Outbound call method and device based on large model

Abstract

The invention discloses an outbound call method and device based on a large model. The method comprises the following steps: determining call information currently obtained by an outbound call system; based on the current user inquiry information, determining a corresponding user prompt word; determining a corresponding system prompt word based on the description text of the preset outbound task and the standard operation flow; determining a corresponding reasoning requirement based on the historical dialogue information; based on the user cue word, the system cue word and the reasoning requirement, taking the user cue word, the system cue word and the reasoning requirement as input of a pre-deployed large model to obtain a response text and pre-estimation information output by the large model; and controlling the outbound robot to output an outbound voice corresponding to the response text, and updating the system prompt word according to the target step. According to the method, the response text output by the large model is used as a reference for the outbound robot to output the outbound voice, and the standard operation process in the system cue word can be updated by means of the prediction information of the large model, so that the response of the outbound robot has diversity and flexibility.

Description

Technical Field

[0001] The present application relates to the field of outbound call technology, and in particular to an outbound call method and device based on a large model. Background Art

[0002] In the field of outbound call technology, the traditional outbound call system architecture can be found in Figure 1 As shown in the figure, the user's voice is first recognized into user text by ASR (Automatic Speech Recognition), and then the user text passes through various intermediate processing modules, such as the semantic understanding module, the process canvas module and other processing modules, and then the final reply text is generated by the text generation module. Finally, the reply text is synthesized into speech by TTS (Text To Speech) and played to the user. Among them, the semantic understanding module is used to perform semantic analysis on the user text. Semantic analysis involves intent recognition and slot extraction. The process canvas module contains a node tree composed of rules, which is used to control the direction of the dialogue process to achieve accurate human-computer dialogue process control.

[0003] However, the outbound call robot replies (which can be understood as response texts) in traditional outbound call systems lack diversity and flexibility, and are unable to cope with users' tricky topics, which hinders the human-computer dialogue process and affects user experience. Summary of the invention

[0004] The present application provides an outbound call method and device based on a large model, the purpose of which is to improve the diversity and flexibility of outbound robot answers in an outbound call system.

[0005] In order to achieve the above objectives, this application provides the following technical solutions:

[0006] An outbound calling method based on a large model, comprising:

[0007] Determine the call information currently obtained by the outbound call system; the call information includes the current user inquiry information and the historical conversation information between the user and the outbound call robot;

[0008] Based on the current user query information, determining a corresponding user prompt word;

[0009] Determine the corresponding system prompt words based on the description text of the preset outbound call task and the standard operating process; the standard operating process includes multiple steps and corresponding detailed description of the process;

[0010] Determining corresponding reasoning requirements based on the historical conversation information;

[0011] Based on the user prompt, the system prompt, and the inference requirements, as the input to the pre-deployed large model to obtain the response text and estimation information output by the large model; the estimation information includes the target step that matches the future user inquiry information;

[0012] Control the outbound robot to output the outbound voice corresponding to the response text, and update the system prompt according to the target step.

[0013] Optionally, updating the system prompt according to the target step includes:

[0014] Based on the historical conversation information, determine the cumulative number of conversations between the user and the outbound robot;

[0015] If the cumulative number of conversations is 0, determine the steps between the first step and the target step in the standard operation process as the steps to be updated; the first step is the step with the first execution order in the standard operation process;

[0016] Update the process detailed description corresponding to the step to be updated in the system prompt.

[0017] Optionally, the method further includes:

[0018] If the cumulative number of conversations is not 0, determine the second step that matches the current user inquiry information;

[0019] Determine the steps between the second step and the target step in the standard operation process as the steps to be updated.

[0020] Optionally, updating the process detailed description corresponding to the step to be updated in the system prompt includes:

[0021] Based on the step to be updated, determine the corresponding first feature vector;

[0022] Determine the second feature vectors corresponding to each process template in the process template library;

[0023] Calculate the similarity between the first feature vector and multiple second feature vectors respectively;

[0024] Based on the process template corresponding to the second feature vector with the highest similarity, determine it as the target process template;

[0025] Use the target process template to replace the process detailed description corresponding to the step to be updated in the system prompt.

[0026] Optionally, after calculating the similarities between the first feature vector and the multiple second feature vectors respectively, the method further includes:

[0027] If the highest similarity among the multiple similarities does not meet the preset threshold, determine the corresponding template reference file based on each process template in the process template library;

[0028] Use the template reference file as the input attachment of the large model, and input the corresponding template reset command to the large model to obtain each new process template output by the large model;

[0029] Use each of the new process templates to replace each process template in the process template library.

[0030] Optionally, the method further includes:

[0031] If the estimated information output by the large model is empty, input the preset retention strategy information into the large model to obtain the user retention text output by the large model;

[0032] Control the outbound robot to output the outbound voice corresponding to the user retention text.

[0033] Optionally, the method further includes:

[0034] If the response text output by the large model contains a specified keyword, notify the human customer service to provide services to the user instead of the outbound robot, and synchronously forward the call information to the human customer service.

[0035] An outbound call device based on a large model, including:

[0036] A call information determination unit, configured to determine the call information currently obtained by the outbound system; the call information includes the current user inquiry information and the historical conversation information between the user and the outbound robot;

[0037] A user prompt determination unit, configured to determine the corresponding user prompt word based on the current user inquiry information;

[0038] A system prompt determination unit, configured to determine the corresponding system prompt word based on the description text of the preset outbound task and the standard operation process; the standard operation process includes multiple steps and corresponding detailed process descriptions;

[0039] An inference requirement determination unit, configured to determine the corresponding inference requirement based on the historical conversation information;

[0040] The large model calling unit is used to obtain the response text and estimation information output by the large model by using the user prompt, the system prompt, and the inference requirement as the input of a pre-deployed large model; the estimation information includes the target step matching the future user inquiry information.

[0041] The control and update unit is used to control the outbound robot to output the outbound voice corresponding to the response text, and update the system prompt according to the target step.

[0042] A storage medium, the storage medium includes a stored program, wherein the program, when run by a processor, executes an outbound call method based on a large model.

[0043] An electronic device, comprising: a processor, a memory, and a bus; the processor is connected to the memory through the bus;

[0044] The memory is used to store a program, and the processor is used to run the program, wherein the program, when run by the processor, executes an outbound call method based on a large model.

[0045] The technical solution provided by this application determines the call information currently obtained by the outbound system. Based on the current user inquiry information, the corresponding user prompt is determined. Based on the description text of the preset outbound task and the standard operation process, the corresponding system prompt is determined. Based on the historical conversation information, the corresponding inference requirement is determined. Using the user prompt, the system prompt, and the inference requirement as the input of a pre-deployed large model to obtain the response text and estimation information output by the large model. Control the outbound robot to output the outbound voice corresponding to the response text, and update the system prompt according to the target step. This application uses the response text output by the large model as a reference for the outbound robot to output the outbound voice, and can update the standard operation process in the system prompt with the help of the prediction information of the large model, so that the response of the outbound robot is diverse and flexible. Description of the Drawings

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0047] Figure 1 It is a schematic diagram of the architecture of a traditional outbound system provided by an embodiment of the present application;

[0048] Figure 2A flowchart of an outbound call method based on a large model provided by an embodiment of the present application;

[0049] Figure 3 A flowchart of another outbound call method based on a large model provided by an embodiment of the present application;

[0050] Figure 4 A flowchart of another outbound call method based on a large model provided by an embodiment of the present application;

[0051] Figure 5 A flowchart of another outbound call method based on a large model provided by an embodiment of the present application;

[0052] Figure 6 An architecture diagram of an outbound call system based on a large model provided by an embodiment of the present application;

[0053] Figure 7 An architecture diagram of an outbound call device based on a large model provided by an embodiment of the present application. Detailed implementation manners

[0054] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0055] In the present application, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. The terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0056] As Figure 2 shown, it is a flowchart of an outbound call method based on a large model provided by an embodiment of the present application, including the following steps.

[0057] S201: Determine the call information currently obtained by the outbound call system.

[0058] Among them, the call information includes the current user's inquiry information and the historical conversation information between the user and the outbound robot.

[0059] It should be noted that the types of the current user's inquiry information and the historical conversation information are text.

[0060] In some examples, after the outbound system receives the user voice currently sent by the user terminal, it uses ASR to convert the user voice into the current user's inquiry information.

[0061] In some examples, the outbound system will save the recordings of each conversation between the user and the outbound robot, and the recordings can be converted into historical conversation information by using ASR. In addition, the outbound system will also save the text information of each conversation between the user and the outbound robot, so the historical conversation information can be directly read from the cache space of the outbound system.

[0062] In some examples, if the user's language is a foreign language or a dialect, the corresponding language recognition model can be used for translation to obtain call information that conforms to the local standard language.

[0063] In a possible implementation manner, assume that the outbound system is used to recommend the membership application of a freight APP for freight drivers. The freight APP is used to provide a platform for freight drivers to undertake freight orders. The text format of the current user's inquiry information can be "Driver: What's the use of the membership?"

[0064] In a possible implementation manner, assume that the outbound system is used to recommend the membership application of a freight APP for freight drivers. The historical conversation information can be "Driver: What's the use of the membership? AI: The membership can enhance your rights and interests. It is recommended that you apply for a VIP membership."

[0065] S202: Determine the corresponding user prompt based on the current user's inquiry information.

[0066] Among them, the content that affects the output of the large model can be called a prompt, and the English translation of the user prompt is user prompt. The user prompt is the question, request or command input to the large model, aiming to obtain the specific response of the large model or complete the specified task, and is the direct reason for triggering the large model to generate a reply or execute an operation.

[0067] In some examples, the user prompt is the direct way for the user to interact with the large model, guiding the direction of the conversation by asking questions, making requests or giving commands to ensure that the large model can understand and generate appropriate replies. Generally speaking, the content of the user prompt allows for diversification and can be in the forms of questions, statements, commands, etc., but the content is required to be clear and understandable so that the large model can accurately understand and make appropriate responses.

[0068] It should be noted that in order to ensure that the content of the user prompt can be accurately understood by the large model, after obtaining the current user's inquiry information, semantic analysis can also be performed on the current user's inquiry information to determine the semantic intention of the current user's inquiry information. Based on this semantic intention, the corresponding user prompt can be determined, which can improve the semantic clarity of the user prompt.

[0069] In some examples, assume that the current user's inquiry information is "What's the deal with the membership?" By performing semantic analysis on this current user's inquiry information, the semantic intention can be determined as "What are the functions of the membership?" Therefore, based on "What are the functions of the membership?", it can be determined as the user prompt.

[0070] S203: Determine the corresponding system prompt based on the description text of the preset outbound task and the standard operation procedure.

[0071] Among them, the standard operation procedure includes multiple steps and corresponding detailed process descriptions.

[0072] In some examples, the description text of the preset outbound task is used to accurately describe the purpose and requirements of the preset outbound task.

[0073] In a possible implementation, assume that the preset outbound task is to recommend the membership application of a freight APP to freight drivers. Then the description text of the preset outbound task can be "Role: You are an invitation expert, and the task is to let the driver purchase the membership through phone outbound calls."

[0074] In some examples, the English translation of the standard operation procedure is Standard Operating Procedure, which can be abbreviated as SOP. It refers to describing the standard operation steps and requirements of the preset outbound task in a unified format, which is used to guide and standardize the output of the large model.

[0075] In a possible implementation, assume that the preset outbound task is to recommend the membership application of a freight APP to freight drivers. The standard operation procedure for this preset outbound task can be "First step: Politely start the conversation; Second step: Ask the driver if they have downloaded the freight APP. If not, gradually guide the driver to download the freight APP; Third step: Help the driver set up the order listening configuration; Fourth step: Guide the driver to operate on the APP and gradually purchase the membership." Correspondingly, there are corresponding detailed process descriptions for each of the four steps shown in the standard operation procedure.

[0076] In a possible implementation, any step in the standard operation process can also add corresponding optional attributes according to the actual situation. That is, the steps with optional attributes are not necessarily steps that must be executed. For example, "Step 2: Ask the driver if they have downloaded the freight APP. If not, guide the driver to download the freight APP step by step; Step 3: Help the driver set the order listening configuration" can be set as optional steps. Then the standard operation process can be set as "Step 1: Politely start the conversation; Step 2 (optional): Ask the driver if they have downloaded the freight APP. If not, guide the driver to download the freight APP step by step; Step 3 (optional): Help the driver set the order listening configuration; Step 4: Guide the driver to operate on the APP and gradually purchase a membership."

[0077] It should be noted that the English translation of the system prompt is "system prompt". The system prompt refers to a set of initial instructions or background information provided to the large model, which is used to guide the behavior mode and response mode of the large model, and can help set the role, tone, knowledge scope, etc. of the large model to ensure that the large model can interact with users in the expected manner.

[0078] In some examples, the main role of the system prompt is to set the basic behavior guidelines, personality characteristics, and ability scope of the large model, similar to an instruction manual or operation manual for the large model. It is usually set before the start of the human-machine conversation and is effective throughout the human-machine conversation process, but it does not directly interact with users.

[0079] In a possible implementation, the content of the system prompt can be specific, such as formulating rules that the large model should follow (such as standard operation processes, etc.), or it can be more general, such as defining the personality traits of the large model.

[0080] S204: Based on the historical conversation information, determine the corresponding inference requirements.

[0081] Among them, adding inference requirements to the prompt of the large model can allow the large model to spend more time thinking and the output results to be more reasonable.

[0082] In some examples, the historical conversation information includes the conversation between the user and the outbound robot. To improve the semantic accuracy of the inference requirements, the historical conversation information can be semantically analyzed to obtain conversation information with clear intentions, and then based on this conversation information, the corresponding inference requirements can be determined.

[0083] S205: Based on the user prompt, the system prompt, and the inference requirements, as the input of the pre-deployed large model, to obtain the response text and estimated information output by the large model.

[0084] Among them, the estimated information includes the target steps that match the future user inquiry information.

[0085] It should be noted that a large model refers to a machine learning model with a large number of parameters and a complex computing structure. It usually uses a vast amount of data for pre-training to obtain general capabilities and can be fine-tuned to adapt to preset outbound tasks through deep learning models.

[0086] In some examples, the large model is fully called the large-scale pre-trained model. The large model generally adopts the Transformer architecture, and the core of the Transformer architecture is the self-attention mechanism, which can dynamically measure the importance of each word in the input sequence and capture long-range dependencies.

[0087] In possible implementation manners, the deployment process of the large model is a technical means familiar to those skilled in the art and will not be elaborated here.

[0088] It should be noted that based on the user prompt words, system prompt words, and inference requirements, as the input of the pre-deployed large model, the large model can output a response text that conforms to the standard operation process and can answer the questions shown in the current user inquiry information. Based on the learning content of the large model itself, the response text output by the large model has high diversity and flexibility.

[0089] It can be understood that inputting the inference requirements into the large model can enable the large model to think for a long time, so that the large model can predict future user inquiry information and the target steps that can answer future user inquiry information.

[0090] In possible implementation manners, based on the user prompt words, system prompt words, and inference requirements, as the input of the pre-deployed large model, the output result format of the large model can be {response text}{next node: the fourth step}, and this next node represents the target step predicted by the large model.

[0091] It should be emphasized that the detailed description of the process corresponding to multiple steps is used as the system prompt words and input into the large model. Compared with the node tree mentioned in the process canvas module in the prior art, there are obvious differences. The large model can output more flexible response texts based on this system prompt word, effectively improving the flexibility of human-computer dialogue.

[0092] S206: Control the outbound robot to output the outbound voice corresponding to the response text, and update the system prompt words according to the target step.

[0093] Among them, after obtaining the response text output by the large model, the outbound system can perform voice synthesis on the response text through TTS to obtain the outbound voice corresponding to the response text, and then control the outbound robot to output the outbound voice corresponding to the response text to complete the response operation to the current user inquiry.

[0094] It should be noted that according to the target steps, the system prompt words are updated so that in the next human-machine dialogue process, the large model can output more diverse response texts based on the updated system prompt words, thereby effectively improving the diversity of human-machine dialogue.

[0095] Optionally, for the implementation process of updating the system prompt words according to the target steps, reference can be made to Figure 3 the steps shown and the corresponding explanatory notes.

[0096] In some examples, the estimated information output by the large model is empty, indicating that the large model predicts that the user has no interest in communicating with the outbound robot. It can be considered that communicating with the user according to the standard operation process can no longer arouse the user's interest. Therefore, the estimated information output by the large model is empty, that is, there is no target step in the standard operation process that matches the future user's inquiry information. Based on this prediction result, emergency retention measures need to be taken to regain the user's interest in communicating again.

[0097] Optionally, if the estimated information output by the large model is empty, the preset retention strategy information is input into the large model to obtain the user retention text output by the large model, and the outbound robot is controlled to output the outbound voice corresponding to the user retention text.

[0098] It can be understood that by inputting the retention strategy information into the large model to obtain the appropriate user retention text output by the large model, and controlling the outbound robot to output the outbound voice corresponding to the user retention text, the user's interest in communicating again can be effectively regained.

[0099] In some examples, based on the input current user inquiry information and historical dialogue information, the large model infers that the user may need a more emotional communication service provided by the artificial customer service. Therefore, the response text output by the large model will contain specified keywords related to this artificial communication scenario, such as "artificial customer service" and "emotional service", that is, the large model determines that providing artificial service can improve the user's dialogue experience and arouse the user's interest in communication.

[0100] Optionally, if the response text output by the large model contains specified keywords, notify the artificial customer service to provide services for the user instead of the outbound robot, and synchronously forward the call information to the artificial customer service.

[0101] It can be understood that the large model can timely understand the user's needs based on the current user inquiry information and historical dialogue information, providing a favorable hint for the scenario where the artificial customer intervenes in the call.

[0102] The process shown in S201 - S206 above uses the current user's inquiry information, historical conversation information, the description text of the preset outbound task, and the standard operation process as the input to the large - model, and uses the response text output by the large - model as a reference for the outbound robot to output outbound voice. Moreover, it can update the standard operation process in the system prompt words with the prediction information of the large - model, so that the responses of the outbound robot are diverse and flexible.

[0103] As Figure 3 shown, it is a schematic flow chart of another outbound call method based on a large - model provided by an embodiment of the present application, including the following steps.

[0104] S301: Based on the historical conversation information, determine the cumulative number of conversations between the user and the outbound robot.

[0105] Among them, each conversation between a user and the outbound robot in the historical conversation information can be determined as corresponding to one counting unit.

[0106] In a possible implementation, if the historical conversation information is "Driver: What's the use of a membership? AI: A membership can enhance your rights and interests. It is recommended that you apply for a VIP membership", then the cumulative number of conversations can be determined to be 1.

[0107] In a possible implementation, if the historical conversation information is empty, then the cumulative number of conversations can be determined to be 0.

[0108] S302: Determine whether the cumulative number of conversations is 0.

[0109] Among them, if the cumulative number of conversations is 0, then execute S303; if the cumulative number of conversations is not 0, then execute S304.

[0110] S303: Determine the steps between the first step and the target step in the standard operation process as the steps to be updated.

[0111] After executing S303, continue to execute S306.

[0112] Among them, the first step is the step with the first execution order in the standard operation process.

[0113] It should be noted that when the cumulative number of conversations is 0, it can be determined that the current user's inquiry is the first time. Generally speaking, the large - model will determine the corresponding response text according to the first step shown in the standard operation process, and the target step shown in the prediction information will be other steps except the first step in the standard operation process.

[0114] In some examples, when the cumulative number of conversations is 0 and the target step output by the large model is the fourth step in the standard operation process, the second and third steps in the standard operation process can be determined as the steps to be updated.

[0115] It should be noted that when the cumulative number of conversations is 0, if the difference in the execution order between the target step output by the large model and the first step is 1, it can be determined that the steps to be updated are empty.

[0116] In a possible implementation, when the cumulative number of conversations is 0, the target step output by the large model is the second step in the standard operation process, and there are no other steps between the first step and the target step, so it is determined that the steps to be updated are empty.

[0117] S304: Determine the second step that matches the current user's inquiry information.

[0118] After executing S304, continue to execute S305.

[0119] Among them, there are corresponding historical output records for the historical outputs of the large model. The second step that matches the current user's inquiry information can be determined by searching for this historical output record.

[0120] S305: Determine the steps in the standard operation process whose execution order is between the second step and the target step as the steps to be updated.

[0121] After executing S305, continue to execute S306.

[0122] Among them, when the cumulative number of conversations is not 0, it can be determined that the current user's inquiry is not the first time.

[0123] In some examples, when the cumulative number of conversations is not 0, the second step is the second step in the standard operation process, and the target step output by the large model is the fourth step in the standard operation process. Then, the third step in the standard operation process can be determined as the step to be updated.

[0124] It should be noted that when the cumulative number of conversations is not 0, if the difference in the execution order between the target step output by the large model and the second step is 1, it can be determined that the steps to be updated are empty.

[0125] In a possible implementation, when the cumulative number of conversations is 0, the second step is the third step in the standard operation process, the target step output by the large model is the fourth step in the standard operation process, and there are no other steps between the second step and the target step, so it is determined that the steps to be updated are empty.

[0126] S306: Update the detailed process description corresponding to the steps to be updated in the system prompt words.

[0127] Among them, after determining the step to be updated, update the detailed process description corresponding to the step to be updated in the system prompt words, so that the large model can provide appropriate response words for subsequent user inquiries according to the updated system prompt words. As the man-machine conversation continues, by updating the detailed process description corresponding to the step to be updated in the system prompt words, the response text output by the large model becomes more diverse and flexible, maintaining a high level of conversion effect for the man-machine conversation service.

[0128] Optionally, for the implementation process of updating the detailed process description corresponding to the step to be updated in the system prompt words, reference can be made to Figure 4 the steps shown and the corresponding explanatory notes.

[0129] For the process shown in S301-S306 above, by accumulating the number of conversations, the system prompt words can be continuously updated to make the response text output by the large model more diverse and flexible, thereby improving the response diversity and flexibility of the outbound robot.

[0130] As Figure 4 shown, it is a schematic flowchart of another outbound call method based on a large model provided by an embodiment of the present application, including the following steps.

[0131] S401: Based on the step to be updated, determine the corresponding first feature vector.

[0132] Among them, the corresponding first feature vector can be obtained by performing feature extraction on the text description of the step to be updated.

[0133] In a possible implementation manner, the implementation process of performing feature extraction on the text description belongs to conventional technical means and will not be elaborated here.

[0134] S402: Determine the second feature vectors corresponding to each process template in the process template library.

[0135] Among them, the process template library is a database created by the outbound system in advance for preset outbound tasks, and each process template in the process template library is a speech template set for each scenario in the outbound conversation process.

[0136] It can be understood that the second feature vectors corresponding to each process template can be obtained by performing feature extraction on the text description of each process template in the process template library.

[0137] S403: Calculate the similarity between the first feature vector and multiple second feature vectors respectively.

[0138] Among them, the similarity calculation method between two feature vectors includes but is not limited to the cosine similarity algorithm.

[0139] S404: Determine the process template corresponding to the second feature vector with the highest similarity as the target process template.

[0140] Among them, determining the process template corresponding to the second feature vector with the highest similarity as the target process template can ensure that the target process template can accurately fit the execution process of the step to be updated, and the words shown in the target process template can handle the human-machine dialogue scenario shown in the step to be updated.

[0141] It should be noted that if each process template in the process template library cannot fully handle the human-machine dialogue scenario shown in the step to be updated, the process template library can also be updated using a large model.

[0142] Optionally, for the implementation process of updating the process template library using a large model, reference can be made to Figure 5 the steps shown and the corresponding explanations.

[0143] S405: Use the target process template to replace the process detailed description corresponding to the step to be updated in the system prompt word.

[0144] Among them, after determining the target process template, the process detailed description corresponding to the step to be updated in the system prompt word can be deleted first, and then the target process template can be loaded into the system prompt word so that the target process template becomes the new process detailed description corresponding to the step to be updated.

[0145] For the process shown in S401 - S405 above, the process detailed description corresponding to the step to be updated in the system prompt word can be updated using the process templates in the process template library to improve the diversity and flexibility of the system prompt word, thereby improving the diversity and flexibility of the response text output by the large model.

[0146] As Figure 5 shown, it is a schematic flowchart of another outbound call method based on a large model provided by an embodiment of the present application, including the following steps.

[0147] S501: If the highest similarity among multiple similarities does not meet the preset threshold, determine the corresponding template reference file based on each process template in the process template library.

[0148] Among them, if the highest similarity among multiple similarities does not meet the preset threshold, it can be determined that each process template in the process template library cannot fully handle the human-machine dialogue scenario shown in the step to be updated. Therefore, it is necessary to reset each process template in the process template library.

[0149] S502: Use the template reference file as the input attachment of the large model and input the corresponding template reset command to the large model to obtain each new process template output by the large model.

[0150] Among them, using each process template in the process template library as an input attachment to the large model can ensure that the large model outputs a suitable new process template.

[0151] In some examples, taking the template reference file as an input attachment to the large model, the template reset command "reset and generate the same number of new process templates based on the attachment as a reference" can be input to the large model.

[0152] S503: Use each new process template to replace each process template in the process template library.

[0153] Among them, using each new process template to replace each process template in the process template library can achieve the rapid update of the process template library.

[0154] In some examples, using the processes shown in S501 - S503, with the help of the large model, the outbound system can achieve the reset and update of the process template. Correspondingly, for the architecture implementation of the outbound system based on the large model, reference can be made to Figure 6 the structural schematic diagram shown.

[0155] The processes shown in the above S501 - S503 can use the large model to generate new process templates to achieve the reset and update of the process templates in the process template library.

[0156] Such as Figure 7 shown, which is a structural schematic diagram of an outbound call device based on a large model provided by an embodiment of the present application, including the following units.

[0157] The call information determination unit 100 is used to determine the call information currently obtained by the outbound system; the call information includes the current user inquiry information and the historical conversation information between the user and the outbound robot.

[0158] The user prompt determination unit 200 is used to determine the corresponding user prompt word based on the current user inquiry information.

[0159] The system prompt determination unit 300 is used to determine the corresponding system prompt word based on the description text of the preset outbound task and the standard operation process; the standard operation process includes multiple steps and corresponding detailed process descriptions.

[0160] The inference requirement determination unit 400 is used to determine the corresponding inference requirement based on the historical conversation information.

[0161] The large model call unit 500 is used to use the user prompt word, the system prompt word, and the inference requirement as the input to the pre - deployed large model to obtain the response text and the prediction information output by the large model; the prediction information includes the target steps matching the future user inquiry information.

[0162] The control and update unit 600 is used to control the outbound voice corresponding to the response text output by the outbound robot and update the system prompt words according to the target step.

[0163] Optionally, the control and update unit 600 is specifically used for: determining the cumulative number of conversations between the user and the outbound robot based on the historical conversation information; if the cumulative number of conversations is 0, determining the steps between the first step and the target step in the standard operation process as the steps to be updated; the first step is the step with the first execution order in the standard operation process; updating the process detailed description corresponding to the step to be updated in the system prompt words.

[0164] Optionally, the control and update unit 600 is further used for: if the cumulative number of conversations is not 0, determining the second step that matches the current user inquiry information; determining the steps between the second step and the target step in the standard operation process as the steps to be updated.

[0165] Optionally, the control and update unit 600 is specifically used for: determining the corresponding first feature vector based on the step to be updated; determining the second feature vectors corresponding to each process template in the process template library; calculating the similarity between the first feature vector and multiple second feature vectors respectively; determining the process template corresponding to the second feature vector with the highest similarity as the target process template; using the target process template to replace the process detailed description corresponding to the step to be updated in the system prompt words.

[0166] Optionally, the control and update unit 600 is further used for: if the highest similarity among multiple similarities does not meet the preset threshold, determining the corresponding template reference file based on each process template in the process template library; using the template reference file as the input attachment of the large model and inputting the corresponding template reset command to the large model to obtain each new process template output by the large model; using each new process template to replace each process template in the process template library.

[0167] Optionally, the control and update unit 600 is further used for: if the estimated information output by the large model is empty, inputting the preset retention strategy information into the large model to obtain the user retention text output by the large model; controlling the outbound robot to output the outbound voice corresponding to the user retention text.

[0168] Optionally, the control and update unit 600 is further used for: if the response text output by the large model contains the specified keyword, notifying the human customer service to provide services for the user instead of the outbound robot and synchronously forwarding the call information to the human customer service.

[0169] Each of the above - shown units uses the current user's inquiry information, historical conversation information, the description text of the preset outbound task, and the standard operation process as the input to the large - model, and uses the response text output by the large - model as a reference for the outbound robot to output outbound voice. Moreover, it can update the standard operation process in the system prompt words with the prediction information of the large - model, so that the responses of the outbound robot are diverse and flexible.

[0170] This application also provides a computer - readable storage medium. The computer - readable storage medium includes a stored program, wherein the program executes the above - provided outbound call method based on the large - model of this application.

[0171] This application also provides an electronic device, including: a processor, a memory, and a bus. The processor is connected to the memory through the bus. The memory is used to store the program, and the processor is used to run the program. When the program runs, it executes the above - provided outbound call method based on the large - model of this application.

[0172] In addition, at least part of the functions described above in the embodiments of this application can be performed by one or more hardware logic components. For example, without limitation, the exemplary types of hardware logic components that can be used include: Field - Programmable Gate Array (FPGA), Application - Specific Integrated Circuit (ASIC), Application - Specific Standard Product (ASSP), System on Chip (SOC), Complex Programmable Logic Device (CPLD), and so on.

[0173] Although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this application. Certain features described in the context of separate embodiments can also be implemented combinatorially in a single embodiment. Conversely, the various features described in the context of a single embodiment can also be implemented separately or in any suitable sub - combination in multiple embodiments.

[0174] The above description is only a preferred embodiment of this application and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of disclosure involved in this application is not limited to the technical solutions formed by the specific combination of the above - mentioned technical features, but also covers other technical solutions formed by any combination of the above - mentioned technical features or their equivalent features without departing from the above - mentioned disclosure concept. For example, the technical solutions formed by mutually replacing the above - mentioned features with other technical features (but not limited to) having similar functions disclosed in this application.

Claims

1. A method for outbound calls based on a large model, characterized in that, Including: Determine the call information currently obtained by the outbound call system; The call information includes the current user's inquiry information and the historical conversation information between the user and the outbound call robot; Based on the current user's inquiry information, determine the corresponding user prompt words; Based on the description text of the preset outbound call task and the standard operation process, determine the corresponding system prompt words; the standard operation process includes multiple steps and corresponding detailed process descriptions; Based on the historical conversation information, determine the corresponding reasoning requirements; Use the user prompt words, the system prompt words, and the reasoning requirements as the input to a pre-deployed large model to obtain the response text and estimated information output by the large model; the estimated information includes the target steps matching the future user inquiry information; Control the outbound call robot to output the outbound call voice corresponding to the response text, and update the system prompt words according to the target steps.

2. The method according to claim 1, wherein Updating the system prompt words according to the target steps includes: Based on the historical conversation information, determine the cumulative number of conversations between the user and the outbound call robot; If the cumulative number of conversations is 0, determine the steps between the first step and the target step in the standard operation process as the steps to be updated; the first step is the step with the first execution order in the standard operation process; Update the detailed process description corresponding to the step to be updated in the system prompt words.

3. The method according to claim 2, characterized in that, The method further includes: If the cumulative number of conversations is not 0, determine the second step that matches the current user's inquiry information; Determine the steps between the second step and the target step in the standard operation process as the steps to be updated.

4. The method according to claim 2, wherein Updating the detailed process description corresponding to the step to be updated in the system prompt words includes: Based on the step to be updated, determine the corresponding first feature vector; Determine the second feature vectors corresponding to each process template in the process template library; Calculate the similarity between the first feature vector and multiple second feature vectors respectively; Based on the process template corresponding to the second feature vector with the highest similarity, determine it as the target process template; Use the target process template to replace the detailed process description corresponding to the step to be updated in the system prompt words.

5. The method according to claim 4, characterized in that After calculating the similarity between the first feature vector and multiple second feature vectors respectively, the method further includes: If the highest similarity among multiple similarities does not meet the preset threshold, determine the corresponding template reference file based on each process template in the process template library; Use the template reference file as the input attachment of the large model, and input the corresponding template reset command to the large model to obtain each new process template output by the large model; Use each new process template to replace each process template in the process template library.

6. The method according to claim 1, wherein The method further includes: If the estimated information output by the large model is empty, input the preset retention strategy information into the large model to obtain the user retention text output by the large model; Control the outbound robot to output the outbound voice corresponding to the user retention text.

7. The method according to claim 1, characterized in that, The method further includes: If the response text output by the large model contains a specified keyword, notify the human customer service to provide services to the user on behalf of the outbound robot, and synchronously forward the call information to the human customer service.

8. An outbound call device based on a large model, characterized in that, It includes: A call information determination unit for determining the call information currently obtained by the outbound system; The call information includes the current user inquiry information and the historical conversation information between the user and the outbound robot; A user prompt determination unit for determining the corresponding user prompt word based on the current user inquiry information; A system prompt determination unit for determining the corresponding system prompt word based on the description text of the preset outbound task and the standard operation process; the standard operation process includes multiple steps and corresponding detailed process descriptions; An inference requirement determination unit for determining the corresponding inference requirement based on the historical conversation information; A large model call unit for using the user prompt word, the system prompt word, and the inference requirement as the input of a pre-deployed large model to obtain the response text and the estimation information output by the large model; the estimation information includes the target step matching the future user inquiry information; A control update unit for controlling the outbound robot to output the outbound voice corresponding to the response text, and updating the system prompt word according to the target step.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein the program, when run by a processor, executes the large model-based outbound call method according to any one of claims 1-7.

10. An electronic device, characterized in that, It includes: A processor, a memory, and a bus; The processor is connected to the memory through the bus; The memory is used to store the program, and the processor is used to run the program, wherein the program, when run by the processor, executes the large model-based outbound call method according to any one of claims 1-7.

Images