System and method for large language model second-order prompt for digital human intelligence

Abstract

The application relates to a system and method for a large language model second-order prompt word for digital human intelligence, and belongs to the technical field of artificial intelligence. The system comprises a user module, a digital human instance object module, an interface module, a second-order prompt word framework module and a first-order prompt word framework module; the user module is signal-connected with the digital human instance object module in a questioning, barrage or dialogue mode; the digital human instance object module calls the second-order prompt word framework module after receiving a signal; the second-order prompt word framework module generates a plurality of groups of first-order prompt words through a large language model; the first-order prompt word framework module calls the large language model to obtain feedback meeting a specific JSON format; the interface module is called to control the digital human to interact; and the digital human instance object module inputs and outputs information through the interface module. The application introduces a second-order prompt word to inspire the generation of a plurality of groups of first-order prompt words, so that the digital human can actively understand and meet the deep needs of the user, instead of passively responding to direct inquiries of the user. Through the enhancement of the initiative and thinking ability of the digital human, the application improves the interactive experience of the user and the digital human.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and relates to a system and method for second-order prompt words in a large language model for digital human intelligence. Background Technology

[0002] Currently, significant progress has been made in the basic technologies of digital humans (including speech understanding, speech output, character modeling, passive question answering, and text-to-speech), demonstrating commercial potential. However, these digital humans suffer from a significant deficiency in "passive intelligence." Existing digital humans primarily rely on user input for interaction, often failing to proactively provide in-depth information and services tailored to user needs. Furthermore, while digital humans can understand and generate language through large models, they still exhibit significant shortcomings in understanding users' deeper intentions and needs, and in handling complex interactive scenarios. This is mainly because existing large language models primarily rely on first-order cues—direct input from the user followed by analysis and feedback from the language model. In this approach, the model's understanding and responsiveness are limited, resulting in insufficient intelligence in digital humans during interactions. While they can provide routine responses, they lack basic human intelligence and the ability to take further action, failing to meet user expectations. Against this backdrop, controlling digital humans through second-order cues within a defined framework can inspire large language models and digital humans to generate a large number of first-order cue words, enabling logical thinking and "proactive intelligence."

[0003] The control and intelligentization technologies for digital humans mainly include speech technology (Text-to-Speech), digital human driving technology (DID API, Sadtalker, LSP), and digital human semantic understanding and communication technology (ChatGPT, Palm 2, Wenxin Yiyan, and other large language models). Among these, large language models are used as the "brain" of the digital human. They use an architecture called Transformer, which can learn how to generate natural, coherent language from large amounts of text data. In principle, this technology and existing technologies are mutually reinforcing, enhancing the intelligence of digital humans, and do not directly compete with each other. A brief introduction to existing technical solutions related to this invention is provided.

[0004] Currently, in a typical use case, users provide a digital human with one or more statements as prompts (first-order prompts), and the digital human then generates responses based on these prompts. The problem with this approach is that the digital human's responses are solely based on the user's prompts, often failing to consider the user's deeper needs and expectations, and lacking extended thinking—in other words, exhibiting clear characteristics of "passive intelligence." To address this issue, some existing technical solutions attempt to use additional labels or metadata during the training of large language models to provide more contextual information. For example, some models might be trained to understand and generate specific labels, such as... <weather>"or" <reminder>However, this approach requires pre-defining all possible labels and manually annotating these labels on the training data, which is very difficult in practice. Some existing technical solutions have also attempted to use reinforcement learning or other machine learning techniques to allow digital humans to learn from user feedback and optimize their responses. However, this approach requires a large amount of interaction data and complex algorithms and computing resources, which is impractical in many real-world applications. Overall, although existing technical solutions have made some progress in certain aspects, many challenges remain in providing a deep and user-expected interactive experience.

[0005] The existing technology has the following technical problems:

[0006] (1) Passive interaction problem: In the traditional digital human interaction mode, digital humans mainly rely on user input to respond, and their interaction behavior is too passive. This limits the initiative of digital humans in providing information and services, and often makes it impossible for them to effectively meet the deeper needs of users.

[0007] (2) Insufficient intelligence: Although existing large language models can understand and generate natural and coherent language, they still have significant shortcomings in understanding users' deeper needs, handling complex interaction scenarios, and providing personalized services. This is mainly because existing models rely primarily on first-order prompts, i.e., direct input provided by users, which limits the model's understanding and response capabilities.

[0008] (3) Insufficient interactive experience: Due to insufficient intelligence and initiative, current digital human interaction methods usually lack rich interactive forms, resulting in an insufficiently rich and vivid interactive experience. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide a system and method for using second-order prompt words in a large language model for digital human intelligence. By introducing second-order prompt words, and inspiring and optimizing the generation of first-order prompt words according to certain conditions and formats, the digital human can achieve a deep understanding and proactive response to user needs, and ensure that the first-order prompt words conform to the JSON data format for API calls. By generating interactive text using first-order prompt words that conform to a specific JSON format, the digital human can directly control various applications and services, achieving a more proactive, intelligent, and richer interactive experience.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A system for second-order prompt words in a large language model for digital human intelligence, the system includes a user module, a digital human instance object module, an interface module, a second-order prompt word framework module, and a first-order prompt word framework module;

[0012] The user module connects to the digital human instance object module via signals through questions, bullet comments, or dialogue. The digital human instance object module receives the signals and provides feedback. The digital human instance object module inputs and outputs information through the interface module, specifically:

[0013] The digital human instance object module utilizes a large language model, such as ChatGPT, Palm 2, ChatGLM2, or Wenxin Yiyan, to understand questions and generate answers. The user module interacts with the digital human to obtain text or language input. The large language model guides the first-order prompting framework module through a second-order prompting framework module, expanding the digital human's thinking logic in a tree structure to form several sets of first-order prompting words. Based on the user module's input and context information, the large language model generates second-order prompting words through the second-order prompting framework module, then converts them into first-order prompting words, analyzes and provides feedback on the user's questions. The generated first-order prompting words, analysis, and feedback results are input back into the large language model, and the responses from the obtained first-order prompting words are either called in blocks to the interface module or fed back to the user module.

[0014] Optionally, by introducing a certain nested dictionary JSON paradigm into the second-order prompt words to guide the first-order prompt words, the first-order prompt words can meet the JSON data format of API calls, enabling the digital human to call APIs or other I / O interface modules.

[0015] Optionally, the user module connects to the digital human instance object module via questioning, bullet comments, or dialogue. After receiving the signal, the digital human instance object module provides feedback. The digital human instance object module inputs and outputs information through the interface module. Specifically, based on JSON nested dictionary format text, second-order prompts guide thinking content, and first-order prompts guide direct dialogue. The large language model text is converted into speech through the Text-to-Speech module, and the speech is adjusted to a set timbre through the SOVITS module. The speech output is converted into digital human lip movements and body language through the SadTalker module. The user's questions can be received and answers can be output in a multimodal manner through API calls.

[0016] Optionally, the large language model guides the first-order prompting word framework module and the user module through the second-order prompting word framework module, specifically as follows:

[0017] The second-order prompt word framework module understands and executes the user module's settings for the digital human. In the given JSON format, the second-order prompt word framework module uses the user-specified role and input as the context and converts it into a token. The token is then fed into the generation model to obtain the basic attributes of the digital human that match the user's settings, including the digital human's personality, knowledge base, and habits.

[0018] The second-order cue word framework module generates corresponding second-order cue words based on the background and task set by the user module to guide the behavior of the digital human.

[0019] The first-order prompt word framework module generates steps to complete the task based on the output of the second-order prompt word framework module, guides the behavior of the digital human, and completes the output in the specified JSON format so that it can interact with the API interface.

[0020] The first-order cue word framework module understands the constraints of the task and generates first-order cue words that meet the constraints.

[0021] A method for second-order prompt words in a large language model for digital human intelligence, the method comprising the following steps:

[0022] S1: The digital human reads the configuration file from the configuration format analyzer and defines the JSON interface paradigm for API calls and other I / O.

[0023] S2: The configuration format analyzer returns configuration information, including nested dictionary JSON paradigms that control multiple interfaces simultaneously;

[0024] S3: The digital human generates second-order prompts based on user input and configuration format, and sends them to the second-order prompt module;

[0025] S4: The second-order prompt word module uses the preset nested dictionary paradigm definition and the user's input language and the packaged token, and inputs the big oracle model to inspire and guide the generation of multiple sets of first-order prompt words, and specifies the first-order prompt words to be inspired by the user's input.

[0026] S5: The large language model outputs several first-order prompt words that satisfy a certain nested dictionary format. Inspired by this, semantic analysis PARSE is performed to obtain first-order prompt words. These first-order prompt words are then input into the first-order prompt word module. The first-order prompt words generated from these second-order prompt words are then converted into interactive text that conforms to a specific JSON format.

[0027] S6: The first-order prompt word module uses the first-order prompt word as the context to convert the input into a token and feed it into the large language model to generate content, resulting in a nested dictionary JSON. By reading commands at the corresponding positions in the nested dictionary, the digital human interacts with applications and services through API calls.

[0028] S7: Based on the nested dictionary JSON input from the Big Prophecy model, the functions are integrated and divided into blocks. The communication text part is extracted by semantic analysis of the nested dictionary and the result is returned to the digital human to realize language communication.

[0029] S8: By performing semantic analysis and parsing on the results of the nested JSON dictionary, the commands required for interaction with each module are obtained, and the commands are substituted into the action module, understanding module, translation module and OFFICE interface module respectively.

[0030] Optionally, the inspiration and guidance of multiple sets of first-order prompts can be in a tree structure.

[0031] The beneficial effects of this invention are as follows:

[0032] Enhancing the initiative of digital humans: By introducing second-order cue words to inspire the generation of first-order cue words, digital humans can proactively understand and meet users' deeper needs, rather than simply passively responding to users' direct inquiries. This enhances the initiative and service capabilities of digital humans and improves the user's interactive experience.

[0033] Enhancing the intelligence of digital humans: By iterating and optimizing second-order prompts based on the output of first-order prompts, digital humans can handle more complex interaction scenarios, better understand user needs and intentions, and thus improve their intelligence.

[0034] Enriching Digital Human Interaction: By generating first-order cue text in a specific JSON format, digital humans can directly control various applications and services, providing a richer and deeper interactive experience. This not only allows users to directly operate various applications through digital humans, but also enables them to provide more personalized services that better meet user expectations based on their needs and context.

[0035] Easy to implement and apply: This invention does not require additional tags or metadata, nor does it require complex algorithms or large amounts of interaction data. Simply by designing and using appropriate second-order prompts, the interactive capabilities and user experience of digital humans can be significantly improved. This makes the invention easy to implement and apply, and it has great potential in practical applications.

[0036] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0037] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0038] Figure 1 This is a system diagram of the present invention;

[0039] Figure 2 A diagram illustrating the attributes of second-order and first-order prompt words;

[0040] Figure 3 Flowchart of objects guiding first-order prompts for second-order prompts;

[0041] Figure 4 This is a logical diagram illustrating how a tree structure inspires first-order prompts for second-order prompts. Detailed Implementation

[0042] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0043] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0044] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the accompanying drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0045] Figure 1 This demonstration showcases the interactive relationships between digital human instances, users, a large language model, a second-order prompting word framework, and a first-order prompting word framework within an intelligent digital human system. Users interact with digital human instances through questions, bullet comments, dialogues, or other means, and the instances receive this input and provide responses. Digital human instances input and output information via API interfaces or other I / O mechanisms. For example, they can process JSON-formatted text, converting text from the large language model to speech using the Text-to-Speech module, adjusting the speech to a specific timbre using the SOVITS module, and converting speech output into digital human lip movements and body language using the SadTalker module. Through API calls, they can receive user questions and output answers in a multimodal manner. Digital human instances are connected to the large language model, utilizing it to understand questions and generate answers. Users can not only ask the large language model specific questions, but the large language model can also guide first-order prompting words through second-order prompting words and actively lead discussions with the user. The large language model generates second-order prompts based on user input and contextual information using a second-order prompt framework. These prompts are then converted into first-order prompts in JSON format, while block-based logic analysis is performed to understand and respond to user questions. The generated first-order prompts and the results of the block-based logic analysis are then fed back into the large language model for further processing and optimization.

[0046] Figure 2 This framework demonstrates the hierarchical relationship between first-order prompts and second-order prompts, and how second-order prompts expand thinking and guide first-order prompts, influencing the behavior and task completion of the digital human. Specifically, second-order prompts control and guide the generation of first-order prompts, enabling them to better meet user needs and contextual requirements. The second-order prompt framework understands and executes user-defined settings for the digital human, including its personality, knowledge base, and habits. Based on the user's background and task settings, the framework generates corresponding second-order prompts to guide the digital human's behavior. It also understands the task's implications and generates second-order prompts that align with those implications. Based on the output of the second-order prompt framework, the first-order prompt framework generates steps to complete the task, guiding the digital human's behavior and producing output in a specified JSON format, enabling interaction with the API. Furthermore, the framework understands the task's constraints and generates first-order prompts that conform to those constraints. The first-order prompts generated by the framework must meet specific JSON format definitions to facilitate API calls and other operations.

[0047] Figure 3 The flowchart for guiding first-order prompts with second-order prompts includes the following steps:

[0048] S1: The digital human reads the configuration file from the configuration format analyzer and defines the JSON interface paradigm for API calls and other I / O.

[0049] S2: The configuration format analyzer returns configuration information, including nested dictionary JSON paradigms that control multiple interfaces simultaneously;

[0050] S3: The digital human generates second-order prompts based on user input and configuration format, and sends them to the second-order prompt module;

[0051] S4: The second-order prompt word module defines the preset nested dictionary paradigm and the user's input language and the packaged token, and inputs the big oracle model to inspire and guide the generation of multiple sets of first-order prompt words (tree structure). Unlike directly answering the user's input (linear structure), this invention inspires first-order prompt words under the specified user input (from which aspects can I solve this problem, how should I answer to make the user happy, etc.).

[0052] S5: The large language model outputs several first-order prompt words that satisfy a certain nested dictionary format. Inspired by this, semantic analysis (PARSE) is performed to obtain first-order prompt words. These first-order prompt words are then input into the first-order prompt word module. By converting the first-order prompt words generated from these second-order prompt words into interactive text that conforms to a specific JSON format, digital humans can achieve a proactive, more intelligent, and richer interactive experience.

[0053] S6: The first-order prompt word module uses the first-order prompt word as the context, converts the input into a token, and feeds it into the large language model to generate content, resulting in a nested dictionary JSON. By reading commands at the corresponding positions in the nested dictionary, the digital human can directly interact with various applications and services through API calls.

[0054] S7: Based on the nested dictionary JSON input from the Big Prophecy model, the functions are integrated and divided into blocks. The communication text part can be extracted through semantic analysis (parse) of the nested dictionary and the result is returned to the digital human to realize language communication.

[0055] S8: By performing semantic analysis (parse) on the results of the nested JSON dictionary, the commands required for interaction with various modules can be obtained. Substituting these commands into modules such as the action module, understanding module, translation module, and OFFICE interface enables functions such as speech synthesis, actions, interaction, and work assistance. This invention's second-order prompt words can generate a large number of sub-first-order prompt words suitable for various requirements. The 'prompt' word can be input again and returned to the large language model, guiding it to provide a response, simultaneously responding with the corresponding action, mood, and self-reflection.

[0056] Figure 4 This is a logical diagram illustrating how a tree structure inspires the logic of first-order prompts for second-order prompts. Unlike first-order prompts, which invoke large language models through linear logic, this invention's second-order prompt framework proposes a tree-like logical structure, enabling broader and more comprehensive proactive intelligence.

[0057] Example 1: Virtual Idols and Desktop Wizards

[0058] Imagine a busy programmer who uses this invention to create an intelligent digital human, a virtual idol who is "cunning, curious, yet loyal to the user" as a personal desktop assistant. This digital human is equipped with a second-order prompting framework, enabling it to proactively engage in deeper and more personalized interactions with the user.

[0059] On a typical weekday morning, when a user turns on their computer, the digital human will proactively greet them and briefly outline the day's work schedule. This process involves a second-order cue word inference module generating second-order cue words based on the user's calendar and historical habits, then using a first-order cue word framework to generate first-order cue words in JSON format. This, in turn, calls the text-to-speech API and uses voice to deliver a slightly witty and humorous greeting to the user.

[0060] During work, users may need to handle a large number of emails and meetings. The digital human will proactively read this information, understand the task requirements, and remind the user when necessary, such as reminding them to reply to an important email or prepare for an upcoming meeting. The second-order cue words involved in this process may include "check the tasks in the email" or "prepare meeting materials," and then generate specific operation commands through the first-order cue word framework.

[0061] When users communicate with clients or bosses, the digital human will generate second-order prompts based on the dialogue content to help users handle potential problems, and even make witty remarks at appropriate times, such as generating second-order prompts like "explain the project progress to the client" or "use a witty metaphor to criticize the boss's decision" to increase the fun of the communication.

[0062] For users' daily lives, the digital human will also generate relevant second-order cue words to show concern for users' physical health. For example, second-order cue words such as "remind users that they ate too little for lunch" or "remind users that they should go to the gym" will be used to generate specific reminder content through the first-order cue word framework to help users maintain a healthy lifestyle.

[0063] Overall, this virtual idol, characterized by being "cunning, curious, yet loyal to users," can better understand and predict user needs through a second-order cue word framework, providing a more proactive, intelligent, and richer interactive experience.

[0064] Example 2: General Personal Work Assistant

[0065] Imagine a user is a typical engineering manager. He / she can customize a digital avatar and voice to suit their preferences, acting as a digital secretary. They can also set up a database of professional knowledge. This digital secretary instance can interact with the user's computer, tablet, and mobile phone, providing insights and suggestions for their daily work and managing numerous tasks and meetings. Their personal digital secretary is integrated with a large language model that controls the digital secretary's second-order cue word framework.

[0066] One morning, when a user wakes up, the digital human proactively reminds him of his schedule for the day: two important meetings, a report to approve, and a business dinner in the evening. These reminders are generated by a second-order cue word inference module based on the user's calendar and historical habits, demonstrating both the digital human's initiative and intelligence.

[0067] When a user is processing emails, they receive one related to an afternoon meeting, which includes a task requiring preparation. The digital human, while reading the email, identifies this task and generates second-order cue words to prompt the user to complete it. Then, based on the second-order cue words, the digital human generates first-order cue words and sends them to the user, reminding them to prepare. In this scenario, the digital human is not merely a passive information provider, but proactively identifies and solves problems using the second-order cue word framework.

[0068] At a business dinner in the evening, the user is required to give a presentation. Before the dinner begins, the digital human proactively provides reference materials and suggestions based on the user's past presentation records and the dinner's theme. Throughout this process, the second-order cue word framework fully demonstrates its proactiveness and intelligence, helping the user better prepare for their presentation.

[0069] This implementation plan demonstrates that the second-order cue word framework for controlling digital humans using large language models is more proactive in understanding and predicting user needs compared to traditional passive intelligence, thereby providing more accurate and personalized services.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.< / reminder> < / weather>

Claims

1. A system for second-order prompt words in a large language model for digital human intelligence, characterized in that: The system includes a user module, a digital human instance object module, an interface module, a second-order prompt word framework module, and a first-order prompt word framework module; The user module connects to the digital human instance object module via signals through questions, bullet comments, or dialogue. The digital human instance object module receives the signals and provides feedback. The digital human instance object module inputs and outputs information through the interface module, specifically: The digital human instance object module uses a large language model to understand questions and generate answers; the user module communicates with the digital human to obtain text or language input; the large language model guides the first-order prompt word framework module through the second-order prompt word framework module, and expands the digital human's thinking logic in a tree structure to form several sets of first-order prompt words. The large language model generates second-order prompts based on the user module's input and context information through the second-order prompt framework module, then converts them into first-order prompts, analyzes and provides feedback on the user's questions; the generated first-order prompts, the analysis and feedback results are then input back into the large language model, and the responses from the obtained first-order prompts will be segmented to call the interface module to execute digital human behavior, or fed back to the user module; By introducing a certain nested dictionary JSON paradigm into the second-order prompt words to guide the first-order prompt words, the first-order prompt words are made to meet the JSON data format of API calls, enabling the digital human to call APIs or other I / O interface modules; The large language model guides the first-order prompting word framework module and the user module through the second-order prompting word framework module, specifically as follows: The second-order prompt word framework module understands and executes the user module's settings for the digital human. In the given JSON format, the second-order prompt word framework module uses the user-specified role and input as the context and converts it into a token. The token is then fed into the generation model to obtain the basic attributes of the digital human that match the user settings, including the digital human knowledge base. The second-order prompt word framework module generates corresponding second-order prompt words based on the background and task set by the user module. It expands the thinking logic of the digital human in a tree structure and inspires and guides the generation of multiple sets of first-order prompt words. The first-order prompt word framework module generates the steps to complete the task based on the output of the second-order prompt word framework module, and expresses the steps as first-order prompt words in a specified JSON format for interaction with the API interface; The first-order prompt word framework module understands the task constraints under the nested dictionary JSON paradigm, generates first-order prompt words that meet the constraints, and converts the first-order prompt words into interactive text that conforms to the JSON format for use by the interface module.

2. The system for second-order prompt words in a large language model for digital human intelligence according to claim 1, characterized in that: The user module connects to the digital human instance object module via signals through questions, bullet comments, or dialogue. The digital human instance object module receives signals and provides feedback. It inputs and outputs information through an interface module. Specifically, it uses JSON-based nested dictionary text, with second-order prompts guiding thought processes and first-order prompts guiding direct dialogue. The Text-to-Speech module converts the large language model text into speech, and the SOVITS module adjusts the speech to a set timbre. The SadTalker module converts the speech output into digital human lip movements and body language. Through API calls, it can receive user questions and output answers in a multimodal manner.

3. A method for using second-order prompt words in a large language model for digital human intelligence based on the system described in claim 1 or 2, characterized in that: The method includes the following steps: S1: The digital human reads the configuration file from the configuration format analyzer and defines the JSON interface paradigm for API calls and other I / O. S2: The configuration format analyzer returns configuration information, including nested dictionary JSON paradigms that control multiple interfaces simultaneously; S3: The digital human generates second-order prompts based on user input and configuration format through the second-order prompt framework module, and sends them to the second-order prompt module; S4: The second-order prompt word module uses the preset nested dictionary paradigm definition and the user's input language and the packaged token, and inputs the large language model to inspire and guide the generation of multiple sets of first-order prompt words, and inspires first-order prompt words under the specified user input; S5: The large language model outputs several first-order prompt words that satisfy a certain nested dictionary format. Inspired by this, semantic analysis PARSE is performed to obtain first-order prompt words under PARSE constraints. These first-order prompt words are then input into the first-order prompt word module. The first-order prompt words generated from these second-order prompt words are then converted into interactive text that conforms to a specific JSON format. S6: The first-order prompt word module uses the first-order prompt word as the context to convert the input into a token and feed it into the large language model to generate content, resulting in a nested dictionary JSON. By reading commands at the corresponding positions in the nested dictionary, the digital human interacts with applications and services through API calls. S7: Based on the nested dictionary JSON input from the large language model, the functions are integrated and divided into blocks. The communication text part is extracted by semantic analysis of the nested dictionary and the result is returned to the digital human to realize language communication. S8: By performing semantic analysis and parsing on the results of the nested JSON dictionary, the commands required for interaction with each module are obtained, and the commands are substituted into the action module, understanding module, translation module and OFFICE interface module respectively. The multiple sets of first-order prompts used for inspiration and guidance form a tree structure.

Images