Sample generation method based on large model, training method of large model and dialogue method

Abstract

The invention provides a sample generation method based on a large model, and relates to the technical field of artificial intelligence, in particular to the technical fields of large language models, agents, human-computer interaction and the like. The sample generation method comprises the steps that a dialogue large model is called to process simulation input information and simulation auxiliary dialogue information, multiple different pieces of simulation reply information are obtained, the simulation auxiliary dialogue information is used for representing the dialogue intention and the personalized intention of the simulation input information, and the multiple pieces of simulation reply information are all used for replying the simulation input information; based on the simulation input information and the simulation auxiliary dialogue information, the multiple pieces of simulation reply information are evaluated, and multiple evaluation results are obtained; and obtaining a target training sample based on the plurality of evaluation results, the simulation auxiliary dialogue information, the simulation input information and the plurality of pieces of simulation reply information. The invention further provides a large model training method and device, a dialogue method and device, an intelligent agent, electronic equipment, a storage medium and a program product.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, in particular to large language models, intelligent agents, human-computer interaction and other technical fields, and can be applied to online education, e-commerce platforms, financial services, medical consultation, hotel and tourism services and other scenarios that require interaction. More specifically, the present disclosure provides a sample generation method based on a large model, a training method for a large model, a dialogue method, a device, an intelligent agent, an electronic device, a storage medium and a program product. Background Art

[0002] With the rapid development of artificial intelligence technology, pre-trained language models based on large models have promoted the development of dialogue systems to a higher level, making them widely used in many fields. How to further improve the flexibility and intelligence of dialogue systems is still a direction worth exploring. Summary of the invention

[0003] The present disclosure provides a sample generation method based on a large model, a training method of a large model, a dialogue method, an apparatus, an intelligent agent, an electronic device, a storage medium and a program product.

[0004] According to one aspect of the present disclosure, a sample generation method based on a large model is provided, comprising: calling a dialogue large model to process simulation input information and simulation-assisted dialogue information to obtain multiple different simulation reply information, wherein the simulation-assisted dialogue information is used to characterize the dialogue intent and personalized intent of the simulation input information, and the multiple simulation reply information are all used to reply to the simulation input information; based on the simulation input information and the simulation-assisted dialogue information, the multiple simulation reply information are respectively evaluated to obtain multiple evaluation results; and based on the multiple evaluation results, the simulation-assisted dialogue information, the simulation input information and the multiple simulation reply information, a target training sample is obtained.

[0005] According to another aspect of the present disclosure, a large model training method is provided, including training the large model using target training samples, wherein the target training samples are generated according to a large model-based sample generation method as disclosed in the present disclosure.

[0006] According to another aspect of the present disclosure, a conversation method is provided, comprising: calling a target big model to process input information and auxiliary conversation information to obtain reply information for replying to the input information, wherein the auxiliary conversation information is used to characterize the conversation intent and personalized intent of the input information; wherein the target big model is trained according to the training method of the big model as disclosed in the present disclosure.

[0007] According to another aspect of the present disclosure, a sample generation device based on a large model is provided, including: a processing module, which is used to call the dialogue large model to process simulation input information and simulation-assisted dialogue information to obtain multiple different simulation reply information, wherein the simulation-assisted dialogue information is used to characterize the dialogue intention and personalized intention of the simulation input information, and the multiple simulation reply information are all used to reply to the simulation input information; an evaluation module, which is used to evaluate the multiple simulation reply information respectively based on the simulation input information and the simulation-assisted dialogue information to obtain multiple evaluation results; and a sample determination module, which is used to obtain a target training sample based on the multiple evaluation results, the simulation-assisted dialogue information, the simulation input information and the multiple simulation reply information.

[0008] According to another aspect of the present disclosure, a large model training device is provided, including: a training module for training the large model using target training samples, wherein the target training samples are generated according to a sample generation device based on the large model as disclosed in the present disclosure.

[0009] According to another aspect of the present disclosure, a dialogue device is provided, including: a dialogue processing module, used to call a target big model to process input information and auxiliary dialogue information, and obtain reply information for replying to the input information, wherein the auxiliary dialogue information is used to characterize the dialogue intent and personalized intent of the input information; wherein the target big model is trained according to a training device such as the big model of the present disclosure.

[0010] According to another aspect of the present disclosure, an intelligent agent is provided, which is configured to execute the method of the present disclosure.

[0011] According to another aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a method as disclosed in the present disclosure.

[0012] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause the computer to execute the method of the present disclosure.

[0013] According to another aspect of the present disclosure, a computer program product is provided, comprising a computer program, wherein the computer program implements the method of the present disclosure when executed by a processor.

[0014] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0016] Figure 1 Schematically illustrates an exemplary system architecture to which a sample generation method, a large model training method, a dialogue method, and related devices according to an embodiment of the present disclosure can be applied;

[0017] Figure 2 The flowchart of the sample generation method according to the embodiment of the present disclosure is schematically shown;

[0018] Figure 3 A schematic diagram of generating a target training sample according to an embodiment of the present disclosure is schematically shown;

[0019] Figure 4 The flowchart of performing multi-objective evaluation according to an embodiment of the present disclosure is schematically shown;

[0020] Figure 5 A schematic diagram schematically shows a fine-tuning sample set according to an embodiment of the present disclosure;

[0021] Figure 6 A flowchart of a large model training method according to an embodiment of the present disclosure is schematically shown;

[0022] Figure 7 A flowchart of a conversation method according to an embodiment of the present disclosure is schematically shown;

[0023] Figure 8 A block diagram schematically shows a sample generating device according to an embodiment of the present disclosure; and

[0024] Fig. 9 A block diagram of a large model training device according to an embodiment of the present disclosure is schematically shown;

[0025] Fig.10 A block diagram of a conversation device according to an embodiment of the present disclosure is schematically shown;

[0026] Fig.11 The structure diagram of the intelligent agent according to the embodiment of the present disclosure is schematically shown;

[0027] Fig.12 A schematic block diagram of an example electronic device that can be used to implement embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0028] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0029] In the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision, disclosure and application of user personal information involved comply with the provisions of relevant laws and regulations, take necessary confidentiality measures, and do not violate public order and good morals.

[0030] In the technical solution of the present disclosure, the user's authorization or consent is obtained before obtaining or collecting the user's personal information.

[0031] Figure 1 An exemplary system architecture to which a sample generation method, a large model training method, a dialogue method, and related devices can be applied according to an embodiment of the present disclosure is schematically illustrated.

[0032] It should be noted that Figure 1 The examples shown are only examples of system architectures to which the embodiments of the present disclosure can be applied, in order to help those skilled in the art understand the technical content of the present disclosure, but do not mean that the embodiments of the present disclosure cannot be used in other devices, systems, environments or scenarios. For example, in another embodiment, the exemplary system architecture to which the sample generation method, large model training method, dialogue method and related devices can be applied may include a terminal device, but the terminal device may implement the sample generation method and device provided by the embodiments of the present disclosure without interacting with the server.

[0033] like Figure 1 As shown, in the application scenario 100, the terminal device 102 of the client can communicate with the server 103 of the service end via the network. The user 101 can send a dialogue request through the interactive interface provided by the terminal device 102, and the terminal device 102 can send the dialogue request to the server 103, so that the server 103 calls the intelligent agent to have a dialogue with the user 101 and outputs a dialogue reply. The server 103 sends the dialogue reply to the terminal device 102, so that the terminal device 102 displays the dialogue reply to the user 101 through the interactive interface.

[0034] The terminal device 102 may be configured with various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers, etc. Various communication client applications may be installed on the terminal device 102, such as instant messaging tools, smart assistant applications, search applications, knowledge reading applications, web browser applications, email clients, and / or social platform software, etc. (only as examples). The user 101 may input a dialogue request in the interactive interface of these client applications, and these client applications will display the generated dialogue response to the user 101.

[0035] Server 103 may be a server that provides various services, such as a background management server that provides support for the content browsed by user 101 using the interactive interface of terminal device 102 (for example only). The background management server may call an intelligent agent to generate an object for dialogue with user 101 in response to the dialogue request received from user 101, and make a dialogue reply according to the dialogue content of the dialogue request, feed back the dialogue reply to terminal device 102, and display it through the interactive interface. Server 103 may also be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). Server 103 may also be a server of a distributed system, or a server combined with blockchain.

[0036] It should be noted that the sample generation method, large model training method, and dialogue method provided in the embodiments of the present disclosure can generally be executed by the server 103. Accordingly, the sample generation device, large model training device, and dialogue device provided in the embodiments of the present disclosure can also be set in the server 103. The sample generation method, large model training method, and dialogue method provided in the embodiments of the present disclosure can also be executed by a server or server cluster that is different from the server 103 and can communicate with the terminal device 102 and / or the server 103. Accordingly, the sample generation device, large model training device, and dialogue device provided in the embodiments of the present disclosure can also be set in a server or server cluster that is different from the server 103 and can communicate with the terminal device 102 and / or the server 103.

[0037] Alternatively, when the terminal device 102 also has computing power and computing resources that meet the requirements, the sample generation method, large model training method, and dialogue method provided in the embodiments of the present disclosure can generally be executed by the terminal device 102. Accordingly, the sample generation device, large model training device, and dialogue device provided in the embodiments of the present disclosure can generally be set in the terminal device 102.

[0038] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0039] It should be noted that the sequence numbers of the operations in the following method are only used as representations of the operations for the purpose of description, and should not be regarded as representing the execution order of the operations. Unless explicitly stated, the method does not need to be executed completely in the order shown.

[0040] In the related art, traditional dialogue systems can be roughly divided into rule-based dialogue systems, retrieval-based dialogue systems, generative dialogue systems, and hybrid dialogue systems.

[0041] Rule-based dialogue systems respond to user input through predefined rules and flowcharts. Typical implementations include decision trees and state machines. Although the rules are simple and easy to implement, they have limited processing capabilities for complex dialogues and multi-round interactions, cannot accurately understand the user's intentions, and are prone to falling into unpredictable dialogue branches. Such systems are suitable for handling simple, repetitive problems, but perform poorly when faced with open-ended or changeable dialogues.

[0042] Retrieval-based dialogue systems match user input with pre-prepared question-answer pairs and directly retrieve the most relevant responses from a preset library. Such systems rely on massive data accumulation and the design of dialogue templates, and are suitable for dealing with common problems and relatively standardized business processes. However, when users' questions go beyond the preset scope or are expressed in a more complex way, such systems are often powerless.

[0043] Generative dialogue systems based on deep learning use large-scale neural networks (such as RNN, LSTM, Transformer, etc.) to generate responses related to user input. Such systems are more flexible in semantic understanding and generation, and can cope with the needs of open-domain dialogues, especially in multi-round dialogue scenarios, showing good context understanding capabilities. However, generative systems also face problems such as unstable quality of generated results, semantic deviation, repeated or irrelevant content, and have high requirements for computing resources and training data.

[0044] Hybrid dialogue system In order to make up for the shortcomings of a single dialogue system, some existing solutions adopt a hybrid architecture, combining rule-based, retrieval, and generation methods to give full play to their respective advantages. For example, use rules or retrieval methods to handle standard questions, and generative methods to deal with complex or open-ended dialogues. Although the hybrid system has improved the accuracy and flexibility of responses to a certain extent, the architecture is complex, the maintenance cost is high, and there is still the problem of poor management of multi-round dialogues.

[0045] The present disclosure provides a sample generation method based on a large model, comprising: calling a dialogue large model to process simulation input information and simulation-assisted dialogue information to obtain multiple different simulation reply information, wherein the simulation-assisted dialogue information is used to characterize the dialogue intention and personalized intention of the simulation input information, and the multiple simulation reply information are all used to reply to the simulation input information; based on the simulation input information and the simulation-assisted dialogue information, the multiple simulation reply information are respectively evaluated to obtain multiple evaluation results; and based on the multiple evaluation results, the simulation-assisted dialogue information, the simulation input information and the multiple simulation reply information, a target training sample is obtained.

[0046] By calling the big dialogue model and combining the simulation input information and the simulation auxiliary dialogue information, multiple different simulation reply information is generated, and then the multiple simulation reply information is further evaluated. By combining the evaluation results, simulation auxiliary dialogue information, simulation input information and simulation reply information, the target training samples are generated. In this way, the big model not only has a rich training set, but also can use the evaluation results and simulation reply information to fully learn the relevant knowledge of the adaptation of simulation reply information to dialogue intentions and personalized intentions, so that the big model can screen and identify replies that are more in line with dialogue intentions and personalized intentions, thereby improving the intelligence of the dialogue.

[0047] Figure 2 The flowchart of the sample generation method according to the embodiment of the present disclosure is schematically shown.

[0048] like Figure 2 As shown, the method includes operations S210 to S230.

[0049] In operation S210, the large dialogue model is called to process the simulation input information and the simulation auxiliary dialogue information to obtain different multiple simulation reply information.

[0050] In operation S220, based on the simulation input information and the simulation auxiliary dialogue information, multiple simulation reply information are evaluated respectively to obtain multiple evaluation results.

[0051] In operation S230, a target training sample is obtained based on the plurality of evaluation results, the simulation auxiliary dialogue information, the simulation input information, and the plurality of simulation reply information.

[0052] According to an embodiment of the present disclosure, the simulation input information may be information that the simulation object may input in a simulated real scene or situation. The simulation input information may be preset manually or generated by a simulation model based on the historical dialogue data and behavior pattern of the simulation object.

[0053] According to an embodiment of the present disclosure, the simulation-assisted dialogue information is used to characterize the dialogue intent and personalized intent of the simulation input information. The dialogue intent may be, for example, a query intent, a knowledge question-and-answer intent, or a small talk intent. The personalized intent may be, for example, the preferences, emotional state, or specific needs of the simulation object.

[0054] According to the embodiments of the present disclosure, multiple simulation reply information are used to reply to the simulation input information. Based on the randomness and diversity of the dialogue model and the sensitivity to different situations and simulation input information, multiple different simulation reply information can be generated when processing the same simulation input information and simulation auxiliary dialogue information.

[0055] For example, the simulated input information is "I want to buy a mobile phone"; the dialogue intent of the simulated auxiliary dialogue information is "query for products", and the personalized intent includes "prefer XX brand", "have requirements for mobile phone performance", etc. After using the large dialogue model to process the simulated input information and simulated auxiliary dialogue information, multiple simulated reply information is obtained, for example, reply 1: "Hello, we have XX brand mobile phones here, you can take a look at this latest model." Reply 2: "Hello, we have multiple brands of mobile phones for you to choose from, you can take a look at these models." Reply 3: "Hello, we have multiple brands of mobile phones for you to choose from, including XX brand. Do you have any specific requirements for the performance of the mobile phone?"

[0056] According to an embodiment of the present disclosure, after obtaining multiple simulation reply information, each simulation reply information is evaluated according to the simulation input information and the simulation auxiliary dialogue information to judge the reply quality of each simulation reply information. After evaluating each simulation reply information, an evaluation result corresponding to each simulation reply information is obtained. The evaluation result can be expressed in the form of a score, a grade, a label or other forms.

[0057] For example, response 1: high accuracy, strong relevance, but average personalization, no further guidance for users, score 75. Response 2: average accuracy, weak relevance, score 70. Response 3: high accuracy, strong relevance, high personalization, and guidance for users to further explain their needs, score 85.

[0058] According to an embodiment of the present disclosure, by calling a large dialogue model and combining simulated input information and simulated auxiliary dialogue information, a plurality of different simulated reply information are generated, and then the plurality of simulated reply information are further evaluated, and by combining the evaluation results, simulated auxiliary dialogue information, simulated input information and simulated reply information, target training samples are generated, so that the large model not only has a rich training set, but also can use the evaluation results and simulated reply information to fully learn the relevant knowledge of the adaptation of simulated reply information to dialogue intentions and personalized intentions, so that the large model can screen and identify replies that are more in line with dialogue intentions and personalized intentions, thereby improving the intelligence of the dialogue.

[0059] The following will be combined Figure 3~Figure 6 The information interaction method provided by the present disclosure is described in detail.

[0060] Figure 3 A schematic diagram of generating target training samples according to an embodiment of the present disclosure is schematically shown.

[0061] like Figure 3 As shown, simulation input information 320 and simulation auxiliary dialogue information 330 can be obtained based on the simulation object 310. The simulation input information 320 and the simulation auxiliary information 330 are input into the dialogue large model 340, and multiple simulation reply information 350 can be output. The multiple simulation reply information 350 are evaluated respectively to obtain multiple evaluation results 360. According to the simulation input information 320, the simulation auxiliary information 330, the multiple simulation reply information 350 and the multiple evaluation results 360, the target training set 370 is obtained.

[0062] According to an embodiment of the present disclosure, when performing Figure 2 The operation S210 shown also includes obtaining simulation auxiliary dialogue information based on the simulation object portrait of the simulation object and the simulation dialogue context information of the simulation input information, wherein the simulation input information is obtained by simulating the simulation object.

[0063] In some exemplary embodiments, the simulation object may be an individual user, an enterprise user, an industry professional, or a group organization, etc. The simulation object has multiple personality characteristics such as identity attributes, living habits, behavioral habits, language habits, personal preferences, etc., and the multiple personality characteristics may constitute a simulation object portrait of the simulation object.

[0064] In some exemplary embodiments, the simulated input information of the simulated object in the current application scenario or context can be simulated through the simulated object portrait, the conversation intention of the current simulated input information can be determined based on the simulated conversation context information of the current simulated input information, and the personalized intention of the current simulated object can be determined based on the simulated object portrait.

[0065] According to the embodiments of the present disclosure, by integrating object portraits into the dialogue generation process, the large model can better understand and predict user needs, thereby providing dialogue content and suggestions that meet the user's personalized needs.

[0066] According to an embodiment of the present disclosure, a simulation object portrait can be obtained by: determining the simulation object's simulation history behavior information and the simulation object's attribute information to obtain the simulation object interest portrait. Based on the simulation object's simulation history conversation information, determining the simulation object interaction portrait. Based on the simulation object interest portrait and the simulation object interaction portrait, obtaining the simulation object portrait.

[0067] In some exemplary embodiments, the simulation history behavior information may be the search records, browsing history, purchase behavior, interest tags, feedback information and other activity information of the simulation object in the past, which may reflect the preferences, habits or behavior patterns of the simulation object. The simulation object attribute information includes the basic attributes of the simulation object, such as age, gender, occupation, hobbies, etc., which may reflect the background characteristics of the simulation object. By combining the simulation history behavior information and the simulation object attribute information, a simulation object interest portrait reflecting the interests and preferences of the simulation object may be constructed.

[0068] In some exemplary embodiments, the simulation history dialogue information may be the dialogue record information of the simulation object in the past with other people or related interactive systems. The simulation history dialogue information can reveal the communication style, language habits, emotional expression and other characteristics of the simulation object. Through the simulation history dialogue information, a simulation object interaction portrait reflecting the interaction characteristics and style of the simulation object can be constructed, which helps to predict the dialogue style and habits of the simulation object in future dialogues.

[0069] In some exemplary embodiments, after obtaining the simulation history behavior information and the simulation history conversation information, the portrait features of each simulation object can be constructed through steps such as data cleaning, feature extraction and modeling.

[0070] According to the embodiments of the present disclosure, by combining the simulation object interest portrait and the simulation interaction portrait, a comprehensive and detailed simulation object portrait can be obtained. The simulation object portrait not only includes the interests and preferences of the simulation object, but also includes the characteristics and style of the simulation object's dialogue, thereby improving the authenticity and reliability of the simulation input information.

[0071] According to an embodiment of the present disclosure, a dialogue big model is called to process simulation input information and simulation auxiliary dialogue information to obtain multiple different simulation reply information, including: generating prompt information of the dialogue big model based on the simulation input information and simulation auxiliary dialogue information; and calling the dialogue big model to process the prompt information to obtain multiple different simulation reply information.

[0072] In some exemplary embodiments, one or more prompt templates may be preset according to the requirements of the large dialogue model and the requirements of the current dialogue scene. Information extraction or integration is performed on the simulation input information and the simulation auxiliary dialogue information. The extraction may be performed by simple splicing or by using natural language processing technology (such as information extraction and semantic analysis). The extracted or integrated information should reflect the user's intention, the characteristics of the simulation scene, or related personalized needs. The extracted or integrated information is then inserted into the corresponding position of the prompt template to generate specific prompt information.

[0073] In some exemplary embodiments, the generated prompt information is input into the dialogue model, and the dialogue model can generate a plurality of different simulated response information according to the prompt information and in combination with its own internal knowledge and algorithms.

[0074] According to an embodiment of the present disclosure, prompt information is comprehensively generated based on simulated input information and simulated auxiliary dialogue information. Utilizing prompt information can help the dialogue model to more accurately understand input information and the user's intentions and personalized needs.

[0075] According to an embodiment of the present disclosure, calling a dialogue model to process prompt information and obtaining different multiple simulation reply information includes: performing an operation of inputting the prompt information into the dialogue model multiple times. Or performing an operation of updating the reply requirement information in the prompt information multiple times and inputting the updated prompt information into the dialogue model, wherein the reply requirement information is used to indicate the expected requirements for the simulation reply information.

[0076] In some exemplary embodiments, the prompt information can be kept unchanged, and the dialogue model can be called multiple times to process the prompt information. The randomness or uncertainty of the dialogue model can be used to allow the dialogue model to output multiple different responses.

[0077] In some exemplary embodiments, before calling the dialog model to process prompt information each time, the reply requirement information in the prompt information may be updated first, and the reply requirement information may include the length, style, content highlights, etc. of the reply. By updating the reply requirement information, the dialog model may be guided to generate simulated reply information of the corresponding requirements.

[0078] According to the embodiments of the present disclosure, the above two methods of obtaining different multiple simulation reply information respectively utilize the randomness and uncertainty of the large dialogue model and the controllability and flexibility of the reply request information, which helps to improve the adaptability and flexibility of the simulation reply information and can meet different simulation scenarios.

[0079] According to an embodiment of the present disclosure, calling the dialogue model to process prompt information includes: calling the dialogue model, identifying the interaction field of the prompt information, and determining the simulation interaction field. Inputting the prompt information into an expert model matching the simulation interaction field to obtain simulation reply information.

[0080] In some exemplary embodiments, corresponding expert models may be set for different interaction fields, and each expert model focuses on dialogue processing in its own specific field, and thus may have a faster response time and more accurate professional field capabilities.

[0081] In some exemplary embodiments, the expert model can be implemented by cutting and optimizing a large basic model. A gate selector can be set, and the input of the gate selector can be the simulation dialogue context information in the prompt information. Different expert models can be switched to respond according to the simulation dialogue context information, thereby providing efficient multi-round dialogue management capabilities.

[0082] According to the embodiments of the present disclosure, by setting an expert model that matches the simulation field, the accuracy and response speed of the dialogue can be effectively improved, while the adaptability to complex dialogue situations can be improved.

[0083] According to an embodiment of the present disclosure, based on the simulation input information and the simulation auxiliary dialogue information, multiple simulation reply information are evaluated respectively to obtain multiple evaluation results, including: inputting the simulation input information, the simulation auxiliary dialogue information and the simulation reply information into a multi-objective evaluation model to obtain multiple different evaluation index results. Based on the multiple evaluation index results, the evaluation result of the simulation reply information is obtained.

[0084] According to the embodiments of the present disclosure, multiple simulated response information can be respectively input into the multi-objective evaluation model with the simulated input information and the simulated auxiliary dialogue information to obtain evaluation results of each of the multiple simulated response information. The evaluation results of each of the multiple simulated response information can facilitate the large dialogue model to learn high-quality simulated response information.

[0085] In some exemplary embodiments, a multi-objective evaluation model (RM) can evaluate multiple evaluation indicators in the conversation process to obtain multiple evaluation indicator results. The multiple evaluation indicators can be, for example, user experience, chat continuation rate, conversion rate, etc. Multiple evaluation result indicators can be quantified. For example, user experience can be analyzed through natural language processing technology to analyze user feedback sentiment and quantify user satisfaction and experience quality. Chat continuation rate can be measured by quantifying whether the simulated reply information can effectively guide the user to continue the conversation and maintain user participation during the conversation. The conversion rate can be measured by whether the conversation can successfully guide the user to the expected business goal (such as purchase, registration, appointment, etc.), and the conversion effect can be scored.

[0086] In some exemplary embodiments, the results of multiple evaluation indicators may be weighted summed, weighted averaged, or otherwise comprehensively calculated to obtain a comprehensive evaluation result.

[0087] According to an embodiment of the present disclosure, multiple evaluation indicators are evaluated on the simulation response information through a multi-objective evaluation model, thereby helping the large model to fully understand the quality and effect of the simulation response information.

[0088] According to an embodiment of the present disclosure, a multi-objective evaluation model is obtained by training in the following manner: based on sample evaluation input information and sample evaluation labels matching the sample evaluation input information, an initial multi-objective evaluation model is trained to obtain a multi-objective evaluation model. The sample evaluation input information includes evaluation input information, evaluation auxiliary dialogue information, and evaluation reply information. The sample evaluation label represents the comprehensive evaluation result of the evaluation reply information for multiple evaluation indicators. The initial multi-objective evaluation model includes an initial indicator weight matrix and a pre-trained evaluation large model. The initial indicator weight matrix includes weight parameters of each of the multiple evaluation indicators.

[0089] In some exemplary embodiments, conversation samples may be collected from relevant historical conversation records or user interaction data sets, and the conversation samples may be cleaned and annotated to obtain evaluation input information, evaluation auxiliary conversation information, evaluation reply information, and corresponding sample evaluation labels. A simulation model may also be used to simulate the conversation between a real user and a human customer service representative to obtain conversation samples, and then the conversation samples may be cleaned and annotated to obtain evaluation input information, evaluation auxiliary conversation information, evaluation reply information, and corresponding sample evaluation labels.

[0090] According to an embodiment of the present disclosure, an initial multi-objective evaluation model is trained based on sample evaluation input information and a sample evaluation label that matches the sample evaluation input information to obtain a multi-objective evaluation model, including: inputting the sample evaluation input information into the initial multi-objective evaluation model to obtain a plurality of sample evaluation index results; obtaining a sample evaluation result based on the plurality of sample evaluation index results and an initial index weight matrix; and adjusting the initial index weights and respective parameters of the pre-trained evaluation large model based on the sample evaluation results and the sample evaluation labels to obtain a multi-objective evaluation model.

[0091] In some illustrative embodiments, sample input information, evaluation auxiliary dialogue information, and evaluation reply information can be input into a pre-trained evaluation large model, so that the pre-trained evaluation large model evaluates the evaluation reply information based on multiple evaluation indicators based on the sample input information and the evaluation auxiliary dialogue information, thereby obtaining sample evaluation indicator results for each of the multiple evaluation indicators. An initial indicator weight matrix can be used to weight the sum of multiple sample evaluation indicator results to obtain an overall sample evaluation result. By comparing the sample evaluation results and the corresponding sample evaluation labels, the initial indicator weights are adjusted to optimize the evaluation effect of the model. At the same time, the parameters of the pre-trained evaluation large model can also be adjusted to further improve the performance and accuracy of the model. After adjustment and optimization in the above steps, an improved multi-objective evaluation model is finally obtained.

[0092] In an illustrative embodiment, the importance of each evaluation indicator may be different based on different dialogue scenarios and business needs. The multi-objective evaluation model can dynamically adjust the impact of each evaluation target on the final score by adjusting the indicator weight, thereby achieving a balance of goals in different scenarios. For example, in a customer service scenario, user experience may be given a higher weight, while in a marketing scenario, conversion rate may account for a larger proportion. This flexible target weight setting enables the multi-objective evaluation model to be adjusted according to specific business needs.

[0093] Figure 4 A schematic diagram of a process for performing multi-objective evaluation according to an embodiment of the present disclosure is schematically shown.

[0094] like Figure 4 As shown, the sample evaluation input information 410 is input into the initial multi-objective evaluation model 420 to obtain multiple sample evaluation index results 430, and the multiple sample evaluation index results 430 and the corresponding multiple weight parameters in the weight matrix 440 are used for calculation, such as weighted summation, to obtain the sample evaluation result 450. The sample evaluation result 450 is compared with the sample evaluation label 460, so as to adjust the weight parameters of the weight matrix 440 and the parameters involved in the initial multi-objective evaluation model 420 until the multi-objective evaluation model is obtained.

[0095] According to the embodiments of the present disclosure, a multi-objective evaluation model is used to evaluate various key indicators in the dialogue process. This helps the big model to fully understand the quality and effect of the simulated reply information, and facilitates the guidance of the big model to generate dialogue content that meets the needs of users and the business goals of merchants. This provides clear feedback and adjustment directions for optimizing dialogue generation.

[0096] According to an embodiment of the present disclosure, a target training sample is obtained based on multiple evaluation results, simulation-assisted dialogue information, simulation input information and multiple simulation reply information, including: obtaining a dialogue sample tuple based on the simulation-assisted dialogue information, simulation input information, simulation reply information and the evaluation results corresponding to the simulation reply information; and obtaining a target training sample based on multiple dialogue sample tuples.

[0097] According to the embodiments of the present disclosure, by combining the simulation-assisted dialogue information, the simulation input information, the simulation reply information and the evaluation result, a dialogue sample tuple can be formed. By collecting a large number of dialogue sample tuples, target training samples for training can be obtained.

[0098] By constructing structured sample multi-tuples, we can effectively organize and manage training data, make conversation information more orderly and easier to manage, and improve data processing efficiency.

[0099] According to an embodiment of the present disclosure, in the operation Figure 2 The step S230 further includes: evaluating the fine-tuning samples in the fine-tuning sample set to obtain an evaluation result set of the fine-tuning sample set, wherein the fine-tuning samples include simulated input information, simulated auxiliary dialogue information and simulated reply information; and splitting the fine-tuning sample set based on the evaluation result set and the evaluation threshold to obtain multiple comparative training sample sets, wherein the evaluation threshold is used to indicate a splitting criterion for splitting the fine-tuning sample set based on the evaluation result set.

[0100] According to an embodiment of the present disclosure, the fine-tuning sample set may be based on the target training samples obtained in step S210 to step S230. The evaluation result set may be based on multiple evaluation results obtained in step S220. Using the evaluation result set and the evaluation threshold, the target training samples may be divided into multiple training sample sets with different dialogue qualities.

[0101] Figure 5 A schematic diagram of a fine-tuning sample set according to an embodiment of the present disclosure is schematically shown.

[0102] like Figure 5 As shown, the sample set 500 is fine-tuned, and the evaluation threshold is set to 80 points. Samples with scores above 80 points can be classified as high-quality samples 501, and samples with scores below 80 points can be classified as medium-quality samples 502. Both high-quality samples 501 and medium-quality samples 502 are used as training samples for comparative training. This allows the large model to distinguish and tend to learn the features of high-quality samples 502.

[0103] According to the embodiments of the present disclosure, through comparative learning, the influence of high-quality dialogue samples can be strengthened, so that the model can more accurately generate responses that meet user needs and merchant service standards, which can effectively solve the defects of the existing model in terms of unstable response quality.

[0104] Figure 6 A flowchart of a large model training method according to an embodiment of the present disclosure is schematically shown.

[0105] like Figure 6 As shown, the method includes operations S610 to S620.

[0106] In operation S610, a target training sample is acquired.

[0107] In operation S620, a large model is trained using the target training sample.

[0108] According to an embodiment of the present disclosure, the target training sample may be generated based on operations S210 to S230, the target training sample may be used to train the dialogue large model, or other large models may be trained using the target training sample.

[0109] In some exemplary embodiments, step S610 may not need to be executed, the target training sample may be a determined training sample, and step S620 may be directly executed to implement training of the large model.

[0110] The implementation method of training the large model can be to input the simulation input information and the simulation auxiliary dialogue information into the dialogue large model, output the sample reply information, use the evaluation results to annotate the simulation reply information, and adjust the parameters of the large model based on the similarity between the sample reply information and the simulation reply information.

[0111] By training the large model with target training samples, the large model can utilize the evaluation results and simulated response information obtained by including auxiliary dialogue information to fully learn the relevant knowledge of how the simulated response information is adapted to dialogue intentions and personalized intentions, thereby improving the intelligence of the dialogue.

[0112] According to an embodiment of the present disclosure, after step S610, the method further includes: performing comparative learning on the pre-trained large model using a plurality of comparative training sample sets to obtain a large model. The samples in the comparative training sample set include simulated input information, simulated auxiliary dialogue information, simulated reply information, and evaluation results for the simulated reply information.

[0113] According to an embodiment of the present disclosure, multiple comparative training sample sets may be target training samples acquired through steps S210 to S230, and then the target training samples are split based on the evaluation result set and the evaluation threshold to obtain multiple comparative training sample sets. For example, the target training samples may be split into high-quality samples and medium-quality samples, and the large model may be fine-tuned. By comparative training of high-quality samples and medium-quality samples, the model may be able to distinguish and tend to learn the features of high-quality samples. Comparative learning strengthens the influence of high-quality dialogue samples, enabling the model to more accurately generate responses that meet user needs and merchant service standards, thereby addressing the defects of existing large models in terms of unstable response quality.

[0114] Figure 7 The flowchart of the dialogue method according to the embodiment of the present disclosure is schematically shown.

[0115] like Figure 7 As shown, the method includes operations S710~S720.

[0116] In operation S710, input information and auxiliary dialogue information are acquired.

[0117] In operation S720, the target large model is called to process the input information and the auxiliary dialogue information to obtain reply information for replying the input information.

[0118] According to an embodiment of the present disclosure, the auxiliary dialogue information is used to characterize the dialogue intent and personalized intent of the input information.

[0119] According to an embodiment of the present disclosure, the target big model may be obtained by training a dialogue big model or other big models according to the method of operations S610 to S620.

[0120] According to an embodiment of the present disclosure, since the target big model is trained according to the operation S610~S620 method, the target big model can output a response that is adapted to the dialogue intent and personalized intent of the input information based on the input information and the auxiliary dialogue information.

[0121] According to an embodiment of the present disclosure, auxiliary dialogue information is obtained based on an object portrait of a target object and dialogue context information of input information.

[0122] The input information may be input by a target object, and the target object may be an individual user, an enterprise user, an industry professional, or a group organization. The target macro model may obtain an object portrait of the target object in response to receiving the input information of the target object.

[0123] According to an embodiment of the present disclosure, the object portrait of the target object can be obtained by determining the historical behavior information and object attribute information of the target object to obtain the object interest portrait. Based on the historical conversation information of the target object, the object interaction portrait is determined. Based on the object interest portrait and the object interaction portrait, the object portrait of the target object is obtained.

[0124] In some exemplary embodiments, during a conversation, the target macro model can call the target object's object portrait in real time, and generate personalized conversation content in combination with the contextual information of the current conversation input information. For example, for users who prefer concise answers, the target macro model will give priority to providing concise replies; while for users who prefer detailed explanations, the target macro model will provide more detailed answers.

[0125] In some exemplary embodiments, the target object's object portrait may not be static, but dynamically updated as the target object's behavior changes. After each conversation, the target big model will adjust the object portrait based on the new interaction data to ensure the real-time and accuracy of the object portrait features. This dynamic update mechanism enables the target big model to continuously learn the target object's new needs and preferences, thereby providing better services in future conversations.

[0126] In some exemplary embodiments, through an in-depth understanding of the target object's object portrait, the target big model can more effectively guide business in the conversation. For example, when serving high-value customers, the target big model can recommend related products or services based on the user's historical purchase records and preferences, thereby improving business conversion rates. In addition, for potential customers, the target big model can provide personalized recommendations and guidance based on the user's browsing behavior and interest tags to promote transactions.

[0127] In some exemplary embodiments, the target big model can not only meet the user's immediate needs but also improve the overall user experience by providing personalized dialogue. For example, in the after-sales service scenario, the target big model can quickly locate the problem and provide corresponding solutions based on the target object's past purchase and problem history. It provides accurate and efficient service methods, enhances user loyalty to the brand, and increases repurchase rate.

[0128] In some exemplary embodiments, in complex conversation scenarios, the target object's object portrait can help the big model better understand the target object's intentions and maintain contextual consistency in multiple rounds of conversations. For example, in cross-channel interactions, the target object can use the object portrait to integrate interactions from different channels into a continuous conversation experience, avoiding information gaps and repeated questions and answers, thereby providing a smoother service process.

[0129] According to an embodiment of the present disclosure, calling the target big model to process input information and auxiliary dialogue information to obtain reply information for replying to the input information includes: calling the dialogue big model to identify the interaction domain of the input information and the auxiliary dialogue information to determine the target interaction domain. Inputting the input information and the auxiliary dialogue information into an expert model matching the target interaction domain to obtain reply information.

[0130] According to the embodiments of the present disclosure, the expert model can be realized by cutting and optimizing the large basic model. Corresponding expert models can be set for different interaction fields, and each expert model focuses on the dialogue processing in its own specific field, so it can have a faster response time and more accurate professional field capabilities. This can effectively improve the accuracy and response speed of the dialogue, while improving the adaptability to complex dialogue situations.

[0131] The dialogue method according to the present disclosure has wide application potential in many fields, such as:

[0132] 1. E-commerce platform

[0133] In the field of e-commerce, it can be used in customer service dialogue, product recommendations, order inquiries, after-sales support and other scenarios. Through the dialogue big model, merchants can achieve all-weather, seamless user interaction, improve customer satisfaction and purchase conversion rate.

[0134] Application Process:

[0135] Users consult on e-commerce platforms through chat windows or voice assistants.

[0136] Input data is received and passed through an expert model for domain recognition (e.g., product recommendations, order inquiries).

[0137] The target large model combines context and history to provide personalized responses through a generation module optimized by contrastive learning.

[0138] The system evaluates user feedback in real time and optimizes dialogue strategies through a multi-objective evaluation model to ensure response quality and customer satisfaction.

[0139] Output personalized recommendations or problem solutions, update user data in the background, and continuously optimize subsequent conversations.

[0140] 2. Online education

[0141] In online education scenarios, it can be used for intelligent teaching assistants, learning guidance, and question answering to improve learning experience and teaching quality. Through personalized learning guidance and real-time question answering, the target big model can help students better master knowledge.

[0142] Application Process:

[0143] Students ask learning-related questions (e.g., course content, homework instructions) through educational platforms or applications.

[0144] The expert model identifies the learning scenario (e.g., mathematics, language learning) and works in conjunction with the target large model.

[0145] Use the large model optimized by contrastive learning to generate personalized teaching suggestions or answers.

[0146] The system evaluates the teaching effectiveness of the dialogue (such as answer accuracy and student understanding) through a multi-objective evaluation model.

[0147] Provide student feedback channels to adjust and optimize model generation strategies in real time to improve the effectiveness of personalized teaching.

[0148] 3. Financial services

[0149] In the field of financial services, it can be used for intelligent financial advisors, customer consultation, risk assessment, etc. The target big model can interact with users through natural language, provide real-time financial information, suggestions and services, and optimize user experience.

[0150] Application Process:

[0151] Users interact with intelligent customer service through banking or financial management applications (such as consulting on financial products and risk assessment).

[0152] The expert model identifies relevant issues in the financial field and combines with the big model for contextual analysis and response generation.

[0153] Generated responses are optimized to ensure conversations meet compliance requirements and provide precise financial advice.

[0154] After the conversation, a multi-objective evaluation model is used to evaluate the effectiveness of the service and the user's satisfaction, and a feedback loop is used to optimize future conversation strategies.

[0155] Output financial advice or answers and update the user's financial profile to support continuous optimization and personalized services.

[0156] 4. Medical consultation

[0157] The present invention can be used in medical consultation scenarios such as patient consultation, health consultation, and medication guidance. The target large model can provide professional health advice and support through natural language understanding of patient questions, reduce the burden on doctors and improve patient satisfaction.

[0158] Application Process:

[0159] Patients consult through medical platforms or applications (such as symptom description, health advice).

[0160] The expert model identifies relevant conversation scenarios in the medical field and connects to the big model for personalized analysis.

[0161] Use the model optimized by contrastive learning to provide professional and accurate health advice or guidance, and optimize the conversation experience based on the multi-objective evaluation model.

[0162] The system monitors the effectiveness of conversations (such as patient understanding and effectiveness of suggestions) in real time and continuously optimizes response strategies.

[0163] Provides specific medical advice or referral recommendations and updates patient health records to support follow-up conversations.

[0164] 5. Hotel and tourism services

[0165] In the field of hotel and tourism services, it can be used for intelligent customer service, reservation management, itinerary recommendations, etc. Through the target big model, merchants can provide personalized service suggestions, such as hotel reservations, travel route recommendations, and emergency handling.

[0166] Application Process:

[0167] Users initiate conversations through hotel or travel platforms (e.g., booking rooms, searching for tourist attractions).

[0168] The expert model identifies the service type and connects to the big model for context understanding and response generation.

[0169] Optimize user experience and ensure accuracy and immediacy of responses, especially in scenarios with complex demands or complaint handling.

[0170] The conversation process is evaluated in real time through a multi-objective evaluation model to ensure service quality and customer satisfaction.

[0171] Output customized service recommendations or booking confirmations, and update users’ historical data to optimize future interactions.

[0172] Figure 8 The block diagram of a sample generating device according to an embodiment of the present disclosure is schematically shown.

[0173] like Figure 8 As shown, the sample generating device 800 includes a processing module 810 , an evaluation module 820 and a sample determining module 830 .

[0174] Processing module 810 is used to call the dialogue model to process the simulation input information and the simulation auxiliary dialogue information to obtain multiple different simulation reply information, wherein the simulation auxiliary dialogue information is used to characterize the dialogue intention and personalized intention of the simulation input information, and the multiple simulation reply information are all used to reply to the simulation input information.

[0175] The evaluation module 820 is used to evaluate the multiple simulation reply information respectively based on the simulation input information and the simulation auxiliary dialogue information to obtain multiple evaluation results.

[0176] The sample determination module 830 is used to obtain a target training sample based on multiple evaluation results, simulation auxiliary dialogue information, simulation input information and multiple simulation reply information.

[0177] According to an embodiment of the present disclosure, the evaluation module 820 includes an evaluation index generation submodule and an evaluation result determination submodule.

[0178] The evaluation index generation submodule is used to input simulation input information, simulation auxiliary dialogue information and simulation response information into the multi-objective evaluation model to obtain different multiple evaluation index results.

[0179] The evaluation result determination submodule is used to obtain the evaluation result of the simulation response information based on multiple evaluation index results.

[0180] According to an embodiment of the present disclosure, the sample generating device 800 further includes a multi-objective evaluation training module.

[0181] A multi-objective evaluation training module is used to train an initial multi-objective evaluation model based on sample evaluation input information and sample evaluation labels matching the sample evaluation input information to obtain a multi-objective evaluation model, wherein the sample evaluation input information includes evaluation input information, evaluation auxiliary dialogue information and evaluation reply information; the sample evaluation labels represent the comprehensive evaluation results of the evaluation reply information for multiple evaluation indicators; the initial multi-objective evaluation model includes an initial indicator weight matrix and a pre-trained evaluation large model; the initial indicator weight matrix includes weight parameters of multiple evaluation indicators.

[0182] According to an embodiment of the present disclosure, the multi-objective evaluation training module includes a sample evaluation index generation submodule, a sample evaluation result determination submodule and a multi-objective evaluation model determination submodule.

[0183] The sample evaluation index generation submodule is used to input the sample evaluation input information into the initial multi-objective evaluation model to obtain multiple sample evaluation index results.

[0184] The sample evaluation result determination submodule is used to obtain the sample evaluation result based on multiple sample evaluation indicator results and the initial indicator weight matrix.

[0185] The multi-objective evaluation model determines the submodule, and adjusts the initial indicator weights and the parameters of the pre-trained evaluation model based on the sample evaluation results and sample evaluation labels to obtain the multi-objective evaluation model.

[0186] According to an embodiment of the present disclosure, the sample determination module includes a sample multi-tuple determination submodule and a target training sample determination submodule.

[0187] The sample multi-group determination submodule is used to obtain a dialogue sample multi-group based on the simulation-assisted dialogue information, the simulation input information, the simulation reply information and the evaluation result corresponding to the simulation reply information.

[0188] The target training sample determination submodule is used to obtain the target training sample based on multiple dialogue sample tuples.

[0189] According to an embodiment of the present disclosure, the large model-based sample generation device further includes a fine-tuning sample evaluation module and a fine-tuning sample splitting module.

[0190] The fine-tuning sample evaluation module is used to evaluate the fine-tuning samples in the fine-tuning sample set to obtain an evaluation result set of the fine-tuning sample set, wherein the fine-tuning samples include simulated input information, simulated auxiliary dialogue information and simulated reply information.

[0191] The fine-tuning sample splitting module is used to split the fine-tuning sample set based on the evaluation result set and the evaluation threshold to obtain multiple comparative training sample sets, wherein the evaluation threshold is used to indicate the splitting standard for splitting the fine-tuning sample set based on the evaluation result set.

[0192] According to an embodiment of the present disclosure, the large model-based sample generation device further includes a simulation-assisted dialogue information generation module.

[0193] The simulation auxiliary dialogue information generation module is used to obtain simulation auxiliary dialogue information based on the simulation object portrait of the simulation object and the simulation dialogue context information of the simulation input information, wherein the simulation input information is obtained by simulating the simulation object.

[0194] According to an embodiment of the present disclosure, the sample generation device based on the large model also includes a simulation interest portrait determination module, a simulation interaction portrait determination module and a simulation object portrait determination module.

[0195] The simulation interest portrait determination module is used to determine the simulation historical behavior information and simulation object attribute information of the simulation object to obtain the simulation object interest portrait.

[0196] The simulation interaction portrait determination module is used to determine the simulation object interaction portrait based on the simulation history dialogue information of the simulation object.

[0197] The simulation object portrait determination module is used to obtain the simulation object portrait based on the simulation object interest portrait and the simulation object interaction portrait.

[0198] According to an embodiment of the present disclosure, the processing module 810 includes a prompt information generation submodule and a simulation reply information generation submodule.

[0199] The prompt information generation submodule is used to generate prompt information of the dialogue model based on the simulation input information and the simulation auxiliary dialogue information.

[0200] The simulation reply information generation submodule is used to call the large dialogue model to process the prompt information and obtain multiple different simulation reply information.

[0201] According to an embodiment of the present disclosure, the reply information generation submodule includes a first execution unit and a second execution unit.

[0202] The first execution unit is used to execute the operation of inputting prompt information into the dialogue model multiple times.

[0203] The second execution unit executes the operation of updating the reply requirement information in the prompt information multiple times and inputting the updated prompt information into the dialogue model, wherein the reply requirement information is used to indicate the expected requirements for the simulation reply information.

[0204] According to an embodiment of the present disclosure, the reply information generation submodule includes a simulation interaction field determination unit and a simulation reply information generation unit.

[0205] The simulation interaction domain determination unit is used to call the dialogue model, identify the interaction domain of the prompt information, and determine the simulation interaction domain.

[0206] The simulation reply information generating unit is used to input the prompt information into an expert model matching the simulation interaction field to obtain the simulation reply information.

[0207] Fig. 9 A block diagram of a large model training device according to an embodiment of the present disclosure is schematically shown.

[0208] like Fig. 9 As shown, the training device of the large model includes a sample acquisition module 910 and a training module 920.

[0209] The sample acquisition module 910 is used to acquire target training samples.

[0210] The training module 920 is used to train the large model using the target training samples, wherein the target training samples are generated according to the above-mentioned sample generating device.

[0211] According to an embodiment of the present disclosure, the training device of the large model also includes a comparative learning module.

[0212] The contrastive learning module is used to perform contrastive learning on the pre-trained large model using multiple contrastive training sample sets to obtain the large model, wherein the samples in the contrastive training sample set include simulated input information, simulated auxiliary dialogue information, simulated reply information and evaluation results for the simulated reply information.

[0213] Fig.10 The block diagram of the dialogue device according to the embodiment of the present disclosure is schematically shown.

[0214] like Fig.10 As shown, the dialogue device 1000 includes an information acquisition module 1010 and a dialogue processing module 1020 .

[0215] The information acquisition module 1010 is used to acquire input information and auxiliary dialogue information.

[0216] The dialogue processing module 1020 is used to call the target big model to process the input information and the auxiliary dialogue information to obtain the reply information used to reply the input information, wherein the auxiliary dialogue information is used to characterize the dialogue intention and personalized intention of the input information; wherein the target big model is trained according to the training device of the above-mentioned big model.

[0217] According to an embodiment of the present disclosure, the dialogue device includes an auxiliary dialogue information determination module.

[0218] The auxiliary dialogue information determination module is used to obtain auxiliary dialogue information based on the object portrait of the target object and the dialogue context information of the input information, wherein the input information is input by the target object.

[0219] According to an embodiment of the present disclosure, the dialogue device also includes an object interest portrait determination module, an object interaction portrait determination module and an object portrait determination module.

[0220] The object interest portrait determination module is used to determine the historical behavior information and object attribute information of the target object to obtain the object interest portrait.

[0221] The object interaction portrait determination module is used to determine the object interaction portrait based on the historical conversation information of the target object.

[0222] The object portrait determination module is used to obtain the object portrait of the target object based on the object interest portrait and the object interaction portrait.

[0223] According to an embodiment of the present disclosure, the dialogue processing module includes a target interaction domain determination submodule and a reply information determination submodule.

[0224] The target interaction domain determination submodule is used to call the large dialogue model, identify the interaction domain of the input information and auxiliary dialogue information, and determine the target interaction domain.

[0225] The reply information determination submodule is used to input the input information and the auxiliary dialogue information into the expert model matching the target interaction domain to obtain the reply information.

[0226] Fig.11The structural block diagram of an intelligent agent according to an embodiment of the present disclosure is schematically shown.

[0227] In the embodiments of the present disclosure, inspired by the von Neumann structure in modern computer theory, such as Fig.11 As shown, the agent 1100 may include multiple core modules: an input module 1110 , a processing module 1120 , and an output module 1130 .

[0228] The input module 1110 is responsible for receiving or sensing information such as queries, requests, instructions, signals, problem information or data from the outside world (such as users or external environment), and converting it into a format that can be understood and processed by the intelligent agent 1100. The input module 1110 is the primary link for the intelligent agent 1100 to interact with the outside world. It enables the intelligent agent 1100 to efficiently and accurately obtain the necessary "sensory" information from the outside world and respond to this information.

[0229] In an example, the input module 1110 may be used to receive input information. The input information may include question information as described above.

[0230] In the example, the processing module 1120 is the core support for the ability of the agent 1100 to handle complex tasks. The processing module 1120 is used to determine the target task based on the input information received by the input module 1110. The large model is determined based on the target task. By calling the large model, the question-answering method based on the large model described above can be executed to obtain output information. The output information may include the target answer information as described above.

[0231] In an example, the output module 1130 may be used to output the output information obtained by the processing module 1120, such as the target answer information described above.

[0232] In an example, the processing module 1120 may include a control unit 1121 , a storage unit 1122 , and an operation unit 1123 .

[0233] During operation, the control unit 1121 will continuously interact with the storage unit 1122, the computing unit 1123 and / or the output module 1130. However, it should be noted that in the embodiment of the present disclosure, the control unit 1121 acts as a single initiator to initiate communication with the storage unit 1130, the computing unit 1123 and / or the output module 1150, and there is no communication coupling between the storage unit 1122, the computing unit 1123 and the output module 1150.

[0234] In the example, the performance of the control unit 1121 may be closely related to the large model based on the agent 1100. In order to give full play to the capabilities of the large language model, the internal structure of the control unit 1121 may be designed to be highly configurable and scalable to cope with various types of tasks and requirements in real scenarios.

[0235] The storage unit 1122 may be responsible for memorizing information such as historical conversations, event flows, etc. The prompt information and information set as described above may be included in the storage unit 1122 .

[0236] In the example, after the intelligent agent 1100 obtains the question information, the intelligent agent 1100 can use the question information to determine the target task. Determine reference information from the information set based on the target task. The reference information can be stored in the storage unit 1122. The intelligent agent 1100 can retrieve relevant prompt information and reference information from the storage unit 1122 and feed it back to the control unit 1121. Then, the control unit 1121 can use the feedback prompt information and reference information to determine the large model based on the target task. The target answer information corresponding to the question information is obtained by calling the large model, and the target answer information is passed to the output module 1130.

[0237] The operation unit 1123 can be regarded as a predefined tool library. The search engine, format conversion tool, etc. mentioned above can be included in the operation unit 1123.

[0238] In the example, when the agent 1100 needs to convert the format of the initial target answer, the relevant search engine or format conversion tool can be called from the operation unit 1123, and the original content information corresponding to the reference segment can be called by the search engine, and the original content information can be fed back to the control unit 1121. Alternatively, the reference segment can be converted by using the format conversion tool to obtain the content of the predetermined modality, and the content of the predetermined modality can be fed back to the control unit 1121. Then, the control unit 1121 can use the original content information or the content information of the predetermined modality fed back to obtain the target answer information, and pass the target answer information to the output module 1130, so that the output module 1130 outputs the target answer information to the user. It can be understood that although the large language model has excellent language understanding and generation capabilities, it is the same as a person, without the help of any tools, and the tasks that can be solved are very limited. When the agent 1100 is given the ability to call tools, it can achieve tasks such as completing mathematical operations with the help of a calculator, completing data analysis with the help of python, and completing weather forecasts with the help of a search engine.

[0239] According to the embodiment of the present disclosure, the intelligent agent 1100 can simply and effectively improve the level of intelligence, and improve flexibility and versatility.

[0240] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0241] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a method as in the embodiment of the present disclosure.

[0242] According to an embodiment of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute a method as described in an embodiment of the present disclosure.

[0243] According to an embodiment of the present disclosure, a computer program product includes a computer program, and when the computer program is executed by a processor, the method according to the embodiment of the present disclosure is implemented.

[0244] Fig.12 A schematic block diagram of an example electronic device 1200 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0245] like Fig.12 As shown, the device 1200 includes a computing unit 1201, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1202 or a computer program loaded from a storage unit 1208 into a random access memory (RAM) 1203. In the RAM 1203, various programs and data required for the operation of the device 1200 can also be stored. The computing unit 1201, the ROM 1202, and the RAM 1203 are connected to each other via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0246] A number of components in the device 1200 are connected to an input / output (I / O) interface 1205, including: an input unit 1206, such as a keyboard, a mouse, etc.; an output unit 1207, such as various types of displays, speakers, etc.; a storage unit 1208, such as a disk, an optical disk, etc.; and a communication unit 1209, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1209 allows the device 1200 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0247] The computing unit 1201 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1201 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1201 performs the various methods and processes described above, such as the sample generation method. For example, in some embodiments, the sample generation method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 1208. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 1200 via the ROM 1202 and / or the communication unit 1209. When the computer program is loaded into the RAM 1203 and executed by the computing unit 1201, one or more steps of the sample generation method described above may be performed. Alternatively, in other embodiments, the computing unit 1201 may be configured to perform the sample generation method in any other appropriate manner (e.g., by means of firmware).

[0248] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0249] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0250] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0251] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0252] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0253] A computer system may include a client and a server. The client and the server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises through computer programs running on respective computers and having a client-server relationship to each other. The server may be a cloud server, a server in a distributed system, or a server combined with a blockchain.

[0254] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0255] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A sample generation method based on a large model, comprising: Calling the big dialogue model to process the simulation input information and the simulation auxiliary dialogue information to obtain a plurality of different simulation reply information, wherein the simulation auxiliary dialogue information is used to characterize the dialogue intention and personalized intention of the simulation input information, and the plurality of simulation reply information are all used to reply to the simulation input information; Based on the simulation input information and the simulation auxiliary dialogue information, respectively evaluate the plurality of simulation reply information to obtain a plurality of evaluation results; and Based on the plurality of evaluation results, the simulation-assisted dialogue information, the simulation input information and the plurality of simulation reply information, a target training sample is obtained.

2. The method according to claim 1, wherein: Based on the simulation input information and the simulation auxiliary dialogue information, the plurality of simulation reply information are evaluated respectively to obtain a plurality of evaluation results, including: Inputting the simulation input information, the simulation auxiliary dialogue information and the simulation reply information into a multi-objective evaluation model to obtain different multiple evaluation index results; and Based on the multiple evaluation index results, an evaluation result of the simulation reply information is obtained.

3. The method according to claim 2, wherein: The multi-objective evaluation model is trained in the following way: Based on sample evaluation input information and sample evaluation labels matching the sample evaluation input information, an initial multi-objective evaluation model is trained to obtain the multi-objective evaluation model, wherein the sample evaluation input information includes evaluation input information, evaluation auxiliary dialogue information and evaluation reply information; the sample evaluation labels represent the comprehensive evaluation results of the evaluation reply information for multiple evaluation indicators; the initial multi-objective evaluation model includes an initial indicator weight matrix and a pre-trained evaluation large model; the initial indicator weight matrix includes weight parameters of each of the multiple evaluation indicators.

4. The method according to claim 3, wherein: The method of training an initial multi-objective evaluation model based on the sample evaluation input information and the sample evaluation label matching the sample evaluation input information to obtain the multi-objective evaluation model includes: Inputting sample evaluation input information into the initial multi-objective evaluation model to obtain multiple sample evaluation index results; Based on the plurality of sample evaluation index results and the initial index weight matrix, obtaining a sample evaluation result; and Based on the sample evaluation results and the sample evaluation labels, the initial indicator weights and the parameters of the pre-trained evaluation model are adjusted to obtain the multi-objective evaluation model.

5. The method according to any one of claims 1 to 4, wherein: The step of obtaining a target training sample based on the plurality of evaluation results, the simulation auxiliary dialogue information, the simulation input information and the plurality of simulation reply information comprises: Obtaining a dialogue sample tuple based on the simulation-assisted dialogue information, the simulation input information, the simulation reply information, and the evaluation result corresponding to the simulation reply information; and Based on a plurality of the dialogue sample tuples, the target training sample is obtained.

6. The method according to any one of claims 1 to 5, further comprising: Evaluating fine-tuning samples in a fine-tuning sample set to obtain an evaluation result set of the fine-tuning sample set, wherein the fine-tuning samples include the simulated input information, the simulated auxiliary dialogue information, and the simulated reply information; and Based on the evaluation result set and the evaluation threshold, the fine-tuning sample set is split to obtain a plurality of comparative training sample sets, wherein the evaluation threshold is used to indicate a splitting criterion for splitting the fine-tuning sample set based on the evaluation result set.

7. The method according to any one of claims 1 to 6, further comprising: The simulation auxiliary dialogue information is obtained based on the simulation object portrait of the simulation object and the simulation dialogue context information of the simulation input information, wherein the simulation input information is obtained by simulating the simulation object.

8. The method according to claim 7, further comprising: Determine the simulation history behavior information and the simulation object attribute information of the simulation object to obtain the simulation object interest portrait; as well as Determining an interaction portrait of the simulation object based on the simulation history dialogue information of the simulation object; as well as Based on the simulation object interest portrait and the simulation object interaction portrait, the simulation object portrait is obtained.

9. The method according to any one of claims 1 to 8, wherein: The calling dialogue model processes the simulation input information and the simulation auxiliary dialogue information to obtain different multiple simulation reply information, including: Based on the simulation input information and the simulation auxiliary dialogue information, generating prompt information of the dialogue macro model; and The large dialogue model is called to process the prompt information to obtain different multiple simulation reply information.

10. The method according to claim 9, wherein: The calling of the dialogue model to process the prompt information and obtain different multiple simulation reply information includes: Execute the operation of inputting the prompt information into the dialogue macro model multiple times; or The operation of updating the reply requirement information in the prompt information multiple times and inputting the updated prompt information into the dialogue model is performed, wherein the reply requirement information is used to indicate the expected requirements for the simulation reply information.

11. The method according to claim 9 or 10, wherein: The calling of the dialog big model to process the prompt information includes: Calling the dialogue macro model, performing interaction field recognition on the prompt information, and determining a simulation interaction field; and The prompt information is input into an expert model matching the simulation interaction field to obtain the simulation reply information.

12. A large model training method, comprising: Use the target training samples to train the large model. Wherein, the target training sample is generated according to the method according to any one of claims 1 to 11.

13. The method according to claim 12, further comprising: The pre-trained large model is subjected to comparative learning using a plurality of comparative training sample sets to obtain the large model, wherein the samples in the comparative training sample sets include the simulated input information, the simulated auxiliary dialogue information, the simulated reply information and an evaluation result for the simulated reply information.

14. A conversation method, comprising: Calling the target large model to process the input information and the auxiliary dialogue information to obtain reply information for replying to the input information, wherein the auxiliary dialogue information is used to characterize the dialogue intent and personalized intent of the input information; Wherein, the target large model is trained according to the method described in claim 12 or 13.

15. The method according to claim 14, further comprising: The auxiliary dialogue information is obtained based on the object portrait of the target object and the dialogue context information of the input information, wherein the input information is input by the target object.

16. The method according to claim 15, further comprising: Determine the historical behavior information and object attribute information of the target object to obtain the object interest portrait; as well as Determine an object interaction profile based on the historical conversation information of the target object; as well as Based on the object interest portrait and the object interaction portrait, an object portrait of the target object is obtained.

17. The method according to any one of claims 14 to 16, wherein: The calling target large model processes the input information and the auxiliary dialogue information to obtain reply information for replying the input information, including: calling the dialogue macro model, performing interaction domain recognition on the input information and the auxiliary dialogue information, and determining a target interaction domain; and The input information and the auxiliary dialogue information are input into an expert model matching the target interaction domain to obtain the reply information.

18. A sample generation device based on a large model, comprising: A processing module, used for calling a large dialogue model to process the simulation input information and the simulation auxiliary dialogue information, and obtaining a plurality of different simulation reply information, wherein the simulation auxiliary dialogue information is used to characterize the dialogue intention and personalized intention of the simulation input information, and the plurality of simulation reply information are all used to reply to the simulation input information; an evaluation module, configured to evaluate the plurality of simulation reply information respectively based on the simulation input information and the simulation auxiliary dialogue information, and obtain a plurality of evaluation results; and The sample determination module is used to obtain a target training sample based on the multiple evaluation results, the simulation auxiliary dialogue information, the simulation input information and the multiple simulation reply information.

19. The device according to claim 18, wherein: The evaluation module comprises: an evaluation index generating submodule, for inputting the simulation input information, the simulation auxiliary dialogue information and the simulation reply information into a multi-objective evaluation model to obtain different multiple evaluation index results; and The evaluation result determination submodule is used to obtain the evaluation result of the simulation reply information based on the multiple evaluation indicator results.

20. The device according to claim 19, wherein The multi-objective evaluation model is obtained by training the following modules: A multi-objective evaluation training module is used to train an initial multi-objective evaluation model based on sample evaluation input information and sample evaluation labels matching the sample evaluation input information to obtain the multi-objective evaluation model, wherein the sample evaluation input information includes evaluation input information, evaluation auxiliary dialogue information and evaluation reply information; the sample evaluation labels represent the comprehensive evaluation results of the evaluation reply information for multiple evaluation indicators; the initial multi-objective evaluation model includes an initial indicator weight matrix and a pre-trained evaluation large model; the initial indicator weight matrix includes weight parameters of each of the multiple evaluation indicators.

21. The device according to claim 20, wherein: The multi-objective evaluation training module includes: A sample evaluation index generation submodule is used to input sample evaluation input information into the initial multi-objective evaluation model to obtain multiple sample evaluation index results; A sample evaluation result determination submodule, configured to obtain a sample evaluation result based on the plurality of sample evaluation index results and the initial index weight matrix; and The multi-objective evaluation model determination submodule adjusts the initial indicator weights and the respective parameters of the pre-trained evaluation model based on the sample evaluation results and the sample evaluation labels to obtain the multi-objective evaluation model.

22. The device according to any one of claims 18 to 21, wherein The sample determination module comprises: a sample multi-group determination submodule, configured to obtain a dialogue sample multi-group based on the simulation-assisted dialogue information, the simulation input information, the simulation reply information, and the evaluation result corresponding to the simulation reply information; and The target training sample determination submodule is used to obtain the target training sample based on a plurality of the dialogue sample tuples.

23. The apparatus according to any one of claims 18 to 22, further comprising: a fine-tuning sample evaluation module, configured to evaluate the fine-tuning samples in the fine-tuning sample set to obtain an evaluation result set of the fine-tuning sample set, wherein the fine-tuning samples include the simulated input information, the simulated auxiliary dialogue information and the simulated reply information; and A fine-tuning sample splitting module is used to split the fine-tuning sample set based on the evaluation result set and the evaluation threshold to obtain multiple comparative training sample sets, wherein the evaluation threshold is used to indicate a splitting criterion for splitting the fine-tuning sample set based on the evaluation result set.

24. The apparatus according to any one of claims 18 to 23, further comprising: The simulation auxiliary dialogue information generation module is used to obtain the simulation auxiliary dialogue information based on the simulation object portrait of the simulation object and the simulation dialogue context information of the simulation input information, wherein the simulation input information is obtained by simulating the simulation object.

25. The apparatus according to claim 24, further comprising: A simulation interest portrait determination module, used to determine the simulation history behavior information and simulation object attribute information of the simulation object to obtain the simulation object interest portrait; as well as A simulation interaction portrait determination module, used to determine a simulation object interaction portrait based on simulation history dialogue information of the simulation object; as well as A simulation object portrait determination module is used to obtain the simulation object portrait based on the simulation object interest portrait and the simulation object interaction portrait.

26. The device according to any one of claims 18 to 25, wherein: The processing module comprises: a prompt information generating submodule, for generating prompt information of the dialogue macromodel based on the simulation input information and the simulation auxiliary dialogue information; and The simulation reply information generation submodule is used to call the dialogue model to process the prompt information and obtain different multiple simulation reply information.

27. The device according to claim 26, wherein: The reply information generation submodule includes: A first execution unit is used to execute the operation of inputting the prompt information into the dialogue macro model multiple times; or The second execution unit executes the operation of updating the reply requirement information in the prompt information multiple times and inputting the updated prompt information into the dialogue model, wherein the reply requirement information is used to indicate the expected requirements for the simulation reply information.

28. The device according to claim 26 or 27, wherein The reply information generation submodule includes: a simulation interaction domain determination unit, configured to call the dialog macro model, identify the interaction domain of the prompt information, and determine the simulation interaction domain; and The simulation reply information generating unit is used to input the prompt information into an expert model matching the simulation interaction field to obtain the simulation reply information.

29. A large model training device, comprising: The training module is used to train the large model using the target training samples. Wherein, the target training sample is generated according to the device according to any one of claims 18 to 28.

30. The apparatus of claim 29, further comprising: A comparative learning module is used to perform comparative learning on a pre-trained large model using a plurality of comparative training sample sets to obtain the large model, wherein the samples in the comparative training sample set include the simulated input information, the simulated auxiliary dialogue information, the simulated reply information and an evaluation result for the simulated reply information.

31. A conversation device, comprising: A dialogue processing module, used for calling the target large model to process input information and auxiliary dialogue information, and obtaining reply information for replying to the input information, wherein the auxiliary dialogue information is used to characterize the dialogue intention and personalized intention of the input information; Wherein, the target large model is trained according to the device as described in claim 29 or 30.

32. The apparatus of claim 31 , further comprising: The auxiliary dialogue information determination module is used to obtain the auxiliary dialogue information based on the object portrait of the target object and the dialogue context information of the input information, wherein the input information is input by the target object.

33. The apparatus of claim 32, further comprising: An object interest portrait determination module is used to determine the historical behavior information and object attribute information of the target object to obtain an object interest portrait; as well as An object interaction portrait determination module, used to determine an object interaction portrait based on historical conversation information of the target object; as well as The object portrait determination module is used to obtain the object portrait of the target object based on the object interest portrait and the object interaction portrait.

34. The device according to any one of claims 31 to 33, wherein: The dialogue processing module comprises: a target interaction domain determination submodule, configured to call the dialog macro model, perform interaction domain identification on the input information and the auxiliary dialog information, and determine a target interaction domain; and The reply information determination submodule is used to input the input information and the auxiliary dialogue information into an expert model matching the target interaction field to obtain the reply information.

35. An artificial intelligence agent configured to perform the method according to any one of claims 14 to 17.

36. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 17.

37. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 17.

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

Images