Public figure speech influence pre-analysis method and system based on multi-agent

Abstract

This invention provides a method and system for pre-analyzing the impact of public figures' statements based on multi-agent intelligence. The method includes: first, a social background generation module retrieves background information and related social issues of the public figure in parallel, constructing a semantic environment that closely resembles reality. Then, the core multi-agent discussion module is activated, driving multiple agents to conduct multiple rounds of simulated discussions. This process is not simple text analysis, but a dynamic simulation of the clash of viewpoints, the evolution of positions, and the fermentation of public opinion in real society. Finally, the consequence analysis module outputs a structured, multi-dimensional risk report based on the simulation results. This transforms an invisible, future public opinion risk into a quantifiable and analyzable "sandbox simulation" result through technological means, achieving a fundamental shift from "passive firefighting" to "proactive fire prevention." Its technological superiority is unmatched by existing post-event negative situation analysis tools.

Description

Technical Field

[0001] This invention relates to the field of information dissemination technology, and in particular to a method and system for pre-analyzing the influence of public figures' statements based on multi-agent systems. Background Technology

[0002] Currently, the general process of intelligent question answering and content generation technology based on Retrieval-augmented Generation (RAG) is as follows: When a user inputs a query or command, the system first converts the query text into vector form and performs a similarity search in a pre-built vector index database to find the most relevant background knowledge or data fragments. Subsequently, this retrieved information, along with the original query, is input into a large language model as context, guiding the model to generate more accurate answers or content that conform to specific domain knowledge. This technology is widely used in fields such as intelligent customer service, sales script generation, and professional data querying. However, from the perspective of analyzing negative situations caused by inappropriate remarks by public figures, it is essentially a tool for information retrieval and auxiliary content generation based on known data.

[0003] In summary, while existing technologies provide powerful tools for text semantic understanding, information retrieval, and preliminary risk identification, their application in analyzing negative situations arising from inappropriate remarks by public figures largely remains at the level of "post-event monitoring" and "static analysis." These solutions generally lack a proactive and comprehensive evaluation framework that can integrate the background of the figure, simulate multiple perspectives, and predict the dynamic dissemination of public opinion, making it difficult to provide comprehensive and forward-looking risk analysis support before the remarks are published. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and system for pre-analysis of the impact of public figures' statements based on multi-agent systems. It proposes innovative solutions from four dimensions: semantic explicit context completion, adaptive system configuration, dynamic propagation simulation, and multi-agent collaborative analysis. This enables proactive early warning, comprehensive prediction, and intelligent intervention for the risks of public figures' statements. It can provide high-precision and operable risk prevention and optimization decision support before the statements are published.

[0005] In a first aspect, embodiments of the present invention provide a method for pre-analysis of the influence of public figures' statements based on multi-agent systems, the method comprising: Obtain user-inputted information about public figures' identities and target statements; By using the constructed social background generator, the identity information of the public figures and the target speech content are retrieved in stages to obtain the background information of the public figures, the cultural background of the topics, and the social public opinion environment data. The background information of the public figures, the cultural background of the topics, and the social public opinion environment data are retrieved and filtered respectively to generate a summary of the background of the figures and a summary of the social background. The personal background summary and the social background summary are integrated to generate social background data; Multiple role-based intelligent agents are constructed, and a corresponding enhancement configuration file is generated for each role-based intelligent agent based on the social background data through the constructed role enhancement configurator, which is used as the intelligent agent role capability configuration in the memory bank of each role-based intelligent agent; Each of the aforementioned role-based intelligent agents determines whether to participate in the current round of speaking. Each of the aforementioned role-based intelligent agents independently generates a speaking intention signal based on the definition of its own role's responsibilities, the semantic relevance of its area of ​​interest to the current speech content, and the novelty of its viewpoint or basis. The discussion results are obtained by conducting multiple rounds of targeted discussions for each of the aforementioned roles and agents; wherein, the discussion results include the statements, changes in positions, and interaction records of each of the aforementioned roles and agents. Based on the discussion results, the potential social impact of public figures' statements is comprehensively evaluated using a constructed consequence analyzer, resulting in consequence analysis results. The results of the consequence analysis were evaluated from seven aspects: legal consequences, necessity of apology, business impact, platform penalties, reputational damage, official supervision, and cyberbullying.

[0006] Secondly, embodiments of the present invention provide a pre-analysis system for the influence of public figures' statements based on multi-agent intelligence, the system comprising: The social background generation module is used to obtain the identity information of public figures and the target speech content input by the user; using the constructed social background generator, the identity information of public figures and the target speech content are retrieved in stages to obtain the background information of public figures and social public opinion environment data; the background information of public figures, topic cultural background and social public opinion environment data are retrieved and filtered respectively to generate a figure background summary and a social background summary; the figure background summary and the social background summary are integrated to generate social background data; The intelligent agent configuration building module is used to build multiple role intelligent agents. The built role enhancement configurator generates a corresponding enhancement configuration file for each role intelligent agent based on the social background data. This configuration file is used as the intelligent agent role capability configuration in the memory bank of each role intelligent agent. Each role intelligent agent obtains targeted search, thinking and generation capabilities from the memory bank. The multi-agent discussion module is used to determine whether each of the aforementioned role agents should participate in the current round of speaking. Each of the aforementioned role agents independently generates a speaking intention signal based on the definition of its own role's responsibilities, the semantic relevance of its area of ​​interest to the current speech content, and the novelty of its viewpoint or basis. The discussion results are obtained by conducting actual discussions through each of the aforementioned role-based intelligent agents; wherein, the discussion results include the statements, changes in positions, and interaction records of each of the aforementioned role-based intelligent agents; The consequences analysis module is used to comprehensively assess the potential social impact of public figures' statements based on the discussion results, using a constructed consequences analyzer, and to obtain the consequences analysis results. The evaluation module is used to evaluate the consequences analysis results and obtain the evaluation results.

[0007] This invention provides a method and system for pre-analyzing the impact of public figures' statements based on multi-agent intelligence, including: acquiring public figure identity information and target statement content input by a user; using a constructed social background generator to retrieve the public figure identity information and target statement content through phased information retrieval to obtain public figure background information and social public opinion environment data; retrieving and filtering the public figure background information, topic cultural background, and social public opinion environment data separately to generate a figure background summary and a social background summary; integrating the figure background summary and the social background summary to generate social background data; constructing multiple role-based intelligent agents, and using a constructed role enhancement configurator to generate corresponding enhancement configuration files for each role-based intelligent agent based on the social background data, using these as the intelligent agent role capability configurations in the memory bank of each role-based intelligent agent; and determining whether each role-based intelligent agent participates in the analysis. In the current round of speaking, each role-based agent independently generates a speaking intent signal based on its own role's responsibilities, the semantic relevance of its area of ​​interest to the current speech content, and the novelty of its viewpoints or evidence. Through actual discussion by each role-based agent, discussion results are obtained. These results include the speaking content, stance evolution, and interaction records of each role-based agent. Based on the discussion results, a constructed consequence analyzer comprehensively assesses the potential social impact of public figures' statements, yielding consequence analysis results. The consequence analysis results are then evaluated to obtain evaluation results. Innovative solutions are proposed from four dimensions: semantic explicit context completion, adaptive system configuration, dynamic propagation simulation, and multi-agent collaborative analysis. This enables proactive early warning, comprehensive prediction, and intelligent intervention for the risks of public figures' statements, providing high-precision and operable risk prevention and optimization decision support before statements are published.

[0008] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0009] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0010] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0011] Figure 1 This is a flowchart of the method for pre-analyzing the influence of public figures' statements based on multi-agent systems, provided in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of a public figure speech influence pre-analysis system based on multi-agent technology provided in Embodiment 2 of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0013] With the rapid development of information dissemination technology, public figures, due to their extensive social influence, enjoy extremely high levels of attention and enormous appeal. Their words and actions often spread rapidly and affect a very wide audience. Therefore, public figures must be cautious in their statements, otherwise they may trigger a large number of negative situations, damage their personal reputation and credibility, affect related commercial endorsements and cooperation relationships, and even potentially impact social stability.

[0014] Given this social reality, a systematic solution is needed to comprehensively and thoroughly review and evaluate the statements to be released by public figures. This application is designed based on this need, aiming to predict the potential impact of statements after their release by using intelligent analysis and public opinion simulation, taking into account multiple dimensions such as social values, public opinion trends, and personal identity. This will help public figures make more rational decisions and avoid unnecessary risks as early as possible.

[0015] While existing technologies have made significant progress in semantic understanding, information extraction, and preliminary risk identification in areas of heightened public attention following negative events, they have not yet proposed effective technologies for intelligent review and risk prediction of public figures' statements before they are published, and thus have the following objective shortcomings.

[0016] Current assessment models for negative statements by public figures are passive and outdated, lacking proactive risk prediction capabilities. Existing technological solutions, regardless of whether they employ semantic understanding, information extraction, or retrieval-enhanced generation mechanisms, essentially analyze, classify, or query related information from existing text content. They are mostly used for post-event analysis and content summarization, exhibiting a "static snapshot" risk assessment characteristic. For example, while existing systems can identify sensitive words in proposed statements or judge the sentiment of comments captured from social networks, their essence is the detection of static content. Such systems cannot predict the deep controversies and public opinion storms that a statement might trigger after being disseminated in a real public opinion environment, based on the current text content. This application constructs a proactive risk prevention and pre-event warning workflow, introducing a simulation stage of dynamic changes in public opinion from multiple parties before the statement is published. This transforms risk management from "passive response" to "proactive prediction," enabling targeted prevention of losses caused by inappropriate statements by public figures and effectively overcoming the lag of existing technologies.

[0017] The lack of a multi-party discussion environment fails to accurately reflect the complex dynamics of communication. Existing solutions typically treat the statements to be analyzed as isolated text units, focusing only on their single semantic attributes while ignoring their discussion process within social networks. However, the statements of public figures are often interpreted, amplified, and antagonized by groups with different stances in an environment where multiple groups coexist, and their impact is shaped by the combined forces of various parties. Existing social network simulation technologies mostly focus on predicting the diffusion trend of macro-level traffic. At the micro-level, they often lack the ability to simulate the refined semantic interactions between different social roles (such as authoritative commentators focusing on hot topics, business competitors, irrational fans, etc.) regarding the analysis of issues, emotional confrontations, and stance reversals based on specific topics. For example, they struggle to capture complex secondary dissemination phenomena such as "comparison and subversion," "counter-attacks," or commercial attacks, thus failing to deeply analyze the implicit information of the statements; or they may violate the law by deeply exploring the privacy of public figures. This application introduces a multi-party discussion mechanism with a social network-like structure, constructs a multi-role intelligent agent group, and simulates the "first dissemination-secondary dissemination" process to further analyze the deeper problems that may exist in the statements.

[0018] The analysis dimensions are singular and susceptible to bias, resulting in insufficient robustness of the assessment results. Most existing solutions rely on a single assessment dimension or a single large model for risk identification. This approach not only leads to insufficient coverage of assessment dimensions, making it difficult to incorporate key influencing factors such as identity consistency, group conflict, and release timing, but also amplifies the inherent biases or knowledge gaps of a single model, reducing the stability and reliability of the system. To overcome this limitation, this application proposes a consequence analyzer based on a large language model. Specifically, it organizes multiple large language models from different vendors or with different architectures (i.e., heterogeneous large models) to conduct parallel analysis from seven dimensions: legal consequences, necessity of apology, commercial impact, platform penalties, reputational damage, official supervision, and cyberbullying. A local importance scoring algorithm is introduced to effectively reduce the bias of a single model and significantly improve the comprehensiveness and accuracy of the assessment.

[0019] This application addresses the shortcomings of traditional numerical models in non-quantitative emotion propagation and social dynamics interpretation by introducing multi-agent semantic simulation and dynamic interaction modeling mechanisms.

[0020] In summary, existing technologies in the field of public figure speech analysis remain at the stage of single-point analysis and passive monitoring, lacking a systematic and interpretable comprehensive evaluation framework. This application addresses these shortcomings by proposing innovative solutions from four dimensions: semantic explicit context completion, adaptive system configuration, dynamic propagation simulation, and multi-model collaborative analysis. These solutions enable full-process prediction and intelligent intervention of speech risks, providing high-precision and actionable risk prevention and optimization decision support before speech is published.

[0021] To facilitate understanding of this embodiment, the embodiments of the present invention will be described in detail below.

[0022] Example 1: Figure 1 The flowchart is provided in Embodiment 1 of the present invention for the pre-analysis of the influence of public figures’ speech based on multi-agent systems.

[0023] Reference Figure 1 The method includes the following steps: Step S101: Obtain the public figure's identity information and the target's statement content input by the user; In step S102, the social background generator in the system obtains the background information of public figures and social public opinion environment data by retrieving public figures' identity information and target speech content in stages. Step S103: The social background generator in the system retrieves and filters the background information of public figures and the social public opinion environment data respectively, and generates a summary of the figure's background and a summary of the social background. Step S104: The social background generator in the system integrates the person's background summary and the social background summary to generate social background data. Step S105: Construct multiple role-based intelligent agents. The role enhancement configurator generates a corresponding enhancement configuration file for each role-based intelligent agent based on social background data. This configuration file is used as the intelligent agent's role capability configuration in the memory bank of each role-based intelligent agent. In other words, it is used as the intelligent agent's memory bank. Each role-based intelligent agent obtains targeted search and thinking capabilities from the memory bank. Step S106: Each role agent determines whether to participate in the current round of speaking. Each role agent independently generates a speaking intention signal based on its own role's responsibility definition, the semantic relevance of its area of ​​interest to the current speech content, and the novelty of its viewpoint or basis. Step S107: Each role agent conducts an actual discussion to obtain the discussion results; wherein, the discussion results include the speech content, position evolution and interaction records of each role agent; Step S108: Based on the discussion results, the potential social impact of public figures' statements is comprehensively evaluated using the constructed consequence analyzer to obtain the consequence analysis results. Step S109: The consequence analyzer in the system evaluates the consequence analysis results and obtains the evaluation results.

[0024] Furthermore, step S103 includes the following steps: Step S201: Perform a background search on the public figure to obtain the first preliminary search results; the background search is conducted using web retrieval. Step S202: Perform title relevance filtering on the first preliminary search results to obtain the filtered first search results; Step S203: Input the filtered first search result into the social background generator to generate a person's background summary; wherein, the person's background summary includes professional background, social role, public image and historical events; Step S204: Perform a public opinion environment retrieval on the social environment data to obtain a second preliminary retrieval result; wherein, the social environment retrieval adopts a web retrieval method. Step S205: Filter the second preliminary search results to obtain the filtered second search results; Step S206: Input the filtered second search results into the social background generator to generate a social background summary; wherein, the social background summary includes social issues, public attitudes, media focus and public opinion environment.

[0025] Furthermore, step S201 includes the following steps: Step S301: Call the social background generator to analyze the background information of public figures and identify the names of the figures and related people; Step S302: Use the person's name as the search object; Step S303: Construct multiple types of search queries based on the search object; Step S304: Perform a search through the network search interface to obtain the search results returned by each query, and use them as the first preliminary search results.

[0026] Furthermore, step S204 includes the following steps: Step S401: Input social public opinion environment data into the social background generator and identify the value analysis results; wherein, the value analysis results include value orientation, attitude stance and viewpoint tendency; Step S402: Based on the identity information of public figures, the social background generator infers the motivation for the remarks based on the value analysis results and judges the intention of the public figures in making the remarks. Step S403: After completing the motivation deduction, analyze the speech content through the social background generator and extract key elements; among which, key elements include the people involved, social circles, core themes and keywords; Step S404: Construct a combined query based on key elements; Step S405: Perform a network search for each query to obtain the search results returned by each query, and use them as the second preliminary search results.

[0027] Specifically, in the first stage, the system executes the social context generation module. This module is used to construct the semantic context of the analysis, providing data support for subsequent discussions.

[0028] The system uses a social background generator to receive user input of public figure identity information and target speech content. Through two parallel retrieval subtasks: figure background retrieval and speech social environment retrieval, it obtains background information of public figures (i.e., public figure profiles) and social opinion environment data.

[0029] In the subtask of background retrieval, the system's social background generator first extracts accurate names of individuals from the input information. The social background generator then uses a large language model to analyze the input text, identifying public figures' names and related individuals. The system uses these extracted names as search targets. The social background generator subsequently constructs three types of search queries. The first type combines a complete description with a "personal profile," the second type combines a complete description with "professional background," and the third type combines a complete description with "controversial events." The complete description retains the professional information and name combinations from the original input, improving search accuracy.

[0030] The social context generator performs searches via a web search interface, returning three results for each query. After obtaining the initial search results, the social context generator performs title relevance filtering. Based on the occurrence of a person's name in the title, the social context generator filters out results irrelevant to the target person.

[0031] The social context generator then inputs the filtered search results into the large language model to generate a personal background summary that includes professional background, social role, public image, and historical events.

[0032] In the subtask of retrieving the social context of speech, the social context generator first constructs a thought chain analysis to understand the deep semantics of the speech. The social context generator inputs the speech content into a large language model, which first analyzes the core values ​​conveyed by the speech, identifying the value orientation, attitude, and viewpoint that the public figure is attempting to convey. The system then combines this with the public figure's identity information, and the model infers the motivation behind the speech based on the value analysis results, determining whether the public figure's purpose in making the speech is commercial promotion, social criticism, personal emotional expression, or other types of motivation. After completing the motivation deduction, the system extracts the core elements of the speech.

[0033] The social context generator analyzes speech content using a large language model, extracting four key elements from public figures' statements: involved figures, social circles, core themes, and keywords. The key elements include involved figures, social circles, the core theme of the statement, and keywords. Keywords refer to semantic identifiers that, in addition to the above three elements, can significantly narrow the search scope or point to a specific context, specifically including: euphemisms for specific events, industry-specific terms, popular internet slang, and time-sensitive descriptions.

[0034] The social background generator constructs combined queries based on the four key elements extracted, including combining a person's name with "controversial events", combining a person's name with "social impact", combining social circles with "hot topics", and combining core themes with "social discussions", generating up to 10 queries.

[0035] The social context generator performs a web search for each query, returning two results per query. After obtaining the initial search results, the social context generator performs relevance filtering.

[0036] The social context generator calls upon a large language model, retaining results highly relevant to the topic of the discourse. Finally, the social context generator inputs the filtered search results into the large language model to generate a social context summary containing social issues, public attitudes, media focus, and the public opinion environment.

[0037] The social context generator employs a thread pool mechanism to execute the two retrieval subtasks in parallel, significantly reducing the overall retrieval time. After integrating the outputs of the two subtasks using a large language model, the social context generator produces complete social context data, including character background summaries, social environment summaries, and structured context data files. The output of this stage will directly serve as the input data for the next stage, providing the basic semantic environment for agent behavior modeling.

[0038] Furthermore, step S105 includes the following steps: Step S501: Parse the strings into the preset interest areas of each role's intelligent agent, and decompose the strings into a keyword list; Step S502: Construct filter prompts; Step S503: Based on the screening criteria, input the role name of the role agent, the target speech content, and the keyword list into the large language model for judgment to obtain the screening results; wherein, the screening results include the screened keywords; Step S504: Input the role name, target speech content, and selected keywords of the role agent into the role enhancement configurator to generate multiple supplementary keywords; Step S505: After merging and deduplicating the selected keywords and multiple supplementary keywords, a set of search keywords is obtained; Step S506: Construct a search query for each keyword in the set of search keywords, and obtain multiple search results returned for each keyword through the web search interface; Step S507: Input the search results into the character enhancement configurator and output structured feature data; wherein, the structured feature data includes a list of focus areas, motivation types, typical positions, descriptions of expression styles, evidence type preferences, and a list of target audiences; Step S508: The structured feature data is integrated into the basic configuration of each role agent, the professional focus areas, behavioral styles and argument preferences of each role agent are updated, and an enhanced configuration file is generated and used as the agent's memory. Each role agent obtains targeted search and thinking capabilities from the memory.

[0039] Specifically, in the second phase, the system uses a role enhancement configurator to execute the agent configuration building module. The main task of this module is to generate personalized enhancement profiles for each role agent based on social context data. The role enhancement configurator in the system constructs 11 role agents, including 9 persuasive roles and 2 fixed-position roles. The 9 persuasive roles are government regulators, legal professionals, media workers, industry commentators, business competitors, industry peers, brokerage spokespeople, brand spokespeople, and rational supporters; these roles may change their positions under sufficient argumentation. The 2 fixed-position roles are fervent supporters and strong detractors; these roles hold fixed views, are not persuasive, and can only choose to continue speaking or remain silent. The role enhancement configurator reads the social context information generated in the previous phase and generates corresponding enhancement configuration data for each role agent, using it as the agent's memory.

[0040] First, the character enhancement configurator filters keywords related to the current statement from the character's preset areas of interest. The configurator parses the string from the character's preset areas of interest, supporting various delimiters, and breaks the string down into a list of keywords.

[0041] The role enhancement configurator then constructs filtering suggestions by inputting the role name, target speech content, and keyword list into the large language model. The model judges based on four filtering criteria: whether the keywords are directly related to the topic of the speech content, whether they are related to the points of controversy that the speech may cause, whether they are related to the role's responsibilities in such events, and whether they can generate valuable search results. The model selects 3 to 5 of the most relevant keywords from the preset keywords.

[0042] The character enhancement configurator determines the sufficiency of the filtered results. If fewer than three keywords are found, the configurator marks the need to generate supplementary keywords. When supplementary keywords are required, the configurator generates supplementary suggestion words.

[0043] The role enhancement configurator inputs the role name, target statement, and selected keywords into a large language model. The model analyzes the core issues of the statement and the typical concerns of the role in similar events, generating 3 to 4 supplementary keywords. Specifically, the model generates keywords for 11 agents primarily from two dimensions: "technical terminology mapping" and "potential risk association." For example, when the agent's role is "legal practitioner," and the target statement involves "commercial product promotion," even if the statement itself does not mention legal terms, the model will automatically generate technical terms such as "false advertising," "advertising law compliance," and "unfair competition" as supplementary keywords to compensate for the lack of professional depth in the initial keywords. The system merges the selected keywords and supplementary keywords, removes duplicates, and forms the final set of search keywords, limiting the total number to no more than 6.

[0044] The character enhancement configurator then performs a web search using the generated keywords. For each keyword, the configurator constructs a search query, retrieves relevant information via a web search interface, and returns three search results for each keyword.

[0045] After aggregating all search results, the Role Enhancement Configurator performs information extraction and feature analysis. The Configurator inputs the search results into a large language model, which analyzes the behavioral characteristics of the role group from six dimensions: focus of attention (identifying issues typically addressed by this group), motivation for speaking (analyzing primary motivations in similar events), stance (judging the distribution of attitudes towards support, opposition, neutrality, or warning), language style (extracting commonly used expressions), evidence sources (identifying preferred types of evidence cited), and target audience (determining the primary groups to whom the voice is addressed). The model outputs structured feature data. To clarify the behavioral logic of each role in a specific public opinion field, this structured feature data is specifically defined as a "role behavior profile." This profile includes: a list of the role's focus of attention regarding the target discourse (e.g., "legal compliance" or "moral bottom line"), the type of deep motivation for participating in the discussion, typical stances in similar events, a personified description of the expression style, preferred types of evidence cited, and a list of expected target audiences.

[0046] The role enhancement configurator finally integrates the extracted structured feature data into the role's basic configuration, updating parameters such as the role's area of ​​interest, behavioral style, and argument preferences, generating a complete enhancement configuration file for each agent. Subsequently, the role enhancement configurator supplements each agent's configuration file with different cognitive reasoning sequences. This sequence defines the state transition logic of the agent instance before generating response data. Based on a preset prior logic template, the role enhancement configurator solidifies a multi-stage thought chain path in the configuration file. This path forces the agent instance to execute preset logical reasoning steps sequentially after receiving input information. These logical operation steps serve as pre-constraints for the large language model to generate text, excluding unstructured random generation patterns. The role enhancement configurator employs a thread pool parallel mechanism to process the configuration building tasks of multiple roles simultaneously, with a maximum concurrency of 4, significantly improving construction efficiency. After completing the configuration building of 11 roles, the system stores the enhancement configuration files in the system. The role enhancement configurator initializes an independent memory bank for each agent instance. This memory bank is divided into a persistent storage partition and a runtime working memory partition on the physical storage medium. The persistent storage partition is used to store the enhanced configuration file generated above, and the runtime working memory partition is used to dynamically record context state data during multiple rounds of interaction. The independent memory bank ensures state isolation and data consistency for different agent instances running in parallel.

[0047] Furthermore, step S108 includes the following steps: Step S601: Extract key information from the discussion results; the key information includes the target statement content, public figure profile, multiple rounds of discussion process, and the final views of each role. Step S602: Construct a summary of public opinion opinions and extract the positions and suggestions of each agent according to their priority.

[0048] Specifically, role priority is allocated based on a dual standard of "social impact and consequence relevance." The system pre-sets a three-tiered priority list: the first tier is "core stakeholders," including government regulators, legal professionals, brand spokespeople, and agency spokespeople. These roles have the ability to directly trigger administrative penalties, legal proceedings, or business terminations, and their statements carry the highest weight; the system must retain all of their warning suggestions. The second tier is "public opinion megaphones," including media workers and industry commentators. These roles determine the scope and characterization of negative situations, and the system primarily extracts their core arguments. The third tier is "representatives of group sentiment," including fervent supporters and strong detractors. These roles primarily reflect the intensity of emotions; the system only performs clustered summaries and does not record each item individually. This hierarchical extraction mechanism ensures that the generated opinion summaries focus on substantive risks, rather than being overwhelmed by a massive amount of invalid emotional noise. Step S603 involves inputting the target speech content, public figure information, and summary of public opinion into a consequence analyzer based on multiple large language models from different vendors or with different architectures for analysis across seven dimensions. To ensure the comprehensiveness of the evaluation and eliminate algorithmic bias from a single model, the system simultaneously invokes multiple heterogeneous large language models with different architectures to perform parallel analysis of the discussion content from seven dimensions. Each model outputs its own analytical viewpoint, constructing an original set of viewpoints.

[0049] Step S604: Perform opinion merging on the opinions in the original opinion set. First, the system uses a pre-trained embedding model to transform all initial opinions into high-dimensional vectors. This is done by calculating the opinions... Opinions The cosine similarity between them quantifies their degree of overlap in the semantic space:

[0050] in Viewpoints and The corresponding embedding vector. When the similarity exceeds a set threshold (default value is 0.65), it indicates that the two viewpoints are semantically highly similar. Subsequently, the system introduces... Solid ( Semantic stability score is used to assess semantic clarity.

[0051] Opinion t of Solid The calculation of the score includes neighborhood proximity. and neighborhood cohesion Two sub-dimensions, specifically defined as:

[0052] in, The weight is the neighborhood proximity. The lower the value, the more likely the viewpoint is to be correct. The more vague the semantic definition of an item (i.e., its natural language expression is too broad and lacks specific facts and risk descriptions), the more likely it is to be replaced or merged during the merger process.

[0053] Neighborhood proximity and neighborhood cohesion The calculations are as follows: Neighborhood proximity :

[0054] in, For the purpose of viewpoint t The k The most similar neighbor, This represents the number of viewpoints selected by the system that have the highest semantic similarity to the target viewpoint. Quantify the target perspective With Top- nearest neighbors , default The average similarity reflects the positional stability of an opinion in the semantic field (i.e., the degree to which the opinion is closely surrounded by neighboring expressions in the semantic space; the higher the positional stability, the more the opinion belongs to the core consensus of the current discussion dimension). Neighborhood cohesion :

[0055] in, For the purpose of viewpoint t The i The and the first j The most similar neighbor, Meaning and proximity of the neighborhood K The similarity between these neighboring nodes is quantified to reflect the tightness of the semantic cluster to which the label belongs.

[0056] The system combines semantic similarity with Solid Fractions, calculating the scores of any two viewpoints. Merger priority index :

[0057] When two viewpoints are extremely similar in meaning ( Similarity (High) and its own definition is unclear ( Solid When the score is low, The value is too large. This means that these two viewpoints are very likely duplicate descriptions of the same vague concept. If so... For each pair of tags, the large language model is invoked to merge these pairs of viewpoints into a single viewpoint. After the deduplication and dimensionality reduction merging processes described above, the core viewpoints that are ultimately retained will be standardized by the system into highly condensed phrases, thus transforming them into "prediction entries" for subsequent modules. Specifically, a viewpoint is a concrete natural language description output by the multi-agent system, while a prediction entry is a structured evaluation benchmark for the viewpoint after refinement; the two represent a mapping relationship between source data and structured features.

[0058] Step S605: The analysis results of the seven dimensions are comprehensively evaluated to obtain the consequences analysis results.

[0059] Furthermore, step S605 includes the following steps: Step S701: Input the analysis results of the seven dimensions into the consequence analyzer for extraction, obtain the two most important consequences, and use them as the core conclusions. Step S702: Standardize the core conclusions to obtain standardized conclusions; Step S703: Based on the quantity and severity of the analysis results of each dimension, the overall risk is divided into multiple severity levels, and a risk assessment report is generated. The risk assessment report includes: detailed consequences, main causes, severity levels, and response recommendations for each dimension.

[0060] Specifically, the third stage is the core component of the system: the multi-agent discussion module. This module can reproduce the reactions and evaluations of target statements by different roles from different perspectives within a virtual social environment, thereby enabling a comprehensive and in-depth assessment of the target statements.

[0061] First, each agent autonomously determines whether to participate in the current round of speaking. Each agent independently generates a speaking intention signal based on its own role's responsibilities, the semantic relevance of its area of ​​interest to the current speech content, and the novelty of its viewpoint or basis.

[0062] If the semantic relevance is judged as "relevant" and the novelty as "new," a signal indicating the agent's intention to "speak" is given; otherwise, a signal indicating "silence" is given. When the agent receives the signal indicating the intention to "speak," the system activates its speaking process; conversely, if the agent believes its own position or information reserves are insufficient to form a valid viewpoint, it will proactively choose silence and generate explanatory text. In situations where fervent supporters or strong critics encounter strong rebuttals or unbalanced viewpoints, a forced silence mechanism may still be triggered to maintain the rationality and balance of the discussion.

[0063] In actual discussions, the system defaults to three rounds, allowing users to adjust the number of rounds through parameter configuration. The first round is the initial stance expression phase, where each role independently forms an opinion or remains silent and provides arguments based on their own knowledge and configuration. The second and subsequent rounds are interactive discussion phases, where each role develops a response strategy after reading others' opinions, including supplementing, refuting, or questioning. Persuasive roles may change their stance under sufficient argumentation, while those with fixed stances, including fervent supporters and strong detractors, maintain their original views or choose to withdraw from the discussion. The system simultaneously monitors the magnitude of stance changes in each round of discussion, ending the discussion early when opinions tend to stabilize.

[0064] The discussion results record the content of each participant's speech, the evolution of their stance, and their interactions. Furthermore, the system has a built-in concurrency optimization mechanism, with a maximum concurrency of 6, and supports dual API channels (primary and backup) to ensure load balancing and stable operation.

[0065] Furthermore, step S109 includes the following steps: Step S801: Obtain the structured dataset; Step S802: Extract the list of main consequences from the consequence analysis results; Step S803: Read the manually labeled cases from the structured dataset and use them as the evaluation benchmark; Step S804: Input the list of main consequences and manually labeled tags into the embedding model to obtain a high-dimensional vector; Step S805: Calculate the cosine similarity matrix between the main consequences list and the manually labeled tags, and construct the similarity mapping relationship between the predicted items and the real tags; Step S806: Based on each predicted entry, identify all matching tags whose similarity exceeds a set threshold to obtain the matching results; Step S807: Calculate the evaluation result based on the matching result.

[0066] Furthermore, step S807 includes the following steps: Step S901: Divide the number of successfully matched predicted entries by the total number of predicted entries to obtain the entry hit rate; Step S902: Collect the similarity scores of all matching predicted entries for each label, and calculate the average similarity of each label; Step S903: Calculate the average similarity of all tags to obtain the average tag similarity. Step S904: Find the maximum similarity with all predicted entries for each label, and average the maximum similarity values ​​after keeping only the maximum similarity values ​​that exceed the threshold to obtain the mean maximum similarity value.

[0067] Specifically, after the multi-agent discussion is completed, the system enters the fourth stage—the consequence analysis module. This module consists of a consequence analyzer, responsible for comprehensively assessing the potential social impact of public figures' statements based on the discussion results. The consequence analyzer in the system first extracts key information from the discussion results file, including the content of the target statement, the identity of the public figure, the multiple rounds of discussion, and the final views of each role. The system then constructs a summary of public opinion, extracting the positions and suggestions of each role according to role priority.

[0068] The consequences analyzer employs a parallel analysis mechanism to simultaneously evaluate consequences across seven dimensions. For each dimension, the analyzer constructs specific analytical prompts, inputting the target statement content, public figure information, and a summary of public opinion into a heterogeneous large-scale model. This model then performs independent reasoning based on the provided information. In the legal consequences analysis dimension, the model assesses the potential legal risks triggered by the statement, including seven categories of legal issues: defamation, insult, incitement of hatred, dissemination of false information, infringement of reputation rights, and violation of advertising laws. The model references professional opinions from legal practitioners and government regulators during discussions to determine whether the statement involves significant legal issues and outputs a specific list of legal consequences. In the apology necessity analysis dimension, the model aims to predict whether the potential negative backlash after the statement's publication warrants a public apology. Specifically, this module does not directly generate an apology letter. Instead, it uses a three-layer decision tree analysis based on the "intensity of social reaction" and the "degree of damage to public order and good morals": The first layer determines whether the statement constitutes substantial harm to a specific group; the second layer, combined with the aforementioned "summary of public opinion," assesses whether not apologizing will lead to further deterioration of the negative situation (such as triggering secondary boycotts); the third layer, based on common sense, determines whether the apology is feasible for repairing the image. The assessment results (including whether an apology is necessary, a clarification is recommended, or no action is needed, and their degree of urgency) will serve as decision-making support information, prompting users to prepare corresponding crisis response plans before publishing their statements, or indirectly encouraging users to modify their original statements to avoid this risk.

[0069] In the business impact analysis dimension, the model assesses the impact of statements on the commercial interests of public figures. The model analyzes five categories of business consequences: endorsement contract risks, loss of brand partnerships, decline in commercial value, changes in market recognition, and impact on investor confidence. The model pays particular attention to the views of spokespeople from brokerage firms, brands, and business competitors in discussions to determine whether statements affect business partnerships. In the platform penalty analysis dimension, the model assesses the management measures that social media platforms may take. The model analyzes five categories of platform penalties: account throttling risks, potential content deletion, account bans, trending topic blocking, and platform warnings. The model infers the likelihood of platform intervention based on the degree of controversy surrounding the content and the intensity of public reaction. In the reputation damage analysis dimension, the model assesses the impact of statements on the social image of public figures. The model analyzes five categories of reputation consequences: decreased public trust, damage to professional image, changes in social evaluation, fragmentation of fan groups, and impact on industry status. The model synthesizes the attitude distribution among various viewpoints to determine the degree and scope of reputation damage. In the official regulatory analysis dimension, the model assesses the regulatory measures that government departments may take. The model analyzes five types of official consequences: regulatory interviews, administrative warnings and penalties, industry rectification requirements, public criticism and notifications, and related policy adjustments. The model places particular emphasis on regulatory suggestions made by government officials during discussions to assess the likelihood of regulatory intervention. In the dimension of cyberbullying risk analysis, the model assesses the potential for cyberbullying arising from online speech. The model analyzes five types of cyberbullying risks: risks associated with searching personal privacy information, the intensity of online abuse, the potential for privacy leaks, the risk of offline harassment, and the impact on mental health. Based on the extreme viewpoints of strong detractors and fervent supporters in discussions, the model assesses the intensity and duration of cyberbullying.

[0070] To ensure the professionalism and depth of the analysis, the consequence analyzer described in this application does not employ a single, universal reasoning logic in its implementation. Instead, it constructs independent analysis operators for the aforementioned seven analytical dimensions. For example, when executing the "legal consequences" operator, the system loads binding prompts containing specific legal provisions (such as the Advertising Law of the People's Republic of China and the elements of the crime of defamation), forcing the model into a "rigorous reasoning mode" to focus on assessing the compliance of the behavior. When executing the "reputation damage" operator, the system loads contextual prompts containing rules for social network sentiment analysis, guiding the model into an "empathy and understanding mode" to focus on assessing the audience's emotional response. This logical heterogeneity achieved through the injection of domain-specific knowledge effectively avoids the knowledge blind spots of a single model when dealing with cross-domain problems.

[0071] Building upon this foundation, to further enhance the reliability of the evaluation results, the consequence analyzer also incorporates a cross-validation process. The consequence analyzer is configured with a multi-model scheduling interface, enabling parallel invocation of heterogeneous large models for analysis tasks of the same dimension. The consequence analyzer distributes the same set of analysis prompts and discussion results data, constructed above, to various heterogeneous models, obtaining multiple independent inference conclusions. These conclusions are then generated through voting or weighted fusion to produce the final consequence description for that dimension. This mechanism leverages the complementary strengths of different models, effectively avoiding algorithmic biases that may exist with a single model. When comprehensively evaluating the analysis results of multiple dimensions generated through the above validation, the system introduces a local importance scoring algorithm to dynamically quantify risk. Based on the hot keywords in the "social background data" generated in the preceding steps, the system dynamically adjusts the weight coefficients of each dimension. The final overall risk level (Mild / Moderate / Severe / Extremely Severe) is not a simple arithmetic average of the severity levels, but rather based on a weighted formula. The calculation yielded the following result. As for the weighting of the legal consequences analysis dimension, The weighting of the dimensions for analyzing the necessity of an apology. As the weighting of the business impact analysis dimension, As the weight of the platform's penalty analysis dimensions, As the weights for the dimensions of reputation damage analysis, As the weighting of official regulatory analysis dimensions, As the weight of the dimensions in the analysis of cyberbullying risk, and ; Risk levels are analyzed from the perspective of legal consequences. To analyze the risk level of the necessity of an apology, Risk levels for business impact analysis dimensions The risk level is determined by the platform's penalty analysis dimensions. For the risk level of reputation damage analysis dimension, Risk levels are determined by official regulatory analysis. The risk level is defined as a dimension of cyberbullying risk analysis, and The range of values ​​is within the enumerated values. The risk levels are 1 for mild, 2 for moderate, 3 for severe, and 4 for extremely severe. Therefore, the overall risk level is... The range of values ​​is , It is mild. Medium. It is serious.

[0072] The system employs a thread pool mechanism to execute the analysis tasks across all seven dimensions simultaneously, improving analysis efficiency. After collecting the analysis results from all seven dimensions, the system performs a comprehensive evaluation. The system inputs the consequence descriptions from all dimensions into a large language model, which extracts the two most significant consequences from each dimension's results as the core conclusions.

[0073] The system then standardizes the extracted core conclusions, that is, summarizes them into highly condensed phrases, thereby simplifying the difficulty of system evaluation.

[0074] Finally, the system estimates the severity level. Based on the number and severity of consequences in each dimension, the model classifies the overall risk into four levels: mild, moderate, severe, or very severe. The system then generates a risk assessment report, which includes detailed consequences, main causes, severity levels, and response recommendations for each dimension.

[0075] The system outputs the analysis results in a structured document containing complete information such as consequence predictions, scope of impact, severity classification, and response recommendations. This module facilitates the transformation from simulation discussion results to inferences about social consequences, providing decision-making references for policymakers, businesses, and media organizations.

[0076] To measure and evaluate the system's performance, this application designs a targeted evaluation module. The main function of this module is to verify the reliability and accuracy of the system's prediction results.

[0077] The system uses a structured dataset for evaluation, organized in CSV format. Each row of the structured dataset represents a real-world example of a public figure's statement, containing three core fields. The first field is the public figure identifier, recording the public figure's identity information, typically a combination of their title and name. The second field is the statement content, recording the complete original text of the statement. The third field is the manually labeled tag field, containing 10 tag columns, numbered from 1 to 10. Each tag records the actual consequences or social reactions of the statement in the real-world context, such as "gender discrimination," "user churn," "defamation," or "misleading teenagers." The number of tags varies depending on the complexity of the case; simple cases may have only 1 or 2 tags, while complex cases may have up to 10 tags. Unused tag columns remain empty.

[0078] The system extracts a list of main consequences from the consequence analysis results file and reads the manually labeled cases from the structured dataset as an evaluation benchmark. The system uses a semantic similarity matching method for evaluation. First, the system inputs the predicted list of main consequences and the manually labeled cases into a text embedding model to generate high-dimensional vector representations.

[0079] The system then inputs the predicted main consequences and manually labeled tags into the embedding model, generating a 1024-dimensional vector representation. The system then calculates the cosine similarity matrix between the prediction results and the labels to construct the prediction entries. With real labels The similarity mapping relationship. The calculation formula is: In the evaluation system of this application, in order to address the inconsistency in text length between "the detailed consequences predicted by the system" and "the short labels annotated by humans", the system first inputs these two types of text into the same embedding model. Through the mapping function of the model, they are uniformly transformed into texts with completely consistent dimensions. and , and The vector is normalized.

[0080] Based on this unified vector shape, the system can apply the cosine similarity formula. The calculation method is located on the left side of the equation: the numerator in the formula. The dot product of vectors measures the degree of overlap between them in a semantic direction; the denominator contains... and The L2 norm (modulus) of the vector is used to normalize the vector, limiting the calculated result to between 0 and 1. The final calculated value is... This is the similarity score between the two in the semantic space, which the system uses to determine whether the prediction result has successfully hit the real label.

[0081] For each predicted entry, the system identifies all matching tags with a similarity exceeding a set threshold, which is 0.75 by default. The system considers predicted entries with similarity reaching the threshold as successful matches and counts the number of matched entries. Based on the matching results, the system calculates three core evaluation metrics. The first metric is the entry hit rate, calculated by dividing the number of successfully matched predicted entries by the total number of predicted entries, reflecting the accuracy of the system's output. The second metric is the average tag similarity. The system collects the similarity scores of all matching predicted entries for each tag, calculates the average similarity for each tag, and then averages the average similarities of all tags, reflecting the semantic closeness between the predicted results and the true tags. The third metric is the maximum similarity mean. The system finds the maximum similarity between each tag and all predicted entries, retains only the maximum similarity exceeding the threshold, and averages them, reflecting the system's optimal predictive ability for each true tag.

[0082] The system writes the evaluation results into a structured file. This file records the case name, prediction item details, matching tag information, similarity score, and the values ​​of three core indicators. Simultaneously, the system updates the case processing status file, recording the processing path and annotation tags for each case. This evaluation mechanism enables automatic verification of the prediction results, providing a quantitative basis for subsequent model optimization and ensuring the reliability and accuracy of the system's predictions.

[0083] This system constructs a complete closed-loop process for analyzing the social impact of public figures' statements, comprising five core stages: public figure profile description, social background generation, role configuration construction, multi-agent discussion, consequence analysis, and effect evaluation. In the social background generation stage, the system uses a social background generator to acquire information about the figure's background and social environment through a parallel retrieval mechanism, providing a semantic context for subsequent analysis.

[0084] During the role configuration construction phase, the system uses a role enhancement configurator to generate enhanced configurations for 11 specialized role agents, and achieves accurate modeling of role behavior characteristics through intelligent keyword filtering and network retrieval.

[0085] The system simulates a real-world public opinion environment during the multi-agent discussion phase. Nine persuasive roles and two roles with fixed stances engage in multiple rounds of interactive discussions based on their respective professional perspectives, leading to the evolution of positions and the clash of viewpoints. In the consequence analysis phase, the system uses a consequence analyzer and a seven-dimensional parallel evaluation mechanism to predict the potential social impact of statements across seven levels: legal, commercial, reputational, platform, regulatory, apology, and cyberbullying. In the effect evaluation phase, the system employs a semantic similarity matching method, quantifying the prediction accuracy through three core indicators: item hit rate, average tag similarity, and the mean of maximum similarity.

[0086] This system automates the entire process of analyzing the impact of public figures' statements, from initial input to multi-dimensional consequence predictions, with each stage forming a closed-loop technology. Through the multi-level reasoning capabilities of a large language model, a multi-agent collaborative analysis mechanism, and a structured evaluation system, the system ensures the scientific rigor, objectivity, and verifiability of the analysis results. This system provides policymakers with a tool for early warning of speech risks and media platforms with a reference for content review, demonstrating broad application value and promising prospects for technology promotion.

[0087] This application transforms the traditional, reactive, and passive post-event analysis model of negative situations into a proactive, pre-event risk simulation and early warning model, providing a crucial "golden window" for public figures' decision-making. Existing technologies, such as sentiment analysis or keyword monitoring tools, are essentially post-event analysis tools, only able to monitor and summarize after statements are published and public opinion is triggered. By this time, public figures are often already in a passive position, missing the best opportunity for intervention. This invention solves this fundamental deficiency through its unique technical solution: public figures can use their statements as input before publishing important statements. The system first uses a social background generation module to retrieve the public figure's background information and related social issues in parallel, deducing and constructing a semantic environment close to reality. Subsequently, the core multi-agent discussion module is activated, driving 11 agents, including "business competitors," "brand spokespersons," and "industry commentators," to conduct multiple rounds of simulated discussions. This process is not a simple text analysis, but a dynamic simulation of the clash of viewpoints, the evolution of positions, and the fermentation of public opinion in real society. Finally, the consequence analysis module outputs a structured, multi-dimensional risk report based on the simulation results. Therefore, this invention transforms an invisible, future public opinion risk into a quantifiable and analyzable "sand table simulation" result through technical means, realizing a fundamental shift from "passive firefighting" to "proactive fire prevention." Its technical superiority is unmatched by existing post-event analysis tools.

[0088] This application upgrades crisis planning from vague, experience-based planning to precise, actionable strategies driven by data.

[0089] Existing technologies, when providing risk warnings, often only offer general conclusions such as "there is a risk of negative consequences," failing to support public figures in developing specific response strategies. This application overcomes this deficiency through its refined role modeling and multi-dimensional consequence analysis. In the multi-agent discussion module, each agent's behavior is driven by its configuration file, enhanced by the agent configuration building module. Serving as the agent's memory, its speech content, stance, and emotions highly simulate the reactions of real-world social roles. For example, public figures can accurately predict potential rifts in partnerships or potential legal risks by observing the strong language of a "brand spokesperson" or the specific regulations cited by a "legal professional" in the simulated discussion. Furthermore, the consequence analysis module provides seven-dimensional analysis reports, such as "Business Impact Analysis," which quantifies the risks of endorsement contracts, and "Reputation Damage Analysis," which reveals the specific impact on the target user group. Therefore, public figures can develop highly targeted crisis plans based on this detailed and supported report. This transforms crisis plans from generic templates into "precision-guided" solutions tailored to the specific risks of their statements, significantly improving the effectiveness and feasibility of the plans.

[0090] Example 2: Figure 2 This is a schematic diagram of a multi-agent-based public figure speech impact pre-analysis system provided in Embodiment 2 of the present invention. This application constructs a social background modeling module, a multi-agent discussion network, and a consequence analysis module, further realizing the dynamic impact analysis of public figures' speech in the social public opinion environment. The overall system operation process consists of four main stages: social background generation, agent configuration construction, multi-agent discussion simulation, and consequence analysis.

[0091] Reference Figure 2 The system includes: The social background generation module is used to obtain the identity information of public figures and the target speech content input by the user; using the constructed social background generator, the identity information of public figures and the target speech content are retrieved in stages to obtain the background information of public figures and social public opinion environment data; the background information of public figures and social public opinion environment data are retrieved and filtered separately to generate a figure background summary and a social background summary; the figure background summary and the social background summary are integrated to generate social background data. The agent configuration building module is used to build multiple role agents. The built role enhancement configurator generates a corresponding enhancement configuration file for each role agent based on social background data, and uses it as the agent role capability configuration in the memory bank of each role agent. The multi-agent discussion module is used to determine whether each role agent should participate in the current round of speaking. Each role agent independently generates a speaking intention signal based on its own role's responsibilities, the semantic relevance of its area of ​​interest to the current speech content, and the novelty of its viewpoint or basis. The actual discussion is conducted through each role agent to obtain the discussion results. The discussion results include the speaking content, position evolution, and interaction records of each role agent. The consequences analysis module is used to comprehensively assess the potential social impact of public figures' statements based on the discussion results, using a constructed consequences analyzer, and obtain the consequences analysis results. The evaluation module is used to assess the consequences analysis results and obtain the evaluation results.

[0092] Compared with the prior art, this application has the following advantages: 1) Proactive risk simulation and early warning workflow: This application addresses the limitations of existing technologies in risk assessment, which are characterized by "static snapshots" and "passive lag," by constructing a proactive risk prevention and early warning workflow. Unlike traditional technologies that retrospectively summarize published texts, this system introduces a multi-party public opinion simulation phase before any statement is published. By dynamically simulating the spread and fermentation process in a virtual social environment, it proactively predicts potential points of contention and public opinion storms. This design shifts the core logic of risk management from "passive response" to "proactive prediction," effectively overcoming the lag in existing technologies when dealing with the non-linear spread and emotional diffusion of public opinion, achieving true risk avoidance and intelligent intervention.

[0093] 2) Multi-agent dynamic game and deep propagation simulation: To address the shortcomings of existing solutions that treat speech as isolated text units and to overcome the limitations of capturing its complex propagation dynamics in real social networks, this application introduces a multi-party discussion mechanism with a social network-like structure. The system constructs a group of intelligent agents comprising diverse roles such as supporters, detractors, business partners, and regulatory bodies. By simulating their interactive game and emotional feedback, the dynamic process of speech propagation from "initial dissemination" to "secondary dissemination" is realistically reproduced. This mechanism can effectively capture complex secondary dissemination phenomena such as "counter-attacks" or commercial attacks, thereby enabling in-depth analysis of the implicit information and deeper issues of speech that cannot be reached by single-text semantic analysis.

[0094] 3) Multidimensional collaborative evaluation and robustness enhancement mechanism: To address the shortcomings of existing technologies that rely on a single assessment dimension or a single large model, leading to insufficient coverage and susceptibility to bias in assessment results, this application proposes a multi-model collaborative large-scale model analysis matrix mechanism. The system organizes multiple heterogeneous large-scale language models to analyze target speech in parallel from seven key dimensions, including legal, commercial, reputational, and regulatory perspectives. A comprehensive assessment mechanism mitigates the inherent biases or knowledge gaps of single models. This multi-dimensional, collaborative analysis design not only incorporates key influencing factors such as identity consistency and group conflict but also significantly improves the comprehensiveness, accuracy, and robustness of the assessment results, ensuring the stability and reliability of risk assessment.

[0095] 4) Dynamic semantic simulation and non-quantitative factor modeling To address the shortcomings of traditional quantitative modeling methods based on indicator systems in explaining complex social phenomena, this application employs a multi-agent semantic simulation and dynamic interaction modeling mechanism. Compared to traditional models that rely on historical data and calculable numerical indicators, this system effectively captures and explains unstructured factors that are difficult to quantify, such as cultural psychology, social emotions, and group interactions, by simulating the cognition, emotions, and interactive behaviors of different social roles. This innovation overcomes the deficiencies of traditional numerical models in explaining non-quantifiable emotion transmission and social dynamics, enabling the model to more deeply understand the intrinsic mechanisms of public opinion fermentation and thus make more accurate predictions about the trajectory of negative situations and potential outbreak points.

[0096] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the multi-agent-based public figure speech influence pre-analysis method provided in the above embodiments.

[0097] This invention also provides a computer-readable medium having processor-executable non-volatile program code, on which a computer program is stored. When the computer program is run by a processor, it executes the steps of the multi-agent-based public figure speech influence pre-analysis method described above.

[0098] The computer program product provided in this embodiment of the invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0099] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0100] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] In the description of this invention, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0102] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A public figure speech influence pre-analysis method based on multi-agent, characterized in that, The method includes: Obtain user-inputted information about public figures' identities and target statements; By using the constructed social background generator, the identity information of the public figures and the target speech content are retrieved in stages to obtain the background information of the public figures, the cultural background of the topics, and the social public opinion environment data. The background information of the public figures, the cultural background of the topics, and the social public opinion environment data are retrieved and filtered respectively to generate a summary of the background of the figures and a summary of the social background. The personal background summary and the social background summary are integrated to generate social background data; Multiple role-based intelligent agents are constructed, and a corresponding enhancement configuration file is generated for each role-based intelligent agent based on the social background data through the constructed role enhancement configurator. This configuration file is used as the intelligent agent role capability configuration in the memory bank of each role-based intelligent agent. Each of the aforementioned role-based intelligent agents determines whether to participate in the current round of speaking. Each of the aforementioned role-based intelligent agents independently generates a speaking intention signal based on the definition of its own role's responsibilities, the semantic relevance of its area of ​​interest to the current speech content, and the novelty of its viewpoint or basis. The discussion results are obtained by conducting actual discussions through each of the aforementioned role-based intelligent agents; wherein, the discussion results include the statements, changes in positions, and interaction records of each of the aforementioned role-based intelligent agents; Based on the discussion results, the potential social impact of public figures' statements is comprehensively evaluated using a constructed consequence analyzer, resulting in consequence analysis results. The results of the consequence analysis were evaluated to obtain the evaluation results. 2.The multi-agent based public figure speech influence pre-analysis method according to claim 1, wherein, The background information of the public figures, the cultural background of the topics discussed, and the data on the social public opinion environment are retrieved and filtered respectively to generate a summary of the figures' backgrounds and a summary of the social backgrounds, including: The background information of the public figures is searched to obtain the first preliminary search results. The first preliminary search results are filtered for title relevance to obtain the filtered first search results. The filtered first search result is input into the social background generator to generate the person's background summary; wherein, the person's background summary includes professional background, social role, public image and historical events; The social opinion environment data is used to perform a speech social environment search to obtain a second preliminary search result; The second preliminary search results are filtered to obtain the filtered second search results; The filtered second search results are input into the social background generator to generate the social background summary; wherein, the social background summary includes social issues, public attitudes, media focus, and public opinion environment. 3.The multi-agent based public figure speech influence pre-analysis method of claim 2, wherein, The background information of the public figures is searched to obtain the first preliminary search results, including: The social background generator is invoked to analyze the background information of the public figure and identify the figure's name and related individuals. Use the names of the individuals mentioned above as the search targets; Construct multiple types of search queries based on the search object; The search is performed through a network search interface, and the search results returned by each query are used as the first preliminary search results. 4.The multi-agent based public figure speech influence pre-analysis method of claim 2, wherein, The social opinion environment data is used for a public discourse environment retrieval to obtain a second preliminary retrieval result, including: The social public opinion environment data is input into the social background generator to identify the value analysis results; wherein, the value analysis results include value orientation, attitude stance and viewpoint tendency; Based on the identity information of the public figures, the social background generator infers the motivation for the statements based on the value analysis results and judges the public figures' intentions in making the statements. After the motivation derivation is completed, the speech content is analyzed through the social background generator to extract key elements; the key elements include the people involved, social circles, core themes and keywords; Construct a combined query based on the key elements; Perform a network search for each query to obtain the search results returned by each query, and use them as the second preliminary search results. 5.The multi-agent based public figure speech influence pre-analysis method of claim 1, wherein, The constructed role enhancement configurator generates corresponding enhancement profiles for each role agent based on the social background data, including: The strings are parsed into preset interest areas for each intelligent agent, and then the strings are decomposed into a list of keywords. Build filter suggestions; Based on the screening criteria, the character agent's name, target speech content, and keyword list are input into the character enhancement configurator for judgment to obtain the screening results; wherein, the screening results include the screened keywords; The character name, target speech content, and filtered keywords of the character agent are input into the character enhancement configurator to generate multiple supplementary keywords; After merging and deduplicating the selected keywords and multiple supplementary keywords, a set of search keywords is obtained; For each keyword in the set of search keywords, construct a search query and obtain multiple search results for each keyword through a web search interface; The search results are input into the character enhancement configurator, which outputs structured feature data; wherein, the structured feature data includes a list of focus areas, motivation types, typical stances, descriptions of expression styles, evidence type preferences, and a list of target audiences; The structured feature data is fused into the basic configuration of each of the intelligent agents, the professional focus areas, behavioral styles and argument preferences of each intelligent agent are updated, and the enhanced configuration file is generated, which is used as the intelligent agent role capability configuration in the memory bank of each intelligent agent. 6.The multi-agent based public figure speech influence pre-analysis method of claim 1, wherein, Based on the discussion results, a comprehensive assessment of the potential social impact of public figures' statements was conducted using a constructed consequence analyzer, yielding consequence analysis results, including: Extract key information from the discussion results; wherein, the key information includes the target statement content, public figure profile, multiple rounds of discussion process, and the final views of each role; Construct a summary of public opinion, and extract the positions and suggestions of each agent according to their priority. The target speech content, the public figure information, and the summary of public opinion are input into a consequence analyzer based on large language models from multiple different vendors or architectures for analysis in seven dimensions. Each model outputs its own analytical viewpoint, thus constructing an original set of viewpoints. Perform view merging on the views in the original view set; The results of the analysis of the seven dimensions are comprehensively evaluated to obtain the consequences analysis results.

7. The multi-agent based public figure speech impact pre-analysis method according to claim 6, characterized in that, The consequences analysis results are obtained by comprehensively evaluating the analysis results from multiple dimensions, including: The analysis results of the seven dimensions are input into the consequence analyzer for extraction, and the two most important consequences are obtained as the core conclusions. The core conclusions are then standardized to obtain standardized conclusions. Based on the quantity and severity of the analysis results for each dimension, the overall risk is divided into multiple severity levels, and a risk assessment report is generated. The risk assessment report includes: detailed consequences, main causes, severity levels, and response recommendations for each dimension. 8.The multi-agent based public figure speech influence pre-analysis method of claim 1, wherein, The consequences analysis results are evaluated to obtain evaluation results, including: Obtain structured datasets; Extract a list of main consequences from the consequences analysis results; Manually labeled cases are read from the structured dataset and used as the evaluation benchmark; The list of main consequences and the manually labeled tags are input into the embedding model to obtain a high-dimensional vector; Calculate the cosine similarity matrix between the list of main consequences and the manually labeled tags, and construct the similarity mapping relationship between the predicted entries and the real tags; Based on each predicted entry, identify all matching tags whose similarity exceeds a set threshold to obtain the matching result; The evaluation result is calculated based on the matching result.

9. The multi-agent based public figure speech impact pre-analysis method according to claim 8, characterized in that, The evaluation result is calculated based on the matching result, including: Divide the number of successfully matched predicted entries by the total number of predicted entries to obtain the entry hit rate; For each tag, collect the similarity scores of all matching predicted entries and calculate the average similarity for each tag; The average similarity of all the aforementioned tags is calculated to obtain the average tag similarity. For each label, the maximum similarity with all predicted entries is found. Only the maximum similarity exceeding the threshold is retained and then averaged to obtain the mean maximum similarity. 10.A public figure speech influence pre-analysis system based on multi-agent, characterized in that, The system includes: The social background generation module is used to obtain the identity information of public figures and the target speech content input by the user; using the constructed social background generator, the identity information of public figures and the target speech content are retrieved in stages to obtain the background information of public figures, the cultural background of topics, and social opinion environment data; the background information of public figures and the social opinion environment data are retrieved and filtered separately to generate a figure background summary and a social background summary; the figure background summary and the social background summary are integrated to generate social background data; The agent configuration building module is used to build multiple role agents. The built role enhancement configurator generates a corresponding enhancement configuration file for each role agent based on the social background data, and uses it as the agent role capability configuration in the memory bank of each role agent. The multi-agent discussion module is used to determine whether each of the aforementioned role agents should participate in the current round of speaking. Each of the aforementioned role agents independently generates a speaking intention signal based on the definition of its own role's responsibilities, the semantic relevance of its area of ​​interest to the current speech content, and the novelty of its viewpoint or basis. The discussion results are obtained by conducting actual discussions through each of the aforementioned role-based intelligent agents; wherein, the discussion results include the statements, changes in positions, and interaction records of each of the aforementioned role-based intelligent agents; The consequences analysis module is used to comprehensively assess the potential social impact of public figures' statements based on the discussion results, using a constructed consequences analyzer, and to obtain the consequences analysis results. The evaluation module is used to evaluate the consequences analysis results and obtain the evaluation results.

Images