A dynamic interaction management method for multi-role AI podcast conversations

Abstract

The application relates to the technical field of AI-generated content management, in particular to a dynamic interaction management method for multi-role AI podcast conversation, which comprises the following steps: obtaining a guest agent of a current round according to a speaking round, inputting the current guest agent with a historical conversation prompt word generated according to a last round of conversation messages; obtaining and playing output content of the guest agent, and inputting a host agent with a host prompt word template; and obtaining and playing the output content. In view of the problems of poor context continuity and weak interaction in the multi-agent turn-by-turn conversation scene in the prior art, the speaking content topic transmission and the viewpoint offset variable injection of the guest agent of the current round are realized by combining the host agent with a specific prompt word template, so that the topic content transmission of the generated content of the guest agent of the next round is realized, and the guest agent of the next round is controlled to provide different viewpoints, thereby making the conversation process more coherent and avoiding the problem of repeatedly outputting the same viewpoint.

Description

Technical Field

[0001] This invention relates to the field of AI-generated content management technology, specifically to a dynamic interactive management method for multi-role AI podcast dialogues. Background Technology

[0002] A podcast is a format that broadcasts media content via audio, featuring discussions among multiple people on a particular topic or interviews with specific individuals, and is available to a general audience. AI podcasts, on the other hand, are audio broadcast programs created in a format similar to existing podcasts, combining large-scale language model technology with text-to-speech (TTS) technology. Their production costs are significantly lower than those of live interviews, and their production cycles are shorter, making them more effective at capturing trending topics.

[0003] In existing technologies, there are technical solutions that use multiple AI agents to play specific roles and take turns speaking.

[0004] For example, patent application CN202411612589.0 provides a method and related apparatus for interview-style conversational information push based on multi-agent collaboration, involving the fields of generative large language models, intelligent agents, and other artificial intelligence technologies. The method includes: determining a target topic of interest to the user; identifying multiple target virtual human intelligent agents within the domain of the target topic and distributing the target topic to each target virtual human intelligent agent, with different virtual human intelligent agents pre-built for different domains; controlling different target virtual human intelligent agents to publish different viewpoints on the target topic, each viewpoint obtained from information acquisition channels associated with the target topic; and summarizing the different viewpoints published by each target virtual human intelligent agent to obtain the information push result centered around the target topic. This solution, through collaboration between the main intelligent agent and the various virtual human intelligent agents, achieves a way to quickly and effectively convey relevant public viewpoints on target topics of interest to users through interviews and conversations. Furthermore, the virtual human intelligent agents are anthropomorphic, making the viewpoints conveyed more user-friendly and easily accepted and understood by users, achieving excellent results in information push scenarios.

[0005] For example, application number CN202410804803.6 provides a dialogue processing method, dialogue method, device, and electronic device based on a large model, relating to the field of artificial intelligence technology, particularly natural language processing, large models, and deep learning, and applicable to scenarios such as intelligent customer service, in-vehicle interaction, and intelligent agents. The specific implementation scheme is as follows: Obtain at least one round of dialogue records in the target dialogue scenario; based on the at least one round of dialogue records, determine the memory content corresponding to any dialogue subject and the content category of the memory content; associate and store the target dialogue scenario, the memory content corresponding to each dialogue subject, and the content category of the memory content to obtain memory data, wherein the memory data is used for dialogue generation in the target dialogue scenario. Determining the memory content and its content category based on the dialogue records extracts effective information from the dialogue records, avoids introducing too much context during dialogue generation, and reduces computational load.

[0006] However, in actual implementation, the inventors found that this type of technical solution is prone to losing some content during long dialogues due to context window issues. For example, identity prompts used to define roles and viewpoints may be lost, and previously discussed dialogues may be repeated, resulting in a lack of coherence in the listening experience. If strong prompts are used to anchor the output content of the agents, the output content of each agent may become independent and spliced ​​together. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, a dynamic interaction management method for multi-role AI podcast dialogues is provided.

[0008] The specific technical solution is as follows: A dynamic interaction management method for multi-role AI podcast dialogues, applicable to multi-agent interaction scenarios, includes: Before the dialogue begins, identity prompts are established and injected into each agent, the speaking turns of each agent are maintained, and a topic overview is obtained and established. The intelligent agents are divided into host intelligent agents and guest intelligent agents; In actual dialogue scenarios, this includes: Step S1: Obtain the guest agent for the current round according to the speaking round, and generate historical dialogue prompts based on the previous round's dialogue messages to input into the current guest agent; Step S2: Obtain and play the output content of the guest agent, and input it into the host agent in combination with the host prompt template; The host prompt template is used for topic transmission and the injection of viewpoint deviations; Step S3: Obtain the output content of the host agent and play it, then return to step S1.

[0009] On the other hand, the identity prompts include personality definition information and identity isolation instructions; The personality definition information includes the character name, voice name, personality prompts, and voice style; The identity isolation instruction includes a positive identity declaration, a reverse identity exclusion information, and output constraints. On the other hand, step S1 includes: Step S11: Obtain the next guest agent according to the speaking round, and mark irrelevant agents according to the obtained guest agent; Step S12: Obtain the dialogue message, and assign a role to each message in the dialogue message according to the irrelevant intelligent agent to form marked dialogue content; Step S13: Process the marked dialogue content using the dialogue prompt template and generate the historical dialogue prompts, which are then input into the guest agent.

[0010] On the other hand, step S2 includes: Step S21: Obtain the guest agent for the next round according to the speaking round, and generate a random variable for opinion differentiation; Step S22: Generate host prompts based on the opinion-differentiated random variable, the output content of the guest agent, and the host prompt template, and add the opinion-differentiated random variable to the host prompts; Step S23: Input the host prompt into the host AI agent.

[0011] On the other hand, it also includes a user command input process, which includes: Step A1: When a user input message is received, update the message identifier for the channel where the dialogue message is located; During the execution of the actual dialogue scenario, the message identifier is verified, and when the message identifier changes, the process switches to processing a new dialogue, while the original dialogue is marked as dialogue history, and the process proceeds to step A1. Step A2: Match the user input message with the name of the intelligent agent to determine whether to specify the intelligent agent to provide feedback; If so, generate corresponding prompt words based on the user input message and input them into the agent to obtain the output content; If not, proceed to step A3: Step A3: Perform summary processing on the dialogue history, and input it together with the user input message into an external large-scale language model for identification and determination of whether it is a related topic; If so, obtain the guest agent for the next round based on the dialogue history, add the user input message to the dialogue message, and then return to step S1; If not, generate a new topic overview based on the user input message and assign a new speaking turn, then return to step S1.

[0012] On the other hand, during the execution of the actual dialogue scenario, the output content of the guest agent and the host agent is written into the pending confirmation history message, and moved to the formal history message after playback is completed. When proceeding to step A1, only the content in the formal historical messages is used as the original dialogue.

[0013] On the other hand, it also includes a guest addition process, which includes: Step B1: When a user's instruction to add a guest is received, the corresponding guest agent is located according to the instruction. Step B2: Detect the end point of the playback of the output content. When the end point of playback is reached, proceed to step B3. Step B3: Add the newly added guest agent to the speaking round, and input the introduction content into the host agent based on the description information of the newly added guest agent and the introduction prompt template to obtain the introduction content and play it, and then return to the original process.

[0014] On the other hand, it also includes a guest removal process, which includes: Step C1: When a guest removal instruction is received, the playback end node of the output content is detected. When the playback end node is reached, proceed to step C2. Step C2: Pause the transmission to the next agent and remove the target agent from the guest removal instruction in the speaking round; Step C3: Determine the next agent using the updated speaking round.

[0015] The above technical solution has the following advantages or beneficial effects: To address the issues of poor contextual coherence and weak interactivity in existing multi-agent turn-based dialogue scenarios, this paper proposes a method that combines the host agent with specific prompt word templates to transmit topics and inject viewpoint shift variables into the speeches of the guest agents in the current round. This enables the transmission of topic content to the content generated by the guest agents in the next round and controls them to provide different viewpoints, thereby making the conversation process more coherent and avoiding the problem of repeatedly outputting the same viewpoints. Attached Figure Description

[0016] Embodiments of the invention will be described more fully with reference to the accompanying drawings. However, the drawings are for illustration and explanation only and do not constitute a limitation on the scope of the invention.

[0017] Figure 1 This is an overall schematic diagram of an embodiment of the present invention; Figure 2 This is a schematic diagram of step S1 in an embodiment of the present invention; Figure 3 This is a schematic diagram of step S2 in an embodiment of the present invention; Figure 4 This is a schematic diagram of the user command input process in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the guest addition process in an embodiment of the present invention; Figure 6 This is a schematic diagram of the guest removal process in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0021] This invention includes: A dynamic interaction management method for multi-role AI podcast dialogues, applicable to multi-agent interaction scenarios, includes: Before the dialogue begins, identity prompts are established and injected into each agent, the speaking turns of each agent are maintained, and a topic overview is obtained and established. The intelligent agents are divided into host intelligent agents and guest intelligent agents; like Figure 1 As shown, in actual dialogue scenarios, the following are included: Step S1: Obtain the guest agent for the current round according to the speaking round, and generate historical dialogue prompts based on the previous round's dialogue messages to input into the current guest agent; Step S2: Obtain and play the output content of the guest agent, and input it into the host agent in combination with the host prompt template; The host prompt template is used for topic transmission and the injection of alternative viewpoints; Step S3: Obtain the output content of the host agent and play it, then return to step S1.

[0022] Specifically, addressing the issues of poor contextual coherence and weak interactivity in existing multi-agent turn-based dialogue scenarios, this paper proposes a method that combines the host agent with specific prompt word templates to transmit topics and inject viewpoint shift variables into the speeches of the guest agents in the current round. This enables the transmission of topic content to the content generated by the guest agents in the next round and controls them to provide different viewpoints. As a result, the conversation becomes more coherent, subsequent guest agents can effectively respond to the previous guest agent, and the problem of repeatedly outputting the same viewpoints is avoided.

[0023] In practical implementation, the aforementioned dynamic interaction management method is primarily configured as a software example within a podcast building system. This system constructs specific agents based on user-inputted instructions and controls these agents to generate corresponding dialogue information. The text-to-speech system then generates the corresponding dialogue content for playback. A common system architecture involves setting up an agent management module, a large language model interface module, a dialogue generation and caching module, and a text-to-speech module on the server side. A specific client is connected to this server to broadcast the final generated audio content, supplemented by text display. This client is installed on the user's computer device, such as a mobile phone or in-vehicle infotainment system. It typically uses a WebSocket long-lived connection as the communication basis, establishing a persistent connection between the client and the server.

[0024] An intelligent agent is a virtual AI character primarily built upon large-scale language model technology, combined with specific prompts, constraints, or RAG fine-tuning techniques. It is used to understand input text information and generate corresponding responses based on prompts. In this scenario, its role is to input the given dialogue message and prompts into a large-scale language model at the backend to obtain the expected text output, which is then returned to the dialogue message channel. Guest agents and host agents are descriptions used to distinguish actual roles in a dialogue scenario. Technically, they are the same or different agents, primarily differentiated by different prompts.

[0025] Identity prompts are used to define the identities of each virtual AI character, and to control the large language model connected to the backend to understand the prompts and output them in a specific way.

[0026] In one embodiment, the identity prompt includes personality definition information and identity isolation instructions; Personality definition information includes character name, voice name, personality cue words, and voice style; Identity segregation instructions include positive identity declaration, reverse identity exclusion information, and output constraints.

[0027] Among them, the role name is used to distinguish different agents in subsequent dialogues, the voice name is used to indicate the voice name to be called by the subsequent text-to-speech module, the personality prompt word is used to provide constraints on output content and speaking style, and the voice style is used to provide the text-to-speech module with fine-tuning of tone.

[0028] Identity isolation instructions are used to inject into agents to forcibly isolate the speech content of each agent, preventing model attention drift when the context window is full. These instructions include: Positive identity declaration: "You are {current character name}"; Reverse identity exclusion: "You are not {other character name 1}, {other character name 2}..."; Output constraint: "Speak your statement directly and do not identify yourself as another character."

[0029] During each round of dialogue, the system maintains the continuity of each agent's perspective by repeatedly injecting identity isolation instructions.

[0030] In real-world dialogue scenarios, the system assigns a channel number to each newly generated dialogue channel and adds corresponding agents to that channel, including a host agent and at least one guest agent. The dialogue messages between the agents are stored within the channel through sequence concatenation. A speaking round is maintained within the dialogue channel to determine the speaking order of each agent. Typically, a host agent is inserted to connect each round of dialogue, and then the guest agent, whose turn is selected from the speaking rounds, speaks.

[0031] At the start of the conversation, a topic overview is injected externally, such as a discussion about a trending product or a phenomenon. This topic overview can be assigned by the system or injected by the user. Then, according to a preset structure, the host agent generates corresponding opening content and an introduction to the agent based on the topic overview and a pre-designed opening prompt template. Subsequently, the conversation enters a cyclical speaking state based on the speaking rounds.

[0032] Furthermore, during the first round of speeches, the opening content generated by the host and the introduction content to the agent are marked, thereby enabling this content in the dialogue messages, so that subsequent guest agents only respond to the topic overview content.

[0033] During the cyclical speaking process, the guest agent for the current round is usually obtained according to the speaking round, historical dialogue prompts are generated based on the previous round's dialogue messages and input into the current guest agent, and then the output content of the guest agent is obtained and played.

[0034] The previous round of dialogue messages is retrieved and determined based on historical messages in the current dialogue channel. The output content is text-based, generated by inputting it into a text-to-speech system and combining it with tags indicating the voice name and voice style from the identity prompts.

[0035] Subsequently, the output of the guest agent is combined with the host prompt template and input into the host agent. This process is used to summarize the output of the guest agent and pass it to the next guest agent, and to randomly inject viewpoint offsets. This step involves generating a random variable to represent the support or extension of the viewpoint, and then the large language model connected to the backend of the host agent predicts the viewpoint content that will be offset under this random variable.

[0036] This viewpoint will be spliced ​​into the output of the host agent, and in the next round of speeches, it will be captured as historical information and injected into the next guest agent as prompt words to achieve continuous content output.

[0037] Similarly, the host agent's output is text-based, which is generated by inputting into a text-to-speech system and combining the tags about voice name and voice style in the identity prompts.

[0038] In actual control, the host agent also sets up a phased behavior state machine to precisely control the differentiated behavior of the host agent in different scenarios (first speech, new topic, follow-up question, designated answer, solo chat mode, discussion end summary) through the combination of multiple Boolean states.

[0039] In one embodiment, such as Figure 2 As shown, step S1 includes: Step S11: Obtain the next guest agent according to the speaking round, and mark irrelevant agents based on the obtained guest agents; Step S12: Obtain dialogue messages, and assign role labels to each message in the dialogue messages according to the irrelevant agents to form labeled dialogue content; Step S13: Process the marked dialogue content using the dialogue prompt template and generate a history dialogue prompt input guest agent.

[0040] Specifically, to achieve better isolation between the viewpoints of different roles, this embodiment, while maintaining the speaking rounds, marks guest agents not selected in the current round as irrelevant agents. Then, it retrieves dialogue messages from the dialogue channel, traces each message back according to the role name marked before it, and re-marks them. The current guest agent's historical speech is marked as the "assistant" role, while the historical speech of other AI guests is uniformly marked as the "user" role, thus obtaining the marked dialogue content. This prevents the large language model from confusing its own identity with the identities of other roles in multi-round dialogues.

[0041] Subsequently, the marked dialogue content is processed in conjunction with the dialogue prompt template, and historical dialogue prompts are generated and input into the guest agent. These dialogue prompts include a portion of the identity prompts, including role name, personality prompts, and identity isolation instructions. The guest agent is then required to generate a response based on the input marked dialogue content, thereby obtaining the output content.

[0042] In one embodiment, such as Figure 3 As shown, step S2 includes: Step S21: Obtain the guest agent for the next round according to the speaking round, and generate random variables for opinion differentiation; Step S22: Generate host prompts based on the opinion differentiation random variable, the output of the guest agent, and the host prompt template, and add the opinion differentiation random variable to the host prompts; Step S23: Input the host prompts into the host AI agent.

[0043] Specifically, after obtaining the guest AI agent, the system also obtains the next guest AI agent according to the speaking round. A random number generator is then used to generate opinion-differentiation random variables, such as agreement, disagreement, supplementation, extension, and expansion, along with corresponding descriptive statements, based on a pre-built opinion offset table. Subsequently, the opinion-differentiation random variables and the output of the guest AI agents are input into the host prompt template to form the host prompts.

[0044] The host prompt template is used to instruct the large language model on the backend of the host agent to summarize the output of the guest agent, and to predict further viewpoints based on the output of the guest agent and the viewpoint differentiation random variable.

[0045] The host prompts are then input into the host agent to obtain the host agent's output.

[0046] In one embodiment, a user instruction input process is also included, such as... Figure 4 As shown, the user command input process includes: Step A1: When a user input message is received, update the message identifier for the dialogue channel where the dialogue message is located; During the execution of the actual dialogue scenario, the message identifier is verified, and when the message identifier changes, the process switches to processing a new dialogue, while the original dialogue is marked as dialogue history, and the process proceeds to step A2. Step A2: Match the user input message with the name of the agent to determine whether to specify an agent to provide feedback; If so, generate corresponding prompt words for the AI ​​agent based on the user's input message to obtain the output content; If not, proceed to step A3: Step A3: Summarize the dialogue history and input it along with the user input message into an external large-scale language model for identification and determination of whether it is a related topic; If so, obtain the guest agent for the next round based on the dialogue history, add the user input message to the dialogue message, and then return to step S1; If not, generate a new topic overview based on the user input message and assign a new speaking turn, then return to step S1.

[0047] Specifically, in order to achieve a better user experience, this embodiment also adds a process of intervention based on user behavior. Specifically, a message identifier is maintained for each dialogue channel, and the message identifier is initialized and assigned a value when a new dialogue is started.

[0048] During the execution of actual dialogue scenarios, the message identifier is validated. This message identifier should remain unchanged. If it changes, it indicates that a user input message has appeared, and further execution rounds need to be paused to proceed to step A2. In practice, the pause point is the point where a single step has been fully executed and the process moves to the next step. If the current step is not completed, execution needs to continue.

[0049] For user input messages, the system matches the input message with the agent's role name to determine if an agent should be specified for feedback. If so, the user input message is concatenated with the user's question prompt template, and the specified agent is then input to obtain and play the feedback output. The system then reverts to the original dialogue scenario to continue retrieving and playing the corresponding content. During this process, a skip flag is set for the prompts input by the specified agent, allowing for direct responses to user input messages without addressing messages from multi-person discussions or previous conversations.

[0050] If no matching agent role name is found, the system further determines whether the user has started a new topic. This includes summarizing the conversation history and inputting it along with the user's input message into an external large language model for identification and determination of whether it is a related topic. The system then returns the content from the large language model.

[0051] When the returned content indicates that no new topic has been started, the user's input message is submitted to the conversation message, and then the response is generated based on the complete context to maintain topic coherence.

[0052] When the returned content indicates that the user has started a new topic, a new topic overview is generated based on the user's input, a new speaking round is assigned, the historical messages in the original chat channel are cleared, and then the process of the first round of conversation is restarted.

[0053] The speaking turn for this new topic can follow the original dialogue message turn, or it can be randomly reassigned.

[0054] In one embodiment, during the execution of an actual dialogue scenario, the output content of the guest agent and the host agent is written into the pending confirmation history message, and then moved to the formal history message after playback is completed. When proceeding to step A2, only the content in the official history messages will be used as the original dialogue.

[0055] Specifically, during actual playback, since the dialogue messages are generated by the backend server in the cloud, while the playback is performed on the client side, there is a certain time delay between the two.

[0056] During the execution of the actual dialogue scenario, the output content of the guest agent and the host agent is first written into the pending confirmation history message as a cache, and after the client has fully played the corresponding output content, it is then transferred to the queue corresponding to the formal history message for storage.

[0057] When transitioning to step A2, only the content in the official history messages will be used as the original dialogue to avoid inserting viewpoints that the user has not heard before into the dialogue history, thereby improving the continuity of the topic.

[0058] In addition, in cases where the user input message does not start a new topic, this embodiment submits the user input message to the official history messages.

[0059] As an optional implementation, when acquiring dialogue messages, the agent can read the pending confirmation history messages and the formal history messages to obtain the complete conversation context.

[0060] In one embodiment, a guest addition process is also included, such as... Figure 5 As shown, the guest addition process includes: Step B1: When a user's instruction to add a guest is received, the corresponding guest agent is located according to the instruction. Step B2: Detect the end point of playback of the output content. When the end point of playback is reached, proceed to step B3. Step B3: Add the newly added guest agent in the speaking round, and input the introduction content into the host agent based on the description information of the newly added guest agent and the introduction prompt template to obtain the introduction content and play it, and then return to the original process.

[0061] Specifically, to enable the addition of guest agents during dialogue, this embodiment pre-maintains several selectable guest agents in the system. Each guest agent is assigned a set of threads connected to a large language model on the backend and a corresponding identity prompt word. When a user's instruction to add a guest is received, the corresponding guest agent is located according to the instruction and added to the waiting queue.

[0062] To maintain the continuity of the dialogue, this embodiment selects two playback end nodes: the nodes where the output content of the host agent and the guest agent has finished playing. The playback end node of the output content is detected. When a playback end node is reached, a new guest agent is added to the speaking round according to the order of the waiting queue. The host agent is then input with the description information of the new guest agent and the introduction prompt template to obtain the introduction content, which is then played. The process then returns to the original flow.

[0063] Similarly, for newly added guest agents, during the first round of speeches, the opening content generated by the host and the introduction content of the agent are marked, thereby enabling this content in the dialogue messages, so that the newly added guest agent only responds to the topic overview content.

[0064] In addition, in cases where a user adds a guest agent while switching topics, the system restores the original guest agent for the reset dialogue channel, moves the newly added guest agent to the end of the speaking round, and clears the waiting queue.

[0065] In one embodiment, a guest removal process is also included, such as... Figure 6 As shown, the guest removal process includes: Step C1: When a guest removal instruction is received, the playback end node of the output content is detected. When the playback end node is reached, proceed to step C2. Step C2: Pause the transmission to the next agent and remove the target agent from the guest removal instruction during the speaking round; Step C3: Determine the next agent using the updated speaking turn.

[0066] Specifically, in order to remove the guest agent, in this embodiment, when a guest removal instruction is received, the playback end node of the output content is detected, including the node where the playback of the output content of the host agent and the guest agent ends.

[0067] When the playback ends, the transmission is paused and passed to the next agent. The target agent in the guest removal command is removed from the speaking round, and then the corresponding content is played again to maintain the continuity of the dialogue.

[0068] Those skilled in the art will understand that various aspects, or possible implementations of various aspects, of the present invention can be embodied as systems, methods, or computer program products. Therefore, various aspects, or possible implementations of various aspects, of the present invention can take the form of entirely hardware embodiments, entirely software embodiments (including firmware, resident software, etc.), or embodiments combining software and hardware aspects, all collectively referred to herein as "circuit," "module," or "system." Furthermore, various aspects, or possible implementations of various aspects, of the present invention can take the form of computer program products, which are computer instructions stored in memory.

[0069] The memory can be a computer-readable signal medium or a computer-readable storage medium. Computer-readable storage media include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or apparatuses, or any suitable combination thereof, such as random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, and portable read-only memory (CD-ROM).

[0070] A processor in a computer reads computer instructions stored in memory, enabling the processor to execute the functional actions specified in each step or combination of steps in a flowchart; and to generate means for implementing the functional actions specified in each block or combination of blocks in a flowchart.

[0071] It should be understood that a processor in a computer can be understood as one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components used to execute the aforementioned computer instructions.

[0072] Computer instructions may be executed entirely on the user's local computer, partially on the user's local computer, as a separate software package, partially on the user's local computer and partially on a remote computer, or entirely on a remote computer or server. It should also be noted that in some alternative implementations, the functions indicated by the steps in the flowchart or the blocks in the block diagram may not occur in the order shown in the diagram. For example, depending on the functions involved, two consecutive steps or blocks may actually be executed approximately simultaneously, or these blocks may sometimes be executed in reverse order.

[0073] Of course, in practical applications, the various components of a computer system are coupled together through a bus system. The bus system is used to enable communication and connection between these components. In addition to the data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0074] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic interaction management method for multi-role AI podcast dialogue, characterized in that, Suitable for multi-agent interaction scenarios, including: Before the dialogue begins, identity prompts are established and injected into each agent, the speaking turns of each agent are maintained, and a topic overview is obtained and established. The intelligent agents are divided into host intelligent agents and guest intelligent agents; In actual dialogue scenarios, this includes: Step S1: Obtain the guest agent for the current round according to the speaking round, and generate historical dialogue prompts based on the previous round's dialogue messages to input into the current guest agent; Step S2: Obtain and play the output content of the guest agent, and input it into the host agent in combination with the host prompt template; The host prompt template is used for topic transmission and the injection of viewpoint deviations; Step S3: Obtain the output content of the host agent and play it, then return to step S1.

2. The dynamic interactive management method according to claim 1, characterized in that, The identity prompts include personality definition information and identity isolation instructions; The personality definition information includes the character name, voice name, personality prompts, and voice style; The identity isolation instructions include positive identity declaration, reverse identity exclusion information, and output constraints.

3. The dynamic interactive management method according to claim 1, characterized in that, Step S1 includes: Step S11: Obtain the next guest agent according to the speaking round, and mark irrelevant agents according to the obtained guest agent; Step S12: Obtain the dialogue message, and assign a role to each message in the dialogue message according to the irrelevant intelligent agent to form marked dialogue content; Step S13: Process the marked dialogue content using the dialogue prompt template and generate the historical dialogue prompts, which are then input into the guest agent.

4. The dynamic interactive management method according to claim 1, characterized in that, Step S2 includes: Step S21: Obtain the guest agent for the next round according to the speaking round, and generate a random variable for opinion differentiation; Step S22: Generate host prompts based on the opinion-differentiated random variable, the output content of the guest agent, and the host prompt template, and add the opinion-differentiated random variable to the host prompts; Step S23: Input the host prompt into the host AI agent.

5. The dynamic interactive management method according to claim 1, characterized in that, It also includes a user command input process, which includes: Step A1: When a user input message is received, update the message identifier of the dialogue channel where the dialogue message is located; During the execution of the actual dialogue scenario, the message identifier is verified, and when the message identifier changes, the process switches to processing a new dialogue, while the original dialogue is marked as dialogue history, and the process proceeds to step A2. Step A2: Match the user input message with the name of the intelligent agent to determine whether to specify the intelligent agent to provide feedback; If so, generate corresponding prompt words based on the user input message and input them into the agent to obtain the output content; If not, proceed to step A3: Step A3: Perform summary processing on the dialogue history, and input it together with the user input message into an external large-scale language model for identification and determination of whether it is a related topic; If so, obtain the guest agent for the next round based on the dialogue history, add the user input message to the dialogue message, and then return to step S1; If not, generate a new topic overview based on the user input message and assign a new speaking turn, then return to step S1.

6. The dynamic interactive management method according to claim 5, characterized in that, During the execution of the actual dialogue scenario, the output content of the guest agent and the host agent is written into the pending confirmation history message, and moved to the formal history message after playback is completed; When proceeding to step A2, only the content in the formal historical messages is used as the original dialogue.

7. The dynamic interactive management method according to claim 1, characterized in that, It also includes a guest addition process, which includes: Step B1: When a user's instruction to add a guest is received, the corresponding guest agent is located according to the instruction. Step B2: Detect the end point of the playback of the output content. When the end point of playback is reached, proceed to step B3. Step B3: Add the newly added guest agent to the speaking round, and input the introduction content into the host agent based on the description information of the newly added guest agent and the introduction prompt template to obtain the introduction content and play it, and then return to the original process.

8. The dynamic interactive management method according to claim 1, characterized in that, It also includes a guest removal process, which includes: Step C1: When a guest removal instruction is received, the playback end node of the output content is detected. When the playback end node is reached, proceed to step C2. Step C2: Pause the transmission to the next agent and remove the target agent from the guest removal instruction in the speaking round; Step C3: Determine the next agent using the updated speaking round.

9. A storage medium comprising computer instructions, characterized in that, When the computer device executes the computer instructions, it performs the dynamic interactive management method as described in any one of claims 1-8.

Images