Training task information processing method and device and storage medium

Abstract

The invention provides a training task information processing method and device and a storage medium, and the method comprises the steps: determining target case information corresponding to a training request and prompt word configuration information for a preset large model in response to the training request; generating a target cue word of a preset large model according to the cue word configuration information and a preset cue word template; processing the target case information according to the target cue word by adopting the preset large model, and generating a target training task corresponding to the training request, the target training task comprising session information of a virtual user, the target training task is used for guiding a trained object to interact with the virtual user according to the session information. According to the method, the expression of the large model is adjusted by adjusting the cue word, different training case scenes are flexibly simulated, the lengthy model iteration constraint is eliminated, and the training task processing efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a training task information processing method, device and storage medium. Background Art

[0002] With the development of the Internet, users are increasingly inclined to purchase goods online through e-commerce platforms. Traditional e-commerce platforms generally provide users with a search function so that users can search for information about goods of interest through the e-commerce platform.

[0003] In e-commerce scenarios, e-commerce customer service is an important bridge for communication with users. Although the problem-solving ability of AI robot customer service is already very strong, the service of human customer service for high-quality and high-value users cannot be completely replaced. The huge number of human customer service still plays a very important role in the task of ensuring member experience. In order for customer service to better serve consumer users, customer service personnel need to be trained in relevant services before they take up their posts. Practical simulation training is to use robots to play the role of e-commerce member users, call in to ask questions and express their demands, so that the customer service staff can interact with the training robot for multiple rounds and find the correct service solution in the interaction, so as to quickly master the service process and service skills of the scenario.

[0004] However, traditional training robots usually perform training tasks based on set script cases, and the training tasks are not very robust for simulation of real scenarios and intelligent dialogue. Summary of the invention

[0005] The main purpose of the embodiments of the present application is to provide a training task information processing method, device and storage medium, which can adjust the performance of the large model by adjusting the prompt words, flexibly simulate different training case scenarios, get rid of the constraints of lengthy model iterations, and improve the efficiency of training task processing.

[0006] In a first aspect, an embodiment of the present application provides a method for processing training task information, comprising: in response to a training request, determining target case information corresponding to the training request and prompt word configuration information for a preset large model; generating target prompt words for the preset large model according to the prompt word configuration information and a preset prompt word template; using the preset large model to process the target case information according to the target prompt words, and generating a target training task corresponding to the training request, the target training task including session information of a virtual user, and the target training task being used to guide the trainee to interact with the virtual user according to the session information.

[0007] In one embodiment, the preset prompt word template includes at least one template slot, and the template slot is determined according to a preset evaluation index for the training operation.

[0008] In one embodiment, the prompt word configuration information includes scene questions and / or fact factors for characterizing training scene features; the preset prompt word template includes background slots; generating target prompt words of a preset large model based on the prompt word configuration information and the preset prompt word template includes: filling the background slots in the preset prompt word template according to the scene questions and / or the fact factors to generate the target prompt words of the preset large model.

[0009] In one embodiment, the prompt word configuration information includes the character personality characteristics of the virtual user; the preset prompt word template includes constraint condition slots; generating the target prompt words of the preset large model according to the prompt word configuration information and the preset prompt word template includes: filling the constraint condition slots in the preset prompt word template according to the character personality characteristics to generate the target prompt words of the preset large model; wherein the character personality characteristics include: one or more of the virtual character's emotional characteristics, comprehension characteristics, and consultation method characteristics.

[0010] In one embodiment, the prompt word configuration information includes the role skill characteristics of the virtual user; the preset prompt word template includes skill slots; generating the target prompt words of the preset large model based on the prompt word configuration information and the preset prompt word template includes: filling the skill slots in the preset prompt word template according to the role skill characteristics to generate the target prompt words of the preset large model.

[0011] In one embodiment, the preset big model is used to process the target case information according to the target prompt word to generate a target training task corresponding to the training request, including: generating a virtual order corresponding to the training request according to the target case information; obtaining the session context corresponding to the training request; and inputting the order information of the virtual order, the session context, the target case information and the target prompt word into the preset big model, so that the preset big model generates the target training task under the guidance of the target prompt word.

[0012] In one embodiment, generating a virtual order corresponding to the training request based on the target case information includes: obtaining a preset order generation tool bound to the target case information; using the order generation tool to process the target case information to generate the virtual order corresponding to the training request.

[0013] In one embodiment, after the target case information is processed according to the target prompt word using the preset large model to generate a target training task corresponding to the training request, it also includes: in response to reply information to the session information, obtaining a preset evaluation rule bound to the target case information for evaluating the training operation, the preset evaluation rule including at least one preset evaluation indicator for evaluating the training operation and a scoring weight corresponding to each of the preset evaluation indicators; determining an evaluation score for the reply information based on the at least one preset evaluation indicator, the scoring weight and the session information.

[0014] In one embodiment, before responding to a training request and determining the target case information corresponding to the training request and the prompt word configuration information for a preset large model, it also includes: responding to an AI material configuration request regarding the target case information, obtaining AI material information for the preset large model, the AI ​​material information including the prompt word configuration information; and binding the AI ​​material information with the target case information.

[0015] In one embodiment, before responding to a training request and determining the target case information corresponding to the training request and the prompt word configuration information for the preset large model, it also includes: responding to the case configuration request, obtaining at least one input preset case configuration information, the preset case configuration information including one or more of the basic configuration, solution configuration, and scoring configuration of the preset case; generating preset case information for training according to the preset case configuration information.

[0016] In a second aspect, an embodiment of the present application provides a method for processing training task information, comprising: in response to a training request regarding a customer service personnel, determining target case information corresponding to the training request and prompt word configuration information for a preset large model, wherein the target case information is case information regarding a commodity order; generating target prompt words for a preset large model based on the prompt word configuration information and a preset prompt word template; using the preset large model to process the target case information according to the target prompt words, and generating a target training task corresponding to the training request, wherein the target training task includes session information of a virtual user, and the target training task is used to guide the customer service personnel to interact with the virtual user based on the session information.

[0017] In a third aspect, an embodiment of the present application provides a task information processing method, comprising: in response to a user's task processing request, determining target case information corresponding to the task processing request and prompt word configuration information for a preset large model; generating target prompt words for the preset large model based on the prompt word configuration information and a preset prompt word template; using the preset large model to process the target case information according to the target prompt words, and generating a target task corresponding to the task processing request, the target task including session information of a virtual user, and the target task being used to guide a specified object to interact with the virtual user based on the session information.

[0018] In a fourth aspect, an embodiment of the present application provides a training task information processing device, including:

[0019] A determination module, configured to determine, in response to a training request, target case information corresponding to the training request and prompt word configuration information for a preset large model;

[0020] A first generating module, used for generating a preset large model of target prompt words according to the prompt word configuration information and a preset prompt word template;

[0021] The second generation module is used to use the preset large model to process the target case information according to the target prompt word, and generate a target training task corresponding to the training request. The target training task includes the session information of the virtual user, and the target training task is used to guide the trainee to interact with the virtual user according to the session information.

[0022] In one embodiment, the preset prompt word template includes at least one template slot, and the template slot is determined according to a preset evaluation index for the training operation.

[0023] In one embodiment, the prompt word configuration information includes scene questions and / or fact factors used to characterize training scene characteristics; the preset prompt word template includes background slots; and the first generation module is used to fill the background slots in the preset prompt word template according to the scene questions and / or the fact factors to generate target prompt words of a preset large model.

[0024] In one embodiment, the prompt word configuration information includes the character personality characteristics of the virtual user; the preset prompt word template includes constraint condition slots; the first generation module is used to fill the constraint condition slots in the preset prompt word template according to the character personality characteristics to generate target prompt words of the preset large model; wherein the character personality characteristics include: one or more of the virtual character's emotional characteristics, comprehension characteristics, and consultation method characteristics.

[0025] In one embodiment, the prompt word configuration information includes the role skill characteristics of the virtual user; the preset prompt word template includes skill slots; the first generation module is used to fill the skill slots in the preset prompt word template according to the role skill characteristics to generate the target prompt words of the preset large model.

[0026] In one embodiment, the second generation module is used to generate a virtual order corresponding to the training request based on the target case information; obtain the session context corresponding to the training request; and input the order information of the virtual order, the session context, the target case information and the target prompt word into the preset large model, so that the preset large model generates the target training task under the guidance of the target prompt word.

[0027] In one embodiment, the second generating module is used to obtain a preset order generating tool bound to the target case information; use the order generating tool to process the target case information to generate the virtual order corresponding to the training request.

[0028] In one embodiment, it also includes: an evaluation module, which is used to obtain the preset evaluation rules bound to the target case information for evaluating the training operation in response to the reply information to the session information, after the target case information is processed according to the target prompt word using the preset large model to generate a target training task corresponding to the training request, wherein the preset evaluation rules include at least one preset evaluation indicator for evaluating the training operation and a scoring weight corresponding to each of the preset evaluation indicators; and determine the evaluation score of the reply information based on the at least one preset evaluation indicator, the scoring weight and the session information.

[0029] In one embodiment, it also includes: a first configuration module, which is used to respond to an AI material configuration request about the target case information, obtain AI material information for the preset large model before determining the target case information corresponding to the training request and the prompt word configuration information for the preset large model in response to the training request, and the AI ​​material information includes the prompt word configuration information; and bind the AI ​​material information with the target case information.

[0030] In one embodiment, it also includes: a first configuration module, which is used to obtain at least one input preset case configuration information in response to a case configuration request before determining the target case information corresponding to the training request and the prompt word configuration information for the preset large model in response to the training request, wherein the preset case configuration information includes one or more of the basic configuration, solution configuration, and scoring configuration of the preset case; and generate preset case information for training according to the preset case configuration information.

[0031] In a fifth aspect, an embodiment of the present application provides an electronic device, including:

[0032] at least one processor; and

[0033] a memory communicatively coupled to the at least one processor;

[0034] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the electronic device to execute the method described in any one of the above aspects.

[0035] In a sixth aspect, an embodiment of the present application provides a cloud device, including:

[0036] at least one processor; and

[0037] a memory communicatively coupled to the at least one processor;

[0038] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the cloud device to execute the method described in any one of the above aspects.

[0039] In a seventh aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, the method described in any one of the above aspects is implemented.

[0040] In an eighth aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements the method described in any of the above aspects when executed by a processor.

[0041] The training task information processing method, device and storage medium provided in the embodiments of the present application respectively obtain the corresponding target case information and the large model prompt word configuration information for a specific training request, and generate the target prompt word corresponding to the training request according to the corresponding prompt word configuration information in combination with the pre-structured prompt word template, and utilize the contextual learning characteristics of the large model to constrain the large model through the target prompt word to simulate the training scenario and generate a matching target training task. The target training task guides the trainee to interact with the virtual user through the session information of the virtual user, thereby realizing the training process, and realizing the adjustment of the performance of the large model by adjusting the prompt word, flexibly simulating different training case scenarios, getting rid of the constraints of lengthy model iterations, and improving the efficiency of training task processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application. Obviously, the drawings described below are some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor.

[0043] Figure 1 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0044] Figure 2 A schematic diagram of an application scenario of a training task information processing system provided in an embodiment of the present application;

[0045] Figure 3A A schematic diagram of the overall architecture of a training task information processing system provided in an embodiment of the present application;

[0046] Figure 3B A schematic diagram of an implementation flow of a training task information processing system provided in an embodiment of the present application;

[0047] Figure 4 A flowchart of a training task information processing method provided in an embodiment of the present application;

[0048] Figure 5 A schematic diagram of a prompt template provided in an embodiment of the present application;

[0049] Fig. 6A A schematic diagram of a training task information processing system provided in an embodiment of the present application;

[0050] Figure 6B A flowchart of a training task information processing method provided in an embodiment of the present application;

[0051] Figure 6C A signaling interaction diagram of an AI case configuration process provided in an embodiment of the present application;

[0052] Fig.6D A schematic diagram of an AI case configuration interface provided in an embodiment of the present application;

[0053] Fig. 7A A schematic diagram of signaling interaction during a customer service training process provided for this application;

[0054] Figure 7B A schematic diagram of the interactive interface of the customer service end during the customer service training process provided for this application;

[0055] Fig. 8A A flowchart of a training task information processing method provided in an embodiment of the present application;

[0056] Figure 8B A flowchart of a task information processing method provided in an embodiment of the present application;

[0057] Fig. 9 A schematic diagram of the structure of a training task information processing device provided in an embodiment of the present application;

[0058] Fig.10 A schematic diagram of the structure of a cloud device provided in an embodiment of the present application.

[0059] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0060] Here, exemplary embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application.

[0061] The term "and / or" in this article is used to describe the association relationship of associated objects, specifically indicating that there may be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0062] 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 this application 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.

[0063] In order to clearly describe the technical solution of the embodiment of the present application, the terms involved in the present application are first defined:

[0064] AI: Artificial Intelligence.

[0065] LLM: Large Language Model, a natural language processing model based on deep learning, which can learn the grammar and semantics of natural language and generate human-readable text. Large language models can handle a variety of natural language tasks, such as text classification, question answering, dialogue, etc., and are an important path to artificial intelligence.

[0066] SFT: Supervised Fine Tuning, a supervised fine-tuning learning, a deep learning strategy, usually used on a pre-trained large language model, using labeled data to fine-tune the model to adapt it to a specific task or domain. API: Application Programming Interface, application programming interface.

[0067] LangChain: is a programming framework that helps use large language models (LLMs) in applications, that is, a framework for combining large language models with external computing and data sources and APIs.

[0068] Prompt: A prompt is a natural language input, similar to a command or instruction, that lets the AI ​​model know what it needs to do.

[0069] Prompt engineering: In the field of large language models, prompt engineering refers to the design and construction of prompts that guide large language models to generate specific types of text. By carefully designing prompts, the model can be guided to generate text output that meets specific needs or contexts. Prompt engineering involves operations such as selecting appropriate text fragments, adjusting text length, and adding specific keywords to influence the text content generated by the model.

[0070] IM: Instant Messaging.

[0071] Online service: generally refers to customer service interacting with members through text online IM to resolve members' demands.

[0072] Intelligent training robot: It provides a practical environment for real online services. Customer service staff can familiarize themselves with service operation procedures and service consultation issues by constantly simulating direct service dialogues with users. This will improve the service quality and efficiency of customer service staff, help new employees quickly master standard knowledge points, and shorten the growth cycle.

[0073] Training simulation environment: provides training practice plug-ins and workbenches equivalent to those used by Xiaoer to serve real members, namely, training workbenches.

[0074] Id: Identity document, ID number, account number, unique code.

[0075] ChatGLM: Chat Global Language Model, chat global model.

[0076] SOP: Standard Operating Procedure, standard operating procedure.

[0077] GPT: Generative Pre-Trained Transformer, a deep learning model for text generation based on the Internet and available data for training. GPT-3.5-turbo, GPT-4 8K, and GPT-4 32K are model versions of GPT.

[0078] QA: Question Answering.

[0079] UDF: User Defined Function, user defined function.

[0080] ASR: Automatic Speech Recognition, speech recognition.

[0081] TTS: Text to Speech, speech synthesis.

[0082] RAG: Retrieval-Augmented Generation, retrieval enhanced generation.

[0083] TPP: The Personalization Platform, personalized algorithm development platform.

[0084] The training task information processing method of the embodiment of the present application can be applied to any field that requires intelligent training.

[0085] Taking the customer service training scenario of an e-commerce platform as an example, e-commerce customer service is an important bridge for communication with users. Although the problem-solving ability of AI robot customer service is already very strong, the service of human customer service for high-quality and high-value users cannot be completely replaced. The huge number of human customer service still plays a very important role in the task of ensuring member experience. In order for customer service to better serve consumer users, customer service personnel need to be trained in relevant services before they take up their posts. Practical simulation training is to use robots to play the role of e-commerce member users, call in to ask questions and express their demands, so that the customer service staff can interact with the training robot for multiple rounds and find the correct service solution in the interaction, so as to quickly master the service process and service skills of the scenario.

[0086] However, traditional training robots usually perform training tasks based on set script cases, and the training tasks are not very robust for simulation of real scenarios and intelligent dialogue.

[0087] In actual scenarios, in order to solve the shortcomings of script case training, a simulation training method based on ChatGLM fine-tuning can be used. The ChatGLM model is fine-tuned using the SFT method, which initially solves the robot's flexible conversation ability. However, the robustness of simulation and intelligent conversation in real scenarios is still poor. This method has some advantages because it uses artificial corpus for fine-tuning, and the generated words are more suitable for the questions and answers of general consumers. This method may be due to overfitting or lack of background information. This method still has the following defects:

[0088] 1. The model after SFT is not robust enough and cannot handle abnormal situations well.

[0089] 2. The follow-up questions and answers are poor, and the generated chat is relatively short and repetitive.

[0090] 3. The scenarios covered by case operations rely on iterative training of models, but the low efficiency of model iteration leads to a long iteration cycle.

[0091] In order to solve at least one of the above problems, an embodiment of the present application provides a training task information processing solution. For a specific training request, the corresponding target case information and the large model prompt word configuration information are respectively obtained, combined with the pre-structured prompt word template, and the target prompt word corresponding to the training request is generated according to the corresponding prompt word configuration information. The context learning characteristics of the large model are utilized, and the target prompt word is used to constrain the large model to simulate the training scenario and generate a matching target training task. The target training task guides the trainee to interact with the virtual user through the virtual user's session information, thereby realizing the training process, and realizing the adjustment of the performance of the large model by adjusting the prompt word, flexibly simulating different training case scenarios, getting rid of the constraints of lengthy model iteration, and improving the efficiency of training task processing.

[0092] Some embodiments of the present application are described in detail below in conjunction with the accompanying drawings. In the case where there is no conflict between the embodiments, the following embodiments and the features in the embodiments can be combined with each other. In addition, the step sequence in the following method embodiments is only an example and not a strict limitation.

[0093] like Figure 1 As shown, this embodiment provides an electronic device 1, including: at least one processor 11 and a memory 12, Figure 1 A processor is taken as an example. The processor 11 and the memory 12 are connected via a bus 10. The memory 12 stores instructions that can be executed by the processor 11, and the instructions are executed by the processor 11 so that the electronic device 1 can execute all or part of the process of the method in the following embodiment, so as to adjust the performance of the large model by adjusting the prompt words, flexibly simulate different training case scenarios, get rid of the constraints of lengthy model iteration, and improve the efficiency of training task processing.

[0094] In one embodiment, the electronic device 1 may be a mobile phone, a tablet computer, a laptop computer, a desktop computer, or a large computing system composed of multiple computers.

[0095] Figure 2 A schematic diagram of an application scenario 200 of a training task information processing system provided in an embodiment of the present application. Figure 2 As shown, the system includes: a server 210 and a terminal 220, wherein:

[0096] Server 210 may be a data platform that provides training task information processing services, such as a customer service training platform in an e-commerce scenario. In actual scenarios, a customer service training platform may have multiple servers 210. Figure 2 In the figure, one server 210 is taken as an example.

[0097] The terminal 220 may be a computer, mobile phone, tablet or other device used by the user to log in to the customer service training platform. There may be multiple terminals 220. Figure 2 Two terminals 220 are taken as an example for illustration.

[0098] The terminal 220 and the server 210 can transmit information via the Internet, so that the terminal 220 can access the data on the server 210. The terminal 220 and / or the server 210 can be implemented by the electronic device 1.

[0099] The training task information processing solution of the embodiment of the present application can be deployed on the server 210, can also be deployed on the terminal 220, or can be partially deployed on the server 210 and partially deployed on the terminal 220. In actual scenarios, it can be selected based on actual needs, and this embodiment does not limit it.

[0100] When the training task information processing solution is fully or partially deployed on the server 210 , a calling interface may be opened to the terminal 220 to provide algorithm support to the terminal 220 .

[0101] The method provided in the embodiment of the present application can be implemented by executing the corresponding software code by the electronic device 1, and can be implemented by exchanging data with the server. The electronic device 1 can be a local terminal device. When the method is run on the server, the method can be implemented and executed based on the cloud interaction system, wherein the cloud interaction system includes a server and a client device.

[0102] In a possible implementation, the method provided in the embodiment of the present application provides a graphical user interface through a terminal device, wherein the terminal device can be the local terminal device mentioned above, or can be the client device in the cloud interaction system mentioned above.

[0103] like Figure 3A As shown, it is a schematic diagram of the overall architecture of the training task information processing system of an embodiment of the present application, taking the customer service training task scenario in the e-commerce scenario as an example, including a conversation interaction module, a QA decision module, a conversation management module, a simulation data management module, a scoring module and an inference proxy (inference proxy module), wherein:

[0104] Conversational interaction module: In the customer service training scenario, after the customer service clerk logs in to the training system and triggers a training request, the system establishes a training session between the customer service clerk and the virtual user, which can be an online session or a hotline session. The virtual user simulates the tone of the consumer user to send questions to the customer service clerk, and the customer service clerk can receive the virtual user's questions and answers (QA) through the training session. For example, the system sends "I want to ship!" through the virtual user, or the customer service clerk proactively sends "Hello, what's the problem?" in the form of text (text) or card (information card) in the training session. The system can reply to the customer service clerk's conversation content through the virtual user.

[0105] The QA decision module is used for prompt module management, parameter management, message processing UDF, memory / cache, result parsing / desensitization and link tracking. It is used to decide the virtual user's reply to the customer service clerk's conversation content based on the conversation context and preset prompt words between the customer service clerk and the virtual user. The QA decision module can transmit information to the conversation interaction module through ASR (hot) or TTS (hot).

[0106] The session management module is connected to the QA decision module and is used for online / hotline management, heartbeat management, and permission management.

[0107] The simulation data management module is connected to the QA decision module and is used to manage the simulation data of the case, mainly including order management, scenario / question / fact factor management and scoring rule management.

[0108] The scoring module is used to score the customer service staff's operations in the training tasks and manage the scoring items, scoring results and scoring reasons.

[0109] Inference proxy, connecting the scoring module, can include idealab (an AI proxy platform) proxy api and tpp scenarios / solutions, where openai (open source AI model) can be called through the idealab proxy api, such as calling gpt-3.5-turbo, gpt-4 8K, gpt-4 32K and other models. It is also possible to orchestrate the real-time reasoning link of the model through tpp scenarios / solutions, such as whale (Alibaba Group's internal large model deployment platform) calling SFT: chatglm-6b, SFT: Qwen, Qwen and other models, where SFT chatglm6b represents the fine-tuned chatGlm-6b model version, Qwen represents the Tongyi Qianwen model, and SFT Qwen represents the fine-tuned Qwen model.

[0110] like Figure 3B As shown, it is a schematic diagram of the implementation process of the training task information processing system of an embodiment of the present application. Taking the e-commerce customer service training scenario as an example, in the actual customer service training scenario, the operational training goal (goal) is first clarified. For example, the training goal can include the assessment of the basic ability and business ability of the customer service staff, where the basic ability can include: empathy, listening, understanding, expression and other abilities. Business abilities can include: pre-sales, expediting, freight, disputes and other business abilities.

[0111] Based on the practical training objectives, the corresponding AI material data can be configured for in-context-learning of large models. AI material data can include: multiple demands that consumers may have (demand 1, demand 2, etc.), the fact factors corresponding to these demands (fact 1, fact 2, etc.), the configuration information of the prompt word project and RAG.

[0112] Based on the configured AI material data, the inference performance of the practical training model is adjusted (model-inference). The training model can complete various agent tasks based on the preset large model base (model), such as member role-playing (agent1), executing the teaching / guidance task of step1 (agent2), and executing the scoring task of step2 for training operations (agent3). The scoring results of the training operations in step3 can be fed back to the system to adjust the system member role-playing process. The member role here can refer to the member users of the e-commerce platform, who are the objects that the customer service staff need to serve.

[0113] The model base can be a fine-tuned model sft-model or a general model (base-model) in the training domain.

[0114] Please see Figure 4 , which is a training task information processing method of an embodiment of the present application, the method can be Figure 1 The electronic device 1 shown is used to perform and can be applied to Figure 2-3B In the application scenario of information search shown in , the performance of the large model can be adjusted by adjusting the prompt words, different training case scenarios can be flexibly simulated, and the constraints of lengthy model iteration can be eliminated to improve the efficiency of training task processing. In this embodiment, the terminal 220 is used as an execution end as an example, and the method includes the following steps:

[0115] Step 401: In response to a training request, target case information corresponding to the training request and prompt word configuration information for a preset large model are determined.

[0116] In this step, the training request may be actively triggered by the trainee, for example, the trainee may be a customer service of an e-commerce platform, and the customer service logs in to the training platform and actively clicks to start training in the human-computer interaction interface to trigger the training request. The training request may also be triggered by the system, for example, the system is triggered according to preset conditions or periodically. The training request may carry the difficulty of the training requirement or a specific training identifier, and the target case information corresponding to the training request and the prompt word configuration information for the preset large model may be determined according to the specified training difficulty or specific training identifier, wherein the target case information may include case knowledge points, case ID information, guidance information for each stage of the case, case dialogue, case association table, case-related scenario information, and the like. The prompt word configuration information is used to adjust the input prompt words of the preset large model, and then adjust the reasoning performance of the preset large model to simulate different training scenarios.

[0117] Step 402: Generate a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template.

[0118] In this step, the preset prompt word template can be set according to the training business needs. Taking the customer service training in the e-commerce scenario as an example, the training purpose can be specified in combination with the e-commerce scenario's customer service capability requirements, and the structured prompt template can be combined with the business background to form a preset prompt word template that matches the training business scenario. By filling the prompt word configuration information corresponding to this training request into the corresponding slot in the preset prompt word template, the target prompt word of the preset large model can be generated. The target prompt word matches this training. Different target prompt words can guide the preset large model to simulate different training scenarios and simulate different virtual objects, thereby improving the richness of the training tasks and improving the efficiency of automatic training.

[0119] In one embodiment, the preset prompt word template includes at least one template slot, and the template slot is determined according to a preset evaluation index for the training operation.

[0120] In this embodiment, the preset prompt word template may include one template slot or multiple template slots. The specific content of the template slot can be flexibly configured. The performance of the large model can be adjusted by adjusting the template slot in the instruction, and different training case scenarios can be flexibly simulated. The preset evaluation index can be an ability assessment index for the trainee. Taking customer service training in an e-commerce scenario as an example, the preset prompt word template can determine the corresponding slot based on the indicator of the business assessment ability of the customer service staff, as shown in Table 1 below, in order to disassemble and abstract the prompt in combination with the business background, an example of a structured prompt template and template slot:

[0121] Table 1

[0122]

[0123]

[0124]

[0125] As can be seen from Table 1, by analyzing the assessment capabilities required for customer service in actual business, the corresponding prompt word template slot can be determined so that the prompt word template can guide the large model to output training tasks for the corresponding assessment capabilities and provide targeted training for customer service.

[0126] According to the above analysis, we can configure Figure 5 The prompt template 1 shown:

[0127] #Role(Role slot):

[0128] You are an online shopping consumer, and you need to express your problems and demands to the platform customer service and evaluate the solutions provided by the platform customer service.

[0129] ##Background (background slot)

[0130] You have encountered the problem {isoStandardProblemName(problem name)}

[0131] {factorList(factor list)}

[0132] ##Goals(Goal slot):

[0133] Describe the shopping problems and demands you encounter

[0134] Respond to the platform customer service's solution and express whether you accept it

[0135] ##Constraints (constraint slot):

[0136] 1. Remember that you are a consumer, not a platform customer service representative.

[0137] 2.{emotion0ption(turn on emotion)}

[0138] 3. Ask the platform customer service for a solution, and clearly express whether you accept the solution given by the platform customer service

[0139] 4. If you don’t get a satisfactory solution, the platform customer service says goodbye, you should ask for a solution

[0140] 5.{comprehension(understanding)}

[0141] 6.{furtherQuestion(further question feature)}

[0142] 7. If you do not understand what the platform customer service is saying, you should make it clear that you do not understand and ask for a more detailed explanation.

[0143] 8. Reply according to the format of "OutputFormat", use Chinese, and the maximum number of characters is 50

[0144] ##Skills(Skill slots):

[0145] 1.{questionDescribe(question description ability)}

[0146] 2. Understand and evaluate the platform customer service solutions

[0147] ##OutputFormat(output slot):

[0148] consumer:

[0149] Among them, the Role slot in prompt template 1 is used to indicate the role positioning of the virtual user. The background slot is used to indicate the actual business scenario of the training, such as the delivery urging scenario, the refund scenario, etc. Different sub-slots can be configured in the Background slot, such as the sub-slot isoStandardProblemName for configuring the question name and the sub-slot factorList for configuring the fact factor. Here, the question refers to the standard question of the corresponding consumer user in a specific business scenario. For example, in the refund scenario, the question can include "I want a refund". The fact factor is used to configure more specific business scenarios. For example, in the refund scenario, fact factor 1: the user does not like the product and refunds, and fact factor 2: the user refunds due to product quality issues. In this way, different fact factors can refine the training scenarios more specifically and guide customer service to give different solutions.

[0150] Constraints (constraint slot) is used to configure the output restrictions of the preset large model. The sub-slot emotion0ption (enable emotion) is used to configure the emotional tendency of the virtual user, such as the virtual user can be angry or happy. The sub-slot comprehension (comprehension) is used to configure the comprehension ability of the virtual user, such as a strong understanding of the solution given by the customer service and understanding what the customer service said, or a weak understanding of the solution given by the customer service and not understanding what the customer service said and needing further questioning. The sub-slot furtherQuestion (question feature) is used to configure whether the virtual user will ask further questions.

[0151] Skills (skill slots) are used to configure the basic skills of virtual users. For example, the questionDescribe sub-slot can be used to configure the description ability of the virtual user, which indicates whether the virtual user can express his or her questions clearly.

[0152] In one embodiment, the prompt word configuration information includes a scene question and / or fact factor for characterizing the training scene characteristics. The preset prompt word template includes a background slot. Step 402 may specifically include: filling the background slot in the preset prompt word template according to the scene question and / or fact factor to generate a preset large model of target prompt words.

[0153] In this embodiment, the scenario question refers to the standard question of the corresponding consumer user in a specific business scenario, such as a refund scenario, the question may include "I want a refund", and the fact factor is used to characterize a more specific business scenario, such as in the refund scenario, fact factor 1: the user does not like the product and refunds, fact factor 2: the user refunds due to product quality issues. In this way, different fact factors can refine the training scenario more specifically and guide customer service to give different solutions. The target prompt word can be constructed by filling the scenario question and / or fact factor in the prompt word configuration information into the corresponding slot in the preset prompt word template. Taking the aforementioned prompt template 1 as an example, the scenario question in the prompt word information is filled into the sub-slot isoStandardProblemName in the Background (background slot) in the prompt template 1, and / or the fact factor in the prompt word information is filled into the sub-slot factorList in the Background (background slot) in the prompt template 1, so as to adjust the prompt model slot according to the pre-configured scenario question and fact factor to adjust the target prompt of the large model.

[0154] In one embodiment, the prompt word configuration information includes the character characteristics of the virtual user. The preset prompt word template includes constraint condition slots. Step 402 may specifically include: filling the constraint condition slots in the preset prompt word template according to the character characteristics to generate the target prompt word of the preset large model.

[0155] In this embodiment, the character personality characteristics are used to characterize the personality characteristics of the virtual character. Taking the customer service training scenario as an example, the character personality characteristics can be used to simulate the personality characteristics of consumers in actual scenarios. Among them, the character personality characteristics include but are not limited to one or more of the emotional characteristics, understanding characteristics, and consultation method characteristics of the virtual character. Emotional characteristics are such as being angry or happy. The understanding characteristics are used to characterize the virtual user's understanding of the solution given by the customer service, such as a strong understanding of the solution given by the customer service, and a clear understanding of what the customer service said, or a weak understanding of the solution given by the customer service, and an inability to understand what the customer service said, and further questioning is required. The consultation method characteristics are used to characterize the inquiry habits of the virtual character, such as whether the virtual character will ask further questions. The target prompt word can be constructed by filling the character personality characteristics in the prompt word configuration information into the constraint condition slot in the preset prompt word template.

[0156] Taking the aforementioned prompt template 1 as an example, the emotional features in the prompt word information are filled into the sub-slot emotion0ption in the Constraints (constraint slot) in prompt template 1, the comprehension features are filled into the sub-slot comprehension in the Constraints (constraint slot) in prompt template 1, and the consultation method features are filled into the sub-slot furtherQuestion in the Constraints (constraint slot) in prompt template 1, so as to adjust the prompt model slots according to the pre-configured character personality characteristics, and then adjust the target prompt of the large model.

[0157] In one embodiment, the prompt word configuration information includes the character skill characteristics of the virtual user. The preset prompt word template includes skill slots. Step 402 may specifically include: filling the skill slots in the preset prompt word template according to the character skill characteristics to generate a target prompt word of the preset large model.

[0158] In this embodiment, the character skill features are used to characterize the skills possessed by the virtual character. Different character skills can simulate different consumer users. For example, a consumer user who is eloquent has relatively strong language expression skills and can usually describe the problems he encounters clearly, which puts relatively less pressure on customer service. For consumer users who are not good at speaking, their language expression skills are often weak, and customer service needs to be more patient. The target prompt word can be constructed by filling the character skill features in the prompt word configuration information into the skill slots in the preset prompt word template. By simulating different consumer users through character skills, customer service training tasks of different difficulty levels can be achieved, the scene diversity of training tasks can be improved, and the training tasks can be closer to real scenes.

[0159] Taking the aforementioned prompt template 1 as an example, the character skill characteristics in the prompt word information are filled into the sub-slot questionDescribe of Skills (skill slot) in prompt template 1 to configure the description ability of the virtual user, and adjust the prompt model slot according to the pre-configured character skill characteristics, thereby adjusting the target prompt of the large model.

[0160] Step 403: Use a preset large model to process the target case information according to the target prompt word, and generate a target training task corresponding to the training request. The target training task includes the session information of the virtual user. The target training task is used to guide the trainee to interact with the virtual user according to the session information.

[0161] In this step, the contextual learning characteristics of the large model are utilized, and the target prompt words are used to constrain the large model to perform training scenario conversation simulation and generate matching target training tasks. The target training tasks guide the trainees to interact with the virtual users through the conversation information of the virtual users, thereby realizing the training process. It is possible to adjust the performance of the large model by adjusting the prompt words, flexibly simulate different training case scenarios, get rid of the constraints of lengthy model iterations, and improve the efficiency of training task processing.

[0162] In one embodiment, step 403 may specifically include: generating a virtual order corresponding to the training request according to the target case information; obtaining a session context corresponding to the training request; inputting the order information, session context, target case information, and target prompt word of the virtual order into a preset macro model, so that the preset macro model generates a target training task under the guidance of the target prompt word.

[0163] In this embodiment, in order to make the scenario of the training task closer to the actual service scenario, after obtaining the target case information corresponding to the training request, the target case information can be analyzed and processed, and a virtual order corresponding to the training request can be generated according to the target case information, wherein the virtual order can include at least one of the following: order number, virtual member information corresponding to the order, order problem, member demand, etc., and the number of virtual orders can be one or more. When using the training task to train the trainee, different application scenarios will generate different conversation contexts. In order to accurately simulate the actual scenario, the conversation context corresponding to the training request can be detected in real time or at a fixed time, and then the order information, conversation context and target case information of the virtual order can be processed using the preset large model, so that the preset large model can generate a target training task closer to the actual scenario under the guidance of the target prompt word.

[0164] In one embodiment, generating a virtual order corresponding to the training request according to the target case information includes: obtaining a preset order generation tool bound to the target case information, and using the order generation tool to process the target case information to generate a virtual order corresponding to the training request.

[0165] In this embodiment, in order to quickly and accurately generate a virtual order that matches the training request, an order generation tool can be bound to the target case information in advance. For example, a suitable order generation tool identifier can be bound to the target case information. In the process of generating the virtual order, the corresponding order generation tool is directly called according to the order generation tool identifier, and the target case information is processed using the order generation tool to generate a virtual order corresponding to the training request, thereby improving the efficiency of virtual order generation.

[0166] In one embodiment, after step 403, the method further includes: in response to reply information to the conversation information, obtaining a preset evaluation rule bound to the target case information for evaluating the training operation, wherein the preset evaluation rule includes at least one preset evaluation indicator for evaluating the training operation and a scoring weight corresponding to each preset evaluation indicator. The evaluation score of the reply information is determined according to the at least one preset evaluation indicator, the scoring weight and the conversation information.

[0167] In this embodiment, in order to accurately conduct a comprehensive evaluation of the training process and to know the ability status of the trainee in the training, the operation of the trainee can be scored. In the interactive operation training based on conversation, the trainee replies to the conversation information of the virtual user to complete the training operation, so the reply information of the trainee in the training process can be used as the scoring object. The target case information is pre-configured with a preset evaluation rule for evaluating the training operation, and the preset evaluation rule includes but is not limited to at least one preset evaluation indicator for evaluating the training operation and the scoring weight corresponding to each preset evaluation indicator. For example, as shown in Table 1 above, the preset evaluation indicator can correspond to the assessment ability, and each assessment ability can be used as a preset evaluation indicator. Each assessment ability is configured with a corresponding scoring weight, and together forms a preset evaluation rule. Then, the evaluation score of the reply information is determined according to the preset evaluation indicator, the scoring weight and the conversation information. The evaluation score can represent the degree of practical pros and cons of the trainee in this training task. Different case information can be configured with different evaluation rules to realize the assessment of different ability information of the trainee.

[0168] In one embodiment, before step 401, the method further includes: in response to an AI material configuration request regarding target case information, obtaining AI material information for a preset large model, the AI ​​material information including prompt word configuration information, and binding the AI ​​material information to the target case information.

[0169] In this embodiment, AI material information refers to material information that can assist the preset large model, such as prompt word configuration information, prompt word template structure information, etc. used to characterize and simulate different training scenarios, including but not limited to scenario labels, fact factors, role personality of virtual users, role skills and other information. Relevant AI material information can be pre-configured for the target case information. For example, the trainer can configure AI material information that matches the business needs for the target case information in advance according to actual business needs, and bind the AI ​​material information configured by the trainer to the target case information, so as to facilitate direct call when the target training task is generated later, thereby improving the generation efficiency of the training task. In this way, by configuring AI material information for the case information, the output of the preset large model can be changed, different training scenarios can be simulated, and different training task difficulties and training goals can be achieved. It gets rid of the disadvantage of relying on model iteration to change the training scenario, and improves the training efficiency.

[0170] In one embodiment, before step 401, the method further includes: in response to a case configuration request, obtaining at least one preset case configuration information input, the preset case configuration information including one or more of a basic configuration, a solution configuration, and a scoring configuration of the preset case, and generating preset case information for training according to the preset case configuration information.

[0171] In this embodiment, one or more preset case information can be pre-configured to form a case information library, so that when a training request is triggered, the target case information matching the training request can be determined from the case information library. The case configuration request can be triggered by a trainer or automatically triggered by the system. Taking the trainer triggering the case configuration request as an example, when the case configuration request is received, at least one preset case configuration information input by the trainer is first obtained. The preset case configuration information includes but is not limited to the basic configuration, solution configuration, and scoring configuration of the preset case. Then, the preset case information for training is generated based on the case configuration information entered by the trainer and stored in the case information library.

[0172] like Fig. 6A As shown, it is a schematic diagram of a training task information processing system provided in an embodiment of the present application. Taking customer service training in an e-commerce scenario as an example, the system may include: xixi-training (a customer service training background application), protocol access, tpp-walle (a model reasoning framework built on the basis of the TPP platform in a training scenario) and platform capabilities, among which:

[0173] The xixi-training part may include an online QA practice module and a hotline QA practice module for providing training practices, a case configuration module for providing case information configuration, a prompt operation module for providing prompt word information configuration, a scoring and evaluation module for providing scoring for training operations, and a case manufacturing module for generating virtual orders for training tasks.

[0174] The protocol access part can be accessed through the xixi-training-protocol protocol.

[0175] tpp-walle can include xlangchain (the basic inference code framework of the large model based on the TPP platform), large model link, prompt project, unified access to multiple services, model AB and inference monitoring modules. Among them, the prompt project is used to generate the target prompt words of the large model according to the prompt word configuration information and prompt word template, the model AB is used to conduct AB experiments on the large model, and the inference monitoring module can monitor the inference process of the large model, find problems in practice and provide feedback.

[0176] The platform capabilities can provide TPP experiment diversion, TPP experiment configuration and whale.

[0177] Through the above system, the aggregation of training scenarios can be achieved through TPP links to realize multi-AB multi-layered structures, such as scenario 37019 (aggregation scenario - multi-AB multi-layered structure). When creating an AB experiment module, one AB experiment can correspond to one business channel, or the solution layer can be unified. It can also manage grayscale rules to control business diversion.

[0178] Through layered configuration, AB experiment diversion can be controlled, and through multi-layer AB configuration, different configurations can be achieved for different channels.

[0179] like Figure 6B As shown, it is a flow chart of a training task information processing method of an embodiment of the present application, taking the training of customer service in an e-commerce scenario as an example, including: case preparation, case operation (operation by operation personnel), case evaluation (operation by labeling technicians), customer service second-hand operation and service solution operation, among which:

[0180] In the case creation part, at the beginning of creation, configure the basic information of the case, such as case name, case description, case label, etc. Then select the training scenario, such as selecting the service process and plan. Then select the order creation tool for the case and simulate the incoming call. You can further practice in the customer service assistant - trial, such as selecting the standard question, manual node and other information, and you can also perform guided configuration. Then the operator configures the prompt slot, sets the score, and the creation is completed to generate the case information.

[0181] The case operation part receives case information, issues QA evaluation tasks for the case information, and evaluates the QA accuracy of the case information. The QA accuracy can be given by the case evaluation part to determine whether the QA accuracy meets the standard. If it does not meet the standard, the large model is instructed to iterate and optimize the model. If it meets the standard, the case information is added to the course package. For the case information in the course package, a training map or training tasks can be created, such as creating a general training task, so that the customer service staff can enter the actual operation scene according to the training map or training task.

[0182] In the case evaluation part, labelers are trained and practiced on QA evaluation tasks based on case information, and perform QA evaluation labeling. They can also perform offline data statistics on QA accuracy and feed back the QA accuracy to case operations.

[0183] When customer service staff start training, they open a session and conduct training through QA or practical guidance (such as manual nodes) in the session. After the session ends, the training operation is scored and the training ends. The practical guidance in the training can be compared and verified with the standard operation to update the order selection and simulated line entry in the case making part and complete information feedback.

[0184] The service plan operation part can monitor and detect the entire process, and if necessary, add, delete or change the service process, manual nodes and other information in the case information to improve the case information.

[0185] like Figure 6C As shown, it is a signaling interaction diagram of an AI case configuration process of an embodiment of the present application, based on Fig. 6A The system shown can respectively implement the basic information configuration, solution configuration, AI material-prompt configuration and case scoring configuration of the relevant training cases, where:

[0186] Basic information configuration includes: the trainer sets the basic information, the training configuration workbench configures the case category according to the basic information, the training service xixi-training generates a scene tree code (identification code) according to the case category, xixi-base (database) returns the number of case categories to the training service xixi-training according to the scene tree code, the training service xixi-training returns the category list to the training configuration workbench according to the number of categories, the training configuration workbench returns the case name, case summary, administrator information and case label to the training service xixi-training, and the training service xixi-training returns a notification of successful basic information setting to the trainer through the done (completed) instruction.

[0187] Solution configuration includes: the trainer enters the search conditions in the training configuration workbench to search for solutions, the training configuration workbench notifies the training service xixi-training to perform the conditional search, the training service xixi-training notifies uservice (user service) to perform the conditional search, uservice returns the retrieved top list (a list of search results sorted in the front preset ranking) to the training service xixi-training, the training service xixi-training merges the top list with the existing case data, the training service xixi-training returns the top service questions / processes / solutions to the training configuration workbench, and the training configuration workbench returns the top service questions / processes / solutions to the trainer. The trainer enters the order-making tool selected for the case in the training configuration workbench, the training configuration workbench searches for the corresponding order-making tool in the order-making pool, and sends the retrieved order-making tool to the training service xixi-training. The training service xixi-training returns the bound order-making tool to the training configuration workbench, and the training configuration workbench notifies the trainer of the order-making tool bound to the case.

[0188] AI material-prompt configuration, including: the trainer operates the example line, the training configuration workbench notifies the training service xixi-training to create a trial session according to the example, the training service xixi-training obtains the trial completed questions and fact factors from the training configuration workbench, and the training service xixi-training returns the determined scenario questions and fact factors to the training configuration workbench. The trainer can perform secondary editing on the returned scenario questions and fact factors, and can edit the target constraints in the prompt word template. The training configuration workbench can render the prompt in real time according to the edited content, and save all configured AI materials to the training service xixi-training. The training service xixi-training notifies the training configuration workbench that the configuration has been completed through the done command.

[0189] Case scoring configuration includes: the trainer sets the scoring through the training configuration workbench, the training configuration workbench sends the scoring weights of various evaluation indicators to the training service xixi-training, the training service xixi-training returns the setting completion to the training configuration workbench, and the training configuration workbench returns the practical configuration success to the trainer.

[0190] Through the above configuration processes, various information configurations of training cases can be completed in advance, which is convenient for direct calling when a training request is triggered.

[0191] like Fig.6DAs shown, it is a schematic diagram of an AI case configuration interface provided by an embodiment of the present application. Taking customer service training in an e-commerce scenario as an example, the configurable items included are: basic information configuration, solution configuration, AI material configuration and scoring configuration. Some guidance information can be displayed to the trainer in the configuration interface to assist the trainer in configuration. For example, in the configuration column of "correct example", the prompt message "Please give a correct example by conducting a practical trial according to the service plan and the reference order" is displayed. If the trainer has configured the correct example, "Currently completed example" can be displayed, and the "Re-example" button can be configured, so that when the trainer wants to change the example, click the "Re-example" button to change the correct example. The trainer can enter the standard questions that consumer users may ask in a specific scenario in the "scenario question" configuration column, such as Question 1: "I changed my mobile phone number, and the previous one is no longer used. I need to change the bound mobile phone number." For scenario questions, one or more corresponding fact factors can be configured to more specifically limit the training scenario. For example, Question 1 can be configured with Fact Factor 1 "The demand is to change the bound mobile phone number." The trainer can edit and rewrite the content of the questions or fact factors to make the questions and / or fact factors more in line with actual training needs.

[0192] The configuration interface can also display the goals and constraints of the large model. For example, based on the training and assessment indicators for customer service, you can determine the goals and constraints, such as: assessing customer service's basic listening ability, empathy, problem location ability, problem solving ability, negotiation and objection handling ability, and emotional soothing ability.

[0193] The configuration interface can also display the AI ​​case configuration bar and give a prompt message "Based on the above information, the system will generate the following prompt for inputting the large model to generate the AI ​​case". The prompt content is as follows:

[0194] #Role:

[0195] You are an online shopping consumer. You need to express your problems and demands to the platform customer service and evaluate the solutions provided by the platform customer service.

[0196] Very angry

[0197] ##Background

[0198] You have changed your mobile phone number and no longer need the previous one, so you need to change the mobile phone number bound to Alipay

[0199] The request is to change the mobile phone number bound to Alipay

[0200] ##Goals:

[0201] Describe the shopping problems and requests you encountered.

[0202] like Fig. 7A As shown, it is a schematic diagram of signaling interaction in a customer service training process provided by this application. Taking customer service training in an e-commerce scenario as an example, it mainly includes: training practice line entry, simulation QA and training scoring, among which:

[0203] The training practice online includes: the customer service clerk starts the practice, triggers the training request on the training workbench, the training workbench responds to the training request, sends the training task id and the corresponding target case id to the training service xixi-training, the training service xixi-training obtains the corresponding business document from the document pool, and initializes the dialogue system state, such as initializing the prompt word, document state, etc. The training service xixi-training requests the virtual member to come online from xixi-online (a workbench online system), and xixi-online returns the training session id to the training service xixi-training. The training service xixi-training notifies the training workbench to complete the simulation online, and the training workbench loads the training session context. After the loading is completed, the training workbench notifies the customer service clerk that the loading is complete.

[0204] Simulation QA includes: the customer service representative sends questions and answers to the training workbench, the training workbench sends questions and answers to xixi-online, xixi-online records the chat content, and can notify the training workbench that the chat has been recorded through the done command. xixi-online pushes messages to the training service xixi-training, which can carry the recorded chat content. The training service xixi-training collects the chat context and sends a request carrying the chat context to the tpp-walle training intelligent service to request the intelligent service. The tpp-walle training intelligent service prepares the pre-scenario configuration, the tpp-walle training intelligent service notifies the diamond configuration platform to obtain the simulation prompt configuration, and the diamond configuration platform sends the obtained simulation prompt configuration to the tpp-walle training intelligent service through the done command. The tpp-walle training intelligent service generates the target prompt based on the simulation prompt configuration and voice chat assembly. The tpp-walle training intelligent service requests the large model service from the whale large model deployment platform. The whale large model deployment platform uses the large model to perform reasoning based on the target prompt and generates the output result of the large model. The whale large model deployment platform returns the model reasoning result to the tpp-walle training intelligent service. The tpp-walle training intelligent service post-processes the model reasoning result and returns the post-processed model reasoning result to the training service xixi-training. The training service xixi-training generates a reply message to the customer service clerk based on the model reasoning result and sends it to xixi-online. xixi-online records the voice chat content based on the reply message. xixi-online generates a robot (virtual user) reply based on the reply message and sends the robot reply to the training workbench. The training workbench sends the robot reply to the customer service clerk for message display.

[0205] In the training scoring part, after the customer service clerk completes the training exercises, the training workbench requests the training service xixi-training to obtain the total score of this training. The training service xixi-training requests the solution iso (Intelligence Service Operation, the intelligent service operation system of Alibaba's passenger transport department) for the SOP instance rendered for this training session. The solution iso returns the solution of this training session to the training service xixi-training. Here, the solution refers to the service solution given by the customer service clerk to the virtual user's problem. The training service xixi-training matches and scores the key nodes of the solution of this training session and determines the total score. The training service xixi-training sends the total score details to the training workbench, and the training workbench returns the actual operation total score details to the customer service clerk.

[0206] like Figure 7B As shown, it is a schematic diagram of the interactive interface of the customer service end in the customer service training process provided by this application. In the customer service end training interface, a training dialog box can be displayed, and the conversation context between the customer service attendant and the virtual user can be displayed in the dialog box. For example:

[0207] Customer service: "Hello, I am a customer service representative. I am happy to serve you. Could you describe your problem and request?"

[0208] Virtual user: "The goods I purchased on your platform have exceeded the agreed delivery time, but the seller has not yet shipped them. What should I do? I hope you can help me urge the seller to ship as soon as possible."

[0209] The training interface can also display an input box, in which the customer service staff can enter the reply information to the virtual user, such as "I'm sorry, it really took up a lot of your time, please wait, the staff will help you verify whether the seller has not fulfilled the delivery agreement."

[0210] The customer service assistant can also be displayed on the training interface, so that the virtual user can obtain the order-related information that the virtual user is inquiring about through the customer service assistant, simulating the conversation interface between customer service and consumer users in real scenarios.

[0211] The above training task information processing method, by abstracting the portrait of the trainee in the training, disassembles and abstracts the prompt in combination with the business background, structures the prompt template and template slot, and uses the In Context Learning feature of the big model to constrain the model to simulate the scene through prompt engineering. The model's performance is adjusted by adjusting the template slot in the instruction, and different training case scenarios are flexibly simulated, which not only gets rid of the rigid script case training mode and the lengthy model iteration capability, but also speeds up the AI ​​case operation efficiency. It also realizes the operationalization of prompt words, greatly improving the operability and operability of the big model in the training scenario.

[0212] According to the technical ideas of the embodiments of the present application, an experimental comparison was made between the schemes using Gpt4 and Qwen-14B based on prompt engineering and ChatGlm-6b (model) using SFT, and the conclusions are shown in Table 2 below:

[0213] Table 2

[0214]

[0215]

[0216] in, and * respectively represent the evaluation scores, The more *, the higher the score. Based on the results of the above experiment, we can get:

[0217] 1. When the general large model of the base is strong enough, the embodiment of the present application uses prompt engineering to constrain the model to complete some specific tasks, and the model performance is even better.

[0218] 2. This application adjusts the performance of the model by adjusting some constraints in the instructions to improve the simulation effect of the large model on specific training scenarios.

[0219] The solution of the embodiment of the present application has at least the following beneficial effects:

[0220] 1. High case operation efficiency: Building the prompt project on the basis of the large base model greatly improves operability. It does not need to rely on model iteration frequently and makes full use of the generalization ability of the base model.

[0221] 2. Real-world scenario simulation: Through prompt's flexible operational configuration, the big model can simulate user problems and demands in different service scenarios, and randomly generate different types of virtual users, such as virtual users with unclear expressions, problems with comprehension, and extremely negative emotions, to restore a more realistic service scenario.

[0222] 3. Improved conversational capabilities: Based on the operational configuration of the prompt, the large model can simulate different user characteristics, such as asking follow-up questions based on the solution, questioning the solution, changing emotions based on the customer service staff's comfort, etc., which can further improve the general capabilities of the customer service staff.

[0223] Please see Fig. 8A , which is a training task information processing method of an embodiment of the present application, the method can be Figure 1 The electronic device 1 shown is used to perform and can be applied to Figure 2-3B In the application scenario of information search shown in , the performance of the large model can be adjusted by adjusting the prompt words, different training case scenarios can be flexibly simulated, and the constraints of lengthy model iteration can be eliminated to improve the efficiency of training task processing. This embodiment takes the terminal 220 as an example of the execution end. Compared with the above embodiment, this embodiment takes the training scenario of the customer service assistant in the e-commerce scenario as an example. The method includes the following steps:

[0224] Step 801: In response to a training request for customer service personnel, target case information corresponding to the training request and prompt word configuration information for a preset large model are determined, where the target case information is case information about a product order.

[0225] Step 802: Generate a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template.

[0226] Step 803: Use the preset large model to process the target case information according to the target prompt words, and generate a target training task corresponding to the training request. The target training task includes the session information of the virtual user. The target training task is used to guide the customer service staff to interact with the virtual user according to the session information.

[0227] For details of each step of the above training task information processing method, please refer to the relevant description of the above embodiment, which will not be repeated here.

[0228] Please see Figure 8B , which is a task information processing method of an embodiment of the present application, the method can be Figure 1 The electronic device 1 shown is used to perform and can be applied to Figure 2-3BIn the application scenario of information search shown in , the performance of the large model can be adjusted by adjusting the prompt words, different task case scenarios can be flexibly simulated, and the constraints of lengthy model iteration can be eliminated to improve the efficiency of task processing. In this embodiment, the terminal 220 is used as the execution end as an example, and the method includes the following steps:

[0229] Step S1: In response to a task processing request from a user, target case information corresponding to the task processing request and prompt word configuration information for a preset large model are determined.

[0230] Step S2: Generate a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template.

[0231] Step S3: Use a preset large model to process the target case information according to the target prompt word, and generate a target task corresponding to the task processing request. The target task includes the session information of the virtual user. The target task is used to guide the specified object to interact with the virtual user according to the session information.

[0232] The task processing request of this embodiment can be a request for various information processing tasks triggered by the user, such as a training request or a conversation request. The specified object here can be specified by the user according to the needs, such as a training object or a conversation object.

[0233] For task processing requests, the corresponding target case information and large model prompt word configuration information are obtained respectively, combined with the pre-structured prompt word template, and the target prompt word corresponding to the task processing request is generated according to the corresponding prompt word configuration information. The context learning characteristics of the large model are utilized to constrain the large model through the target prompt word to simulate the task scenario and generate a matching target task. The target task guides the specified object to interact with the virtual user through the virtual user's session information, thereby realizing the task processing process. It is realized that the performance of the large model can be adjusted by adjusting the prompt word, and different task case scenarios can be flexibly simulated, which gets rid of the constraints of lengthy model iteration and improves the efficiency of task processing.

[0234] Please see Fig. 9 , which is a training task information processing device 900 of an embodiment of the present application, which can be applied to Figure 1 The electronic device 1 shown can be applied to Figure 2-3B In the application scenario of information search shown in , the performance of the large model can be adjusted by adjusting the prompt words, different training case scenarios can be flexibly simulated, and the constraints of lengthy model iteration can be eliminated to improve the efficiency of training task processing. The device includes: a determination module 901, a first generation module 902 and a second generation module 903, and the functional principles of each module are as follows:

[0235] The determination module 901 is used to determine, in response to a training request, target case information corresponding to the training request and prompt word configuration information for a preset large model.

[0236] The first generating module 902 is used to generate a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template.

[0237] The second generation module 903 is used to use a preset large model to process the target case information according to the target prompt word, and generate a target training task corresponding to the training request. The target training task includes the session information of the virtual user. The target training task is used to guide the trainee to interact with the virtual user according to the session information.

[0238] In one embodiment, the preset prompt word template includes at least one template slot, and the template slot is determined according to a preset evaluation index for the training operation.

[0239] In one embodiment, the prompt word configuration information includes a scene question and / or fact factor for characterizing the training scene characteristics. The preset prompt word template includes a background slot. The first generation module 902 is used to fill the background slot in the preset prompt word template according to the scene question and / or fact factor to generate a target prompt word of the preset large model.

[0240] In one embodiment, the prompt word configuration information includes the character characteristics of the virtual user. The preset prompt word template includes a constraint condition slot. The first generation module 902 is used to fill the constraint condition slot in the preset prompt word template according to the character character characteristics to generate a target prompt word of the preset large model. The character character characteristics include: one or more of the emotional characteristics, understanding characteristics, and consultation method characteristics of the virtual character.

[0241] In one embodiment, the prompt word configuration information includes the character skill characteristics of the virtual user. The preset prompt word template includes skill slots. The first generation module 902 is used to fill the skill slots in the preset prompt word template according to the character skill characteristics to generate a target prompt word of the preset large model.

[0242] In one embodiment, the second generation module 903 is used to generate a virtual order corresponding to the training request according to the target case information. The session context corresponding to the training request is obtained. The order information, session context, target case information and target prompt word of the virtual order are input into the preset macro model, so that the preset macro model generates the target training task under the guidance of the target prompt word.

[0243] In one embodiment, the second generating module 903 is used to obtain a preset order generating tool bound to the target case information, and use the order generating tool to process the target case information to generate a virtual order corresponding to the training request.

[0244] In one embodiment, it further includes: an evaluation module, which is used to process the target case information according to the target prompt word using a preset large model to generate a target training task corresponding to the training request, and then, in response to the reply information to the session information, obtain the preset evaluation rules bound to the target case information for evaluating the training operation, wherein the preset evaluation rules include at least one preset evaluation indicator for evaluating the training operation and the scoring weight corresponding to each preset evaluation indicator. The evaluation score of the reply information is determined according to the at least one preset evaluation indicator, the scoring weight and the session information.

[0245] In one embodiment, it further includes: a first configuration module, which is used to respond to an AI material configuration request about the target case information, obtain AI material information for the preset large model before determining the target case information corresponding to the training request and the prompt word configuration information for the preset large model in response to the training request, and the AI ​​material information includes the prompt word configuration information. Bind the AI ​​material information with the target case information.

[0246] In one embodiment, the method further includes: a first configuration module, which is used to obtain at least one preset case configuration information input in response to the case configuration request before determining the target case information corresponding to the training request and the prompt word configuration information for the preset large model in response to the training request, wherein the preset case configuration information includes one or more of the basic configuration, solution configuration, and scoring configuration of the preset case. The preset case information for training is generated according to the preset case configuration information.

[0247] For a detailed description of the above-mentioned training task information processing device 900, please refer to the description of the relevant method steps in the above-mentioned embodiment. Its implementation principle and technical effect are similar, and will not be repeated here in this embodiment.

[0248] Fig.10 The following is a schematic diagram of a cloud device 100 provided in an exemplary embodiment of the present application. The cloud device 100 can be used to run the method provided in any of the above embodiments. Fig.10 As shown, the cloud device 100 may include: a memory 1004 and at least one processor 1005, Fig.10 A processor is taken as an example.

[0249] The memory 1004 is used to store computer programs and can be configured to store various other data to support operations on the cloud device 100. The memory 1004 can be an object storage service (OSS).

[0250] Memory 1004 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0251] The processor 1005 is coupled to the memory 1004 and is used to execute the computer program in the memory 1004 to implement the solution provided by any of the above method embodiments. The specific functions and technical effects that can be achieved are not repeated here.

[0252] Furthermore, if Fig.10 The cloud device also includes: a firewall 1001, a load balancer 1002, a communication component 1006, a power supply component 1003 and other components. Fig.10 Only some components are shown schematically, which does not mean that the cloud device only includes Fig.10 Components shown.

[0253] In one embodiment, the above Fig.10 The communication component 1006 in is configured to facilitate wired or wireless communication between the device where the communication component 1006 is located and other devices. The device where the communication component 1006 is located can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G, LTE (Long Term Evolution, Long Term Evolution, referred to as LTE), 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component 1006 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 1006 also includes a near field communication (Near Field Communication, referred to as NFC) module to facilitate short-range communication. For example, the NFC module can be based on Radio Frequency Identification (Radio Frequency Identification, referred to as RFID) technology, Infrared Data Association (Infrared Data Association, referred to as IrDA) technology, Ultra Wide Band (UWB) technology, Bluetooth (bluetooth, referred to as BT) technology and other technologies to achieve.

[0254] In one embodiment, the above Fig.10 The power supply component 1003 provides power to various components of the device where the power supply component 1003 is located. The power supply component 1003 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device where the power supply component is located.

[0255] An embodiment of the present application further provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the method of any of the aforementioned embodiments is implemented.

[0256] An embodiment of the present application also provides a computer program product, including a computer program, which implements the method of any of the aforementioned embodiments when executed by a processor.

[0257] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0258] The above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform some steps of the methods of various embodiments of the present application.

[0259] It should be understood that the above-mentioned processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the application can be directly embodied as being executed by a hardware processor, or can be executed by a combination of hardware and software modules in the processor. The memory may include a high-speed RAM (Random Access Memory) memory, and may also include non-volatile storage NVM (Nonvolatile memory, NVM for short), such as at least one disk storage, and can also be a USB flash drive, a mobile hard disk, a read-only memory, a disk or an optical disk, etc.

[0260] The above storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special computer.

[0261] An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist as discrete components in an electronic device or a main control device.

[0262] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, clothing or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, clothing or device. In the absence of more restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, clothing or device including the element.

[0263] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0264] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods of each embodiment of the present application.

[0265] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of user data and other information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0266] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A training task information processing method, characterized in that: include: In response to a training request, determining target case information corresponding to the training request and prompt word configuration information for a preset large model; Generate a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template; The preset large model is used to process the target case information according to the target prompt word to generate a target training task corresponding to the training request, wherein the target training task includes the session information of the virtual user, and the target training task is used to guide the trainee to interact with the virtual user according to the session information.

2. The method according to claim 1, characterized in that: The preset prompt word template includes at least one template slot, and the template slot is determined according to a preset evaluation index for a training operation.

3. The method according to claim 1, characterized in that The prompt word configuration information includes scenario labels and / or fact factors for characterizing training scenario features; The preset prompt word template includes a background slot; The step of generating a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template includes: The background slots in the preset prompt word template are filled according to the scene question and / or the fact factor to generate the target prompt words of the preset large model.

4. The method according to claim 2 or 3, characterized in that: The prompt word configuration information includes the character characteristics of the virtual user; The preset prompt word template includes a constraint condition slot; the step of generating a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template includes: Fill the constraint condition slots in the preset prompt word template according to the character's personality characteristics to generate a target prompt word for a preset large model; The character characteristics include one or more of the virtual character's emotional characteristics, understanding characteristics, and consulting style characteristics.

5. The method according to claim 2 or 3, characterized in that: The prompt word configuration information includes the role skill characteristics of the virtual user; The preset prompt word template includes a skill slot; the step of generating a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template includes: The skill slots in the preset prompt word template are filled according to the skill characteristics of the character to generate target prompt words of a preset large model.

6. The method according to claim 1, characterized in that The step of using the preset large model to process the target case information according to the target prompt word to generate a target training task corresponding to the training request includes: generating a virtual order corresponding to the training request according to the target case information; Obtaining a session context corresponding to the training request; The order information of the virtual order, the conversation context, the target case information and the target prompt word are input into the preset large model, so that the preset large model generates the target training task under the guidance of the target prompt word.

7. The method according to claim 6, characterized in that The generating a virtual order corresponding to the training request according to the target case information includes: Obtaining a preset order generation tool bound to the target case information; The target case information is processed using the order generation tool to generate the virtual order corresponding to the training request.

8. The method according to claim 1, characterized in that After the target case information is processed according to the target prompt word using the preset large model to generate a target training task corresponding to the training request, the method further includes: In response to reply information to the conversation information, obtaining a preset evaluation rule bound to the target case information for evaluating the training operation, wherein the preset evaluation rule includes at least one preset evaluation indicator for evaluating the training operation and a scoring weight corresponding to each of the preset evaluation indicators; An evaluation score for the reply information is determined according to the at least one preset evaluation indicator, the scoring weight, and the conversation information.

9. The method according to claim 1, characterized in that: Before responding to the training request and determining the target case information corresponding to the training request and the prompt word configuration information for the preset large model, the method further includes: In response to an AI material configuration request regarding the target case information, obtaining AI material information for the preset large model, the AI ​​material information including prompt word configuration information; Bind the AI ​​material information with the target case information.

10. The method according to claim 1, characterized in that Before responding to the training request and determining the target case information corresponding to the training request and the prompt word configuration information for the preset large model, the method further includes: In response to the case configuration request, obtaining at least one preset case configuration information input, wherein the preset case configuration information includes one or more of a basic configuration, a solution configuration, and a scoring configuration of the preset case; Preset case information for training is generated according to the preset case configuration information.

11. A training task information processing method, characterized in that: include: In response to a training request for customer service personnel, determining target case information corresponding to the training request and prompt word configuration information for a preset large model, wherein the target case information is case information about a product order; Generate a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template; The preset large model is used to process the target case information according to the target prompt word to generate a target training task corresponding to the training request, wherein the target training task includes the session information of the virtual user, and the target training task is used to guide the customer service staff to interact with the virtual user according to the session information.

12. A task information processing method, characterized in that: include: In response to a task processing request from a user, determining target case information corresponding to the task processing request and prompt word configuration information for a preset large model; Generate a preset large model of target prompt words according to the prompt word configuration information and the preset prompt word template; The preset large model is used to process the target case information according to the target prompt word to generate a target task corresponding to the task processing request, wherein the target task includes the session information of the virtual user, and the target task is used to guide the designated object to interact with the virtual user according to the session information.

13. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the electronic device to perform the method described in any one of claims 1 to 12.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the processor executes the computer-executable instructions, the method according to any one of claims 1 to 12 is implemented.

15. A computer program product, characterized in that The method comprises a computer program, which implements the method according to any one of claims 1 to 12 when being executed by a processor.

Images