Scenario-based AIGC content generation method, system and storage medium

Through the scenario-based AIGC content generation method, through task combination and model automatic matching, the problem that dialog boxes are difficult to handle complex information is solved, and high-quality and stable content generation is achieved in a commercial environment, meeting the user needs in complex scenarios. .

CN118569390BActive Publication Date: 2024-11-15SHANGHAI FARILUAN INTELLIGENT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411047870.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-15
Estimated Expiration
2044-08-01

AI Technical Summary

Technical Problem

The existing dialog box method is difficult to effectively manage and process complex structured information, especially in a commercial environment, and cannot continuously and stably output high-quality results and cannot meet the user needs in complex scenarios.

Method used

Provide a scenario-based AIGC content generation method, which can obtain user interaction data, determine task combinations and model requests, call matching models to process pending data, generate target content, and support serial and parallel combinations of multiple tasks, including input, Preprocessing, intelligent processing and output steps, automatically match and optimize models, and achieve high-quality content generation.

Benefits of technology

It realizes the ability to meet user needs without perception in complex business scenarios, continuously and stably output high-quality content, automatically make model selections, simplify user operations, and improve the efficiency and quality of AI workflows.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118569390B_ABST
    Figure CN118569390B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of artificial intelligence technology, and specifically provides a scenario-based AIGC content generation method and system, which aims to solve the problem that the interaction between humans and AI in related technologies is mostly in a dialog-style question-and-answer mode, and it is impossible to continuously and stably output high-quality answers based on complex scenarios. To this end, the present invention provides a scenario-based AIGC content generation method and system, including obtaining user interaction data, the interaction data at least including scenario task requests and data to be processed; determining a corresponding task combination according to the scenario task request and the data to be processed, and generating a corresponding model request; calling a matching model based on the model request, processing the data to be processed, and generating target content.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a scenario-based AIGC content generation method, system and storage medium. Background Art

[0002] At present, a large number of human and AI interactions are conducted through dialogue and dialog boxes, one question and one answer. There are many difficulties in using AI in commercial scenarios with this dialogue mode. For example, scenarios where AI is used in commercial environments often involve a large amount of complex background information, based on which judgments must be made, and the existing dialog box method cannot effectively manage and input a large amount of information, especially structured information. Some scenarios involve workflows, where the results of the previous step are passed to the next AI operation after manual intervention. This scenario is difficult to handle with a dialog box method and cannot meet user needs. A simple question-and-answer method cannot guarantee the continuous and stable output of high-quality results.

[0003] Therefore, a scenario-based content generation technology is needed to solve problems that cannot be solved by previous conversation methods. Summary of the invention

[0004] In order to solve at least one of the above technical problems, according to one aspect of the present invention, a scenario-based AIGC content generation method is provided, comprising: obtaining user interaction data, the interaction data at least including scenario task requests and data to be processed; determining a corresponding task combination according to the scenario task request and the data to be processed, and generating a corresponding model request; calling a matching model based on the model request, processing the data to be processed, and generating target content; wherein, at the user end, the user triggers a request, selects a scenario or task definition, and imports or inputs or adds data to be processed, that is, generates user interaction data; the method of determining a corresponding task combination according to the scenario task request and the data to be processed, and generating a corresponding model request, comprises: having a scenario definition set produced based on a scenario development platform, having a task definition set produced based on a task development platform, defining a corresponding relationship between any scenario and a plurality of task combinations, each task including a plurality of execution steps, a scenario being formed by a serial and parallel combination of a plurality of tasks, supporting a user to add a plurality of tasks in a scenario, and defining a serial and parallel relationship between tasks; determining a scenario corresponding to the scenario task request and a corresponding task combination, wherein the scenario task request includes a task type and a scenario type; generating a model request according to a plurality of execution steps in the corresponding task combination.

[0005] Optionally, in the above method, the task combination includes the execution order of each task, wherein the output result of the execution of one task is the input of the execution of another task, or multiple tasks are executed simultaneously, or the output results of multiple tasks are the input of the execution of other tasks; the multiple execution steps include: input step, preprocessing step, intelligent processing step, post-processing step and output step.

[0006] Optionally, in the above method, in the input step, the input data type includes but is not limited to one or more of the following: documents, voices, pictures, videos, URLs, data in different formats, three-dimensional models; in the preprocessing step, a preset algorithm or a matching model is called to perform desensitization processing, content identification and / or data cleaning on the data to be processed; in the intelligent processing step, a corresponding model is matched according to the preprocessed data to be processed and the scene task request, and the corresponding model is called to process the data to be processed; in the post-processing step, in response to the user's operation type, the output result of the intelligent processing step is processed corresponding to the operation type, and the operation type includes one or more of the following: retry, like, copy, translate, forward, save, summarize, classify, error, language type, wherein when the operation type is retry, the intelligent processing step is executed again; in the output step, the target content is output in a predetermined format, and the predetermined format includes one or more of the following: text, audio and video, digital human, machine instruction.

[0007] Optionally, in the above method, the input step receives data from a first terminal, and the output step outputs the result to a second terminal, wherein the first terminal and the second terminal both include one of the following device types: mobile phone, computer, AR glasses, car, server.

[0008] Optionally, the above method further includes: when the execution of the previous task is completed, after receiving the user's modification or confirmation operation, entering the execution of the next task; adjusting the execution of the next task according to the user's modification or confirmation operation.

[0009] Optionally, in the above method, calling the matching model based on the model request includes: for each model request, calculating a model that matches both the data type of the data to be processed and the scenario task request through a recommendation algorithm.

[0010] Optionally, in the above method, each model is scored according to the domain type of the scene in the scene task request, the data type and text length of the data to be processed, the format and text length of the target content, user level information and / or the current load rate of the model; and the matching model is selected based on the scoring results, wherein the domain type includes general domains and specific domains.

[0011] Optionally, in the above method, before calculating the matching model, the scenario task request is added to a priority queue, and queues at the same level execute the request in order of request time; and computing resources are allocated according to preset weights corresponding to each priority queue.

[0012] Optionally, the above method further includes: predicting a matching task type based on the data to be processed; if the predicted task type is inconsistent with the type in the scenario task request, prompting to change the task type or directly use the matching task type.

[0013] Optionally, the above method also includes: storing the current scene task request and the target content, the task combination information and model information used to generate the target content in correspondence as instances; and executing the corresponding task combination again when receiving an access request from the user.

[0014] According to the second aspect of the present invention, a storage medium is also provided, which stores one or more programs, and the one or more programs can be executed by one or more processors to implement the scenario-based AIGC content generation method described in any of the above technical solutions.

[0015] According to the third aspect of the present invention, a scenario-based content generation system is also provided, including: a processor and a memory, wherein the processor is used to execute a scenario-based content generation program stored in the memory to implement the scenario-based AIGC content generation method described in any of the above technical solutions.

[0016] In some embodiments of the present application, after receiving the user's interaction data, task combinations and model requests are automatically generated, and appropriate models are called to process the user's complex needs to meet the user's needs in various complex scenarios without prompting the user and asking the user to make decisions. The whole process is seamless and can continuously and stably output high-quality answers. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings, in which:

[0018] Figure 1 is a schematic diagram of a scenario-based content generation system according to an embodiment of the present invention;

[0019] Figure 2 is a schematic diagram of a scenario-based content generation process according to an embodiment of the present invention;

[0020] Figure 3 is a flowchart of a scenario-based AIGC content generation method according to an embodiment of the present invention;

[0021] Figure 4 is a schematic diagram of a user input interface according to an embodiment of the present invention;

[0022] Figure 5 is a schematic diagram of a user output interface according to an embodiment of the present invention;

[0023] Figure 6 is a schematic diagram of a user input interface according to an embodiment of the present invention;

[0024] Figure 7 is a schematic diagram of a user output interface according to an embodiment of the present invention;

[0025] Figure 8 is a schematic diagram of a result saving interface according to an embodiment of the present invention;

[0026] Fig. 9 is a schematic diagram of a scenario constituted by parallel execution of tasks according to an embodiment of the present invention;

[0027] Fig.10 is a schematic diagram of performing comparison processing and display of two versions of input content according to an embodiment of the present invention;

[0028] Fig.11 is a comparative schematic diagram of processing the same input content in different ways according to an embodiment of the present invention;

[0029] Fig.12 It is a schematic diagram of a process of receiving user modification during serial task execution according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The preferred embodiments of the present application are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present application and are not used to limit the scope of protection of the present application. Those skilled in the art can make adjustments to them as needed to adapt to specific application scenarios.

[0031] It should be noted that in the description of this application, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the drawings. This is only for the convenience of description, and does not indicate or imply that the relevant devices or components must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on this application. In addition, the ordinal numbers "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0032] In addition, it should be noted that in the description of this application, unless otherwise clearly specified and limited, the terms "installation" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For those skilled in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0033] Figure 1 It is a schematic diagram of a scenario-based AIGC content generation system according to an embodiment of the present invention. Generative artificial intelligence AIGC (Artificial Intelligence Generated Content). The scenario development platform (ScenarioFactory) is a low-code development platform that supports the definition of AI Scenario, and has an AI scenario definition set produced based on the scenario development platform. An AI scenario is formed by a serial and parallel combination of N AI tasks (AI Task). The task development platform (Task Factory) is a low-code development platform that supports custom AI Tasks, and has an AI task definition set produced based on the task development platform. Support operators to customize commercial AI Tasks. Through the low-code development interface, users can add multiple tasks in a scenario and define the serial and parallel relationship between tasks.

[0034] In a specific application, on the user side, the user triggers a request, selects a scenario or task definition, and imports or inputs or adds documents or data to be processed, that is, the user's interaction data is generated. The scenario or task selected by the user can be understood as a scenario task request for the interaction data, and the imported document or data can be understood as the data to be processed in the interaction data. The system submits the interaction data to the scenario execution engine. The scenario execution engine will call the corresponding tasks in sequence through the task execution engine according to the interaction data.

[0035] The task execution engine will convert the input data to be processed, the type of task selected, and the definition of the task into a model request, and input the model request to the model calling module. The model calling module will intelligently call the model platform (Model Farm) in the background. In a specific application, suppose the task of the commercial AI application selected by the user is "reading and understanding policy documents", and the input content is a policy document, such as the 2023 Government Work Report of Pudong New Area, Shanghai. The task type is "policy document interpretation". After passing this information to the model calling module, the model calling module recognizes that it is a long text interpretation task, and will not regard it as an AI task such as short article writing, dialogue, video scene interpretation, big data BI interpretation, etc., so as to call the most appropriate model to process it.

[0036] In the specific calling process, the model request will enter a priority queue, and the queues at the same level will be executed in a first-come, first-served time order. Secondly, a weighted resource sharing mechanism is adopted. High-priority queues will be executed first according to a weight (such as 80%) of resources, and low-priority queues will use fewer resources. When the high-priority queue is empty, the low-priority queue can use all resources. When one queue is empty, the other queue will use 100% of the resources. When both queues have requests, the high-priority and low-priority queues will share resources according to the set weights. It should be understood that multi-level settings are supported, not limited to the high and low levels in this embodiment.

[0037] The model warehouse maintains multiple different types of AI models, including three typical categories of models, such as 1. LLM models with different characteristics and sizes; 2. Specific domain models trained based on industry data; 3. Other multimodal models that process images, audio and video data (such as audio and video models). Flexible model recommendation and automatic matching are achieved through the model call module. The specific recommendation and matching process will be explained below.

[0038] For each model request, automatic matching is performed through a multi-dimensional recommendation algorithm. For example, based on the field recommendation algorithm, it is determined whether the service content (that is, the acquired user interaction data) belongs to a general field or a specific field (such as medical, education, law, etc.), and each model is scored.

[0039] Based on the content modality recommendation algorithm, each model has a recommendation score based on the format of the input and output content (text, pictures, videos, audio, data), etc.

[0040] The recommendation algorithm is based on the text length and output content length. Some models support 128K content length, while others only support 8K but the answers are more accurate. Each model has a recommendation score.

[0041] Based on the user's level and payment preference, there is a cost-effectiveness recommendation model, and multiple models will also be scored.

[0042] Finally, integrate the recommendation scores and compile a list of recommended order for the system to select.

[0043] The task execution engine will return the task execution results to the scene execution engine. The scene execution engine will retain a set of scene instances for a long time and return the real-time results to the user at the same time.

[0044] A task instance contains all the information required for task execution. Users can return to access historical AI execution results at any time and re-execute and re-output AI tasks.

[0045] Combine the following Figure 1 , Figure 2 and Figure 3 , further illustrating the scenario-based content generation process of the present invention.

[0046] like Figure 3 As shown, in this embodiment, a scenario-based AIGC content generation method according to the present invention includes: step 302, obtaining the user's interaction data, the interaction data at least includes a scenario task request and data to be processed; step 304, determining the corresponding task combination according to the scenario task request and the data to be processed, and generating a corresponding model request; step 306, calling the matching model based on the model request, processing the data to be processed, and generating the target content. Through the above method, the cross-domain and cross-enterprise generalization implementation capabilities of multiple AI workflows can be quickly realized, and the ability to realize complex business processes in parallel with multiple AI Task strings is realized. At the same time, the ability to call and seamlessly connect complex multi-model, multi-modal, and multi-domain models is also realized, so that users only need to focus on the tasks at hand, and the system will automatically make the best choice of model.

[0047] Specifically, Figure 1In the scenario development platform, scenario 1 and scenario 2 are defined, and the task development platform defines tasks A, B, C, and D. For example, the execution of scenario 1 requires a combination of tasks A, B, and D, and these tasks can be in a serial relationship, in a parallel relationship, or in a series-parallel combination relationship. For example, task A and task B are connected in parallel, and the output result is given to task D. Users can also define the series-parallel relationship between these tasks according to their actual needs. The execution of scenario 2 requires a task combination of tasks B, C, and D, and these tasks can be in a serial relationship or in a parallel relationship. Therefore, the task combination includes the execution order of each task, where the output result of one task execution is the input of another task execution, or multiple tasks are executed simultaneously, or the output results of multiple tasks are the input of other task executions. This defines the corresponding relationship between any scenario and multiple task combinations.

[0048] Each task consists of multiple execution steps, see Figure 2 , Task A or Task B includes five fixed execution steps: input step, preprocessing step, intelligent processing step, post-processing step and output step. Figure 2 In the scenario shown, the implementation of the scenario is completed by the combined execution of task A and task B, and the execution result of task A is used as the input of task B, and the output of task B is the final result. The combination relationship is a serial relationship.

[0049] In the user interface, a trigger action of the user is received, and a scenario task request of the user is obtained, such as inputting a sales contract and reviewing a contract. The system can determine the scenario and the corresponding task combination corresponding to the scenario task request, wherein the scenario task request includes a task type and / or a scenario type. For example, the task type is reviewing a contract, and the scenario type is for example, a contract or a sales contract.

[0050] After determining the task combination, a model request is generated according to the multiple execution steps in the corresponding task combination. The task execution engine sends the model request to the model calling module, and the model calling module calls the appropriate model from the model repository for use.

[0051] like Figure 2 As shown, the multiple execution steps in each task include: an input step, a preprocessing step, an intelligent processing step, a post-processing step, and an output step.

[0052] In the input step, 1. Support different types of input content formats, including but not limited to:

[0053] Documents (documents can be entered via keyboard, as well as in TXT, WORD, PDF and other formats); Audio (meeting recordings, music, TV and movie recordings and other files can be entered); Pictures (document photos, pictures, etc. can be entered); Videos (mobile phone videos, security camera videos, promotional videos, movie and TV content); URLs (website addresses, mobile public accounts, Weibo, Douyin accounts, etc.); Data (excel, CSV, data interfaces, etc.); and it can also be expanded to support other data types, such as 3D models.

[0054] 2. Support the expansion of multiple terminal inputs through Web API, including but not limited to: PC, mobile phone, sensor (support input of Internet of Things device's network address, account, password, etc. in the input step), device, robot, and other devices that can initiate calls to Web API.

[0055] In the preprocessing step, the preprocessing of the input content is supported and executed before the main task processing in the intelligent processing step. The main processing content includes at least the following aspects:

[0056] 1. For documents, audio and video content, perform sensitive word detection, such as approximate search of sensitive word vector space based on sensitive word blacklist; 2. Desensitization of privacy information; 3. Audio and video text conversion, text recognition, and image content interpretation. 1) It is based on the understanding of audio and video content and combined with the application of business scenarios. For example, the analysis of meeting minutes, the statement of the views of both parties, and the next task arrangement of both parties, as well as the content minutes of court trials, the content summary and labels of online video courses, and the analysis of emotions and value propositions of competitive posters. 2) In the preprocessing step, the purpose is to extract the content in the audio and video to prepare for business analysis. 4. For the input URL, automatically read the web page content and make HTML presentation. Prepare for the next step of analyzing the website content, key points and propositions. 5. For the address of the database, automatically read the data and prepare for the next step of processing: 1) Read the data and do simple cleaning, 2) Interpret and judge based on the data.

[0057] The AI ​​processing step is the main purpose and main task of this content generation process application, while the pre-processing step and post-processing step are auxiliary processing.

[0058] The main processing object types of the intelligent processing step include text processing, data processing, and multimedia processing. The intelligent processing step is the key to achieving continuous, stable, and high-quality AI content. The intelligent processing step supports multiple AI agents, one of which is responsible for determining whether the content input by the user matches the selected Task type. If not, the user can be prompted to change the Task type, or the system can directly replace it with a suitable Task type. By training a domain model to make a Document Type Classifier, the main document types in the domain can be identified. Construct a knowledge graph relationship between document types and task types. If the input content type and task type do not match (for example, a lease contract is entered, but a divorce agreement review is selected), it can be predicted with a high probability that the user actually wants to perform a lease contract review. If the prediction cannot be accurate, an error prompt is returned to the user.

[0059] One AI Agent will select the model to use based on the current task (LLM (Large Language Model) of different sizes, domain model, multimedia model), one AI Agent will be responsible for prompt changes and result evaluation, and continuous tuning will be achieved by introducing a random walk mechanism. Another AI Agent will monitor the content generated and returned to ensure that it does not contain sensitive words. If it does, it can output failure.

[0060] Therefore, the ability to call and seamlessly connect complex multi-model, multi-modal, and multi-domain models is realized, so that users only need to focus on the task at hand, and the system will automatically make the best choice of model. The ability to implement complex business processes by serializing and parallelizing multiple AI tasks is realized.

[0061] For the output of the intelligent processing step, perform one or more of the following post-processing steps:

[0062] 1. Retry. If the user is not satisfied with the first result, he can click Retry, and the system will execute AI Processing again, and return the new result and history for comparison. Retry means that the customer is not satisfied with the result this time; 2. Like; 3. Copy; 4. Translate; 5. Forward; 6. Save. In addition to retrying, other operations reflect the user's satisfaction and acceptance of the results. This information will be saved and used in RLHF (Reinforced Learning from Human Feedback). The system's RLHF is not only for domain fine-tuning models, but also for prompt guidance. Through the initial prompt given by the operator and based on the generation capability of AI, different versions of slightly differentiated prompts are generated, and the prompt is optimized through enhanced learning through user feedback. The ability to permanently save AI task calls, search across the entire domain, continue to use, and resume transfers from breakpoints is realized.

[0063] In the output step (AI Output), the output step is the output end and the receiving end of the AI ​​Task execution result. AIOutput supports a variety of media, including text, audio and video, digital humans, machine instructions, data, etc. The input and output can be on different machines.

[0064] The output steps can be connected to PCs, mobile devices, ARVR glasses, smart devices, robots, databases, etc. according to the needs of commercial scenarios. The system can see the output content through the web page and the content is sent to the specified device terminal.

[0065] The above details the workflow of the content generation method. Through the pre-set five-step AI work processing mechanism, automatic identification of tasks, automatic matching of models, automatic tuning of prompts, automatic quality monitoring, and reinforcement learning based on user feedback, high-quality and stable output results are achieved to avoid the hallucination of large models.

[0066] Based on the same development framework, it can support multiple media such as text, audio and video, digital humans, data processing, etc., and adapt to multiple models (general LLM, domain large model, visual large model, etc.)

[0067] A generalized AI Task implementation mechanism is designed, which can receive data from multiple device ends and return the results to multiple scalable device ends, so that LLM applications are not limited to copywriting work, but can be extended to real production environments (such as automatic analysis reports of trial records, access to the trial recording system, and output to mobile phones; such as visual signal access of AR glasses, background processing, and output on the glasses end).

[0068] Based on the same development environment, different task processing mechanisms can be developed in the background, unifying the development of multi-scenario AI tasks. In the execution of an AI Task, based on the multi-AI Agent mechanism, multi-angle task quality assurance and automatic model optimization mechanism can be achieved.

[0069] Integrating user interaction into every aspect of AI Task not only realizes the improvement of quality through human intervention, but also realizes the supervised learning process.

[0070] In terms of user interaction, the entire content generation call method is concentrated in two AI Editors, one of which handles input (IA, input AIDitor) and the other handles output (OA, output AIditor). Figure 4 , Figure 5 , Figure 6 Figure 7 and Figure 8 The user interaction method and practical application scenario examples of the present invention are described.

[0071] Figure 4 It is a simplified version of the input editor interface. Figure 4 The sticky note in the middle represents the type of task to be processed in the input step. Figure 4 The label below is the task type to be processed in the pre-processing step (Pre-Processing), and the IA edit box displays the pre-processed input content. Figure 5 It is a simplified version of the output editor interface. Figure 5 The upper label defines the task type processed by the AI ​​Processing step (such as 4 tasks, task! represents the task type selected or being executed). The subsequent AI Processing steps will be processed according to the selected task type. Figure 5 The OA edit box (Output) shows the processed output content. The lower label of Figure 5 shows the types of tasks that need to be processed in the post-processing step (Post-Processing), such as summary, translation, classification, like, and error reporting. Figures 6 to 8 It is a screenshot of a specific application scenario.

[0072] User in Figure 6 The interface IA shown in the figure imports the multimodal files to be processed (Word, PDF, photos, videos, audios, databases, web addresses, etc.), and the contract content imported in this figure. The multimodal management is achieved through the upper labels of IA (such as typing, files). The lower labels of IA contain multiple pre-processing operations, such as translation (Chinese, English, Japanese, French), digital number (the current displayed number of words is 1344), etc.

[0073] like Figure 7 As shown, the name of the tab page in the output OA interface defines the various AI operations to be performed on the content, that is, the task type is defined, such as (format review, contract summary, text-to-picture, picture-to-text, and other AI operations). The edit box of the output interface contains the content generated after AI processing. This figure shows the result of the format review of the contract, which can support various formats such as TEXT, video, picture, audio, and digital human. The lower label of the edit box in the OA interface is the post-processing task type for the output. Including being dissatisfied with the previous AI answer and outputting it again. Like, copy, forward, translate, etc., which can be expanded as needed. At the same time, the past execution results are continuously saved, and through the stateless mechanism, the intermediate results of each task execution can be retained at any time, revisited at any time, and resumed at breakpoints. As Figure 8 As shown, the interface shows that the execution history of previous patent service contracts, guarantee contracts, property contracts, and transportation contracts is saved. Click on one of the contracts to see the previous processing status.

[0074] Those skilled in the art should understand that the input editor interface and the output editor interface according to the present invention can be flexibly designed according to the scene requirements, and are not limited to the layout shown in the figure. The overall layout of the input editor interface and the output editor interface can be arranged up and down, down and up, left and right, right and left, and can display languages ​​from right to left according to local cultural differences. The labels on the editor can be arranged up and down, left and right, or one on the left and one on the bottom, and can be flexibly set.

[0075] As described above, the scenario-based content generation method of the present invention supports serial and parallel processing of multiple tasks to complete more complex commercial scenarios, as described below.

[0076] Fig. 9 The displayed interface is based on a user input content 902 , and has several associated independent and parallel (4 task tags) tasks, forming a scenario 904 . Fig.10 There are two versions of input content displayed. In the input boxes at the upper left 102 and the upper right 104, there can be original text translation, old version of the law versus new version of the law, Huangpu District annual report versus Pudong District annual report, 2023 annual report versus 2024 annual report. The lower box 106 is an analysis of the above-mentioned comparison content, which can be a difference analysis of the two contents, analysis of the same items, analysis of advantages and disadvantages, analysis of characteristics, comparison analysis of specific values, etc. The comparison result is the target content finally output by the scenario-based AIGC content generation method according to the present invention.

[0077] Fig.11What is shown is that for the same input content, such as the upper frame 110, it can be processed in two different ways, such as the lower right frame 114 and the lower left frame 112, to compare the difference in the effect of the processed content. The two different ways can be the comparison of the processing results of an 8K model of a large LLMModel language model and the processing results of a 128K model, or the comparison of a 7B model and a 13B model, or the comparison of a general model and a domain model, or the comparison of different versions of prompts, or the comparison of different embedding technologies. The comparison result is the target content finally output by the scenario-based AIGC content generation method according to the present invention. This means that through the comparison of different processing methods, users can choose the appropriate processing method to obtain satisfactory output results.

[0078] Fig.12 What is presented is that the content to be processed is input on the input interface 120, and the execution result of the first TASK is passed to the second TASK as input. After the user makes modifications and confirmations on the first TASK interface 122, the second Task is triggered, as shown in the interface 124, and so on. For example, the first TASK generates an article summary, and the second TASK generates detailed content based on the article summary. The first TASK proposes the key points of analysis for the input content. The second TASK draws specific analysis conclusions based on the input content and the analysis framework. The analysis conclusion is the target content finally output by the scenario-based AIGC content generation method according to the present invention.

[0079] Therefore, the technical solution of the present invention provides a general, commercially available scenario-based content generation solution, which uses a corresponding model between scenarios and tasks, and a fixed multi-step task execution method, so as to obtain high-quality and stable output results based on specific scenarios, and avoid the boundless fantasy of artificial intelligence. Through intelligent model calling, automatically matching models (general language models of different versions and sizes, domain models, and multimodal models), selecting general or domain models according to task classification, or selecting model size according to the length of input content, or selecting multimodal models according to multimedia types, or automatically balancing loads according to the intensity of model use to avoid overloading a model. The ability to quickly implement scenario implementation. There are many and diverse AI scenarios that can be commercialized, and it is necessary to be able to quickly and cost-effectively support the expansion of multiple scenarios while ensuring the quality of answers for each scenario. Through the scenario development platform and the task development platform, low-code rapid development of task and scenario definitions is achieved. In addition, operators or customers can also customize new tasks. At the same time, it is necessary to be able to support serial and parallel processing of multiple AI tasks to complete more complex commercial scenarios. Multimodal switching. In the processing of some specific tasks, fine-tuning models or other modal models for specific fields may be used. It is necessary to be able to switch seamlessly between general models, multimodal models and domain models. Through the encapsulation of AI Tasks and the use of AI agent mechanisms, seamless switching of multiple models can be achieved based on content classification, task classification, media classification, and content length, without the user noticing. Past execution results are continuously saved. Through the stateless mechanism, the intermediate results of each task execution can be retained at any time, revisited at any time, and resumed at breakpoints.

[0080] According to another embodiment of the present invention, a storage medium is also provided, which stores one or more programs, and the one or more programs can be executed by one or more processors to implement the scenario-based AIGC content generation method according to the present invention.

[0081] In one embodiment, a scenario-based content generation system according to the present invention includes: a processor and a memory, wherein the processor is used to execute a scenario-based content generation program stored in the memory to implement a scenario-based AIGC content generation method. The processor may be various types of processors, such as but not limited to a central processor unit (CPU), a digital signal processor (DSP), a neural network processor (NPU), a graphics processor (GPU), a field programmable gate array (FPGA), and an application-specific integrated circuit (ASIC). In this embodiment, the number of processors may be one or more. The memory includes non-volatile memories such as flash memory, ROM, hard disk drives, disks, and optical disks, and various types of dynamic random access memories and active memories. The memory is coupled to the processor so that the processor can read and write information from / to the memory. In an alternative solution, the memory may be integrated into the processor. In another alternative solution, the processor and the memory may exist in the form of discrete components.

[0082] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A scenario-based AIGC content generation method, characterized in that: include: Acquire user interaction data, where the interaction data at least includes a scenario task request and data to be processed; Determine a corresponding task combination according to the scenario task request and the data to be processed, and generate a corresponding model request; Based on the model request, a matching model is called to process the data to be processed to generate target content; Among them, on the user side, the user triggers a request, selects a scenario or task definition, and imports or inputs or adds data to be processed, that is, generates user interaction data; The step of determining a corresponding task combination according to the scenario task request and the data to be processed, and generating a corresponding model request, includes: It has a scenario definition set produced based on the scenario development platform and a task definition set produced based on the task development platform. It defines the correspondence between any scenario and a combination of multiple tasks. Each task includes multiple execution steps. A scenario is formed by a serial and parallel combination of multiple tasks. It supports users to add multiple tasks in a scenario and define the serial and parallel relationship between tasks. Determine a scene and a corresponding task combination corresponding to the scene task request, wherein the scene task request includes a task type and a scene type; A model request is generated according to the multiple execution steps in the corresponding task combination.

2. The scenario-based AIGC content generation method according to claim 1, characterized in that: The task combination includes the execution order of each task, wherein the output result of the execution of one task is the input of the execution of another task, or multiple tasks are executed simultaneously, or the output results of multiple tasks are the input of the execution of other tasks; The multiple execution steps include: an input step, a pre-processing step, an intelligent processing step, a post-processing step and an output step.

3. The scenario-based AIGC content generation method according to claim 2, characterized in that: In the input step, the input data type includes one or more of the following: documents, voice, pictures, videos, websites, data in different formats, and three-dimensional models; In the preprocessing step, a preset algorithm or a matching model is called to perform desensitization, content identification and / or data cleaning on the data to be processed; In the intelligent processing step, the corresponding model is matched according to the pre-processed data to be processed and the scene task request, and the corresponding model is called to process the data to be processed; In the post-processing step, in response to the user's operation type, the output result of the intelligent processing step is processed corresponding to the operation type, and the operation type includes one or more of the following: retry, like, copy, translate, forward, save, summarize, classify, error report, language type, wherein when the operation type is retry, the intelligent processing step is executed again; In the output step, the target content is output in a predetermined format, and the predetermined format includes one or more of the following: text, audio and video, digital human, and machine instructions.

4. The scenario-based AIGC content generation method according to claim 3 is characterized in that: The input step receives data from a first terminal, and the output step outputs a result to a second terminal, wherein the first terminal and the second terminal both include one of the following device types: a mobile phone, a computer, AR glasses, a car, and a server.

5. The scenario-based AIGC content generation method according to claim 2, characterized in that: Also includes: At the end of the execution of the previous task, after receiving the modification or confirmation operation of the user, the execution of the next task begins; Adjust the execution of the next task based on the user's modification or confirmation operation.

6. The scenario-based AIGC content generation method according to claim 1, characterized in that: The calling of the matching model based on the model request includes: For each model request, a model that matches both the data type of the data to be processed and the scenario task request is calculated through a recommendation algorithm.

7. The scenario-based AIGC content generation method according to claim 6, characterized in that: Scoring each model according to the domain type of the scenario in the scenario task request, the data type and text length of the data to be processed, the format and text length of the target content, user level information and / or the current load rate of the model; The matching model is selected according to the scoring result, wherein the field type includes a general field and a specific field.

8. The scenario-based AIGC content generation method according to claim 6 is characterized in that: Before calculating the matching model, the scenario task request is added to the priority queue, and the queues at the same level are executed in the order of request time; Computing resources are allocated according to the preset weights corresponding to each priority queue.

9. The scenario-based AIGC content generation method according to any one of claims 1 to 8, characterized in that: Also includes: Predicting a matching task type according to the data to be processed; If the predicted task type is inconsistent with the type in the scenario task request, a prompt is given to change the task type or to directly use the matching task type.

10. The scenario-based AIGC content generation method according to any one of claims 1 to 8, characterized in that: Also includes: The current scenario task request and the target content, the task combination information and the model information used when generating the target content are stored as instances in correspondence; When an access request is received from the user, the corresponding task combination is executed again.

11. A storage medium, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the scenario-based AIGC content generation method according to any one of claims 1 to 10.

12. A scenario-based content generation system, characterized in that: include: A processor and a memory, wherein the processor is used to execute a scenario-based content generation program stored in the memory to implement the scenario-based AIGC content generation method according to any one of claims 1 to 10.