Open topic discussion process simulation and analysis method and system based on large language model

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses an open topic discussion process simulation and analysis method based on a large language model, which adds a multi-layer memory management function and improves the storage and query efficiency of session information. And a tool management module is arranged to provide efficient tool calling. And a topic sensing module is arranged to monitor the session state in real time and prevent topic deviation. And a collaborative planning management module is arranged to efficiently plan the collaboration among the intelligent agents, dynamically distribute tasks and ensure the completion quality. The action management module is arranged to coordinate actions of all agents, and invalid or repeated communication is avoided. Modeling of the intelligent agent is realized based on the big five personality characteristics, a mapping relation between personality dimensions and behaviors is established, and quantification of the personality dimensions of the intelligent agent is realized based on the behaviors in the discussion process. And designing a double-layer verification scheme, and verifying the consistency of personality characteristics and agent behaviors from the combination of single-dimensional characteristics and multi-dimensional personality characteristics of large five personalities.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method and system for simulating and analyzing open-topic discussion processes based on a large language model. Background Technology

[0002] Large Language Model (LLM)-based intelligent agents have been gradually developed and widely applied in fields such as medicine, education, and science and technology. LLM-based multi-agent systems provide support for the analysis and research of simulating real human discussions on development topics.

[0003] Open-ended discussions are a prevalent form of communication and exchange, characterized by diverse answers, broad participation, flexibility, and the need for in-depth thinking. However, the mechanisms by which factors such as differences in participant abilities, roles, and discussion management influence the outcomes of open-ended discussions require further exploration. LLM-based agents can leverage the reasoning, text generation, and process simulation capabilities of LLM to simulate discussion processes and define discussion roles, providing a novel research perspective for exploring the influencing mechanisms of open-ended discussions. However, the simulation and analysis of open-ended discussion processes based on multi-agent systems still face numerous challenges, including establishing quantitative standards for agent personality consistency assessment, evaluating agent role settings and behavioral consistency, and assessing the quality of open-ended discussion process simulation and outcomes.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0005] Simulation and analysis of open-topic discussion processes based on multi-agent systems still face many challenges, such as establishing quantitative standards for agent personality consistency assessment, agent role setting and behavioral consistency evaluation, and quality assessment of open-topic discussion process simulation and results. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method for simulating and analyzing open-topic discussion processes based on a large language model.

[0007] This invention is implemented as follows: A method for simulating and analyzing open-topic discussion processes based on a large language model includes:

[0008] Step 1: Build a multi-agent system. To ensure that multiple agents can have effective discussions on open topics to simulate the real human discussion process, we design a multi-agent architecture that can efficiently manage discussion information, coordinate collaboration between agents, generate personalized responses, and improve the quality of discussions through the use of efficient tools.

[0009] Step two involves two aspects: agent modeling design based on the Big Five personality traits and verification of consistency between personality traits and behavior. A two-layer verification scheme is adopted in the consistency verification method to verify the consistency between the agent's personality and behavior. This ensures that under the Big Five personality-based modeling approach, each agent can understand its assigned personality traits and exhibit corresponding behaviors.

[0010] Step 3: Based on the summary report content generated by the agent as the data source, and through the analysis and summarization of domestic and foreign text quality assessment methods, and combined with the characteristics of agent-generated content, a quality assessment index system for summary reports of multi-agent open topic discussions is constructed; at the same time, the weights of each index in the summary report quality assessment index system are determined by using expert scoring and analytic hierarchy process.

[0011] Step four: Simulate the multi-agent open-topic discussion process of LLM and evaluate the results.

[0012] Furthermore, the multi-agent architecture described in step one is designed from five aspects: perception, memory, planning, action, and tools, as well as meeting the non-functional requirements of efficiency, scalability, and robustness; the overall research scheme is as follows:

[0013] (1) Memory management; Create a hierarchical memory management system that combines short-term memory (STM) and long-term memory (LTM). Short-term memory is responsible for storing the content of the previous few rounds of speech in the current session to ensure that the agent can understand and respond to the current dialogue and avoid the agent's speech content being incoherent and inconsistent. Long-term memory is used to store the entire history of the topic discussion. For example, when the dialogue continues for multiple stages, the agent needs to obtain the previous historical content to maintain consistency and avoid speech being disjointed or repeated.

[0014] (2) Topic awareness management; In order to prevent the topic discussion of the agent from diverging and to keep the discussion goals consistent, this paper sets up topic tracking (TT) to dynamically monitor the core issues of the current discussion, to ensure that the agent communicates around the topic, to keep the discussion context coherent, and to avoid going off-topic or getting confused;

[0015] (3) Collaborative planning and management; identify task objectives and assign tasks to appropriate agents, design collaboration rules and strategies between agents to ensure that topic discussions or task progress are carried out in an orderly manner; monitor the progress of discussions and adjust the direction of discussions to ensure that they do not deviate from the objectives;

[0016] (4) Action management; ensure that the agent can generate actions with individual characteristics to simulate the diversity and role differences in human discussions, while maintaining the logical consistency of the discussion and the efficiency of the actions;

[0017] (5) Tool management; add relevance filtering and semantic retrieval tools; establish a unified tool call interface to facilitate intelligent agents to call, and also support the addition of new tools to adapt to different scenario needs and have high scalability.

[0018] Furthermore, the specific method for intelligent agent personality modeling design described in step two is as follows:

[0019] First, based on relevant psychological theories and research, we establish a mapping relationship between the five dimensions of the Big Five personality traits and behavior, as well as specific quantification methods;

[0020] To make the description more concise, we use E to represent extraversion, A to represent agreeableness, O to represent openness, N to represent neuroticism, and C to represent conscientiousness. From the five dimensions of the Big Five personality traits, we select one level for each dimension (high, medium, low) to form a combination containing all five dimensions. Each combination represents a unique personality trait with different behavioral patterns and emotional tendencies. Based on the factors considered in the previous section regarding agent personality design, we need to consider the coordination between combinations and the mapping relationship between dimensions and behaviors when combining personality dimensions. Regarding agent personality modeling, the prompt template contains five parts: a description of the Big Five personality dimensional combinations and their corresponding behavioral manifestations, a description of language style, a task description, and an example.

[0021] Furthermore, in the two-layer verification design of the intelligent agent's personality, the first layer is a single-dimensional verification, and the second layer is a multi-dimensional interaction verification; the specific method is as follows:

[0022] (1) Single-dimensional verification: The five dimensions of the Big Five personality were set to two levels, high and low; the control variable method was used, and only one dimension was changed in each experiment; and the frequency of the agent’s corresponding behavior during open topic discussion was counted.

[0023] (2) In the multidimensional interaction verification, a 3×2 factor design is adopted, namely, three personality trait combinations and two dimension levels: High and Low. The irrelevant dimension is set as the medium dimension level. Based on relevant psychological theories, this method selects to verify whether the behavior of the agent under the interaction combination of three personality traits is consistent with the psychological theory, such as the interaction combination of extraversion + agreeableness + neuroticism (E+A+N), the interaction combination of openness + extraversion + agreeableness (O+E+A), the interaction combination of extraversion + agreeableness + neuroticism (E+A+N), and the interaction combination of openness + agreeableness + agreeableness (O+E+A). The correlation between the interaction combination of (E+A+N) and (O+E+A) and "conflict" and "innovation" behaviors is explored, and the correlation between the interaction of (A+N+C) and (O+E+C) and "cooperative" and "leadership" behaviors is explored.

[0024] Furthermore, the quality score of the open-topic discussion summary report in step three is derived from evaluating the text quality of the summary report across five dimensions; specifically as follows:

[0025] (1) When verifying the relevance of the task, in order to verify the similarity between the open topic and the summary report text, sentence vector embedding is performed first, and then the similarity between the summary report text and the dynamic topic embedding vector is calculated. .in This represents the sentence vector of the summary report text. Sentence vectors for dynamic topics;

[0026]

[0027] (2) When performing textual logical coherence verification, the semantic similarity between adjacent sentences is mainly calculated, among which... Indicating the cohesion between sentences, Indicates structural integrity. This represents the mean similarity between adjacent sentences. The standard deviation of similarity The sentence vector representing the current sentence;

[0028]

[0029]

[0030]

[0031] (3) When assessing information richness, the overlap of technical terms between the assessment report and the dynamic topic is calculated separately, and the precision, recall, and F1 score are calculated based on the technical terms; lexical diversity is calculated by calculating the ratio of the number of deduplicated words to the threshold; the relevant calculation formulas are as follows, where Indicates TF-IDF coverage. Indicates the coverage rate of technical terms. This indicates content diversity;

[0032]

[0033]

[0034]

[0035] (4) The fluency of sentence expression is evaluated using a large language model, and sentence dependency relations are parsed. Semantic naturalness is captured through perplexity. The specific formula is as follows, where Indicates grammatical fluency, Indicates naturalness. Indicates overall fluency. Indicates the number of sentences. Indicates the level of confusion;

[0036]

[0037]

[0038]

[0039] (5) The report text content is evaluated using a combination of two methods: sentiment bias extraction and large language modeling. Sentiment bias analysis primarily calculates the sentiment polarity of the text, while large language modeling is used to generate loss and term ratio to assess content authenticity. The relevant calculation formulas are shown below. Indicates the reliability of the content. This indicates a neutral stance. This indicates overall credibility; the content reliability metric combines loss and terminology ratio. The lower the loss and the higher the terminology ratio, the more reliable the content. The closer to 1; |Polarity| represents the absolute value of sentiment polarity. The more neutral the sentiment polarity, the better. The closer to 1; It is overall credibility, which is composed of content reliability ( and neutrality of position ( The weighted sum of ).

[0040]

[0041]

[0042]

[0043] Finally, the overall score was calculated by having experts score each of the five dimensions, determining the weights for each dimension, and then summing the weighted scores for each dimension to arrive at the total quality score for the summary report. .

[0044] Furthermore, the simulation of the LLM multi-agent open-topic discussion process and result evaluation involves: First, research on the impact of different agent role combinations on the quality of the summary report during the open-topic discussion process; designing agent roles for different groups based on the Big Five personality model; and designing agent combination strategies; simulating the discussion process with a specific open topic as the discussion task, generating a discussion summary report, and evaluating and analyzing its quality according to an evaluation index system; Second, introducing a timely feedback agent during the discussion process to explore its impact on the quality of the summary report; designing a feedback strategy for the timely feedback agent, including feedback intensity, timing, and content; specifically:

[0045] (1) Strategy design of agent role combination; Based on the above analysis of different combinations of group members, agent roles are designed in the open topic discussion process. According to the common characteristics of group member combinations in reality, agents with different roles are designed, such as innovators, executors, challengers, coordinators, etc., and agents with different roles are combined into different groups, such as complementary groups, conflict groups, balanced groups and collaborative groups, etc.; Through simulation of the open topic discussion process of multiple agents, the impact of different agent role combinations on the quality of the summary report generated by the open topic discussion is explored and analyzed.

[0046] (2) Timely feedback strategy design for the intelligent agent; First, design the feedback intensity type of the intelligent agent, such as weak, medium, and strong; specify different feedback content corresponding to different intensities; such as the number of feedback questions, the level of detail, the tone of feedback, and the intensity of instructions; set different numbers of feedback questions and different tone of feedback guidance for different feedback intensity levels; it is also necessary to determine the analysis indicators of the feedback of the intelligent agent, that is, from which aspects to provide feedback, such as the coverage of viewpoints, arguments, and evidence in the agent's speech; then design the feedback timing of the intelligent agent, and intervene in the discussion of the open topic at different times; explore and analyze the impact of different timing interventions of the intelligent agent on the quality of the summary report of the open topic discussion;

[0047] The workflow of the feedback agent in step four is as follows:

[0048] Data collection and preprocessing stage: Import the current discussion topic and background; adjust the feedback template according to the assigned feedback intensity level; confirm the current discussion round and feedback timing type;

[0049] (2) Content analysis phase: Extract all agent speech content from the current round (real-time feedback) or the last three rounds (stage feedback); use topic modeling to identify topics, arguments, evidence and sentiment; complete the evaluation of multi-agent speech content according to the dimensions of content analysis; generate feedback priorities; provide some aspects that need the most improvement according to the intensity, such as providing an improvement direction when the feedback intensity type is weak;

[0050] (3) Feedback generation stage: Select the appropriate language template according to the feedback intensity level; organize the feedback content according to the "affirmation-guidance-suggestion" structure; apply the preset tone and vocabulary standards; control the length of the feedback within the prescribed range.

[0051] Another objective of this invention is to provide a simulation and analysis system for open-topic discussion processes based on a large language model, comprising:

[0052] The module is designed to build a multi-agent system. To ensure that multiple agents can have effective discussions on open topics to simulate the real human discussion process, a multi-agent architecture is designed that can efficiently manage discussion information, coordinate collaboration between agents, generate personalized responses, and improve the quality of discussions through the use of efficient tools.

[0053] The verification module is used to carry out work on two aspects: the modeling design of intelligent agents based on the Big Five personality traits and the verification of consistency between personality traits and behavior. The consistency verification method adopts a two-layer verification scheme to verify the consistency between the personality and behavior of the intelligent agents. It ensures that under the modeling method based on the Big Five personality traits, each intelligent agent can understand its assigned personality traits and exhibit corresponding behaviors.

[0054] The analysis module is used to construct a quality assessment index system for summary reports of multi-agent open-topic discussions, based on the data source of summary reports generated by agents. This system is achieved by analyzing and summarizing methods for assessing the quality of texts both domestically and internationally, and by combining the characteristics of the content generated by agents. At the same time, the module uses expert scoring and the analytic hierarchy process to determine the weights of each index in the summary report quality assessment index system.

[0055] The simulation module is used to simulate the multi-agent open-topic discussion process and result evaluation in LLM.

[0056] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the open topic discussion process simulation and analysis method based on a large language model.

[0057] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the open-topic discussion process simulation and analysis method based on a large language model.

[0058] Another objective of this invention is to provide an information data processing terminal for implementing the open topic discussion process simulation and analysis system based on a large language model.

[0059] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0060] Existing multi-agent open-topic discussion simulation and analysis techniques based on Large Language Models (LLM) face three major technical bottlenecks: challenges in agent consistency and quantification, difficulty in controlling multi-agent collaboration efficiency and discussion quality, and a lack of a scientific system for evaluating the quality of discussion results. To address these issues, this invention proposes a multi-agent architecture design for open-topic discussion simulation.

[0061] This invention proposes a Big Five personality trait modeling method for open-topic discussions and a two-layer personality verification mechanism for intelligent agents. Based on the Big Five personality theory in psychology, a three-dimensional mapping system of "personality dimension - behavioral performance - quantification method" is established. The five dimensions—Extraversion (E), Agreeableness (A), Openness (O), Neuroticism (N), and Conscientiousness (C)—are combined at high, medium, and low levels to form various unique personalities. Each personality corresponds to a specific language style, typical behavior, and quantification formula. Simultaneously, standardized prompt templates are designed, including five parts: personality dimension combination, behavioral description, and language style, ensuring that the intelligent agent accurately understands the personality settings. A two-layer consistency verification scheme is also proposed. Single-dimensional verification mainly uses the controlled variable method, setting a single personality dimension to a high / low level while keeping other dimensions at a medium level. The frequency of corresponding behaviors of the intelligent agent is statistically analyzed to verify the linear relationship between a single dimension and behavior. The multidimensional interaction verification adopts a 3×2 factor design, selecting typical personality combinations such as E+A+N and O+E+A to explore the impact of multidimensional interaction on complex behaviors such as "conflict," "innovation," "cooperation," and "leadership." By quantifying the content of the speeches, the consistency between personality combinations and behaviors is analyzed and verified.

[0062] This invention constructs a quality evaluation index system for summary reports of open-topic discussions based on multi-agent systems and proposes a scientific weight allocation method. The system comprises five dimensions: task relevance, logical coherence, information richness, fluency of expression, and content credibility, with each dimension employing a quantitative formula. The invention proposes a scientific weight allocation method for the report quality evaluation indicators, combining expert scoring with the analytic hierarchy process (AHP) to determine the weights of each evaluation dimension. A weighted summation is then used to obtain the total quality score of the summary report, ensuring the objectivity and rationality of the evaluation results.

[0063] The expected benefits and commercial value of the technical solution of this invention after transformation include direct economic benefits, indirect commercial value, and industry benefits. Direct economic benefits include the development of standardized multi-agent discussion simulation and analysis products, covering core areas such as education, corporate training, market research, and scientific research. In the education field, it can provide customized simulated discussion scenarios for university critical thinking courses and primary and secondary school critical thinking training, charging a service fee per school per year. In the corporate training field, it provides simulation training tools for cross-departmental collaboration and customer communication scenarios, charging according to licensing and procurement costs. In the scientific research field, it provides a quantitative research platform for disciplines such as social psychology and organizational behavior, charging according to different versions of the platform, generating direct economic value. Regarding indirect commercial value and industry benefits, the technical solution of this invention can complement existing large language model products, enhancing product differentiation and competitiveness, and helping to seize the initiative in the intelligent collaboration track. The related technology transformation of this invention can significantly reduce the time and labor costs of discussion simulation in various industries, significantly improving the overall efficiency of the industry.

[0064] This invention addresses the limitations of existing multi-agent frameworks, such as "quantification of agent personality consistency," "multi-agent collaboration management," and "multi-dimensional quantitative evaluation of discussion results." It proposes a multi-agent architecture design for open-topic discussion simulation. This invention innovatively constructs an integrated technical solution encompassing "Big Five personality modeling + dual-layer verification," "five-dimensional collaboration architecture + full-process management," and "five-dimensional evaluation system + weight optimization." This achieves a scientific and standardized end-to-end process for multi-agent open-topic discussions, from personality simulation and process management to result evaluation, filling a gap in this technological field both domestically and internationally. This invention improves system efficiency, reduces information redundancy, and enhances the quality of open-topic discussions, increasing the efficiency of collaboration among multiple agents. Regarding the quality evaluation framework for multi-agent open-topic discussion summary reports, it addresses the current lack of scientific and multi-dimensional evaluation standards, providing a quantifiable evaluation method for related research. Furthermore, it innovatively introduces a timely feedback agent into the multi-agent interaction process, exploring and analyzing the impact of different agent role combinations and the introduction of a timely feedback agent on the results of open-topic discussions, providing a new perspective and empirical foundation for research in the field of artificial intelligence. This research not only helps to solve the limitations of depth and interaction in existing role simulations, but also provides theoretical basis and practical guidance for building more natural and efficient human-computer collaboration systems, and promotes the application of LLM-based intelligent agents in complex social scenarios.

[0065] This invention explicitly solves a core technical challenge that has long existed in the field of multi-agent open-topic discussions, a challenge that the industry has universally desired to overcome but has yet to succeed in. For a long time, the industry has hoped to achieve "human-like" multi-agent discussion simulations—ensuring stable and non-drifting agent personalities, achieving efficient collaboration and high-quality discussions, and accurately evaluating discussion effectiveness. However, this has been hampered by difficulties in interdisciplinary integration, balancing collaborative management, and designing evaluation systems. Specifically, in the problem of personality consistency, various personality modeling methods have been tried, but none have achieved a precise mapping from theory to behavior to quantification; in multi-agent collaboration, balancing "diversity of viewpoints" and "orderliness of topics" has always been difficult, resulting in either uncontrolled topics or singular viewpoints; and in discussion evaluation, the lack of scientific quantitative standards makes it impossible to measure effectiveness. This invention, through the integration of technologies from psychology and artificial intelligence, designs a proprietary architecture and method to successfully solve the above problems and realize the industry's long-standing technical demands. Attached Figure Description

[0066] Figure 1 This is a flowchart of the open-topic discussion process simulation and analysis method based on a large language model provided in the embodiments of the present invention.

[0067] Figure 2 This is a structural block diagram of an open-topic discussion process simulation and analysis system based on a large language model, provided in an embodiment of the present invention.

[0068] Figure 3 This is a multi-agent overall architecture diagram for open topic discussion simulation provided in an embodiment of the present invention.

[0069] Figure 4 This is a memory management architecture diagram provided in an embodiment of the present invention.

[0070] Figure 5 This is a topic analysis architecture diagram provided in an embodiment of the present invention.

[0071] Figure 6 Radar chart comparing scores for each group and each dimension. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0073] like Figure 1 As shown in the figure, the method for simulating and analyzing open-topic discussion processes based on a large language model provided by this invention includes the following steps:

[0074] S101, to build a multi-agent system, in order to ensure that multiple agents can have effective discussions on open topics to simulate the real human discussion process, a multi-agent architecture is designed that can efficiently manage discussion information, coordinate cooperation between agents, generate personalized responses, and improve the quality of discussion through the use of efficient tools.

[0075] The multi-agent architecture described in S101 is designed from five aspects: perception, memory, planning, action, and tools, while also meeting non-functional requirements for efficiency, scalability, and robustness. The overall research scheme is as follows:

[0076] (1) Memory Management. A hierarchical memory management system combining short-term memory (STM) and long-term memory (LTM) is created. Short-term memory is responsible for storing the content of the previous few rounds of speech in the current session, ensuring that the agent can understand and respond to the current dialogue and avoid incoherence and inconsistency in the agent's speech. Long-term memory is used to store the entire history of the topic discussion. For example, when the dialogue continues for multiple stages, the agent needs to obtain the previous historical content to maintain consistency and avoid disjointed or repetitive speech.

[0077] (2) Topic awareness management. To prevent the agents from going off-topic and to keep the discussion goals consistent, this paper sets up topic tracking (TT) to dynamically monitor the core issues of the current discussion, ensuring that the agents communicate around the topic, maintain the continuity of the discussion context, and avoid going off-topic or getting confused.

[0078] (3) Collaborative planning and management. Identify task objectives and assign tasks to appropriate agents. Design collaboration rules and strategies between agents to ensure that topic discussions or task progress proceed in an orderly manner. Monitor discussion progress and adjust the direction of discussions to ensure that they do not deviate from the objectives.

[0079] (4) Action Management. Ensure that the agent can generate actions with individual characteristics to simulate the diversity and role differences in human discussions, while maintaining logical consistency and efficiency in the discussion.

[0080] (5) Tool Management. Add relevance filtering and semantic retrieval tools. Establish a unified tool call interface to facilitate intelligent agents to call the tools, while also supporting the addition of new tools to adapt to different scenario needs and have high scalability.

[0081] S102 focuses on two aspects: agent modeling design based on the Big Five personality traits and verification of consistency between personality traits and behavior. A two-layer verification scheme is employed to verify the consistency between the agent's personality and behavior. This ensures that, under the Big Five personality-based modeling approach, each agent can understand its assigned personality traits and exhibit corresponding behaviors.

[0082] The specific method for intelligent agent personality modeling and design described in S102 is as follows:

[0083] First, based on relevant psychological theories, a mapping relationship is established between the five dimensions of the Big Five personality traits and behavior, along with specific quantification methods, as shown in Table 1 below.

[0084] Table 1. Personality and Behavior Mapping and Quantification

[0085]

[0086] To make the description more concise, we use E to represent extraversion, A to represent agreeableness, O to represent openness, N to represent neuroticism, and C to represent conscientiousness. From the five dimensions of the Big Five personality traits, we select one level for each dimension (high, medium, and low) to form a combination containing all five dimensions. Each combination represents a unique personality trait with different behavioral patterns and emotional tendencies. Based on the factors considered in the previous section regarding agent personality design, the coordination between combinations and the mapping relationship between dimensions and behaviors must be considered when combining personality dimensions. Regarding agent personality modeling, the prompt template contains five parts: a description of the Big Five personality trait combination and its corresponding behavioral performance, a description of language style, a task description, and an example. The specific prompt template design for the agent is shown in Table 2 below.

[0087] Table 2. Prompt Templates for Agent Personality Design

[0088]

[0089] Finally, Table 3 shows the specific content of the prompt word design for a certain agent. [Medium Extraversion, High Openness, High Agreeableness, Low Neuroticism, High Conscientiousness] indicates that the agent's Big Five personality traits are moderate extraversion, high openness, high agreeableness, low neuroticism, and high conscientiousness. In addition to the five dimensions of the Big Five personality traits, the table also includes descriptive information on the specific behavioral characteristics, language style, and examples for each dimension.

[0090] Table 3. Specific examples of prompts for the intelligent agent.

[0091]

[0092] In the two-layer verification design of intelligent agent personality, the first layer is single-dimensional verification, and the second layer is multi-dimensional interaction verification. The specific method is as follows:

[0093] (1) Single-dimensional validation: The five dimensions of the Big Five personality traits were set to two levels, high and low. Using the controlled variable method, only one dimension was changed in each experiment. The frequency of the corresponding behaviors of the agents during open-ended topic discussions was also counted.

[0094] (2) A 3×2 factorial design was adopted in the multidimensional interaction verification, namely, three personality trait combinations and two dimension levels: High and Low. The irrelevant dimension was set to the medium dimension level. Based on relevant psychological theories, this study selected to verify whether the behavior of the agent under the interaction combination of three personality traits was consistent with psychological theories, such as the interaction combination of extraversion + agreeableness + neuroticism (E+A+N), openness + extraversion + agreeableness (O+E+A), extraversion + agreeableness + neuroticism (E+A+N), and openness + agreeableness + agreeableness (O+E+A). The correlation between the interaction combinations of (E+A+N) and (O+E+A) and "conflict" and "innovative" behaviors was explored, and the correlation between the interaction combinations of (A+N+C) and (O+E+C) and "cooperative" and "leadership" behaviors was explored.

[0095] This invention primarily quantifies and statistically analyzes the content of an agent's speech. It uses the semantic similarity between the spoken content and reference statements as a basis for scoring "conflict-making," "innovative," and "cooperative" behaviors. For "leadership" behavior, it combines semantic analysis (dominance, organization, motivation), speaking frequency, and response rate as the scoring criteria for the agent's "leadership" behavior in open-topic discussions. The response rate is calculated by counting the number of times the agent receives a response during their speech. Finally, three-way ANOVA and heatmaps are used to demonstrate the impact of different personality combinations on the above four behaviors.

[0096] S103, based on the summary report content generated by intelligent agents as the data source, and through the analysis and summarization of domestic and international text quality assessment methods, and combined with the characteristics of intelligent agent-generated content, constructs a quality assessment index system for summary reports of multi-agent open-topic discussions. Simultaneously, it uses expert scoring and the analytic hierarchy process (AHP) to determine the weights of each index in the summary report quality assessment index system.

[0097] The quality score for the open-topic discussion summary report in S103 is derived from evaluating the text quality of the summary report across five dimensions. Specifically:

[0098] (1) When verifying the relevance of the task, in order to verify the similarity between the open topic and the summary report text, sentence vector embedding is performed first, and then the similarity between the summary report text and the dynamic topic embedding vector is calculated. .in This represents the sentence vector of the summary report text. Sentence vectors for dynamic topics.

[0099]

[0100] (2) When performing textual logical coherence verification, the semantic similarity between adjacent sentences is mainly calculated, among which... Indicating the cohesion between sentences, Indicates structural integrity. This represents the mean similarity between adjacent sentences. The standard deviation of similarity This represents the sentence vector of the current sentence.

[0101]

[0102]

[0103]

[0104] (3) When assessing information richness, the overlap of technical terms between the report and the dynamic topic is mainly calculated, and precision, recall, and F1 score are calculated based on the technical terms. Lexical diversity is also calculated by determining the ratio of the number of unique words to a threshold. The relevant calculation formulas are as follows, where... Indicates TF-IDF coverage. Indicates the coverage rate of technical terms. It indicates the diversity of content.

[0105]

[0106]

[0107]

[0108] (4) The fluency of sentence expression is evaluated using a large language model, and sentence dependency relations are parsed. Semantic naturalness is captured through perplexity. The specific formula is as follows, where Indicates grammatical fluency, Indicates naturalness. Indicates overall fluency. Indicates the number of sentences. Indicates the level of confusion.

[0109]

[0110]

[0111]

[0112] (5) The report text content is evaluated using a combination of two methods: sentiment bias extraction and large language modeling. Sentiment bias analysis primarily calculates the sentiment polarity of the text, while large language modeling is used to generate loss and term ratio to assess content authenticity. The relevant calculation formulas are shown below. Indicates the reliability of the content. This indicates a neutral stance. This indicates overall credibility; the content reliability metric combines loss and terminology ratio. The lower the loss and the higher the terminology ratio, the more reliable the content. The closer to 1; |Polarity| represents the absolute value of sentiment polarity. The more neutral the sentiment polarity, the better. The closer to 1; It is overall credibility, which is composed of content reliability ( and neutrality of position ( The weighted sum of ).

[0113]

[0114]

[0115]

[0116] Finally, the overall score was calculated by having experts score each of the five dimensions, determining the weights for each dimension, and then summing the weighted scores for each dimension to arrive at the total quality score for the summary report. .

[0117] S104, simulating the multi-agent open-topic discussion process in LLM and evaluating the results. First, the impact of different agent role combinations on the quality of the summary report during the open-topic discussion is investigated. Based on the Big Five personality model, agent roles for different groups are designed, and agent combination strategies are designed. Using a specific open topic as the discussion task, the discussion process is simulated, a discussion summary report is generated, and its quality is evaluated and analyzed according to an evaluation index system. Second, a timely feedback agent is introduced during the discussion process to explore its impact on the quality of the summary report. Feedback strategies for the timely feedback agent are designed, including feedback intensity, timing, and content. Specifically:

[0118] (1) Strategy Design for Agent Role Combinations. Based on the above analysis of different combinations of group members, agent roles are designed during the open-topic discussion process. Different roles are designed based on common group member combinations in reality, such as innovator, executor, challenger, and coordinator. Agents with different roles are then combined into different groups, such as complementary groups, conflict groups, balanced groups, and collaborative groups. Through simulation of the open-topic discussion process involving multiple agents, the impact of different agent role combinations on the quality of the summary report generated from the open-topic discussion is explored and analyzed.

[0119] (2) Timely Feedback Strategy Design for the Feedback Agent. First, design the feedback intensity types for the feedback agent, such as weak, medium, and strong. Specify different feedback content corresponding to different intensities, such as the number of feedback questions, the level of detail, the tone of the feedback, and the intensity of the instruction. Different feedback intensity levels correspond to different numbers of feedback questions and different guiding tones. It is also necessary to determine the analysis indicators for the feedback agent's feedback, i.e., from which aspects feedback should be provided, such as the coverage of viewpoints, arguments, and evidence in the agent's statements. Then, design the feedback timing for the feedback agent, intervening in the discussion of open topics at different times. Explore and analyze the impact of different timing interventions on the quality of the open topic discussion summary report.

[0120] The specific strategies for the S104 agent role combination are shown in Table 4 below:

[0121] Table 4 Role Combination Table

[0122]

[0123] The workflow of the feedback agent in S104 is as follows:

[0124] Data collection and preprocessing stage: Import the current discussion topic and background; adjust the feedback template according to the assigned feedback intensity level; confirm the current discussion round and feedback timing type.

[0125] (2) Content Analysis Phase: Extract all agent statements from the current round (real-time feedback) or the last three rounds (stage feedback); use topic modeling to identify topics, arguments, evidence, and sentiment; evaluate the statements of multiple agents based on the dimensions of content analysis; generate feedback priorities. Based on the intensity, provide the aspects that need the most improvement, such as providing an improvement direction when the feedback intensity type is weak.

[0126] (3) Feedback generation stage: Select the appropriate language template according to the feedback intensity level; organize the feedback content according to the "affirmation-guidance-suggestion" structure; apply the preset tone and vocabulary standards; control the length of the feedback within the prescribed range.

[0127] like Figure 2 As shown, the open-topic discussion process simulation and analysis system based on a large language model provided in this embodiment of the invention includes:

[0128] The module is designed to build a multi-agent system. To ensure that multiple agents can have effective discussions on open topics to simulate the real human discussion process, a multi-agent architecture is designed that can efficiently manage discussion information, coordinate collaboration between agents, generate personalized responses, and improve the quality of discussions through the use of efficient tools.

[0129] The verification module is used to carry out work on two aspects: the modeling design of intelligent agents based on the Big Five personality traits and the verification of consistency between personality traits and behavior. The consistency verification method adopts a two-layer verification scheme to verify the consistency between the personality and behavior of the intelligent agents. It ensures that under the modeling method based on the Big Five personality traits, each intelligent agent can understand its assigned personality traits and exhibit corresponding behaviors.

[0130] The analysis module is used to construct a quality assessment index system for summary reports of multi-agent open-topic discussions, based on the data source of summary reports generated by agents. This system is achieved by analyzing and summarizing methods for assessing the quality of texts both domestically and internationally, and by combining the characteristics of the content generated by agents. At the same time, the module uses expert scoring and the analytic hierarchy process to determine the weights of each index in the summary report quality assessment index system.

[0131] The simulation module is used to simulate the multi-agent open-topic discussion process and result evaluation in LLM.

[0132] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the open topic discussion process simulation and analysis method based on a large language model.

[0133] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the open-topic discussion process simulation and analysis method based on a large language model.

[0134] Another objective of this invention is to provide an information data processing terminal for implementing the open topic discussion process simulation and analysis system based on a large language model.

[0135] Specific implementation of the present invention:

[0136] Multi-agent systems are distributed systems composed of multiple autonomous agents. Each agent can autonomously perceive its environment, make decisions through learning, and achieve its own goals or collaborate to complete the overall goal. Open-topic discussions, as a form of communication without a fixed agenda, are widespread in human society. With the rapid development of LLM (Limited Learning Model), LLM-based multi-agent systems provide fundamental support for simulating real-world human discussions on open topics. Multiple agents can play different roles, generating multi-dimensional perspectives through rules of cooperation and competition to simulate the discussion process of real humans.

[0137] The Big Five Personality Theory: As a widely accepted personality description framework in contemporary psychology, it extracts five core personality traits—extraversion, agreeableness, conscientiousness, neuroticism, and openness—through lexical analysis and factor analysis. Specific personality descriptions are shown in Table 5. This theory not only possesses cross-cultural consistency and stability but also provides scientific and quantifiable methods for personality assessment. Therefore, it has been widely applied in personality assessment, career choice, and mental health research, becoming a theoretical cornerstone of modern personality research.

[0138] Table 5: Description of the Five Key Personality Traits

[0139]

[0140] Multi-agent architecture for open-topic discussion simulation

[0141] like Figure 3 As shown, the multi-agent architecture in this example consists of 5 parts.

[0142] (1) Memory Module. It mainly consists of Short-Term Memory (STM), Long-Term Memory (LTM), and summary management, optimizing information storage and retrieval to make discussions more efficient. Its memory management architecture diagram is shown below. Figure 4 As shown. Specifically:

[0143] 1) Short-term memory. Primarily used to store the content of the most recent N rounds of conversation. It uses a default sliding window storage method, treating the most recent N rounds of conversation as context to improve the coherence of the group chat context in the short term. Specifically, it includes:

[0144] First, store the most recent N rounds of dialogue. This represents the t-th round of dialogue messages.

[0145]

[0146] Then, when the current conversation exceeds N rounds, the earliest conversation will be discarded. Important information will be stored in long-term memory to prevent the loss of key information.

[0147] 2) Long-term memory. Stores complete dialogue history, retrieves relevant information through semantic retrieval, and supports cross-scenario knowledge accumulation. Specifically, this includes:

[0148] First, store the complete conversation in the database.

[0149] Then, in order to capture semantic similarity and effectively retrieve content with different expressions but similar meanings, it is necessary to calculate a semantic vector for each dialogue and store it in a vector database.

[0150]

[0151] Finally, when the agent needs to review historical dialogues, it performs semantic retrieval. By calculating similarity, it finds historical records with similar meanings.

[0152] 3) Summary Management. The semantic summarization tool in the utility module is used to periodically summarize short-term memory, extracting key information and storing it in long-term memory to reduce redundant memory. Specifically, this includes:

[0153] First, the semantic summarization tool is invoked, and BERT is used to periodically generate summaries of the dialogue content every 10 rounds.

[0154] Then, when the agent queries historical session records, it prioritizes providing summary versions to improve query efficiency.

[0155] (2) Tool Management Module. This module includes relevance filtering tools, semantic retrieval tools, semantic summarization tools, and deduplication tools. Specifically:

[0156] 1) Relevance filtering tools are used to filter information irrelevant to the current topic or task, ensuring the discussion remains focused. They are primarily used for topic analysis and tracking in the perception module, and for response filtering in the action module. The specific steps are as follows:

[0157] Step 1: Receive the response text from the agents to be filtered and the current topic, and preprocess the received content.

[0158] Step 2: Semantic Embedding Generation. Using Sentence-BERT, embeddings of the agent's response information and the topic are generated, where... This indicates the embedding of the agent's response text. This indicates that the current topic is embedded. This represents the text content of the agent's response. Sentence-BERT() represents the topic content of the current conversation and implements sentence-level vector embedding.

[0159]

[0160]

[0161] Step 3: Calculate relevance. Calculate relevance using cosine similarity and filter text below the threshold θ.

[0162] Step 4: Output the filtered text and relevance score.

[0163] 2) Semantic retrieval tools primarily retrieve and query semantically relevant information from long-term memory and external databases. This improves the efficiency of agents querying historical records because semantic retrieval utilizes natural language understanding and vector matching techniques to find semantically related information, rather than relying solely on keyword matching. Compared to traditional full-text search or SQL keyword matching methods, semantic retrieval can identify information with different expressions but similar semantics. The specific steps are as follows:

[0164] Step 1: Receive query and context information to prepare data for subsequent processing. Clean the query string.

[0165] Step 2: Generate query embeddings. Use Sentence-BERT to generate query and context embeddings, where α represents the fusion weight. Indicates query embedding. Indicates context embedding, This represents the vector embedding used to generate the query.

[0166]

[0167]

[0168] Step 3: Retrieve Candidates. Retrieve candidate information from long-term memory or external databases. Based on query embedding. Search for candidates.

[0169] Step 4: Calculate the semantic similarity between the query embedding and the candidate embeddings using cosine similarity. Finally, sort the candidate results from highest to lowest similarity and output them.

[0170] 3) Semantic summarization tools primarily perform semantic-level summarization of long texts or multi-turn dialogues. Through semantic understanding and information extraction, they provide efficient information processing capabilities for multi-agent systems. They are mainly applied to information compression and storage optimization, key information extraction and focusing, and contextual understanding and coherence support. In terms of information compression and storage optimization, they compress lengthy texts or multi-turn dialogues into concise summaries, thereby reducing storage and computational overhead. Regarding key information extraction and focusing, they extract key information from complex texts or dialogues, helping the system focus on core content. In terms of contextual understanding and coherence support, they primarily preserve the contextual semantics of the dialogue through semantic summarization, supporting the coherence of subsequent tasks. Its key components include the following two aspects:

[0171] First, all sentences are embedded, and the average of all sentence embeddings is taken to generate the document embedding. The specific formula is shown below, where... Vector embeddings representing documents Let N represent the vector embedding of the i-th sentence, and N represent the total number of sentences.

[0172]

[0173] Second, sentence importance is scored by calculating the cosine similarity between each sentence and the document, which is then used as an importance score. Finally, the sentences are sorted according to their importance scores, and the top k sentences are selected as the summary. Score the importance of the sentence.

[0174]

[0175] (3) The topic awareness module is mainly divided into two parts: topic analysis and topic tracking. Topic analysis is primarily responsible for obtaining current dialogue context information from the short-term memory module and historical conversation information from the long-term memory module. It extracts core topics or keywords to support subsequent topic tracking, task allocation, and agent actions. Its corresponding modules are as follows: Figure 5 As shown. The specific operation process is as follows:

[0176] 1) Data Preprocessing. Dialogue information is obtained from the memory management module and transformed into structured data through text cleaning, word segmentation and magnetic tagging, and named entity recognition to prepare for subsequent analysis.

[0177] 2) Keyword Extraction. High-frequency and important keywords are extracted from the preprocessed text to serve as a preliminary representation of the topic. The specific steps are as follows:

[0178] Step 1: Calculate the TF-IDF score for each word and extract important keywords.

[0179] Step 2: Use Sentence-BERT to generate document embeddings and word embeddings, and calculate the cosine similarity between word embeddings and document embeddings. The calculation formula is as follows, where Representative word embedding, It is document embedding.

[0180]

[0181] Step 3: Select the words with the highest similarity as keywords.

[0182] Step 4: Combine the results of TF-IDF and Sentence-BERT to sort the keywords.

[0183] 3) Semantic Embedding Generation. If the acquired information is textual, it is converted into semantic embeddings to capture contextual semantics, providing a foundation for topic modeling and tool invocation. If the acquired information is multi-turn dialogue, a sliding window is used to generate embeddings for each turn of the dialogue.

[0184] 4) Topic modeling: Extract potential topics from the acquired text information or multi-turn dialogue information to generate topic distribution.

[0185] 5) External knowledge enhancement. The semantic retrieval tool in the tool module is invoked to retrieve relevant historical information from long-term memory, obtain the information with the highest similarity, supplement semantic information, and enhance the comprehensiveness of topic analysis.

[0186] 6) Use the relevance filtering tool in the tools module to calculate the relevance of keywords or topics to the current conversation. Filter out keywords with similarity below the threshold.

[0187] 7) Results Integration and Output. Keywords, topic distribution, and external information are merged to generate final topic information, and metadata such as timestamps and sources are added for easy tracking and storage. Finally, the integrated structured information is passed to the memory management and action planning management modules.

[0188] (4) Collaborative Planning and Management Module. Collaborative planning requires data from the action module, tool module, perception module, and memory module to coordinate the behavior of multiple agents and ensure the dynamic, adaptable, and efficient nature of the planning. Its core includes the following aspects:

[0189] 1) Task decomposition. Break down complex tasks into executable subtasks. Specific steps:

[0190] Step 1: Use semantic understanding to analyze the target and extract keywords.

[0191] Step 2: When analyzing the grammatical structure of the target sentence based on keywords, focus on the key part, noun phrases, and output the relationship between verbs and noun phrases.

[0192] Step 3: Based on the dependency analysis output, map verbs to action templates, and use noun phrases and keywords as objects to generate subtasks.

[0193] Step 4: Remove duplicates and perform relevance filtering. Remove duplicate or irrelevant subtasks to make the final subtasks more reasonable. Finally, generate the final subtasks.

[0194] 2) Agent allocation. This mainly involves allocating tasks based on the matching degree between the agent's capabilities and the sub-task requirements, to ensure that tasks are assigned to suitable agents.

[0195] 3) Dynamic Adjustment. Feedback is collected using a relevance filtering tool based on real-time acquisition of each round of speeches. The goal achievement is calculated based on the relevance score and task completion rate. If the achievement rate is lower than the target value, tasks are reassigned or new tasks are added. Achievement represents the task achievement rate, Ti represents a single task, Completed() represents the set of completed tasks, and Relevance(Ti) represents the relevance score of a single completed task. This indicates the total number of tasks.

[0196] The entire section is the sum of the relevance scores for all completed tasks.

[0197]

[0198] (5) Action Module. Action management is mainly used to coordinate the actions of various agents to serve the overall discussion goals and avoid ineffective or repetitive communication. It mainly includes the following parts:

[0199] 1) Deduplication mechanism. Using the relevance filtering tool in the tools module, the semantic similarity of the agent's responses is calculated, and responses below the threshold are filtered out.

[0200] 2) Conflict handling. Calculate the semantic similarity of the agents' statements to detect whether there are conflicts. When the agents' viewpoints conflict, an arbitration mechanism is introduced.

[0201] 3) Control the pace of the discussion. Recall historical information from long-term memory, combine it with topic tracking results, judge the rationality of viewpoints, guide the agent to focus on the current topic, and avoid going off-topic.

[0202] 2. Verify the consistency of the agent's personality traits.

[0203] To verify the correlation between single dimensions and behavior, and to prevent simultaneous changes in multiple personality dimensions leading to mutual influence, this experiment used controlled variables to control for the influence of other factors. Other irrelevant variables were controlled at a moderate level. Based on agent personality modeling, each dimension was divided into high and low levels. Two agents were created for each run: one with a high level and one with a low level. This was repeated 30 times, creating 60 agents per dimension, corresponding to 30 high-level and 30 low-level agents. A total of 300 agents were created across the five dimensions. Three topic scenarios were selected based on an open topic list: "How to solve the traffic problem in a city where 80% of residents rely on private cars but pollution exceeds standards by 50%?", "How to balance the contradiction between ecological protection and infrastructure construction?", and "How can artificial intelligence improve medical services (such as disease diagnosis and health management)?". Each agent ran independently to ensure separate statistical analysis of the five dimensions of the Big Five personality traits. This experiment used one-way ANOVA and correlation coefficients to verify the linear behavioral relationship between personality scores and behavioral performance under two personality combinations: low and high. The experiment verified five dimensions and behaviors, including: openness (O) and innovative behavior; extraversion (E) and proactive behavior; agreeableness (A) and supportive behavior; conscientiousness (C) and summarizing behavior; and neuroticism (N) and negative emotions.

[0204] To verify the interaction effects among the three personality dimensions of the agent, the correlation between "conflict-making" and "innovative" behaviors was verified using the interaction combinations E+A+N and O+E+A. The correlation between "cooperative" and "leadership" behaviors was verified using the interaction combinations O+E+C and A+N+C.

[0205] In the first type of experiment, two interaction combinations, E+A+N and O+E+A, were selected to verify "conflictual" and "creative" behaviors. Each dimension had two levels, high and low. There were a total of 16 possible combinations of agent personality traits, with 8 combinations from each of the two types. Thirty experiments were conducted under each of the three open-ended topics. Each experiment involved three agents discussing an open-ended topic. The agent personality traits were randomly selected from the 16 possible combinations of E+A+N and O+E+A, and each round of topic discussion consisted of 12 rounds.

[0206] Another experiment selected "cooperation" and "leadership" as the validation targets. "Cooperation" refers to the degree to which an agent exhibits cooperative behavior in a group chat, such as supporting or agreeing with others. "Leadership" refers to the degree to which an agent exhibits leadership behavior in a group chat, such as actively guiding the discussion, offering directional suggestions, or coordinating other agents. Similarly, two levels were set for each agent: high and low. This experiment also conducted 30 trials in each of the three open-topic scenarios. Each trial involved three agents discussing an open topic. The agent's personality traits were randomly selected from 16 possible interaction combinations (O+E+C and A+N+C), and each trial consisted of 12 rounds of topic discussion.

[0207] I. Specific application areas or related products of this invention.

[0208] Based on the core technological advantages of this invention—"personality stability simulation, efficient collaborative management, and precise quantitative assessment"—its application areas cover multiple industries such as education, enterprise services, scientific research, and market research. It can be transformed into various standardized or customized products, such as simulated training for university debate competitions / critical thinking courses, cultivation of critical thinking in primary and secondary schools, and simulation of foreign language oral dialogue scenarios, improving teaching effectiveness and enhancing user experience. In the interdisciplinary field of artificial intelligence and social psychology, research on "the relationship between team structure and discussion effectiveness," experiments on "the impact of feedback mechanisms on collaborative efficiency" in organizational behavior, and exploration of "the diffusion patterns of viewpoints in different personality roles" in communication studies, computer simulation methods are used to achieve social science verification.

[0209] To analyze the impact of agent combinations with different personality traits on the results of open-topic discussions, this invention designed a series of experiments, creating six agents with different roles. To highlight the characteristics of different combinations and prevent excessive overlap in the personality traits of agents within each group, this experiment selected agents with different roles and personality traits to form different subgroups, such as complementary groups, conflict groups, and balanced groups, and used these four subgroups as experimental groups. Collaborative groups and single-agent groups were set up as control groups.

[0210] To ensure consistent experimental conditions across groups, all extraneous variables were controlled, and all agents were standardized in design. Aside from differences in agent personality traits, all other configuration information remained consistent, with a uniform response time window (within 60 seconds). Each output was limited to 150-250 words, and the same large model version and parameter settings were used. This experiment selected "Super City Transportation Solutions: Prioritizing Autonomous Driving or Optimizing Comprehensive Public Transportation Coverage?" as the unified discussion topic. To ensure each agent focused solely on the chat task, a neutral agent was assigned to each group to summarize and provide a neutral overview of the other agents' statements. This neutral agent's responsibilities were clearly defined by its prompts; it was not to participate in the discussion, nor to offer subjective evaluations or suggestions. This neutral agent was initialized separately from the other agents. The neutral agent was responsible for retrieving all agents' statements from the long-term memory, providing a comprehensive summary and generating a summary report. The system environment was reset before each group's experiment to ensure isolation of experimental conditions and eliminate the influence of other factors.

[0211] Each group conducted independent topic discussions. Since the results of a single experiment may be affected by randomness, to ensure numerical stability, each group conducted 5 topic discussions in this experiment. Finally, the quality scores of the summary reports generated from each discussion were recorded, and the average score of the 5 summary reports was calculated. The specific experimental groups are shown in Table 6 below.

[0212] Table 6 Role Grouping

[0213]

[0214] Figure 6 Radar chart comparing scores for each group and each dimension.

[0215] Experiments reveal that agents based on role combinations have a significant advantage in information richness. Complementary combinations show a prominent advantage in information richness, and execution-oriented role division is more conducive to improving overall performance. Innovators are a key driver of information richness but may cause topic deviation. Conservatives and coordinators excel in textual logic.

[0216] Table 7. Statistics of scores for each group and dimension

[0217]

[0218] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0219] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for simulating and analyzing open-topic discussion processes based on a large language model, characterized in that, Includes the following steps: Construct a multi-agent system with perception management, memory management, planning management, action management, and tool management; Each intelligent agent is given quantifiable personality parameters based on the Big Five personality traits, and a consistency constraint mechanism between personality and behavior is formed through single-dimensional verification and multi-dimensional interactive verification. After the agent completes the open topic discussion, a multi-dimensional quality evaluation model is constructed based on the semantic matching, logical coherence, information richness, language fluency and content credibility of the dynamic topic representation and the summary text representation to automatically score the discussion results. During the discussion process, a feedback agent with the ability to adjust the feedback intensity, control the feedback timing, and generate feedback content is introduced to dynamically intervene in the multi-agent discussion process. By comparing the quality scores of discussion results after feedback intervention with those without feedback, the causal relationship between feedback mechanisms, personality combination mechanisms, and discussion quality is analyzed, thereby achieving a systematic simulation and analysis of the process and results of open-topic discussions.

2. The method as described in claim 1, characterized in that, The personality consistency constraint mechanism includes: The five personality dimensions—extraversion, agreeableness, openness, neuroticism, and conscientiousness—were set to high, medium, and low levels, respectively. By controlling for individual personality dimensions while keeping other dimensions constant, the frequency of corresponding behaviors was statistically analyzed, and the effects of each dimension were verified using analysis of variance and correlation analysis. By setting up multiple personality dimension combinations and keeping irrelevant dimensions at a moderate level, the study statistically analyzed behavioral synergies and verified the impact of multi-dimensional interactions on conflict-prone, innovative, cooperative, and leadership behaviors, thereby verifying the consistency between personality and behavior.

3. The method as described in claim 1, characterized in that, The feedback agent includes: Feedback templates are set according to three intensity levels: weak, medium, and strong. Intervene in the discussion process at two different times: real-time round feedback and phase round feedback. Based on the content of the agent's speech in the current round or the last 3 rounds, conduct topic analysis, argument analysis and sentiment analysis to determine the priority feedback items; Feedback content is generated in the order of affirmation, guidance, and suggestions, and its length is controlled to guide the direction and quality of the discussion.

4. A method for evaluating the quality of a multi-agent open-topic discussion summary report using the method described in claim 1, characterized in that, include: The summary report text and dynamic topic text are converted into semantic vectors and their similarity is calculated to obtain a task relevance score; The mean and standard deviation of the semantic similarity between adjacent sentences are calculated and combined with the number of sentences to obtain a logical coherence score; The information richness score is obtained by calculating the overlap ratio of technical terms, term precision, term recall, and the proportion of deduplicated words. A language fluency score is obtained based on the proportion of syntactic dependency relations and the perplexity of the language model. Content credibility scores are obtained based on sentiment polarity bias and language model generation loss; The above five categories of scores are weighted and summed according to preset weights to obtain the overall quality score of the summary report.

5. The method as described in claim 4, characterized in that, The task relevance score is obtained by calculating the cosine similarity between the semantic vector of the summary report and the semantic vector of the dynamic topic.

6. The method as described in claim 4, characterized in that, The information richness score is obtained by weighted combination of technical term coverage ratio, technical term precision, technical term recall, and deduplication ratio.

7. A simulation and analysis system for open-topic discussion processes based on a large language model, characterized in that, include: The multi-agent management module is used to build multi-agent systems with perception management, memory management, planning management, action management, and tool management. The personality modeling and verification module is used to model the Big Five personality traits of an intelligent agent and verify the consistency between personality and behavior through single-dimensional and multi-dimensional interaction methods. The quality evaluation module is used to perform multi-dimensional text quality scoring on the summary reports generated by the agent. The feedback control module is used to generate feedback content and intervene in the discussion process according to feedback intensity, timing, and template. The results analysis module is used to analyze the relationship between different personality combinations, feedback strategies, and discussion quality.

8. The system as described in claim 7, characterized in that, The multi-agent management module includes a short-term memory unit and a long-term memory unit. The short-term memory unit is used to store the content of the most recent rounds of dialogue, while the long-term memory unit is used to store the entire discussion history to maintain contextual consistency.

9. The system as described in claim 7, characterized in that, The multi-agent management module includes a topic tracking unit, which continuously calculates the current discussion topic and monitors the degree of deviation from the discussion. When the deviation exceeds a threshold, the feedback control module is triggered to intervene.

10. The system as described in claim 7, characterized in that, The personality modeling and verification module supports configuring agents as innovators, executors, challengers, and coordinators, and forming different discussion groups according to complementary, conflicting, balanced, and synergistic combinations.

Images