generating candidate topic data based on the acquired data.

[0060] In this way, the task processing status forms a data closed loop between the model processing node and the tool calling node, so that the tool results obtained from different data sources can continuously affect the subsequent model processing, reduce the lack of data evidence caused by insufficient single input information or discontinuous data status across rounds, and improve the reusability of multi-source data in the candidate topic generation process.

[0061] The following combination Figure 2 The steps in steps S200 to S212, as well as other optional steps, are described in detail.

[0062] Step S200 The system receives a topic recommendation request, which includes the object identifier of the target content creator.

[0063] Topic recommendation requests can be received via synchronous or asynchronous interfaces. A topic recommendation request must at least carry an object identifier to distinguish the target content creator, and may also carry task type, task processing mode, checkpoint storage backend parameters, or other parameters.

[0064] In some embodiments, the object identifier is an identifier data in the form of a positive integer. After receiving the object identifier, the topic recommendation server 100 can perform data type validation and numerical validity validation on the object identifier. If the validation passes, the object identifier can be written into the context data of this task. If the validation fails, a processing result indicating an error in the request parameter can be returned, without proceeding to the subsequent model processing flow.

[0065] Synchronous request processing is suitable for scenarios where candidate topic data needs to be returned in the current request connection. Asynchronous request processing is suitable for scenarios with many rounds of tool calls, long model processing times, or a large number of concurrent requests.

[0066] Step S202 The document describes the acquisition of tool descriptions for multiple data tools, wherein the multiple data tools are used to acquire content creator association data and hot topic data of the target content creator.

[0067] Tool description information can be obtained from the tool registry, tool gateway, or tool configuration file. Each data tool can correspond to a tool identifier, function description, parameter structure, and output structure. The target model can identify which data tools can obtain account characteristics, recent content, followed object data, personalized content service status, or trending data based on the tool description information.

[0068] Local data tools can be directly accessed by the topic recommendation server 100. For example, local data tools may include a rule loading tool for loading processing rules and a hotspot aggregation tool for obtaining hotspot data based on content categories. Remote data tools can access account data services, content data services, follower statistics services, service status data sources, or hotspot data sources through the unified tool gateway.

[0069] In some embodiments, local data tools may have a pre-configured internal tool identifier. The internal tool identifier may be a string identifier, a numeric identifier, or a composite identifier that can uniquely identify the corresponding data tool.

[0070] Step S204 The task processing status is constructed based on the topic recommendation request and the tool description information.

[0071] Task processing status can be stored using a status object, a message list, or a graph execution status. When constructing a task processing status, topic recommendation requests, object identifiers, tool descriptions, processing rules, system constraints, content category hotspot configurations, and current model messages can be written into the status object. The task processing status can also be associated with a processing volume statistician, a processing link recorder, object identifier context variables, and task processing status checkpoints. By associating these objects, the same object identifier can be passed between processing nodes, node inputs and outputs can be recorded, and task status snapshots can be saved when needed.

[0072] In an optional embodiment, constructing a task processing status based on the topic recommendation request and the tool description information includes: Step S300: Based on the task type corresponding to the topic recommendation request, determine the main processing rule and at least one available processing rule from the processing rule registry.

[0073] Step S302: Obtain the rule content of the main processing rule and the summary information of the at least one available processing rule.

[0074] Step S304: Obtain the hotspot data source configuration information corresponding to each of the multiple content categories.

[0075] Step S306: Construct the task processing status based on the rule content, the summary information, the hotspot data source configuration information, and the tool description information.

[0076] The processing rule registry can store processing rule identifiers, task types, rule descriptions, and the storage location of rule content. The main processing rule can be the processing rule directly corresponding to the currently selected topic recommendation task. Other available processing rules can be extended rules that the target model can further read when needed. Writing the complete content of the main processing rule to the task processing status allows the target model to obtain the main processing objective of the current task; writing other available processing rules in summary form to the task processing status controls the initial context length and allows the target model to choose whether to further load corresponding rules based on current processing needs.

[0077] Hotspot data source configuration information can be mapped according to content categories. For example, different content categories can each be configured with one or more hotspot data source nodes. Hotspot data source configuration information can include content category identifiers, data source node identifiers, and node priorities.

[0078] The above method allows for the dynamic selection of processing rules corresponding to the task type within a single task, while simultaneously providing the target model with a range of available hot data sources. This ensures that the task processing status takes into account both the current task constraints and content category differences, reducing redundancy caused by indiscriminately writing all rule content and all hot data sources into the model context.

[0079] In some embodiments, the content creator-related data includes at least one of the following: content creator characteristic data, content category data, historical content data, recent content data, follower size data, follower hierarchy data, follower interaction data, dedicated content service activation status, dedicated content service restriction status, content publishing restriction status, and historical content performance data. The above data can be implemented through step S400. In step S400, the topic recommendation server 100 can invoke one or more data tools based on the object identifier to obtain at least one type of data from the content creator-related data. Content creator characteristic data may include account name, account description, main content category, or other data used to characterize content creation features. Historical content data may include the title, tags, description, content category, and publication time of published content. Recent content data may be a preset number of content data or content within a preset time range that is relatively close to the current time. Follower size data may represent the number of objects following the target content creator. Follower hierarchy data may represent the statistical results corresponding to different follower hierarchy levels obtained according to predetermined hierarchy rules. Follower interaction data may include aggregated statistical data corresponding to comments, bullet comments, favorites, resource interactions, or other interactive behaviors. The "Exclusive Content Service Activation" status indicates whether the target content creator has the capability to publish exclusive content. The "Exclusive Content Service Restriction" status indicates whether the exclusive content service is currently restricted. The "Content Publishing Restriction" status indicates whether the target content creator's content publishing function is limited. Historical content performance data can include playback data, interaction data, collection data, resource interaction data, comment data, and exclusive content response data. Not all types of data need to be obtained in every topic recommendation process; the target model can select the necessary data tools based on the current task processing status. By limiting different categories of data to content creator-related data, the model can process inputs covering multiple data dimensions such as content category, historical content, recent content, target audience characteristics, and service status, thereby reducing input bias caused by generating topics solely based on a single hot topic or historical content.

[0080] In optional embodiments, the content creator association data and the trending data are obtained through at least one of the following methods: Step S500: In response to the topic recommendation request, the data tool is invoked to obtain the target data.

[0081] Step S502: Collect target data in advance according to a preset cycle and store it in the feature data storage space.

[0082] Step S504: Obtain the target data from the vector data storage space or offline feature service.

[0083] Step S506: Obtain target data through a hybrid feature service combining real-time and offline data.

[0084] The target data refers to either the content creator's associated data or the trending data. The obtained data is then written into the task processing status.

[0085] Real-time invocation is suitable for scenarios where hot topics change rapidly or service status requires immediate confirmation. Pre-collection can periodically update content creator features, historical content vectors, or statistical features of followed objects. Vector data storage space can be used to filter historical content or hot topic entries based on semantic similarity. Offline feature services can provide aggregated features with longer computation cycles but lower change frequencies. Hybrid feature services can combine offline-maintained content creator features, historical content vectors, and statistical features of followed objects with the latest online hot topic data and service status, balancing data timeliness and response speed. By supporting multiple data acquisition methods, appropriate data paths can be selected based on data change frequency and processing latency requirements, reducing waiting time caused by real-time remote invocation of all data while maintaining the ability to acquire the latest hot topic and status data.

[0086] In an optional embodiment, the method for obtaining the hotspot data includes: Step S600: Determine the target content category corresponding to the target content creator based on the content creator association data.

[0087] Step S602: Based on the target content category, determine multiple target hotspot data sources from the hotspot data source configuration information.

[0088] Step S604: In parallel, call the data tools corresponding to the multiple target hotspot data sources to obtain multiple hotspot data results.

[0089] Step S606: Aggregate the multiple hotspot data results to obtain hotspot data corresponding to the target content category.

[0090] The target content category can be determined based on the account's main content category, historical content distribution, or recent content distribution. When multiple content categories exist, the main content category can be selected, or multiple candidate content categories can be retained and their respective trending data obtained. Target trending data sources can include internal rankings from content provision systems, search term data sources, high-frequency comment term data sources, high-frequency bullet screen term data sources, online community trending data sources, industry news data sources, vertical domain information sources, or manually maintained trending data sets. When aggregating multiple trending data results, fields from different data sources can be mapped to unified fields, such as trending title, source identifier, ranking, publication time, popularity metric, and content category. Aggregation results can be written back to the task processing status using a unified data object or structured array. By selecting trending data sources according to the target content category and performing parallel calls and unified aggregation on multiple data sources, the amount of trending data unrelated to the target content creator's content category entering the model context can be reduced, while also decreasing the cumulative waiting time caused by serial calls to multiple data sources.

[0091] In some embodiments, the hotspot data source configuration information can be statically configured or dynamically generated based on the target content creator's associated data. When dynamically determining the target hotspot data source, the matching degree of the data source corresponding to each of multiple candidate hotspot data sources can be determined based on at least one of the target content creator's target content category, historical content distribution, recent content performance data, and interaction data of the followed objects; then, based on the data source matching degree, one or more target hotspot data sources are selected from the multiple candidate hotspot data sources. In this way, when the content direction or recent content performance of the target content creator changes, the collection scope of hotspot data can be adjusted accordingly, reducing the situation where fixed hotspot data sources do not match the current content direction.

[0092] Step S206 The target model is invoked to process the task processing state, and the model processing result is obtained.

[0093] When invoking the target model, system rules, current messages, descriptions of available tools, and obtained tool results from the task processing status can be converted into model input. The target model can output ordinary candidate topic text, or it can output tool invocation information including tool identifiers and tool invocation parameters.

[0094] When the target model outputs tool call information, it indicates that the data in the current task processing state is insufficient to generate candidate topics, or that the target model needs to obtain specific types of data. When the target model outputs candidate topic generation information, it indicates that the target model has generated candidate topic content that can be structured and processed based on the current state.

[0095] It should be noted that this application does not limit the target model to a specific model structure. The target model can be a generative model that supports function calls, tool calls, or structured instruction output. Model inference can be performed locally by the topic recommendation server 100 or through the model service 400.

[0096] Step S208 If the model processing result includes tool call information, the target data tool is called according to the tool call information to obtain the tool result. The tool result is written into the task processing status, and the target model is called again based on the updated task processing status.

[0097] The topic recommendation server 100 can parse tool call information from the model processing results to determine the required data tools and call parameters. After the tool results are returned, they can be encapsulated as tool result messages and appended to the task processing status. The updated task processing status is then input into the target model, enabling it to read the previous round of tool results and continue processing. Model processing and tool calls can be looped multiple times. For example, the target model can first call the account data tool to obtain content creator characteristics, then call the recent content tool to obtain historical content, and then call the follower data tool and the hot topic data tool. The results of each round of tools can be retained in the task processing status.

[0098] In an optional embodiment, the target data tool is invoked based on the tool invocation information to obtain the tool result, including: Step S700: Extract the tool identifier and tool call parameters from the tool call information.

[0099] Step S702: Determine the target data tool based on the tool identifier.

[0100] Step S704: Send the tool call parameters to the target data tool through the unified tool distribution interface.

[0101] Step S706: If the tool call information includes multiple tool call tasks, submit the multiple tool call tasks to the concurrent execution resource, and copy the context data corresponding to the target content creator for each tool call task.

[0102] Step S708: Aggregate the tool results returned by each of the multiple tool call tasks.

[0103] The unified tool distribution interface can distinguish between local and remote data tools based on their registration information. Local data tools can execute within the current process. Remote data tools can be invoked through hypertext transfer interfaces, remote procedure call interfaces, graph query interfaces, database query interfaces, data warehouse query interfaces, message subscription interfaces, or feature service interfaces.

[0104] Concurrent execution resources can be thread pools, asynchronous coroutine executors, batch remote call components, or server-side aggregation components. Context data can include object identifiers, task identifiers, or processing chain identifiers. Copying context data for each tool invocation task allows each concurrently executing data tool to obtain data corresponding to the current target content creator, without needing to repeatedly construct the entire context in each tool invocation parameter.

[0105] Tool result aggregation can form a unified result object according to the tool call order, tool identifier, or data category. When a single tool call fails, the corresponding error information can be written into the task processing status as part of the tool result, enabling the target model to decide whether to use other tools or continue processing based on existing data.

[0106] By uniformly distributing local and remote data tools and executing multiple tool invocation tasks concurrently, the time for acquiring multi-source data can be shortened while maintaining consistency in tool invocation interfaces. Furthermore, by replicating context data, the possibility of object identifier loss or mismatch between concurrent tasks can be reduced.

[0107] In some embodiments, multiple tool invocation tasks can be executed concurrently via thread pools, asynchronous coroutines, batch remote calls, streaming tool invocations, or server-side aggregation tools. For tool invocation tasks that can be split into multiple independent data shards, each data shard can be processed separately and the results of multiple shards can be aggregated. When the same remote service supports batch requests, multiple tool invocation parameters can be combined into a batch request to obtain multiple tool results at once. For tool services that can return results incrementally, the corresponding results can be written to the task processing status when partial tool results are received. Through these methods, the concurrent execution method can be selected based on the interface capabilities and data scale of the data tools, reducing the cumulative waiting time caused by the serial execution of multiple data tools.

[0108] In an optional embodiment, the method further includes: Step S800: Record the cumulative number of tool call rounds corresponding to the target model.

[0109] Step S802: When the cumulative number of tool calls reaches the preset upper limit, write the result output constraint into the task processing status.

[0110] Step S804: Based on the task processing state with the stated output constraints, call the target model to obtain the candidate topic generation information.

[0111] The cumulative number of tool call rounds can be increased each time the model processes results, including tool call information. Output constraints can indicate when to stop further calls to data tools and generate candidate topics based on the currently obtained data.

[0112] In one exemplary configuration, the preset maximum number of rounds can be 15, the timeout for a single remote data tool call can be 60 seconds, and the timeout for a single target model request can be 300 seconds. These values ​​are only one engineering configuration and can be adjusted based on model processing latency, the number of tools, and available runtime resources.

[0113] When a data tool times out or returns an error, the abnormal tool result can be written to the task processing status. The target model can choose to call an alternative tool, ignore the corresponding data, or generate candidate topics based on the existing tool results.

[0114] By setting an upper limit on the number of tool call rounds, the iterative processing between the model and the data tool can be completed within a limited processing range, reducing the situation where tool call loops occupy computing resources for a long time, while preserving the target model's ability to generate results based on the obtained data.

[0115] In some embodiments, the processing method for determining the target data tool is not limited to dynamically outputting tool invocation information by the target model. Alternatively, the tool invocation order can be generated according to pre-configured tool invocation rules; the target data tool can be determined based on the task processing status using decision trees or classification models; the tool invocation action can be determined using a strategy model obtained from historical recommendation feedback; or a planning processing component can first generate a tool invocation plan, and then the execution processing component can execute the tool invocation plan item by item. In other embodiments, content creator feature data, recent content data, attention object data, and hot topic data can be obtained sequentially according to a predetermined data acquisition order, and then the aggregated data can be provided to the target model. These alternative methods change the scheduling method of the target data tool but do not change the processing objective of supporting candidate topic generation through multi-source data.

[0116] As an alternative, iterative processing between models and data tools can be organized using state diagrams, workflow engines, directed acyclic graphs of tasks, combinations of planners and executors, or custom task runtime. Multiple processing roles can also handle content creator characteristics, hot data, specific content adaptation information, and title generation separately, with the results integrated and processed nodes generating candidate topics. These alternative methods do not change the processing objective of supporting candidate topic generation through multi-source data.

[0117] Step S210 If the model processing result includes candidate topic generation information, candidate topic data is generated based on the candidate topic generation information and the tool results accumulated in the task processing state.

[0118] The candidate topic generation information can be free text, structured text, or intermediate results output by the target model. The topic recommendation server 100 can associate the candidate topic generation information with the content creator association data and hot topic data accumulated in the task processing status to generate candidate topic data.

[0119] When generating candidate topics, each candidate topic can be associated with corresponding recommendation criteria. For example, recommendation criteria can point to recent content from the target content creator, trending items corresponding to the target content category, or statistical characteristics of the target audience. Candidate topic data can also include data fields for subsequent structured processing and sorting.

[0120] In some embodiments, the candidate topic data includes multiple candidate topics and at least two of the following information corresponding to each candidate topic: topic title, recommendation basis, content generation perspective, title example, related hot topic information, estimated popularity information, exclusive content adaptation information, and exclusive content generation strategy information. The above scheme can be implemented through step S900. In step S900, a candidate topic data object can be created for each candidate topic, and at least two types of related information can be written into the corresponding candidate topic data object. The topic title can be used to characterize the theme of the candidate topic. The recommendation basis can be used to record the content creator-related data or hot topic data used to generate the candidate topic. The content generation perspective can be used to indicate the content organization direction of the candidate topic. The title example can be used as candidate title data for subsequent processing or display. Related hot topic information can include hot topic entry titles, hot topic data source identifiers, or ranking lists. Estimated popularity information can be represented by enumeration values, level values, or numerical values. Exclusive content adaptation information can indicate the degree of adaptation between the candidate topic and the exclusive content format. The proprietary content generation strategy information can indicate how candidate topics can be expanded into in-depth content, series content, Q&A content, resource content, real-time interactive extended content, or other proprietary content formats. By associating recommendation criteria, trending information, and content generation information with candidate topics, the candidate topic data can include not only topic text but also the data fields required for subsequent structured reading, sorting, evaluation, or interface display.

[0121] In an optional embodiment, the candidate topic data is structured candidate topic data; the method for generating the candidate topic data includes: Step S1000: Obtain the accumulated raw results of the tools in the task processing state and the candidate topic text generated by the target model.

[0122] Step S1002: Input the original results of the tool and the candidate topic text into the structured processing module.

[0123] Step S1004: Based on the preset candidate topic data structure, the original results of the tool and the candidate topic text are extracted in a structured manner to obtain the structured candidate topic data.

[0124] The structured processing module can call the same target model, independent structured models, or information extraction components. By inputting the raw results of the tool into the structured processing module, the module can supplement fields such as object identifiers, content categories, or service status from the tool results when the final text of the target model is missing some fields.

[0125] The preset candidate topic data structure may include a content creator information field, a recommendation criteria field, and a candidate topic list field. Each field can be represented by a string identifier, a numeric identifier, or a predefined field code; the specific field names do not constitute a limitation on the preset candidate topic data structure.

[0126] Structured candidate topic data can take the form of structured objects, key-value pair collections, structured data, or tabular records. The structured processing module can split multiple candidate topics in the candidate topic text into an array of objects, and write the title, basis, perspective, title example, and adaptation information of each candidate topic into the corresponding fields.

[0127] By simultaneously utilizing both the tool's original results and candidate topic texts to perform structured extraction, the field gaps caused by the lack of contextual information when extracting fields solely from free text can be reduced, and the candidate topic data can be stably read by terminals, evaluation modules, or other downstream programs.

[0128] In an optional embodiment, the method further includes: Step S1100: Based on the preset candidate topic data structure, perform data parsing and field validation on the structured candidate topic data.

[0129] Step S1102: If the structured candidate topic data fails data parsing or field validation, generate validation feedback information to characterize the reason for the failure.

[0130] Step S1104: Write the verification feedback information into the task processing status and call the target model to generate corrected candidate topic data.

[0131] Step S1106: Re-execute data parsing and field validation on the corrected candidate topic data until the field validation is passed or the preset retry conditions are met.

[0132] The field validation may include at least one of the following: object identifier field validation, content category field validation, service status field validation, popularity field validation, required field validation, field data type validation, numerical validity validation, and field length validation.

[0133] The object identifier field validation verifies whether the field is a valid object identifier. The content category field validation and service status field validation verify whether the corresponding value belongs to a preset enumeration set. The popularity field validation verifies whether the popularity information is a preset level or a valid value. Required field validation determines whether necessary fields such as candidate topic titles and recommendation criteria are missing.

[0134] Validation feedback can include the location of data parsing failure, field name, expected data type, actual data type, illegal values, or field length information. The target model can correct only the corresponding fields based on the validation feedback without regenerating all tool results.

[0135] In one exemplary configuration, the maximum number of structured corrections can be set to 10. If the validation still fails after reaching the maximum number of structured corrections, a structured processing failure message can be returned.

[0136] Structured output methods can also employ data structure constraint output, function call constraint output, syntax constraint decoding, finite state parsers, or traditional information extraction models. Corrections following validation failures can be performed by the original target model, a lightweight correction model, rule repair components, field completion components, or a review workbench.

[0137] By writing the reason for parsing or field validation failure back to the task processing status and making targeted corrections, the structured results can gradually meet the preset data structure within a limited retry range, reducing downstream processing anomalies caused by format errors or illegal fields.

[0138] In an optional embodiment, generating the candidate topic data includes: Step S1200: For multiple candidate topics to be sorted, obtain the following information: hot topic relevance, content creator matching degree, exclusive content suitability, target audience suitability, content sample matching degree, duplication risk information, and content security risk information.

[0139] Step S1202: The hot topic relevance, content creator matching degree, exclusive content adaptation degree, target audience adaptation degree and content sample matching degree are fused together, and the fusion result is corrected according to the duplication risk information and the content security risk information to obtain the topic score corresponding to each of the candidate topics to be ranked.

[0140] Step S1204: Determine the order of the multiple candidate topics to be sorted based on the topic selection score.

[0141] The aforementioned rating dimensions can be output by the target model or calculated by independent rating components based on tool results. Each rating dimension can be normalized first to allow ratings with different dimensions to participate in the fusion.

[0142] In one exemplary implementation, let the target content creator be u, the candidate topic be t, and the target content category be c. The candidate topic score can be expressed as: TopicScore(t,u)=α×H(t,c)+β×M(t,u)+γ×C(t,u)+δ×F(t,u)+η×B(t,u)-ε×R(t,u)-ζ×S(t).

[0143] Wherein, H(t,c) represents the relevance to trending topics, M(t,u) represents the matching degree between content creators, C(t,u) represents the adaptation degree of specific content, F(t,u) represents the adaptation degree of the target audience, B(t,u) represents the matching degree of content samples, R(t,u) represents the correction value corresponding to duplicate risk information, and S(t) represents the correction value corresponding to content security risk information. α, β, γ, δ, η, ε, and ζ represent the fusion parameters for the corresponding scoring dimensions. These fusion parameters can be configured in a fixed manner, by content category, by the scale of the target audience, by the content creation stage, or updated based on feedback data.

[0144] In some embodiments, the fusion parameters corresponding to different scoring dimensions can be configured according to the processing objectives of the candidate topics. When it is necessary to improve the correlation between candidate topics and current hot topics, the fusion parameters corresponding to hot topic relevance and target audience fit can be increased; when it is necessary to improve the fit between candidate topics and specific content formats, the fusion parameters corresponding to specific content fit can be increased; when it is necessary to improve the consistency between candidate topics and the existing content direction of the target content creator, the fusion parameters corresponding to content creator matching degree and content sample matching degree can be increased, and the correction degree of duplication risk information on the scoring results can be improved; when the content security risk corresponding to the target content category is high, the correction parameters corresponding to content security risk information can be increased, or a hard threshold can be set for content security risk information. The fusion parameters can also be configured according to the content creation stage of the target content creator, or automatically updated based on posterior content samples and evaluation results. Through the above methods, the degree of influence of different scoring signals on the ranking order can be adjusted without changing the meaning of each scoring dimension.

[0145] By fusing positive matching signals and using duplicate risk and content security risk to correct the score, the ranking of candidate topics can simultaneously consider hot topic relevance, existing content from content creators, adaptation of exclusive content, and risk information, reducing the excessive influence of a single popularity indicator on the ranking results.

[0146] In an optional embodiment, the method for obtaining the hotspot relevance includes: Step S1300: Obtain the text matching degree between the candidate topics to be sorted and each of the multiple hot items.

[0147] Step S1302: Obtain the data source priority and ranking of each hot item.

[0148] Step S1304: Based on the text matching degree, the data source priority, and the ranking decay parameter corresponding to the ranking of the list, generate the hot spot contribution value corresponding to each hot spot item.

[0149] Step S1306: Aggregate the multiple hotspot contribution values ​​to obtain the hotspot relevance.

[0150] Text matching accuracy can be obtained through text vector similarity, keyword matching accuracy, or the target model's judgment results. Data source priority can be used to characterize the applicability of popular data sources to the target content category. Ranking can indicate the position of popular items within their respective data sources.

[0151] In some embodiments, hotspot relevance can be expressed as: H(t,c)=Normalize(ΣSim(t,h_i)×P(node_i)×RankDecay(rank_i)).

[0152] Where h_i represents the i-th hot topic item, Sim(t,h_i) represents the text matching degree between the candidate topic and the hot topic item, P(node_i) represents the priority of the hot topic data source node, and RankDecay(rank_i) represents the ranking decay parameter corresponding to the ranking. The priority of the priority data source node can be 1.0, and the priority of the ordinary data source node can be 0.6 to 0.8; the ranking decay parameter can be 1 / log2(rank_i+1). The above values ​​and expressions are only implementation examples.

[0153] By using text matching degree, data source priority, and ranking list to determine the contribution value of hot topics, we can distinguish the semantic relevance of different hot topics and their relative importance in the data source, making the hot topic relevance more reflective of the relationship between candidate topics and the current hot topics of the target content category.

[0154] In an optional embodiment, the method for obtaining the content creator matching degree includes: Step S1400: Obtain the semantic matching degree between the candidate topics to be sorted and at least one of the following: content category, content type, historical content title, historical content tag, historical content description, and recent content data of the target content creator.

[0155] Step S1402: Determine the content creator matching degree based on the semantic matching degree.

[0156] The content category, content type, and historical content text of the target content creator can be combined into a content creator description text. Then, the semantic match degree between the candidate topics and the content creator description text can be calculated. The content creator match degree can be expressed as: M(t,u)=Normalize(Sim(topic_text(t),profile_text(u)+recent_archives(u))).

[0157] Where topic_text(t) represents candidate topic text, profile_text(u) represents content creator feature text, and recent_archives(u) represents a collection of recent content texts. The above expression is only used to illustrate one matching method.

[0158] When there is a lot of historical content, we can first use vector filtering to obtain some historical content that is close to the candidate topics, and then calculate the semantic matching degree based on the filtering results, so as to reduce the context inflation caused by all historical content participating in the model processing at the same time.

[0159] By using content categories, historical content, and recent content to determine the matching degree of content creators, the ranking results can take into account the existing content direction of the target content creators, reducing the situation where candidate topics deviate significantly from their long-term content categories or recent creative directions.

[0160] In an optional embodiment, the method may further include: Step S1500: Obtain the exclusive content service status of the target content creator, the content depth corresponding to the candidate topics to be ranked, and the adaptation degree of the exclusive content extension form corresponding to the candidate topics to be ranked. Step S1502: Determine the adaptation degree of the exclusive content based on the status of the exclusive content service, the content depth, and the adaptation degree of the exclusive content extension form; Step S1504: Determine the suitability of the target content creator based on at least one of the following: the size of the target content creator's target audience, the level of the target audience, and the interaction data of the target audience.

[0161] Customized content extensions may include in-depth content, series content, Q&A content, resource content, real-time interactive extended content, or supplementary content. In some embodiments, customized content adaptability can be expressed as: C(t,u)=StateWeight(service_state)×FormatFit(t)×DepthFit(t).

[0162] Among them, StateWeight(service_state) represents the state weight corresponding to the exclusive content service state, FormatFit(t) represents the degree of adaptation between the candidate topic and the exclusive content extension form, and DepthFit(t) represents the degree of content depth adaptation of the candidate topic.

[0163] When the dedicated content service has been activated, a higher state weight can be configured; when it has not been activated, an intermediate state weight can be used and adaptation instructions can be output; when the service is in a restricted state, the corresponding state weight can be reduced.

[0164] In some embodiments, when determining the suitability of exclusive content, the publicly available content generation information and the exclusive content generation information corresponding to the candidate topics can also be obtained, and the degree of correlation between the publicly available content generation information and the exclusive content generation information can be determined. This degree of correlation can characterize whether the exclusive content can provide further data analysis, content expansion, supplementary information, question-and-answer extensions, or serialized content based on the publicly available content, rather than forming a content result that duplicates the publicly available content. The topic recommendation server can use the degree of correlation, along with the content depth and the suitability of the exclusive content expansion form, to determine the suitability of the exclusive content. Through this method, it is possible to distinguish between content duplication and content progression during the candidate topic ranking process, enabling the exclusive content suitability information to reflect the data relationship between the publicly available content and the exclusive content.

[0165] The target audience fit score can be used to differentiate the appropriate range of topics for different target audience sizes and interaction characteristics. For example, when the target audience is large, the fit score for broadly covered trending topics can be increased; when the target audience is small, the fit score for topics with high verticality and low competition can be increased. The above configuration is just an example.

[0166] By separately determining the content-specific fit and the target audience fit, the ranking of candidate topics can take into account the current service status of the target content creator and the statistical characteristics of its target audience, avoiding ranking based solely on content history and trending data.

[0167] In an optional embodiment, the method for obtaining the content sample matching degree includes: Step S1600: Obtain the semantic matching degree between the candidate topics to be sorted and the historical content samples or subsequent published content samples of the target content creator; Step S1602: Obtain the content performance data corresponding to the historical content sample or the subsequent published content sample; Step S1604: Determine the content performance weight based on the content performance data; Step S1606: Generate the content sample matching degree based on the semantic matching degree and the content performance weight.

[0168] The post-release content sample can be content actually released by the target content creator after the candidate topics are generated. It is used to verify the relationship between the candidate topics and the actual creative direction in subsequent evaluations or model adjustments. Content performance data can include at least one of the following: playback data, interaction data, collection data, resource interaction data, comment data, and personalized content response data. Content performance data can be normalized first to form content performance weights.

[0169] In some embodiments, the content performance data used to determine content performance weights may further include at least one of the following: content completion statistics, content entry selection statistics, follower relationship change statistics, personalized content access status change statistics, and anomaly feedback statistics. Content completion statistics may correspond to the proportion of target content being fully played or fully read; content entry selection statistics may correspond to the proportion of target content entry points being selected; and anomaly feedback statistics may correspond to the number or proportion of anomaly information reports generated for the target content. These indicators can be normalized separately and then combined with playback data, interaction data, favorites data, comment data, or personalized content response data to determine content performance weights. By introducing posterior statistical data from different processing stages, the effectiveness of real content samples can be characterized from multiple dimensions such as content access, content interaction, and anomaly feedback.

[0170] In some embodiments, the content sample matching degree can be expressed as: B(t,u)=BacktestMatch(t,archives_u)×PerformanceWeight(archives_u).

[0171] Here, BacktestMatch(t,archives_u) represents the semantic matching degree between the candidate topic and the real content sample set, and PerformanceWeight(archives_u) represents the content performance weight determined based on the content performance data.

[0172] By using both semantic matching degree and content performance weight to generate content sample matching degree, samples that are close to the actual content direction of the target content creator and have high content performance can have a greater impact on the ranking of candidate topics.

[0173] In an optional embodiment, the method further includes: Step S1700: Determine the duplication risk information based on at least one of the following: degree of duplication between the candidate topics to be ranked and the recent content data of the target content creator, degree of hot topic saturation, degree of content category matching, degree of content positioning deviation, and degree of adaptation to the exclusive content generation strategy; Step S1702: Based on at least one of the following: topic title, recommendation basis, content generation angle, title example, related hot topic information, and exclusive content generation strategy information of the candidate topics to be ranked, determine the content security risk information; Step S1704: If the content security risk information does not meet the preset content output conditions, remove the corresponding candidate topic from the candidate topic data or adjust the arrangement of the corresponding candidate topic.

[0174] The degree of repetition can be determined by the semantic similarity between candidate topics and recent content titles, tags, or descriptions. The degree of trending topic saturation can be determined based on pre-defined trending topic saturation indicators. The degree of content positioning deviation can be determined based on the differences between candidate topics and the long-term content categories or directions of the target content creator.

[0175] Content security risk information can be generated by evaluation models, classification models, or rule components. Risk dimensions may include expressions that do not meet preset content requirements, antagonistic expressions, false or exaggerated expressions, risks of unauthorized use of content, inappropriate expressions targeting individuals or groups, content unsuitable for minors, exaggerated medical or financial information, and incompatibility with content rules.

[0176] In some embodiments, content security risk information may further include verification results of whether the dedicated content service guidance information conforms to preset output rules. For example, it can be verified whether the dedicated content generation strategy information in the candidate topics contains descriptions that exceed the actual access range of the target content resource, descriptions that are inconsistent with the dedicated content service status, or guidance information that does not meet preset content output requirements. If the verification result does not meet the preset content output conditions, the corresponding descriptions can be removed, rewritten, or processed in descending order. Through the above methods, the dedicated content service-related information in the candidate topics can be kept consistent with the actual service status and preset output rules.

[0177] If the content security assessment results in a Boolean value and the result is "fail", the candidate topic can be removed directly; if the assessment result is a numerical risk score, the ranking position can be lowered based on the risk score.

[0178] By incorporating duplication and content security risks into the topic selection scoring process, candidate topics that are duplicated with recent content, oversaturated with trending topics, or do not meet the preset content output conditions can be sorted, corrected, or filtered, thereby improving the effectiveness and output stability of the final candidate topic set.

[0179] Step S212 Output the candidate topic data.

[0180] Outputting candidate topic data can include returning structured candidate topic data as a synchronous interface response, or it can include storing structured candidate topic data in an asynchronous task record and returning it in response to subsequent result query requests.

[0181] Candidate topic data can be read and displayed by the content creator terminal 200, or further processed by the evaluation system 600, ranking component, or other downstream services. Outputting candidate topic data does not require the topic recommendation server 100 to directly execute content publishing or other business actions.

[0182] In one interface interaction implementation, the content creator terminal 200 can display a topic recommendation page. The topic recommendation page can refresh candidate topic data daily, weekly, or otherwise, displaying 10 candidate topics each time, along with the recommendation criteria, content generation perspective, and title example for each candidate topic. It should be noted that the refresh cycle and display quantity mentioned above are merely interface configuration examples and do not constitute a limitation on the topic recommendation method.

[0183] The following is combined Figure 3 This describes the evaluation loop after the candidate topic data is generated.

[0184] Obtain the results of candidate topic execution, including profile, recommend_basis, and recommended_topics. Obtain execution records and step snapshots, including run input / output and step input / output, to record input and output data during the recommendation process. Obtain historical / posterior content data sets, including historical content sample data such as titles, sections, playback, and interactions of the target content creators' actual content, to be used for subsequent evaluation of the matching degree between candidate topics and actual creation trajectories. Based on the above data, create an Eval Template and an EvalConfig. The Eval Template is used to configure the evaluation prompt word template and boolean / numeric schema; the Eval Config is used to bind the agent / run / step and configure variable mapping relationships to determine the evaluation target, target processing node, and the extraction method of the data required for evaluation. Variables are extracted from the runtime records through variable mapping relationships in EvalConfig and JSONPath. For example, candidate topic data is extracted from the recommendation output, target content creator's partition and followed object information is extracted from the profile, and recent content data or hot topic data is extracted from data tool results. Further, a backend Eval Instance is created based on the Eval Template and EvalConfig. The evaluation instance executes according to the pending→running→success / error state flow, where pending indicates pending execution, running indicates execution, success indicates successful execution, and error indicates an execution error. During the evaluation process, a dual-channel evaluation is performed. The first channel is a content security evaluation, assessing the potential risks of information such as the title, reason, creative perspective, and exclusive content service strategy to obtain a security evaluation result. The second channel is a backtesting of the target content creator's content data, evaluating the degree of matching between the candidate topics and the target content creator's actual creative trajectory based on real content sample data such as candidate topics, the target content creator's historical content data, and post-hoc content data to obtain sample matching results. After the evaluation is completed, the evaluation results are persisted, recording information such as score, reason, token, prompt, and original output to form a structured evaluation result. Based on the content security evaluation results and sample matching results, an evaluation feedback loop is entered. Specifically, when the security evaluation result does not meet the preset output conditions, the corresponding candidate topic can be intercepted, or a candidate topic rewriting instruction can be generated based on the security evaluation result.When the sample matching results meet the preset output conditions, the processing rules, the fusion parameter weights corresponding to the topic score, or the evaluation template can be updated based on the sample matching results to achieve continuous optimization of the recommendation strategy. The processing rules can include descriptions of recommendation rules, rules for using trending topics, descriptions of dedicated content service strategies, and content security constraints.

[0185] Through the above process, the recommendation basis, evaluation process and evaluation results in the candidate topic generation process can be recorded and verified, so that the candidate topics can not only be generated based on the model, but also be output controlled after content security assessment and verification of real creation trajectory, thus forming a closed loop process of recommendation generation, risk detection, effect verification and strategy optimization.

[0186] In an optional embodiment, after generating the candidate topic data, the method further includes: Step S1800: Obtain the recommended running records corresponding to the candidate topic data; Step S1802: Based on the evaluation configuration data, extract candidate topic information, recommendation evidence information, and content creator association information from the recommended operation record; Step S1804: Obtain real content sample data corresponding to the information associated with the content creator; Step S1806: Perform content security assessment based on the candidate topic information and the recommendation evidence information to obtain the security assessment result; Step S1808: Perform content sample matching evaluation based on the candidate topic information and the real content sample data to obtain sample matching results; Step S1810: Based on the security evaluation results and the sample matching results, perform output control on the candidate topic data.

[0187] Recommended evidence information may include trending data, recent content data, content creator characteristic data, and tool call results used when the model generates candidate topics. Real content sample data may include historical content samples, posterior published content samples, content samples from similar content creators, samples of the same content category, or manually annotated samples.

[0188] Content security assessment can evaluate candidate topic titles, recommendation criteria, content generation perspectives, title examples, hot topic references, and customized content generation strategies. Content sample matching assessment can match candidate topics with real content samples based on titles, tags, descriptions, content categories, publication times, and content performance data.

[0189] By setting up two evaluation channels—content security evaluation and content sample matching evaluation—candidate topic data can be verified from two aspects: the degree of matching between content output conditions and real content trajectory. The evaluation results can then be used for subsequent output control.

[0190] In an optional embodiment, based on the evaluation configuration data, candidate topic information, recommendation evidence information, and content creator association information are extracted from the recommendation operation record, including: Step S1900: Determine the evaluation target, target processing node, and variable mapping relationship based on the evaluation configuration data, wherein the evaluation target includes the recommended running process or the target processing steps in the recommended running process; Step S1902: Determine the data source of the variable to be extracted as the input or output data of the recommended running record based on the variable mapping relationship; Step S1904: Extract the candidate topic information, the recommendation evidence information, and the content creator association information from the corresponding input or output data through data path expressions.

[0191] Evaluation objectives can include process-level evaluation and step-level evaluation. Process-level evaluation can be used to evaluate a complete topic recommendation process; step-level evaluation can be used to evaluate structured processing nodes, tool call nodes, or other target processing steps.

[0192] Variable mapping relationships can include variable names, data source types, and data path expressions. Data source types can indicate whether data is extracted from input or output data. Data path expressions can use structured data path expressions.

[0193] In some embodiments, the evaluation configuration data may further include a target processing entity identifier and a start / stop status. The target processing entity identifier can be used to determine which topic recommendation processing entity generated the recommended run record to be evaluated; the start / stop status can be used to control whether the corresponding evaluation configuration participates in the creation of evaluation instances. When multiple evaluation configurations match the same recommended run record, the evaluation configuration to be executed can be determined based on the target processing entity identifier, the evaluation target, and the target processing node, and evaluation configurations in a disabled state can be skipped. In this way, evaluation tasks can be configured separately for different processing entities or different processing nodes, and disabled evaluation configurations can be prevented from continuing to generate evaluation instances.

[0194] For example, a list of candidate topics can be extracted from the output data of the recommended running records, content category and target audience information can be extracted from the content creator feature fields, recent content and hot data can be extracted from the tool results, and a set of real content samples can be extracted from the historical records.

[0195] Data path expressions can also be replaced by structured query statements, feature service queries, vector filtering, log retrieval, data warehouse queries, or event stream aggregation rules.

[0196] By adopting a mapping relationship between evaluation targets and variables, the same evaluation template can be applied to different recommended running processes or processing steps, and the evaluation system does not need to rely on fixed field positions to read evaluation variables.

[0197] In an optional embodiment, performing the content security assessment or the content sample matching assessment includes: Step S2000: Create an evaluation instance based on the evaluation template and the evaluation configuration data, and set the evaluation instance to the pending execution state; Step S2002: Update the evaluation instance to the execution state, and generate evaluation input data based on the candidate topic information, the recommendation evidence information, and the content creator association information; Step S2004: Call the evaluation model to process the evaluation input data; Step S2006: When the evaluation template enables evaluation tool invocation, perform iterative evaluation tool invocation and generate evaluation model input based on the evaluation tool invocation results; Step S2008: Obtain the structured evaluation results output by the evaluation model, wherein the structured evaluation results include the scoring type, the scoring value, and the scoring reason; Step S2010: Perform data parsing and data structure verification on the structured evaluation results, and update the evaluation instance to a success state or an error state according to the verification results.

[0198] The evaluation template can define evaluation prompts, evaluation models, scoring types, scoring ranges, output data structures, whether tool calls are enabled, and available evaluation tools. Scoring types can be Boolean or numerical. Boolean scores can be used to indicate whether preset output conditions are met; numerical scores can be used to indicate risk level or sample matching degree.

[0199] In some embodiments, the evaluation template may further include an evaluation tool service address. The evaluation tool service address can be used to determine the tool service or tool gateway accessed by the evaluation model when making tool invocations. If tool invocation is not enabled in the evaluation template, the evaluation tool service address can be ignored; if tool invocation is enabled in the evaluation template, evaluation tool description information can be obtained based on the evaluation tool service address and provided to the evaluation model. In this way, the evaluation template can simultaneously constrain the scoring output range and the data source of the evaluation tool.

[0200] Structured evaluation results may include a rating type field, a rating value field, and a rating reason field. The rating type field indicates whether the rating result is in Boolean or numerical form, the rating value field records the corresponding rating result, and the rating reason field records the basis for forming the rating result.

[0201] In one exemplary configuration, the evaluation tool can be invoked a maximum of 6 times, and the maximum length of a single evaluation tool result written to the evaluation context can be 8000 characters. If a network timeout or server error occurs during the evaluation model, it can be retried exponentially, with a maximum of 3 retries. The above parameters are for illustrative purposes only.

[0202] Content security assessment can be performed using a rule engine, sensitive information sets, content classification models, multimodal security models, content verification interfaces, manual review, or a combination of multiple assessment models. Content sample matching assessment can be performed based on historical content, subsequently published content, content from similar content creators, high-performing content in the same content category, manually labeled samples, or feedback from online experiments.

[0203] By establishing evaluation instance states and performing structured validation on the evaluation output, each evaluation process can have clear inputs, outputs, and final states, and the situation where the evaluation model output format does not conform to the preset structure can be incorrectly adopted.

[0204] In an optional embodiment, the method further includes: Step S2100: Record the evaluation template identifier and evaluation template version corresponding to the evaluation instance; Step S2102: Record at least one of the following: the evaluation input data, the original output of the evaluation model, the processing volume usage data, the evaluation time, the structured evaluation result, and the error message; Step S2104: Store the evaluation input data and the original output in the target storage space; Step S2106: Store reference information pointing to the target storage space in the evaluation instance record.

[0205] The target storage space can be a key-value store, an object store, or other data storage space suitable for storing long text. Evaluation instance records can save template identifiers, template versions, scoring results, and references to the original inputs and outputs.

[0206] When the content of the evaluation template changes, a new template version can be generated. Historical evaluation instances continue to be associated with the template version used during their execution to avoid interpreting old evaluation results using a new template.

[0207] In some embodiments, the version information used to trace evaluation criteria is not limited to the evaluation template version. Version information may also include at least one of the following: content validation rule version, evaluation model version, output control strategy version, evaluation configuration snapshot identifier, or experiment group version. Evaluation instances can store information about the versions actually used, or store references to the corresponding version configurations. Even after changes to evaluation rules, evaluation models, or output control strategies, historical evaluation instances remain associated with the version information actually used during their execution. This approach distinguishes evaluation results generated by different rules, models, or strategies, reducing confusion caused by interpreting historical evaluation results using the current configuration.

[0208] By recording the evaluation template version, original input and output, and processing volume data, the basis for the evaluation results can be traced back, and the results formed by different evaluation criteria can be distinguished, thereby improving the traceability of the evaluation loop.

[0209] In an optional embodiment, output control is performed on the candidate topic data based on the security evaluation results and the sample matching results, including at least one of the following: Step S2200: If the security evaluation result does not meet the preset content output conditions, the corresponding candidate topic is blocked. Step S2202: Generate candidate topic rewriting instructions based on the security evaluation results; Step S2204: Adjust the order of candidate topics according to the security assessment results and the sample matching results; Step S2206: Send the candidate topics to the review processing queue; Step S2208: Update the processing rules for generating the candidate topic data, the fusion parameters for determining the order of candidate topics in the candidate topic data, or the evaluation template for performing the evaluation based on the sample matching results.

[0210] One or more of the above steps can be selected for execution. If the Boolean security assessment fails, candidate topics can be directly blocked or a rewrite can be triggered. Numerical risk scores can serve as a ranking correction signal. When the sample matching result is low, the ranking position of the corresponding candidate topic can be lowered.

[0211] The candidate topic rewriting instruction can include the risk dimensions that need to be modified and the recommendation basis to be retained, allowing the target model to modify candidate topics without losing all instrumental evidence. The review processing queue can be used to save candidate topics that require further processing.

[0212] In some embodiments, updating the processing rules based on the sample matching results may include updating at least one of the following: description information of the exclusive content generation strategy, rules for using hot data, rules for controlling duplicate candidate topics, and content security constraints. Updating the fusion parameters corresponding to the topic score based on the sample matching results may include updating the fusion parameters or correction parameters corresponding to the content sample matching degree, content security risk information, and duplicate risk information, respectively. Specifically, when multiple low-matching candidate topics deviate from the existing content direction of the target content creator, the fusion parameter corresponding to the content sample matching degree can be increased; when candidate topics have many content security risks, the correction parameter corresponding to the content security risk information can be increased; when candidate topics have a high degree of duplication with recent content, the correction parameter corresponding to the duplicate risk information can be increased. Through the above methods, the evaluation results can be applied to explicit rule items and scoring signals, enabling subsequent candidate topic generation and ranking to be adjusted based on the identified deviation reasons.

[0213] The sample matching results can also be fed back into the evaluation template. For example, the reasons for deviations in candidates with low matching scores can be analyzed, and the matching dimensions or scoring range used in subsequent evaluations can be adjusted.

[0214] By involving the evaluation results in interception, rewriting, sorting, and parameter updates, the evaluation process can be expanded from simply recording results to becoming part of the control of candidate topic output, allowing candidate topics that do not meet the preset content output conditions or have a low degree of matching with real content samples to be processed accordingly.

[0215] The following is combined Figure 4 Explain synchronous processing and asynchronous reliable execution.

[0216] The initial state is no record, only the external task identifier exists. When an asynchronous task is submitted for the first time, an interactive lock (Redis NX+TTL) is used to prevent the concurrent creation of the same logical task, and an asynchronous task record is created, causing the task to enter the Pending (waiting to be executed) state.

[0217] After entering the Pending state, the AgentResult is saved, the publishing index is written, and then the task message is published to the task queue. If message publishing fails or the task exits abnormally during the publishing process, the task message can be republished through the publishing compensation mechanism to ensure that the task can be executed normally.

[0218] After receiving the task message, the Worker updates the task status from Pending to Running via CAS pending→running. Upon entering the Running state, an execution identifier is generated, and the status thread identifier is set to equal the execution identifier. This identifier is used to associate the checkpoint (task processing status checkpoint) and trace (processing chain record) corresponding to this execution.

[0219] If an object initiates a cancellation request before the task has started execution, i.e. the cancellation request succeeds before CAS pending→running, then the task status is updated to Canceled (cancelled state) via CAS pending→canceled; if the task has already entered the Running (execution state), the cancellation request is rejected to avoid interrupting ongoing model or tool calls and causing inconsistencies in the state.

[0220] Once a task enters the Running state, it executes the asynchronous task. When the asynchronous task executes successfully, it enters the Completed state; when the asynchronous task fails, it enters the Failed state.

[0221] Upon entering the Completed state, the system writes structured results and saves the task processing status checkpoint identifier, indicating that the task has been completed normally. Upon entering the Failed state, the system logs error information, which can be subsequently corrected upon successful completion.

[0222] When an asynchronous task is resubmitted, a new queue batch identifier is generated for this submission. When the Worker receives a task message, it compares the queue batch identifier in the message with the queue batch identifier in the current task record; if they do not match, the old message is discarded directly. This ensures that when a new batch is repeatedly generated, the old message / old state is discarded, preventing old messages from affecting the execution of the current task.

[0223] When task states compete for space, a success-first competition rule applies. Specifically, a result that enters the Failed state later cannot override a successful result that has already entered the Completed state; a successful result that enters the Completed state later can correct a previous Failed state to a Completed state, thereby improving the availability of the final task result.

[0224] In an optional embodiment, the method may further include: Step S2300: If the topic recommendation request is a synchronous processing request, the candidate topic data is returned after the generation of the candidate topic data is completed. Step S2302: If the topic recommendation request is an asynchronous processing request, create an asynchronous task record, set the asynchronous task record to a pending execution state, and return the task identifier corresponding to the asynchronous task record. Step S2304: In response to the result query request for the task identifier, return the task status of the asynchronous task record and the candidate topic data generated when the task is completed.

[0225] Synchronous processing is suitable for requests with few model processing rounds or short tool call times. Asynchronous processing is suitable for scenarios with multiple tool calls, long-context model generation, or high concurrent request volumes. Asynchronous task records can be established before the task message is published, allowing the result query interface to return a pending status before the task begins execution. After the task is completed, structured candidate topic data can be written to the asynchronous task record.

[0226] By providing both synchronous and asynchronous processing paths, the appropriate return method can be selected based on the task processing time, and long-running tasks do not need to continuously occupy synchronous request connections.

[0227] In an optional embodiment, the asynchronous task record includes: The task identifier remains unchanged externally; Queue batch identifier used to distinguish different task submission batches; Execution identifiers used to distinguish different actual execution processes; Used to associate task processing status checkpoints and processing link records with status thread identifiers; Task status, task results, error messages, and task processing status checkpoint information; Specifically, when the asynchronous task corresponding to the asynchronous task record is started to be executed, the execution identifier is generated, and the state thread identifier is set as the execution identifier.

[0228] The above scheme corresponds to steps S2400 to S2406.

[0229] Step S2400: Create a task identifier that remains unchanged externally and a queue batch identifier used to distinguish submission batches.

[0230] Step S2402: Generate an execution identifier when the asynchronous task actually begins execution.

[0231] Step S2404: Set the state thread identifier to the execution identifier.

[0232] Step S2406: Associate the task status, task result, error message, and task processing status checkpoint information with the corresponding identifiers.

[0233] The task identifier is used to identify the logical task externally during the processing of the same logical task; the queue batch identifier is used to distinguish different submission batches of the logical task; the execution identifier is used to distinguish different execution processes actually started by the worker node; and the status thread identifier is used to associate the task processing status checkpoint and processing link record corresponding to an execution process.

[0234] The task identifier remains unchanged during the resubmission of the same logical task, allowing result queryers to continue using the same task identifier. The queue batch identifier is updated with each resubmission to distinguish between old and new task messages. The execution identifier is generated when the worker node actually begins processing to distinguish between different actual execution processes. When the status thread identifier matches the execution identifier, the checkpoints and processing chain records for that execution can be bound to the same execution process.

[0235] In one exemplary configuration, task processing status checkpoints can be retained for 7 days to support troubleshooting, runtime log association, or recovery. The retention period can be adjusted based on storage resources and traceability requirements.

[0236] By separating task identifiers, queue batch identifiers, execution identifiers, and status thread identifiers, logical tasks, message batches, actual execution processes, and checkpoint chains can be identified separately, reducing data confusion between different execution processes when resubmitting or when old messages arrive late.

[0237] In an optional embodiment, the method for creating and submitting the asynchronous task corresponding to the asynchronous task record includes: Step S2500: Create a prompt cross-exclusion flag corresponding to the task flag; Step S2502: Store the asynchronous task record and write the asynchronous task record into the task index to be published; Step S2504: Publish the task message corresponding to the asynchronous task record to the task queue; Step S2506: If the task message is not confirmed for publication, scan the index of tasks to be published and republish the task message; Step S2508: If the asynchronous task corresponding to the same task identifier is resubmitted, generate a new queue batch identifier. Step S2510: After obtaining a task message from the task queue, compare the queue batch identifier carried in the task message with the queue batch identifier in the asynchronous task record. Step S2512: If the batch identifiers of the two queues are inconsistent, terminate the execution of the task corresponding to the task message.

[0238] A cross-exclusion flag can be created through a key-value store operation that has write capability if it does not exist. In one exemplary configuration, the validity period of the cross-exclusion flag can be 30 seconds to limit the concurrent creation of multiple task records for the same task flag within a short period of time.

[0239] The pending task index can record tasks that have been persisted but whose message publication has not yet been confirmed. In one exemplary configuration, if a task record is not confirmed for publication within 15 seconds after being written, the task message can be republished by the scanning task.

[0240] In some embodiments, task message publishing anomalies may include a message publishing interface returning a failure, the message publishing result not being confirmed, and the task record being written but the publishing process being interrupted before the task message submission is complete. For the above anomalies, the corresponding record in the pending task index remains in a pending-publishing state. The compensation scan task can determine the asynchronous tasks that have not yet been confirmed for publishing based on the pending task index and reconstruct the task message corresponding to the current queue batch identifier for publishing. After confirming that the task message has entered the task queue, the corresponding record can be removed from the pending task index or updated to a published state. Through the above timing, the situation where tasks remain in a pending state for a long time due to task records being persisted but task messages not entering the task queue can be reduced.

[0241] The queue batch identifier comparison occurs before the worker node executes the target model or data tool. If the queue batch identifier carried in the task message does not match the current queue batch identifier stored in the asynchronous task record, it indicates that the message belongs to an old submission batch, and the worker node can directly terminate the corresponding execution.

[0242] In some embodiments, the identifier used to distinguish task submission batches is not limited to a queue batch identifier. The identifier can also be implemented using a task version number, retry batch number, idempotent identifier, task pipeline identifier, or a monotonically increasing isolation token. Before executing a task message, a worker node can compare the identifier carried in the task message with the current identifier stored in the asynchronous task record, and continue execution only if they match. In this way, execution isolation between new and old task batches can be maintained when using different task queues or data storage systems.

[0243] The index of tasks to be published can be implemented using transaction outgoing mailboxes, message transaction records, event logs, dual-write consistency verification records, or timed replay records. The task queue can be implemented using streaming message queues, cloud task queues, or database task tables.

[0244] By persisting task records and the index to be published before publishing messages, tasks to be published can be rediscovered if message publishing fails or the publishing process terminates abnormally. Isolating different submission batches using queue batch identifiers can reduce the possibility of older task messages overwriting the results of newer batches.

[0245] In an optional embodiment, updating the task status recorded in the asynchronous task includes: Step S2600: The task status is updated from pending execution to execution status through a comparison and swap operation; Step S2602: If the task is executed successfully, the task status is updated from execution status to completion status by comparison and exchange operation, and written into the candidate topic data; Step S2604: In the event of task execution failure, the task status is updated from execution status to failure status through a comparison and swap operation, and error information is written. Step S2606: In response to a cancellation request for an asynchronous task in a pending execution state, the task state is updated to a cancelled state by a comparison and swap operation; Step S2608: If the asynchronous task corresponding to the asynchronous task record is already in the execution state, reject the cancellation request; Step S2610: If the task status is completed, do not overwrite the completed status with a subsequent failure status. Step S2612: If a failure status is written first and then valid candidate topic data is obtained, the failure status is corrected to a completion status.

[0246] The compare-and-swap operation can simultaneously verify the current task status and queue batch identifier in the database update conditions. The status update is only accepted if the preceding state matches the current database state.

[0247] Work nodes, cancellation requests, and recurring task messages may simultaneously attempt to update the task status. By using a compare-and-swap operation, state changes such as from pending to executed, executed to completed, executed to failed, and pending to cancelled can be made to satisfy preset preceding states.

[0248] Cancel operations only block tasks that have not yet entered the running state. Worker nodes can reject cancellation requests once their state has been successfully updated to the running state, reducing inconsistencies in intermediate states caused by forced termination during model or tool calls.

[0249] In other embodiments, asynchronous task cancellation can also employ an in-run collaborative cancellation method. The topic recommendation server can set cancellation status flags for target model calls and data tool calls. The model call component or data tool can read the cancellation status flags between processing nodes and end subsequent processing at a processing boundary that allows safe termination. Asynchronous tasks can also be associated with execution leases. If the execution lease exceeds a preset validity period and is not renewed, other worker nodes can regain the task execution rights. Different asynchronous tasks can also be associated with task priorities, and the scheduling order of tasks to be executed can be adjusted according to task priorities. The in-run cancellation, lease processing, and priority scheduling described above are all alternative implementations of the cancellation and scheduling mechanisms and do not constitute a necessary limitation on the rules for canceling tasks in the pending execution state.

[0250] In one exemplary configuration, when a final state write fails, a backoff retry can begin after 100 milliseconds, with a maximum backoff time of 30 seconds, and continue to retry during the process's lifetime until the completion state, failure state, or cancellation state is persisted.

[0251] In some embodiments, after obtaining candidate topic data or error information, the worker node can first write the corresponding final state to the asynchronous task record, and then perform normal acknowledgment on the task messages in the task queue. If the final state writing temporarily fails, the worker node can keep the task message in an unacknowledged state and re-execute the final state writing according to the backoff strategy. Normal acknowledgment of the task message is then performed after the target final state in the completed, failed, or canceled state has been persisted. By making the final state persisted before the task message acknowledgment, the state loss caused by the task message being removed from the task queue but the task record not yet saving the final state can be reduced.

[0252] Compare-and-swap operations can be implemented using database optimistic lock version numbers, pessimistic locks, distributed locks, lease tokens, isolation tokens, vector clocks, or simply appending event streams.

[0253] By restricting the preceding states of state updates and adopting a completion state priority rule, state competition among worker nodes, cancellation requests, and delayed messages can be coordinated, reducing the possibility of tasks that have already generated valid candidate topic data being overwritten by later-arriving failure information.

[0254] In an optional embodiment, the method further includes: Step S2700: Record at least one of the following according to the task processing node: node input data, node output data, tool call result, target model processing result, processing volume usage data, task execution status, and task processing status checkpoint; Step S2702: Record at least one of the following: tool call exception, target model call exception, structured processing exception, and task status update exception; Step S2704: According to the data tools, candidate topics or topic scoring dimensions, the data sources and intermediate processing results corresponding to the candidate topics are linked and recorded to form a candidate topic evidence chain.

[0255] Node-level records allow you to correlate model inputs, model outputs, tool calls, and structured results in the order of execution during a recommendation run. Processing volume data can be used to identify anomalies in context length or model call volume. Task processing status checkpoints allow you to review the status of specific nodes after execution.

[0256] The evidence chain for candidate topics can link candidate topic identifiers, trending items, content creator-related data, tool call results, and scoring signals. For example, it can record which trending items a candidate topic originates from, which historical content it semantically matches, and the corresponding scoring dimensions.

[0257] Task processing status checkpoints can be stored in key-value databases, relational databases, embedded databases, object storage, vector databases, or event logs. Processing link records can be stored by a link tracing system, performance metrics system, log system, or performance record database.

[0258] By establishing node-level operation records and candidate topic evidence chains, the corresponding processing nodes can be located when tool anomalies, model anomalies, structure failures, or task state competition occur, and the data sources and intermediate processing results used by candidate topics can be traced back.

[0259] In addition to the above-described embodiments, the acquisition of multi-source data, inference and arrangement, candidate generation, structured processing, evaluation, and asynchronous scheduling in this application can all adopt other technical methods that can achieve the same data processing purpose.

[0260] In terms of data collection, the unified tool call interface can be replaced by direct calls to data interfaces, graph query interfaces, database views, data warehouse queries, message subscriptions, or offline feature services. Hot topic data can come from internal rankings of content provision systems, hot search terms, high-frequency comment terms, high-frequency bullet screen terms, online community hot topics, industry news, vertical field information sources, or manually maintained hot topic collections.

[0261] For inference orchestration, state machines, workflow engines, business process engines, directed acyclic graphs (DAGs) for tasks, or custom task runtimes can be used. A single-objective model can also be replaced by multiple processing models executing collaboratively.

[0262] For candidate topic generation, template generation, rule combination, keyword expansion, knowledge graph path processing, vector filtering, collaborative filtering, or supervised learning models can be used. Candidate topic ranking can employ explicit scoring models, learned ranking models, classification models, pairwise ranking models, or content response prediction models.

[0263] For structured processing, data structure-constrained output, function call-based strongly constrained output, syntax-constrained decoding, finite-state parsers, or traditional information extraction models can be employed. Structured data fields may also include at least one of the following: related hot topic information, estimated popularity information, specific content adaptation information, specific content generation strategy information, and content output risk warning information. All of the above fields are optional fields for candidate topic data and do not constitute a mandatory set of fields for candidate topic data.

[0264] In terms of evaluation, the output can take the form of Boolean scores, numerical scores, multidimensional scores, rating scales, risk label sets, ranking preference pairs, natural language diagnostic information, or executable correction instructions. The evaluation results can be used as offline diagnostic information, or as real-time output access control, ranking features, rewrite trigger conditions, or review trigger conditions during the recommendation process.

[0265] For asynchronous scheduling, the index of tasks to be published can use transactional mailboxes, message transactions, event sourcing logs, dual-write consistency checkers, or timed replays; state updates can use optimistic locking, pessimistic locking, distributed locking, lease tokens, isolation tokens, or simply appending event streams.

[0266] In another implementation, content creator characteristics, hotspot matching results, and candidate topics can be pre-computed periodically, and only lightweight sorting and structured return can be performed when a request arrives.

[0267] In another implementation, all tool data can be obtained in real time after the request arrives, and multiple rounds of model-tool interaction can be performed to improve the timeliness of hot data.

[0268] In another implementation, content creator characteristics, historical content vectors, and target audience characteristics can be maintained offline, while only the latest trending data and exclusive content service status can be obtained online.

[0269] In another implementation, multi-source features can be used as model inputs to train a dedicated topic recommendation model, without explicitly calling data tools during online processing.

[0270] In another implementation, candidate topics can be generated first, and then filtered, sorted, or rewritten by content security assessment and content sample matching assessment.

[0271] In another implementation, a human-machine collaborative processing approach can be adopted. After the topic recommendation server generates candidate topics, recommendation criteria, content generation perspectives, and exclusive content generation strategies, it can provide the candidate topic data to the review processing workbench. The review processing workbench can receive retention, deletion, sorting adjustment, or text correction information for candidate topics and store the corresponding feedback data in association with candidate topic identifiers, task identifiers, and candidate topic evidence chains. The topic recommendation server can adjust processing rules, fusion parameters, or model input data based on feedback data corresponding to multiple tasks, and can also use the feedback data as a data source for subsequent model training or evaluation configuration updates. Through the above methods, human feedback can be converted into recordable, associative, and reusable processed data, which can be used for adjustments in the subsequent candidate topic generation process.

[0272] To make this application easier to understand, an exemplary application is provided below. In this exemplary application: S1, the topic recommendation server 100 receives the object identifier and checkpoint storage backend parameters from the target content creator, and performs positive integer verification on the object identifier. Synchronous processing generates the current call's state thread identifier, while asynchronous processing generates a task identifier that remains unchanged externally.

[0273] S2, the recommended server 100 initializes the processing volume statistician, processing link recorder, and object identifier context variables, and creates or reuses the checkpoint storage backend parameters. Asynchronous processing also creates asynchronous task records and sets the task status to pending execution.

[0274] S3, the topic recommendation server 100 reads the processing rule registry based on the task type, loads the complete content of the main processing rule, and writes summaries of other available processing rules into the task processing status. Simultaneously, it writes the hotspot data source configuration information corresponding to multiple content categories.

[0275] S4, the topic recommendation server 100 constructs a task instruction based on the object identifier. The task instruction is used to instruct the target model to generate candidate topics after obtaining the content creator's feature data, recent content data, data of the objects of interest, and the status of the exclusive content service.

[0276] S5, the topic recommendation server 100 inputs the task processing status and tool description information into the target model. The target model outputs candidate topic generation information or a model message including tool invocation information.

[0277] S6, when the model message includes multiple tool call tasks, the topic recommendation server 100 extracts the tool identifier and call parameters from the model message, distributes them uniformly to local data tools and remote data tools, and executes multiple tool call tasks through concurrent execution resources.

[0278] S7, when calling the hot data tool, the topic recommendation server 100 reads multiple hot data source nodes according to the target content category, obtains multiple hot data results in parallel, and aggregates the multiple results into a unified structured data.

[0279] S8, the topic recommendation server 100 encapsulates the tool results into a tool result message and writes it back to the task processing status, then calls the target model again. When the cumulative tool calls reach the preset limit, a result output constraint is written to the task processing status.

[0280] S9, the target model, integrates content creator characteristics, recent content, audience size, dedicated content service status, and trending data to generate multiple candidate topics. Each candidate topic includes a topic title, recommendation criteria, content generation perspective, title example, related trending topics, estimated popularity, dedicated content adaptation information, and dedicated content generation strategy information.

[0281] S10, the topic recommendation server 100 inputs the raw results from the tool and the candidate topic text into the structured processing module to generate structured candidate topic data, and validates the object identifier, object name, object display image, content category, content type, service status, popularity, required fields, and numerical validity. If the validation fails, the reason for the failure is fed back to the target model for correction.

[0282] S11: Synchronous processing directly returns structured candidate topic data. Asynchronous processing updates the task status to complete by comparing and exchanging operations, and writes the structured candidate topic data and task processing status checkpoint identifier. The result query interface returns the task status and result based on the task identifier.

[0283] S12, when a task fails, a comparison and swap operation is performed to write the failure status and error information. If a cancellation request is received before a task has started, the pending execution status can be updated to the cancelled status. When old task messages arrive, invalid executions are discarded according to the queue batch identifier.

[0284] S13, After the recommended operation process or target processing steps are completed, the evaluation system 600 creates an evaluation instance based on the evaluation configuration data. It extracts candidate topics, tool evidence, content creator characteristics, and real content samples from the operation records, and performs content security evaluation and content sample matching evaluation respectively. The evaluation results are structured into Boolean values ​​or numerical values ​​and corresponding scoring reasons, and are used for candidate topic interception, rewriting, sorting correction, fusion parameter updates, or evaluation template updates.

[0285] Content creators no longer need to manually describe complete account characteristics and historical content; they can automatically obtain multi-source data based on object identifiers and generate candidate topic data during the iterative calling of models and tools. Structured processing, evaluation closed loop, and asynchronous reliable execution mechanisms further process the result data format, result quality, and task status, respectively.

[0286] Example 2 Figure 5 A flowchart illustrating a topic recommendation method according to Embodiment 2 of this application is shown schematically. This embodiment can be applied to a content creator terminal 200. The content creator terminal 200 can communicate with a topic recommendation server 100 via network 700 to send topic recommendation requests to the topic recommendation server 100, receive structured candidate topic data returned by the topic recommendation server 100, and perform local data reading and display control based on the structured candidate topic data. Figure 5 As shown, the topic selection recommendation method may include steps S2800 to S2808, wherein: Step S2800: Obtain the object identifier of the target content creator.

[0287] Step S2802: Generate a topic recommendation request based on the object identifier, and send the topic recommendation request to the topic recommendation server.

[0288] Step S2804: Receive structured candidate topic data returned by the topic recommendation server. The structured candidate topic data includes multiple candidate topic entries. Each candidate topic entry includes a candidate topic title, recommendation basis, content generation angle, and / or exclusive content adaptation information.

[0289] Step S2806: Read the candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information corresponding to each candidate topic item from the structured candidate topic data, so as to generate candidate topic display information corresponding to each candidate topic item.

[0290] Step S2808: Control the display of candidate topic display information corresponding to each of the aforementioned candidate topic entries.

[0291] In this embodiment, the content creator terminal 200 can receive structured candidate topic data with a preset field structure and read the field values ​​corresponding to different candidate topic entries from the structured candidate topic data. The content creator terminal 200 can generate candidate topic display information corresponding to each candidate topic entry according to the read field values, and then provide the candidate topic display information to the terminal's display component for display. Through the above method, different types of candidate topic data returned by the topic recommendation server 100 can be read by the content creator terminal 200 according to the corresponding fields, so that the candidate topic title, recommendation basis, content generation angle, and exclusive content adaptation information can establish a corresponding relationship according to the candidate topic entries, reducing the impact of changes in free text structure on terminal data reading and display processing, and improving the stability of candidate topic data in the client processing process.

[0292] The following combination Figure 5 The steps in steps S2800 to S2808, as well as other optional steps, are described in detail.

[0293] Step S2800: Obtain the object identifier of the target content creator.

[0294] An object identifier can be data used to distinguish a target content creator in a content delivery system or topic recommendation server 100. The object identifier can be a numeric identifier, a string identifier, or a composite identifier. In some embodiments, the content creator terminal 200 can read the object identifier from currently associated content creator information or obtain it from pre-acquired and stored object identifier data. The target content creator can be the content creator currently requesting candidate topic data using the content creator terminal 200, or the content creator selected by the content creator terminal 200. The object identifier enables the topic recommendation server 100 to obtain content creator association data corresponding to the target content creator and generate candidate topic data accordingly.

[0295] Step S2802: Generate a topic recommendation request based on the object identifier, and send the topic recommendation request to the topic recommendation server.

[0296] The content creator terminal 200 can write an object identifier into the topic recommendation request. The topic recommendation request can take the form of a structured request message, API call parameters, or other data format that can be read by the topic recommendation server 100. The topic recommendation request can be synchronous or asynchronous. A synchronous topic recommendation request corresponds to a request that receives the processing result through the current request connection. An asynchronous topic recommendation request corresponds to a request that first obtains an asynchronous task identifier, and then queries the task status and processing result based on the asynchronous task identifier.

[0297] The content creator terminal 200 can send a topic recommendation request to the topic recommendation server 100 via a network interface. After receiving the topic recommendation request, the topic recommendation server 100 can execute the multi-source data acquisition, target model processing, data tool invocation, and candidate topic data generation processes described in Embodiment 1 based on the object identifier.

[0298] Step S2804: Receive structured candidate topic data returned by the topic recommendation server. The structured candidate topic data includes multiple candidate topic entries. Each candidate topic entry includes a candidate topic title, recommendation basis, content generation angle, and / or exclusive content adaptation information.

[0299] Structured candidate topic data can take the form of objects, arrays, lists, tables, or other data formats organized according to preset fields. For example, a JSON object can be used, where multiple candidate topic items can be organized into a collection of candidate topic items.

[0300] Candidate topic titles can be used to represent the theme content of the corresponding candidate topic. Recommendation criteria can be used to represent the content creator's associated data, trending data, or their processing results used to generate the corresponding candidate topic. Content generation perspective can be used to represent the content organization direction adopted when generating target content around the corresponding candidate topic. Specific content adaptation information can be used to represent the compatibility between the corresponding candidate topic and the specific content service scenario. Different candidate topic entries can each have corresponding candidate topic titles, recommendation criteria, content generation perspectives, and specific content adaptation information. The content creator terminal 200 can determine the correspondence between the field values ​​in the same candidate topic entry based on the data boundaries or hierarchical relationships of the candidate topic entries.

[0301] Step S2806: Read the candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information corresponding to each candidate topic item from the structured candidate topic data to generate candidate topic display information.

[0302] The content creator terminal 200 can parse the structured candidate topic data according to the field structure corresponding to the structured candidate topic data. The parsing process may include identifying the set of candidate topic entries, traversing the candidate topic entries, and reading the field values ​​in each candidate topic entry according to the field identifier. For each candidate topic entry, the content creator terminal can create a corresponding candidate topic display information object and write the candidate topic title, recommendation basis, content generation angle, and exclusive content adaptation information corresponding to that candidate topic entry into the candidate topic display information object. The candidate topic display information can be data that can be read by the content creator terminal.

[0303] Step S2808: Control the display of the candidate topic information.

[0304] The content creator terminal 200 can generate corresponding display content based on the candidate topic display information. The candidate topic display information corresponding to multiple candidate topic items can be displayed on the same display interface, or they can be displayed separately in different display areas.

[0305] For each candidate topic, the content creator's terminal can display the candidate topic title, recommendation criteria, content generation perspective, and exclusive content adaptation information. The specific display location, size, and style of the above information do not constitute a limitation of this application.

[0306] In an optional embodiment, the structured candidate topic data further includes content creator information corresponding to the target content creator. Steps S2806 and S2808 may also include steps S2900 and S2902, wherein: Step S2900: Generate content creator display information based on the content creator information.

[0307] Step S2902: Control the display of the content creator's information.

[0308] Content creator information can include data on the scale of followers, content category information, content direction information, and the status of dedicated content services. The scale of followers represents the number of people following the target content creator or the corresponding range of those people. Content category information represents the category to which the content published by the target content creator belongs. Content direction information represents the main themes or characteristics of the target content creator's content. The status of dedicated content services represents whether the target content creator has the corresponding dedicated content service capability, or whether the corresponding service is restricted. The content creator terminal can read the above content creator information from the structured candidate topic data and generate content creator display information according to the type of each field. The content creator display information can be displayed simultaneously with the candidate topic display information, or it can be displayed before or after the candidate topic display information.

[0309] By displaying content creator information, a recognizable display association can be established between candidate topic data and its corresponding target content creator, reducing confusion between candidate topic data from different content creators during terminal display.

[0310] In some embodiments, each candidate topic entry may further include a title example, associated trending topics information, estimated popularity information, and dedicated content generation strategy information. The title example can be used to represent an exemplary content title generated based on the corresponding candidate topic. Associated trending topics information can be used to represent trending themes or data associated with the corresponding candidate topic. Estimated popularity information can be used to represent the popularity level or popularity value corresponding to the corresponding candidate topic. Dedicated content generation strategy information can be used to represent the content organization method when generating dedicated content around the corresponding candidate topic.

[0311] Content creators can read title examples, related hot topics, estimated popularity, and exclusive content generation strategies from structured candidate topic data. They can then write the read information, along with the candidate topic title, recommendation criteria, content generation perspective, and exclusive content adaptation information of the corresponding candidate topic item, into the candidate topic display information.

[0312] By reading and combining different fields on a per-candidate-topic-item basis, the correspondence between title examples, related hot topic information, estimated popularity information, and exclusive content generation strategy information and the corresponding candidate-topics can be maintained, enabling the terminal to process candidate-topic-items with different information dimensions according to a unified data structure.

[0313] In an optional embodiment, if the topic recommendation request is an asynchronous topic recommendation request, the method may further include: Step S3100: Receive the asynchronous task identifier returned by the topic recommendation server.

[0314] Step S3102: Based on the asynchronous task identifier, send at least one task status query request to the topic recommendation server.

[0315] Step S3104: Receive the asynchronous task status returned by the topic recommendation server.

[0316] Step S3106: If the asynchronous task status is complete, receive the structured candidate topic data corresponding to the asynchronous task identifier.

[0317] An asynchronous task identifier can be used to distinguish the current topic recommendation task among multiple asynchronous tasks. After receiving the asynchronous task identifier, the content creator's terminal can include it in subsequent task status query requests, enabling the topic recommendation server to return the task status corresponding to the asynchronous task. The content creator's terminal can determine whether to send further task status query requests based on the returned task status after sending one. By using asynchronous task identifiers to query task status and processing results, the content creator's terminal can avoid continuously waiting for multiple rounds of data tool calls and model processing to complete within a single request connection, thereby reducing the continuous occupation of the current request connection by long-running topic recommendation tasks.

[0318] In an optional embodiment, the asynchronous task state includes at least one of a pending state, an executing state, a completed state, a failed state, and a cancelled state. Step S3104 may further include steps S3200 and S3202, wherein: Step S3200: If the asynchronous task status is pending execution or execution, continue to send a task status query request based on the asynchronous task identifier.

[0319] Step S3202: If the asynchronous task is in a failed state, receive the error information returned by the topic recommendation server.

[0320] The "Pending Execution" status indicates that the asynchronous task has not yet started. The "Execution" status indicates that the topic recommendation server is performing multi-source data acquisition, target model processing, or data tool invocation. The "Completion" status indicates that structured candidate topic data has been generated and is available for reading. The "Failure" status indicates that the asynchronous task has not generated any usable results. The "Cancellation" status indicates that the asynchronous task was cancelled before it began execution. Content creators can choose the subsequent processing path based on the different asynchronous task statuses. In the "Pending Execution" or "Execution" status, users can continue to query the task status; in the "Completion" status, they can receive structured candidate topic data; in the "Failure" status, they can receive error messages and end the current query process; in the "Cancellation" status, they can end the current query process.

[0321] By controlling subsequent queries or result reception based on different task states, the task query process on the content creator's terminal can be kept in correspondence with the task execution status in the topic recommendation server, reducing invalid queries or reading results before the task is completed.

[0322] In an optional embodiment, when the asynchronous task is in a pending execution state, steps S3300 to S3304 may be further included, wherein: Step S3300: Receive a cancellation operation for the asynchronous task corresponding to the asynchronous task identifier.

[0323] Step S3302: In response to the cancellation operation, a task cancellation request including the asynchronous task identifier is sent to the topic recommendation server.

[0324] Step S3304: Receive the cancellation status returned by the topic recommendation server.

[0325] A cancellation operation can be a request received by the content creator's terminal to stop an asynchronous task that is about to be executed. The content creator's terminal can write the asynchronous task identifier into the task cancellation request so that the topic recommendation server can determine the target asynchronous task that needs to be cancelled.

[0326] The topic recommendation server can update the target asynchronous task to a canceled state while it is still in a pending state. If the target asynchronous task has already entered the execution state, the topic recommendation server can maintain the execution state and return the corresponding processing status to reduce inconsistencies in task state caused by interrupting ongoing model processing or tool calls.

[0327] By sending and processing cancellation requests only for asynchronous tasks that have not yet entered the execution state, the cancellation operation can be matched with the task state transition on the server side, reducing state conflicts when cancellation requests and task start operations occur concurrently.

[0328] In an optional embodiment, when the topic recommendation request is a synchronous topic recommendation request, the content creator's terminal can receive structured candidate topic data through a request connection that sends a synchronous topic recommendation request.

[0329] Synchronous topic recommendation requests are suitable for topic recommendation tasks that can be completed within the processing time corresponding to the current request connection. After the topic recommendation server generates structured candidate topic data, it can directly return the structured candidate topic data through the current request connection without generating an asynchronous task identifier for the client to query. By synchronously returning structured candidate topic data, the processing step of the content creator terminal 200 sending a separate task status query request can be reduced, making it suitable for topic recommendation tasks with relatively short processing times.

[0330] In an optional embodiment, the structured candidate topic data may further include an arrangement order of multiple candidate topic entries. The content creator's terminal can control the display of candidate topic information corresponding to each candidate topic entry according to the arrangement order.

[0331] The order of arrangement can be determined by the topic recommendation server based on candidate topic scores or output control results. The content creator's terminal can maintain the arrangement order in the structured candidate topic data, or it can read the sequence data used to represent the arrangement order and provide different candidate topic display information to the display component according to the sequence data.

[0332] By displaying candidate topic entries according to the sorting order generated by the server, consistency between the candidate topic generation or sorting results and the client display order can be maintained, reducing the order changes caused by the client reorganizing candidate topic entries.

[0333] In an optional embodiment, the method may further include steps S3400 to S3404, wherein: In step S3400, in response to the periodic refresh condition, a topic recommendation request including the object identifier is resent to the topic recommendation server.

[0334] Step S3402: Receive the updated structured candidate topic data returned by the topic recommendation server.

[0335] Step S3404: Based on the updated structured candidate topic data, generate updated candidate topic display information corresponding to each candidate topic entry.

[0336] Periodic refresh conditions can be conditions that are met when a preset refresh cycle is reached. The preset refresh cycle can be set according to the update requirements of the candidate topic data. For example, the preset refresh cycle can be a daily cycle or a weekly cycle. The above cycles are only examples and are not intended to limit the specific time length of the periodic refresh conditions.

[0337] After the content creator's terminal resends the topic recommendation request, the topic recommendation server can re-obtain the target content creator's content creator association data and current trending data, and generate updated structured candidate topic data. The content creator's terminal can reread the updated structured candidate topic data and update the candidate topic display information according to the updated candidate topic entries. By periodically obtaining updated structured candidate topic data, the candidate topic information displayed on the client can be adapted to changes in the target content creator's recent content data and trending data.

[0338] In an optional embodiment, the method may further include steps S3500 to S3504, wherein: Step S3500: Receive selection, feedback, or correction operations for the target candidate topic items.

[0339] Step S3502: Generate candidate topic feedback data based on the selection operation, the feedback operation, or the correction operation.

[0340] Step S3504: Send the candidate topic feedback data to the topic recommendation server.

[0341] The selection operation can be used to represent the target content creator choosing a target candidate topic from multiple candidate topic items. The feedback operation can be used to represent the target content creator providing corresponding feedback information for the candidate topic items. The correction operation can be used to represent adjusting the candidate topic title, content generation angle, or other editable information in the candidate topic items.

[0342] Candidate topic feedback data can be used to characterize the manipulated candidate topic items and their corresponding selection, feedback, or correction results. Content creator terminals can send candidate topic feedback data to the topic recommendation server, allowing the server to record the feedback results or use it for subsequent adjustments to processing rules, model parameters, or candidate topic ranking methods. By forming structured feedback data from the selection, feedback, or correction results received by the content creator terminal and sending it to the topic recommendation server, the client's interaction results can be further read and processed by the server, providing a data foundation for subsequent adjustments to the candidate topic generation process.

[0343] like Figure 6 As shown, in an exemplary graphical user interface, the content creator terminal 200 can display a topic recommendation page. The topic recommendation page may include a page prompt area, a content creator information area, and a candidate topic area. Specifically, the page prompt area can be used to display the page name of the topic recommendation page and generation result prompt information. The generation result prompt information can be used to indicate that the candidate topic data on the page was generated by the target model, and the displayed candidate topic data is used to provide content generation references to the target content creator. The specific text content of the page name and generation result prompt information does not constitute a limitation of this application.

[0344] The content creator information area can be used to display content creator information corresponding to the target content creator. In the exemplary interface shown, the content creator information area is displayed in the form of information cards before the candidate topic area. The information card may include at least one of the following: object display image, object name, object identifier, target audience size information, content category information, content direction information, and exclusive content service status.

[0345] For example, the content creator terminal 200 can read the object name, object identifier, target audience size data, content category information, content direction information, and exclusive content service status of the target content creator from the structured candidate topic data, and write the read information into the corresponding display position in the content creator information card.

[0346] Among them, the scale of the target audience can be displayed in the form of quantity or quantity range; the content category information can display the content section or content category corresponding to the target content creator; the content direction information can display the main content theme of the target content creator; and the exclusive content service status can display whether the corresponding service is in an activated, unactivated, or restricted state.

[0347] By displaying the above information in the same content creator information card, a clear correspondence can be established between the candidate topic data currently displayed on the terminal and the corresponding target content creator.

[0348] The candidate topic area can include at least one candidate topic card. Each candidate topic card corresponds to one candidate topic entry in the structured candidate topic data. Multiple candidate topic cards can be displayed sequentially according to their order in the structured candidate topic data.

[0349] The candidate topic card may include a candidate topic identification area, a candidate topic title area, an indicator marking area, and a topic description area.

[0350] The candidate topic identification area can be used to display the sequential identifier of a candidate topic item among multiple candidate topic items. For example, it can use a serial number, sorting identifier, or other identifier format that can distinguish candidate topic items. The candidate topic title area can be used to display the candidate topic title. The candidate topic title can be a summary of the corresponding candidate topic theme, or it can be an exemplary title generated based on the corresponding candidate topic.

[0351] The indicator tag area can be used to display at least one topic indicator tag corresponding to a candidate topic item. In the exemplary interface shown in the figure, the topic indicator tags include an estimated popularity tag and a content-specific adaptation tag.

[0352] The estimated popularity tag can be generated based on the estimated popularity information in the candidate topic entries and can be displayed using one of multiple popularity levels. The content-specific adaptation tag can be generated based on the content-specific adaptation information in the candidate topic entries and can be displayed using one of multiple adaptation levels.

[0353] It should be noted that the popularity level and the adaptation level can be displayed using text, numerical values, graphic symbols, or a combination thereof, and their specific display style does not constitute a limitation of this application.

[0354] The topic selection description area may include at least one of the following: a recommendation basis display area, a content generation perspective display area, and a dedicated content generation strategy display area.

[0355] The recommendation criteria display area can be used to display the recommendation criteria corresponding to candidate topic entries. In some embodiments, the recommendation criteria may include explanatory information related to the target content creator's historical content data, recent content data, content category information, content direction information, data related to the target audience, or hot topic data.

[0356] The content generation angle display area can be used to display the content generation angle corresponding to the candidate topic. The content generation angle can be used to characterize the approach, content organization direction, or content development method adopted when generating target content around the candidate topic.

[0357] The dedicated content generation strategy display area can be used to display the dedicated content generation strategy information corresponding to the candidate topic entries. The dedicated content generation strategy information can be used to characterize the content depth, content format, or content extension direction when candidate topics are used to generate dedicated content.

[0358] In some embodiments, candidate topic cards may also display related hot topic information and title examples. Related hot topic information may be placed in the candidate topic title area, indicator marker area, or topic description area. Title examples may be placed in the content generation angle display area or a separately set title example display area.

[0359] The content creator terminal 200 can write different field values ​​into the corresponding display areas of the candidate topic cards based on the field identifiers in the structured candidate topic data. For example, the content creator terminal 200 can write the candidate topic title into the candidate topic title area, the estimated popularity information into the estimated popularity marker, the exclusive content adaptation information into the exclusive content adaptation marker, the recommendation basis into the recommendation basis display area, and the content generation angle into the content generation angle display area.

[0360] Therefore, multiple fields in each candidate topic item can be mapped to the corresponding display area in the same candidate topic card, and multiple candidate topic items can be mapped to different candidate topic cards, which can reduce the incorrect association of fields between different candidate topic items.

[0361] In some embodiments, the candidate topic area may include page description information. This page description information can be used to characterize the content category, content direction, or specific content generation target corresponding to the currently displayed candidate topic. The page description information can be generated based on the content creator information corresponding to the target content creator, or it can be generated based on the common topic characteristics corresponding to multiple candidate topic entries.

[0362] In some embodiments, the content creator terminal 200 can display a preset number of candidate topic cards on the same topic recommendation page. For example, ten candidate topic cards can be displayed. The number is merely an interface example provided in the figure and is not intended to limit the number of candidate topic entries generated or displayed each time.

[0363] When the number of candidate topic cards exceeds the number that the current display area can display at the same time, the content creator terminal 200 can arrange multiple candidate topic cards in sequence in the candidate topic area, and adjust the currently displayed candidate topic cards in response to browsing operations on the candidate topic area.

[0364] It should be noted that the aforementioned page prompt area, content creator information area, candidate topic area, and the display areas in the candidate topic cards are merely exemplary display methods. Each display area can be split, merged, or adjusted according to the terminal display size or display status, and is not intended to limit specific interface layout, color, font, size, or decorative style.

[0365] In other embodiments, interactive controls for topic selection can be set in the candidate topic card or at an associated location on the candidate topic card. These interactive controls may include at least one of a selection control, a feedback control, and a correction control. The content creator terminal 200 can receive operations on the interactive controls in the target candidate topic card, determine the target candidate topic item based on the operations, and generate candidate topic feedback data corresponding to the target candidate topic item.

[0366] To make this application easier to understand, a client-side application example is provided below in conjunction with the exemplary graphical user interface described above.

[0367] S1, the content creator terminal 200 obtains the object identifier of the current target content creator and sends an asynchronous topic recommendation request including the object identifier to the topic recommendation server 100.

[0368] S2, the content creator terminal 200 receives the asynchronous task identifier returned by the topic recommendation server 100, and queries the asynchronous task status based on the asynchronous task identifier.

[0369] S3, when the asynchronous task is in the pending or executing state, the content creator terminal 200 continues to query the asynchronous task status; when the asynchronous task is in the completed state, the content creator terminal 200 receives the structured candidate topic data.

[0370] S4, the content creator terminal 200 reads at least one of the following from the structured candidate topic data: object name, object identifier, scale data of the target content creator, content category information, content direction information, and exclusive content service status, and generates a content creator information card.

[0371] S5, the content creator terminal 200 identifies multiple candidate topic items from the structured candidate topic data and creates a corresponding candidate topic card for each candidate topic item.

[0372] S6, for each candidate topic card, the content creator terminal 200 writes the candidate topic title of the corresponding candidate topic item into the candidate topic title area, converts the estimated popularity information and exclusive content adaptation information into estimated popularity markers and exclusive content adaptation markers respectively, and writes the recommendation basis, content generation angle and exclusive content generation strategy information into the corresponding topic description area respectively.

[0373] S7, the content creator terminal 200 displays the content creator information card and multiple candidate topic cards on the topic recommendation page according to the arrangement order of multiple candidate topic items.

[0374] S8, the content creator terminal 200 receives a browsing operation for the candidate topic area and adjusts the currently displayed candidate topic card according to the browsing operation.

[0375] S9, when a topic selection interaction control is set, the content creator terminal 200 receives the selection operation, feedback operation or correction operation for the target candidate topic card, generates candidate topic feedback data and sends it to the topic recommendation server 100.

[0376] S10, when the periodic refresh conditions are met, the content creator terminal 200 resends the topic recommendation request, and regenerates or updates the content creator information card and candidate topic card based on the updated structured candidate topic data.

[0377] In the above application example, the content creator information in the structured candidate topic data is configured into a content creator information card, each candidate topic entry is mapped to a corresponding candidate topic card, and the fields in each candidate topic entry are further mapped to the corresponding display area in the candidate topic card. This maintains the correspondence between the server output fields, the terminal display objects, and the display areas, enabling the terminal to read and display different types of candidate topic information according to a preset structure.

[0378] Example 3 Figure 7A block diagram of a topic selection recommendation device according to Embodiment 3 of this application is illustrated schematically. This device can be divided into one or more program modules, which are stored in a storage medium and executed by one or more processors to complete the embodiments of this application. The program module referred to in the embodiments of this application refers to a series of computer program instruction segments capable of performing a specific function. The following description will specifically introduce the function of each program module in this embodiment. For example... Figure 7 As shown, the topic recommendation device 7000 may include a receiving module 7010, a construction module 7020, a processing module 7030, a calling module 7040, a generation module 7050, and an output module 7060.

[0379] The receiving module 7010 is used to receive a topic recommendation request, wherein the topic recommendation request includes the object identifier of the target content creator.

[0380] The construction module 7020 is used to obtain the tool description information corresponding to each of the multiple data tools, and to construct the task processing status based on the topic recommendation request and the tool description information. The multiple data tools are used to obtain the content creator association data and hot topic data of the target content creator.

[0381] The processing module 7030 is used to call the target model to process the task processing status and obtain the model processing result.

[0382] The calling module 7040 is used to call the target data tool according to the tool calling information when the model processing result includes tool calling information, obtain the tool result, and write the tool result into the task processing status.

[0383] The processing module 7030 is also used to invoke the target model again based on the updated task processing status.

[0384] The generation module 7050 is used to generate candidate topic data based on the candidate topic generation information and the tool results accumulated in the task processing state, when the model processing result includes candidate topic generation information.

[0385] Output module 7060 is used to output the candidate topic data.

[0386] It should be noted that the device in this embodiment can be used to execute the steps in the above method embodiment one. The processing module 7030, the construction module 7020, the processing module 7030, the calling module 7040, the generation module 7050, and the output module 7060 correspond to the steps in embodiment one. The relevant technical details can be referred to the above method embodiment, and will not be repeated here.

[0387] Example 3 Figure 8 A block diagram of a topic selection recommendation device according to Embodiment 3 of this application is illustrated schematically. This device can be divided into one or more program modules, which are stored in a storage medium and executed by one or more processors to complete the embodiments of this application. The program module referred to in the embodiments of this application refers to a series of computer program instruction segments capable of performing a specific function. The following description will specifically introduce the function of each program module in this embodiment. For example... Figure 8 As shown, the topic recommendation device 8000 may include an acquisition module 8010, a sending module 8020, a receiving module 8030, a processing module 8040, and a control module 8050.

[0388] Module 8010 is used to obtain the object identifier of the target content creator. The sending module 8020 is used to generate a topic recommendation request based on the object identifier and send the topic recommendation request to the topic recommendation server. The receiving module 8030 is used to receive structured candidate topic data returned by the topic recommendation server. The structured candidate topic data includes multiple candidate topic entries, and each candidate topic entry includes a candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information. The processing module 8040 is used to read the candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information corresponding to each candidate topic item from the structured candidate topic data, so as to generate candidate topic display information corresponding to each candidate topic item; The control module 8050 is used to control the display of candidate topic information corresponding to each of the aforementioned candidate topic entries.

[0389] It should be noted that the device in this embodiment can be used to execute the steps in the above method embodiment two. The acquisition module 8010, the sending module 8020, the receiving module 8030, the processing module 8040 and the control module 8050 correspond to the steps in embodiment two. The relevant technical details can be referred to the above method embodiment, and will not be repeated here.

[0390] Example 5 Figure 9 This illustration schematically depicts a hardware architecture diagram of a computer device suitable for implementing a topic selection recommendation method according to Embodiment 5 of this application. In some embodiments, the computer device 10000 may be a rack server, blade server, tower server, cabinet server, cloud computing instance, or a server cluster consisting of multiple servers. Figure 9 As shown, computer device 10000 may include a memory 10010, a processor 10020, and a network interface 10030 that are communicatively connected via a system bus. Wherein: The memory 10010 includes at least one type of computer-readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 10010 may be an internal storage module of a computer device 10000, such as the hard disk or memory of the computer device 10000. In other embodiments, the memory 10010 may also be an external storage device of the computer device 10000, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 10000. Of course, the memory 10010 may also include both the internal storage module and the external storage device of the computer device 10000. In this embodiment, the memory 10010 is typically used to store the operating system and various application software installed on the computer device 10000, such as the program code of the method described above. Furthermore, the memory 10010 can also be used to temporarily store various types of data that have been output or will be output.

[0391] In some embodiments, processor 10020 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other chip. Processor 10020 is typically used to control the overall operation of computer device 10000, such as performing control and processing related to data interaction or communication with computer device 10000. In this embodiment, processor 10020 is used to run program code stored in memory 10010 or process data.

[0392] Network interface 10030 may include a wireless network interface or a wired network interface, which is typically used to establish a communication link between computer device 10000 and other computer devices. For example, network interface 10030 is used to connect computer device 10000 to an external terminal via a network, establishing a data transmission channel and communication link between computer device 10000 and the external terminal. The network may be an intranet, the Internet, Global System for Mobile Communication (GSM), Wideband Code Division Multiple Access (WCDMA), 4G network, 5G network, Bluetooth, Wi-Fi, or other wireless or wired networks.

[0393] It should be pointed out that, Figure 9 Only computer devices with components 10010-10030 are shown; however, it should be understood that it is not required to implement all of the shown components, and more or fewer components may be implemented instead.

[0394] In this embodiment, the text-to-semantic processing method stored in memory 10010 can also be divided into one or more program modules and executed by one or more processors (such as processor 10020) to complete the embodiments of this application.

[0395] Example 6 This application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the steps in the above method embodiments.

[0396] In this embodiment, the computer-readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the computer-readable storage medium can be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the computer-readable storage medium can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device. Of course, the computer-readable storage medium can also include both the internal storage unit and the external storage device of the computer device. In this embodiment, the computer-readable storage medium is typically used to store the operating system and various application software installed on the computer device, such as the program code of the text-to-semantic processing method in this embodiment. In addition, the computer-readable storage medium can also be used to temporarily store various types of data that have been output or will be output.

[0397] Example 5 This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps described in the method embodiments above.

[0398] Obviously, those skilled in the art should understand that the modules or steps of the embodiments of this application described above can be implemented using general-purpose computer devices. They can be centralized on a single computer device or distributed across a network of multiple computer devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computer device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of this application are not limited to any particular combination of hardware and software.

[0399] It should be noted that the above are merely preferred embodiments of this application and do not limit the scope of patent protection of this application. Any equivalent structural or procedural changes made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of this application.

Claims

1. A method for recommending topics, characterized in that, Applied to a server, the method includes: Receive a topic recommendation request, wherein the topic recommendation request includes the object identifier of the target content creator; Obtain tool description information corresponding to each of the multiple data tools, wherein the multiple data tools are used to obtain content creator association data and hot topic data of the target content creator; The task processing status is constructed based on the topic recommendation request and the tool description information; The target model is invoked to process the task processing state, and the model processing result is obtained; If the model processing result includes tool call information, the target data tool is called according to the tool call information to obtain the tool result. The tool result is written into the task processing status, and the target model is called again based on the updated task processing status. If the model processing result includes candidate topic generation information, candidate topic data is generated based on the candidate topic generation information and the tool results accumulated in the task processing state. Output the candidate topic data.

2. The method according to claim 1, characterized in that, The target data tool is invoked based on the tool invocation information to obtain the tool results, including: Extract the tool identifier and tool call parameters from the tool call information; The target data tool is determined based on the tool identifier; The tool invocation parameters are sent to the target data tool through a unified tool distribution interface; When the tool invocation information includes multiple tool invocation tasks, the multiple tool invocation tasks are submitted to the concurrent execution resource, and the context data corresponding to the target content creator is copied for each tool invocation task; Aggregate the tool results returned by each of the multiple tool invocation tasks.

3. The method according to claim 1, characterized in that, The method further includes: Record the cumulative number of tool calls corresponding to the target model; When the cumulative number of tool calls reaches the preset maximum number of rounds, a result output constraint is written into the task processing status. The target model is invoked based on the task processing state with the stated output constraints to obtain the candidate topic generation information.

4. The method according to claim 1, characterized in that, The method further includes: Obtain the exclusive content service status of the target content creator, the content depth corresponding to the candidate topics to be ranked, and the adaptation degree of the exclusive content extension form corresponding to the candidate topics to be ranked; The adaptability of the exclusive content is determined based on the status of the exclusive content service, the content depth, and the adaptability of the exclusive content extension form. The fit of the target content creator is determined based on at least one of the following: the size of the target audience, the level of the target audience, and the interaction data of the target audience.

5. The method according to claim 1, characterized in that, After generating the candidate topic data, the method further includes: Obtain the recommended execution records corresponding to the candidate topic data; Based on the evaluation configuration data, candidate topic information, recommendation evidence information, and content creator association information are extracted from the recommended operation records. Obtain real content sample data corresponding to the information associated with the content creator; Based on the candidate topic information and the recommended evidence information, a content security assessment is performed to obtain the security assessment results; Based on the candidate topic information and the real content sample data, a content sample matching evaluation is performed to obtain the sample matching result; Based on the security assessment results and the sample matching results, output control is performed on the candidate topic data.

6. The method according to claim 5, characterized in that, Performing the content security assessment or the content sample matching assessment includes: An evaluation instance is created based on the evaluation template and the evaluation configuration data, and the evaluation instance is set to the pending execution state. The evaluation instance is updated to the execution state, and evaluation input data is generated based on the candidate topic information, the recommendation evidence information, and the content creator association information; The evaluation model is invoked to process the evaluation input data; When the evaluation template enables evaluation tool invocation, iterative evaluation tool invocation is performed, and evaluation model input is generated based on the evaluation tool invocation results; Obtain the structured evaluation results output by the evaluation model, wherein the structured evaluation results include the rating type, rating value, and rating reason; The structured evaluation results are parsed and validated, and the evaluation instance is updated to a success or error state based on the validation results.

7. The method according to claim 1, characterized in that, Also includes: If the topic recommendation request is a synchronous processing request, the candidate topic data is returned after the generation of the candidate topic data is completed; If the topic recommendation request is an asynchronous processing request, create an asynchronous task record, set the asynchronous task record to a pending execution state, and return the task identifier corresponding to the asynchronous task record; In response to a result query request for the task identifier, the task status of the asynchronous task record and the candidate topic data generated when the task is completed are returned.

8. The method according to claim 7, characterized in that, The methods for creating and submitting the asynchronous task corresponding to the asynchronous task record include: Create a prompt cross-exclusion flag corresponding to the task identifier; Store the asynchronous task records and write them into the task index to be published; Publish a task message corresponding to the asynchronous task record to the task queue; If the task message is not confirmed for publication, scan the index of tasks to be published and republish the task message; When an asynchronous task with the same task identifier is resubmitted, a new queue batch identifier is generated. After retrieving a task message from the task queue, the queue batch identifier carried in the task message is compared with the queue batch identifier in the asynchronous task record. If the batch identifiers of the two queues are inconsistent, the execution of the task corresponding to the task message will be terminated.

9. The method according to claim 7, characterized in that, Updating the task status recorded in the asynchronous task includes: By comparing and swapping operations, the task status is updated from the pending state to the executing state; If the task is executed successfully, the task status is updated from the execution status to the completion status through a comparison and exchange operation, and then written into the candidate topic data; In the event of task execution failure, the task status is updated from execution status to failure status through a comparison and swap operation, and an error message is written. In response to a cancellation request for an asynchronous task that is in a pending state, the task status is updated to a cancelled state by a comparison and swap operation; If the asynchronous task corresponding to the asynchronous task record is already in the execution state, the cancellation request is rejected; If the task status is completed, a subsequent failure status will not be used to overwrite the completed status. If a failure status is written first and then valid candidate topic data is obtained, the failure status is corrected to a completion status.

10. A method for recommending topics, characterized in that, Applied to content creator terminals, the method includes: Obtain the object identifier of the target content creator; A topic recommendation request is generated based on the object identifier, and the topic recommendation request is sent to the topic recommendation server. Receive structured candidate topic data returned by the topic recommendation server. The structured candidate topic data includes multiple candidate topic entries. Each candidate topic entry includes a candidate topic title, recommendation basis, content generation angle, and / or exclusive content adaptation information. Read the candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information corresponding to each candidate topic item from the structured candidate topic data, so as to generate candidate topic display information corresponding to each candidate topic item; The control displays the candidate topic display information corresponding to each of the aforementioned candidate topic entries.

11. A topic recommendation device, characterized in that, Applied to a server, the device includes: The receiving module is used to receive topic recommendation requests, wherein the topic recommendation request includes the object identifier of the target content creator; The module is used to obtain the tool description information corresponding to each of the multiple data tools, and to construct the task processing status based on the topic recommendation request and the tool description information. The multiple data tools are used to obtain the content creator association data and hot topic data of the target content creator. The processing module is used to call the target model to process the task processing status and obtain the model processing result; The calling module is used to call the target data tool according to the tool calling information when the model processing result includes tool calling information, obtain the tool result, and write the tool result into the task processing status; The processing module is also used to invoke the target model again based on the updated task processing status; The generation module is used to generate candidate topic data based on the candidate topic generation information and the tool results accumulated in the task processing state, when the model processing result includes candidate topic generation information. The output module is used to output the candidate topic data.

12. A topic recommendation device, characterized in that, The device, applied to a content creator's terminal, includes: The acquisition module is used to obtain the object identifier of the target content creator; The sending module is used to generate a topic recommendation request based on the object identifier and send the topic recommendation request to the topic recommendation server. The receiving module is used to receive structured candidate topic data returned by the topic recommendation server. The structured candidate topic data includes multiple candidate topic entries, and each candidate topic entry includes a candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information. The processing module is used to read the candidate topic title, recommendation basis, content generation angle and / or exclusive content adaptation information corresponding to each candidate topic item from the structured candidate topic data, so as to generate candidate topic display information corresponding to each candidate topic item; The control module is used to control the display of candidate topic information corresponding to each of the aforementioned candidate topic entries.

13. A computer device, characterized in that, include: At least one processor; and A memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method of any one of claims 1 to 10.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 10.

Images