Task processing method, model training method, intelligent conversation and model training platform

Abstract

The embodiment of the invention provides a task processing method, a model training method and an intelligent dialogue and model training platform. The task processing method comprises the steps of obtaining initial prompt information corresponding to a target task; the execution process of the target task is disassembled based on the initial prompt information through a natural language model, multiple target sub-tasks are obtained, target prompt information corresponding to the target task is generated, and the target prompt information comprises prompt information corresponding to the multiple target sub-tasks; and determining target task models of specifications corresponding to the plurality of target sub-tasks, and correspondingly executing the plurality of target sub-tasks by using the plurality of target task models based on the prompt information corresponding to the plurality of target sub-tasks to obtain task execution results of the target tasks. By disassembling the execution process of the target task and calling the target task models of different specifications, computing resources can be optimized, the model calling cost can be reduced, and the system efficiency can be improved.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of computer technology, and in particular to a task processing method, a model training method, an intelligent dialogue, and a model training platform. Background Art

[0002] Large Language Models (LLMs) are increasingly used in various application scenarios due to their powerful natural language understanding and generation capabilities. When dealing with complex tasks, large language models can understand and execute complex judgment logic, thereby completing many tasks that could only be done by humans in the past.

[0003] However, the training and maintenance costs of large language models, as well as the computing resource costs and time costs required in the task execution process are all high. Therefore, there are still problems that need to be optimized, such as long information analysis time and high analysis costs. Summary of the invention

[0004] In view of this, an embodiment of this specification provides a task processing method. One or more embodiments of this specification also involve a model training method, a method for processing a dialogue analysis task, an intelligent dialogue platform, a model training platform, a computing device, a computer-readable storage medium, and a computer program product to solve the technical defects existing in the prior art.

[0005] According to one aspect of an embodiment of this specification, a task processing method is provided, including: Get the initial prompt information corresponding to the target task; Using a natural language model, based on the initial prompt information, the execution process of the target task is disassembled to obtain multiple target subtasks, and target prompt information corresponding to the target task is generated, wherein the target prompt information includes prompt information corresponding to the multiple target subtasks; Determine a target task model with corresponding specifications of multiple target subtasks, and use the multiple target task models to execute the multiple target subtasks based on prompt information corresponding to the multiple target subtasks to obtain a task execution result of the target task.

[0006] In this way, by obtaining the initial prompt information corresponding to the target task, a data basis is provided for the subsequent decomposition of the execution process of the target task; by determining the target task models of the corresponding specifications of multiple target subtasks, task models of different specifications have different requirements for computing resources. By selecting a target task model with appropriate specifications for each subtask, it is possible to avoid using overly complex or resource-intensive models to process simple subtasks, thereby optimizing the utilization of computing resources, reducing the model call cost and improving the operating efficiency of the system; by decomposing the target task into multiple target subtasks, and multiple target subtasks corresponding to clear prompt information and target task models of corresponding specifications, a multi-agent scheduling method can be used to execute multiple target subtasks, thereby improving the flexibility and efficiency of task processing, and improving the task processing efficiency and reducing the model call cost on the basis of ensuring the accuracy of task results, thereby improving the overall performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is an application architecture diagram of a task processing provided by an embodiment of this specification; Figure 2 is a flowchart of a task processing method provided by an embodiment of this specification; Figure 3 is a flow chart of a model training method provided by an embodiment of this specification; Figure 4 is a flowchart of a method for processing a conversation analysis task provided by an embodiment of this specification; Figure 5 It is a processing flow diagram of a method for processing a conversation analysis task provided by an embodiment of this specification; Figure 6 It is a structural diagram of an intelligent dialogue platform provided by an embodiment of this specification; Figure 7 This is a schematic diagram of a front-end interface of an intelligent dialogue platform provided by an embodiment of this specification; Figure 8 It is a structural diagram of a model training platform provided by an embodiment of this specification; Fig. 9 is a structural diagram of a task processing device provided by an embodiment of this specification; Fig.10 It is a structural schematic diagram of a model training device provided by an embodiment of this specification; Fig.11 It is a structural diagram of a device for processing a conversation analysis task provided by an embodiment of this specification; Fig.12 It is a structural block diagram of a computing device provided by an embodiment of this specification. DETAILED DESCRIPTION

[0008] Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.

[0009] The terms used in one or more embodiments of this specification are only for the purpose of describing a specific embodiment, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items. It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein can be interpreted as "at the time of..." or "when..." or "in response to determination".

[0010] In addition, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0011] In one or more embodiments of this specification, a large model refers to a deep learning model with large-scale model parameters, which usually contains hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than 10 trillion model parameters. A large model can also be called a foundation model / foundation model. The large model is pre-trained with large-scale unlabeled corpus to produce a pre-trained model with more than 100 million parameters. This model can adapt to a wide range of downstream tasks, and the model has good generalization ability, such as a large-scale language model (LLM), a multi-modal pre-training model, etc. When the big model is used in practice, only a small number of samples are needed to fine-tune the pre-trained model and it can be applied to different tasks. The big model can be widely used in natural language processing (NLP), computer vision and other fields. Specifically, it can be applied to computer vision tasks such as visual question answering (VQA), image caption (IC), image generation, as well as natural language processing tasks such as text-based sentiment classification, text summary generation, and machine translation. The main application scenarios of the big model include digital assistants, intelligent robots, search, online education, office software, e-commerce, intelligent design, etc.

[0012] First, the terms involved in one or more embodiments of this specification are explained.

[0013] prompt: It is the prompt information, which can include input text or questions provided to the large model to guide the model to generate specific responses or outputs.

[0014] Natural language model: a computer model based on artificial intelligence technology that can understand, process and generate natural language.

[0015] Agent: It is an intelligent entity with characteristics such as autonomy, interactivity, responsiveness and initiative, also known as an agent, which can perform assigned tasks.

[0016] In this specification, a task processing method is provided. This specification also involves a model training method, a method for processing a dialogue analysis task, an intelligent dialogue platform, a model training platform, a computing device, a computer-readable storage medium, and a computer program product, which are described in detail one by one in the following embodiments.

[0017] Considering the huge amount of model parameters of the large model and the limited computing resources of the mobile terminal, the task processing method provided in the embodiment of the present application can be applied to Figure 1 The application scenarios shown are not limited to these. Figure 1 In the application scenario shown, the large model is deployed in the server 104, and the server 104 can be connected to one or more client devices 102 through a local area network connection, a wide area network connection, an Internet connection, or other types of data networks. The client device 102 here may include but is not limited to: smart phones, tablet computers, laptops, PDAs, personal computers, smart home devices, vehicle-mounted devices, etc. The client device 102 can interact with the user through a graphical user interface to call the large model, thereby implementing the method provided in the embodiments of this specification.

[0018] In an embodiment of the present specification, a system composed of a client device and a server can perform the following steps: the client device receives task information of a conversation analysis task input by a user, and task information of multiple preset subtasks of the conversation analysis task, and sends the task information of the conversation analysis task and the task information of multiple preset subtasks of the conversation analysis task to the server. The server uses a natural language model to integrate the task information of the conversation analysis task and the task information of multiple preset subtasks, obtains integrated prompt information corresponding to the conversation analysis task, rewrites the integrated prompt information, and obtains initial prompt information corresponding to the conversation analysis task; uses a natural language model to disassemble the execution process of the conversation analysis task based on the initial prompt information, obtains multiple target subtasks, and generates target prompt information corresponding to the conversation analysis task, wherein the target prompt information includes prompt information corresponding to multiple target subtasks; determines the conversation analysis task models corresponding to the specifications of the multiple target subtasks, and uses the multiple conversation analysis task models to execute the multiple target subtasks based on the prompt information corresponding to the multiple target subtasks, and obtains the task execution result of the conversation analysis task; and feeds back the task execution result to the client. It should be noted that, when the operating resources of the client device can meet the deployment and operating conditions of the large model, the embodiments of the present application can be carried out in the client device.

[0019] See also Figure 2 , Figure 2 A flowchart of a task processing method provided according to an embodiment of the present specification is shown, which specifically includes the following steps.

[0020] Step 202: Obtain initial prompt information corresponding to the target task.

[0021] The embodiments of this description are applied to applications, platforms or systems with task processing functions, such as intelligent customer service platforms, intelligent legal consulting platforms, online medical diagnosis systems, virtual character dialogue applications, and code writing question and answer systems. Furthermore, they can also be applied to applications, platforms or systems that perform dialogue analysis tasks, such as intelligent dialogue platforms that can perform dialogue quality inspection on customer service dialogues.

[0022] The target task is a specific problem or requirement that needs to be solved through model reasoning. The target task usually involves extracting knowledge from given information, understanding the background, performing logical analysis, and then completing reasoning to obtain task results. The target task can be a dialogue analysis task. The target task may include multiple preset subtasks, and different preset subtasks have different task types or task goals, such as summary, extraction, analysis, quality inspection, event type recognition, etc. Different target tasks may involve different fields, generally have high complexity, and need to be analyzed and processed by large language models. For example, in a customer service dialogue scenario, users and customer service have a dialogue on how to check account balances and how to pay fees. The target task is to conduct a dialogue analysis of the dialogue content in this scenario, including key information extraction, question-answer pair extraction, identification of problem solutions, service quality inspection, event type recognition and other multi-dimensional dialogue analysis tasks. Among them, key information extraction is to screen out information that is of great value for subsequent analysis and decision-making from a large amount of dialogue text. This information can usually summarize the core content of the dialogue or is related to specific task goals. For example, in customer service conversations, key information such as customer needs and product problems are extracted to quickly respond to and solve problems; in market research conversations, key content such as consumers' evaluation and suggestions on products are extracted to provide a basis for product improvement and market strategy adjustment. Question and answer pair extraction is used to extract content presented in the form of questions and answers in the conversation text to form structured question and answer pair data. Identifying problem solutions can be understood as analyzing the content of the conversation, determining the problems raised and the corresponding solutions, so as to help sort out the problem handling situation in the task process. For example, in after-sales service conversations, product failure problems reported by customers and solutions provided by customer service are identified to optimize the after-sales service process and improve customer satisfaction; in project discussion conversations, problems encountered in the project and solutions proposed by team members are identified to provide guarantees for the smooth progress of the project. Service quality inspection can be understood as evaluating the service quality of service personnel through analysis of the content of the conversation, checking whether the service process meets the specifications, whether it meets customer needs, and whether it can quickly and specifically solve customer problems. For example, in an online customer service chat scenario, quality inspection of the chat records between customer service and users can supervise service quality and promptly discover and correct problems in the service. Event type identification can also be understood as labeling and classifying the content of the conversation. It can determine the type of event involved based on the content of the conversation, such as a complaint event, consultation event, suggestion event, etc., so as to classify and manage the event and subsequently process it.

[0023] The initial prompt information can be understood as the prompt initially input to the natural language model, that is, the initial prompt information of the target task. It can be used to guide the natural language model to analyze and process the target task, and contains basic information and requirements about the target task. In an optional embodiment of the present specification, the initial prompt information can be obtained by defining the target task. Further, the task definition of the target task can include: uploading the conversation content, specifying preset subtasks such as conversation quality inspection, information mining, causal analysis, specifying the conversation scene, the character settings involved in the conversation, the goals of the parties to the conversation, etc. Specifically, the uploaded conversation content can be in text format or other formats such as audio format. The preset subtasks can be understood as subtasks of different types and different goals. The conversation scene can be understood as the scene involved in the conversation content, such as after-sales service, pre-sales consultation, operation method guidance, etc. The character settings involved in the conversation can be used to indicate what kind of people the parties to the conversation are, such as customer service and customers, sales staff and consumers, etc. The goal of the dialogue parties can be understood as the purpose that the dialogue parties want to achieve through the dialogue. For example, the customer wants to know how to query the account information through the dialogue, and the customer service needs to tell the customer how to correctly query the account information through the dialogue. The above content can be set by the front-end user according to the needs of the actual application. Optionally, the preset subtasks such as dialogue quality inspection, information mining, and causal analysis can be pre-set by the platform and provided to the user for selection, or they can be set by the user according to their own needs.

[0024] In practical applications, by defining the target task, the initial prompt information of the target task can be obtained according to the task setting information defined for the target task, including preset subtasks, dialogue scenes, character settings involved in the dialogue, goals of the dialogue parties, etc., as well as the uploaded dialogue content. For example, an example of initial prompt information is given below.

[0025] “{Preset subtask} Label, summary, keyword, question and answer pair, problem solution, service quality inspection, key information, label classification.

[0026] {Task scenario} After-sales customer service of a large e-commerce platform communicates with customers to solve various problems encountered by customers during the shopping process, such as product quality problems, logistics delays, return and exchange requirements, etc. By analyzing the conversation, it is determined whether the customer service follows the service specifications, whether the customer problems are effectively solved, and whether there are any violations.

[0027] {Character settings involved in the dialogue} "Customer service": E-commerce platform employees who have received professional training and are familiar with the platform's after-sales policies and processes. They are responsible for communicating with customers, solving customer problems, and maintaining the platform's image and customer satisfaction. "Customer": Consumers who encounter problems after purchasing goods on the e-commerce platform and hope to solve the problem by communicating with customer service and obtain a satisfactory solution.

[0028] {Goals of all parties in the dialogue} "Customer service": follow the platform's service standards and use polite and professional language to communicate with customers. Accurately understand customer problems and provide reasonable solutions based on after-sales policies. Try to meet customer needs, improve customer satisfaction, and avoid complaints. "Customer": Clearly express the problems you encounter, such as damaged goods, inconsistent sizes, and delayed receipt of goods. Get positive responses and effective solutions from customer service, such as returns, replacements, compensation for losses, etc. Be satisfied with the solutions and safeguard your own legitimate rights and interests. " Obtaining the initial prompt information corresponding to the target task provides a data basis for subsequent decomposition to obtain multiple target subtasks, and then executing multiple target subtasks separately by calling different specification model combinations.

[0029] Step 204: using a natural language model, based on the initial prompt information, disassemble the execution process of the target task, obtain multiple target subtasks, and generate target prompt information corresponding to the target task, wherein the target prompt information includes prompt information corresponding to multiple target subtasks.

[0030] In practical applications, on the basis of obtaining the initial prompt information corresponding to the target task, the execution process of the target task can be disassembled based on the natural language model and the initial prompt information to obtain multiple target subtasks and generate target prompt information corresponding to the target task. Specifically, the natural language model is a model based on machine learning or deep learning technology that can understand, process and generate human natural language, such as the GPT series, BERT and other models, which can be used to perform various analyses and operations on natural language texts. The target subtask is a plurality of smaller and relatively independent subtasks obtained by disassembling the execution process of the target task, and each subtask is used to complete a component of the target task. The target subtask can also be understood as each execution step in the execution process of the target task, which is obtained by disassembling the execution process of the target task step by step. The target prompt information can be understood as the prompt information obtained after disassembling the execution process of the target task. The target prompt information can be understood as a prompt information set including multiple prompt information, including prompt information corresponding to each target subtask, which is used to guide the execution of the target subtask.

[0031] In an optional embodiment of this specification, using a natural language model, based on the initial prompt information, the execution process of the target task is disassembled to obtain multiple target subtasks, and the target prompt information corresponding to the target task is generated, which may include: using the natural language model to parse the initial prompt information, determine the execution process of the target task, disassemble the execution process of the target task, and obtain multiple target subtasks; based on the task types of multiple target subtasks, generate prompt information corresponding to multiple target subtasks, and obtain target prompt information corresponding to the target task. In practical applications, the natural language model can determine the execution process of the target task by parsing the initial prompt information, and the execution process may contain task definitions and processes of multiple subtasks. The natural language model can disassemble the execution process to obtain multiple target subtasks. Further, according to the task type of each target subtask obtained by disassembly, the prompt information corresponding to each target subtask can be generated, so as to obtain the target prompt information corresponding to the target task. The target prompt information can be used to guide the execution of each target subtask, and provide a data basis for the subsequent execution of multiple target subtasks by combining and calling different specification models.

[0032] Step 206: Determine target task models corresponding to the specifications of the multiple target subtasks, and use the multiple target task models to execute the multiple target subtasks based on the prompt information corresponding to the multiple target subtasks to obtain the task execution result of the target task.

[0033] In actual applications, after decomposing multiple target subtasks and generating prompt information corresponding to the multiple target subtasks, a target task model with corresponding specifications of the multiple target subtasks can be determined, and the multiple target task models can be used to execute the multiple target subtasks based on the prompt information corresponding to the multiple target subtasks to obtain the task execution result of the target task.

[0034] Specifically, the task type can be used to describe the nature and characteristics of the task. Different task types may require different processing methods and models of different specifications (for example, different sizes of models such as 7B and 72B). In the case where the target task is a conversation analysis task, the task type can specifically be the task type of subtasks under various dimensions included in the conversation analysis task, such as summary, extraction, problem solution, service quality inspection, label classification, etc. The task model can be understood as a model designed for different task types, with a specific structure and function, and capable of completing tasks of the corresponding type. Based on different task types, the task model can have different model specifications, for example, it can be a small model of 7B or a large model of 72B. The target task model can be understood as a model selected from task models of different specifications according to the task type of the target subtask, and used to execute the corresponding target subtask. The task execution result can be understood as the processing result of the target task obtained after executing multiple target subtasks through multiple target task models.

[0035] In an optional embodiment of the present specification, determining target task models of specifications corresponding to multiple target subtasks may include: determining target task models corresponding to the multiple target subtasks from task models of different specifications based on the task types of the multiple target subtasks. In practical applications, the natural language model may determine target task models of corresponding specifications for each target subtask from task models of different specifications based on the task types of each target subtask.

[0036] In another optional embodiment of the present specification, determining target task models of corresponding specifications for multiple target subtasks may also include: determining target task models corresponding to multiple target subtasks from task models of different specifications based on the amount of task resources corresponding to the multiple target subtasks. In practical applications, the natural language model may also determine the target task models corresponding to each target subtask based on the amount of task resources required for each target subtask during the task execution process. It should be noted that in addition to the task type and the amount of task resources, the natural language model may also determine the target task models of corresponding specifications for multiple target subtasks based on other factors such as more contextual information and the model performance of task models of different specifications.

[0037] Optionally, the processing result of the target task may specifically include a result obtained by combining the subtask processing results of multiple target subtasks, or may include the subtask processing results corresponding to multiple target subtasks. In the actual implementation process, the target task is generally a relatively complex task that needs to be inferred and processed by a large model. However, if the entire task execution process uses a large model, there will be problems such as high model call cost, reduced task processing efficiency, and high model training and maintenance costs.

[0038] For example, in the application scenario of conversation quality inspection, conversation quality inspection aims to evaluate the performance of customer service or sales in the service process, and then improve and optimize the service process based on the problems found. Since a quality inspection item usually involves multiple task concepts and contains complex judgment logic, it is often necessary to use the largest size model to understand the definition in order to obtain a relatively good effect. However, in actual scenarios, there may be a situation where the content to be identified usually has a very low ratio to the actual conversation data, which may be 1:100 or even more. For example, if you want to check a type of violation of customer service (such as asking for user private information, insulting users, mocking users, showing impatience, etc.), there may be only 10-100 calls out of 100,000 calls. A large number of conversations do not have such problems. Using the largest size model will obviously consume unnecessary costs.

[0039] According to an optional embodiment of the present specification, obtaining initial prompt information corresponding to the target task may include the following steps: obtaining task information of the target task and task information of multiple preset subtasks for the target task; using a natural language model to integrate the task information of the target task and the task information of multiple preset subtasks to obtain integrated prompt information corresponding to the target task, and rewriting the integrated prompt information to obtain initial prompt information corresponding to the target task.

[0040] Specifically, task information can be understood as various descriptive contents related to the task, including the task's objectives, scope, constraints, background information, etc. In the case where the target task is a dialogue analysis task, task information can specifically include dialogue content, preset subtasks, scenes involved in the dialogue, character settings, goals of the dialogue parties, etc.

[0041] Optionally, the target task may include multiple preset subtasks. The task information of the preset subtasks may be predefined by the platform or designed by the user. Optionally, the natural language model may integrate the task information of the target task and the task information of the multiple preset subtasks to obtain integrated prompt information corresponding to the target task, and may rewrite the integrated prompt information to obtain initial prompt information corresponding to the target task.

[0042] Specifically, the integrated prompt information can be understood as a collection of information obtained by merging the task information of the target task with the task information of multiple preset subtasks through a natural language model, integrating key content related to various tasks. The initial prompt information can be understood as the output obtained after rewriting the integrated prompt information, which is used to guide the subsequent further analysis, disassembly and execution of the target task, and provide a data basis for the entire task processing process.

[0043] Optionally, the natural language model may organize the task information of the target task and the task information of the preset subtasks in a format required by the model, and usually converts the information in various formats into a text string and combines them.

[0044] For example, when the target task is a conversation analysis task, the task information of the target task often includes the conversation content, which can be a conversation text or a voice (such as a call recording), etc. The natural language model can convert the audio data into text data first, and then combine the text data.

[0045] Furthermore, based on the integrated prompt information, the rewriting goals and rules can be determined, such as making the language more concise and clear, conforming to a specific language style (such as a formal business style, a popular science style that is easy to understand), highlighting key information, etc. Natural language processing tools can be called or natural language models can be used again for rewriting. For example, the synonym replacement and sentence conversion functions in the text editing tool are called to adjust some of the vocabulary and sentence structures in the integrated prompt information; or the natural language model rewrites the integrated prompt information by setting specific instructions (such as "rewrite this text into a more concise expression") and outputs the rewritten initial prompt information.

[0046] By applying this embodiment, by obtaining and integrating task information, it can be ensured that key information is not missed; the natural language model's information integration process is not a simple splicing, but a fusion based on semantic understanding, which can eliminate ambiguity and contradictions between information, making the initial prompt information semantics clearer and more accurate, and providing a reliable basis for subsequent task processing; the rewriting process can optimize the integrated information according to actual needs, so that the initial prompt information is more in line with specific task scenarios and execution requirements.

[0047] According to an optional embodiment of the present specification, obtaining task information of a target task and task information of multiple preset subtasks for the target task may include the following steps: receiving task information of the target task and task information of multiple preset subtasks for the target task input by a user front end.

[0048] Specifically, the user front end can be understood as an interface for users to interact with the system, which can be in the form of a web page, a mobile application, a desktop software, etc. Users input information and receive system feedback through the front-end interface to achieve communication with the system. Task information can be task description information in the form of text or voice, or it can be a task identifier. In the case where the task information is a task identifier, the task description information corresponding to the task identifier can be obtained from the information library according to the task identifier.

[0049] According to another optional embodiment of the present specification, obtaining task information of the target task and task information of multiple preset subtasks for the target task may also include the following steps: obtaining task information of the target task and task information of multiple preset subtasks for the target task through a preset interface.

[0050] Specifically, the preset interface may be a preconfigured interface for obtaining task information. In practical applications, multiple groups of target tasks and multiple preset subtasks corresponding to the target tasks may be preconfigured. The task information of the target task and the task information of multiple preset subtasks for the target task may be pre-stored in other systems, data platforms or information libraries, such as a data storage platform for intelligent customer service. The task information of the target task and the task information of multiple preset subtasks for the target task may be obtained by calling the preset interface.

[0051] According to an optional embodiment of the present specification, after integrating the task information of the target task and the task information of multiple preset subtasks to obtain the initial prompt information corresponding to the target task, it also includes: feeding back the initial prompt information corresponding to the target task to the user front end; receiving feedback information for the initial prompt information fed back by the user front end; and using a natural language model to rewrite the initial prompt information corresponding to the target task based on the feedback information to obtain adjusted initial prompt information corresponding to the target task.

[0052] Feedback information can be the modification opinions, supplementary content or confirmation information proposed by the user based on his / her own needs and judgment after receiving the initial prompt information. For example, if the user thinks that the time arrangement of a subtask in the initial prompt information is unreasonable, or a key requirement is not reflected, he / she can express these ideas through feedback information. The adjusted initial prompt information can be understood as the information obtained by rewriting the initial prompt information based on the user feedback information using the natural language model, which is more in line with the user's actual needs and expectations and can provide more accurate guidance for the subsequent task execution.

[0053] Exemplarily, the initial prompt information may be "In the after-sales communication scenario of a certain large e-commerce platform, customer service must follow the specifications and provide professional services to solve user shopping problems, and users hope to effectively solve their own problems. Please analyze the conversation to determine whether the customer service follows the specifications, effectively solves the problem, and whether there are any violations"; the initial prompt information is fed back to the user front end, and the front desk operator can input feedback information for the initial prompt information, "It is not clear what the specific violations include, which leads to some ambiguity in the judgment." Further, based on the feedback information, the initial prompt information can be rewritten as "In the after-sales communication scenario of a certain large e-commerce platform, customer service must follow the specifications and provide professional services to solve user shopping problems, and users hope to effectively solve their own problems. Violations include asking for user private information, insulting users, and not handling problems according to after-sales policies. Please analyze the conversation to determine whether the customer service follows the specifications, effectively solves the problem, and whether there are any violations." More accurate initial prompt information can be obtained through multiple interactions and feedback.

[0054] In actual applications, after the user completes the information input on the front-end interface, the data can be transmitted to the back-end server through the network protocol (such as HTTP, HTTPS). The front-end code will encapsulate the data entered by the user into a suitable format (such as JSON or HTML) and send it to the specified server interface. After the back-end server receives the data entered by the user, it can use the natural language model to assemble and rewrite the task information of the target task and the task information of multiple preset subtasks, and encapsulate the generated initial prompt information into a suitable format (such as JSON or HTML) and send it to the user front-end. After the user front-end receives the initial prompt information, it can use technologies such as HTML, CSS and JS to present the initial prompt information in an attractive and easy-to-read manner, such as displaying it in the form of paragraphs, lists, etc. on a web page. Furthermore, a feedback interface can be designed on the user front-end so that users can easily enter feedback information. For example, a text box and a submit button are added below the initial prompt information display page so that users can enter their opinions and suggestions on the initial prompt information. After the user enters information and submits it on the feedback interface, the front-end code can encapsulate the feedback information into a suitable format and send it to the back-end server through the network protocol. After receiving the feedback information, the backend server can use the natural language model to rewrite the initial prompt information corresponding to the target task based on the feedback information to obtain the adjusted initial prompt information corresponding to the target task. It can also be stored in the database for subsequent processing. At the same time, the feedback information can be preliminarily classified and marked, such as into modification suggestions, supplementary information, etc.

[0055] In an optional implementation of the present specification, the initial prompt information and the feedback information can be used as input and passed to the natural language model. The natural language model performs semantic analysis, content modification and generation on the initial prompt information according to the requirements of the feedback information, and outputs the adjusted initial prompt information.

[0056] By applying this embodiment, by receiving the information input by the user's front end and allowing the user to provide feedback and adjust the initial prompt information, the user's professional knowledge and actual needs can be fully utilized to ensure that the initial prompt information accurately reflects the actual situation of the target task; by processing the feedback information using a natural language model, the initial prompt information can be rewritten and optimized more intelligently. The model can understand the semantics and intention of the feedback information, and reasonably modify and supplement the initial prompt information to improve the accuracy and completeness of the information. It can also achieve personalized customization of task information. Different users may have different needs and concerns for tasks. In this way, the personalized needs of users can be met and user satisfaction can be improved.

[0057] In actual applications, after interaction and feedback with the front-end operator, the initial prompt information of the target task that is finally adjusted can be obtained. Furthermore, the execution process of the target task can be disassembled based on the initial prompt information using a natural language model to obtain multiple target subtasks and generate target prompt information corresponding to the target task, wherein the target prompt information includes prompt information corresponding to multiple target subtasks.

[0058] Optionally, the natural language model can perform in-depth semantic analysis on the initial prompt information. The input text is segmented, the sentence is split into individual words or phrases, and the part of speech, word form and other information of each word are determined. At the same time, the natural language model can combine the large amount of knowledge and language patterns learned from its own pre-training to understand the overall semantics and intention of the initial prompt information. Based on the understanding of semantics, the model will extract the key elements of the target task. Including the task subject, task action, task object, and possible constraints such as time, place, and conditions. The model will also identify the goal and expected result of the task. The extraction of these key elements provides an important basis for the subsequent task decomposition.

[0059] Furthermore, based on the extracted key elements, the model will deduce the entire task process to complete the target task according to the existing task mode and logical reasoning ability, and divide the entire task process into multiple sub-steps. In the process of deriving the steps, the model will consider the sequence and logical dependencies between the sub-steps to ensure the rationality and coherence of the sub-steps. For example, if the data input of a sub-step is the output result of another sub-step, the execution order of the other sub-step is before the sub-step. Each derived sub-step is then further refined into a specific target sub-task. Each target sub-task has relatively independent goals and operation contents, which together constitute a specific work unit to complete the target task.

[0060] For each generated target subtask, the model will conduct another in-depth analysis to understand the specific goals, objects and operations involved in the target subtask. The model will also consider the dependencies between subtasks and other subtasks, as well as the position and role of subtasks in the entire target task. Furthermore, based on the analysis results of the target subtask, the model will use natural language generation technology to generate prompt information corresponding to the target subtask. The prompt information will clarify the specific requirements of the subtask, the expected results, and the resources or conditions that may be required.

[0061] In actual applications, after decomposing the execution process of the target task into multiple target subtasks and generating prompt information corresponding to the multiple target subtasks, it is possible to determine the target task models corresponding to the multiple target subtasks from task models of different specifications based on the task types of the multiple target subtasks, and use the multiple target task models to execute the multiple target subtasks based on the prompt information corresponding to the multiple target subtasks to obtain the task execution result of the target task.

[0062] Optionally, the natural language model can determine the target task models corresponding to the multiple target subtasks from the task models of different specifications based on the task types of the multiple target subtasks. For example, for simple target subtasks, such as extracting key information and checking whether there is any violation information, a small model is selected for execution; for complex target subtasks, such as determining whether the customer service has effectively solved the user's problem, a large model is selected for execution. The natural language model can adopt a multi-agent scheduling scheme to assign each target subtask and the target task model of the corresponding specification of the target subtask to different agents, and determine the calling and combination order of multiple agents according to the dependency relationship between each target subtask, and obtain the task execution result of the target task after the scheduling process of multiple agents is completed. After the scheduling process of multiple agents is completed, the task execution results of each agent can be combined to obtain the overall task result; or, the task results of different agents can also be displayed separately. It can be determined according to the needs of the actual application.

[0063] According to an optional embodiment of the present specification, multiple target subtasks include a retrieval enhancement generation subtask; utilizing multiple target task models, based on prompt information corresponding to the multiple target subtasks, correspondingly executing multiple target subtasks to obtain a task execution result of the target task may include the following steps: utilizing the target task model corresponding to the retrieval enhancement generation subtask, based on prompt information corresponding to the retrieval enhancement generation subtask, retrieving target knowledge from an external knowledge base, and under the guidance of the prompt information, using the target knowledge as enhanced knowledge to generate a task execution result of the retrieval enhancement generation subtask, and determining the task execution result of the target task based on the task execution result of the retrieval enhancement generation subtask.

[0064] Specifically, the retrieval enhancement generation subtask can be understood as a target subtask that needs to be search enhanced. Exemplarily, for some proper nouns or special needs, it is necessary to call an external knowledge base to further understand their meaning. The prompt information corresponding to the retrieval enhancement generation subtask can be understood as text information used to guide the target task model of the retrieval enhancement generation subtask, which can include key content such as task requirements, restrictions, and expected output forms, which can help the model clarify how to retrieve information from the knowledge base and how to generate results based on this information. The external knowledge base can be understood as a pre-built database that stores a large amount of knowledge. It can contain information in various fields, such as encyclopedic knowledge, professional literature, industry reports, etc. There can be more than one external knowledge base. The external knowledge base is the source of information retrieval and provides factual basis for text generation. The target knowledge can be understood as the knowledge content related to the task retrieved from the external knowledge base according to the prompt information corresponding to the retrieval enhancement generation subtask. This knowledge may be specific facts, cases, data, etc., which are used for subsequent model reasoning and task processing. The task execution result of the retrieval enhancement generation subtask can be understood as the search enhancement information obtained by using the target knowledge as enhanced knowledge. The task execution result of the retrieval enhancement generation subtask can be used to enrich and enhance the prompt information of the retrieval enhancement generation subtask, thereby helping the target task model to more accurately execute the target task and obtain the execution result of the target task.

[0065] In an optional implementation of the present specification, the target task model corresponding to the retrieval enhancement generation subtask can be loaded. For example, a pre-trained model that integrates the retrieval module and the generation module is loaded into the memory. The target task model can be configured according to the task requirements, and the retrieval parameters can be set, such as determining the amount of information to be retrieved from the external knowledge base, the specific settings of the retrieval algorithm; setting the generation parameters, such as the maximum length of text generation, the temperature parameter of the generation (controlling the randomness of the generated text), etc. Further, the prompt information corresponding to the retrieval enhancement generation subtask is obtained from the task scheduling system or the previous task processing flow. The retrieval module of the target task model analyzes the prompt information, extracts key information, and uses this key information to search in the external knowledge base. For example, if the external knowledge base is a document-based database, the retrieval module will filter out documents related to key information from a large number of documents through technologies such as keyword matching and semantic retrieval. The retrieved documents are sorted and screened, and the most relevant documents are determined as target knowledge according to indicators such as relevance and timeliness.

[0066] Furthermore, the retrieved target knowledge is input into the generation module of the target task model together with the prompt information. The generation module uses a pre-trained language model architecture, such as the Transformer architecture, to process the input information; based on the guidance of the prompt information and combined with the target knowledge, it generates the task execution result corresponding to the retrieval enhancement generation subtask. If the retrieval enhancement generation subtask is a key component of the entire target task, then the result it generates will be directly used as part of the target task execution result; if there are other subtasks, this result will be integrated with the results of other subtasks to comprehensively determine the final execution result of the target task.

[0067] By applying this embodiment, the model can obtain the latest and most accurate factual basis by retrieving the target knowledge from the external knowledge base. Compared with relying solely on the pre-trained knowledge of the model itself, the results generated based on the real-time updated external knowledge base can reflect the latest knowledge rather than outdated pre-trained knowledge; the model is guided to focus on the retrieval and generation of task-related knowledge through prompt information, which reduces the possibility of generating irrelevant or erroneous information and improves the accuracy and pertinence of the results; the rich information sources of the external knowledge base make the generated results richer and more comprehensive, and make the results more in-depth and broad. For tasks in different fields and different types, as long as the external knowledge base covers the knowledge in the relevant fields, the retrieval enhancement generation subtask can flexibly adapt to the task requirements by adjusting the prompt information, without integrating all knowledge into the pre-training of the model, reducing the amount of data and computing resources required for model training. By relying on an external knowledge base, the model can obtain the latest knowledge in real time through retrieval on the basis of a smaller scale of pre-training, thereby reducing training costs and time while ensuring performance.

[0068] According to an optional embodiment of the present specification, after obtaining the task execution result of the target task, the following steps may also be included: integrating the target task execution log of at least one target task, wherein any target task execution log is obtained by integrating multiple target subtasks, the prompt information corresponding to the multiple target subtasks, and the subtask execution results in steps according to the execution order between the multiple target subtasks based on the target prompt information and the task execution result of the target task.

[0069] Specifically, the target task execution log can be understood as a document or data set used to record information related to the target task execution process, including relevant information from the start to the end of the task. The target task execution log may include multiple target subtasks obtained after the execution process of the target task is disassembled, as well as prompt information and task execution results corresponding to the multiple target subtasks. Log sources may include database tables, file storage systems (such as server local folders, cloud storage), specific log management tools, etc. Data collection standards can be used to specify standards such as the time dimension and task type dimension of the collected log data. In practical applications, when the execution process of the target task is disassembled into multiple target subtasks based on the initial prompt information, each target subtask obtained by the disassembly can be recorded; further, the natural language model will determine the target task model of the corresponding specifications of the multiple target subtasks, and based on the generated target prompt information, a multi-agent scheduling scheme is used to execute the multiple target subtasks.

[0070] Optionally, integrating the target task execution log of at least one target task may include: determining at least one log source and data collection standard; collecting and integrating the log data of at least one target task from at least one log source based on the data collection standard to obtain the target task execution log of at least one target task. The target task execution log may include information corresponding to the task execution process of the target task, and may also include other information such as the initial prompt information of the target task, target prompt information, and task execution results. The target task execution log can be understood as a log obtained based on the integration of the real online task execution process, and the data in the log is obtained based on the execution process record and integration of the target task.

[0071] In the process of integrating the target task execution log, based on the execution order of multiple target subtasks in the multi-agent scheduling scheme and the prompt information corresponding to each target subtask, multiple target subtasks can be recorded in the form of "step1-prompt1; step2-prompt2; ...; stepN-promptN" (N is a positive integer greater than or equal to 1). Furthermore, in the multi-agent scheduling process, the subtask execution results of each target subtask can also be associated with the target subtask and the prompt information, and the information corresponding to the task execution process of the target task obtained by the final integration can be "step1-prompt1-result1; step2-prompt2-result2; ...; stepN-promptN-resultN".

[0072] By applying this embodiment, by integrating the target task execution logs of at least one target task, a data basis can be provided for the subsequent analysis of the logs generated by online calls, which is helpful for the subsequent analysis of indicators such as recognition rate and accuracy of existing solutions, and can optimize the natural language model based on abnormal cases generated by real logs.

[0073] According to an optional embodiment of the present specification, after integrating the task execution log of at least one target task, the following steps may also be included: based on the task execution log of at least one target task, analyzing the execution process of at least one target task, and determining the execution process indicator of at least one target task; adjusting the target prompt information of at least one target task according to the execution process indicator of at least one target task; and constructing an adjusted task execution log of at least one target task based on the adjusted target prompt information corresponding to the target task.

[0074] Optionally, analyzing the execution process of at least one target task may include: analyzing multiple target subtasks obtained by disassembling the target task during the execution process of the target task. Further, it may include analyzing the prompt information and task execution results corresponding to each target subtask, and calculating indicators such as recognition rate and accuracy rate.

[0075] The execution process indicators may specifically include task execution indicators of multiple target subtasks corresponding to the target task, such as recognition rate, accuracy rate, etc. The feedback information may include abnormal cases (Bad cases) in multiple target subtasks and corresponding reasons.

[0076] Optionally, adjusting target prompt information of at least one target task according to an execution process indicator of at least one target task may include: adjusting target prompt information of at least one target task when the execution process indicator does not meet a preset indicator requirement.

[0077] Specifically, the preset indicator requirements can be understood as specific numerical requirements for indicators such as recognition rate and accuracy. In practical applications, the natural language model can independently determine whether the execution process indicators meet the preset indicator requirements. If not, the natural language model can trigger automatic adjustment and optimization of the target prompt information. Furthermore, the natural language model can continue to execute the online target task based on the optimized prompt, thereby generating a new task execution log.

[0078] By applying this embodiment, the execution process of at least one target task is analyzed based on the task execution log of at least one target task, and problems existing in the task execution process can be accurately located based on the indicators obtained by the analysis and in-depth study of the log, such as a subtask execution time being too long or a high error rate. By continuously integrating task execution logs, analyzing execution process indicators, and adjusting target prompt information, a closed loop of continuous improvement can be formed, which is conducive to improving the efficiency and quality of task execution.

[0079] According to an optional embodiment of the present specification, adjusting the target prompt information of at least one target task according to the execution process indicator of at least one target task may include the following steps: sending the task execution log and execution process indicator of at least one target task to the developer front end; receiving feedback information of multiple target subtasks of at least one target task fed back by the developer front end; and rewriting the target prompt information corresponding to the target task based on the feedback information by using a natural language model to obtain adjusted target prompt information corresponding to the target task.

[0080] In actual applications, the task execution log and execution process indicators of at least one target task can be sent to the developer front end, and the developer front end can manually screen out abnormal cases and give corresponding reasons. Furthermore, the natural language model can be used to rewrite the target prompt information corresponding to the target task based on the feedback information to obtain the adjusted target prompt information corresponding to the target task. Optionally, the natural language model can automatically rewrite and optimize the prompt information corresponding to the abnormal target subtask based on the abnormal case and the corresponding reason, so as to realize the automatic optimization of the prompt.

[0081] By applying this embodiment, the developer front end can screen out abnormal cases based on the execution process indicators and give corresponding reasons and suggestions; by rewriting the target prompt information through the natural language model, adjustments can be made to the problems, thereby optimizing the entire task execution process.

[0082] According to an optional embodiment of the present specification, after constructing an adjusted task execution log of at least one target task based on the adjusted target prompt information corresponding to the target task, the following steps may also be included: sending the adjusted task execution log of at least one target task to the developer front end; receiving the task execution log of at least one target task annotated as fed back by the developer front end, wherein the task execution log includes the annotated target prompt information corresponding to the target task; constructing a test set based on the task execution log of at least one target task annotated, wherein the test set is used to test different versions of target prompt information corresponding to at least one target task.

[0083] Specifically, the adjusted task execution log of at least one target task can be understood as a task execution log reconstructed based on the adjusted target prompt information, which records the relevant information of the whole task execution process obtained by executing at least one target task online after prompt optimization, and can mainly include multiple target subtasks obtained by decomposing each target task, prompt information corresponding to multiple target subtasks, and task execution results. The task execution log of at least one target task after annotation can be understood as the task execution log obtained by the developer marking and annotating the key information in the log according to his own professional knowledge and experience after receiving the task execution log. The test set can be understood as a data set constructed by extracting relevant data from the annotated task execution log, which can be used to test different versions of target prompt information corresponding to the target task to evaluate its performance and effect. Different versions of target prompt information can be understood as target prompt information versions with different expressions, emphases or optimization directions generated for the target task. Through the test set, combined with the automatic evaluation system, etc., the current version with better effect can be screened out.

[0084] Optionally, marking the target prompt information may include marking the prompt information of multiple target subtasks obtained by decomposing the target task, and the marking content may specifically include evaluation of the prompt information, improvement suggestions, applicable scenario descriptions, etc.

[0085] By applying this embodiment, by sending the adjusted task execution log of at least one target task to the developer front end, it can help discover problems in the target prompt information, such as unclear expression, inaccurate guidance, etc., based on the developer's annotation of the task execution log, and make targeted improvements in subsequent tests, thereby improving the quality and effectiveness of the target prompt information. By receiving the task execution log of at least one target task that has been annotated and fed back by the developer front end, a test set is constructed based on the task execution log of at least one target task that has been annotated, and the test set can be used to test different versions of the target prompt information, so that the advantages and disadvantages of each version can be intuitively compared, and by analyzing the test results, the target prompt information version with current performance better than other versions can be screened out, and the task execution process can be further optimized to improve the efficiency and accuracy of task execution. In the future, by continuously collecting real online logs, marking and updating the test set, the target prompt information can be continuously optimized and iterated, and the performance of the model in actual tasks can be continuously improved.

[0086] According to an optional embodiment of the present specification, after constructing a test set based on the task execution log of at least one target task after annotation, the following steps may also be included: returning to execute and integrate the target task execution log of at least one target task until determining the task execution log of the target task whose execution process indicator exceeds a preset threshold, and obtaining the standard version of the target prompt information corresponding to the target task.

[0087] Specifically, the standard version of the target prompt information can be understood as the target prompt information version that is currently most effective in guiding the execution of the target task and making the task execution result reach a higher quality (such as the accuracy meets the preset requirements) after a series of evaluations and screenings from multiple different versions of the target prompt information. It can also be understood as the target prompt information with the highest accuracy that can be determined at present. Different versions of prompt information can be understood as target prompt information with different content, structure or focus generated by various means (such as developer feedback, model optimization, etc.) for the target task. The preset threshold can be understood as a pre-set standard value used to determine whether the execution process indicator of the target task reaches an acceptable or expected level. When the execution process indicator exceeds the threshold, it is considered that the task execution effect has reached a certain requirement. Accuracy can be understood as an indicator used to measure the correctness of the task execution result in an automatic evaluation system.

[0088] Optionally, the preset threshold may be set according to the current highest accuracy rate, and each time the accuracy rate is improved, the improved accuracy rate may be determined as the preset threshold value.

[0089] In an optional implementation of the present specification, multiple different versions of target prompt information generated for the target task can be collected and organized. These versions may be generated based on different optimization strategies, developer feedback, or model adjustments; ensure that a test set based on the annotated target task execution log has been constructed, and the test set contains relevant data for evaluating the task execution results, such as input samples, expected outputs, etc. An automatic evaluation system can be used for evaluation, and each version of the target prompt information is applied to the test set in turn. For each input sample in the test set, a natural language model is used to generate the task execution result based on the corresponding version of the target prompt information.

[0090] The automatic evaluation system can compare the generated task execution results with the expected output in the test set according to the preset evaluation criteria, and calculate the accuracy of each version of the prompt information on the test set. For example, in a target task of text sentiment analysis, the automatic evaluation system will determine whether the results of the model's text sentiment classification based on different versions of prompt information are consistent with the actual sentiment tendency, and calculate the correct classification ratio as the accuracy. Furthermore, the accuracy of each version of the prompt information is recorded, and the change of the accuracy with the prompt information version is analyzed. The accuracy curve can be drawn to intuitively show the accuracy change trend of different versions of the prompt information. By observing the change of the accuracy, find the law of rising or falling accuracy. For example, the accuracy of some versions of the prompt information may be significantly improved after adding specific guidance content; while the accuracy of other versions may decrease due to overly complex expressions. Compare the accuracy of each version of the prompt information with the preset threshold; if the accuracy of a version of the prompt information exceeds the preset threshold, and no better accuracy appears in the subsequent version (or the improvement is not obvious), it is considered that the prompt information of this version may be a candidate for the standard version; if the accuracy of all versions of the prompt information does not exceed the preset threshold, it is necessary to return to the step of integrating the target task execution log. This includes collecting more task execution logs, obtaining more developer feedback, further optimizing the model, and then regenerating different versions of the target prompt information and evaluating them again. When a version with an accuracy exceeding the preset threshold is found, the version can be further verified and evaluated. An additional verification data set (different from the test set) can be used to verify the stability and reliability of its accuracy again. If the accuracy of the version remains at a high level and stable during the verification process, it can be determined that the target prompt information of this version is the target prompt information of the standard version corresponding to the target task. The target prompt information of the standard version can be replaced online. Specifically, the automatic evaluation system can be understood as a system that can automatically evaluate the execution results of the target task. According to the preset evaluation standards and algorithms, it calculates the relevant indicators of task execution, such as accuracy and recall, to measure the effects of different versions of prompt information. Furthermore, through the multi-agent scheduling scheme, the 7B, 72B and other large and small models can be combined and called to view the changes in the overall accuracy. Subsequently, the overall indicators can be adjusted by optimizing the large and small models separately, and finally a balance between cost and effect can be achieved.

[0091] By using this embodiment, through the test set combined with the automatic evaluation system, it is possible to continuously compare the accuracy of different versions of prompt information, and screen out the standard version based on the preset threshold, to ensure that the most effective prompt information is used when executing the target task, thereby improving the accuracy of task execution, reducing errors and deviations, and improving the overall quality of task execution. The evaluation process of the automatic evaluation system for different versions of prompt information also provides feedback to the natural language model. The natural language model can understand the impact of different prompt information on the task execution results based on the changes in accuracy, and then perform targeted optimization.

[0092] See also Figure 3 , Figure 3 A flow chart of a model training method provided by an embodiment of the present specification is shown, which specifically includes the following steps.

[0093] Step 302: Based on the test set, train target task models corresponding to multiple target subtasks, wherein the test set is constructed based on the task execution log of at least one target task after annotation, and the task execution log of at least one target task is obtained by executing at least one target task based on a natural language model.

[0094] Specifically, the target task model can be understood as a large model or a small model that executes the target subtask. In practical applications, since the execution process of the target task can be broken down into multiple target subtasks, multiple target subtasks can be executed through a multi-agent scheduling scheme. Therefore, multiple target subtasks can be executed in combination with large and small model calls, so that large models or small models corresponding to multiple target subtasks can be trained based on the test set.

[0095] Optionally, based on the prompt information corresponding to the target subtask in the test set and the task execution result label annotated for the target subtask, the target task model of the corresponding specification of the target subtask can be trained separately, so as to optimize the target task model in a targeted manner.

[0096] By applying this embodiment, by training the target task models corresponding to multiple target subtasks based on the test set, it is possible to fine-tune the target task models for executing different target subtasks individually, thereby improving the overall execution effect of the target task. Moreover, through the task execution method of calling a combination of large and small models, a balance between cost and effect can be achieved, and the model calling and training costs can be reduced on the basis of improving the task execution efficiency and accuracy.

[0097] See also Figure 4 , Figure 4 A flowchart of a method for processing a conversation analysis task provided by an embodiment of the present specification is shown, which specifically includes the following steps.

[0098] Step 402: Receive task information of a conversation analysis task input by a user front end, as well as task information of a plurality of preset subtasks of the conversation analysis task.

[0099] Step 404: Using the natural language model, integrate the task information of the conversation analysis task and the task information of multiple preset subtasks to obtain integrated prompt information corresponding to the conversation analysis task, rewrite the integrated prompt information, and obtain initial prompt information corresponding to the conversation analysis task.

[0100] Step 406: Using the natural language model, based on the initial prompt information, the execution process of the conversation analysis task is disassembled to obtain multiple target subtasks, and target prompt information corresponding to the conversation analysis task is generated, wherein the target prompt information includes prompt information corresponding to multiple target subtasks.

[0101] Step 408: Determine a dialog analysis task model of specifications corresponding to the multiple target subtasks, and use the multiple dialog analysis task models to execute the multiple target subtasks based on the prompt information corresponding to the multiple target subtasks to obtain the task execution result of the dialog analysis task.

[0102] Step 410: Feedback the task execution result to the user front end.

[0103] Specifically, the conversation analysis task can be understood as a task of analyzing various conversation data (such as conversations between customer service and customers, conversations on social platforms, etc.), and the analysis content may include but is not limited to key information extraction, sentiment analysis, intent recognition, etc. The conversation analysis task model can be understood as a model selected from task models of different specifications for different target subtasks to perform corresponding subtasks, such as a named entity recognition model for key information extraction, a text classification model for sentiment analysis, etc. It should be noted that the specific implementation of steps 402 to 410 can refer to the specific implementation of steps 202 to 206 above. This manual will not go into details here.

[0104] By applying this embodiment, by integrating the task information and preset subtask information of the conversation analysis task, it is possible to ensure that key information is not missed during the task processing process, providing a more comprehensive basis for subsequent analysis and execution; by selecting a suitable conversation analysis task model according to the task type of the target subtask, different conversation analysis models have different model specifications, required computing resources and calling costs, so that each subtask can be processed by the most suitable model, further improving the accuracy and efficiency of task execution and balancing the model calling cost.

[0105] According to an optional embodiment of the present specification, the task information of the conversation analysis task includes the conversation content information, and the task execution result includes at least one of the key information of the conversation content information, the solution corresponding to the conversation problem included in the conversation content information, at least one set of question-answer pairs included in the conversation content information, the service quality inspection result for the conversation content information, and the event type corresponding to the conversation content information. Specifically, the conversation content information may include the conversation text, or may include the conversation voice, etc.

[0106] By applying this embodiment, by executing the conversation analysis task, the task execution results of each preset subtask contained in the conversation analysis task can be obtained. For example, the preset subtasks may include key information extraction, solution extraction, question and answer pair extraction, service quality inspection and label classification, thereby better meeting the user's conversation analysis needs for the conversation content.

[0107] The following combination Figure 5 , taking the application of the processing method of the dialogue analysis task provided in this specification in the intelligent dialogue platform as an example, the processing method of the dialogue analysis task is further explained. Among them, Figure 5 A processing flow diagram of a method for processing a conversation analysis task provided in an embodiment of the present specification is shown, specifically including Step 1-Step 5.

[0108] Step 1: Define the conversation analysis task. Specifically, you can define upstream tasks and conversation contexts. Upstream tasks can be understood as various types of preset subtasks included in conversation analysis tasks. Each upstream task can correspond to a preset prompt, and users can also define upstream tasks and corresponding prompts according to their own needs. Upstream tasks can include conversation quality inspection, information mining, causal analysis, skill training, etc. The conversation context can include scene elements, character settings, target strategies, emotional tendencies, etc. In actual applications, by setting upstream tasks, uploading conversation content, and defining conversation contexts, you can complete the definition of conversation analysis tasks, and provide a data basis for the input of large models and the subsequent execution of conversation analysis tasks.

[0109] Step 2: Offline task optimization and disassembly. Specifically, the large model automatically assembles and rewrites the prompt according to the task definition completed in Step 1, and after interaction and feedback from the front-end operator, it is updated through multiple iterations to form the final prompt. For complex prompts (such as a task that includes dozens of task definitions and processes), the large model will decompose the entire task process into multiple sub-step calls and combinations based on the complex prompt, and form a multi-agent scheduling plan based on the dependencies between the sub-steps, and generate a corresponding prompt for each sub-step. The sub-steps 1-n obtained by the disassembly will be recorded in the real online log to facilitate subsequent verification of the sub-steps and the corresponding prompts of the sub-steps.

[0110] Step 3: Online enhancement module. Specifically, by using the RAG online enhancement module, search enhancement is performed for some requirements that require search (such as the understanding of some specific nouns), so as to complete the task with higher quality. Optionally, retrieval enhancement can be performed through external knowledge bases such as web search, enterprise knowledge base, positive and negative case base, etc.

[0111] Step 4: Log analysis. Specifically, you can collect real online logs to obtain various types of information in the entire process of executing the conversation analysis task. This includes the sub-steps obtained by disassembling the task process, the prompts corresponding to each sub-step, and the task execution results corresponding to each sub-step. The specific information can be determined according to the needs of the actual application. Furthermore, you can feed back the logs to the front-end developers, who will screen the bad cases and give corresponding reasons, and input the bad cases and corresponding reasons into the big model so that the big model can automatically optimize the prompts corresponding to the sub-steps with problems.

[0112] Step 5: Test set construction. Specifically, the real online logs can be marked and accumulated to form a test set. Specifically, the execution results of each sub-step can be marked, and the prompts corresponding to each sub-step can also be marked, etc. The specific requirements can be determined according to the actual application needs.

[0113] Step 6: Solution evolution. Specifically, you can apply the test set built in Step 5, and combine it with the automatic evaluation system to let the model automatically optimize the prompt based on the error data. Furthermore, you can view the changes in accuracy based on multiple prompt optimization versions until you select a prompt version with the highest accuracy. If the overall accuracy is the highest, you can replace the online version of the prompt with this version of the prompt to achieve prompt optimization. At the same time, you can try to call a combination of large and small models such as 7B and 72B to view changes in overall accuracy. Subsequently, you can adjust the overall indicators by optimizing the large and small models separately, and ultimately achieve a balance between cost and effect.

[0114] Corresponding to the above method embodiment, this specification also provides an intelligent dialogue platform embodiment, see Figure 6 , Figure 6 FIG. 1 shows a schematic diagram of the structure of an intelligent dialogue platform provided by an embodiment of this specification. Figure 6 As shown, the intelligent dialogue platform 600 includes a user front-end interface 602 and a first response unit 604 .

[0115] User front-end interface 602: used to receive task information of a conversation analysis task input by a user front-end, and task information of a plurality of preset subtasks for the conversation analysis task.

[0116] The first response unit 604 is used to use a natural language model to integrate the task information of the conversation analysis task and the task information of multiple preset subtasks, obtain the integrated prompt information corresponding to the conversation analysis task, rewrite the integrated prompt information, and obtain the initial prompt information corresponding to the conversation analysis task; use the natural language model to disassemble the execution process of the conversation analysis task based on the initial prompt information to obtain multiple target subtasks, and generate target prompt information corresponding to the conversation analysis task, wherein the target prompt information includes prompt information corresponding to the multiple target subtasks; based on the task types of the multiple target subtasks, determine the conversation analysis task models corresponding to the multiple target subtasks from task models of different specifications, and use the multiple conversation analysis task models to execute the multiple target subtasks based on the prompt information corresponding to the multiple target subtasks to obtain the task execution result of the conversation analysis task; and feed back the task execution result to the user front end.

[0117] In the embodiments of this specification, the intelligent dialogue platform performs dialogue analysis tasks according to user needs, realizes personalized model services, provides users with an efficient, flexible and easy-to-use model service platform, and improves user experience.

[0118] The above is a schematic scheme of an intelligent dialogue platform of this embodiment. It should be noted that the technical scheme of the intelligent dialogue platform and the technical scheme of the processing method of the above-mentioned dialogue analysis task belong to the same concept, and the details not described in detail in the technical scheme of the intelligent dialogue platform can be referred to the description of the technical scheme of the processing method of the above-mentioned dialogue analysis task.

[0119] The following combination Figure 7 , further explaining the intelligent dialogue platform. Figure 7 A schematic diagram of the front-end interface of an intelligent dialogue platform provided by an embodiment of the present specification is shown. The front-end interface of the intelligent dialogue platform may include a task definition panel on the left and a result display panel on the right. In the task definition panel on the left, you can select the preset subtasks to be executed by checking each instruction under the dialogue analysis dimension. You can also customize the instructions according to your own needs. You can upload dialogue information in plain text format or in audio format. After clicking the test button, the large model will execute the dialogue analysis task based on the task information customized by the user through the front-end interface, and display the task execution results in the result display panel on the right when the execution is completed. The task execution results can be displayed by category according to the preset subtasks checked by the user.

[0120] Corresponding to the above method embodiment, this specification also provides a model training platform embodiment, see Figure 8 , Figure 8 FIG. 1 shows a schematic diagram of a model training platform provided by an embodiment of the present specification. Figure 8 As shown, the model training platform 800 includes a training interface 802 and a second response unit 804 .

[0121] Training interface 802: used to receive a first training request sent by a first terminal, wherein the first training request carries a test set, the test set is constructed based on the task execution log of at least one target task after annotation, and the task execution log of at least one target task is obtained by executing at least one target task based on a natural language model.

[0122] The second response unit 804 is used to train target task models corresponding to multiple target subtasks based on the test set, and the target task model is used to execute the target subtask of the task processing method.

[0123] The model training platform 800 is an integrated virtual platform with a model training function, which is used to train a target task model adapted to a target subtask to solve different types of target subtasks. The model training platform 800 includes multiple interfaces and units, which can receive corresponding training requests and complete corresponding model training.

[0124] The training interface 802 is a component of the model training platform 800, and is used to receive a first training request from a first terminal (such as a developer or administrator). The training interface 802 is responsible for receiving and parsing the training request, and passing the sample data to the second response unit 804 for processing. In this way, the model training platform 800 can train a high-performance target task model based on a large amount of sample data.

[0125] The second response unit 804 is the core processing unit of the model training platform 800, responsible for executing model training and target subtasks. It trains or fine-tunes target task models of different specifications based on the received training requests and sample data.

[0126] By applying this embodiment, the combined call of large and small models can be realized through the multi-agent scheduling solution, and the large and small models can be optimized separately according to the change of the overall accuracy to realize the adjustment of the overall index, and finally achieve a balance between cost and effect.

[0127] The above is a schematic scheme of a model training platform of this embodiment. It should be noted that the technical scheme of the model training platform and the technical scheme of the above-mentioned model training method belong to the same concept, and the details not described in detail in the technical scheme of the model training platform can be referred to the description of the technical scheme of the above-mentioned model training method.

[0128] Corresponding to the above method embodiment, this specification also provides a task processing device embodiment, Fig. 9 FIG. 1 shows a schematic diagram of a task processing device provided by an embodiment of the present specification. Fig. 9 As shown, the device comprises: The acquisition module 902 is configured to acquire initial prompt information corresponding to the target task.

[0129] Generation module 904: using a natural language model, based on the initial prompt information, disassembles the execution process of the target task, obtains multiple target subtasks, and generates target prompt information corresponding to the target task, wherein the target prompt information includes prompt information corresponding to multiple target subtasks.

[0130] Execution module 906: Determine target task models corresponding to the specifications of the multiple target subtasks, and use the multiple target task models to execute the multiple target subtasks accordingly based on the prompt information corresponding to the multiple target subtasks to obtain the task execution result of the target task.

[0131] Optionally, multiple target subtasks include a retrieval enhancement generation subtask; the execution module 906 is further configured to: utilize the target task model corresponding to the retrieval enhancement generation subtask, based on the prompt information corresponding to the retrieval enhancement generation subtask, retrieve target knowledge from an external knowledge base, and under the guidance of the prompt information, use the target knowledge as enhanced knowledge to generate a task execution result of the retrieval enhancement generation subtask, and determine the task execution result of the target task based on the task execution result of the retrieval enhancement generation subtask.

[0132] By applying this embodiment, by selecting a target task model with appropriate specifications, it is avoided to use a model that is too complex or consumes too much resources to process simple subtasks. The above is a schematic scheme of a task processing device of this embodiment. It should be noted that the technical scheme of the task processing device and the technical scheme of the above-mentioned task processing method belong to the same concept. For the details not described in detail in the technical scheme of the task processing device, please refer to the description of the technical scheme of the above-mentioned task processing method.

[0133] Corresponding to the above method embodiment, this specification also provides a model training device embodiment, Fig.10 FIG. 1 shows a schematic diagram of a model training device provided by an embodiment of the present specification. Fig.10 As shown, the device comprises: Training module 1002: is configured to train target task models corresponding to multiple target subtasks based on a test set, wherein the test set is constructed based on a task execution log of at least one target task after annotation, and the task execution log of at least one target task is obtained by executing at least one target task based on a natural language model.

[0134] By applying this embodiment, the target task model can be fine-tuned individually through the training module, thereby improving the overall execution effect of the target task. The above is a schematic scheme of a model training device of this embodiment. It should be noted that the technical scheme of the model training device and the technical scheme of the above-mentioned model training method belong to the same concept. For details not described in detail in the technical scheme of the model training device, please refer to the description of the technical scheme of the above-mentioned model training method.

[0135] Corresponding to the above method embodiment, this specification also provides a processing device embodiment for a conversation analysis task. Fig.11 FIG. 1 is a schematic diagram showing a structure of a device for processing a conversation analysis task provided by an embodiment of the present specification. Fig.11 As shown, the device comprises: Receiving module 1102: configured to receive task information of a conversation analysis task input by a user front end, and task information of a plurality of preset subtasks of the conversation analysis task.

[0136] Rewriting module 1104: is configured to utilize a natural language model to integrate the task information of the conversation analysis task and the task information of multiple preset subtasks, obtain integrated prompt information corresponding to the conversation analysis task, rewrite the integrated prompt information, and obtain initial prompt information corresponding to the conversation analysis task.

[0137] Disassembly module 1106: is configured to utilize a natural language model to disassemble the execution process of the conversation analysis task based on the initial prompt information, obtain multiple target subtasks, and generate target prompt information corresponding to the conversation analysis task, wherein the target prompt information includes prompt information corresponding to multiple target subtasks.

[0138] Determination module 1108: is configured to determine a dialogue analysis task model of corresponding specifications of multiple target subtasks, and utilize multiple dialogue analysis task models to obtain the task execution result of the dialogue analysis task based on the prompt information corresponding to the multiple target subtasks and correspondingly execute the multiple target subtasks.

[0139] Feedback module 1110: configured to feed back the task execution result to the user front end.

[0140] By applying this embodiment, by determining the module, each subtask can be processed by the most suitable model, thereby improving the accuracy and efficiency of task execution and balancing the model call cost. The above is a schematic scheme of a processing device for a conversation analysis task in this embodiment. It should be noted that the technical scheme of the processing device for the conversation analysis task and the technical scheme of the processing method for the conversation analysis task mentioned above belong to the same concept. For details not described in detail in the technical scheme of the processing device for the conversation analysis task, please refer to the description of the technical scheme of the processing method for the conversation analysis task mentioned above.

[0141] Fig.12 The block diagram of a computing device according to an embodiment of the present specification is shown. The components of the computing device 1200 include but are not limited to a memory 1210 and a processor 1220. The processor 1220 is connected to the memory 1210 via a bus 1230, and the database 1250 is used to store data.

[0142] The computing device 1200 also includes an access device 1240 that enables the computing device 1200 to communicate via one or more networks 1260. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 1240 may include one or more of any type of network interface (e.g., a network interface card (NIC)) of wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a world-wide interoperability for microwave access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, and a near field communication (NFC).

[0143] In one embodiment of the present specification, the above components of the computing device 1200 and Fig.12 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Fig.12The computing device structure block diagram shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art may add or replace other components as needed. The computing device 1200 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.) or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC, Personal Computer). The computing device 1200 may also be a mobile or stationary server. Among them, the processor 1220 is used to execute the following computer executable instructions, which implement the steps of the above method when executed by the processor. The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above method belong to the same concept, and the details of the technical scheme of the computing device that are not described in detail can all be referred to the description of the technical scheme of the above method.

[0144] An embodiment of this specification also provides a computer-readable storage medium, which stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps of the above method are implemented. The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical solution of the storage medium and the technical solution of the above method belong to the same concept, and the details of the technical solution of the storage medium that are not described in detail can be referred to the description of the technical solution of the above method.

[0145] An embodiment of this specification also provides a computer program product, including a computer program / instruction, which implements the steps of the above method when executed by a processor. The above is a schematic scheme of a computer program product of this embodiment. It should be noted that the technical solution of the computer program product and the technical solution of the above method belong to the same concept, and the details of the technical solution of the computer program product that are not described in detail can be found in the description of the technical solution of the above method.

[0146] The above describes a specific embodiment of the present specification. Other embodiments are within the scope of the attached claims. In some cases, the actions or steps recorded in the claims can be performed in a different order from that in the embodiments and still achieve the desired results. In addition, the process depicted in the drawings does not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-tasking and parallel processing are also possible or may be advantageous. The computer instructions include computer program code, which may be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of patent practice. For example, in some regions, according to patent practice, the computer-readable medium does not include electric carrier signals and telecommunication signals. It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.

[0147] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments. The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and changes can be made according to the contents of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technical personnel in the relevant technical field can well understand and use this specification. This specification is only limited by the claims and their full scope and equivalents.

Claims

1. A task processing method, comprising: Get the initial prompt information corresponding to the target task; Using a natural language model, based on the initial prompt information, the execution process of the target task is disassembled to obtain a plurality of target subtasks, and target prompt information corresponding to the target task is generated, wherein the target prompt information includes prompt information corresponding to the plurality of target subtasks; Determine target task models of specifications corresponding to the multiple target subtasks, and use the multiple target task models to execute the multiple target subtasks accordingly based on prompt information corresponding to the multiple target subtasks to obtain task execution results of the target tasks.

2. According to the method of claim 1, the step of obtaining initial prompt information corresponding to the target task comprises: Acquire task information of a target task and task information of multiple preset subtasks for the target task; The natural language model is used to integrate the task information of the target task and the task information of the plurality of preset subtasks to obtain integrated prompt information corresponding to the target task, and the integrated prompt information is rewritten to obtain initial prompt information corresponding to the target task.

3. The method according to claim 2, after integrating the task information of the target task and the task information of the plurality of preset subtasks to obtain the initial prompt information corresponding to the target task, further comprises: Feedback the initial prompt information corresponding to the target task to the user front end; Receive feedback information from the user front end in response to the initial prompt information; The natural language model is used to rewrite the initial prompt information corresponding to the target task based on the feedback information to obtain adjusted initial prompt information corresponding to the target task.

4. The method according to any one of claims 1 to 3, after obtaining the task execution result of the target task, further comprising: Integrate the target task execution log of at least one target task, wherein any target task execution log is obtained by integrating the multiple target subtasks, the prompt information corresponding to the multiple target subtasks and the subtask execution results in steps according to the execution order between the multiple target subtasks based on the target prompt information of the target task and the task execution result.

5. The method according to claim 4, after integrating the task execution log of at least one target task, further comprising: Analyzing the execution process of the at least one target task based on the task execution log of the at least one target task, and determining an execution process indicator of the at least one target task; Adjusting target prompt information of the at least one target task according to an execution process indicator of the at least one target task; Based on the adjusted target prompt information corresponding to the target task, an adjusted task execution log of the at least one target task is constructed.

6. The method according to claim 5, wherein adjusting the target prompt information of the at least one target task according to the execution process indicator of the at least one target task comprises: Sending the task execution log and execution process indicators of the at least one target task to the developer front end; receiving feedback information of a plurality of target subtasks of the at least one target task fed back by the developer front end; The target prompt information corresponding to the target task is rewritten based on the feedback information by using the natural language model to obtain adjusted target prompt information corresponding to the target task.

7. The method according to claim 6, after constructing the adjusted task execution log of the at least one target task based on the adjusted target prompt information corresponding to the target task, further comprising: Sending the adjusted task execution log of the at least one target task to the developer front end; receiving a task execution log of the at least one target task that is annotated and fed back by the developer front end, wherein the task execution log includes target prompt information corresponding to the target task that is annotated; A test set is constructed based on the labeled task execution log of the at least one target task, wherein the test set is used to test different versions of target prompt information corresponding to the at least one target task.

8. The method according to claim 7, after constructing the test set based on the task execution log of the at least one target task after annotation, further comprises: Return the target task execution log for executing the integrated at least one target task until determining the task execution log of the target task whose execution process indicator exceeds the preset threshold, and obtain the target prompt information of the standard version corresponding to the target task.

9. The method according to any one of claims 1 to 8, wherein the plurality of target subtasks include a retrieval enhancement generation subtask; The using of multiple target task models, based on the prompt information corresponding to the multiple target subtasks, and correspondingly executing the multiple target subtasks to obtain the task execution result of the target task includes: Utilizing the target task model corresponding to the retrieval enhancement generation subtask, based on the prompt information corresponding to the retrieval enhancement generation subtask, the target knowledge is retrieved from an external knowledge base, and under the guidance of the prompt information, the target knowledge is used as enhanced knowledge to generate the task execution result of the retrieval enhancement generation subtask, and based on the task execution result of the retrieval enhancement generation subtask, the task execution result of the target task is determined.

10. A model training method, comprising: Based on the test set, the target task models corresponding to the multiple target subtasks are trained, wherein the test set is constructed according to the task processing method according to any one of claims 1-9.

11. A method for processing a dialogue analysis task, applied to an intelligent dialogue platform, comprising: Receive task information of a conversation analysis task input by a user front end, as well as task information of a plurality of preset subtasks for the conversation analysis task; Using a natural language model, integrating the task information of the conversation analysis task and the task information of the plurality of preset subtasks, obtaining integrated prompt information corresponding to the conversation analysis task, and rewriting the integrated prompt information to obtain initial prompt information corresponding to the conversation analysis task; Using the natural language model, based on the initial prompt information, the execution process of the dialogue analysis task is disassembled to obtain multiple target subtasks, and target prompt information corresponding to the dialogue analysis task is generated, wherein the target prompt information includes prompt information corresponding to the multiple target subtasks; Determine a dialog analysis task model of specifications corresponding to the multiple target subtasks, and use the multiple dialog analysis task models to execute the multiple target subtasks accordingly based on prompt information corresponding to the multiple target subtasks, so as to obtain a task execution result of the dialog analysis task; Feedback the task execution result to the user front end.

12. According to the method according to claim 11, the task information of the conversation analysis task includes conversation content information, and the task execution result includes key information of the conversation content information, solutions corresponding to the conversation problems included in the conversation content information, at least one set of question-answer pairs included in the conversation content information, service quality inspection results for the conversation content information, and at least one event type corresponding to the conversation content information.

13. An intelligent dialogue platform, comprising a user front-end interface and a first response unit; The user front-end interface is used to receive task information of a conversation analysis task input by a user front-end, and task information of a plurality of preset subtasks for the conversation analysis task; The first response unit is used to integrate the task information of the dialogue analysis task and the task information of the plurality of preset subtasks by using a natural language model, obtain integrated prompt information corresponding to the dialogue analysis task, rewrite the integrated prompt information, and obtain initial prompt information corresponding to the dialogue analysis task; Utilizing the natural language model, based on the initial prompt information, the execution process of the conversation analysis task is disassembled to obtain multiple target subtasks, and target prompt information corresponding to the conversation analysis task is generated, wherein the target prompt information includes prompt information corresponding to the multiple target subtasks; determining conversation analysis task models of corresponding specifications of the multiple target subtasks, and utilizing multiple conversation analysis task models, based on the prompt information corresponding to the multiple target subtasks, correspondingly executing the multiple target subtasks to obtain task execution results of the conversation analysis task; and feeding back the task execution results to the user front end.

14. A model training platform, comprising a training interface and a second response unit; The training interface is used to receive a first training request sent by a first terminal, wherein: The first training request carries a test set, and the test set is constructed and obtained by the task processing method according to any one of claims 1 to 9; The second response unit is used to train the target task models corresponding to the multiple target subtasks based on the test set, and the target task model is used to execute the target subtask of the task processing method according to any one of claims 1-9.

15. A computing device comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, the steps of the method according to any one of claims 1 to 12 are implemented.

16. A computer-readable storage medium storing a computer program / instruction, wherein the computer program / instruction, when executed by a processor, implements the steps of the method according to any one of claims 1 to 12.

17. A computer program product, comprising a computer program / instruction, which implements the steps of the method according to any one of claims 1 to 12 when executed by a processor.

Images