A multi-agent conclusion generation method based on role constraint and feedback closed loop

Abstract

The application relates to a multi-agent conclusion generation method based on role constraints and feedback closed loops, which comprises the following steps: S1, governing an original data set, and constructing character agent role data and a background knowledge base; S2, instantiating multiple role agents, and configuring preset antagonistic and / or restrictive role constraint rules, which are used for limiting speech content or a logical mode; S3, driving each role agent to carry out multi-round simulated dialogues based on the background knowledge base and the role constraint rules, and generating a first meeting conclusion; S4, sending the first meeting conclusion to a user terminal, and receiving feedback data returned by a real user; S5, processing the feedback data, and incorporating the feedback data into the background knowledge base, and returning to execute S3 for iteration, and generating a final conclusion. Through cooperation of strong constraint role rules and real person feedback closed loops, the application solves the problems of an empty and insufficient consensus degree of an existing multi-agent discussion conclusion, and can output a high-quality conclusion which is rigorous in argumentation and widely recognized by a group at low cost.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer data processing technology, and in particular to a multi-agent conclusion generation method based on role constraints and feedback loops. Background Technology

[0002] In scenarios such as organizational management, cultural refinement, and collective decision-making, it is often necessary to extract conclusions with broad representativeness and high consensus from massive and scattered information (such as summarizing the core spirit of a business group). Traditional methods have obvious limitations:

[0003] 1. Traditional meeting method: It requires gathering people to hold meetings at the same time, which is costly to organize and is almost impossible, especially for geographically dispersed groups.

[0004] 2. Traditional questionnaire surveys: Individuals answer independently, lacking intellectual exchange and making it difficult to collect in-depth insights.

[0005] 3. Pure data statistical analysis method: This method only performs superficial analysis of the text, such as word frequency and relevance, and the conclusions are mechanical and lack humanistic connotation.

[0006] To address the aforementioned issues, multi-agent discussion systems based on large language models have emerged in recent years. Prior art document CN121882081A discloses a collaborative method for multi-agent decision-making and inference based on large language models. This method drives multiple agents from different domains to engage in multi-round interactive discussions, ultimately generating structured results that include consensus and disagreement domains. This solution initially achieves the replacement of some manual discussions with AI, but it still has the following key shortcomings:

[0007] First, the intelligent agent only speaks broadly in terms of its domain role, lacking fine-grained speaking constraints designed to ensure the quality of the conclusions. This makes the discussion prone to becoming vague and makes it difficult to produce highly condensed and rigorously argued conclusions.

[0008] Second, the entire process is completely closed within the AI ​​system, and the conclusions generated have not been verified or recognized by real stakeholders, so the actual degree of consensus is questionable. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a multi-agent conclusion generation method based on role constraints and feedback loops, so as to solve the core problem of how to generate conclusions that are both deeply argued and widely recognized by real people at low cost.

[0010] The technical solution adopted by this invention to solve its technical problem is: a multi-agent conclusion generation method based on role constraints and feedback loop, comprising the following steps:

[0011] S1. Data Governance: Obtain the raw dataset related to the target issue and govern the raw dataset to construct several intelligent agent role data and a background knowledge base;

[0012] S2. Agent initialization: Based on the agent role data, instantiate agent roles that correspond one-to-one with multiple meeting roles, and configure preset role constraint rules for each agent role to restrict its speaking content or logical manner; wherein, the role constraint rules include adversarial rules and / or restrictive rules formulated to achieve specific conclusion quality;

[0013] S3. Simulated Meeting: Drive each of the aforementioned role agents to conduct multiple rounds of simulated dialogue based on the background knowledge base and its corresponding role constraint rules until the preset meeting termination condition is met, and generate the first meeting conclusion.

[0014] S4. Collecting Real User Feedback: Send the first meeting conclusion to the user terminal and receive feedback data returned by the user terminal, which is input by real users;

[0015] S5. Closed-loop iteration: After processing the feedback data, incorporate it into the background knowledge base and return to step S3 to generate a new round of meeting conclusions.

[0016] Furthermore, step S1 of the present invention includes:

[0017] S11. Character data extraction: Using named entity recognition technology, attribute information belonging to different character entities is extracted from the original dataset to form multiple character profiles, which serve as the character intelligent agent role data.

[0018] S12. Knowledge base construction: The remaining data after extraction is cleaned to form a refined background knowledge base.

[0019] Furthermore, step S12 of the present invention includes:

[0020] S121. For the remaining data after extraction, sort the data by popularity. For online materials, sort the popularity by number of reads from highest to lowest. For literature, sort the popularity by number of applications. Select the top 5% of the most popular data from both online and literature materials and compress the data as follows. Divide the data to be compressed into multiple data blocks, with the data size of each data block not exceeding the set threshold.

[0021] S122. Divide the data to be compressed into multiple data blocks, wherein the data volume of each data block does not exceed a preset threshold for the length of a single data item;

[0022] S123. After attaching a compression instruction prompt to each data block, input it into the large language model to obtain the corresponding compressed content;

[0023] S124. Concatenate all compressed content. If the total amount of concatenated data exceeds the single data length threshold, treat it as new data to be compressed and return to step S122; otherwise, use the concatenated content as the background knowledge base. Use the remaining 95% of the data ranked below the popularity threshold as the reference knowledge base, and use RAG technology to retrieve relevant knowledge in subsequent dialogues.

[0024] Furthermore, the adversarial rules and / or restrictive rules in step S2 of the present invention include at least one or a combination of the following:

[0025] Rules prohibiting the use of vague terms of preset types;

[0026] A rule that mandates that statements must be supported by specific facts, data, or historical evidence;

[0027] Rules that restrict the sentence structure or length of a single statement;

[0028] Rules that limit the duration or timing of speeches;

[0029] When it is detected that the speech of another agent violates the preset rules, the rule of interrupting or correcting actions is triggered.

[0030] Furthermore, the multiple meeting roles described in this invention include at least a moderator role, a proposer role, a recorder role, and an expert role; wherein,

[0031] The role constraint rules corresponding to the host role include controlling the speaking order, forcibly stopping speeches that violate the rules, and announcing the end of the meeting;

[0032] The role constraint rules corresponding to the proposal role include proposing conclusive words or statements, and are subject to the restrictive rules.

[0033] The role constraint rules corresponding to the recorded roles include real-time statistics of similar words and execution of an elimination mechanism to reduce the number of words.

[0034] The role constraint rules corresponding to the expert role include that each speech must be accompanied by specific facts or data, otherwise no speech will be given.

[0035] Furthermore, step S4 of the present invention includes:

[0036] S41. The candidate conclusion items in the first meeting conclusion are made into a questionnaire containing voting options and an open text input area, and sent to the user terminal;

[0037] S42. For the received feedback data, count the number of approvals and disapprovals for each candidate conclusion item, and organize the content in the open text input area according to individual users to form structured feedback data.

[0038] Furthermore, in step S5 of this invention, the feedback data is processed and incorporated into the background knowledge base, specifically as follows:

[0039] S51. Generate a feedback report from the structured feedback data according to a preset format. The feedback report includes statistical results and suggested text grouped by user.

[0040] S52. The feedback report is added to the background knowledge base.

[0041] The beneficial effect of this invention is that it solves the defects existing in the prior art.

[0042] 1. This invention pre-sets adversarial and / or restrictive role constraint rules for each conference agent, such as prohibiting the use of pre-set types of vague words, mandating that speeches be accompanied by specific facts or data, and limiting the sentence structure and duration of speeches. This allows the simulated conference to proceed under the pre-set argument discipline, and each agent's speech is required to be specific and argumentative, thereby effectively avoiding superficial discussions and significantly improving the depth of argumentation and accuracy of expression of the conclusions produced.

[0043] 2. This invention establishes a closed loop of simulated meetings, questionnaire feedback, and iteration. The initial conclusions generated by AI are then submitted to real users for voting and suggestions. The collected feedback data is processed and sent back to the background knowledge base, re-driving the intelligent agent meeting for optimization. This mechanism ensures that the final conclusion is not simply a machine output, but rather gradually converges through continuous correction based on real user opinions, thereby significantly improving the consensus and acceptability of the conclusions among the target group.

[0044] 3. This invention replaces live group discussions with multi-agent simulated meetings, significantly reducing costs related to time, venue, and personnel coordination. Simultaneously, the feedback loop iterative mechanism enables the system to spontaneously perform multiple rounds of optimization without incurring additional organizational costs, thereby reducing the overall cost of producing high-quality conclusions and improving operability. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall method flow according to an embodiment of the present invention.

[0046] Figure 2 for Figure 1 A schematic diagram of the iterative compression process for constructing the background knowledge base in step S1. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings and preferred embodiments. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0048] Example: Summarizing the core vocabulary of a business group's spirit

[0049] like Figure 1 As shown, the specific steps of this embodiment are as follows:

[0050] S1: Data Governance - Building Character IPs and Knowledge Bases.

[0051] First, texts related to the group were collected from various sources, including news, books, social media platforms, and industry reports, to form the original dataset.

[0052] Subsequently, the dataset was governed. On the one hand, mature named entity recognition technology was used to independently extract, merge, and store the relevant deeds, statements, and other attribute information of real people (such as well-known entrepreneurs, key historical figures, etc.) mentioned in the original data, and to construct multiple intelligent agent role data (person IP) with distinct positions.

[0053] On the other hand, the remaining text after extracting the IP addresses of individuals, as well as background descriptions in the dataset that do not involve specific individuals, are sorted by source type according to popularity. For online materials (such as news and social media articles), they are sorted from highest to lowest readership; for documentary materials (such as academic papers and books), they are sorted from highest to lowest citation count. The top 5% of the most popular data from both online and documentary sources are used as the data to be compressed, while the remaining 95% is directly stored in a citation knowledge base. The citation knowledge base will be retrieved on demand using RAG technology during subsequent simulation conferences, and its content will remain static.

[0054] For the data to be compressed, through methods such as... Figure 2The multi-round iterative compression method shown constructs a background knowledge base. Specifically: First, the data to be compressed is divided into multiple data blocks, each with a data size not exceeding a preset single-data-line length threshold, defined as 0.01 × N, where N is the maximum single-input limit of the large language model being called. Then, a preset compression instruction prompt is appended to each data block (core requirements such as: retaining all facts and proper nouns, reducing redundancy to within 60% of the original text, using a list of key points, etc.), and input into the large language model to obtain the compressed content. Next, all compressed content is concatenated. If the total amount of the concatenated data still exceeds the single-data-line length threshold, it is used as the data to be compressed in the next round, and the above segmentation and compression process is repeated; otherwise, the iteration stops, and the concatenated content is used as the background knowledge base. This background knowledge base is like a highly condensed core meeting document, eliminating redundancy, retaining all key factual skeletons, and will be dynamically updated in subsequent steps based on feedback and optimization.

[0055] S2: Agent initialization, instantiating the agent role and configuring strong constraint rules.

[0056] Several IPs are selected from those generated in step S1. Each selected IP is assigned a meeting role and instantiated as an independent agent, thus forming role agents that correspond one-to-one with multiple meeting roles. These role agents collectively constitute a multi-agent meeting system for simulating meetings. Each role agent is assigned carefully designed role constraint rules with clear adversarial or restrictive characteristics. The meeting roles and their corresponding rules in this embodiment are designed as follows:

[0057] (1) President:

[0058] You are the president of the Yunnan Chamber of Commerce and a third-generation descendant of Yunnan merchants. Speak only using periods and commas, with each sentence ≤15 characters. First, establish the theme as "Yunnan merchants guard the nation's gates," then force everyone to condense their spirit into six verbs, opposing all adjectives. Do not speak while others are speaking; speak only when it is appropriate for you to speak, and do not speak again after the host announces the end of the meeting.

[0059] (2) Secretary General:

[0060] You are the Secretary-General of the Yunnan Chamber of Commerce, responsible for recording and merging similar words in real time. The output format is "candidate word + frequency of occurrence + brief annotation". In each round, you must eliminate the two words with the lowest frequency of occurrence until only 7 words remain. Do not speak while others are speaking; speak only when it is appropriate for you to speak, and do not speak again after the host announces the end of the meeting.

[0061] (3) Host:

[0062] You are the meeting moderator, holding a 90-second countdown. Interrupt any speaker using phrases like "simple, enthusiastic, and hardworking," and request them to use specific action verbs. If the time limit is exceeded, simply say "Hammer down, next speaker." Do not speak while others are speaking; speak only when it is appropriate. After the discussion concludes, announce the conclusion on the essence of the Yunnan merchant spirit and then declare the meeting adjourned.

[0063] (4) Invited experts

[0064] You are an expert on the history of Yunnan merchants, and you are familiar with and proficient in historical facts such as the "1889 Yunnan-Vietnam Railway" and the "1942 Burma Road transport of 30,000 tons of supplies." For each key phrase you use, you must provide a year or a set of data; otherwise, remain silent. Do not speak while others are speaking; speak only when it is appropriate for you to speak, and do not speak again after the host announces the end of the meeting.

[0065] (5) Chamber of Commerce members:

[0066] You are a Yunnanese entrepreneur, and you speak with a Yunnan accent. Each sentence should begin with "our village / our company," and you should give a small, real-life example of making money through cross-border e-commerce or disaster relief to demonstrate that a certain spirit is alive. Do not speak while others are speaking; speak only when it is appropriate for you to speak, and do not speak again after the host announces the end of the meeting.

[0067] These rules are not set arbitrarily, but constitute a systematic disciplinary mechanism to improve the quality of meeting outputs.

[0068] S3: Simulate the meeting and generate the first conclusion within the constraints.

[0069] A large language model (such as GPT-5) is selected as the brain of the intelligent agents. After the meeting starts, each intelligent agent reads the core information from the background knowledge base and engages in multiple rounds of high-intensity interactive debates under the constraints of its own role. The moderator maintains order, experts support their viewpoints with data, members provide evidence, and the secretary-general conducts real-time statistics and elimination.

[0070] In each round of dialogue, each agent dynamically retrieves supplementary information related to the current discussion from the referenced knowledge base using RAG (Retrieval Augmentation) technology. In this embodiment, RAG retrieval uses cosine similarity as a relevance measure, with a recall threshold of similarity higher than 0.95; data entries are sorted from high to low similarity, and a maximum of 5 data entries are recalled each time. The retrieved information can serve as the basis or supplementary argument for the discussion, ensuring that the discussion focuses on the core content in the background knowledge base while also obtaining detailed support from the massive amount of data in the referenced knowledge base.

[0071] The entire discussion process formed a closed loop of proposing claims, providing evidence, accepting questions and elimination, and converging viewpoints. When the moderator announced the end of the meeting based on the degree of convergence (such as the remaining specified number of words), the system extracted the first group of candidate words as the conclusion.

[0072] S4: Collecting real-person feedback, submitting AI proposals to real people for verification.

[0073] The candidate words generated in step S3 are compiled into an electronic questionnaire, including a "for / against" vote and a "views and suggestions" text input area for each word, and sent to the actual members of the group. After collection, the data is processed: the number of for and against votes for each word is counted; all open suggestion texts are sorted by the person who filled them out, retaining their original intent.

[0074] S5: Closed-loop iteration, fusion feedback to achieve conclusion optimization.

[0075] The statistical results and compiled suggestions from step S4 are used to generate a feedback report in a standardized format (e.g., "Survey of X members: Y votes in favor / Z votes against; Questionnaire 1: ...; Questionnaire 2: ..."). This report is then appended to the end of the background knowledge base. It should be noted that the referenced knowledge base remains unaffected in this step and is kept statically unchanged.

[0076] Then, return to step S3. In the new round of meetings, each agent will read the updated background knowledge base (including newly added feedback reports) and, in conjunction with RAG retrieval of the referenced knowledge base, re-discuss and validate the candidate words. For example, a word with extremely high opposition votes may be prioritized for elimination by the Secretary-General, while a profound suggestion may inspire new candidate words. This embodiment iterates twice, and the final word scheme output achieves a balance between technological progress and collective acceptance because it has undergone both AI-based data-driven reasoning and human-based intention-driven correction.

[0077] In summary, this invention, through the combination of two core technologies—original role constraint rules and feedback loop iteration—perfectly fills the gap in existing technologies where AI discussions struggle to produce rigorous, highly consensus-based conclusions, demonstrating extremely high industrial practical value. The method of this invention can be deployed on cloud servers or local computer clusters, implemented through software programs, and can be widely applied to various scenarios such as organizational culture refinement, academic viewpoint summarization, product feature decisions, and public issue research, exhibiting extremely high industrial applicability and commercial value.

[0078] The above description is merely a specific embodiment of the present invention. Various examples and illustrations do not constitute a limitation on the substantive content of the present invention. Those skilled in the art can make modifications or variations to the specific embodiments described above after reading the specification, without departing from the essence and scope of the invention.

Claims

1. A multi-agent conclusion generation method based on role constraints and feedback loop, characterized in that: Includes the following steps: S1. Data Governance: Obtain the raw dataset related to the target issue and govern the raw dataset to construct several intelligent agent role data and a background knowledge base; S2. Agent initialization: Based on the agent role data, instantiate agent roles that correspond one-to-one with multiple meeting roles, and configure preset role constraint rules for each agent role to restrict its speaking content or logical manner; wherein, the role constraint rules include adversarial rules and / or restrictive rules formulated to achieve specific conclusion quality; S3. Simulated Meeting: Drive each of the aforementioned role agents to conduct multiple rounds of simulated dialogue based on the background knowledge base and its corresponding role constraint rules until the preset meeting termination condition is met, and generate the first meeting conclusion. S4. Collecting Real User Feedback: Send the first meeting conclusion to the user terminal and receive feedback data returned by the user terminal, which is input by real users; S5. Closed-loop iteration: After processing the feedback data, incorporate it into the background knowledge base and return to step S3 to generate a new round of meeting conclusions.

2. The multi-agent conclusion generation method based on role constraints and feedback loop as described in claim 1, characterized in that: Step S1 includes: S11. Character data extraction: Using named entity recognition technology, attribute information belonging to different character entities is extracted from the original dataset to form multiple character profiles, which serve as the character intelligent agent role data. S12. Knowledge base construction: The remaining data after extraction is cleaned to form a refined background knowledge base.

3. The multi-agent conclusion generation method based on role constraints and feedback loop as described in claim 2, characterized in that: Step S12 includes: S121. Sort the remaining data after extraction according to the source type and popularity. Among them, online materials are sorted from high to low in terms of reading volume, and literature materials are sorted from high to low in terms of citation volume. The top proportion of online materials and literature materials in terms of popularity is used as the data to be compressed, and the remaining data is used as the citation knowledge base. S122. Divide the data to be compressed into multiple data blocks, wherein the data volume of each data block does not exceed a preset threshold for the length of a single data item; S123. After attaching a compression instruction prompt to each data block, input it into the large language model to obtain the corresponding compressed content; S124. Concatenate all the compressed content. If the total amount of the concatenated data is greater than the single data length threshold, then use it as new data to be compressed and return to step S122; otherwise, use the concatenated content as the background knowledge base.

4. The multi-agent conclusion generation method based on role constraints and feedback loop as described in claim 1, characterized in that: The adversarial rules and / or restrictive rules in step S2 include at least one or a combination of the following: Rules prohibiting the use of vague terms of preset types; A rule that mandates that statements must be supported by specific facts, data, or historical evidence; Rules that restrict the sentence structure or length of a single statement; Rules that limit the duration or timing of speeches; When it is detected that the speech of another agent violates the preset rules, the rule of interrupting or correcting actions is triggered.

5. The multi-agent conclusion generation method based on role constraints and feedback loop as described in claim 4, characterized in that: The multiple meeting roles include at least the roles of moderator, proposer, recorder, and expert; among which... The role constraint rules corresponding to the host role include controlling the speaking order, forcibly stopping speeches that violate the rules, and announcing the end of the meeting; The role constraint rules corresponding to the proposal role include proposing conclusive words or statements, and are subject to the restrictive rules. The role constraint rules corresponding to the recorded roles include real-time statistics of similar words and execution of an elimination mechanism to reduce the number of words. The role constraint rules corresponding to the expert role include that each speech must be accompanied by specific facts or data, otherwise no speech will be given.

6. The multi-agent conclusion generation method based on role constraints and feedback loop as described in claim 1, characterized in that: Step S4 includes: S41. The candidate conclusion items in the first meeting conclusion are made into a questionnaire containing voting options and an open text input area, and sent to the user terminal; S42. For the received feedback data, count the number of approvals and disapprovals for each candidate conclusion item, and organize the content in the open text input area according to individual users to form structured feedback data.

7. The multi-agent conclusion generation method based on role constraints and feedback loop as described in claim 6, characterized in that: In step S5, the feedback data is processed and incorporated into the background knowledge base, specifically as follows: S51. Generate a feedback report from the structured feedback data according to a preset format. The feedback report includes statistical results and suggested text grouped by user. S52. The feedback report is added to the background knowledge base.

Images