Session simulation method and device, storage medium and program product

Abstract

The embodiment of the invention provides a session simulation method and device, a storage medium and a program product. The method comprises the steps of obtaining target real session data matched with a simulation session intention; performing desensitization processing on privacy information of a real user in the target real session data to obtain target real session desensitization data; generating a first prompt word according to the target real session desensitization data; inputting the first cue word into a large language model, so that the large language model extracts simulation input information of at least one dimension from the target real session desensitization data according to the first cue word; generating a second cue word according to the simulation input information; and inputting the second prompt word into the simulation user robot, so that the simulation user robot and the customer service perform a natural language simulation session. The second prompt word is automatically generated through the large language model based on the real session data, a real session scene can be covered, and the capability of simulating a real user to perform session interaction by the simulation user robot is improved.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a session simulation method, device, storage medium and program product. Background Art

[0002] In the field of customer service training, with the continuous development of machine learning technology, using simulated user robots equipped with machine learning models to simulate real users and interact with customer service has become an emerging customer service training method. In this customer service training method, the simulated user robot talks to customer service in the user's tone, thereby helping customer service to practice responding to various user issues and improving customer service's adaptability and communication skills in actual work.

[0003] In actual applications, the simulation of real users is achieved by manually writing corresponding prompt words for the simulated user robot. However, this method will result in limited ability of the simulated user robot to simulate real users, and the final customer service training effect is not good. Summary of the invention

[0004] Embodiments of the present application provide a conversation simulation method, device, storage medium, and program product to enhance the ability of a simulated user robot to simulate a real user for conversational interaction.

[0005] In a first aspect, the present application provides a session simulation method, the method comprising:

[0006] Acquire target real session data that matches the simulation session intention, where the target real session data is session data generated by a real user having a real session with customer service;

[0007] Desensitizing the privacy information of the real user in the target real session data to obtain target real session desensitized data;

[0008] Generate a first prompt word according to the target real conversation desensitized data;

[0009] Inputting the first prompt word into a large language model, so that the large language model extracts at least one dimension of simulated input information from the target real conversation desensitized data according to the first prompt word;

[0010] generating a second prompt word according to the simulation input information;

[0011] The second prompt word is input into the simulated user robot so that the simulated user robot and the customer service have a simulated conversation in natural language.

[0012] In a second aspect, the present application provides a conversation simulation device, the device comprising:

[0013] An acquisition module is used to acquire target real session data that matches the simulation session intention, where the target real session data is session data generated by a real user having a real session with a customer service representative;

[0014] A processing module, used for desensitizing the privacy information of the real user in the target real session data to obtain the target real session desensitized data;

[0015] A generating module, configured to generate a first prompt word according to the target real conversation desensitized data; input the first prompt word into a large language model, so that the large language model extracts at least one dimension of simulated input information from the target real conversation desensitized data according to the first prompt word; and generate a second prompt word according to the simulated input information;

[0016] The conversation module is used to input the second prompt word into the simulated user robot so that the simulated user robot and the customer service can have a simulated conversation in natural language.

[0017] In a third aspect, the present application provides a computing device comprising a storage component and a processing component; the storage component stores one or more computer program instructions, the computer program instructions are called and executed by the processing component, and the processing component executes the one or more computer program instructions to implement the session simulation method as described in the first aspect.

[0018] In a fourth aspect, the present application provides a computer-readable storage medium storing a computer program, wherein the computer program is executed by a computer to implement the session simulation method as described in the first aspect.

[0019] In a fifth aspect, the present application provides a computer program product storing a computer program, which implements the session simulation method as described in the first aspect when executed by a computer.

[0020] In the customer service training scenario, the customer service staff can practice dealing with various user problems by interacting with the simulated user robot in conversation, thereby improving their adaptability and communication skills in actual work. The conversation simulation solution provided in the embodiment of the present application is used to improve the simulation level of the simulated user robot on the real user, so that the simulated conversation interaction between the customer service staff and the simulated user robot is closer to the interaction between the real user and the customer service staff in the real conversation scenario, thereby achieving better customer service training results.

[0021] Specifically, in the conversation simulation method provided in the embodiment of the present application, first, the target real conversation data matching the simulation conversation intention is obtained, wherein the target real conversation data is the conversation data generated by the real user and the customer service having a real conversation. Since the target real conversation data contains the privacy information of the real user, such as: real user identity information, payment information, etc., in order to ensure the information security of the real user, the privacy information of the real user in the target real conversation data will be desensitized to obtain the target real conversation desensitized data corresponding to the target real conversation data after desensitization. Afterwards, a first prompt word is generated according to the target real conversation desensitized data; the first prompt word is input into the large language model, so that the large language model extracts at least one dimension of simulation input information from the target real conversation desensitized data according to the first prompt word. Finally, a second prompt word is generated according to the simulation input information; the second prompt word is input into the simulation user robot, so that the simulation user robot and the customer service have a natural language simulation conversation.

[0022] In this solution, on the one hand, the simulated input information is extracted from the target real conversation desensitized data corresponding to the target real conversation data, which can ensure that the simulated input information has the conversation characteristics reflected by the real conversation, so that when the simulated user robot uses the second prompt words generated by the simulated input information to conduct a simulated conversation with the customer service, the simulated user robot is closer to the real user and has a better conversation simulation effect. On the other hand, the method of automatically extracting the simulated input information from the target real conversation desensitized data through a large language model and generating the second prompt words for inputting into the simulated user robot can effectively improve the generation efficiency of the second prompt words, thereby facilitating the configuration of the simulated user robot with the second prompt words corresponding to various simulated conversation intentions, covering various conversation scenarios, and meeting the diverse customer service training needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 A flowchart of a conversation simulation method provided in an embodiment of the present application;

[0025] Figure 2 A schematic diagram of a scenario of a simulation input information acquisition process provided in an embodiment of the present application;

[0026] Figure 3 A flowchart of obtaining target real session data provided by an embodiment of the present application;

[0027] Figure 4 A schematic diagram of a real conversation intention provided in an embodiment of the present application;

[0028] Figure 5 A flowchart of another session simulation method provided in an embodiment of the present application;

[0029] Figure 6 A flowchart of another session simulation method provided in an embodiment of the present application;

[0030] Figure 7 A flowchart of a third prompt word generation method provided in an embodiment of the present application;

[0031] Figure 8 A structural diagram of a conversation simulation device provided in an embodiment of the present application;

[0032] Fig. 9 For Figure 8 A schematic diagram of the structure of an electronic device corresponding to the conversation simulation device provided in the illustrated embodiment. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0034] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0035] 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 the embodiments of the present 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.

[0036] It should be noted that the technical solution of the embodiment of the present application is applicable to a network virtual environment, and the user described generally refers to a "virtual user". A real user can register a user account in the server through a registration method to obtain a user identity in the network environment. Optionally, the user can be any role that can interact with the customer service, such as a consumer, a merchant, etc.

[0037] The conversation simulation solution provided in the embodiment of the present application can be applied to customer service training scenarios. In customer service training scenarios, customer service personnel can practice responding to various user questions by interacting with the simulated user robot in conversation, thereby improving their adaptability and communication skills in actual work.

[0038] For ease of description, in the embodiments of the present application, the conversation between the customer service and the real user is called a real conversation, and the conversation between the customer service and the simulated user robot is called a simulated conversation. Similar to the real conversation, the customer service can interact with the simulated user robot through a variety of conversation channels such as telephone, online chat, email, social media, etc. in a simulated conversation.

[0039] With the continuous development of machine learning technology, the simulated user robot is equipped with a machine learning model. The machine learning model in the simulated user robot simulates the behavior and language expression of real users through pre-configured dialogue scripts and prompts that match various conversation scenarios, generates simulated conversation messages that conform to real user reactions in different conversation scenarios, and realizes simulated interaction between the simulated user robot and customer service. Among them, the prompt is used to guide the machine learning model to generate conversation content that meets the expectations. The prompt can be a simple phrase, sentence or complex paragraph.

[0040] Among them, optionally, the machine learning model configured in the simulated user robot includes but is not limited to: a large language model (Large Language Model l, LLM for short) for natural language processing, such as: a bidirectional encoder model from Transformers (Bid i rect iona l Encoder Representation from Transformers, BERT for short), etc.

[0041] In the traditional solution, the prompt words corresponding to various conversation scenarios pre-configured in the simulated user robot are generated in the following way: first, conversations corresponding to different conversation scenarios are manually written; then, based on the manually written conversations, prompt words corresponding to each conversation are manually written.

[0042] It is understandable that the conversation scenarios corresponding to real conversations are often complex and changeable, and real users are different from person to person. In the same conversation scenario, different real users have different interaction needs with customer service, and the resulting conversations are also different. For example, in the return scenario of the same product, some users may request a return and refund, while some users may request only a refund. The method of manually writing conversations corresponding to different conversation scenarios is often difficult to cover real conversation scenarios and various real user needs in real conversation scenarios, resulting in limited simulation effects of the final simulated user robot. In addition, the method of manually writing conversations and prompt words often takes a lot of time and effort, with high labor costs and low efficiency.

[0043] To solve at least one of the above technical problems, an embodiment of the present application provides a conversation simulation method: for any type of simulated conversation intention, first obtain target real conversation data matching the simulated conversation intention from a real conversation between a real user and customer service, wherein the simulated conversation intention can be described by conversation information of multiple dimensions such as conversation scenarios and user demands to be solved by the conversation; then, desensitize the privacy information of the real user in the target real conversation data to obtain target real conversation desensitized data; thereafter, extract at least one dimension of simulated input information from the target real conversation desensitized data through a large language model; finally, generate prompt words for inputting a simulated user robot according to the simulated input information, so that the simulated user robot and the customer service can conduct a natural language simulated conversation. In this solution, the simulated input information is extracted from the target real conversation desensitized data corresponding to the target real conversation data, which can ensure that the simulated input information has the conversation characteristics reflected by the real conversation scenario, so that when the simulated user robot uses the prompt words generated by the simulated input information to conduct a simulated conversation with the customer service, the simulated user robot is closer to the real user and has a better conversation simulation effect; the method of automatically extracting the simulated input information from the target real conversation desensitized data through a large language model and generating prompt words for inputting into the simulated user robot can effectively improve the efficiency of prompt word generation, thereby facilitating the configuration of prompt words corresponding to various simulated conversation intentions for the simulated user robot, covering various conversation scenarios, and meeting diverse customer service training needs.

[0044] 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.

[0045] The session simulation method provided in the embodiment of the present application can be executed by an electronic device, which can be a terminal device such as a PC, a laptop, a smart phone, or a server. The server can be a physical server including an independent host, or a virtual server, or a cloud server or a server cluster.

[0046] Figure 1 A flow chart of a session simulation method provided in an embodiment of the present application, such as Figure 1 As shown, the following steps may be included:

[0047] 101. Obtain target real session data that matches the simulation session intention, where the target real session data is session data generated by a real user having a real session with customer service.

[0048] The intention of a simulated conversation refers to the purpose or goal to be achieved through a simulated conversation. For example, customer service personnel conduct simulated conversations with simulated user robots to train or test their ability to resolve specific user demands in specific conversation scenarios.

[0049] Optionally, the simulation session intention may be described by at least one of the following session-related intention description information: session scenario, user demand, demand type, solution, etc.

[0050] Among them, the conversation scenario refers to the specific situation in which the customer service staff has a conversation with the user. For example, in the e-commerce shopping scenario, the conversation scenario may include: online shopping consultation, logistics inquiry, member rights consultation, merchant problem complaint, etc. User demands refer to the needs, questions or requests raised by the user in the conversation with the customer service staff. In actual applications, user demands can be associated with standard questions, and multiple similar user demands can be described in a unified manner through standard questions, such as: return issues, complaint issues, information consultation issues, etc. Demand types include: single demands and multiple demands. A single demand means that the user only raises one specific user demand in a conversation, and a multiple demand means that the user raises multiple specific user demands in a conversation. Solution refers to the solution provided by the customer service staff to solve the user demand.

[0051] Similar to the simulated session having the simulated session intent, the real session also has the real session intent. In the embodiment of the present application, the target real session data matching the simulated session intent is the session data corresponding to the target real session whose real session intent is the same or similar to the simulated session intent.

[0052] Among them, the target real conversation data contains various information used to understand and analyze the interaction characteristics between real users and customer service (such as: real users' emotions, demands, etc.), which can be used to help the simulated user robot simulate real users.

[0053] Optionally, the target real conversation data includes but is not limited to the following types of information: real conversation records between real users and customer service, real orders involved in real conversations, real user demands summarized by customer service, real solutions provided by customer service, etc.

[0054] The process of acquiring the target real conversation data will be described below. This embodiment mainly describes the process of acquiring prompt words for inputting into the simulated user robot based on the target real conversation data.

[0055] 102. Desensitize the privacy information of the real user in the target real session data to obtain the target real session desensitized data.

[0056] Since the target real session data contains the real user's private information, such as the real user's identity information, payment information, real order information, etc., in order to ensure the information security of the real user and avoid the impact of session simulation on the real user, the real user's private information in the target real session data will be desensitized to obtain the target real session desensitized data corresponding to the target real session data after the desensitization processing.

[0057] The privacy information of the real user in the target real conversation data is desensitized, including but not limited to data replacement, data masking, data deletion, etc. For example, the order number, name, address, etc. of the real user in the real order contained in the target real conversation data can be replaced to generate a simulated order. The simulated order can be output to the customer service so that the customer service can conduct a simulated conversation with the simulated user robot based on the information in the simulated order.

[0058] It is understandable that the target real conversation desensitized data contains the same data types as the target real conversation data, except that the target real conversation desensitized data does not contain the privacy information of the real user. Therefore, optionally, the target real conversation desensitized data includes but is not limited to the following types of information: real conversation records between real users and customer service, simulated orders corresponding to real orders involved in real conversations, real user demands summarized by customer service, real solutions provided by customer service, etc.

[0059] 103. Generate a first prompt word according to the target real conversation desensitized data.

[0060] Among them, the first prompt word is used to guide the large language model to extract at least one dimension of simulated input information from the target real conversation desensitization data.

[0061] 104. Input the first prompt word into the large language model, so that the large language model extracts at least one dimension of simulated input information from the target real conversation desensitized data according to the first prompt word.

[0062] It is understandable that in order to enable the simulated user robot to simulate a real user and conduct a simulated conversation with customer service that matches the intention of the simulated conversation, it is necessary to provide the simulated user robot with at least one dimension of information for simulating the real user, which is referred to as simulated input information in this embodiment.

[0063] The dimensions of the simulation input information can be customized based on actual session simulation requirements.

[0064] Optionally, the simulated input information of at least one dimension can be obtained from the target real conversation desensitized data manually. However, this method often has the problems of high manpower consumption and low efficiency.

[0065] Therefore, the embodiment of the present application provides a method for automatically obtaining at least one dimension of simulated input information through a machine learning model to improve the efficiency of simulated input information. Specifically, at least one dimension of simulated input information is extracted from the target real conversation desensitized data that matches the simulated conversation intention through a large language model.

[0066] During the specific implementation process, the first prompt word can be generated according to the target real conversation desensitized data, for example: the target real conversation desensitized data is filled into the preset first prompt word template to generate the first prompt word, wherein the first prompt word template is configured with a slot for filling the target real conversation desensitized data. In addition, the first prompt word template is also configured with a part for describing the task of extracting simulated input information, which clarifies which dimensions of simulated input information need to be extracted and how to extract it. Based on this, after the first prompt word generated according to the target real conversation desensitized data is input into the large language model, it can guide the large language model to extract and output at least one dimension of simulated input information from the target real conversation desensitized data.

[0067] Optionally, generating a first prompt word according to the target real conversation desensitized data includes: generating a first prompt word for guiding the large language model to extract at least one dimension of simulated input information from the target real conversation desensitized data according to the target real conversation desensitized data; or generating first prompt words corresponding to at least one dimension according to the target real conversation desensitized data, wherein each first prompt word is used to guide the large language model to extract a corresponding dimension of simulated input information from the target real conversation desensitized data.

[0068] For ease of understanding, combined Figure 2 The above simulation input information acquisition process is explained. Figure 2 A schematic diagram of a scenario of a simulation input information acquisition process provided in an embodiment of the present application, such as Figure 2 As shown, it is assumed that it is necessary to obtain three dimensions of simulated input information from the target real conversation desensitized data, namely: user background (for example, information on personal characteristics such as account membership level, shopping preferences, shopping habits, etc.), user order problems, and user primary and secondary demands (i.e., main demands and secondary demands).

[0069] like Figure 2 As shown in the figure, in the first method of obtaining simulated input information, user background, user order problems, and user primary and secondary demands can be extracted from the target real conversation desensitized data through a first prompt word, that is, a first prompt word describes the extraction task of simulated input information with three dimensions. Based on the input first prompt word, the large language model can output simulated input information of three dimensions, namely user background, user order problems, and user primary and secondary demands, at one time.

[0070] like Figure 2 As shown, in the second method of obtaining simulated input information, three dimensions of simulated input information can be extracted from the target real conversation desensitized data through three different first prompt words, for example: the first prompt word (1) is used to extract user background from the target real conversation desensitized data, the first prompt word (2) is used to extract user order problems from the target real conversation desensitized data, and the first prompt word (3) is used to extract user primary and secondary demands from the target real conversation desensitized data. Afterwards, the large language model outputs user background based on the input first prompt word (1), outputs user order problems based on the input first prompt word (2), and outputs user primary and secondary demands based on the input first prompt word (3).

[0071] In the above two simulation input information acquisition methods, when the simulation input information to be acquired has more than one dimension, the method of generating first prompt words corresponding to at least one dimension can effectively avoid mutual interference between the extraction tasks of simulation input information of different dimensions, and improve the accuracy of the extracted simulation input information; in addition, the large language model can be called in parallel to execute simulation input information extraction tasks corresponding to multiple first prompt words, thereby improving the acquisition efficiency of at least one dimension of simulation input information.

[0072] 105. Generate a second prompt word according to the simulation input information of at least one dimension.

[0073] 106. Input the second prompt word into the simulated user robot so that the simulated user robot can conduct a simulated conversation with the customer service in a natural language.

[0074] Finally, the simulated input information of at least one dimension output by the large language model can be filled into the preset slot of the second prompt word template to generate the second prompt word. The second prompt word is used to guide the simulated user robot to simulate the real user and conduct a simulated conversation in natural language with the customer service to realize the simulated conversation intention.

[0075] In the embodiment of the present application, when the customer service wants to conduct a simulation conversation corresponding to a certain simulation conversation intention to achieve a certain training demand, first obtain the target real conversation data matching the simulation conversation intention, and desensitize the privacy information of the real user in the target real conversation data to obtain the target real conversation desensitized data; then, generate a first prompt word according to the target real conversation desensitized data, and input the first prompt word into the large language model, so that the large language model extracts at least one dimension of simulation input information from the target real conversation desensitized data according to the first prompt word; finally, generate a second prompt word according to at least one dimension of simulation input information, and input the second prompt word into the simulation user robot, so that the simulation user robot and the customer service conduct a natural language simulation conversation to achieve the simulation conversation intention. Since the simulation input information is extracted from the target real conversation desensitized data corresponding to the target real conversation data, it can be ensured that the simulation input information has the conversation characteristics reflected by the real conversation, so that when the simulation user robot uses the second prompt word generated by the simulation input information to conduct a simulation conversation with the customer service, the simulation user robot is closer to the real user and has a better conversation simulation effect. In addition, by automatically extracting simulated input information from the target real conversation desensitized data through a large language model and generating second prompt words for input into the simulated user robot, the efficiency of generating the second prompt words can be effectively improved, thereby facilitating the configuration of second prompt words corresponding to various simulated conversation intentions for the simulated user robot, covering various conversation scenarios and meeting diverse customer service training needs.

[0076] The above describes the overall process of the conversation simulation method provided in the embodiment of the present application. Next, the process of obtaining the target real conversation data that matches the simulation conversation intention is specifically described.

[0077] Figure 3 A flowchart of obtaining target real session data provided by an embodiment of the present application, such as Figure 3 As shown, the following steps may be included:

[0078] 301. According to the correlation between the real session intention corresponding to each real session data in the real session database and the simulated session intention, at least one real session intention having a correlation greater than a set correlation threshold is screened out from the real session intention corresponding to each real session data.

[0079] 302. According to the number of real sessions corresponding to at least one real session intention, determine target real session data matching the simulation session intention from the real session data corresponding to at least one real session intention.

[0080] In the embodiment of the present application, the real conversation data corresponding to the real conversation between the real user and the customer service will be stored in the real conversation database after the real conversation ends. In order to facilitate the subsequent screening of the target real conversation data matching the simulation conversation intention from the real conversation database, each real conversation data in the real conversation database can also be associated with the real conversation intention.

[0081] Optionally, the real conversation intent and the simulated conversation intent use the same description rules when describing, for example: the conversation scenario, user demands and other intent description information used to describe the conversation intent corresponds to the same classification method.

[0082] In the specific implementation process, the correlation between the real conversation intent and the simulated conversation intent corresponding to each real conversation data in the real conversation database can be determined by keyword matching, text similarity calculation, semantic similarity calculation, etc. Then, according to the relationship between the correlation and the set correlation threshold, at least one real conversation intent with a correlation greater than the set correlation threshold is screened out from the real conversation intent corresponding to each real conversation data.

[0083] Optionally, the above correlation threshold can be customized.

[0084] It is understandable that when there is only one real session intention with a correlation greater than a set correlation threshold, the target real session data matching the simulation session intention can be directly screened out from the real session data corresponding to the real session intention. When there are more than one real session intentions with a correlation greater than the set correlation threshold, the real session intentions with a correlation greater than the set correlation threshold can be further screened to narrow the acquisition scope of the target real session data, so as to determine the target real session data matching the simulation session intention from the real session data corresponding to the at least one real session intention.

[0085] As an optional way to determine the target real session data, the number of real sessions corresponding to at least one real session intention can be used as a screening condition, that is, based on the number of real sessions corresponding to at least one real session intention, the target real session data matching the simulated session intention is determined from the real session data corresponding to at least one real session intention.

[0086] It is easy to understand that the more real conversations a real conversation intent corresponds to, the more real users are generally concerned about the conversation scenarios and user demands corresponding to the real conversation intent. The real conversations corresponding to high-frequency real conversation intents can better simulate the conversation interactions of real users in actual applications.

[0087] In a specific implementation process, a visualization method can be combined to first display the sorting result of at least one real session intent sorted according to the corresponding real session quantity; then, in response to a confirmation operation of the target real session intent according to the sorting result, the target real session data matching the simulated session intent is determined from the real session data corresponding to the target real session intent. The target real session intent is one of the at least one real session intent.

[0088] Optionally, the number of real sessions corresponding to each real session intent can be counted from different time dimensions, such as one day, one week or one month. When sorting, you can customize the time dimension corresponding to the number of real sessions for reference, and sort according to the number of real sessions corresponding to at least one real session intent in the same time dimension.

[0089] For ease of understanding, combined Figure 4 An exemplary description is given. Figure 4 A schematic diagram of a real conversation intention provided in an embodiment of the present application, such as Figure 4 As shown in the figure, it is assumed that among the real conversation intentions corresponding to the real conversation data in the real conversation database, there are three real conversation intentions with correlations greater than the set correlation threshold, namely, real conversation intentions Figure 1 , real conversation meaning Figure 2 and real conversation Figure 3 Among them, the real conversation Figure 1 There are 7,000 real sessions in the past day, 50,000 real sessions in the past 7 days, and 180,000 real sessions in the past 30 days; Figure 2 There are 8,000 real sessions in the past day, 40,000 real sessions in the past 7 days, and 190,000 real sessions in the past 30 days; Figure 3 There are 9,000 real sessions in the past 1 day, 60,000 real sessions in the past 7 days, and 170,000 real sessions in the past 30 days.

[0090] If the real conversations corresponding to the real conversation intents in one day are sorted in descending order, the sorting results are as follows: Figure 4 As shown in the figure, they are the real conversation meanings. Figure 3 , real conversation meaning Figure 2 and real conversation Figure 1Based on the ranking results, if the real conversation is confirmed Figure 3 If the target is the real conversation intention, then Figure 3 Among the real session data corresponding to 9000 real sessions in the past day, target real session data matching the simulation session intention is determined.

[0091] In fact, the real conversation data corresponding to the target real conversation intent may include various types of real conversation data, such as: real conversation data of real users' problems being solved / unsolved, real conversation data of real users being satisfied / dissatisfied with customer service, real conversation data of real users with more / fewer conversation rounds with customer service, etc.

[0092] Among them, in the real conversation data where the real user's problem is not solved and / or the real user is dissatisfied with the customer service, there may be some information that is not conducive to the simulated user robot simulating the real user and the customer service to conduct a simulated conversation, such as: emotional conversation messages output by the real user or the customer service. Such information may interfere with the simulation effect of the simulated user robot and is not conducive to training the customer service to solve the real user's problems better and faster. In addition, if the real conversation data has a small number of conversation rounds, it may not be possible to obtain accurate simulation input information for the simulated user robot to simulate the real user.

[0093] In order to extract accurate simulated input information that will not interfere with the simulated user robot's simulation of a real user from the target real conversation desensitized data corresponding to the target real conversation data, in an optional embodiment, the target real conversation data that matches the simulated conversation intention can be determined from the real conversation data corresponding to the target real conversation intention according to preset conversation screening conditions.

[0094] Optionally, the conversation screening condition includes at least one of the following: the real user's problem is solved, the real user's satisfaction with the customer service is greater than a first set threshold, and the conversation rounds between the real user and the customer service are greater than a second set threshold.

[0095] In the embodiment of the present application, based on the correlation between the real conversation intention and the simulated conversation intention corresponding to each real conversation data, at least one real conversation intention whose correlation with the simulated conversation intention is greater than the set correlation threshold is screened out; based on the number of real conversations corresponding to at least one real conversation intention, a target real conversation intention with a large number of corresponding real conversations is screened out; based on the preset conversation screening conditions, the target real conversation data whose real user's problem is solved, the real user's satisfaction with the customer service is greater than the first set threshold, and the conversation rounds between the real user and the customer service are greater than the second set threshold are screened out from the real conversation data corresponding to the target real conversation intention. Through multiple screening methods, high-quality target real conversation data that is related to the simulated conversation intention, corresponds to the high-frequency real conversation intention, and meets the preset conversation screening conditions can be screened out from the real conversation database, which is conducive to extracting accurate and non-interfering simulation input information of the simulated user robot simulating the real user from the target real conversation desensitized data corresponding to the target real conversation data, so that the simulated user robot has a better conversation simulation effect when conducting a simulated conversation with the customer service.

[0096] In the above embodiment, it is mentioned that the target real session data can be obtained from the real session database. In actual applications, as new real sessions are continuously generated between real users and customer service, correspondingly, new real session data will be continuously stored in the real session database. It is understandable that the real session data newly stored in the real session database often has better timeliness, that is, it can better reflect the actual problems faced by real users at the moment and the solutions to the problems, and is more meaningful for customer service training.

[0097] For example, for the same user return and refund problem, the customer service's daily strategy for solving the return and refund problem may be different from the solution strategy adopted in the large-scale shopping festival scenario. If during the large-scale shopping festival, the real conversation data in the non-large-scale shopping festival scenario is still used as the target real conversation data to generate the second prompt word for input into the simulated user robot, the simulated user robot may not be able to simulate the interaction behavior of the real user in the large-scale shopping festival scenario, so the customer service and the simulated user robot have a simulated conversation, which cannot help the customer service practice to deal with user problems in the large-scale shopping festival scenario.

[0098] In order to ensure the timeliness of the target real conversation data, in an optional embodiment, in response to the update of the real conversation data contained in the real conversation database, the target real conversation data matching the simulation conversation intention is reacquired; then, the second prompt word is updated according to the reacquired target real conversation data, that is, the second prompt word is regenerated according to the reacquired target real conversation data, and if the newly generated second prompt word is different from the previous second prompt word, the old second prompt word is replaced with the newly generated second prompt word. The generation process of the second prompt word can refer to the aforementioned embodiment, which will not be repeated here.

[0099] Optionally, a mechanism for triggering the retrieval of target real session data matching the simulation session intent from the real session database may be customized. For example, based on a preset fixed time interval (e.g., 12 hours, one day, one week, etc.), the retrieval of target real session data matching the simulation session intent from the real session database may be automatically triggered; or, based on a manually triggered target real session data retrieval operation, the retrieval of target real session data matching the simulation session intent from the real session database may be triggered.

[0100] In the embodiment of the present application, in response to the update of the real conversation data contained in the real conversation database, by re-acquiring the target real conversation data that matches the simulation conversation intention, it is possible to ensure that the target real conversation data matches the actual problems and problem-solving strategies faced by the current real user, that is, to ensure the timeliness of the target real conversation data. Therefore, the second prompt word based on the updated target real conversation data that is re-acquired can also guide the simulation user robot to better simulate the real user.

[0101] In actual applications, one of the goals of the simulated conversation between customer service and the simulated user robot is to evaluate the customer service capability, that is, the customer service's ability to solve user problems.

[0102] Based on this, an embodiment of the present application provides another session simulation method. Figure 5 A flowchart of another session simulation method provided in an embodiment of the present application, such as Figure 5 As shown, the following steps may be included:

[0103] 501. During the simulated conversation between the customer service and the simulated user robot, the simulated solution corresponding to the simulated conversation is determined according to the system decision factors corresponding to the real solution in the target real conversation desensitized data and the manual decision factors determined by the customer service.

[0104] 502. Determine the customer service capabilities based on real solutions and simulated solutions.

[0105] To facilitate understanding, the process of determining a solution during a conversation between customer service and a user (including a real user and a simulated user robot) is first explained.

[0106] Generally, customer service determines the solution to a user's problem based on a standard operating procedure (SOP). The process of determining the solution is usually a multi-level decision-making process.

[0107] Specifically, SOP corresponds to a multi-level decision tree. Each node on the decision tree represents a judgment condition, each branch represents a possible decision path, and the final leaf node represents a specific solution. In the process of determining the solution to the user's problem, the system decision factor and the manual decision factor are used to make judgments on each decision node to finally output the solution corresponding to the user's problem.

[0108] The system decision factor is determined by a preset algorithm based on the user's behavior data. The user behavior data includes but is not limited to user history data (such as: user's past consultation records, complaint records, browsing records, purchase records, search records, purchase preferences, etc.), user order records (such as: current order status, logistics information, etc.) and other static or dynamic behavior data.

[0109] The manual decision factor is determined by the customer service who is talking to the user. In other words, the manual decision factor is determined by the customer service through observation and questioning of the user and through subjective judgment, including but not limited to the user's tone, emotion, specific demands, etc.

[0110] Based on the above introduction, in the process of determining the solution, the system decision factor is essentially determined objectively based on the behavioral data generated by the user, and is not affected by the subjective behavior of the customer service. Therefore, in the process of evaluating the service capabilities of the customer service through conversation simulation, the key is to evaluate whether the artificial decision factors determined in the process of the customer service and the simulated user robot are accurate and whether the correct solution can be obtained.

[0111] Since the system decision factor is indispensable for determining the solution, in order to more realistically simulate the process of having a conversation with a real user, in an embodiment of the present application, the target real conversation desensitized data also includes: a real solution, and a system decision factor corresponding to the real solution.

[0112] Therefore, during the simulated conversation between the customer service and the simulated user robot, the simulated solution corresponding to the simulated conversation can be determined based on the system decision factors corresponding to the real solution in the target real conversation desensitized data and the manual decision factors determined by the customer service. Furthermore, the service capability of the customer service can be determined based on the real solution and the simulated solution.

[0113] In an embodiment of the present application, the system decision factors corresponding to the real solutions in the target real conversation desensitized data are used together with the artificial decision factors determined by the customer service conducting the simulated conversation to jointly determine the simulated solution, thereby realizing the simulation of the process of obtaining the real solution during the real conversation. The artificial decision factors determined by the customer service during the simulated conversation and the artificial decision factors determined during the real conversation have the same influence on obtaining the final solution, thereby enabling a more accurate evaluation of the customer service's service capabilities.

[0114] In the above embodiment, the conversation simulation method provided by the embodiment of the present application is described by taking the simulation conversation of the user robot and the customer service in natural language as an example. However, in a real conversation, in addition to sending natural language conversation messages, the real user may also send image conversation messages, such as order screenshots, product damage images and other evidence images.

[0115] In traditional solutions, since the simulated user robot does not have the ability to process evidence images, it cannot send evidence images during the simulated conversation and may send some content that is inconsistent with the actual situation or completely fictitious. For example, when the customer service asks the simulated user robot to send a certain type of evidence image, the simulated user robot cannot provide the customer service with the evidence image on the one hand, and on the other hand, it will output fictitious information that is inconsistent with the actual situation, such as "Haven't I already sent the evidence image?"

[0116] In order to further improve the simulation user robot's simulation capability of real users, the present application provides another conversation simulation method. Figure 6 A flowchart of another session simulation method provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, the following steps may be included;

[0117] 601. During a simulated conversation between a customer service representative and a simulated user robot based on a simulated order in the target real conversation desensitized data, if the simulated conversation message contains a target keyword, a third prompt word generated based on a simulated evidence image in the target real conversation desensitized data is obtained, and the third prompt word is used to obtain a target simulated evidence image that matches the target keyword from the simulated evidence image.

[0118] The target real session data includes a real evidence image. By desensitizing the privacy information of the real user in the real evidence image, a simulated evidence image that does not contain the privacy information of the real user can be obtained.

[0119] Optionally, the target keyword can be customized according to the category of evidence image used in actual applications, for example, the target keyword can be: screenshot, photo, voucher, proof, evidence, etc.

[0120] Optionally, the simulated conversation message includes the target keyword, including: the simulated conversation message sent by the customer service and / or the simulated user robot includes the target keyword. For example, the following simulated conversation message "Please provide photos of the damaged goods, thank you!" sent by the customer service to the simulated user robot includes the target keyword "photo".

[0121] In practical applications, the third prompt word can be generated in real time when the simulation conversation message contains the target keyword, or can be generated and stored in advance before the simulation conversation. In the embodiment of the present application, the timing of generating the third prompt word is not limited.

[0122] The following combination Figure 7 The process of generating the third prompt word based on the simulated evidence image in the target real conversation desensitized data is described in detail.

[0123] Figure 7 A flowchart of a third prompt word generation method provided in an embodiment of the present application, such as Figure 7 As shown, the following steps may be included:

[0124] 701. Input the simulated evidence image in the target real conversation desensitized data into the image recognition model, so that the image recognition model outputs the evidence category corresponding to the simulated evidence image.

[0125] Optionally, the machine learning model configured in the simulated user robot also includes: an image recognition model, wherein the image recognition model includes but is not limited to a multimodal large model (Multimodal Large Mode) capable of processing data in various forms such as images, texts, and videos.

[0126] Optionally, a variety of different evidence categories can be pre-defined, such as: order screenshots, session record screenshots, shopping cart screenshots, photos of damaged goods, etc., and image training samples corresponding to each evidence category can be provided during the training phase of the image recognition model so that the image recognition model has corresponding classification capabilities. Alternatively, the image recognition model can be made to output the evidence category corresponding to the simulated evidence image by inputting prompt words into the image recognition model, wherein the prompt words input into the image recognition model can include the image features of the evidence image of each evidence category.

[0127] 702. Generate a fourth prompt word according to the evidence category and the real conversation record included in the target real conversation desensitized data, where the fourth prompt word is used to guide the large language model to extract description information corresponding to the simulated evidence image from the real conversation record according to the evidence category.

[0128] It is understandable that for a certain real evidence image, before and after the real user sends the real evidence image to the customer service, the conversation between the real user and the customer service will revolve around this real evidence image, that is, the context of this real evidence image will include the description information of the real evidence image. Among them, the description information of the real evidence image is also the description information of the corresponding simulated evidence image.

[0129] For example, suppose that the real conversation record between a real user and customer service includes the following conversation messages:

[0130] “Customer service: ‘Please provide a screenshot of the order for product xx’

[0131] Real user: '[Order screenshot of xx product]'

[0132] Customer service: "OK, we have received a screenshot of your order for product xx and will process a shipping refund for you right away."

[0133] It can be seen that the real session record contains information used to describe the order screenshots, such as "for xx product" and "for handling freight refunds". Among them, the order screenshot is also the image category of the evidence image.

[0134] Based on this, according to the evidence category corresponding to the simulated evidence image and the real conversation records contained in the target real conversation desensitized data, a fourth prompt word can be generated for extracting the descriptive information corresponding to the simulated evidence image from the real conversation records according to the evidence category, such as "The user sent an order screenshot to the customer service. Please describe the order screenshot in about 10 words based on the conversation records between the user and the customer service."

[0135] 703. Input the fourth prompt word into the large language model so that the large language model outputs description information corresponding to the simulated evidence image.

[0136] 704. Generate a third prompt word according to the description information corresponding to the simulated evidence image.

[0137] Under the guidance of the fourth prompt word, the large language model can output the description information corresponding to each simulated evidence image. After that, the image acquisition path (such as: image download address, etc.) corresponding to each simulated evidence image is associated with its corresponding description information, and the preset slot in the third prompt word template is filled to generate the third prompt word.

[0138] 602. Input a third prompt word into the simulated user robot, so that the simulated user robot sends a target simulated evidence image to the customer service.

[0139] Since the third prompt word contains the descriptive information corresponding to each simulated evidence image and the image acquisition path corresponding to the simulated evidence image, the simulated user robot can be guided to obtain the target simulated evidence image matching the target keyword from the simulated evidence image and send the target simulated evidence image to the customer service.

[0140] In the embodiment of the present application, the image recognition model is first used to identify the evidence category of the simulated evidence image in the target real conversation desensitized data, and then the large language model is used to extract the description information of the simulated evidence image according to the evidence category and the real conversation record contained in the target real conversation desensitized data, and then the third prompt word for obtaining the target simulated evidence image matching the target keyword from the simulated evidence image is generated based on the description information of the simulated evidence image. Under the guidance of the third prompt word, the simulated user robot can send the target simulated evidence image corresponding to the target keyword to the customer service when the target keyword is included in the simulated conversation message. In other words, in addition to being able to conduct a natural language simulated conversation with the customer service, the simulated user robot also has the ability to send the correct simulated evidence image to the customer service. By expanding the simulated evidence image sending capability of the simulated user robot, the authenticity of the simulated user robot simulating real users can be further improved, and the ability and service quality of the customer service staff to solve user problems can be improved.

[0141] The conversation simulation device of one or more embodiments of the present application will be described in detail below. Those skilled in the art will appreciate that these devices can be configured using commercially available hardware components through the steps taught in this solution.

[0142] Figure 8 A structural diagram of a conversation simulation device provided in an embodiment of the present application is shown in FIG. Figure 8 As shown, the device includes: an acquisition module 11, a processing module 12, a generation module 13 and a conversation module 14.

[0143] The acquisition module 11 is used to acquire target real session data that matches the simulation session intention, where the target real session data is session data generated by a real user having a real session with customer service.

[0144] The processing module 12 is used to perform desensitization processing on the privacy information of the real user in the target real session data to obtain the target real session desensitized data.

[0145] The generation module 13 is used to generate a first prompt word according to the target real conversation desensitized data; input the first prompt word into a large language model so that the large language model extracts at least one dimension of simulated input information from the target real conversation desensitized data according to the first prompt word; and generate a second prompt word according to the simulated input information.

[0146] The conversation module 14 is used to input the second prompt word into the simulated user robot so that the simulated user robot and the customer service can have a simulated conversation in natural language.

[0147] Optionally, the acquisition module 11 is specifically used to screen out at least one real session intention whose correlation is greater than a set correlation threshold from the real session intentions according to the correlation between the real session intention and the simulated session intention corresponding to each real session data in the real session database; and determine, according to the number of real sessions corresponding to the at least one real session intention, target real session data matching the simulated session intention from the real session data corresponding to the at least one real session intention.

[0148] Optionally, the acquisition module 11 is further specifically used to display a sorting result of the at least one real session intention sorted according to the corresponding number of real sessions; in response to a confirmation operation on the target real session intention according to the sorting result, determine target real session data matching the simulated session intention from the real session data corresponding to the target real session intention, and the target real session intention is one of the at least one real session intention.

[0149] Optionally, the acquisition module 11 is further specifically used to determine, according to a preset conversation screening condition, target real conversation data that matches the simulated conversation intention from the real conversation data corresponding to the target real conversation intention; wherein the conversation screening condition includes at least one of the following: the real user's problem is solved, the real user's satisfaction with the customer service is greater than a first set threshold, and the conversation rounds between the real user and the customer service are greater than a second set threshold.

[0150] Optionally, the generating module 13 is specifically configured to generate first prompt words corresponding to the at least one dimension according to the target real conversation desensitized data.

[0151] Optionally, the target real conversation desensitized data includes: a real solution, and a system decision factor corresponding to the real solution, wherein the system decision factor is determined by a preset algorithm based on the behavior data of the real user. The conversation module 14 is specifically used to determine the simulation solution corresponding to the simulation conversation according to the system decision factor and the manual decision factor determined by the customer service during the simulation conversation between the customer service and the simulated user robot; and determine the service capability of the customer service according to the real solution and the simulation solution.

[0152] Optionally, the processing module 12 is specifically used to generate a simulated order corresponding to the real order contained in the target real session data, and a simulated evidence image corresponding to the real evidence image, and the simulated order and the simulated evidence image do not contain the privacy information of the real user.

[0153] Optionally, the generation module 13 is also used to obtain a third prompt word generated according to the simulation evidence image if the simulation conversation message contains target keywords during the process of the customer service conducting a simulation conversation with the simulated user robot based on the simulated order, and the third prompt word is used to obtain a target simulation evidence image that matches the target keyword from the simulation evidence image.

[0154] Correspondingly, the conversation module 14 is further used to input the third prompt word into the simulated user robot, so that the simulated user robot sends the target simulated evidence image to the customer service.

[0155] Optionally, the generation module 13 is further specifically used to input the simulated evidence image into an image recognition model so that the image recognition model outputs an evidence category corresponding to the simulated evidence image; generate a fourth prompt word based on the evidence category and the real conversation record contained in the target real conversation desensitized data, and the fourth prompt word is used to guide the large language model to extract the description information corresponding to the simulated evidence image from the real conversation record according to the evidence category; input the fourth prompt word into the large language model so that the large language model outputs the description information; and generate the third prompt word based on the description information.

[0156] Optionally, the generating module 13 is further specifically configured to obtain a third prompt word generated according to the simulation evidence image if the simulation conversation message sent by the customer service contains a target keyword.

[0157] Optionally, the acquisition module 11 is further configured to reacquire target real conversation data matching the simulation conversation intention in response to an update of the real conversation data contained in the real conversation database; and update the second prompt word according to the reacquired target real conversation data.

[0158] Figure 8 The device shown can execute the steps introduced in the aforementioned embodiments. For detailed execution process and technical effects, please refer to the description in the aforementioned embodiments, which will not be repeated here.

[0159] In one possible design, the above Figure 8 The structure of the conversation simulation device shown can be implemented as an electronic device, such as Fig. 9 As shown, the electronic device may include: a memory 21, a processor 22, and a communication interface 23. The memory 21 stores executable codes, and when the executable codes are executed by the processor 22, the processor 22 can at least implement the session simulation method provided in the above embodiments.

[0160] In addition, an embodiment of the present application provides a non-temporary machine-readable storage medium, on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor can at least implement the session simulation method provided in the aforementioned embodiment.

[0161] An embodiment of the present application provides a computer program product, including: a computer program, when the computer program is executed by a processor of an electronic device, the processor is caused to execute the session simulation method provided in the aforementioned embodiment.

[0162] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. Those of ordinary skill in the art may understand and implement the present invention without creative effort.

[0163] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by adding a necessary general hardware platform, and of course can also be implemented by combining hardware and software. Based on such an understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a computer product, and the present application can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A conversation simulation method, characterized in that: include: Acquire target real session data that matches the simulation session intention, where the target real session data is session data generated by a real user having a real session with customer service; Desensitizing the privacy information of the real user in the target real session data to obtain target real session desensitized data; Generate a first prompt word according to the target real conversation desensitized data; Inputting the first prompt word into a large language model, so that the large language model extracts at least one dimension of simulated input information from the target real conversation desensitized data according to the first prompt word; generating a second prompt word according to the simulation input information; The second prompt word is input into the simulated user robot so that the simulated user robot and the customer service have a simulated conversation in natural language.

2. The method according to claim 1, characterized in that The step of obtaining target real session data matching the simulation session intention includes: According to the correlation between the real session intention and the simulated session intention corresponding to each real session data in the real session database, at least one real session intention whose correlation is greater than a set correlation threshold is screened out from the real session intentions; According to the number of real sessions respectively corresponding to the at least one real session intention, target real session data matching the simulation session intention is determined from the real session data respectively corresponding to the at least one real session intention.

3. The method according to claim 2, characterized in that The determining, according to the number of real sessions respectively corresponding to the at least one real session intention, target real session data matching the simulation session intention from the real session data respectively corresponding to the at least one real session intention, comprises: Displaying a sorting result of the at least one real session intention sorted according to the corresponding number of real sessions; In response to a confirmation operation on a target real session intention according to the sorting result, target real session data matching the simulated session intention is determined from the real session data corresponding to the target real session intention, and the target real session intention is one of the at least one real session intention.

4. The method according to claim 3, characterized in that The step of determining target real session data matching the simulation session intention from the real session data corresponding to the target real session intention includes: According to preset conversation screening conditions, target real conversation data matching the simulated conversation intention is determined from the real conversation data corresponding to the target real conversation intention; wherein the conversation screening conditions include at least one of the following: the real user's problem is solved, the real user's satisfaction with the customer service is greater than a first set threshold, and the conversation rounds between the real user and the customer service are greater than a second set threshold.

5. The method according to claim 1, characterized in that The step of generating a first prompt word according to the target real conversation desensitized data includes: According to the target real conversation desensitized data, a first prompt word corresponding to the at least one dimension is generated.

6. The method according to any one of claims 1 to 5, characterized in that The target real conversation desensitized data includes: a real solution, and a system decision factor corresponding to the real solution, wherein the system decision factor is determined by a preset algorithm based on the behavior data of the real user, and the method further includes: During the simulation conversation between the customer service and the simulated user robot, determining a simulation solution corresponding to the simulation conversation according to the system decision factor and the manual decision factor determined by the customer service; The service capability of the customer service is determined according to the real solution and the simulation solution.

7. The method according to claim 1, characterized in that The desensitizing the privacy information of the real user in the target real session data includes: Generate a simulated order corresponding to the real order included in the target real session data, and a simulated evidence image corresponding to the real evidence image, wherein the simulated order and the simulated evidence image do not contain the privacy information of the real user.

8. The method according to claim 7, characterized in that The method further comprises: During the process of the customer service conducting a simulated conversation with the simulated user robot based on the simulated order, if the simulated conversation message contains a target keyword, a third prompt word generated according to the simulated evidence image is obtained, and the third prompt word is used to obtain a target simulated evidence image matching the target keyword from the simulated evidence image; The third prompt word is input into the simulated user robot so that the simulated user robot sends the target simulated evidence image to the customer service.

9. The method according to claim 8, characterized in that The step of generating a third prompt word according to the simulated evidence image includes: Inputting the simulated evidence image into an image recognition model so that the image recognition model outputs the evidence category corresponding to the simulated evidence image; Generate a fourth prompt word according to the evidence category and the real conversation record included in the target real conversation desensitized data, wherein the fourth prompt word is used to guide the large language model to extract description information corresponding to the simulated evidence image from the real conversation record according to the evidence category; Inputting the fourth prompt word into the large language model so that the large language model outputs the description information; The third prompt word is generated according to the description information.

10. The method according to claim 8, characterized in that If the simulation conversation message contains the target keyword, obtaining a third prompt word generated according to the simulation evidence image includes: If the simulated conversation message sent by the customer service contains the target keyword, a third prompt word generated according to the simulated evidence image is obtained.

11. The method according to claim 1, characterized in that: The method further comprises: In response to an update of the real session data contained in the real session database, reacquiring target real session data matching the simulation session intention; The second prompt word is updated according to the reacquired target real conversation data.

12. A computing device, characterized in that: It comprises a storage component and a processing component; the storage component stores one or more computer program instructions, the computer program instructions are called and executed by the processing component, and the processing component executes the one or more computer program instructions to implement the session simulation method as described in any one of claims 1-11.

13. A computer-readable storage medium, characterized in that: A computer program is stored, and the computer program is executed by a computer to implement the conversation simulation method according to any one of claims 1 to 11.

14. A computer program product, characterized in that A computer program is stored, and when the computer program is executed by a computer, the conversation simulation method according to any one of claims 1 to 11 is implemented.

Images