Socialized intelligent agent and apparatus for opinion personalized guidance

Abstract

This invention discloses a social intelligent agent and device for personalized opinion guidance, comprising: constructing a multi-agent social opinion simulation environment; after a user posts a comment, quantitatively evaluating the comment based on position relevance, argument direction, value recognition, and expression tendency to obtain a user attitude score; if the score is below a threshold, an adaptive guidance module is triggered; constructing a dynamic memory pool to store dialogue data; encoding data features and clustering by mean to obtain different persuasion strategy categories; reviewing different categories of dialogue data using a large language model; adaptively learning targeted rhetoric, logic, and emotion from successful cases; extracting a reusable set of persuasion strategy prompts; the adaptive guidance module dynamically adapting to the user's personalized expression and cognitive characteristics based on this prompt set to achieve personalized guidance; updating the dynamic memory pool and optimizing agent performance after each iteration; re-guiding failed users to minimize the impact on the network environment.

Description

Technical Field

[0001] This invention relates to the field of social networks, and more particularly to a social intelligent agent and device for personalized opinion guidance.

[0002] Background Technology: With the rapid development of artificial intelligence, large language models have achieved considerable success in natural language processing and computer vision. These achievements have made it possible to build human-like text and decision-making-capable social intelligent agents on social media based on large language models. Today, social media has become a core information hub and a center for public opinion, playing a crucial role in shaping public perception and influencing social decision-making.

[0003] Users on social media platforms can take actions such as liking, commenting, and sharing. Specifically, users can express their opinions under a post on a social media platform, that is, post comments to express their thoughts. Different users will hold different views, and there will be instances where users fail to understand the core content of a post, or are "led astray" by others and make inappropriate remarks. This phenomenon can further influence the views of others, and in the long run, it may even trigger a public opinion crisis and have an impact on society.

[0004] However, existing research on social media simulation platforms largely focuses on predicting group behavior, such as consensus generation, opinion polarization, and herding effects, while lacking research on individual behavior, specifically on the consequences of users making inappropriate statements. Persuasive intervention, a cutting-edge technology that can change users' opinions through multi-round dialogue in social media simulation platforms, is of great significance for public opinion discovery and real-time control.

[0005] However, there are still three challenges in achieving persuasive intervention: (1) There is a lack of opinion guidance datasets required for user persuasive intervention. Existing datasets have relatively simple topics and fragmented content; (2) User opinions in social media simulation platforms are difficult to quantify and evaluate. Existing methods focus on coarse-grained classification (support or opposition), and it is difficult to capture the multidimensional characteristics of user attitudes (including differences in position relevance, argument direction, value recognition and expression tendency); (3) Existing methods use predefined persuasion strategies. The persuasion strategies are simple and difficult to adjust according to the user's real-time cognitive state, making it difficult to change the user's opinion.

[0006] To address the aforementioned challenges and achieve persuasive intervention in public opinion guidance—that is, engaging in multi-round dialogues with users to change their original viewpoints, reducing and eliminating inappropriate speech on social media as much as possible, nipping public opinion crises in the bud, purifying social media platforms, and maintaining social stability—the following steps are crucial. First, collecting authentic dialogue data reflecting user interactions around specific topics is essential. Simultaneously, an automated user opinion evaluation module is introduced. This module is triggered when users post comments on social media platforms, assessing whether the comments are misleading. If a comment is determined to be misleading or have potential impact, an adaptive guidance module is activated to conduct persuasive intervention based on public opinion. This module adjusts its persuasion strategy in real time based on user feedback, generating persuasive content and engaging in multi-round dialogues with users to clarify the core arguments of the topic and correct user misunderstandings. By combining authentic dialogue data with the real-time user opinion evaluation module and the adaptive guidance module, this system aims to reduce the spread of inappropriate speech, purify social media platforms, and maintain a stable online environment. Summary of the Invention

[0007] This invention provides a social intelligent agent and device for personalized opinion guidance. The invention aims to achieve persuasive intervention in public opinion guidance, that is, to engage in multi-round dialogues with users to change their original viewpoints, reduce and eliminate inappropriate remarks on social media as much as possible, nip public opinion crises in the bud, purify social media platforms, and maintain a stable online environment. See the description below for details:

[0008] Firstly, a social intelligent agent oriented towards personalized opinion guidance, the intelligent agent comprising:

[0009] After a user posts a comment, the comment is quantitatively evaluated from the perspectives of position relevance, argumentation direction, value recognition, and expression tendency. The user's attitude score is calculated and compared with a threshold to determine whether adaptive guidance should be triggered.

[0010] A dynamic memory pool for storing dialogue interaction data is constructed. Different persuasion strategy categories are obtained by feature encoding and mean clustering of the interaction data. A large language model is used to extract the persuasion strategies used by samples in different persuasion strategy categories into a set of reusable persuasion strategy prompts.

[0011] When the adaptive guidance receives user comments, it generates guidance content using different persuasion strategies based on a set of reusable persuasion strategy prompts. It then determines which persuasion strategy generates guidance content with a higher probability of success based on a Gaussian distribution of perplexity, and selects the guidance content with the highest probability of success as the final output of this module. The user then comments again based on the guidance content.

[0012] The agent optimization mechanism writes new interaction samples into a dynamic memory pool after each iteration, retrains the agent based on the data in the dynamic memory pool, optimizes the agent's performance, and guides users who have not been successfully guided again, minimizing the impact on the network environment.

[0013] The relevance of the position refers to assessing the degree and direction of the connection between user comments and the topic's position; the directionality of the argument refers to assessing the degree to which the argumentation logic in the user comments supports or opposes the topic's position.

[0014] The value recognition level refers to the degree to which user comments agree with the values ​​behind the topic; the expression tendency refers to the attitude tendency implied in the language expression.

[0015] The reusable persuasion strategy cue set is as follows:

[0016] The persuasion strategies used by samples in different persuasion strategy categories are extracted into a set of reusable persuasion strategy prompts using a large language model. The perplexity of each sample under different persuasion strategy categories is calculated, and after obtaining the perplexity of each sample under different categories, a Gaussian distribution model is performed.

[0017] The intelligent agent further includes:

[0018] We crawl initial data from posts and comments on specific topics on real-world short video platforms and initialize a social media simulation platform based on this initial data.

[0019] Second aspect, a social device for personalized opinion guidance, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method described in any of the first aspects.

[0020] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.

[0021] The beneficial effects of the technical solution provided by this invention are:

[0022] 1. This invention enables persuasive intervention against users in a social media simulation platform. The intelligent agent adaptively learns and selects persuasive strategies and generates persuasive content based on real-time user information, thereby achieving persuasive intervention against users with inappropriate views on social media platforms, timely detection and control of public opinion, and reduction of potential impact on the network environment.

[0023] 2. To address the challenge of the lack of opinion guidance datasets required for persuasive intervention, this invention constructs an opinion guidance dataset. The dataset contains metadata, video descriptions, comment texts, likes, and other information related to specific topics on real-world short video platforms, covering diverse and novel topics.

[0024] 3. This invention addresses the challenge of quantifying and evaluating user opinions on social media platforms by establishing a multi-dimensional and more granular method for quantifying opinion evaluation. It achieves accurate quantitative evaluation of user opinions from four dimensions: position relevance, argumentation direction, value recognition, and expression tendency.

[0025] 4. To address the difficulty of static persuasion strategies in existing inventions, this invention designs an adaptive guidance module that can adaptively select persuasion strategies. Specifically, it learns persuasion strategies from previous dialogue interaction data, performs Gaussian distribution modeling based on each persuasion strategy, and selects the persuasion content with the highest probability of success based on user information for output.

[0026] 5. Considering the possibility of failure in persuasion interventions against users, this invention introduces an agent optimization mechanism. After each iteration, the samples in the memory pool are updated, and the persuasion strategy is re-clustered and relearned. Data is also selected from the dynamic memory pool for retraining / fine-tuning to continuously improve the agent's persuasion ability. Subsequently, users who were not successfully guided before are guided again, so that users with inappropriate views on social media platforms can change their views as much as possible, minimizing the existence of inappropriate remarks and purifying social media network platforms. Attached Figure Description

[0027] Figure 1 A flowchart for a social intelligent agent that is guided by personalized viewpoints;

[0028] Figure 2 A framework diagram for a social intelligent agent that is guided by personalized viewpoints;

[0029] Figure 3 A flowchart for optimizing the mechanism for intelligent agents. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0031] To fill the gap in existing social media simulation platforms regarding individual user behavior research, this invention develops a social media simulation platform capable of persuasive intervention in public opinion guidance. By persuading users holding inappropriate views, this platform enables users to change their opinions, thereby allowing for the timely detection and control of public opinion and reducing harm to the online environment. To address the challenges in the background technology, this invention first constructs an opinion dataset, then builds a social media simulation platform based on this dataset, allowing users to like, comment, and share. Subsequently, based on the principles of conversational cooperation and cognitive psychology, a quantitative evaluation mechanism for user opinions is established from four dimensions, providing a more granular quantitative evaluation of user opinions. Next, an adaptive guidance module intelligent agent can generate persuasive content with the highest probability of successful guidance. Simultaneously, this invention deploys and implements an intelligent agent optimization mechanism to continuously improve the agent's persuasive ability, enabling it to guide users who were previously unsuccessful, ensuring that users holding inappropriate views change their opinions as much as possible, minimizing the presence of inappropriate speech, and minimizing the impact on the online environment.

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0033] Example 1

[0034] A social intelligent agent that guides personalized viewpoints, see [link / reference]. Figure 1 , Figure 2 and Figure 3 The method includes the following steps:

[0035] Step 101: The data acquisition module is used to crawl initial data such as posts and comments on specific topics from real-world short video platforms, and to initialize the OASIS platform based on the initial data. [2] This will make it a social media simulation platform with a TikTok-like style;

[0036] Step 102: The opinion evaluation module is used to quantitatively evaluate a user's comment from four dimensions after the user posts it: position relevance (evaluating the degree of relevance and directionality between the user's comment and the topic's position), argumentation directionality (evaluating the degree to which the user's argumentation logic supports or opposes the topic's position), value identification (evaluating the degree to which the user's comment identifies with the values ​​behind the topic), and expression tendency (evaluating the implicit attitude tendency in the language expression). The user's attitude score is calculated and compared with a threshold to determine whether to trigger the adaptive guidance module. If the user's attitude score is higher than the threshold, it indicates that the user does not need persuasion intervention, that is, the statement is normal and has no adverse effects. If the user's attitude score is lower than the threshold, it indicates that the user needs persuasion intervention, that is, the user holds inappropriate opinions.

[0037] Step 103: The adaptive guidance module constructs a dynamic memory pool to store dialogue interaction data. It encodes the interaction data features and uses mean clustering to obtain different persuasion strategy categories, each containing numerous samples. Then, the large language model GPT-4o is used to extract the persuasion strategies used by the samples in different persuasion strategy categories into a reusable persuasion strategy cue set. Next, the perplexity of each sample under different persuasion strategy categories is calculated, and Gaussian distribution modeling is performed after obtaining the perplexity of each sample under different categories. When the adaptive guidance module receives user comments, it generates guidance content using different persuasion strategies based on the reusable persuasion strategy cue set obtained above. Based on the Gaussian distribution of perplexity, it determines which persuasion strategy generates guidance content with a higher probability of success, selects the guidance content with the highest probability of success as the module's final output, and inputs it to the user. The user will then post comments based on this guidance content.

[0038] Step 104: Agent optimization mechanism. After each iteration, new interaction samples are written into the dynamic memory pool. The agent is retrained / fine-tuned based on the data in the dynamic memory pool to minimize the loss function and optimize the network parameters, thereby improving the persuasiveness of subsequent iterations. This allows the agent to guide users who were not successfully guided in the previous iterations again, ensuring that users with inappropriate views change their views as much as possible, minimizing the existence of inappropriate remarks and minimizing the impact on the network environment.

[0039] In summary, this embodiment of the invention achieves persuasive intervention guided by public opinion through steps 101-104. First, a social media simulation platform with a TikTok-like style is built based on initial data. On this simulation platform, a guiding agent is built, which can generate persuasive content based on persuasion strategies learned from previous dialogue interaction data, and engage in multiple rounds of dialogue with users who hold inappropriate views, as screened by the user opinion evaluation module, so that users change their original views. At the same time, considering the situation where user persuasion fails, an agent optimization mechanism is added to optimize the agent's performance, so that in the next iteration, it can guide users who were not successfully guided before, so that users with inappropriate views can change their views as much as possible, minimize the existence of inappropriate remarks, and minimize the impact on the network environment.

[0040] Example 2

[0041] The following section uses specific calculation formulas... Figure 1 , Figure 2 and Figure 3 The following example further illustrates the solution in Embodiment 1, as detailed in the description below:

[0042] 201: Based on scheduled web crawling or calling the public API of the target short video platform to obtain information such as metadata of posts related to a specific topic, video descriptions, comment text, and number of likes, a data set for constructing opinion-guided discussions around a specific topic is built; based on the opinion-guided dataset, a social media simulation platform is initialized, and a social media simulation platform with a TikTok style is constructed, on which users can like, comment, and forward; the user commenting process in the social media simulation platform is defined as follows in this embodiment of the invention:

[0043]

[0044] In this embodiment of the invention, a large language model is used for user simulation, and the user's large language model is parameterized as follows: and expressed as ; This refers to user comments generated during the interaction process; This indicates the prompt words input to the large model; This represents the content of the user profile; This indicates the content of the post; Indicates the sequence length; This indicates the words or phrases being generated at the current moment. This indicates words that have already been generated.

[0045] 202: After a user comments on the platform, the opinion evaluation module is automatically triggered to quantitatively evaluate the comment from four dimensions: relevance of stance, direction of argumentation, value consensus, and expression bias. This process is defined as follows:

[0046]

[0047]

[0048] In this embodiment of the invention, a large language model is used to quantitatively evaluate user opinions, and the evaluation large language model is parameterized as follows: and expressed as , Indicates the user to be evaluated In time The Dimensional support score, ; This indicates the prompt words input to the large model; in this embodiment of the invention, the four dimensions are summarized into four scoring rules. This is used for fine-grained evaluation of each dimension; in this embodiment of the invention, the quantified scores of the four dimensions are fused. The score indicates the user's final attitude and support.

[0049] 203: If the user's final opinion quantification score is less than the threshold T, a persuasion intervention is triggered, and the user enters the persuasion intervention interface. The persuasion intervention of the agent in the social media simulation platform is defined as follows:

[0050]

[0051] In this embodiment of the invention, a large language model is used for persuasive intervention, and the guiding large language model is parameterized as follows: and expressed as ; This refers to the guiding content generated during the interaction; This indicates the prompt words input to the large model; This indicates a record of previous interactions with the user; This indicates the words or phrases being generated at the current moment. This indicates previously generated words. 204: The persuasive content used by the agent during the persuasion intervention process is generated based on a certain persuasion strategy, which is obtained from previous dialogue interaction data. In the persuasion strategy learning and selection stage, this embodiment of the invention designs a dynamic memory pool to store the dialogue interaction data between the user and the agent. Subsequently, this embodiment of the invention filters the dialogue interaction data stored in the dynamic memory pool to obtain a triple consisting of <user profile-user comment-persuasion content>. Then, the triple consisting of <user profile-user comment-persuasion content> is feature-encoded, wherein the user comment sequence is... , Indicates the sequence length; the sequence of guiding content generated by the agent is , Indicates sequence strength; user profile sequence is , Indicates the sequence length; subsequently, embodiments of the present invention address the sequence. and Using the SBERT model for vector encoding, comment sequence vectors are obtained respectively. and guiding content sequence vector The embodiments of the present invention address the sequence. One-hot encoding is performed to obtain the user profile sequence vector. Vector; where, and Indicates the length of the vector;

[0052] Next, in this embodiment of the invention, the three vectors obtained in the above steps are weighted and fused to obtain a fused vector. ,in This represents the weight of the vector, and The system encodes all triples stored in the dynamic memory pool to obtain a fusion vector set. =[ ], where M represents the number of elements in the set; subsequently, the embodiments of the present invention cluster the above-obtained fusion vector set to obtain different persuasion strategy categories, the process of which is defined as follows:

[0053]

[0054] in, Represents the set of fusion vectors The elements in Indicates the cluster center. The number of persuasion strategy categories is indicated; subsequently, in this embodiment of the invention, the large language model GPT-4o is used to analyze and summarize the different persuasion strategy categories obtained by clustering, and a persuasion strategy cue set is obtained.

[0055] 205: To enable the agent to adaptively generate guiding content based on persuasion strategies, this embodiment of the invention models the different persuasion strategy categories obtained from clustering using a Gaussian distribution; firstly, it calculates the perplexity (PPL) of the samples contained in each persuasion strategy category, defined as follows:

[0056]

[0057] Furthermore, in this embodiment of the invention, after obtaining the specific values ​​of the perplexity of each sample under each strategy category, a Gaussian distribution model is performed:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] in, Indicates the first Persuasion strategies; Represent a binary variable, Indicates the first The number of samples that successfully guided a particular persuasion strategy. Indicates the first The number of samples where the persuasion strategy failed.

[0065] Furthermore, based on Gaussian distribution modeling and Bayes' theorem, the probability of successful persuasion under different perplexity PPL values ​​can be calculated:

[0066]

[0067] When user information is input into the agent, the agent generates a persuasive content based on each persuasion strategy and calculates the PPL value of each persuasive content. Then, based on Gaussian distribution modeling, it calculates the success probability of the persuasive content generated by different persuasion strategies, selects the persuasive content with the highest probability as the final output, and sends it to the persuasion intervention interface.

[0068] 206: The user interacts with the agent in multiple rounds, and the user comments again based on the persuasive content generated by the agent. Finally, the opinion evaluation module is triggered again, and persuasive interventions may continue.

[0069] 207: In each iteration, due to insufficient sample size in the memory pool and noise interference, persuasion intervention may fail. Therefore, the samples in the current dynamic memory pool are screened for quality and the agent is retrained / fine-tuned to improve the persuasion ability of subsequent iterations. This is then used in the next iteration to guide users who were not successfully guided before, so that users with inappropriate views can change their views as much as possible, minimizing the existence of inappropriate remarks and minimizing the impact on the network environment.

[0070] 208: Regarding step 207, during the quality screening process, for users who require guidance, the initial user comments... The corresponding attitude score is The user is guided by the intelligent agent. The comment that was published again is And the comments published again The corresponding attitude score is If the degree of change of the fraction Greater than the threshold In other words, the guiding content generated by the intelligent agent. If it can significantly improve users' opinions, then... and These are combined to form a training dataset.

[0071] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0072] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0073] Example 3

[0074] A social networking device for personalized opinion guidance includes a processor and a memory. The memory stores program instructions, and the processor invokes the program instructions stored in the memory to cause the device to perform the following method steps in Embodiment 1:

[0075] After a user posts a comment, the comment is quantitatively evaluated from the perspectives of position relevance, argumentation direction, value recognition, and expression tendency. The user's attitude score is calculated and compared with a threshold to determine whether adaptive guidance should be triggered.

[0076] A dynamic memory pool for storing dialogue interaction data is constructed, and different persuasion strategy categories are obtained by encoding the features of the interaction data and clustering by mean.

[0077] When the adaptive guidance receives user comments, it generates guidance content using different persuasion strategies based on a set of reusable persuasion strategy prompts. It then determines which persuasion strategy generates guidance content with a higher probability of success based on a Gaussian distribution of perplexity, and selects the guidance content with the highest probability of success as the final output of this module. The user then comments again based on the guidance content.

[0078] The agent optimization mechanism writes new interaction samples into a dynamic memory pool after each iteration, retrains the agent based on the data in the dynamic memory pool, optimizes the agent's performance, and guides users who have not been successfully guided again, minimizing the impact on the network environment.

[0079] Among them, position relevance refers to: assessing the degree of correlation and directionality between user comments and the topic position; the directionality of argumentation refers to assessing the degree to which the argumentation logic in user comments supports or opposes the topic position.

[0080] Among them, value recognition is: assessing the degree to which user comments agree with the values ​​behind the topic; the expression tendency is assessing the attitude tendency implied in the language expression.

[0081] The set of reusable persuasion strategy prompts includes:

[0082] The persuasion strategies used by samples in different persuasion strategy categories are extracted into a set of reusable persuasion strategy prompts using a large language model. The perplexity of each sample under different persuasion strategy categories is calculated, and after obtaining the perplexity of each sample under different categories, a Gaussian distribution model is performed.

[0083] Intelligent agents also include:

[0084] We crawl initial data from posts and comments on specific topics on real-world short video platforms and initialize a social media simulation platform based on this initial data.

[0085] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0086] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0087] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0088] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0089] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0090] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0091] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0092] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0093] References:

[0094] [1]Lu J, Yu H, Xu S, et al. Prolonged reasoning is not all you need: Certainty-based adaptive routing for efficient llm / mllm reasoning[J]. arXivpreprint arXiv:2505.15154, 2025.

[0095] [2]Yang Z, Zhang Z, Zheng Z, et al. OASIS: Open Agents SocialInteraction Simulations on a Large Scale[J].

[0096] [3]Chen M, Guo B, Wang H, et al. The future of cognitive strategy-enhanced persuasive dialogue agents: new perspectives and trends[J]. Frontiers of Computer Science, 2025, 19(5): 195315.

[0097] [4]Grice H P. Logic and conversation: The William James lectures[J].Unpublished manuscript, Harvard, 1968.

[0098] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0099] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A social intelligent agent oriented towards personalized opinion guidance, characterized in that, The intelligent agent includes: After a user posts a comment, the comment is quantitatively evaluated from the perspectives of position relevance, argumentation direction, value recognition, and expression tendency. The user's attitude score is calculated and compared with a threshold to determine whether to trigger the adaptive guidance module. A dynamic memory pool for storing dialogue interaction data is constructed. Different persuasion strategy categories are obtained by feature encoding and mean clustering of the interaction data. A large language model is used to extract the persuasion strategies used by samples in different persuasion strategy categories into a set of reusable persuasion strategy prompts. After calculating the perplexity of each sample under different persuasion strategy categories, a Gaussian distribution model is performed. When the adaptive guidance module receives user comments, it generates guidance content using different persuasion strategies based on a set of reusable persuasion strategy prompts. Based on the Gaussian distribution of perplexity, it determines which persuasion strategy generates guidance content with a higher probability of success. The guidance content with the highest probability of success is selected as the final output of the module and input to the user. The user will then post comments based on this guidance content. The agent optimization mechanism writes new interaction samples into the dynamic memory pool after each iteration, retrains the agent based on the data in the dynamic memory pool, optimizes the agent's performance, and guides users who have not been successfully guided again, minimizing the impact on the network environment. The interaction data feature encoding and mean clustering are as follows: feature encoding is performed on the triples composed of <user profile-user comment-persuasive content>, and the user comment sequence is... , Indicates the sequence length; the sequence of guiding content generated by the agent is , Indicates sequence strength; user profile sequence is , Indicates the sequence length; for the sequence and Vector encoding is performed to obtain comment sequence vectors. and guiding content sequence vector ; For sequences One-hot encoding is performed to obtain the user profile sequence vector. ; and The vector length is represented by a weighted sum of the three vectors to obtain the fused vector. ,in This represents the weight of the vector, and The system encodes all triples stored in the dynamic memory pool to obtain a fusion vector set. =[ ], where M represents the number of elements in the set; clustering the fusion vector set yields different persuasion strategy categories; Retraining the agent based on dynamic memory pool data involves: quality screening of samples in the current dynamic memory pool and retraining / fine-tuning the agent for use in the next iteration; and re-guiding users who were previously unsuccessful. During quality screening, for users requiring guidance, the initial user comments... The corresponding attitude score is The user is guided by the intelligent agent. The comment that was published again is And the comments published again The corresponding attitude score is If the degree of change of the fraction Greater than the threshold Guide content generated by the intelligent agent If it can improve the user's perspective, then... and These are combined to form a training dataset.

2. The social intelligent agent for personalized opinion guidance according to claim 1, characterized in that, The relevance of the position refers to assessing the degree and direction of the connection between user comments and the topic's position; the directionality of the argument refers to assessing the degree to which the argumentation logic in the user comments supports or opposes the topic's position.

3. The social intelligent agent for personalized opinion guidance according to claim 1, characterized in that, The value recognition level is defined as the degree to which user comments agree with the values ​​behind the topic; the expression tendency is defined as the attitude tendency implied in the language expression.

4. A social intelligent agent for personalized opinion guidance according to claim 1, characterized in that, The set of reusable persuasion strategy prompts is as follows: The persuasion strategies used by samples in different persuasion strategy categories are extracted into a set of reusable persuasion strategy prompts using a large language model. The perplexity of each sample under different persuasion strategy categories is calculated, and after obtaining the perplexity of each sample under different categories, a Gaussian distribution model is performed.

5. A social intelligent agent for personalized opinion guidance according to claim 1, characterized in that, Based on this, the intelligent agent further includes: We crawl initial data from posts and comments on specific topics on real-world short video platforms and initialize a social media simulation platform based on this initial data.

6. A social device for personalized opinion guidance, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-5.

Images