Digital human generation method and device and storage medium

Through machine learning algorithms, analyzing user needs and generating personalized digital people's appearance, voice and interactive behavior, solving the problem of lack of personalization and insufficient interactive experience of existing digital people's customized services, and improving digital people's processing capabilities and personalized customization level.

CN118916461APending Publication Date: 2024-11-08SHANGHAI SHENGTONG INFORMATION TECH CO LTD

Patent Information

Application Number
CN202410984924.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing digital human customization services lack personalization and flexibility, cannot meet users' highly personalized needs, and lack interactive experience, knowledge updates and learning abilities.

Method used

The user interface obtains the digital life needs of the target user, uses machine learning algorithms to analyze the user's preference characteristics, emotional characteristics and behavioral pattern characteristics, and generates digital appearance data, voice data, interactive behavior data and knowledge base that meets the needs, thereby building a Personalized digital people.

Benefits of technology

The personalized customization level and processing capabilities of digital people have been improved, so that they can better meet user needs, provide natural and smooth interactive experience and accurate knowledge and information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916461A_ABST
    Figure CN118916461A_ABST
Patent Text Reader

Abstract

The invention provides a digital human generation method and device and a storage medium. The method comprises the following steps: obtaining a digital human generation demand of a target user through a user interface, wherein the digital human generation demand comprises an appearance demand, a character demand, a knowledge field demand, a language style demand and an interaction mode demand of the digital human; and analyzing the digital human generation demand by adopting a machine learning algorithm to obtain preference features, emotion features and behavior pattern features of the target user. Processing the preference features, the emotion features and the behavior pattern features of the target user based on a deep learning model, and generating appearance data, sound data, interactive behavior data, an action library and a knowledge library of the digital person corresponding to the digital person generation demand, and further generating a target digital person corresponding to the digital person generation demand. According to the method provided by the invention, the processing capability and individuation capability of the digital human are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device and storage medium for generating a digital human. Background Art

[0002] With the rapid development of artificial intelligence technology, digital humans have become important applications in many fields such as virtual reality, augmented reality, games, education, customer service, etc. However, existing digital human customization services often lack personalization and flexibility, and cannot meet users' needs for highly personalized customization. At the same time, they also have defects such as insufficient interactive experience, insufficient knowledge updating and learning ability, and limited data processing capabilities.

[0003] Therefore, how to improve the processing power and personalization of digital humans is an urgent problem that needs to be solved. Summary of the invention

[0004] The present application provides a method, device and storage medium for generating a digital human, so as to solve the problem of poor processing capability and personalization of the digital human.

[0005] In a first aspect, the present application provides a method for generating a digital human, comprising:

[0006] Acquiring the target user's digital human generation requirements through the user interface, wherein the digital human generation requirements include the digital human's appearance requirements, personality requirements, knowledge domain requirements, language style requirements, and interaction method requirements;

[0007] Using a machine learning algorithm to analyze the digital human generation needs, and obtain the target user's preference characteristics, emotional characteristics, and behavioral pattern characteristics;

[0008] Processing the target user's preference characteristics, the emotional characteristics, and the behavioral pattern characteristics based on a deep learning model to generate the digital human's appearance data, sound data, interactive behavior data, action library, and knowledge library corresponding to the digital human generation requirements;

[0009] A target digital human corresponding to the digital human generation requirement is generated according to the digital human's appearance data, the sound data, the interactive behavior data, the action library, and the knowledge library.

[0010] Optionally, the use of a machine learning algorithm to analyze the digital human generation requirements to obtain the target user's preference characteristics, emotional characteristics, and behavioral pattern characteristics includes:

[0011] Using natural language processing technology to identify the digital human generation requirements, and obtaining a keyword set in the digital human generation requirements, wherein the keyword set represents the focus and requirements of the target user;

[0012] Performing lexical analysis and syntactic analysis on each keyword in the keyword set, and using a pre-built semantic knowledge base to use a clustering analysis algorithm to extract topic data related to the preferences of the target user from the keyword set;

[0013] Using a sentiment classification algorithm to classify the sentiment annotations in the keyword set, determining the overall sentiment tendency of the target user towards the digital human, wherein the sentiment annotations are generated by sentiment annotation of each keyword in the keyword set using a sentiment annotation tool; the overall sentiment tendency includes sentiment intensity and sentiment motivation;

[0014] Based on the behavior pattern recognition algorithm, the historical interaction behaviors of the target user are collected, and the behavior pattern recognition and classification are performed on each keyword in the keyword set to determine the interaction mode and behavior habits between the target user and the digital human;

[0015] The preference characteristics, emotional characteristics, and behavioral pattern characteristics of the target user are determined based on the topic data, the overall emotional tendency, and the interaction mode and behavioral habits.

[0016] Optionally, the processing of the target user's preference characteristics, the emotional characteristics, and the behavioral pattern characteristics based on the deep learning model to generate the digital human's appearance data, sound data, interactive behavior data, action library, and knowledge base corresponding to the digital human generation requirements includes:

[0017] Extracting the target user's appearance preference features, voice preference features, and knowledge domain features for the digital human from the preference features;

[0018] Using a generative adversarial network and a variational autoencoder to process the target user's preference characteristics for the digital person's appearance, and obtaining the digital person's appearance data, wherein the appearance data includes the digital person's facial feature data, body feature data, age data, gender data, and skin color data;

[0019] Using speech synthesis technology to process the target user's voice preference characteristics for the digital human, and obtaining the digital human's voice data, wherein the voice data includes language style data and emotional style data;

[0020] According to the knowledge domain feature, obtaining a knowledge base of the knowledge domain and / or related knowledge domains corresponding to the knowledge domain feature, wherein the knowledge base is composed of knowledge extracted from multiple knowledge domains and multiple text data of the knowledge domains;

[0021] The behavior pattern features are input into an interactive behavior model based on a reinforcement learning algorithm, and the interactive behavior data and the action library are generated based on the behavior pattern features. The actions included in the action library are obtained from a pre-built digital human action library based on the behavior pattern features. The pre-built digital human action library is generated by capturing and converting real human actions through a motion capture algorithm and a motion generation algorithm.

[0022] Optionally, generating a target digital human corresponding to the digital human generation requirement according to the digital human's appearance data, the sound data, the interactive behavior data, the action library, and the knowledge library includes:

[0023] Selecting appropriate 3D modeling technology according to the appearance data to construct a three-dimensional appearance model of the digital human, and generating the appearance effect of the target digital human through material rendering processing and lighting rendering processing;

[0024] Acquire the voice intonation, voice speed and voice emotion of the target digital person according to the sound data, and generate a voice library of the target digital person, wherein the voice in the voice library is the voice after noise reduction and enhancement processing;

[0025] Design the dialogue logic and response method of the target digital person based on the interactive behavior data, and train the dialogue logic and response method through deep learning technology, so that the target digital person can query and extract knowledge information corresponding to the input from the knowledge base based on the user's input and make corresponding responses;

[0026] Based on the actions included in the action library, and the dialogue logic and response mode of the target digital person, an association relationship between the actions and the dialogue logic and response mode of the target digital person is constructed to synchronize the actions with the dialogue logic and response mode;

[0027] The three-dimensional appearance model, the voice library, the dialogue logic, the response method, and the association between the actions in the action library and the dialogue logic and the response method are integrated to debug and generate the target digital human corresponding to the digital human generation requirements.

[0028] Optionally, the method further includes:

[0029] Acquiring the dialogue information of the target user, wherein the dialogue information is used to call the target digital human to output response information corresponding to the dialogue information;

[0030] Inputting the dialogue information into the large language model of the target digital human to generate the response information;

[0031] Feeding back the generated response information to the target user through the target digital human;

[0032] The step of inputting the dialogue information into the large language model of the target digital human to generate the response information includes:

[0033] Using named entity recognition and keyword extraction technology to semantically understand the conversation information, and obtain the intention characteristics, demand characteristics, and emotional characteristics of the target user;

[0034] Processing the intention characteristics and demand characteristics of the target user through a cluster analysis algorithm, obtaining the knowledge base category corresponding to the intention characteristics and demand characteristics of the target user, and determining the target knowledge base corresponding to the intention characteristics and demand characteristics from the knowledge base set according to the knowledge base category;

[0035] Performing knowledge retrieval in the target knowledge base to obtain candidate response content corresponding to the intention feature and the demand feature;

[0036] Generate candidate response information by reasoning and integrating the candidate response contents;

[0037] According to the candidate response information and the emotional feature, determine the language expression and tone corresponding to the emotional feature, and adjust the candidate response information by the language expression and tone to generate the response information;

[0038] The step of performing knowledge retrieval in the target knowledge base to obtain candidate response content corresponding to the intention feature and the demand feature includes:

[0039] Preliminarily constructing a target knowledge graph in the target knowledge base, wherein the nodes of the target knowledge graph are knowledge entities, the edges of the target knowledge graph are relationships between the knowledge entities, and the knowledge entities and the relationships are generated by preprocessing and annotation of a natural language processing model;

[0040] Retrieving the knowledge entities corresponding to the intention feature and the demand feature in the target knowledge graph by a depth-first search algorithm, and obtaining candidate knowledge entities and associated knowledge entities associated with the candidate knowledge entities;

[0041] According to the candidate knowledge entity and the relationship between the candidate knowledge entity and the associated knowledge entity, the candidate knowledge entity and the associated knowledge entity are integrated to generate the candidate response content.

[0042] Optionally, the method further includes:

[0043] Processing the intention characteristics and demand characteristics of the target user by a cluster analysis algorithm, obtaining knowledge base categories corresponding to the intention characteristics and demand characteristics of the target user, and determining a first target knowledge base and a second target knowledge base corresponding to the intention characteristics and demand characteristics from a knowledge base set according to the knowledge base categories, wherein the first target knowledge base and the second target knowledge base have different knowledge fields;

[0044] According to the first knowledge domain of the first target knowledge base and the second knowledge domain of the second target knowledge base, decomposing the intention characteristics and demand characteristics of the target user into first intention characteristics and demand characteristics and second intention characteristics and demand characteristics, wherein the first intention characteristics and demand characteristics are related to the first knowledge domain, and the second intention characteristics and demand characteristics are related to the second knowledge domain;

[0045] Performing knowledge retrieval on the first intention feature and the demand feature in the first target knowledge base to obtain first candidate response content corresponding to the first intention feature and the demand feature;

[0046] and performing knowledge retrieval on the second intention feature and the demand feature in the second target knowledge base to obtain second candidate response content corresponding to the second intention feature and the demand feature;

[0047] Merging the first candidate response content and the second candidate response content to generate the candidate response content;

[0048] The step of fusing the first candidate response content and the second candidate response content to generate the candidate response content includes:

[0049] Acquire a first knowledge entity corresponding to the first candidate response content, and a second knowledge entity corresponding to the second candidate response content;

[0050] determining an association type between the first knowledge entity and the second knowledge entity by crossing an edge between the first target knowledge base and the second target knowledge base;

[0051] Determine, according to the association type and the mapping relationship between the association type and the content fusion strategy, a target content fusion strategy for fusing the first candidate response content and the second candidate response content;

[0052] According to the target content fusion strategy, the first candidate response content and the second candidate response content are fused to generate the candidate response content.

[0053] Optionally, the step of fusing the first candidate response content and the second candidate response content according to the target content fusion strategy to generate the candidate response content includes:

[0054] According to the target content fusion strategy, the first candidate response content and the second candidate response content are fused to generate candidate response content to be evaluated for quality;

[0055] Generating an evaluation score by performing consistency evaluation and accuracy evaluation on the candidate response content to be quality evaluated;

[0056] If the evaluation score is less than a preset score threshold, then according to the target content fusion strategy, the first candidate response content and the second candidate response content are re-fused;

[0057] If the evaluation score is greater than or equal to the preset score threshold, the candidate response content is generated.

[0058] In a second aspect, the present application provides an electronic device, comprising: a processor, a communication interface and a memory, wherein the processor is communicatively connected to the communication interface and the memory respectively;

[0059] The memory stores computer-executable instructions;

[0060] The communication interface communicates and interacts with external devices;

[0061] The processor executes the computer-executable instructions stored in the memory to implement the method as described in any one of the first aspects.

[0062] In a third aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, they are used to implement the method as described in any one of the first aspects.

[0063] In a fourth aspect, the present application provides a computer program product, comprising a computer program, which implements any one of the methods in the first aspect when executed by a processor.

[0064] The digital human generation method, device and storage medium provided by the present application obtain the digital human generation needs of the target user through the user interface, and use the machine learning algorithm to analyze the digital human generation needs to obtain the preference characteristics, emotional characteristics, and behavior pattern characteristics of the target user. The digital human generation needs are analyzed by the machine learning algorithm to obtain the preference characteristics, emotional characteristics, and behavior pattern characteristics of the target user, and the target digital human corresponding to the digital human generation needs is generated according to the digital human's appearance data, the sound data, the interactive behavior data, the action library, and the knowledge base, thereby improving the personalized customization level of the digital human and the processing ability of the digital human. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0066] Figure 1 A schematic diagram of a flow chart of a method for generating a digital human provided in an embodiment of the present application;

[0067] Figure 2 A schematic diagram of a flow chart of another method for generating a digital human provided in an embodiment of the present application;

[0068] Figure 3 A flowchart of another method for generating a digital human provided in an embodiment of the present application;

[0069] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0070] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0071] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0072] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below in conjunction with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0073] Figure 1 A flowchart of a method for generating a digital human provided in an embodiment of the present application is shown below. Figure 1 As shown, the method may include:

[0074] S101. Obtaining the digital human generation needs of the target user through the user interface.

[0075] The digital human generation requirements include the digital human's appearance requirements, personality requirements, knowledge field requirements, language style requirements, and interaction method requirements.

[0076] The server can provide users with an intuitive user interface on which they can describe in detail their various expectations for the digital human. This interface may include multiple options and input boxes, allowing users to clearly express their specific requirements for the digital human's appearance, personality, knowledge domain, language style, and interaction method.

[0077] For example, users can select appearance features, such as hair color, eye shape, body shape, etc.; specify personality traits, such as cheerful, calm, humorous, etc.; select knowledge fields, such as history, science, art, etc.; determine language styles, such as formal, humorous, friendly, etc.; and describe the interaction methods they want, such as dialogue, operation, etc. For example, a user may want to create a digital person with an Asian face, cheerful personality, good at music knowledge, using a humorous language style, and interacting through voice and gestures. In this way, the server can accurately understand the user's needs and provide clear guidance for the subsequent digital person generation process.

[0078] S102: Analyze the digital human generation needs using a machine learning algorithm to obtain the target user's preference characteristics, emotional characteristics, and behavioral pattern characteristics.

[0079] The server uses natural language processing technology to analyze the text input by the user and extract keywords and semantic information. Then, the server uses machine learning algorithms, such as cluster analysis and classification algorithms, to process and model this information. By learning and training on a large amount of user data, the server's algorithm can identify the user's preference characteristics, emotional characteristics, and behavioral pattern characteristics.

[0080] For example, if a user frequently mentions music and painting in their descriptions, the server's algorithm may identify their preference for art. If a user's language is full of positive emotional words, the server's algorithm may determine that they have positive emotional characteristics. In addition, by analyzing the user's historical data, the server can also discover the user's behavioral pattern characteristics, such as whether they pay attention to details in communication, whether they tend to take the initiative to initiate conversations, etc.

[0081] S103, processing the preference characteristics, the emotional characteristics, and the behavioral pattern characteristics of the target user based on the deep learning model, and generating the appearance data, sound data, interactive behavior data, action library, and knowledge base of the digital human corresponding to the digital human generation requirements.

[0082] The server uses deep learning models, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), to generate the digital human's appearance data, voice data, and action library. These models are able to learn the appearance, voice, and action features of real humans, and adjust and optimize them according to the target user's preference characteristics and emotional characteristics. For example, if the target user wants the digital human to have a specific hairstyle and clothing style, the server can generate appearance data that meets the requirements by adjusting the parameters of the GAN model. In terms of voice, the server uses speech synthesis technology to generate the digital human's voice data. According to the target user's emotional characteristics, the server can adjust the tone, speed, and emotional color of the voice to make it more in line with the user's needs. In terms of interactive behavior data, the server uses knowledge graphs and natural language processing technologies to build a knowledge base for the digital human. This knowledge base contains rich knowledge and information, enabling the digital human to answer various questions from users and provide relevant suggestions and guidance. In addition, the server also designs the interactive behavior data of the digital human based on the behavioral pattern characteristics of the target user, so that it can interact with the user naturally and smoothly.

[0083] S104: Generate a target digital human corresponding to the digital human generation requirement based on the digital human's appearance data, the sound data, the interactive behavior data, the action library, and the knowledge library.

[0084] The server applies the generated appearance data, voice data, and action library to the digital human model, giving it a specific appearance, voice, and action. At the same time, the server designs the interaction logic and process of the digital human based on the interaction behavior data, so that it can react and answer accordingly based on the user's input. With the support of the knowledge base, the digital human can answer the user's questions and provide relevant knowledge and information. In order to ensure the quality and performance of the digital human, the server also conducts rigorous testing and optimization on it. The server uses various evaluation indicators to evaluate the performance of the digital human, such as accuracy, naturalness, fluency, etc. Based on the evaluation results, the server adjusts and optimizes the digital human to improve its performance and user experience. Finally, the server generates a digital human that meets the needs of the target user, with a unique appearance, voice, and personality, which can interact with the user naturally and smoothly and provide valuable services and support.

[0085] The method provided in the embodiment of the present application obtains the digital human generation demand of the target user through the user interface, and uses a machine learning algorithm to analyze the digital human generation demand to obtain the preference characteristics, emotional characteristics, and behavior pattern characteristics of the target user. The machine learning algorithm is used to analyze the digital human generation demand to obtain the preference characteristics, emotional characteristics, and behavior pattern characteristics of the target user, and a target digital human corresponding to the digital human generation demand is generated according to the digital human's appearance data, the sound data, the interactive behavior data, the action library, and the knowledge base, thereby improving the personalized customization level of the digital human and the processing ability of the digital human.

[0086] In a possible implementation, the aforementioned step S102 may include the following sub-steps:

[0087] S1021. Using natural language processing technology to identify the digital human generation requirements, and obtaining a keyword set in the digital human generation requirements.

[0088] The keyword set represents the concerns and needs of the target users.

[0089] The server first uses a lexical analyzer to analyze the digital human generation requirements input by the user. The lexical analyzer will divide the text into words, punctuation marks, and other lexical elements, and mark the part of speech for each word, such as noun, verb, adjective, etc. Then, the syntactic analyzer will further analyze the text after lexical analysis to determine the structure and grammatical relationship of the sentence. The syntactic analyzer will identify the subject, predicate, object and other components in the sentence and analyze the relationship between them. Then, the named entity recognizer will identify the named entities in the text, such as names of people, places, and names of organizations. Finally, the keyword extractor will extract the keywords that best represent the topic of the text from the text. The keyword extractor can use statistical-based methods or semantic-based methods to extract keywords. Through the above steps, the server can obtain a set of keywords from the digital human generation requirements, which represent the focus and needs of the target users.

[0090] For example, the user inputs a digital human generation requirement of "I want a digital human that looks young and energetic, has a cheerful personality, is knowledgeable, and can have interesting conversations with me." The server uses natural language processing technology to analyze the requirement and extracts a keyword set {young and energetic, cheerful personality, knowledgeable, interesting conversation}.

[0091] S1022: Perform lexical analysis and syntactic analysis on each keyword in the keyword set, and use a pre-built semantic knowledge base and a clustering analysis algorithm to extract topic data related to the preferences of the target user from the keyword set.

[0092] The server first performs lexical analysis on each keyword in the keyword set to determine its part of speech, stem, and other lexical features. Then, the server performs syntactic analysis on each keyword to determine its role and grammatical relationship in the sentence. Next, the server uses a pre-built semantic knowledge base, which contains a large number of words, concepts, and the semantic relationships between them. The server matches the keywords in the keyword set with the concepts in the semantic knowledge base to obtain richer semantic information. Finally, the server uses a clustering analysis algorithm to cluster the keyword set and group similar keywords together. Through the above steps, the server can extract topic data related to the preferences of the target user from the keyword set.

[0093] Continuing with the above example, the server performs lexical analysis and syntactic analysis on the keyword set {young and energetic, cheerful, knowledgeable, interesting conversation} to determine the part of speech and grammatical relationship of each keyword. Then, the server uses the pre-built semantic knowledge base to match these keywords with the concepts in the knowledge base. Next, the server uses a clustering analysis algorithm to cluster the keyword set and group similar keywords together. Finally, the server extracts topic data related to the target user's preferences from the clustering results, such as "appearance characteristics", "personality characteristics", "knowledge level", "communication skills", etc.

[0094] S1023: using a sentiment classification algorithm to classify the sentiment annotations in the keyword set, and determining the overall sentiment tendency of the target user towards the digital human.

[0095] The sentiment annotation is generated by sentiment annotation of each keyword in the keyword set through a sentiment annotation tool, and the overall sentiment tendency includes sentiment intensity and sentiment motivation.

[0096] The server first uses the sentiment annotation tool to perform sentiment annotation on each keyword in the keyword set, such as positive, negative, or neutral. Then, the server uses the sentiment classification algorithm to classify the sentiment-annotated keyword set to determine the target user's overall sentiment tendency towards the digital human. The sentiment classification algorithm can use a dictionary-based method, a machine learning-based method, or a deep learning-based method. The server can also determine the target user's emotional intensity towards the digital human, such as strong, medium, or slight. Finally, the server can also determine the target user's emotional motivation towards the digital human, such as like, hate, or have no emotion. Through the above steps, the server can determine the target user's overall emotional tendency towards the digital human, including emotional intensity and emotional motivation.

[0097] Taking the above example, the server uses the emotion annotation tool to perform emotion annotation on each keyword in the keyword set {young and energetic, cheerful, knowledgeable, interesting conversation}, such as "young and energetic" is marked as positive, "cheerful" is marked as positive, "knowledgeable" is marked as positive, and "interesting conversation" is marked as positive. Then, the server uses the emotion classification algorithm to classify the emotion-annotated keyword set and determines that the target user's overall emotional tendency towards the digital human is positive. In addition, the server can also determine that the target user's emotional intensity towards the digital human is strong and the emotional motivation is liking.

[0098] S1024. Based on a behavior pattern recognition algorithm, the historical interaction behaviors of the target user are collected, and behavior pattern recognition and classification are performed on each keyword in the keyword set to determine the interaction mode and behavior habits between the target user and the digital human.

[0099] The server first uses a behavior pattern recognition algorithm to analyze the historical interaction behavior of the target user and extract the user's behavior pattern characteristics. The behavior pattern recognition algorithm can use a rule-based method, a machine learning method, or a deep learning method. Then, the server matches each keyword in the keyword set with the behavior pattern characteristics to determine the association between the keyword and the behavior pattern. Finally, based on the association between the keyword and the behavior pattern, the server determines the interaction method and behavior habits between the target user and the digital human.

[0100] For example, suppose the historical interaction behaviors of the target user include frequently asking questions to the digital person, liking to have conversations with the digital person, and giving positive feedback to the digital person's answers. The server uses a behavior pattern recognition algorithm to analyze these behaviors and extract the behavior pattern characteristics of the target user, such as "asking questions", "conversation", "positive feedback", etc. Then, the server matches each keyword in the keyword set {young and energetic, cheerful, knowledgeable, interesting conversation} with the behavior pattern characteristics, and finds that "conversation" and "interesting conversation" match the behavior pattern characteristics of the target user. Therefore, the server determines that the target user's interaction method with the digital person is conversation, and the behavior habit is liking to have interesting conversations with the digital person.

[0101] S1025. Determine the preference characteristics, emotional characteristics, and behavioral pattern characteristics of the target user based on the topic data, the overall emotional tendency, and the interaction mode and behavioral habits.

[0102] The server first determines the target user's preference characteristics for the digital human based on the subject data, such as appearance characteristics, personality characteristics, knowledge level, communication ability, etc. Then, the server determines the target user's emotional characteristics for the digital human based on the overall emotional tendency, such as liking, positivity, etc. Finally, the server determines the target user's behavioral pattern characteristics based on the interaction method and behavioral habits, such as asking questions, dialogue, positive feedback, etc. Through the above steps, the server can determine the target user's preference characteristics, emotional characteristics, and behavioral pattern characteristics, which will provide important references for the subsequent generation of digital humans that meet the needs of the target users.

[0103] Based on the results of the above examples, the server determines that the target user's preference characteristics are young and energetic appearance, cheerful personality, extensive knowledge, and strong communication skills; emotional characteristics are liking and being positive; and behavioral pattern characteristics are liking to ask questions, have conversations, and give positive feedback. These characteristics will provide important references for the subsequent generation of digital humans that meet the needs of the target user.

[0104] In a possible implementation, the aforementioned step S103 may include the following sub-steps:

[0105] S1031. Extracting the target user's appearance preference features, voice preference features, and knowledge domain features for the digital human from the preference features.

[0106] The server first analyzes and processes the target user's preference features. Preference features may include the user's specific requirements and preferences for the digital human's appearance, voice, and knowledge domain. The server uses natural language processing technology and machine learning algorithms to extract and classify preference features. By learning and training a large amount of user data, the server can identify the preference patterns and features of different users. The server then stores and manages the extracted appearance preference features, voice preference features, and knowledge domain features for use in subsequent digital human generation.

[0107] For example, suppose the target user's preference features include "young and energetic appearance", "gentle and friendly voice", and "strong interest in history and scientific knowledge". By analyzing these features, the server extracts the appearance preference features as "young" and "energetic", the voice preference features as "gentle" and "friendly", and the knowledge domain features as "history" and "science".

[0108] S1032: Use a generative adversarial network and a variational autoencoder to process the target user's appearance preference features for the digital person, and obtain the appearance data of the digital person.

[0109] The appearance data includes the facial feature data, body feature data, age data, gender data, and skin color data of the digital person.

[0110] The server uses deep learning technologies such as generative adversarial networks and variational autoencoders to process the target user's preference characteristics for the appearance of digital humans. GAN consists of a generator and a discriminator. The generator tries to generate realistic digital human appearance, while the discriminator determines whether the generated appearance is real. VAE can generate new samples by learning the potential representation of data. The server inputs the target user's appearance preference characteristics into GAN or VAE, allowing it to learn and generate digital human appearance data that meets these characteristics. This appearance data may include detailed information such as facial features, hairstyle, and skin color.

[0111] Taking the generation of a young and energetic appearance as an example, the server inputs the target user's appearance preference features "young" and "energetic" into the GAN. The GAN generator will generate a digital human image with young appearance features based on these features, such as smooth skin, bright eyes, and energetic expressions. The discriminator will determine whether the generated image meets the characteristics of a young and energetic appearance. Through continuous training and optimization, GAN can generate increasingly realistic young and energetic digital human appearances.

[0112] S1033: Use speech synthesis technology to process the target user's voice preference characteristics for the digital human, and obtain the voice data of the digital human.

[0113] The sound data includes language style data and emotion style data.

[0114] The server uses speech synthesis technology to convert the target user's preferred characteristics of the digital human's voice into actual voice data. Speech synthesis technology can generate voices with a specific style by synthesizing voice features such as rhythm, pitch, and timbre. The server inputs the target user's preferred characteristics of the voice, such as a gentle and friendly voice, into the speech synthesis system. The speech synthesis system will generate corresponding voice data based on these characteristics, including language style data and emotional style data.

[0115] For example, if the target user wants the digital human to have a gentle and friendly voice, the server will input the voice preference feature into the speech synthesis system. The speech synthesis system will adjust the rhythm, pitch and timbre of the voice according to the gentle and friendly features to generate voice data that sounds gentle and friendly. These voice data can include different tones, speech speeds and emotional expressions to meet the user's requirements for the digital human's voice.

[0116] S1034. According to the knowledge domain feature, obtain a knowledge domain corresponding to the knowledge domain feature and / or a knowledge base of related knowledge domains.

[0117] The knowledge base is composed of knowledge extracted from multiple knowledge fields and text data in the multiple knowledge fields.

[0118] The server determines the corresponding knowledge domain and / or related knowledge domain according to the knowledge domain characteristics of the target user. Then, the server obtains the knowledge and information related to these knowledge domains from the knowledge base. The knowledge base can be constructed based on multiple knowledge domains and knowledge extracted from text data of multiple knowledge domains. The server can use natural language processing technology and information retrieval technology to search and extract knowledge related to the needs of the target user from the knowledge base.

[0119] For example, if the target user's knowledge domain characteristics include history and science, the server will obtain knowledge and information related to history and science from the knowledge base. The knowledge base may contain historical events, scientific principles, scientists' biographies, etc. The server can use natural language processing technology to analyze and organize this knowledge so that the digital human can answer the user's questions about history and science and provide relevant knowledge and explanations.

[0120] S1035: Input the behavior pattern features into an interactive behavior model based on a reinforcement learning algorithm, and generate the interactive behavior data and the action library based on the behavior pattern features.

[0121] The actions included in the action library are obtained from a pre-built digital human action library according to the behavior pattern characteristics, and the pre-built digital human action library is generated by capturing and converting real human actions through a motion capture algorithm and a motion generation algorithm.

[0122] The server inputs the target user's behavior pattern characteristics into the interactive behavior model based on the reinforcement learning algorithm. The reinforcement learning algorithm learns the optimal behavior strategy by interacting with the environment and based on the reward signal. The interactive behavior model generates corresponding interactive behavior data and action library based on the input behavior pattern characteristics.

[0123] For the generation of the action library, the server will obtain relevant actions from the pre-built digital human action library according to the characteristics of the behavior pattern. The pre-built digital human action library is generated by capturing and converting real human actions through motion capture algorithms and motion generation algorithms. These actions can include various postures, gestures, facial expressions, etc.

[0124] For example, if the target users' behavioral pattern characteristics indicate that they tend to actively interact with digital humans and like to use rich gestures and expressions, the interactive behavior model based on the reinforcement learning algorithm will generate corresponding interactive behavior data based on these characteristics, such as the digital human actively initiating conversations, using various gestures to emphasize key points, and showing different expressions to express emotions.

[0125] At the same time, the action library will contain actions related to active interaction and rich gesture expressions. These actions may be obtained from the performances of real people through motion capture technology, and processed and transformed by motion generation algorithms to adapt to the expression of digital people.

[0126] In a possible implementation, the aforementioned step S104 may include the following sub-steps:

[0127] S1041. Select an appropriate 3D modeling technology according to the appearance data to construct a three-dimensional appearance model of the digital human, and generate the appearance effect of the target digital human through material rendering processing and lighting rendering processing.

[0128] The server first selects the appropriate 3D modeling technology based on the appearance data, such as polygon modeling, surface modeling, or voxel modeling. Then, the selected modeling technology is used to build a three-dimensional appearance model of the digital human, including facial features, body proportions, hairstyle, etc. Next, material rendering is performed to add realistic materials to the surface of the digital human, such as skin, clothing, hair, etc. The selection and adjustment of materials can be based on the needs and preferences of the user. Finally, through lighting rendering, appropriate lighting effects are added to the digital human, such as ambient light, direct light, shadows, etc., to enhance the realism and three-dimensionality of the digital human.

[0129] For example, suppose the appearance data shows that the target digital human has short hair, blue eyes, and white skin. The server selects polygon modeling technology to build a 3D appearance model of the digital human, using appropriate polygon shapes and textures to simulate short hair and skin. Then, a blue material is selected for the digital human's eyes, and the glossiness and reflectivity of the material are adjusted to make it look more realistic. Next, appropriate lighting conditions are set, such as simulating natural light or indoor lighting, to add shadow and highlight effects to the digital human, so that it can present a natural appearance under different lighting environments.

[0130] S1042: Acquire the voice intonation, voice speed and voice emotion of the target digital person according to the sound data, and generate a voice library of the target digital person.

[0131] The speech in the speech database is speech that has been processed by noise reduction and enhancement.

[0132] The server first analyzes and processes the sound data to extract features such as voice intonation, voice speed, and voice emotion. Voice intonation can be determined by analyzing parameters such as voice pitch, sound intensity, and duration. Voice speed can be determined by calculating the number of voice frames or syllables per unit time. Voice emotion can be determined by analyzing features such as the emotional color of the voice, changes in intonation, and changes in speech speed. Then, based on the extracted features, speech synthesis technology is used to generate the voice library of the target digital person. Speech synthesis technology can convert text into speech, and can adjust the intonation, speech speed, and emotion of the speech according to user needs.

[0133] For example, suppose the sound data shows that the target digital person's voice has a gentle tone, a moderate speaking speed, and a positive emotion. The server generates a voice library of the target digital person using speech synthesis technology, and adjusts the speech synthesis parameters so that the generated voice has a gentle tone, a moderate speaking speed, and a positive emotion. The generated voice is then stored in the voice library for use when needed.

[0134] S1043. Design the dialogue logic and response method of the target digital person based on the interactive behavior data, and train the dialogue logic and response method through deep learning technology, so that the target digital person can query and extract the knowledge information corresponding to the input from the knowledge base based on the user's input and make corresponding responses.

[0135] The server first designs the target digital person's dialogue logic and response method based on the interactive behavior data. The dialogue logic can include different dialogue modes such as greetings, questions, answers, and explanations. The response method can include different response forms such as text response, voice response, and expression response. Then, deep learning technology is used to train the dialogue logic and response method. Deep learning technology can automatically extract patterns and rules in the dialogue and generate corresponding responses by learning from a large amount of dialogue data. During the training process, supervised learning or reinforcement learning methods can be used to continuously adjust and optimize the dialogue logic and response method based on user input and feedback.

[0136] For example, suppose the interactive behavior data shows that the target digital human needs to be able to answer questions about history and science and be able to have simple daily conversations with users. The server designs the dialogue logic and response method of the target digital human based on these requirements. For example, when the user asks about historical events, the digital human can provide relevant historical background, event process and impact information; when the user asks about scientific knowledge, the digital human can provide relevant scientific principles, experimental results and applications. Then, deep learning technology is used to train the dialogue logic and response method. By learning from a large amount of historical and scientific texts, the digital human can understand the user's questions and query and extract relevant knowledge information from the knowledge base to make corresponding responses.

[0137] S1044. Based on the actions included in the action library, as well as the dialogue logic and response method of the target digital person, an association relationship between the actions and the dialogue logic and response method of the target digital person is constructed to synchronize the actions with the dialogue logic and response method.

[0138] The server first constructs the association between actions and dialogue logic and response methods based on the actions included in the action library and the dialogue logic and response methods of the target digital person. The association between actions and dialogue logic and response methods may include the correspondence between actions and dialogue content, the correspondence between actions and dialogue emotions, etc. Then, based on the constructed association, corresponding actions are added to the dialogue logic and response methods of the target digital person. The addition of actions can be achieved through animation production or motion capture and other technologies.

[0139] For example, suppose the action library includes nodding, shaking head, smiling, frowning and other actions, and the target digital person's dialogue logic and response methods include affirmative answer, negative answer, happy, unhappy, etc. The server builds the association between actions and dialogue logic and response methods based on this information. For example, when the target digital person makes an affirmative answer, a nodding action is added at the same time; when the target digital person makes a negative answer, a shaking head action is added at the same time; when the target digital person is happy, a smiling action is added at the same time; when the target digital person is unhappy, a frowning action is added at the same time. Then, corresponding actions are added to the dialogue logic and response methods of the target digital person through animation production or motion capture technology to synchronize the actions with the dialogue logic and response methods.

[0140] S1045, integrating the three-dimensional appearance model, the voice library, the dialogue logic, the response method, and the association between the actions in the action library and the dialogue logic and the response method, and debugging to generate the target digital human corresponding to the digital human generation requirements.

[0141] The server first integrates the 3D appearance model, voice library, dialogue logic, response method, and the relationship between actions in the action library and dialogue logic and response method to form a complete digital human system. Then, the integrated digital human system is debugged and optimized to ensure its performance and stability. The debugging and optimization content can include the accuracy of the model, the naturalness of the voice, the fluency of the dialogue, the coordination of the action, etc. Finally, according to the needs and feedback of the users, the digital human system is further adjusted and improved to generate the target digital human that meets the needs of the users.

[0142] For example, suppose that the integrated digital human system has some problems in performance and stability, such as unnatural voice, unsmooth conversation, uncoordinated movements, etc. The server debugs and optimizes these problems, and improves the performance and stability of the digital human system by adjusting speech synthesis parameters, optimizing conversation logic and response methods, and improving motion capture technology. Then, according to user needs and feedback, the digital human system is further adjusted and improved, for example, more expressions and movements are added, and the emotional expression of voice is optimized, so as to generate a more realistic and natural target digital human.

[0143] The method provided in the embodiment of the present application can realize the generation of personalized digital humans according to the needs of target users. The appearance of digital humans is improved through technologies such as 3D modeling, material rendering, and lighting rendering. The dialogue logic and response methods of digital humans are trained through deep learning technology, so that they can intelligently understand user input and respond accordingly. Speech synthesis technology and action library enable digital humans to express voice intonation, speech speed and emotion, and enhance the interaction effect with users. Digital humans can also query and extract cross-domain knowledge from the knowledge base to provide comprehensive and accurate information. The synchronization of actions and dialogues makes their performance more natural and coordinated. In addition, according to user needs and feedback, digital humans can also be debugged and optimized to continuously improve their performance and user experience. .

[0144] Optionally, the above method may also include the target digital human providing real-time feedback on the content of the response information according to the dialogue information input by the target user. Figure 2 A flowchart of another method for generating a digital human provided in an embodiment of the present application. Figure 2 As shown, the method may also include:

[0145] S201: Acquire the conversation information of the target user.

[0146] The dialogue information is used to call the target digital human to output response information corresponding to the dialogue information.

[0147] The server can obtain the target user's conversation information in a variety of ways. First, the server can receive the text or voice information entered by the user on the client through a network connection. Users can use various devices, such as mobile phones, computers, etc., to communicate with the server through the network. After receiving the user's input, the server decodes and parses it and converts it into a format that can be processed by the computer. In addition, the server can also obtain the user's non-verbal information, such as expressions, actions, etc., through sensors, cameras and other devices. This non-verbal information can provide more clues about the user's emotions and intentions, helping the server to better understand the user's needs. Finally, the server can also obtain user-related information from other data sources, such as the user's historical conversation records, preference settings, etc. By comprehensively analyzing this information, the server can have a more comprehensive understanding of the user's background and needs, and thus provide more accurate and personalized responses.

[0148] For example, suppose the target user has a conversation with the target digital person through a chat application on his mobile phone. The user enters a text in the chat application: "Hello, I want to learn about the application of artificial intelligence in the medical field." After receiving the user's input, the server decodes and parses it and converts it into a format that can be processed by the computer. The server can also obtain the user's facial expression information, such as smiling, frowning, etc., through the mobile phone's camera to understand the user's emotional state. At the same time, the server can also obtain relevant information from the user's historical conversation records, such as whether the user has mentioned topics in the medical field before, and what specific aspects are of interest. By comprehensively analyzing this information, the server can understand the user's needs more accurately and generate more targeted responses.

[0149] S202: Input the dialogue information into the large language model of the target digital human to generate the response information.

[0150] The server inputs the acquired conversation information into the big language model of the target digital human. The big language model is a natural language processing model based on deep learning, which can understand and analyze the input text and generate corresponding responses. The big language model is usually trained using a large-scale corpus to learn the patterns and rules of the language. During the training process, the model learns how to generate appropriate responses based on the input text to improve the accuracy and quality of the answers. When the server inputs the conversation information into the big language model, the model parses and understands the input text and generates corresponding responses based on the learned knowledge and patterns. The response generation process usually involves multiple steps such as lexical analysis, syntactic analysis, and semantic understanding to ensure that the generated response conforms to grammatical and semantic rules.

[0151] Specifically, this step can be implemented through the following sub-steps:

[0152] S2021. Use named entity recognition and keyword extraction technology to perform semantic understanding on the conversation information to obtain the intention characteristics, demand characteristics, and emotional characteristics of the target user.

[0153] The server uses named entity recognition technology to identify entities with specific meanings from the conversation information, such as names of people, places, and organizations. These entities can provide important information about the topic and background of the conversation. Then, the server uses keyword extraction technology to extract keywords that best represent the user's intentions and needs from the conversation information. These keywords can help the server quickly understand the user's main concerns. Next, the server performs semantic analysis on the named entities and keywords to obtain a deeper semantic understanding. Semantic analysis can include techniques such as part-of-speech tagging, syntactic analysis, and semantic role tagging to determine the relationship between named entities and keywords and their role in the sentence. Finally, based on the results of semantic understanding, the server extracts the target user's intention features, demand features, and emotional features. Intent features can include the user's questions, requests, or instructions, etc.; demand features can include the user's needs for specific information or services; and emotional features can include the user's emotional state, such as happiness, sadness, anger, etc.

[0154] S2022. Process the intention characteristics and demand characteristics of the target user through a cluster analysis algorithm, obtain the knowledge base category corresponding to the intention characteristics and demand characteristics of the target user, and determine the target knowledge base corresponding to the intention characteristics and demand characteristics from the knowledge base set based on the knowledge base category.

[0155] The server preprocesses the intent features and demand features of the target user and converts them into a format suitable for cluster analysis. Preprocessing can include operations such as data cleaning, feature extraction, and standardization. Then, the server uses a clustering analysis algorithm to cluster the preprocessed intent features and demand features. The clustering analysis algorithm can divide the data into different clusters based on their similarity, and each cluster represents a knowledge base category. Based on the clustering results, the server determines the knowledge base category to which the intent features and demand features of the target user belong. Finally, the server selects the corresponding target knowledge base from the knowledge base set based on the knowledge base category. The target knowledge base is a database containing knowledge and information related to the knowledge base category.

[0156] S2023. Perform knowledge retrieval in the target knowledge base to obtain candidate response content corresponding to the intention characteristics and demand characteristics.

[0157] The server constructs a query statement based on the intent features and demand features. The query statement is an expression used to search for relevant knowledge and information in the target knowledge base. Then, the server uses the knowledge retrieval engine to execute the query statement in the target knowledge base. The knowledge retrieval engine can find matching knowledge and information in the target knowledge base based on the content of the query statement. The knowledge retrieval engine can use indexing, full-text search or other technologies to improve retrieval efficiency. Finally, the server obtains candidate response content corresponding to the intent features and demand features based on the retrieval results. Candidate response content refers to knowledge and information found in the target knowledge base that may meet user needs.

[0158] Specifically, this step can be implemented through the following sub-steps:

[0159] S1. Construct a target knowledge graph in the target knowledge base in advance.

[0160] Among them, the nodes of the target knowledge graph are knowledge entities, the edges of the target knowledge graph are the relationships between the knowledge entities, and the knowledge entities and the relationships are generated by preprocessing and annotation of a natural language processing model.

[0161] The server uses a natural language processing model to preprocess and annotate the knowledge entities and relationships in the target knowledge base. The preprocessing and annotation process includes operations such as named entity recognition, part-of-speech tagging, and semantic analysis of knowledge entities to extract key information and semantic features of knowledge entities. The server then uses the annotated knowledge entities and relationships to build a target knowledge graph. The target knowledge graph is a graph-based data structure in which nodes represent knowledge entities and edges represent relationships between knowledge entities. The server can use a graph database or other graph processing technologies to store and manage the target knowledge graph. When building the target knowledge graph, the server needs to ensure the accuracy and completeness of knowledge entities and relationships to improve the effect and quality of knowledge retrieval.

[0162] Exemplarily, assume that the target knowledge base contains knowledge about fields such as history, science, culture, and technology. The server first uses a natural language processing model to preprocess and annotate this knowledge, and extracts the knowledge entities and relationships therein. For example, in the field of history, knowledge entities can include historical events, historical figures, historical periods, etc., and relationships can include time sequence, causal relationships, geographical locations, etc. Then, the server uses the annotated knowledge entities and relationships to construct a target knowledge graph. In the knowledge graph, historical events can be used as nodes, and the time sequence relationships between them can be represented by edges. Similarly, knowledge entities such as historical figures and historical periods can also be used as nodes, and the relationships between them can be represented by edges. By constructing the target knowledge graph, the server can organize the knowledge in the knowledge base in an intuitive and easy-to-understand way, which facilitates subsequent knowledge retrieval and application.

[0163] S2. Retrieve the knowledge entities corresponding to the intention feature and the demand feature in the target knowledge graph through a depth-first search algorithm to obtain candidate knowledge entities and associated knowledge entities associated with the candidate knowledge entities.

[0164] The server represents the target knowledge graph as a graph data structure, where nodes represent knowledge entities and edges represent relationships between knowledge entities. The server then uses a depth-first search algorithm to traverse the graph starting from the starting node, recursively visiting neighbor nodes along the edges until it reaches the target node or cannot proceed further. During the traversal process, the server records the visited nodes and edges to avoid repeated visits. When the server finds a knowledge entity that corresponds to the intent feature and the demand feature, it marks it as a candidate knowledge entity. At the same time, the server also obtains associated knowledge entities associated with the candidate knowledge entity, which can provide more relevant information and knowledge. Finally, the server returns the candidate knowledge entity and the associated knowledge entity to subsequent processing steps.

[0165] S3. According to the candidate knowledge entity and the relationship between the candidate knowledge entity and the associated knowledge entity, the candidate knowledge entity and the associated knowledge entity are integrated to generate the candidate response content.

[0166] The server integrates and organizes the candidate knowledge entities and the associated knowledge entities according to the relationships between them. The integration and organization process may include operations such as classifying, sorting, and filtering the knowledge entities to extract the most relevant and valuable information. Then, the server uses natural language generation technology to convert the integrated knowledge entities and relationships into natural language text to generate candidate response content. Natural language generation technology can generate natural language text that conforms to grammatical and semantic rules based on the characteristics and context of knowledge entities and relationships. Finally, the server returns the candidate response content to the target user as a response to his query.

[0167] S2024. Generate candidate response information by reasoning and integrating the candidate response contents.

[0168] The server first analyzes and understands the candidate response content and extracts the key information and knowledge points. Then, the server uses the reasoning algorithm to reason and integrate these key information and knowledge points. The reasoning algorithm can derive new conclusions and information based on known knowledge and rules. Finally, the server generates candidate response information based on the results of reasoning and integration. Candidate response information refers to the response content after reasoning and integration, which is more logical and accurate.

[0169] S2025. Determine the language expression and tone corresponding to the emotional characteristics based on the candidate response information and the emotional characteristics, and adjust the candidate response information through the language expression and tone to generate the response information.

[0170] The server determines the corresponding language expression and tone based on the emotional characteristics. Different emotional characteristics may require different language expressions and tones to convey. For example, positive emotional characteristics may require the use of cheerful and friendly language expressions and tones; negative emotional characteristics may require the use of comforting and caring language expressions and tones. Then, the server adjusts the candidate response information based on the determined language expressions and tones. The adjustments may include vocabulary selection, sentence adjustment, tone changes, and other operations to make the response information more in line with the requirements of the emotional characteristics. Finally, the server generates the final response information and sends it to the target user.

[0171] S203: Feedback the generated response information to the target user through the target digital human.

[0172] The server sends the response information generated by the target digital human to the client through the network connection, and the client displays the response information to the target user. The server can use various communication protocols and technologies, such as HTTP, WebSocket, etc., to transmit the response information to the client. After receiving the response information, the client selects an appropriate method to display it to the user based on the user's device and settings. For example, if the user is using a mobile client, the response information can be displayed to the user in the form of text, voice or video. If the user is using a computer client, the response information can be displayed to the user in the form of pop-ups, message prompts, etc. The response information is fed back to the target user through the target digital human, so that the user can obtain the required information in a timely manner and further interact with the target digital human.

[0173] For example, the server sends the response generated by the large language model, "Artificial intelligence has many applications in the medical field, such as medical imaging diagnosis, disease prediction, and drug development. Medical imaging diagnosis uses artificial intelligence technology to analyze and diagnose medical images to help doctors detect diseases more accurately. Disease prediction is to predict the occurrence and development of diseases by analyzing patients' medical data and take preventive measures in advance. Drug development uses artificial intelligence technology to accelerate the drug development process and improve research and development efficiency and success rate." to the user's mobile client through a network connection. After receiving the response information, the mobile client displays it to the user in text form. The user can see the response of the target digital person on the mobile phone screen and continue to talk to the target digital person.

[0174] The method provided in the embodiment of the present application, by inputting the dialogue information into the large language model of the target digital person, utilizes the powerful language understanding and generation capabilities of the large language model, and can quickly and accurately generate responses that meet the user's needs. At the same time, by processing the intention features and demand features through the cluster analysis algorithm, the user's questions can be classified into different knowledge base categories, improving the accuracy and efficiency of knowledge retrieval. Knowledge retrieval is performed in the target knowledge base, candidate response content is obtained, and candidate response information is generated through reasoning and integration, which further improves the quality and accuracy of the response, thereby improving the processing ability of the target digital person and improving the accuracy of the generated response information.

[0175] Optionally, the above method may also include the target digital human providing real-time feedback on the content of the response information based on the dialogue information input by the target user through cross-domain integration. Figure 3 A flowchart of another method for generating a digital human provided in an embodiment of the present application is shown below. Figure 3 As shown, the method may also include:

[0176] S301. Process the intention characteristics and demand characteristics of the target user through a cluster analysis algorithm to obtain the knowledge base categories corresponding to the intention characteristics and demand characteristics of the target user, and determine the first target knowledge base and the second target knowledge base corresponding to the intention characteristics and demand characteristics from the knowledge base set according to the knowledge base categories.

[0177] The first target knowledge base and the second target knowledge base are in different knowledge fields.

[0178] The server first preprocesses the target user's intent features and demand features, including data cleaning, feature extraction, and standardization. Then, clustering analysis algorithms, such as the K-Means algorithm, are used to cluster the preprocessed data. The purpose of clustering is to group similar intent features and demand features together for subsequent processing.

[0179] During the clustering process, the server will iteratively calculate the data according to the preset clustering parameters, such as the number of clusters, the number of iterations, etc., until the clustering conditions are met. After the clustering is completed, the server will determine the knowledge base category corresponding to each cluster based on the clustering results. These knowledge base categories can be divided according to knowledge fields, problem types or other related factors.

[0180] Next, the server selects a first target knowledge base and a second target knowledge base corresponding to the intention feature and the demand feature from the knowledge base set according to the determined knowledge base category. The first target knowledge base and the second target knowledge base have different knowledge fields to meet the cross-domain knowledge needs of the target user.

[0181] Exemplarily, assume that the target user's intention feature is to learn about healthy eating and fitness knowledge, and the demand feature is to obtain advice on weight loss and balanced nutrition. After preprocessing these features, the server uses the K-Means algorithm for clustering. The clustering results show that these features can be divided into two clusters: one cluster is related to healthy eating, and the other cluster is related to fitness. Based on the clustering results, the server determines two knowledge base categories: a healthy eating knowledge base and a fitness knowledge base. Then, the server selects the first target knowledge base and the second target knowledge base corresponding to these two knowledge base categories from the knowledge base set. The first target knowledge base contains knowledge about healthy eating, such as food nutrition, eating habits, etc.; the second target knowledge base contains knowledge about fitness, such as types of exercise, fitness plans, etc.

[0182] In this way, the server can accurately select the corresponding knowledge base according to the target user's intention characteristics and demand characteristics, providing a basis for subsequent knowledge retrieval and response generation.

[0183] S302. Decompose the intention characteristics and demand characteristics of the target user into first intention characteristics and demand characteristics and second intention characteristics and demand characteristics according to the first knowledge domain of the first target knowledge base and the second knowledge domain of the second target knowledge base.

[0184] The first intention feature and the demand feature are related to the first knowledge domain, and the second intention feature and the demand feature are related to the second knowledge domain.

[0185] The server first analyzes the knowledge domains of the first target knowledge base and the second target knowledge base to determine the specific knowledge scope covered by each knowledge base. Then, based on the results of the domain knowledge analysis, the target user's intention features and demand features are decomposed into first intention features and demand features related to the first knowledge domain, and second intention features and demand features related to the second knowledge domain.

[0186] During the decomposition process, the server will perform reasonable decomposition and division based on the specific content of the intent features and demand features, as well as the knowledge structure of the knowledge base. For example, the intent features and demand features about healthy eating can be decomposed into the first intent features and demand features about food nutrition, eating habits, etc.; the intent features and demand features about fitness can be decomposed into the second intent features and demand features about exercise types, fitness plans, etc.

[0187] Finally, the server will establish an association between the first intent feature and the demand feature and the first target knowledge base, and an association between the second intent feature and the demand feature and the second target knowledge base. In this way, in the subsequent knowledge retrieval process, the server can accurately retrieve relevant knowledge and information from the corresponding knowledge base based on the association relationship.

[0188] For example, continuing with the above example, the first target knowledge base is the healthy diet knowledge base, and its knowledge areas include food nutrition, eating habits and other aspects; the second target knowledge base is the fitness knowledge base, and its knowledge areas include exercise types, fitness plans and other aspects.

[0189] The server decomposes the target user's intent feature "learn about healthy eating and fitness knowledge" into the first intent feature "learn about food nutrition and eating habits" related to the healthy eating knowledge base, and the second intent feature "learn about exercise types and fitness plans" related to the fitness knowledge base. At the same time, the demand feature "get advice on weight loss and balanced nutrition" is decomposed into the first demand feature "get advice on weight loss about food nutrition and eating habits" related to the healthy eating knowledge base, and the second demand feature "get advice on balanced nutrition about exercise types and fitness plans" related to the fitness knowledge base.

[0190] In this way, the server is able to decompose the target user's complex intentions and needs into specific and actionable sub-goals, providing a clearer direction for subsequent knowledge retrieval and response generation.

[0191] S303. Perform knowledge retrieval on the first intention feature and demand feature in the first target knowledge base to obtain first candidate response content corresponding to the first intention feature and demand feature; and perform knowledge retrieval on the second intention feature and demand feature in the second target knowledge base to obtain second candidate response content corresponding to the second intention feature and demand feature.

[0192] The server first searches the first target knowledge base using knowledge retrieval techniques, such as keyword search, semantic search, etc., based on the first intent feature and demand feature. During the search process, the server matches and filters the knowledge in the knowledge base according to the preset search strategy and algorithm to find knowledge and information related to the first intent feature and demand feature.

[0193] Similarly, the server performs knowledge retrieval in the second target knowledge base according to the second intention feature and the demand feature. The retrieval results may include knowledge and information in the form of text, image, video, etc.

[0194] Next, the server will sort and analyze the search results, extracting the first candidate response content corresponding to the first intent feature and demand feature, and the second candidate response content corresponding to the second intent feature and demand feature. These candidate response contents can be specific knowledge, suggestions, solutions, etc.

[0195] For example, in the healthy diet knowledge base, the server performs knowledge retrieval based on the first intention feature "understanding food nutrition and eating habits" and the first demand feature "obtaining weight loss advice on food nutrition and eating habits". The retrieval results may include relevant knowledge such as food calories, nutritional components, principles of healthy diet, and some dietary advice during weight loss.

[0196] Similarly, in the fitness knowledge base, the server performs knowledge retrieval based on the second intent feature "understanding exercise types and fitness plans" and the second demand feature "obtaining nutritionally balanced recommendations for exercise types and fitness plans". The retrieval results may include relevant knowledge such as different types of exercise, recommendations for exercise intensity and time, and nutritionally balanced fitness diets.

[0197] In this way, the server can obtain knowledge and information related to the target user's intention characteristics and demand characteristics from different knowledge bases, providing rich materials for subsequent response generation.

[0198] S304: Merge the first candidate response content and the second candidate response content to generate the candidate response content.

[0199] The server merges the first candidate response content and the second candidate response content to generate a comprehensive candidate response content. The merging method can be determined according to the specific situation, which can be a simple merger or integration based on certain rules or algorithms.

[0200] During the fusion process, the server will consider the relationship and logic between the first candidate response content and the second candidate response content to ensure that the fused content is coherent and consistent. For example, if the first candidate response content provides knowledge about food nutrition and eating habits, and the second candidate response content provides knowledge about exercise types and fitness plans, the server will integrate these contents to form a comprehensive suggestion on weight loss and nutritional balance.

[0201] Next, the server will conduct a quality assessment on the fused candidate response content to ensure its accuracy, completeness and readability. If the assessment result is not ideal, the server will optimize and adjust the candidate response content to improve its quality and effectiveness.

[0202] Specifically, the above steps can be implemented through the following sub-steps:

[0203] S3041. Obtain a first knowledge entity corresponding to the first candidate response content, and a second knowledge entity corresponding to the second candidate response content.

[0204] The server uses natural language processing technology to analyze the first candidate response content and the second candidate response content. Natural language processing technology can help the server identify entities in the text, such as people, places, organizations, events, etc. Through lexical analysis, syntactic analysis, and semantic analysis of the text, the server can extract key information from the text and convert it into a form that can be understood by the computer. Next, the server uses knowledge graph technology to process the extracted key information. A knowledge graph is a semantic network that represents the relationship between entities in the form of a graph. The server can match the extracted key information with the entities in the knowledge graph to determine the relationship between them. Finally, the server stores the successfully matched entities in the knowledge base for subsequent processing and use.

[0205] S3042: Determine the type of association between the first knowledge entity and the second knowledge entity by crossing the edge of the first target knowledge base and the second target knowledge base.

[0206] The server uses knowledge graph technology to construct a knowledge graph of the first target knowledge base and the second target knowledge base. The knowledge graph is a semantic network that represents the relationship between entities in the form of a graph. The server can represent the entities and relationships in the first target knowledge base and the second target knowledge base as two different subgraphs, respectively, and then connect them through cross-edges. Next, the server uses a graph algorithm to analyze the connected knowledge graph to determine the type of association between the first knowledge entity and the second knowledge entity. Graph algorithms can help the server identify nodes and edges in the graph and calculate the relationship between them. For example, the server can use the shortest path algorithm to calculate the shortest path between the first knowledge entity and the second knowledge entity, thereby determining the type of association between them. Finally, the server stores the determined association type in the knowledge base for subsequent processing and use.

[0207] S3043. Determine a target content fusion strategy for fusing the first candidate response content and the second candidate response content according to the association type and the mapping relationship between the association type and the content fusion strategy.

[0208] The server uses the knowledge graph technology to construct the knowledge graph of the first target knowledge base and the second target knowledge base. The knowledge graph is a semantic network that represents the relationship between entities in the form of a graph. The server can represent the entities and relationships in the first target knowledge base and the second target knowledge base as two different subgraphs, respectively, and then connect them through cross-edges. Next, the server uses a graph algorithm to analyze the connected knowledge graph to determine the type of association between the first knowledge entity and the second knowledge entity. The graph algorithm can help the server identify the nodes and edges in the graph and calculate the relationship between them. For example, the server can use the shortest path algorithm to calculate the shortest path between the first knowledge entity and the second knowledge entity to determine the type of association between them. Then, the server uses the mapping relationship between the association type and the content fusion strategy to determine the target content fusion strategy for fusing the first candidate response content and the second candidate response content. The mapping relationship between the association type and the content fusion strategy is a predefined rule that corresponds different association types to different content fusion strategies. Finally, the server stores the determined target content fusion strategy in the knowledge base for subsequent processing and use.

[0209] S3044. According to the target content fusion strategy, fuse the first candidate response content and the second candidate response content to generate the candidate response content.

[0210] The server first uses the knowledge graph technology to construct a knowledge graph of the first target knowledge base and the second target knowledge base. The knowledge graph is a semantic network that represents the relationship between entities in the form of a graph. The server can represent the entities and relationships in the first target knowledge base and the second target knowledge base as two different subgraphs, respectively, and then connect them through cross-edges. Next, the server uses a graph algorithm to analyze the connected knowledge graph to determine the type of association between the first knowledge entity and the second knowledge entity. The graph algorithm can help the server identify the nodes and edges in the graph and calculate the relationship between them. For example, the server can use the shortest path algorithm to calculate the shortest path between the first knowledge entity and the second knowledge entity to determine the type of association between them. Then, the server uses the mapping relationship between the association type and the content fusion strategy to determine the target content fusion strategy for fusing the first candidate response content and the second candidate response content. The mapping relationship between the association type and the content fusion strategy is a predefined rule that corresponds different association types to different content fusion strategies. Finally, the server fuses the first candidate response content and the second candidate response content according to the target content fusion strategy to generate the candidate response content. The specific implementation method of content fusion may vary depending on the target content fusion strategy.

[0211] For example, if the target content fusion strategy is "direct splicing", the server can directly splice the first candidate response content and the second candidate response content to generate the candidate response content. If the target content fusion strategy is "association fusion", the server can associate the related entities in the first candidate response content and the second candidate response content, and generate the candidate response content based on the association relationship. If the target content fusion strategy is "geographic fusion", the server can fuse the geographic location information in the first candidate response content and the second candidate response content, and generate the candidate response content based on the fused geographic location information.

[0212] Optionally, the fusing the first candidate response content and the second candidate response content according to the target content fusion strategy to generate the candidate response content may also include:

[0213] According to the target content fusion strategy, the first candidate response content and the second candidate response content are fused to generate candidate response content to be quality evaluated. The server first obtains the target content fusion strategy, which specifies how to fuse the first candidate response content and the second candidate response content. Then, the server uses the corresponding fusion algorithm or method to fuse the first candidate response content and the second candidate response content. The fusion process can consider various factors, such as the relevance, importance, and completeness of the content. After the fusion is completed, the server generates candidate response content to be quality evaluated. This candidate response content is the result of the fusion and needs to be further evaluated and optimized.

[0214] An evaluation score is generated by performing consistency evaluation and accuracy evaluation on the candidate response content to be quality evaluated. The server performs consistency evaluation and accuracy evaluation on the candidate response content to be quality evaluated. The consistency evaluation mainly checks whether the candidate response content is consistent with the intentions and needs of the target user, and whether there are any contradictions or inconsistencies. The accuracy evaluation checks whether the information contained in the candidate response content is accurate, complete, and in line with the actual situation. The specific method of the evaluation can be selected and customized according to the actual situation. For example, natural language processing technology can be used to analyze and understand the candidate response content and compare it with the intentions and needs of the target user; knowledge bases or other data sources can also be used to verify and verify the information in the candidate response content.

[0215] If the evaluation score is less than the preset score threshold, the first candidate response content and the second candidate response content are reintegrated according to the target content fusion strategy. The server compares the evaluation score with the preset score threshold. If the evaluation score is less than the preset score threshold, it means that the quality of the candidate response content does not meet the requirements and needs to be reintegrated. The server selects a suitable fusion algorithm or method based on the target content fusion strategy to reintegrate the first candidate response content and the second candidate response content. The reintegration process can be adjusted and optimized according to the specific situation to improve the quality of the candidate response content.

[0216] For example, suppose the preset score threshold is 80 points, and the evaluation score is 70 points, which is less than the preset score threshold. The server selects a more complex fusion method based on the target content fusion strategy, and reintegrates the first candidate response content "The weather is fine today" and the second candidate response content "Suitable for outdoor activities". The new fusion method takes into account more factors, such as weather temperature, wind speed, etc., as well as the type and safety of outdoor activities. After reintegration, a new candidate response content is generated: "Today's weather is fine, the temperature is moderate, and the wind is light, which is very suitable for outdoor sports such as running, cycling, etc., but please pay attention to safety."

[0217] If the evaluation score is greater than or equal to the preset score threshold, the candidate response content is generated. If the evaluation score is greater than or equal to the preset score threshold, it means that the quality of the candidate response content meets the requirements and can be used as the final candidate response content. The server organizes and optimizes the candidate response content to make it more in line with the user's needs and expectations. Then, the server returns the candidate response content to the target user, or displays and applies it on a related system or platform.

[0218] Figure 4 The structure diagram of an electronic device provided in an embodiment of the present application. The electronic device is used to execute the aforementioned digital human generation method, for example, it can be the aforementioned server. Figure 4 As shown, the electronic device 400 may include: at least one processor 401 , a memory 402 , and a communication interface 403 .

[0219] The memory 402 is used to store programs. Specifically, the programs may include program codes, and the program codes include computer operation instructions.

[0220] The memory 402 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.

[0221] The processor 401 is used to execute the computer-executable instructions stored in the memory 402 to implement the method described in the above method embodiment. The processor 401 may be a CPU, or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0222] The processor 401 can communicate and interact with an external device through the communication interface 403, and the external device can be, for example, a user's terminal device. In a specific implementation, if the communication interface 403, the memory 402 and the processor 401 are implemented independently, the communication interface 403, the memory 402 and the processor 401 can be connected to each other through a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc., but it does not mean that there is only one bus or one type of bus.

[0223] Optionally, in a specific implementation, if the communication interface 403, the memory 402 and the processor 401 are integrated on a chip, the communication interface 403, the memory 402 and the processor 401 can communicate through an internal interface.

[0224] The present application also provides a computer-readable storage medium, which may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, and other media that can store program codes. Specifically, the computer-readable storage medium stores program instructions, and the program instructions are used for the methods in the above embodiments.

[0225] The present application also provides a program product, which includes an execution instruction, which is stored in a readable storage medium. At least one processor of a computing device can read the execution instruction from the readable storage medium, and at least one processor executes the execution instruction so that the computing device implements the above method.

[0226] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for generating a digital human, characterized in that: The method comprises: Acquiring the target user's digital human generation requirements through the user interface, wherein the digital human generation requirements include the digital human's appearance requirements, personality requirements, knowledge domain requirements, language style requirements, and interaction method requirements; Using a machine learning algorithm to analyze the digital human generation needs, and obtain the target user's preference characteristics, emotional characteristics, and behavioral pattern characteristics; Processing the target user's preference characteristics, the emotional characteristics, and the behavioral pattern characteristics based on a deep learning model to generate the digital human's appearance data, sound data, interactive behavior data, action library, and knowledge library corresponding to the digital human generation requirements; A target digital human corresponding to the digital human generation requirement is generated according to the digital human's appearance data, the sound data, the interactive behavior data, the action library, and the knowledge library.

2. The method according to claim 1, characterized in that The use of a machine learning algorithm to analyze the digital human generation needs to obtain the target user's preference characteristics, emotional characteristics, and behavioral pattern characteristics includes: Using natural language processing technology to identify the digital human generation requirements, and obtaining a keyword set in the digital human generation requirements, wherein the keyword set represents the focus and requirements of the target user; Performing lexical analysis and syntactic analysis on each keyword in the keyword set, and using a pre-built semantic knowledge base to use a clustering analysis algorithm to extract topic data related to the preferences of the target user from the keyword set; Using a sentiment classification algorithm to classify the sentiment annotations in the keyword set, determining the overall sentiment tendency of the target user towards the digital human, wherein the sentiment annotations are generated by sentiment annotation of each keyword in the keyword set using a sentiment annotation tool; the overall sentiment tendency includes sentiment intensity and sentiment motivation; Based on the behavior pattern recognition algorithm, the historical interaction behaviors of the target user are collected, and the behavior pattern recognition and classification are performed on each keyword in the keyword set to determine the interaction mode and behavior habits between the target user and the digital human; The preference characteristics, emotional characteristics, and behavioral pattern characteristics of the target user are determined based on the topic data, the overall emotional tendency, and the interaction mode and behavioral habits.

3. The method according to claim 1, characterized in that: The method of processing the target user's preference characteristics, the emotional characteristics, and the behavioral pattern characteristics based on the deep learning model to generate the digital human's appearance data, sound data, interactive behavior data, action library, and knowledge library corresponding to the digital human generation requirements includes: Extracting the target user's appearance preference features, voice preference features, and knowledge domain features for the digital human from the preference features; Using a generative adversarial network and a variational autoencoder to process the target user's preference characteristics for the digital person's appearance, and obtaining the digital person's appearance data, wherein the appearance data includes the digital person's facial feature data, body feature data, age data, gender data, and skin color data; Using speech synthesis technology to process the target user's voice preference characteristics for the digital human, and obtaining the digital human's voice data, wherein the voice data includes language style data and emotional style data; According to the knowledge domain feature, obtaining a knowledge base of the knowledge domain and / or related knowledge domains corresponding to the knowledge domain feature, wherein the knowledge base is composed of knowledge extracted from multiple knowledge domains and multiple text data of the knowledge domains; The behavior pattern features are input into an interactive behavior model based on a reinforcement learning algorithm, and the interactive behavior data and the action library are generated based on the behavior pattern features. The actions included in the action library are obtained from a pre-built digital human action library based on the behavior pattern features. The pre-built digital human action library is generated by capturing and converting real human actions through a motion capture algorithm and a motion generation algorithm.

4. The method according to claim 1, characterized in that: The step of generating a target digital human corresponding to the digital human generation requirement according to the digital human's appearance data, the sound data, the interactive behavior data, the action library, and the knowledge library comprises: Selecting appropriate 3D modeling technology according to the appearance data to construct a three-dimensional appearance model of the digital human, and generating the appearance effect of the target digital human through material rendering processing and lighting rendering processing; Acquire the voice intonation, voice speed and voice emotion of the target digital person according to the sound data, and generate a voice library of the target digital person, wherein the voice in the voice library is the voice after noise reduction and enhancement processing; Design the dialogue logic and response method of the target digital person based on the interactive behavior data, and train the dialogue logic and response method through deep learning technology, so that the target digital person can query and extract knowledge information corresponding to the input from the knowledge base based on the user's input and make corresponding responses; Based on the actions included in the action library, and the dialogue logic and response mode of the target digital person, an association relationship between the actions and the dialogue logic and response mode of the target digital person is constructed to synchronize the actions with the dialogue logic and response mode; The three-dimensional appearance model, the voice library, the dialogue logic, the response method, and the association between the actions in the action library and the dialogue logic and the response method are integrated to debug and generate the target digital human corresponding to the digital human generation requirements.

5. The method according to any one of claims 1 to 4, characterized in that: The method further comprises: Acquiring the target user's dialogue information, wherein the dialogue information is used to call the target digital human to output response information corresponding to the dialogue information; Inputting the dialogue information into the large language model of the target digital human to generate the response information; Feeding back the generated response information to the target user through the target digital human; The step of inputting the dialogue information into the large language model of the target digital human to generate the response information includes: Using named entity recognition and keyword extraction technology to perform semantic understanding on the conversation information, and obtain the intention characteristics, demand characteristics, and emotional characteristics of the target user; Processing the intention characteristics and demand characteristics of the target user through a cluster analysis algorithm, obtaining the knowledge base category corresponding to the intention characteristics and demand characteristics of the target user, and determining the target knowledge base corresponding to the intention characteristics and demand characteristics from the knowledge base set according to the knowledge base category; Performing knowledge retrieval in the target knowledge base to obtain candidate response content corresponding to the intention feature and the demand feature; Generate candidate response information by reasoning and integrating the candidate response contents; According to the candidate response information and the emotional feature, determine the language expression and tone corresponding to the emotional feature, and adjust the candidate response information by the language expression and tone to generate the response information; The step of performing knowledge retrieval in the target knowledge base to obtain candidate response content corresponding to the intention feature and the demand feature includes: Preliminarily constructing a target knowledge graph in the target knowledge base, wherein the nodes of the target knowledge graph are knowledge entities, the edges of the target knowledge graph are relationships between the knowledge entities, and the knowledge entities and the relationships are generated by preprocessing and annotation of a natural language processing model; Retrieving the knowledge entities corresponding to the intention feature and the demand feature in the target knowledge graph by a depth-first search algorithm, and obtaining candidate knowledge entities and associated knowledge entities associated with the candidate knowledge entities; According to the candidate knowledge entity and the relationship between the candidate knowledge entity and the associated knowledge entity, the candidate knowledge entity and the associated knowledge entity are integrated to generate the candidate response content.

6. The method according to claim 5, characterized in that The method further comprises: Processing the intention characteristics and demand characteristics of the target user by a cluster analysis algorithm, obtaining knowledge base categories corresponding to the intention characteristics and demand characteristics of the target user, and determining a first target knowledge base and a second target knowledge base corresponding to the intention characteristics and demand characteristics from a knowledge base set according to the knowledge base categories, wherein the first target knowledge base and the second target knowledge base have different knowledge fields; According to the first knowledge domain of the first target knowledge base and the second knowledge domain of the second target knowledge base, decomposing the intention characteristics and demand characteristics of the target user into first intention characteristics and demand characteristics and second intention characteristics and demand characteristics, wherein the first intention characteristics and demand characteristics are related to the first knowledge domain, and the second intention characteristics and demand characteristics are related to the second knowledge domain; Performing knowledge retrieval on the first intention feature and the demand feature in the first target knowledge base to obtain first candidate response content corresponding to the first intention feature and the demand feature; and performing knowledge retrieval on the second intention feature and the demand feature in the second target knowledge base to obtain second candidate response content corresponding to the second intention feature and the demand feature; Merging the first candidate response content and the second candidate response content to generate the candidate response content; The step of fusing the first candidate response content and the second candidate response content to generate the candidate response content includes: Acquire a first knowledge entity corresponding to the first candidate response content, and a second knowledge entity corresponding to the second candidate response content; determining an association type between the first knowledge entity and the second knowledge entity by crossing an edge between the first target knowledge base and the second target knowledge base; Determine, according to the association type and the mapping relationship between the association type and the content fusion strategy, a target content fusion strategy for fusing the first candidate response content and the second candidate response content; According to the target content fusion strategy, the first candidate response content and the second candidate response content are fused to generate the candidate response content.

7. The method according to claim 6, characterized in that The step of fusing the first candidate response content and the second candidate response content according to the target content fusion strategy to generate the candidate response content includes: According to the target content fusion strategy, the first candidate response content and the second candidate response content are fused to generate candidate response content to be evaluated for quality; Generating an evaluation score by performing consistency evaluation and accuracy evaluation on the candidate response content to be quality evaluated; If the evaluation score is less than a preset score threshold, then according to the target content fusion strategy, the first candidate response content and the second candidate response content are re-fused; If the evaluation score is greater than or equal to the preset score threshold, the candidate response content is generated.

8. An electronic device, characterized in that: include: A processor, a communication interface and a memory, wherein the processor is communicatively connected to the communication interface and the memory respectively; The memory stores computer-executable instructions; The communication interface communicates and interacts with external devices; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 7 when executed by a processor.

10. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 7 when being executed by a processor.

Citation Information

Patent Citations

  • Digital human generation method and device and digital human generation system

    CN116543082A

  • Digital human campus scene system based on sentiment analysis and large model driving

    CN117725180A

  • Digital human implementation method and system based on user portrait

    CN117743608A

  • Automatically generating a custom artificially intelligent (AI) character based on a user-provided description of the ai character

    US20240193839A1

Cited By

  • Method and system for generating digital human video

    CN119180895A

  • Training method and system for generating 5D digital human based on AIGC and medium

    CN119378647A

  • Training method, system and medium for generating 5D digital human based on AIGC

    CN119378647B

  • Construction method, device and equipment of intelligent public security digital person and storage medium

    CN119379875A

  • Interactive digital human generation method and system based on artificial intelligence

    CN119600159A