Dialogue abstract generation method and device, electronic equipment, medium and program product

Abstract

The embodiment of the invention provides a dialogue abstract generation method and device, electronic equipment, a medium and a program product, and relates to the fields of artificial intelligence, cloud technologies and the like, in particular to the field of large models and natural language processing in the field of artificial intelligence. The method can comprise the steps of obtaining target dialogue information input by a target object; target prompt information corresponding to the target dialogue information is determined, the target prompt information comprises a key field and is used for prompting generation of a dialogue abstract based on the key field, and the steps of generating summary information of the target dialogue information based on the target dialogue information and the target prompt information are executed through a large-scale language model; and based on the target prompt information, extracting field information of each target key field from the target dialogue information, and taking summary information of the target dialogue information and the field information of each target key field as summary information of the target dialogue information. According to the method provided by the embodiment of the invention, a high-quality dialogue abstract can be generated.

Description

Technical Field

[0001] This application belongs to the field of computer technology and may involve fields such as artificial intelligence, cloud technology, large models, natural language processing, etc. Specifically, this application relates to a conversation summary generation method, device, electronic device, medium and program product. Background Art

[0002] The conversation summary is the key information in the conversation. During the conversation process, in some cases, the historical conversation may be between the intelligent customer service and the user. When the manual customer service continues the conversation with the user, in order to better communicate, it is necessary to understand the historical conversation between the intelligent customer service and the user. Or even if the conversation between the intelligent customer service and the user ends, in some cases, the manager still needs to understand the content of the conversation to determine whether the intelligent customer service solves the user's problem. Or, in some cases, it is also necessary to obtain the conversation between the user and the customer service (manual customer service or intelligent customer service) for subsequent data analysis.

[0003] In the above process, it is usually necessary to browse all historical conversation records to review the historical conversation content between the user and the customer service. However, this method of reviewing the historical conversation content by browsing all historical conversation records is highly complex, time-consuming, and inefficient.

[0004] Therefore, how to generate a high-quality conversation summary corresponding to the conversation information based on the conversation information becomes a key issue. Summary of the Invention

[0005] The purpose of the embodiments of the present application is to provide a method, apparatus, electronic device, medium, and program product for generating a conversation summary corresponding to conversation information. To achieve the above purpose, the technical solutions provided by the embodiments of the present application are as follows:

[0006] In a first aspect, a method for generating a conversation summary is provided, comprising:

[0007] Get the target dialogue information input by the target object;

[0008] Determining target prompt information corresponding to the target conversation information, wherein the target prompt information includes key fields and is used to prompt generation of a conversation summary based on the key fields, wherein the key fields include at least one first key field;

[0009] Large-scale language models can do the following:

[0010] generating summary information of the target dialogue information based on the target dialogue information and the target prompt information; and

[0011] Extracting, based on the target prompt information, field information of each target key field from the dialog information, the target key field including the at least one first key field and at least one second key field, the at least one second key field being obtained based on the at least one first key field;

[0012] The summary information of the target conversation information and the field information of each target key field are used as the summary information of the target conversation information.

[0013] In one possible implementation, determining the at least one second key field includes:

[0014] Get multiple preset fields;

[0015] Determining at least one second key field from each of the preset fields based on at least one of a relevance between each of the preset fields and the target domain or a similarity between each of the preset fields and at least one of the first key fields;

[0016] The target domain is the application domain to which the target dialogue information belongs.

[0017] In another possible implementation, extracting field information of at least one second key field from the target dialog information based on the target prompt information includes:

[0018] Identifying preset information for the target prompt information;

[0019] If the target prompt information includes the preset information, field information of the at least one second key field is generated based on the target dialogue information and the target prompt information.

[0020] In another possible implementation, determining at least one second key field from each of the preset fields based on at least one of a correlation between each of the preset fields and the target domain or a similarity between each of the preset fields and at least one of the first key fields includes:

[0021] Among the preset fields, at least one preset field having a correlation with the target field greater than a preset correlation and a similarity with at least one of the first key fields greater than a first preset similarity is determined as at least one of the second key fields.

[0022] In another possible implementation, the plurality of preset fields belong to a target field set corresponding to the target field;

[0023] Determining at least one second key field from each of the preset fields based on similarity between each of the preset fields and at least one of the first key fields includes:

[0024] For each first key field, a field in the target field set having a similarity with the first key field greater than a second preset similarity is determined to obtain a field corresponding to each first key field as at least one second key field corresponding to the target conversation information.

[0025] In another possible implementation, determining the target prompt information corresponding to the target dialogue information includes any one of the following:

[0026] Acquire configuration information related to the target prompt information input by the target object, and construct target prompt information corresponding to the target dialogue information based on the configuration information;

[0027] An application field corresponding to the target dialogue information is determined, prompt information related to the application field is determined from various pre-configured prompt information, and the determined prompt information is used as the target prompt information.

[0028] In another possible implementation, the configuration information is determined by at least one of the following methods:

[0029] In response to a prompt information construction instruction triggered by the target object received through the configuration interface, displaying configuration prompt information related to the application field corresponding to the target dialogue information;

[0030] The target object receives configuration information related to the target prompt information based on the configuration prompt information.

[0031] In another possible implementation, the large-scale language model is trained in the following manner:

[0032] Acquire multiple training samples, wherein each training sample includes: sample conversation information, a conversation summary sample corresponding to the sample conversation information, and prompt information corresponding to the sample conversation information, wherein the conversation summary sample includes: summary information in the sample conversation information and field information of at least one sample key field in the sample conversation information;

[0033] Based on the multiple training samples, an initial large-scale language model is trained to obtain the large-scale language model, wherein the training loss of the large-scale language model is determined based on the difference between the conversation summary sample corresponding to each of the training samples and the conversation summary generated by the model.

[0034] In another possible implementation, the large-scale language model is obtained by:

[0035] Obtain multiple trained candidate large-scale language models;

[0036] determining the performance of each of the candidate large-scale language models;

[0037] The large-scale language model is determined from the candidate large-scale language models based on performance of the candidate large-scale language models.

[0038] In another possible implementation, the performance of any candidate large-scale language model is determined by:

[0039] Obtaining answers to frequently asked questions, wherein the answers to frequently asked questions include at least one question and an answer to each question;

[0040] For each of the questions, based on the question, generating model input information corresponding to the question, the model input information including at least one of the dialogue information or the summary information;

[0041] For each of the questions, input the model input information corresponding to the question into any of the target large-scale language models to obtain a predicted answer, and

[0042] Perform similarity matching between the predicted answer and the answer corresponding to the question to obtain a matching result;

[0043] Based on the matching results corresponding to each question, the performance of any of the large-scale language models is determined.

[0044] In a second aspect, a device for generating a conversation summary is provided, the device comprising:

[0045] An acquisition module is used to obtain target dialogue information input by a target object;

[0046] A first determining module is configured to determine target prompt information corresponding to the target conversation information, wherein the target prompt information includes key fields and is configured to prompt generation of a conversation summary based on the key fields, wherein the key fields include at least one first key field;

[0047] The model execution processing module is used to perform the following operations using a large-scale language model:

[0048] generating summary information of the target dialogue information based on the target dialogue information and the target prompt information; and

[0049] Extracting, based on the target prompt information, field information of each target key field from the dialog information, the target key field including the at least one first key field and at least one second key field, the at least one second key field being obtained based on the at least one first key field;

[0050] The summary information of the target conversation information and the field information of each target key field are used as the summary information of the target conversation information.

[0051] In a possible implementation, the apparatus further includes: a second determining module, wherein:

[0052] When determining the at least one second key field, the second determining module is specifically configured to:

[0053] Get multiple preset fields;

[0054] Determining at least one second key field from each of the preset fields based on at least one of a relevance between each of the preset fields and the target domain or a similarity between each of the preset fields and at least one of the first key fields;

[0055] The target domain is the application domain to which the target dialogue information belongs.

[0056] In another possible implementation, when the model execution processing module extracts field information of at least one second key field from the target dialogue information based on the target prompt information, it is specifically configured to:

[0057] Identifying preset information for the target prompt information;

[0058] If the target prompt information includes the preset information, field information of the at least one second key field is generated based on the target dialogue information and the target prompt information.

[0059] In another possible implementation, when the second determination module determines at least one second key field from each of the preset fields based on at least one of a correlation between each of the preset fields and the target domain or a similarity between each of the preset fields and at least one of the first key fields, it is specifically configured to:

[0060] Among the preset fields, at least one preset field having a correlation with the target field greater than a preset correlation and a similarity with at least one of the first key fields greater than a first preset similarity is determined as at least one of the second key fields.

[0061] In another possible implementation, the plurality of preset fields belong to a target field set corresponding to the target field;

[0062] When the second determining module determines at least one second key field from each of the preset fields based on the similarity between each of the preset fields and at least one of the first key fields, it is specifically configured to:

[0063] For each first key field, a field in the target field set having a similarity with the first key field greater than a second preset similarity is determined to obtain a field corresponding to each first key field as at least one second key field corresponding to the target conversation information.

[0064] In another possible implementation, the first determining module is specifically configured to perform any of the following when determining the target prompt information corresponding to the target dialogue information:

[0065] Acquire configuration information related to the target prompt information input by the target object, and construct target prompt information corresponding to the target dialogue information based on the configuration information;

[0066] An application field corresponding to the target dialogue information is determined, prompt information related to the application field is determined from various pre-configured prompt information, and the determined prompt information is used as the target prompt information.

[0067] In another possible implementation, the configuration information is determined by at least one of the following methods:

[0068] In response to a prompt information construction instruction triggered by the target object received through the configuration interface, displaying configuration prompt information related to the application field corresponding to the target dialogue information;

[0069] The target object receives configuration information related to the target prompt information based on the configuration prompt information.

[0070] In another possible implementation, the large-scale language model is trained in the following manner:

[0071] Acquire multiple training samples, wherein each training sample includes: sample conversation information, a conversation summary sample corresponding to the sample conversation information, and prompt information corresponding to the sample conversation information, wherein the conversation summary sample includes: summary information in the sample conversation information and field information of at least one sample key field in the sample conversation information;

[0072] Based on the multiple training samples, an initial large-scale language model is trained to obtain the large-scale language model, wherein the training loss of the large-scale language model is determined based on the difference between the conversation summary sample corresponding to each of the training samples and the conversation summary generated by the model.

[0073] In another possible implementation, the large-scale language model is obtained by:

[0074] Obtain multiple trained candidate large-scale language models;

[0075] determining the performance of each of the candidate large-scale language models;

[0076] The large-scale language model is determined from the candidate large-scale language models based on performance of the candidate large-scale language models.

[0077] In another possible implementation, the performance of any candidate large-scale language model is determined by:

[0078] Obtaining answers to frequently asked questions, wherein the answers to frequently asked questions include at least one question and an answer to each question;

[0079] For each of the questions, based on the question, generating model input information corresponding to the question, the model input information including at least one of the dialogue information or the summary information;

[0080] For each of the questions, input the model input information corresponding to the question into any of the target large-scale language models to obtain a predicted answer, and

[0081] Perform similarity matching between the predicted answer and the answer corresponding to the question to obtain a matching result;

[0082] Based on the matching results corresponding to each question, the performance of any of the large-scale language models is determined.

[0083] In a third aspect, an embodiment of the present application further provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method for generating a conversation summary provided by any possible implementation of the first aspect.

[0084] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the method for generating a conversation summary provided by any possible implementation of the first aspect.

[0085] In a fifth aspect, an embodiment of the present application further provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the method for generating a conversation summary provided by any possible implementation of the first aspect.

[0086] The beneficial effects of the technical solution provided by the embodiments of the present application are as follows:

[0087] Embodiments of the present application provide a conversation summary generation method, apparatus, electronic device, and storage medium. In the embodiments of the present application, after obtaining target conversation information input by a target object and determining target prompt information corresponding to the target conversation information, summary information for the target conversation information can be generated using a large-scale language model and based on the target prompt information. Furthermore, at least one second key field is heuristically obtained based on a first key field in the target prompt information, and field information of the first key field and field information of the second key field are extracted from the target conversation information to obtain summary information of the target conversation information based on the summary information of the target conversation information, the field information of the first key field, and the field information of the second key field. That is, in the embodiments of the present application, the generated summary information not only includes a summary of the target conversation information and field information of the first key field, but also heuristically obtains field information of the second key field. Thus, with a single prompt information, higher-quality summary information that meets the requirements of the target object can be generated, thereby generating an experience for the target object. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments of the present application.

[0089] Figure 1a A schematic diagram of the structure of a conversation summary generation system provided in an embodiment of the present application;

[0090] Figure 1b A schematic diagram of the structure of another conversation summary generation system provided in an embodiment of the present application;

[0091] Figure 1c A schematic diagram of the structure of another conversation summary generation system provided in an embodiment of the present application;

[0092] Figure 1d A flowchart of a method for generating a conversation summary provided in an embodiment of the present application;

[0093] Figure 1e A flowchart of another method for generating a conversation summary provided in an embodiment of the present application;

[0094] Figure 2 This is an example diagram of generating a conversation summary during a practical application process in an embodiment of the present application;

[0095] Figure 3a This is a schematic diagram of a method for generating a conversation summary using a telephone conference as an example in an embodiment of the present application;

[0096] Figure 3b This is a flowchart of an application process of a method for generating a conversation summary in the actual application process of an embodiment of the present application;

[0097] Figure 3c This is a flowchart of a training process in a method for generating a conversation summary during the actual application of an embodiment of the present application;

[0098] Figure 4 This is a structural diagram of a conversation summary generation device according to an embodiment of the present application;

[0099] Figure 5 This is a schematic diagram of the device structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0100] The following describes the embodiments of the present application in conjunction with the accompanying drawings. It should be understood that the embodiments described below in conjunction with the accompanying drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0101] Those skilled in the art will understand that, unless otherwise stated, the singular forms "a", "an", "said", and "the" used herein may also include plural forms. It should be further understood that the terms "including" and "comprising" used in the embodiments of the present application mean that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements, and / or components, but do not exclude implementation as other features, information, data, steps, operations, elements, components, and / or combinations thereof supported by the present technical field. It should be understood that when we say that an element is "connected" or "coupled" to another element, the element can be directly connected or coupled to the other element, or it can refer to the element and the other element establishing a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used here can include wireless connection or wireless coupling. The term "and / or" used here indicates at least one of the items defined by the term, for example, "A and / or B" can be implemented as "A", or as "B", or as "A and B". When describing multiple (two or more) items, if the relationship between the multiple items is not clearly defined, the multiple items may refer to one, multiple or all of the multiple items. For example, the description of "parameter A includes A1, A2, A3" can be implemented as parameter A including A1 or A2 or A3, and can also be implemented as parameter A including at least two of the three items A1, A2, and A3.

[0102] The embodiments of the present application are to provide a method, device, electronic device, and storage medium for generating a conversation summary in a human-computer interaction scenario, in which, upon obtaining conversation information, a higher-quality summary result can be generated through a single request, so that the requesting party can better and faster understand the conversation information.

[0103] In order to better understand and illustrate the solutions provided by the embodiments of the present application, some technical terms involved in the embodiments of the present application are briefly introduced below.

[0104] Large Language Model (LLM): A large-scale language model is a natural language processing model based on deep learning. It can learn the syntax and semantics of natural language to generate human-readable text. A "language model" is an AI model that processes language (or symbolic systems), identifies patterns within them, and automatically generates content that conforms to these patterns based on prompts.

[0105] Conversation summarization: A special case of text summarization, conversation summarization focuses on conversational data. Conversational data can take many forms, such as meetings, casual chats, emails, debates, and customer service. While these different forms of conversation summarization have different applications in their respective fields, their core goal is to capture key information from a conversation and facilitate rapid understanding of its core content.

[0106] Prompt learning: Unify all downstream tasks into pre-training tasks; use specific templates to convert downstream task data into natural language, fully tapping the capabilities of the pre-trained model. Essentially, this involves designing a template that closely matches the upstream pre-training task. This template design unlocks the potential of the upstream pre-trained model, allowing it to better complete downstream tasks with minimal need for labeled data. This key approach involves three steps:

[0107] 1. Design the task of pre-training language model;

[0108] 2. Design input template style (Prompt Engineering);

[0109] 3. Design label styles and how to map model outputs to labels (Answer Engineering).

[0110] Prompt: Prompt text, representing a command or instruction, is used to instruct the large language model to perform an action or generate an output, that is, to instruct the model what action to take or what output to generate when performing a specific task.

[0111] NLG (Natural Language Generation): Natural language generation uses artificial intelligence and natural language processing to convert non-natural language data into a natural language form that is easy to understand and use. Its main purpose is to reduce the communication gap between humans and machines and convert non-language format data into a language format that humans can understand.

[0112] Among them, the solutions provided in the embodiments of this application relate to the field of artificial intelligence. Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to achieve the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0113] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, pre-trained models, operating / interaction systems, and mechatronics. Pre-trained models, also known as large models or basic models, can be fine-tuned and widely applied to downstream tasks across various AI disciplines. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0114] The solutions provided in the embodiments of the present application may specifically relate to natural language processing, machine learning / deep learning and other fields in the field of AI. Natural language processing (NLP) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that can enable effective communication between humans and computers using natural language. Natural language processing involves natural language, that is, the language people use in daily life, and is closely related to linguistic research; it also involves computer science and mathematics. The pre-training model, an important technology for model training in the field of artificial intelligence, is developed from the large language model (Large Language Model) in the field of NLP. After fine-tuning, the large language model can be widely used in downstream tasks. Natural language processing technology generally includes text processing, semantic understanding, machine translation, robot question answering, knowledge graphs and other technologies.

[0115] Optionally, the data processing involved in the methods provided in the embodiments of this application (e.g., obtaining a training set during the large language model training phase, processing target conversation information using the trained large language model, etc.) can be implemented based on cloud computing. Cloud computing is a computing model that distributes computing tasks across a resource pool consisting of a large number of computers, enabling various application systems to obtain computing power, storage space, and information services as needed.

[0116] It should be noted that in the optional embodiments of the present application, the target dialogue information and target prompt information and other related data involved, when the embodiments in the present application are applied to specific products or technologies, need to obtain the permission or consent of the object, and the collection, use and processing of the relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions. In other words, if the embodiments of the present application involve data related to the object, it needs to be obtained through the authorization and consent of the object, the authorization and consent of the relevant departments, and in compliance with the relevant laws, regulations and standards of the country and region. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject needs to be obtained. The embodiments also need to be implemented with the authorization and consent of the object.

[0117] In the related art, conversation summaries may include: (1) Problem / Solution Summary: a feature specific to call centers that provides a summary of problems and solutions in conversations between customer service agents and customers; (2) Section Title Summary: provides suggested titles of input conversation sections; and (3) Narrative Summary: provides call notes, meeting notes, or chat summaries of input conversations.

[0118] Function (1) is primarily applicable to customer service-consumer online chat scenarios. It extracts customer service scripts and user questions, such as customer service question-customer answer and customer service question-customer answer, from the conversation content. This can be used for subsequent analysis of user hot issues or building a customer service script library to optimize customer service robots. Function (2) title summaries significantly detract from information, and function (3) narrative summaries can be long, making it difficult for users to find the key points.

[0119] In order to solve the above technical problems, the embodiment of the present application proposes a heuristic-based dialogue summary generation scheme, which is applied to dialogue scenarios (such as: meeting summaries, customer service work order summaries, etc.). In various dialogue-based business scenarios, due to the length of the dialogue, it is very important for the business party to "reduce" the information. Extracting the core information in the dialogue can greatly accelerate the subsequent follow-up costs. In the embodiment of the present application, the LLM model is used to unify the two tasks of dialogue summarization (generative summary and extractive summary), and heuristics are added to the prompt template through prompt learning, so that the induction model can extract key field information while generating the dialogue summary when the LLM outputs, so that the LLM model can better match the needs of the business. That is, in the embodiment of the present application, the generation and extraction functions are integrated together through prompt information, and heuristic information prompts are innovatively used to help the conversation summary extract key information, so that the extracted fields contained in the obtained summary information are no longer limited to the provided field range, and LLM can also find the key fields in the field that do not appear in the prompt information according to the prompt information; the embodiment of the present application also introduces the Frequently-Asked Questions (FAQ) summary task to achieve quality verification, help the model learn the key information in the conversation, and expand the applicable scenarios of LLM. The embodiment of the present application innovatively proposes a universal conversation solution that can ensure that the business party obtains a higher quality summary result within a single request.

[0120] The embodiments of the present application can be applied to conversational business scenarios to refine lengthy conversation information in business scenarios such as telephone, conference, and instant messaging chat, thereby accelerating subsequent follow-up costs.

[0121] Further, see Figure 1a , Figure 1a This is a structural diagram of a conversation summary generation system provided by an embodiment of the present application. Figure 1a The conversation summary generation system shown includes a first server 110 and a terminal 120 .

[0122] The terminal 120 may include, but is not limited to, one or more of various desktop computers, laptop computers, smartphones, tablet computers, IoT devices, portable wearable devices, or immersive image display devices. IoT devices may be one or more of smart speakers, smart TVs, smart air conditioners, or smart car devices. Portable wearable devices may be one or more of smart watches, smart bracelets, or head-mounted devices. Immersive image display devices include, but are not limited to, augmented reality (AR) devices and virtual reality (VR) devices.

[0123] Among them, the server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), as well as big data and artificial intelligence platforms.

[0124] Specifically, the first server 110 can function as a training server, obtaining multiple training sets containing multiple training samples and performing training to obtain a trained large-scale language model. The trained large-scale language model can then be deployed on the first server 110 or the terminal 120. When it is necessary to generate summary information for target conversation information, the terminal 120 can obtain target prompt information and target conversation information, and then generate a summary of the target conversation information based on the trained large-scale language model.

[0125] It is understood that if the trained large-scale language model is deployed on first server 110, after terminal 120 obtains the target prompt information and target conversation information, it can send the target prompt information and target conversation information to first server 110. First server 110 then generates summary information of the target conversation information based on the target prompt information. Terminal 120 then receives the summary information of the target conversation information fed back by first server 110 and displays the summarized information. For example, if the recommended object is Advertisement 1, terminal 120 displays Advertisement 1. If the trained large-scale language model is deployed on terminal 120, terminal 120 can directly call the trained large-scale language model to generate the summary information of the target conversation information, or it can display the summarized information after generating the summary information of the target conversation information.

[0126] It should be noted that the target conversation information may be obtained through the terminal 120 or from other devices; this is not limited here.

[0127] In some scenarios, training can also be performed through the first server 110, and the trained large-scale language model can be deployed on the first server 110; in addition, training can also be performed through the terminal 120, and the trained target large language model can be deployed on the terminal 120; in addition, the first server 110 and the terminal 120 server can also be trained collaboratively, which is not limited here.

[0128] See also Figure 1b , Figure 1b This is a structural diagram of another conversation summary generation system provided by an embodiment of the present application. Figure 1bThe conversation summary generation system shown includes a first server 110, a terminal 120, and a second server 130. In this embodiment, the first server 110 can serve as a training server, and the second server 130 can serve as a conversation summary generation server. That is, the second server 130 generates a conversation summary for the target conversation information.

[0129] In some scenarios, the first server 110 and at least one of the terminal 120 or the second server 130 may also perform collaborative training, which is not limited here.

[0130] Specifically, a large-scale language model can be trained by the first server 110 and deployed in the second server 130. Then, the target prompt information and target dialogue information can be obtained through the terminal 120, and the target prompt information and target dialogue information can be sent to the second server 130. The second server 130 can then generate summary information for the terminal 120. That is, after the second server 130 generates the summary information of the target dialogue information, it will feedback the summary information to the terminal 120.

[0131] See also Figure 1c , Figure 1c This is a schematic diagram of the structure of another object recommendation system provided by the embodiment of the present application. Figure 1c The object recommendation system shown includes a first server 110 , a terminal 120 , and multiple second servers 130 .

[0132] The first server 110 can train a model to obtain summary information of target conversation information in each scenario and deploy each target large language model on a different second server 130. After obtaining the target conversation information through the terminal 120, the application scenario of the target conversation information can be determined, and the target conversation information and target prompt information can be sent to the corresponding second server 130. The second server 130 that receives the target prompt information and target conversation information processes the target conversation information using the deployed target large-scale language model to obtain summary information corresponding to the target conversation information, and then feeds the summary information back to the terminal 120.

[0133] It is understandable that the above large-scale language systems are some example scenarios and do not constitute all scenarios for implementing the technical solutions of the embodiments of the present application.

[0134] The following describes several exemplary embodiments to illustrate the technical solutions of the embodiments of the present application and the technical effects produced by the technical solutions of the present application. It should be noted that the following embodiments can refer to, learn from, or combine with each other, provided that there is no conflict between them. For the same terms, similar features, and similar implementation steps in different embodiments, the description will not be repeated.

[0135] Figure 1d A method for generating a conversation summary provided by an embodiment of the present application is shown, wherein the method can be executed by a terminal or by a server. For example, the method can be deployed in an application running in the terminal, which can provide human-computer interaction services and output summary information in the target conversation information to the user. That is, when the user obtains the target conversation information and target prompt information input by the user through the human-computer interaction interface of the application (the user interface of the application), summary information corresponding to the target conversation information can be generated and fed back to the user. The operation of generating the summary information corresponding to the target conversation information can be executed in the user terminal or by the server. That is, the trained large-scale language model can be deployed on the terminal or on the server side. Optionally, it can be deployed on the server side to reduce the resource occupation of the application on the terminal. Deployment on the server side also facilitates the update or upgrade of the model.

[0136] like Figure 1d As shown, the conversation summary generation method provided in the embodiment of the present application may include the following steps S101-S103.

[0137] S101: Acquire target dialogue information input by a target object.

[0138] In the embodiment of the present application, the target conversation information input by the target object can be entirely in text format, or in voice format, or partly in voice format and partly in text format. In the embodiment of the present application, if the target conversation information input by the target object includes voice format conversation information, the voice format target conversation information can be converted into text format to obtain the text format target conversation information.

[0139] Furthermore, the target object may also input target dialogue information in a video format, from which dialogue information in a voice format is extracted, and the dialogue information in the voice format is converted into target dialogue information in a text format.

[0140] S102: Determine target prompt information corresponding to the target dialogue information.

[0141] In the embodiment of the present application, different dialogue messages may correspond to the same prompt information, and different dialogue messages may correspond to different prompt information. In the embodiment of the present application, if different dialogue messages correspond to different prompt information, dialogue messages in the same application field may correspond to the same prompt information, or dialogue messages in the same application field may correspond to different prompt information.

[0142] Furthermore, if different dialogue information corresponds to the same prompt information, that is, the prompt information is fixed; if the dialogue information of the same application field corresponds to the same prompt information, that is, one field corresponds to one prompt information, after obtaining the target dialogue information, determine the application field to which the target dialogue information belongs, so as to determine the prompt information corresponding to the application field as the target prompt information corresponding to the target dialogue information; if the dialogue information of the same application field can correspond to different prompt information, other methods can be used to determine the target prompt information corresponding to the target dialogue information. Furthermore, the prompt information corresponding to different dialogue information can be input by the target object, or obtained according to the configuration of the target object, which is not limited in the embodiments of the present application.

[0143] Among them, the target prompt information includes key fields and is used to prompt the generation of a conversation summary based on the key fields. The key fields include at least one first key field. In an embodiment of the present application, at least one first key field is a key field carried in the target prompt information to prompt the extraction of field information corresponding to each first key field from the target conversation information. That is to say, the conversation summary generated based on the target prompt information contains at least field information corresponding to each first key field in at least one first key field.

[0144] It should be noted that step S102 can be executed before step S101, after step S101, or simultaneously with step S101, which is not limited in the embodiments of the present application.

[0145] S103. Perform the following operations using the large-scale language model: generate summary information of the target dialogue information based on the target dialogue information and the target prompt information; and extract field information of each target key field from the dialogue information based on the target prompt Western Sydney.

[0146] The summary information of the target conversation information and the field information of each target key field are used as the summary information of the target conversation information.

[0147] After obtaining the summary information of the target conversation information and extracting the field information of each target key field in the above manner, the summary information of the target conversation information and the field information of each target key field are used as the summary information of the target conversation information. In other words, the summary information includes: the summary information of the target conversation information and the field information of each target key field, wherein the target key field includes at least one first key field and at least one second key field, and the at least one second key field is obtained based on the at least one first key field; in this embodiment of the application, the target prompt information is used to prompt the generation of a conversation summary based on the target key field in the target conversation information, and the target key field includes at least one first key field and at least one second key field. That is, the generated summary information of the target conversation information includes the summary information of the target conversation information, the field information of the first key field, and the field information of the second key field.

[0148] Specifically, in an embodiment of the present application, the target prompt information includes at least one first key field, and at least one second key field can be heuristically obtained based on the at least one first key field, so that when the field information of the first key field is extracted from the target conversation information, the field information of the second key field can also be extracted.

[0149] It should be noted that the summary information of the target conversation information and the field information of the target key fields (the field information of the first key field and the field information of the second key field) can be displayed separately, that is, displayed through different paragraphs, and the field information of the first key field and the field information of the second key field can also be included in the generated summary information of the target conversation information. Optionally, in real applications, the solution of using LLM to integrate the generation and extraction functions through target prompt information to help the conversation summary extract key information is more in line with business needs. That is, using LLM to generate a summary of the target conversation information and extract the field information of the first key field and the field information of the second key field from it to obtain summary information, so that the summary information generated by the embodiment of the present application can be more in line with business needs, and the quality of the summary information is higher.

[0150] Furthermore, if the summary information of the target conversation information is generated by the server, the server sends the generated summary information of the target conversation information to the terminal so that the terminal can display it; if the summary information of the target conversation information is generated by the terminal, the terminal can directly display it after generating the summary information of the target conversation information.

[0151] Furthermore, it can be seen from the above embodiment that the generated summary information includes field information of each target key field, and the target key field includes a first key field and a second key field, that is, the generated summary information includes field information of the first key field, and further, the target key field may also include at least one second key field; that is, the field information of the target key field included in the target summary information generated by the above embodiment includes not only field information of the first key field, but also field information of at least one second key field. In other words, if the target prompt information only includes the first key field, it can be heuristically prompted that the trained large-scale language model is not limited to the provided field range based on the extraction of field information of the first key field, but can also find important field outputs in the field that do not appear in the prompt information according to the prompt, that is, it can also extract field information of the second key field, that is, the generated summary information can include: a conversation summary, field information of each first key field, and field information of each second key field, so as to further improve the quality of summary information generation and enhance the experience of the target object.

[0152] Specifically, the method of determining at least one second key field may include: step Sa and step Sb, such as Figure 1e As shown, Figure 1e Only the processing flow of step Sa and step Sb is shown;

[0153] It should be noted that step Sa and step Sb may be performed before step S103 or during the execution of step S103, which is not limited in the embodiment of the present application. Step Sa and step Sb are as follows:

[0154] Step Sa, obtain multiple preset fields.

[0155] For the embodiments of the present application, the electronic device can store multiple preset fields locally, that is, when determining at least one second key field, multiple preset fields can be obtained from the local storage, or the trained large-scale language model can include multiple preset fields.

[0156] Step Sb: determining at least one second key field from each preset field based on at least one of the relevance of each preset field and the target domain, or the similarity between each preset field and at least one first key field.

[0157] The target domain is the application domain to which the target dialogue information belongs.

[0158] Specifically, in an embodiment of the present application, after obtaining multiple preset fields, the correlation between each preset field and the application field to which the target dialogue information belongs is determined, that is, whether each preset field belongs to the application field to which the target dialogue information belongs is determined; and / or, the similarity between each preset field and at least one first key field is determined, wherein determining the similarity between each preset field and at least one first key field may specifically include: determining the similarity between each preset field and each first key field, wherein a first key field may have at least one similar preset field, or may not have a similar preset field, which is not limited in the embodiment of the present application.

[0159] In other words, the target prompt information may only contain the first key field, indicating that field information of the first key field is included when generating conversation summary information based on the target conversation information. At least one second key field may also be heuristically determined based on the first key field, so that when generating field information containing the first key field based on the target conversation information, field information of at least one second key field may also be included. For example, if the first key field in the target prompt information includes the contact name, the second key field may include the contact phone number.

[0160] Specifically, based on the correlation between each preset field and the target field, or at least one of the similarities between each preset field and at least one first key field, at least one second key field is determined from each preset field. Specifically, it can include: determining at least one preset field in each preset field whose correlation with the target field is greater than the preset correlation and whose similarity with at least one first key field is greater than the first preset similarity as at least one second key field.

[0161] That is to say, after determining the application field to which the target conversation information belongs, that is, the target field, at least one preset field is determined from each preset field whose correlation with the target field is greater than the preset correlation and whose similarity with at least one first key field is greater than the first preset similarity, and these at least one preset field is determined as at least one second key field.

[0162] Furthermore, multiple preset fields belong to the target field set corresponding to the target field; that is, the multiple preset fields obtained themselves belong to the target field set corresponding to the target field. At this time, when determining at least one second key field of the target conversation, it is only necessary to determine from the multiple preset fields a field whose similarity with the first key field is greater than the second preset similarity. That is, in step Sb, based on the similarity between each preset field and at least one first key field, at least one second key field is determined from each preset field. Specifically, it can include: for each first key field, determining the field in the target field set whose similarity with the first key field is greater than the second preset similarity, so as to obtain the fields corresponding to each first key field, as the at least one second key field corresponding to the target conversation information.

[0163] Specifically, if the target prompt information contains only one first key field, at this time, a field whose similarity with the first key field is greater than the second preset similarity is determined from the target field set as at least one second key field corresponding to the target dialogue information. Of course, there may not be a field in the target field set whose similarity with the first key field is greater than the second preset similarity; if the target prompt information contains at least two first key fields, at this time, for each first key field, a field whose similarity with the first key field is greater than the second preset similarity is determined from the target field set, wherein there may not be a second key field for each first key field, or there may be at least two second key fields for one first key field, which is not limited in the embodiments of the present application.

[0164] In order to further improve the experience of the target object and generate higher quality summary information that meets the requirements of the target object, that is, when the target object needs to include the field information of the first key field and the field information of the second key field in the generated summary information, then when generating the summary information, the second key field is heuristically determined and the field information of the second key field is extracted; if the target object only needs the first key field, the generated summary information can only include the field information of the first key field, and does not include the field information of the second key field.

[0165] That is, in step S103, based on the target prompt information, field information of at least one second key field is extracted from the target dialogue information, which may specifically include: identifying preset information of the target prompt information; if the target prompt information contains preset information, generating field information of at least one second key field based on the target dialogue information and the target prompt information.

[0166] In an embodiment of the present application, a large-scale language model can identify target recognition information to identify whether the target prompt information contains preset information. If it is identified that the target prompt information contains preset information, the intention of the target object is represented by extracting the field information of the first key field from the target dialogue information, and the field information of the second key field can be further extracted. If it is identified that the target prompt information does not contain preset information, the intention of the target object is represented by extracting the field information of the first key field from the target dialogue information, and there is no need to heuristically extract the field information of the second key field. That is, in an embodiment of the present application, by identifying whether the target prompt information contains preset information, the intention of the target object can be accurately determined. While improving the quality of generated summary information, the user experience of the target object can also be further improved.

[0167] For example, the preset information may include preset words such as "etc." and "at least", that is, the target prompt information is "Extract key field information mentioned in the conversation, the fields include: car model, consulting questions, contacts, car owners, frame numbers, dealers, etc.", or "Extract key field information mentioned in the conversation, the fields include at least: car model, consulting questions, contacts, car owners, frame numbers, dealers". When it is recognized that the target prompt information contains preset words such as "etc." or "at least", when extracting the field information of the key fields of "car model, consulting questions, contacts, car owners, frame numbers, dealers" from the target conversation information, it is also possible to heuristically extract field information of other fields related to these fields.

[0168] Furthermore, in step S103, based on the target prompt information, field information of at least one second key field is extracted from the target dialogue information, which may specifically include: based on the target prompt information, determining the intention information of the target object; if the intention information represents that the field information of the second key field needs to be extracted from the target summary information, then generating field information of at least one second key field, that is, the summary information generated at this time includes not only the field information of the first key field, but also the field information of the second key field; conversely, if the intention information represents that the field information of the second key field does not need to be extracted from the target summary information, then the summary information generated at this time does not include the field information of the second key field.

[0169] Furthermore, for the target dialogue information and the target prompt information, the field information of generating at least one second key field can be specifically described in the above embodiment, which will not be repeated here.

[0170] Furthermore, it can be seen from the above embodiment that when generating summary information corresponding to the target conversation information, it is necessary to determine the target prompt information corresponding to the target conversation information based on the target prompt information. Specifically, any of the following (methods 1 and 2) may be included, wherein:

[0171] Method 1: Obtain configuration information related to the target prompt information input by the target object, and construct the target prompt information corresponding to the target dialogue information based on the configuration information.

[0172] For the embodiments of the present application, in one possible implementation, the target object can input configuration information related to the target prompt information through the human-computer interaction interface (input interface of the application) corresponding to the trained large-scale language model, and then obtain the target prompt information corresponding to the target dialogue information based on the configuration information, so as to display it through the human-computer interaction interface; wherein, the target prompt information corresponding to the target dialogue information obtained based on the configuration information can be generated by the trained large-scale language model, or can be generated by other means (for example, by other models); further, the target object can also adjust the target prompt information corresponding to the obtained target dialogue information to obtain adjusted target prompt information for subsequent generation of summary information of the target dialogue information;

[0173] For example, the configuration information input by the target object may include: "conversation summary; vehicle model, inquiry question, contact person, vehicle owner, vehicle frame number, dealer, etc." Based on this configuration information, the target prompt information can be generated as follows:

[0174] "Please generate a work order summary based on the following telephone conversation. Requirements: 1. The expression should be fluent and the key points should be highlighted; 2. Summarize the entire conversation first; 3. Then summarize 1-2 key words in the conversation; 4. Extract the key field information mentioned in the conversation, including: vehicle model, inquiry question, contact person, owner, chassis number, dealer, etc."

[0175] In another possible implementation, the electronic device may display various configuration information so that the target object can select from the various configuration information and generate target prompt information corresponding to the target conversation information based on the selected configuration information. The specific method for generating the target prompt information corresponding to the target conversation information is described in detail in the above embodiment and will not be repeated here.

[0176] Specifically, in the embodiment of the present application, the configuration information is determined by the following methods (method 3 and method 4), wherein:

[0177] Mode 3: In response to a prompt information construction instruction triggered by a target object received through the configuration interface, configuration prompt information related to the application field corresponding to the target dialogue information is displayed.

[0178] That is to say, the target object can trigger a virtual button or input a trigger instruction in the configuration interface to trigger the prompt information construction instruction to display configuration prompt information related to the application field corresponding to the target dialogue information, wherein the configuration prompt information is used to prompt the target object to enter accurate configuration information.

[0179] For example, the configuration prompt information can be a configuration prompt template to prompt the target object to select specific configuration information. For example, in the car company scenario, there is a high level of attention to the model, frame number, etc., and the configuration information can be set based on the model, frame number, etc. as inspiration words.

[0180] Furthermore, the configuration prompt information may not be targeted at a certain application field, that is, the configuration prompt template is a general prompt template to prompt the target which configuration information to select.

[0181] Mode 4: receiving configuration information related to the target prompt information input by the target object based on the configuration prompt information.

[0182] Furthermore, the target user may input (select) configuration information related to the configuration prompt information according to the configuration prompt information.

[0183] Method 2: Determine the application field corresponding to the target dialogue information, determine the prompt information related to the application field from various pre-configured prompt information, and use the determined prompt information as the target prompt information.

[0184] Specifically, prompt information corresponding to each field can also be pre-configured. For example, for the automotive field, prompt information 1 is pre-configured; for the food field, prompt information 2 is pre-configured; and so on. After determining the application field corresponding to the target dialogue information, the prompt information related to the field is determined from each pre-configured prompt information. For example, if the application field corresponding to the target dialogue information is the automotive field, then prompt information 1 is determined to be the target prompt information.

[0185] Furthermore, after determining the target prompt information, it can be displayed directly to generate summary information of the target conversation information instead of through the human-computer interaction interface (input interface of the application) corresponding to the trained large-scale language model; or, after determining the target prompt information, it can also be displayed through the human-computer interaction interface (input interface of the application) corresponding to the large-scale language model. Of course, during the display process, the target object can also adjust the prompt information so that the summary information corresponding to the target conversation information can be generated based on the adjusted prompt information. Furthermore, in an embodiment of the present application, multiple prompt information can be pre-configured, and then the target prompt information can be configured for it according to the habits of the target object. Alternatively, the target object can select a prompt information from the pre-configured multiple prompt information as the target prompt information.

[0186] Furthermore, as described above, the target conversation information can be input by the target object. In addition, the target conversation information can also be imported into the large-scale language model, that is, it can be in a file format. Alternatively, the target object can also select a dialog box with a certain contact through a selection operation to import the conversation information with the contact into the large-scale language model, so as to generate summary information through the large-scale language model. For example, Figure 2 As shown, Figure 2 All conversations can be displayed on the left side of the , for example, the conversations corresponding to Visitor 1, Visitor 2, Visitor 3, Visitor 4 and Visitor 5 respectively. The target object can select one of the dialog boxes, such as the dialog box of Visitor 3, to Figure 2 The middle part shows the historical message records of visitor 3, and then you can click "Create Work Order" to enter the target conversation information and prompt information to generate a summary of the target conversation information.

[0187] Specifically, the large-scale language model is trained in the following manner: obtaining multiple training samples; and training an initial large-scale language model based on the multiple training samples to obtain a large-scale language model.

[0188] Each training sample includes: sample conversation information, a conversation summary sample corresponding to the sample conversation information, and prompt information corresponding to the sample conversation information. The conversation summary sample includes: summary information in the sample conversation information, and field information of at least one sample key field in the sample conversation information.

[0189] Specifically, training samples are constructed from various text conversations (if audio is present, such as from phone calls or meetings, Automatic Speech Recognition (ASR) is used for text). Multiple prompts are constructed based on the collected conversation text. The prompts first explain the task, followed by detailed requirements. Related tasks include conversation summaries, key fields, and high-quality FAQs. Subsequent tasks are expandable, such as summary titles.

[0190] For example, let's define conversation information as D, prompt information as P, and conversation summaries, key fields, and high-quality FAQs as T1, T2, and T3, respectively. Capital letters represent sets. A set of conversations d in D corresponds to the answers to t1, t2, and t3. A requirement in P corresponds to any combination of {t1, t2, t3}. Training data is constructed based on these combinations.

[0191] For example, the corresponding p for a set of conversations d might include: Please generate a work order summary based on the following phone conversation content. Requirements: 1. Fluent expression with clear key points; 2. Summarize the entire conversation first; 3. Then summarize one or two key words; 4. Extract key fields mentioned in the conversation, including: vehicle model, inquiry question, contact person, owner, vehicle frame number, dealer, etc. The output corresponding to the conversation summary and key fields is the data of t1 and t2.

[0192] For example, the corresponding p for a conversation group d might include: Given the following conversation record between XXXX customer service and a customer, please summarize the conversation. Requirements: 1. Retain key information; 2. Extract high-quality questions and answers. The output of the conversation summary and high-quality FAQ is data at t1 and t3.

[0193] Furthermore, the training loss of the large-scale language model is determined based on the difference between the conversation summary sample corresponding to each training sample and the conversation summary generated by the model.

[0194] Specifically, training is performed using the training examples described in the previous examples. During training, the model iteratively adjusts parameter values ​​until it can correctly predict the next token based on a sequence of previous input tokens. To achieve this, the model uses self-learning techniques that teach it to adjust parameters to maximize the probability of correctly predicting the next token given the training examples. After fine-tuning, the LLM can be easily adapted to perform multiple tasks using relatively small supervised datasets.

[0195] Furthermore, through the training method shown in the above embodiment, a large-scale language model can be trained to generate summaries. In order to improve the quality of summary generation, multiple candidate large-scale language models can be trained to determine a large-scale language model for summarizing the target dialogue information. The model architectures of the candidate large-scale language models can be different, that is, models with different model architectures are trained to obtain candidate large-scale language models; or, the candidate large-scale language models can be models trained at different stages (each candidate large-scale language model can be trained using a different sample size for the initial model).

[0196] Specifically, a large-scale language model is obtained by: obtaining multiple trained candidate large-scale language models; determining the performance of each candidate large-scale language model; and determining a trained large-scale language model from among the candidate large-scale language models based on the performance of each candidate large-scale language model. In this embodiment of the present application, the candidate large-scale language model with the highest performance is determined from among the candidate large-scale language models and used as the large-scale language model for subsequent summarization of the target conversation information.

[0197] Specifically, the performance of any candidate large-scale language model is determined in the following manner: obtain answers to frequently asked questions; for each question, generate model input information corresponding to the question based on the question; for each question, input the model input information corresponding to the question into any target large-scale language model to obtain a predicted answer, and perform similarity matching between the predicted answer and the answer corresponding to the question to obtain a matching result; based on the matching results corresponding to each question, determine the performance of any large-scale language model.

[0198] The FAQ includes at least one question and an answer corresponding to each question; the model input information includes at least one of dialogue information and summary information. In an embodiment of the present application, dialogue format information and summary format information corresponding to each question in the FAQ are generated and respectively input into any target large-scale language model to obtain a predicted answer. The predicted answer is then matched with the answer corresponding to the question in the FAQ. If the matching result indicates a high degree of match, the corresponding large-scale language model is determined to have high performance.

[0199] Furthermore, whether the answers to the FAQs can be obtained by using the original conversation and the conversation summary (abstract) as input can better promote the improvement of training data.

[0200] The following is a method for generating a conversation summary using a specific example, taking the conversation information as a telephone conference audio as an example. Figure 3aAs shown, the called or input conference call audio is converted into dialogue text through ASR and input into LLM. Figure 3a The article gives examples of prompt information in two different scenarios, from which the target prompt information is determined, and then the target prompt information is also input into the LLM. The corresponding conversation summary can be output through the LLM. The conversation summary is the conversation summary of the current request and is output in the form of a text paragraph; the field information of the key fields is also output. Specifically, while outputting the field information of the key fields of the current request, if there are keyword prompts in the target prompt information, the key fields in the conversation information can be heuristically extracted. Furthermore, high-quality FAQs can be extracted, that is, customer service scripts and customer problems such as customer service questions and answers, customer questions and answers can be extracted from the conversation content, which can be used for subsequent analysis of customer hot issues and construction of a customer service script library to optimize customer service robots.

[0201] It should be noted that the input dialogue and target prompt information of the LLM may be as shown in FIG3 .

[0202] The following combination Figure 3b and Figure 3c The implementation process of an optional embodiment of the present application is described, and the implementation process includes a model training phase and a model application phase. Figure 3b As shown, the model application stage may include: step S11, step S12 and step S13, as shown in FIG. Figure 3c As shown, the model training phase may include: step S21 and step S22. The steps of the above two phases are introduced below. Among them, the model application phase is as follows:

[0203] Step S11: Acquire target dialogue information input by the target object.

[0204] Optionally, as an example, the target conversation information input by the target object may be as shown in Table 1 below, wherein part of the conversation information included in Table 1 is conversation information between a customer service representative and a customer in a telephone scenario.

[0205] Table 1

[0206]

[0207]

[0208] Step S12: Determine target prompt information corresponding to the target conversation information, wherein the target prompt information is used to prompt generation of a conversation summary based on target key fields in the target conversation information, wherein the key fields include at least one first key field.

[0209] Specifically, the target prompt information corresponding to the target dialogue information can be input by the target object into the human-computer interaction interface corresponding to the LLM, or can be imported into the human-computer interaction interface through other means. The following table 2 is used as an example to obtain the target prompt information for the above table 2, where:

[0210] Table 2

[0211]

[0212] Among them, the first key fields in Table 2 above are "car model", "consulting question", "contact person", "owner", "frame number" and "dealer".

[0213] Step S13: Based on the target dialogue information and the target prompt information, and using the trained large-scale language model, generate summary information of the target dialogue information, wherein the summary information includes: summary information of the target dialogue information, field information of each first key field, and field information of at least one second key field obtained by the heuristic.

[0214] In an embodiment of the present application, the target dialogue information and target prompt information are input into the large-scale language model to obtain summary information including the target dialogue information, field information of each first key field, and field information of at least one second key field obtained by the heuristic method, and the summary information is output in the human-computer interaction interface. The summary information generated based on Table 1 and Table 2 above can be shown in Table 3, where:

[0215] Table 3

[0216]

[0217] Among them, "Contact number: 138XXXXXXXX" is the field information of the second key field obtained heuristically.

[0218] Continue, as Figure 3c As shown, the model training phase may include steps S21 and S22, wherein:

[0219] Step S21: Acquire multiple training samples.

[0220] Each training sample includes: sample conversation information, a conversation summary sample corresponding to the sample conversation information, and prompt information corresponding to the sample conversation information. The conversation summary sample includes: summary information in the sample conversation information and field information of at least one sample key field in the sample conversation information.

[0221] Step S22: Based on multiple training samples, the initial large-scale language model is trained to obtain a trained large-scale language model.

[0222] The training loss of the large-scale language model is determined based on the difference between the conversation summary sample corresponding to each training sample and the conversation summary generated by the model.

[0223] It is understandable that the various optional embodiments provided in this application may be implemented individually or in combination when there is no conflict between the embodiments during actual implementation.

[0224] Based on the same principle as the method provided in the embodiment of the present application, the embodiment of the present application also provides a device for generating a conversation summary, such as Figure 4 As shown, the conversation summary generating device 40 may include:

[0225] An acquisition module 41 is used to acquire target dialogue information input by a target object;

[0226] A first determining module 42 is configured to determine target prompt information corresponding to the target conversation information, wherein the target prompt information includes key fields and is configured to prompt generation of a conversation summary based on the key fields, wherein the key fields include at least one first key field;

[0227] The model execution processing module 43 is configured to perform the following operations using the trained large-scale language model:

[0228] generating summary information of the target dialogue information based on the target dialogue information and the target prompt information; and

[0229] Extracting field information of each target key field from the target conversation information based on the target prompt information, where the target key field includes at least one first key field and at least one second key field, where the at least one second key field is obtained based on the at least one first key field;

[0230] The summary information of the target conversation information and the field information of each target key field are used as the summary information of the target conversation information.

[0231] In a possible implementation of the embodiment of the present application, the apparatus 40 further includes: a second determining module, wherein:

[0232] When determining at least one second key field, the second determining module is specifically configured to:

[0233] Get multiple preset fields;

[0234] Determining at least one second key field from each preset field based on at least one of a relevance of each preset field to the target domain or a similarity between each preset field and at least one first key field;

[0235] The target domain is the application domain to which the target dialogue information belongs.

[0236] In another possible implementation of the embodiment of the present application, when the model execution processing module 43 extracts field information of at least one second key field from the target dialogue information based on the target prompt information, it is specifically configured to:

[0237] Identify preset information for target prompt information;

[0238] If the target prompt information includes preset information, field information of at least one second key field is generated based on the target dialogue information and the target prompt information.

[0239] In another possible implementation of the embodiment of the present application, the second determination module, when determining at least one second key field from each preset field based on at least one of a correlation between each preset field and the target field or a similarity between each preset field and at least one first key field, is specifically configured to:

[0240] At least one preset field among the preset fields, which has a correlation with the target field greater than a preset correlation and a similarity with at least one first key field greater than a first preset similarity, is determined as at least one second key field.

[0241] In another possible implementation of the embodiment of the present application, the plurality of preset fields belong to a target field set corresponding to the target domain;

[0242] The second determining module is specifically configured to: when determining at least one second key field from each preset field based on the similarity between each preset field and the at least one first key field,

[0243] For each first key field, a field in the target field set having a similarity with the first key field greater than a second preset similarity is determined to obtain fields corresponding to each first key field as at least one second key field corresponding to the target conversation information.

[0244] In another possible implementation of the embodiment of the present application, the first determining module 42 may be configured to perform any of the following when determining the target prompt information corresponding to the target conversation information:

[0245] Obtaining configuration information related to the target prompt information input by the target object, and constructing the target prompt information corresponding to the target dialogue information based on the configuration information;

[0246] The application field corresponding to the target dialogue information is determined, prompt information related to the application field is determined from various pre-configured prompt information, and the determined prompt information is used as the target prompt information.

[0247] In another possible implementation of the embodiment of the present application, the configuration information is determined in the following manner:

[0248] In response to a prompt information construction instruction triggered by a target object received through the configuration interface, displaying configuration prompt information related to the application field corresponding to the target dialogue information;

[0249] The receiving target object inputs configuration information related to the target prompt information based on the configuration prompt information.

[0250] In another possible implementation of the embodiment of the present application, the large-scale language model is trained by the following method:

[0251] Acquire multiple training samples, where each training sample includes: sample conversation information, a conversation summary sample corresponding to the sample conversation information, and prompt information corresponding to the sample conversation information, wherein the conversation summary sample includes: summary information in the sample conversation information and field information of at least one sample key field in the sample conversation information;

[0252] Based on multiple training samples, an initial large-scale language model is trained to obtain a large-scale language model, wherein the training loss of the large-scale language model is determined based on the difference between the conversation summary sample corresponding to each training sample and the conversation summary generated by the model.

[0253] In another possible implementation of the embodiment of the present application, the large-scale language model is obtained by:

[0254] Obtain multiple trained candidate large-scale language models;

[0255] Determine the performance of each candidate large-scale language model;

[0256] Based on the performance of each candidate large-scale language model, a large-scale language model is determined from among the candidate large-scale language models.

[0257] In another possible implementation of the embodiment of the present application, the performance of any candidate large-scale language model is determined by:

[0258] Get FAQs, which include at least one question and the corresponding answer to each question;

[0259] For each question, based on the question, generate model input information corresponding to the question, where the model input information includes at least one of the dialogue information or the summary information;

[0260] For each question, input the model input information corresponding to the question into any target large-scale language model to obtain the predicted answer, and

[0261] Perform similarity matching between the predicted answer and the answer corresponding to the question to obtain a matching result;

[0262] Determine the performance of any large-scale language model based on the matching results corresponding to each question.

[0263] It should be noted that the first determination module 41 and the second determination module can be the same module or different modules, which is not limited in the embodiment of the present application.

[0264] An embodiment of the present application provides a conversation summary generation device. In this embodiment, after obtaining target conversation information input by a target object and determining target prompt information corresponding to the target conversation information, summary information for the target conversation information can be generated using a large-scale language model and based on the target prompt information. Furthermore, at least one second key field is heuristically obtained based on a first key field in the target prompt information, and field information of the first key field and the second key field are extracted from the target conversation information to obtain summary information of the target conversation information based on the summary information of the target conversation information, the field information of the first key field, and the field information of the second key field. That is, in this embodiment of the present application, the generated summary information not only includes a summary of the target conversation information and the field information of the first key field, but also heuristically obtains the field information of the second key field. Thus, with a single prompt information, higher-quality summary information that meets the requirements of the target object can be generated, thereby generating an experience for the target object.

[0265] The device of the embodiment of the present application can execute the method provided by the embodiment of the present application, and its implementation principle is similar. The actions performed by each module in the device of each embodiment of the present application correspond to the steps in the method of each embodiment of the present application. For the detailed functional description of each module of the device, please refer to the description in the corresponding method shown in the previous text, and will not be repeated here.

[0266] An embodiment of the present application further provides an electronic device, comprising at least one processor configured to execute the steps of the method provided in any optional embodiment of the present application. Optionally, the electronic device may further comprise a transceiver and / or a memory coupled to the processor, wherein the memory stores a computer program, and when the processor executes the computer program, the solution provided in any optional embodiment of the present application may be implemented. Optionally, the electronic device may be a user terminal or a server.

[0267] Figure 5 A schematic diagram of the structure of an electronic device applicable to the embodiment of the present application is shown in FIG. Figure 5 As shown, for example, the electronic device may be a server or a user terminal, and the electronic device may be used to implement the method provided in any embodiment of the present application.

[0268] like Figure 5 As shown in FIG, the electronic device 2000 may mainly include at least one processor 2001 ( Figure 5 1 ), memory 2002, communication module 2003 and input / output interface 2004 and other components, optionally, the components can be connected and communicated through bus 2005. It should be noted that, Figure 5 The structure of the electronic device 2000 shown in the figure is merely illustrative and does not constitute a limitation on the electronic devices to which the method provided in the embodiments of the present application is applicable.

[0269] Memory 2002 can be used to store operating systems and application programs, etc. Application programs can include computer programs that implement the methods described in the embodiments of the present invention when called by processor 2001, and can also include programs for implementing other functions or services. Memory 2002 can be ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices that can store information and computer programs, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited to these.

[0270] The processor 2001 is connected to the memory 2002 via the bus 2005 and implements corresponding functions by calling the application program stored in the memory 2002. The processor 2001 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof, which can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the present disclosure. The processor 2001 can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0271] The electronic device 2000 can be connected to a network via a communication module 2003 (which may include, but is not limited to, components such as a network interface) to communicate with other devices (such as a user terminal or a server) via the network to implement data interaction, such as sending data to or receiving data from other devices. The communication module 2003 may include a wired network interface and / or a wireless network interface, etc., that is, the communication module may include at least one of a wired communication module and a wireless communication module.

[0272] The electronic device 2000 can be connected to the required input / output devices, such as a keyboard, a display device, etc., through the input / output interface 2004. The electronic device 2000 itself can have a display device, and can also be connected to other external display devices through the interface 2004. Optionally, a storage device, such as a hard disk, can also be connected through the interface 2004, so that data in the electronic device 2000 can be stored in the storage device, or data in the storage device can be read, and data in the storage device can also be stored in the memory 2002. It can be understood that the input / output interface 2004 can be a wired interface or a wireless interface. Depending on the actual application scenario, the device connected to the input / output interface 2004 can be a component of the electronic device 2000, or it can be an external device connected to the electronic device 2000 when needed.

[0273] Bus 2005, used to connect the various components, may include a path for transmitting information between the components. Bus 2005 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. Depending on their function, bus 2005 may be categorized as an address bus, a data bus, a control bus, or the like.

[0274] Optionally, for the solution provided in the embodiment of the present application, the memory 2002 can be used to store a computer program for executing the solution of the present application, and be run by the processor 2001. When the processor 2001 runs the computer program, the actions of the method or device provided in the embodiment of the present invention are implemented.

[0275] Based on the same principle as the method provided in the embodiment of the present application, the embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the corresponding content of the aforementioned method embodiment can be implemented.

[0276] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the corresponding content of the aforementioned method embodiment can be implemented.

[0277] It should be noted that the terms "first," "second," "third," "fourth," "1," "2," etc. (if any) in the specification and claims of this application and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than that shown or described in the drawings.

[0278] It should be understood that, although each operation step is indicated by arrows in the flowchart of the embodiment of the present application, the order of implementation of these steps is not limited to the order indicated by the arrows. Unless otherwise clearly stated herein, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flowchart can be performed in other orders according to demand. In addition, some or all of the steps in each flowchart can include multiple sub-steps or multiple stages based on actual implementation scenarios. Some or all of these sub-steps or stages can be executed at the same time, and each sub-step or stage in these sub-steps or stages can also be executed at different times respectively. Under different scenarios at the execution time, the execution order of these sub-steps or stages can be flexibly configured according to demand, and the embodiment of the present application does not limit this.

[0279] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0280] The above description is only an optional implementation method for some implementation scenarios of this application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the solution of this application, the use of other similar implementation methods based on the technical ideas of this application also falls within the protection scope of the embodiments of this application.

Claims

1. A method for generating a conversation summary, characterized in that: The method comprises: Get the target dialogue information input by the target object; Determining target prompt information corresponding to the target conversation information, wherein the target prompt information includes key fields and is used to prompt generation of a conversation summary based on the key fields, wherein the key fields include at least one first key field; Large-scale language models can do the following: generating summary information of the target dialogue information based on the target dialogue information and the target prompt information; and Extracting, based on the target prompt information, field information of each target key field from the target conversation information, the target key field including the at least one first key field and at least one second key field, the at least one second key field being obtained based on the at least one first key field; The summary information of the target conversation information and the field information of each target key field are used as the summary information of the target conversation information.

2. The method according to claim 1, characterized in that The method of determining the at least one second key field includes: Get multiple preset fields; Determining at least one second key field from each of the preset fields based on at least one of a relevance between each of the preset fields and the target domain or a similarity between each of the preset fields and at least one of the first key fields; The target domain is the application domain to which the target dialogue information belongs.

3. The method according to claim 2, characterized in that Extracting field information of at least one second key field from the target dialogue information based on the target prompt information includes: Identifying preset information for the target prompt information; If the target prompt information includes the preset information, field information of the at least one second key field is generated based on the target dialogue information and the target prompt information.

4. The method according to claim 2, characterized in that The determining at least one second key field from each of the preset fields based on at least one of the relevance of each of the preset fields and the target domain or the similarity of each of the preset fields and at least one of the first key fields includes: Among the preset fields, at least one preset field having a correlation with the target field greater than a preset correlation and a similarity with at least one of the first key fields greater than a first preset similarity is determined as at least one of the second key fields.

5. The method according to claim 2, characterized in that The plurality of preset fields belong to a target field set corresponding to the target field; Determining at least one second key field from each of the preset fields based on similarity between each of the preset fields and at least one of the first key fields includes: For each first key field, a field in the target field set having a similarity with the first key field greater than a second preset similarity is determined to obtain a field corresponding to each first key field as at least one second key field corresponding to the target conversation information.

6. The method according to claim 1, characterized in that The determining of the target prompt information corresponding to the target dialogue information includes any one of the following: Acquire configuration information related to the target prompt information input by the target object, and construct target prompt information corresponding to the target dialogue information based on the configuration information; An application field corresponding to the target dialogue information is determined, prompt information related to the application field is determined from various pre-configured prompt information, and the determined prompt information is used as the target prompt information.

7. The method according to claim 6, characterized in that The configuration information is determined in the following manner: In response to a prompt information construction instruction triggered by the target object received through the configuration interface, displaying configuration prompt information related to the application field corresponding to the target dialogue information; The target object receives configuration information related to the target prompt information based on the configuration prompt information.

8. The method according to claim 1, characterized in that The large-scale language model is trained in the following way: Acquire multiple training samples, wherein each training sample includes: sample conversation information, a conversation summary sample corresponding to the sample conversation information, and prompt information corresponding to the sample conversation information, wherein the conversation summary sample includes: summary information in the sample conversation information and field information of at least one sample key field in the sample conversation information; Based on the multiple training samples, an initial large-scale language model is trained to obtain the large-scale language model, wherein the training loss of the large-scale language model is determined based on the difference between the conversation summary sample corresponding to each of the training samples and the conversation summary generated by the model.

9. The method according to any one of claims 1 to 8, characterized in that The large-scale language model is obtained by: Obtain multiple trained candidate large-scale language models; determining the performance of each of the candidate large-scale language models; The large-scale language model is determined from the candidate large-scale language models based on performance of the candidate large-scale language models.

10. The method according to claim 9, characterized in that The performance of any candidate large-scale language model is determined in the following way: Obtaining answers to frequently asked questions, wherein the answers to frequently asked questions include at least one question and an answer to each question; For each of the questions, based on the question, generating model input information corresponding to the question, the model input information including at least one of the dialogue information or the summary information; For each of the questions, input the model input information corresponding to the question into any of the target large-scale language models to obtain a predicted answer, and Perform similarity matching between the predicted answer and the answer corresponding to the question to obtain a matching result; Based on the matching results corresponding to each question, the performance of any of the large-scale language models is determined.

11. A conversation summary generating device, characterized in that: The device comprises: An acquisition module is used to obtain target dialogue information input by a target object; A first determining module is configured to determine target prompt information corresponding to the target conversation information, wherein the target prompt information includes key fields and is configured to prompt generation of a conversation summary based on the key fields, wherein the key fields include at least one first key field; The model execution processing module is used to perform the following operations using a large-scale language model: generating summary information of the target dialogue information based on the target dialogue information and the target prompt information; and Extracting, based on the target prompt information, field information of each target key field from the target conversation information, the target key field including the at least one first key field and at least one second key field, the at least one second key field being obtained based on the at least one first key field; The summary information of the target conversation information and the field information of each target key field are used as the summary information of the target conversation information.

12. An electronic device, characterized in that: The electronic device includes a memory and a processor, wherein a computer program is stored in the memory, and the processor executes the method for generating a conversation summary according to any one of claims 1 to 10 when running the computer program.

13. A computer-readable storage medium, characterized in that The storage medium stores a computer program, which, when executed by a processor, implements the method for generating a conversation summary according to any one of claims 1 to 10.

14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method for generating a conversation summary according to any one of claims 1 to 10 are implemented.

Images