Multi-agent-based multi-round interactive content wind direction guiding method

Abstract

The invention relates to the technical field of natural language processing, in particular to a multi-round interactive content wind direction guiding method based on multiple agents. Through cooperative work of the inquiry behavior agent, the chat behavior agent and the guide behavior agent, the intention of the user is more accurately understood, and proper feedback is quickly generated, so that the communication efficiency is improved. The planning agent selects the optimal response according to the user preference and the interaction history, the user is effectively guided to carry out the next round of interaction, the interaction process is promoted to be smoothly carried out, it is ensured that each time of interaction can meet the user requirement to the maximum degree, and the guiding accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method for guiding the direction of multi-round interactive content based on multiple agents. Background Art

[0002] The purpose of the interactive content direction guidance method is to enhance the user experience and promote effective communication between users and products to ensure that users can achieve their expected goals efficiently and smoothly.

[0003] Traditional interactive content direction guidance methods include attribute-based guidance methods and generation-based guidance methods. Attribute-based methods: This type of method clarifies user preferences by asking attribute-related questions, and the user answers with "yes / no". This type of method focuses on when to ask and what questions to ask before deciding the direction of guidance. For example, some studies model the entire multi-round guidance process as a path reasoning problem on a user-item-attribute graph. However, due to the limited communication between the system and the user, the attribute-based method lacks flexibility. The user cannot actively ask questions, and the system can only reply through manually annotated templates, resulting in reduced accuracy of content guidance. Generation-based methods: This type of method allows users and systems to communicate using free-form natural language. Most existing generative dialogues or multi-round comment replies rely on knowledge graphs for reasoning, resulting in poor generalization of the model when dealing with actual scenarios, resulting in low accuracy of content guidance and low user satisfaction.

[0004] It can be seen that the guidance method in the above-mentioned prior art is too mechanical and rigid, which leads to reduced guidance accuracy, thereby causing poor customer satisfaction, because it is not suitable for the current interactive method with higher requirements. Summary of the invention

[0005] In order to avoid and overcome the technical problems existing in the prior art, the present invention provides a multi-agent-based multi-round interactive content direction guidance method. The present invention effectively improves the accuracy of the guidance content through flexible and changeable interactive modes.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A multi-round interactive content direction guidance method based on multi-agents includes the following guidance steps:

[0008] S1. Based on the principle of natural language processing, we build inquiry behavior agents, chat behavior agents, guidance behavior agents and planning agents for human-computer interaction.

[0009] S2, select an interaction event and start human-computer interaction. In the current round of interaction, the inquiry behavior agent, the chat behavior agent, and the guidance behavior agent generate corresponding candidate responses according to the user input, user profile, and interaction history;

[0010] S3, the planning agent selects a candidate response that meets the requirements from the candidate responses generated in step S2 according to the interaction history and user preferences as feedback to the user input in the current round of interaction, and uses the candidate response as the final response in the current round of interaction to guide the user's input in the next round of interaction;

[0011] S4. Complete multiple rounds of human-computer interaction in sequence according to the contents of step S2 and step S3 to form a complete interaction process, and obtain the user's opinion on the interaction event based on the interaction process.

[0012] As a further solution of the present invention: the inquiry behavior agent, the small talk behavior agent, the guidance behavior agent and the planning agent share the same agent architecture: a memory module, an instruction module and a response module;

[0013] Memory module: used to save and update interaction history and user profiles;

[0014] Instruction module: used to store task instructions of inquiry behavior agent, chat behavior agent and guidance behavior agent;

[0015] Response module: Generates corresponding candidate responses based on the interaction history and user profile in the memory module and the task instructions in the instruction module.

[0016] As a further solution of the present invention, the process of generating candidate responses is as follows:

[0017] S21. Before starting the interaction: select the interaction event and store it in the interaction history;

[0018] S22. The memory module receives the user input in the current round of interaction and feeds it back to the instruction module; the instruction module instructs the response module to generate candidate responses for the inquiry behavior agent, the chat behavior agent and the guidance behavior agent corresponding to the user input based on the saved task instructions and in combination with the interaction history and the user profile.

[0019] As a further solution of the present invention: the process of determining the final response is as follows:

[0020] S31. In the current round of interaction, for the three candidate responses that have been generated, the memory module first reviews the saved interaction history to confirm whether the three candidate responses have been used as final responses in the previous interaction process;

[0021] S32: If yes, then regenerate three candidate responses according to the content of step S22, and make a judgment according to step S31; otherwise, proceed to step S33;

[0022] S33. The response module evaluates user preferences based on the interaction history and user profiles stored in the memory module; if the preferences are sufficient, the corresponding candidate responses are selected as the final responses through the final response model; if the preferences are insufficient, the candidate responses with the largest user information gain are selected as the final responses.

[0023] As a further solution of the present invention: during the interaction process, the user will actively evaluate each round of interaction; if the active evaluation is unsatisfactory, the interaction trajectory of this round will be counted, and the error summary corresponding to the interaction trajectory will be input into the final response model to continue the subsequent interaction.

[0024] As a further solution of the present invention: the candidate responses of the inquiry behavior agent are:

[0025] R ask =π ask (I ask , D h , U P );

[0026] Among them, R ask represents the candidate responses generated by the query behavior agent; π ask I represents the policy function of the query behavior agent; ask Indicates the task instruction to ask the behavior agent; D h Represents the interaction history; U P Represents a user profile;

[0027] The candidate responses for the chat behavior agent are:

[0028] R chat =π chat (I chat , D h , U P );

[0029] Among them, R chat represents the candidate responses generated by the chat behavior agent; π chat I represents the policy function of the chat behavior agent; chat represents the task instructions of the chat behavior agent;

[0030] The candidate responses for guiding the behavior agent are:

[0031] R guid =π guid (I guid , D h , U P);

[0032] Among them, R guid represents the candidate responses generated by the guiding behavior agent; π guid I represents the policy function that guides the behavior of the agent; guid Represents the task instructions that guide the behavior of the agent.

[0033] As a further solution of the present invention: the final response model is expressed as follows:

[0034] R s =π plan (I plan ,R ask ,R guid ,R chat ,D h ,A h ,E s );

[0035] Among them, R s represents the candidate response finally selected, i.e., the final response; π plan I represents the policy function of the planning agent; plan is the task instruction of the planning agent; E s Indicates an error summary.

[0036] As a further solution of the present invention: the interaction trajectory is expressed as follows:

[0037]

[0038] in, represents the interaction trajectory formed by the interaction history from when the user was dissatisfied in the i-th round of interaction to when the user was dissatisfied in the t-th round of interaction; if the t-th round of interaction was the first time when the user was dissatisfied, i = 1; Represents the user profile at the i-th round of interaction; represents the final response in the i-th round of interaction; Represents user feedback during the i-th round of interaction; Represents the user profile at the tth round of interaction; represents the final response in the tth round of interaction; Represents the user feedback during the tth round of interaction;

[0039] The error summary is as follows:

[0040]

[0041] Among them, π reflect Represents the policy function for error summarization.

[0042] As a further solution of the present invention: the task instruction of the inquiry behavior agent is: to elicit the user's preferences by asking questions; the task instruction of the guiding behavior agent is: to help users form a comprehensive understanding by providing diverse information and viewpoints from different angles, thereby reducing the impact of the information cocoon; the task instruction of the chat behavior agent is: to chat with the user to understand the user's preferences; the task instruction of the planning agent is: please select the most appropriate response from the candidate responses of the inquiry behavior agent, the guiding behavior agent and the chat behavior agent based on the current round of interaction.

[0043] As a further solution of the present invention: the task instruction of error summary is: write down the reasons for user dissatisfaction according to the interaction trajectory.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The present invention realizes the refined division of labor in human-computer interaction by constructing multiple functional modules such as inquiry behavior agent, chat behavior agent, guidance behavior agent and planning agent. Each agent independently generates candidate responses according to its own responsibilities (such as inquiry, chat, guiding users or planning responses). This division of labor and cooperation can more accurately capture the user's intentions and needs, thereby generating feedback that is closer to the user's expectations. It is built based on the principles of natural language processing, which means that it can deeply understand the user's natural language input, including semantics, context and emotion. This deep understanding ability enables the system to more accurately parse user intentions, reduce misunderstandings and ambiguities, and improve the accuracy of guidance. When generating candidate responses, the system not only relies on the current user input, but also combines user profiles (such as personal information, preferences, etc.) and previous interaction history. This context-aware capability enables the system to continuously learn and adapt to the personalized needs of different users, further enhancing the accuracy and personalization level of guidance. The planning agent is responsible for selecting the response that best meets the user's preferences and interaction history requirements from multiple candidate responses as the final feedback. This process optimizes decisions by comprehensively considering various factors (such as user feedback, historical success rates, etc.), ensuring that each interaction can maximize user needs and improve the accuracy of guidance.

[0046] 2. When guidance fails, the present invention dynamically adjusts the dialogue behavior plan, generates guidance suggestions and correction experience, so as to optimize the accuracy of subsequent dialogue and guidance.

[0047] 3. The planning agent of the present invention controls the dialogue process by reviewing the dialogue history, evaluating user preference information and selecting the most appropriate candidate response to ensure the diversity and rationality of the dialogue.

[0048] 4. The system of the present invention can effectively collect user preference information through questioning, guidance and chatting in multi-round conversations, and dynamically adjust the recommendation strategy according to user feedback to improve the accuracy of guidance and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a structural schematic diagram of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] 1. System Overview

[0052] Users express their opinions and comments through natural language, and the system responds to the user's request and may ask for more details or engage in small talk to get the user's point of view. After multiple rounds of interactive dialogue, the system attempts to guide the user to form a more comprehensive point of view or change the user's extreme point of view. If the user accepts the guidance, the dialogue ends; otherwise, the system will provide multiple guidance strategies and summarize them before the end of the dialogue. If the user accepts the point of view guided by the system, the dialogue session is considered a successful sample and the user will end the dialogue. If the dialogue reaches the maximum round threshold, but the system still fails to provide a satisfactory point of view or suggestion to the user, the system will adopt a fallback strategy and provide a list of points of view containing multiple recommendations before the end of the dialogue. At this time, if the system cannot guide successfully before reaching the maximum round threshold, the dialogue will be considered an unsuccessful sample.

[0053] 2. Multi-agent behavior planning

[0054] In order to achieve an engaging interaction, the system needs to use multiple dialogue behaviors. Therefore, it is necessary to effectively plan the dialogue flow by selecting appropriate dialogue behaviors for each round. Dialogue flow planning plays a vital role in improving the efficiency and attractiveness of the system. To accomplish this complex planning task, the present invention designs a multi-agent behavior planning framework, such as Figure 1 As shown, the planning framework consists of multiple behavior agents and a planning agent.

[0055] 2.1 Behavior Agent

[0056] An engaging multi-round interactive content orientation system should include multiple dialogue behaviors in the multi-round interaction. Therefore, the present invention designs three types of behavior agents to implement various dialogue behaviors, including inquiry behavior agent, guiding behavior agent and chat behavior agent.

[0057] Inquiry agent: aims to elicit user preferences and needs by asking questions. For example, if a user expresses interest in a certain type of movie, the inquiry agent may further ask the user about their preference for specific directors or actors in that type of movie.

[0058] Chatting behavior agent: Improve user interest and better understand user preferences by chatting with users. These chats not only make the conversation more natural, but also inadvertently collect more user preference information.

[0059] Guiding agent: responsible for providing users with diverse information and actively guiding the direction of the conversation, so as to guide users to form a more comprehensive view. For example, if a user is interested in a certain policy, the guiding agent can provide opinions and data supporting and opposing the policy; when the user only focuses on one aspect, the guiding agent can ask related questions from different angles to encourage the user to think about other possibilities.

[0060] 2.2. Agent Architecture

[0061] Each behavioral agent shares the same agent architecture, which consists of three modules: memory module, command module and response module.

[0062] Memory module: Contains interaction history and user profile. Interaction history records the conversation between the user and the system in previous rounds. User profile records the user's current needs and browsing history.

[0063] Instruction module: contains the agent's task instructions. Different agents have different task instructions.

[0064] The task instruction for the inquiry behavior agent is: “You should elicit the user’s preferences by asking questions.”

[0065] The task instructions for the agent guiding the behavior are: "Your task is to help users form a more comprehensive understanding by providing diverse information and different perspectives, thereby reducing the impact of information cocoons."

[0066] The task instruction of the chat behavior agent is: "You should chat with users to understand their preferences. You can express your appreciation for certain item elements to guide the conversation towards these elements and thus obtain the user's preferences."

[0067] Response module: The response module is responsible for generating candidate responses based on the memory module and the instruction module. Each agent generates a candidate response based on its own instructions, interaction history, and user profile.

[0068] The candidate responses of the query behavior agent are shown in formula (1):

[0069] Rask =π ask (I ask , D h , U P ) (1)

[0070] The candidate responses of the chat behavior agent are shown in formula (2):

[0071] R chat =π chat (I chat , D h , U P ) (2)

[0072] The candidate responses of the guided behavior agent are shown in formula (3):

[0073] R guid =π guid (I guid , D h , U P ) (3)

[0074] These candidate responses are passed to the planning agent, which selects the most suitable candidate response as the final response of the system based on the current dialogue context. Through the collaboration of the above agents, the system can dynamically adjust the dialogue behavior in multi-round dialogues, thereby providing more personalized and precise guidance services.

[0075] 3. Planning agent

[0076] In order to control the flow of dialogue, it is crucial to select the most appropriate candidate response as the system response. The planning agent is responsible for selecting the most appropriate response for the current dialogue situation in multiple dimensions such as informativeness and participation through multi-step reasoning to ensure the rationality and effectiveness of the dialogue process. Its tasks include reviewing the interaction history, evaluating user preference information, and selecting the most appropriate candidate response to guide users to form a more comprehensive view. For example, when the available user preference information is insufficient, the planning agent infers which candidate response can potentially bring higher user information gain, and then selects the candidate response of a certain behavior agent as the final response. The planning agent also uses an agent architecture, and the functions of the agent in the planning agent are as follows:

[0077] Memory module: contains the following:

[0078] Interaction history: Contains user utterances and the system’s responses in previous turns. This information helps understand the context and progression of the conversation.

[0079] Interaction history: records the dialogue behaviors chosen by the planning agent in previous dialogue turns. This helps avoid repeating the same behaviors and ensures diversity and fluency of the dialogue.

[0080] Guided experience: Experience learned from user feedback in past rounds. These experiences are used to improve decisions in future conversations.

[0081] Instruction module: The instruction module specifies the role and background knowledge of the planning agent. The specific planning agent instruction is: "Please select the most appropriate response from the candidate responses of the inquiry agent, guidance agent, and small talk agent based on the current conversation context."

[0082] Planning module: The planning module makes decisions through multi-step reasoning and selects the response that best suits the current dialogue situation.

[0083] The processing steps of the planning agent are as follows:

[0084] Step 1: Review the interaction history and avoid repeating the same actions in multiple turns. This helps maintain conversation diversity and user interest.

[0085] Step 2: Determine whether the available user preferences are sufficient based on the interaction history. If sufficient, the planning agent selects candidate responses to guide the agent based on the preference information expressed by the user, further guiding the user to form a more comprehensive opinion.

[0086] Step 3: If the available user preferences are insufficient, determine which candidate response is more appropriate or can bring higher user information gain. For example, when the user preference information is insufficient, the response of asking the agent may be more appropriate because it can elicit more user preference information.

[0087] After multiple steps of reasoning, the planning agent will output a choice, and then determine the final system response through the final response model. The final response model is shown in formula (4):

[0088] R s =π plan (I plan , R ask , R guid , R chat , D h , A h , E s ) (4)

[0089] 4. User feedback experience summary

[0090] Users provide feedback when interacting with the interactive wind direction guidance system (MCCI). These feedbacks include the user's active evaluation of the system response and also implicitly express the user's satisfaction with the guidance. Therefore, the present invention incorporates the experience of user feedback into both the behavior agent and the planning agent instructions. The main purpose is to dynamically adjust the dialogue behavior when the guidance fails to improve the decision-making of future dialogues.

[0091] When the user is dissatisfied, it indicates that there is a problem with the process control of the system in the previous rounds of interaction. Therefore, the present invention first collects multiple rounds of interaction trajectories, including user profiles, final responses, and user feedback. When the system provides a failed guide in the tth round, the system will reflect on the trajectory from the i-th round to the tth round, where the i-th round is the round immediately following the last failed recommendation. If the tth round is the first failed recommendation, i=1. The interaction trajectory is shown in formula (5):

[0092]

[0093] When booting fails, an error summary is generated as shown in formula (6):

[0094]

[0095] Error summaries are natural language descriptions of errors made by the acting and planning agents in a conversation. The instructions for generating error summaries are as follows: “Based on past interaction trajectories, write a few sentences explaining why the user was dissatisfied with your recommendations.”

[0096] To avoid repeating the same mistakes, the mechanism generates guidance suggestions and corrective experiences to indicate how each agent should respond in the next round.

[0097] Based on the previous content, the planning agent can dynamically adjust the dialogue behavior in multi-round dialogues to ensure the rationality and effectiveness of the dialogue process, thereby guiding users to form a more comprehensive perspective.

[0098] 5. Main steps of the guidance method

[0099] The guiding steps of the present invention are as follows:

[0100] S1. Based on the principles of natural language processing, we construct inquiry behavior agents, chat behavior agents, guidance behavior agents, and planning agents for human-computer interaction.

[0101] S2. Select an interaction event and start human-computer interaction. In the current round of interaction, the inquiry behavior agent, the chat behavior agent, and the guidance behavior agent generate corresponding candidate responses based on user input, user profile, and interaction history.

[0102] The process of generating candidate responses is as follows:

[0103] S21. Before starting the interaction: select the interaction event and store it in the interaction history;

[0104] S22. The memory module receives the user input in the current round of interaction and feeds it back to the instruction module; the instruction module instructs the response module to generate candidate responses for the inquiry behavior agent, the chat behavior agent and the guidance behavior agent corresponding to the user input based on the saved task instructions and in combination with the interaction history and the user profile.

[0105] S3. The planning agent selects a candidate response that meets the requirements as feedback to the user input in the current round of interaction based on the interaction history and user preferences, and uses the candidate response as the final response in the current round of interaction to guide the user's input in the next round of interaction.

[0106] The final response is determined as follows:

[0107] S31. In the current round of interaction, for the three candidate responses that have been generated, the memory module first reviews the saved interaction history to confirm whether the three candidate responses have been used as final responses in the previous interaction process;

[0108] S32: If yes, then regenerate three candidate responses according to the content of step S22, and make a judgment according to step S31; otherwise, proceed to step S33;

[0109] S33, the response module evaluates the user's preference based on the interaction history and user profile stored in the memory module; if the preference is sufficient, the corresponding candidate response is selected as the final response through the final response model; if the preference is insufficient, the candidate response with the largest user information gain is selected as the final response. During the interaction process, the user will actively evaluate each round of interaction; if the active evaluation is unsatisfactory, the interaction trajectory of this round will be counted, and the error summary corresponding to the interaction trajectory will be input into the final response model at the same time to continue the subsequent interaction.

[0110] S4. Complete multiple rounds of human-computer interaction in sequence according to the contents of step S2 and step S3 to form a completed interaction process, and obtain the user's opinion on the interaction event based on the interaction process.

[0111] 6. Embodiment 1

[0112] 6.1. Scene Setting

[0113] User: Xiao Ming.

[0114] System: Multi-Cross Interactive Content Direction System (MCCI).

[0115] 6.2. Initial Dialogue

[0116] System (chatting behavior agent): "Hi, Xiao Ming! How are you doing recently? Are there any good movies you want to talk about?"

[0117] Xiao Ming: "Well, I'm quite interested in science fiction movies recently."

[0118] 6.3. First round of interaction

[0119] Memory module: records user profile (Xiao Ming is interested in science fiction movies) and interaction history (Xiao Ming mentioned his interest in science fiction movies).

[0120] Ask the agent (generates candidate responses based on its task instructions): "It sounds like you are very interested in science fiction movies! Do you have any favorite science fiction movie directors or actors?"

[0121] Planning Agent: Reviewing the interaction history (no repeated behavior), evaluates the user’s preference information (the user has expressed an interest in science fiction movies), but the preference information is insufficient. Therefore, the Planning Agent selects the candidate response of the Inquiry Agent as the final response.

[0122] System (asking the behavior agent): "It sounds like you are very interested in science fiction movies! Do you have any favorite science fiction movie directors or actors?"

[0123] Xiao Ming: "I like Christopher Nolan's movies very much. His "Interstellar" is one of my favorite science fiction movies."

[0124] 6.4. Second round of interaction

[0125] Memory module: updates user profile (Xiao Ming likes Christopher Nolan and his work "Interstellar") and records new interaction history.

[0126] Guided Behavior Agent (generates candidate responses based on its task instructions): "Christopher Nolan is indeed a very good director, and his works often explore complex themes such as time and memory. In addition to Interstellar, have you seen his other works? For example, Inception or The Prestige?"

[0127] Planning agent: Review the interaction history (the previous round was to ask the behavior agent) and evaluate the user preference information (enough). The planning agent selects candidate responses that guide the behavior agent to further guide the user to form a more comprehensive opinion.

[0128] System (guided behavior agent): "Christopher Nolan is indeed an excellent director, and his works often explore complex themes such as time and memory. In addition to "Interstellar", have you seen his other works? For example, "Inception" or "The Prestige"? "

[0129] Xiao Ming: "Wow, I haven't seen Inception. It sounds interesting. I will go find it and watch it."

[0130] 6.5. The third round of interaction

[0131] Memory module: Continue to update user profiles and interaction history.

[0132] Small talk agent (generating candidate responses based on instructions): "Inception is indeed a very brain-burning movie. Remember to prepare paper and pen when watching it, otherwise you may not be able to follow the plot! Haha, just kidding. Are you interested in other aspects of science fiction movies? For example, science fiction novels or science fiction games?"

[0133] Planning agent: Evaluates the current conversation context and believes that the user has a comprehensive understanding of science fiction movies and may be interested in other science fiction content. The planning agent selects candidate responses for the small talk behavior agent to maintain the naturalness of the conversation and the user's interest.

[0134] System (chatting behavior agent): "Inception is indeed a very brain-burning movie. Remember to prepare paper and pen when watching it, otherwise you may not be able to follow the plot! Haha, just kidding. Are you interested in other aspects of science fiction movies? For example, science fiction novels or science fiction-themed games?"

[0135] Xiao Ming: "Yes, I also like reading science fiction novels, especially the works of Liu Cixin."

[0136] 6.6 Subsequent Interaction

[0137] The interaction will continue, and the system will dynamically adjust the interaction behavior based on Xiao Ming's answers and feedback to ensure the diversity and effectiveness of the conversation. At the same time, the system will also record and analyze user feedback to better meet user needs and improve user satisfaction in future conversations.

[0138] 7. Embodiment 2

[0139] In order to verify the effectiveness of the present invention, experiments were conducted to simulate the interactive guidance process using the MovieLens dataset.

[0140] 7.1 Experimental Setup

[0141] In the experiments, all MCCI-C variants and the ChatGPT baseline use the gpt-3.5-turbo-0613 version, and the MCCI-L variants and the Llama2 baseline use the open source Llama-2-70b-chat-hf version.

[0142] 7.2 Evaluation Metrics

[0143] In order to quantitatively evaluate the performance of MCCI, three automated metrics are designed:

[0144] Success Rate: In order to evaluate the overall efficiency, the success rate of all test set samples is calculated:

[0145]

[0146] Among them, Success_Rate represents the success rate. N is the number of samples in the test set, N su is the number of successful samples.

[0147] Hit Ratio@K: Calculates the number of hits for the first K items in the recommendation list, where K = 5 or 10. Also includes the number of successful samples in the final calculation:

[0148]

[0149] Among them, Hit_Ratio represents the hit rate. un Indicates the number of unsuccessful samples. When the target item appears in the recommendation list, hit(i) = 1. Otherwise, hit(i) = 0.

[0150] Average Turns: To evaluate efficiency, calculate the average number of turns in a conversation:

[0151]

[0152] Among them, Average_Turns represents the average number of turns. Ti is the number of rounds in sample i.

[0153] 7.3 Dataset

[0154] Experiments were conducted on the MovieLens dataset, a benchmark dataset for movie recommendations. The dataset contains 162,541 users, 62,423 movies, and 25,000,095 ratings. To build a user simulator, the interaction sequence was truncated to 20 items, which were used to create a browsing history. The 21st item was used as the prediction target item. The browsing history was then filtered to keep only items belonging to the same category. 100 samples were randomly selected and their browsing history length was ensured to be no less than 5.

[0155] 7.4. Comparison Methods: To evaluate the effectiveness of MCCI, it is compared with four strong interactive recommendation baseline methods:

[0156] KBRD: Contains a recommendation module and a Transformer-based dialogue module. It leverages external knowledge graphs to enhance the semantics of entities mentioned in the dialogue history.

[0157] BARCOR: Uses BART’s bidirectional encoder as the recommendation module and the autoregressive decoder as the dialogue module.

[0158] ChatGPT: A closed-source LLM from OpenAI. Using version gpt-3.5-turbo-0613, a strong single agent was built based on the same prompt.

[0159] Llama2: An open source LLM that uses the Llama2 model with 7 billion parameters to build a single-agent approach to planning dialogue flows and compare it to a multi-agent framework.

[0160] Two variants of MCCI are proposed: MCCI-C and MCCI-L, which use ChatGPT and Llama2 as backbones respectively. In order to verify the effectiveness of each module in the MCCI framework, an ablation model based on MCCI-C is used in the experiments.

[0161] MCCI w / o FER: The Feedback Experience Summary (FER) module in MCCI was removed to verify the overall effectiveness of the reflection mechanism in MCCI.

[0162] 7.5 Main Results

[0163] Table 1 shows the performance comparison of the proposed MCCI and the baseline methods on four indicators. It can be seen that the LLM-based interaction guidance methods (e.g., ChatGPT and Llama2) significantly outperform the traditional dialogue recommendation methods (e.g., BARCOR and KBRD) in performance. This is because BARCOR and KBRD rely on knowledge graphs and have limitations in successfully recommending items that are not in the knowledge graph; on the other hand, they rely on smaller-scale generative models, which have weak reasoning and natural language understanding capabilities.

[0164] Table 1 Comparison results

[0165]

[0166]

[0167] MCCI-C outperforms its base LLM ChatGPT in all indicators. This shows that with the help of the MCCI framework, LLM-based interactive recommendation can be improved in user preference collection and recommendation accuracy. Although MCCI-L outperforms Llama2 in all recommendation accuracy indicators (e.g., success rate and hit rate), it is observed that the average number of rounds of MCCI-L is slightly higher than that of Llama2. Presumably, this may be because the MCCI method is more cautious in recommending, which gives it an advantage in success rate and recommendation accuracy. In contrast, single-agent methods (e.g., ChatGPT and Llama2) tend to recommend items to users directly and frequently. Although this behavior improves efficiency, it also increases the risk of user dissatisfaction and is not conducive to capturing user preferences.

[0168] Furthermore, while Llama2 outperforms ChatGPT as a single agent, MCCI-L lags behind MCCI-C in all metrics. We speculate that this phenomenon is due to Llama2’s weak ability to follow complex instructions, which leads to poor results in the proposed multi-agent framework.

[0169] As can be seen from Table 1, the LLM-based interactive guidance shows better performance in comparison with traditional methods. MCCI-C is significantly better than ChatGPT in terms of success rate, average number of rounds, hit rate @5, and hit rate @10, indicating that the multi-agent framework and user feedback perception reflection mechanism have a significant effect on improving recommendation performance.

[0170] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A multi-agent-based multi-round interactive content direction guidance method, characterized in that: The following boot steps are included: S1. Based on the principle of natural language processing, we build inquiry behavior agents, chat behavior agents, guidance behavior agents and planning agents for human-computer interaction. S2, select an interaction event and start human-computer interaction. In the current round of interaction, the inquiry behavior agent, the chat behavior agent, and the guidance behavior agent generate corresponding candidate responses according to the user input, user profile, and interaction history; S3, the planning agent selects a candidate response that meets the requirements from the candidate responses generated in step S2 according to the interaction history and user preferences as feedback to the user input in the current round of interaction, and uses the candidate response as the final response in the current round of interaction to guide the user's input in the next round of interaction; S4. Complete multiple rounds of human-computer interaction in sequence according to the contents of step S2 and step S3 to form a complete interaction process, and obtain the user's opinion on the interaction event based on the interaction process.

2. The method for guiding content direction based on multi-round interaction of multiple agents according to claim 1, characterized in that: The inquiry agent, the chat agent, the guidance agent, and the planning agent share the same agent architecture: a memory module, an instruction module, and a response module; Memory module: used to save and update interaction history and user profiles; Instruction module: used to store task instructions of inquiry behavior agent, chat behavior agent and guidance behavior agent; Response module: Generates corresponding candidate responses based on the interaction history and user profile in the memory module and the task instructions in the instruction module.

3. The method for guiding content direction based on multi-round interaction of multiple agents according to claim 2, characterized in that: The process of generating candidate responses is as follows: S21. Before starting the interaction: select the interaction event and store it in the interaction history; S22, the memory module receives the user input in the current round of interaction and feeds it back to the instruction module; The instruction module instructs the response module to generate candidate responses for feedback of the user input corresponding to the inquiry behavior agent, the chat behavior agent and the guidance behavior agent according to the stored task instructions in combination with the interaction history and the user profile.

4. The method for guiding content direction based on multi-round interaction of multiple agents according to claim 3 is characterized in that: The final response is determined as follows: S31. In the current round of interaction, for the three candidate responses that have been generated, the memory module first reviews the saved interaction history to confirm whether the three candidate responses have been used as final responses in the previous interaction process; S32: If yes, then regenerate three candidate responses according to the content of step S22, and make a judgment according to step S31; otherwise, proceed to step S33; S33, the response module evaluates the user preference based on the interaction history and user profile stored in the memory module; If the preference is sufficient, the corresponding candidate response is selected as the final response through the final response model; If the preference is insufficient, the candidate response with the largest user information gain is selected as the final response.

5. The method for guiding content direction based on multi-round interaction of multiple agents according to claim 4, characterized in that: During the interaction process, users will actively evaluate each round of interaction; if the active evaluation is unsatisfactory, the interaction trajectory of this round will be counted, and the error summary corresponding to the interaction trajectory will be input into the final response model to continue subsequent interactions.

6. A method for guiding content direction based on multi-round interaction based on multiple agents according to any one of claims 1 to 5, characterized in that: The candidate responses for the interrogation behavior agent are: R ask =π ask (I ask ,D h ,U P ); Among them, R ask represents the candidate responses generated by the query behavior agent; π ask I represents the policy function of the query behavior agent; ask Indicates the task instruction to ask the behavior agent; D h Represents the interaction history; U P Represents a user profile; The candidate responses for the chat behavior agent are: R chat =π chat (I chat ,D h ,U P ); Among them, R chat represents the candidate responses generated by the chat behavior agent; π chat I represents the policy function of the chat behavior agent; chat represents the task instructions of the chat behavior agent; The candidate responses for guiding the behavior agent are: R guid =π guid (I guid ,D h ,U P ); Among them, R guid represents the candidate responses generated by the guiding behavior agent; π guid I represents the policy function that guides the behavior of the agent; guid Represents the task instructions that guide the behavior of the agent.

7. The method for guiding content direction based on multi-round interaction of multiple agents according to claim 6, characterized in that: The final response model is expressed as follows: R s π plan (I plan ,R ask ,R guid ,R chat ,D h ,AM h ,E s )4 Among them, R s represents the candidate response finally selected, i.e., the final response; π plan I represents the policy function of the planning agent; plan is the task instruction of the planning agent; E s Indicates an error summary.

8. The method for guiding content direction based on multi-round interaction of multiple agents according to claim 7, characterized in that: The interaction trajectory is represented as follows: in, represents the interaction trajectory formed by the interaction history from when the user was dissatisfied in the i-th round of interaction to when the user was dissatisfied in the t-th round of interaction; if the t-th round of interaction was the first time when the user was dissatisfied, i = 1; Represents the user profile at the i-th round of interaction; represents the final response in the i-th round of interaction; Represents the user feedback during the i-th round of interaction; Represents the user profile at the tth round of interaction; represents the final response in the tth round of interaction; Represents the user feedback during the tth round of interaction; The error summary is as follows: Among them, π reflect Represents the policy function for error summarization.

9. The method for guiding content direction based on multi-round interaction of multiple agents according to claim 8, characterized in that: The task instructions for the inquiry behavior agent are: to elicit user preferences by asking questions; the task instructions for the guiding behavior agent are: to help users form a comprehensive understanding by providing diverse information and perspectives from different angles, thereby reducing the impact of information cocoons; the task instructions for the chat behavior agent are: to chat with users to understand their preferences; the task instructions for the planning agent are: please select the most appropriate response from the candidate responses of the inquiry behavior agent, guiding behavior agent and chat behavior agent based on the current round of interaction.

10. The method for guiding content direction based on multi-round interaction of multiple agents according to claim 9, characterized in that: The task instruction for error summary is: based on the interaction trajectory, write down the reasons for user dissatisfaction.

Images