Network violent behavior prediction method and system

Abstract

The invention provides a network violent behavior prediction method and system. The method comprises the following steps: acquiring input data corresponding to a prediction task; wherein the input data corresponding to the prediction task comprises user-corresponding portrait features obtained based on user social data, the user-corresponding portrait features comprise base row portrait features, psychological portrait features of a mechanism center and psychological portrait features of an attribute center, and the user social data comprises user content footprints and user content interaction behaviors; and inputting the obtained input data, the to-be-discussed topic scene information and a preset task request corresponding to the prediction task into a large language model, and outputting to obtain the possibility that the user has a violent behavior under the to-be-discussed topic scene information. According to the method, the hidden network violent behaviors can be accurately identified, and the user portraits are constructed based on multi-level analysis of the user social data, so that accurate prediction of the network violent behaviors is realized by combining the user portraits and the large language model.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for predicting cyberbullying behavior. Background Technology

[0002] With the widespread use of the internet, cyberbullying has become a global social problem. A survey of 3,591 respondents nationwide, as reported in the "Report on the Governance of Cyberbullying," revealed that over 60% of respondents claimed to have experienced cyberbullying; the "2020 National Research Report on Internet Use by Minors" indicated that 19.5% of minors reported experiencing cyberbullying. These data demonstrate that cyberbullying not only affects adults but also poses a serious threat to minors.

[0003] Cyberbullying typically manifests as excessive verbal abuse, and filtering such speech through innovative internet technologies has become a research hotspot. Current research on the identification and prediction of cyberbullying primarily includes the following two technological paradigms: The first paradigm is the traditional classification paradigm based on sensitive words and feature engineering. The sensitive word filtering system in this paradigm simplifies the identification of cyberbullying to literal scanning and matching: it compares online speech samples character by character with a static dictionary text maintained by the system; once a match is found, the identified text is marked as violent content. This approach works in explicit abusive scenarios, but it is completely ineffective in handling variations such as metaphors, irony, and homophones. For example, under this traditional paradigm, phrases like "kill the game" in sports commentary and the popular internet slang "laughing myself to death" are often misjudged as violent content, while seemingly mild but actually offensive expressions like "send you to the top of the trending topics" easily pass the review. Although traditional machine learning frameworks attempt to compensate for the shortcomings of sensitive word matching by using statistical features such as Term Frequency-Inverse Document Frequency (TF-IDF) or N-gram, their improvement schemes still remain at the "bag of characters" level: using TF-IDF or N-gram to extract statistical features from online speech samples, and then using Support Vector Machine (SVM), Random Forest or Naive Bayes to perform binary classification on the extracted statistical features. Although this type of improvement method has low computational cost, it also only stays at the surface vocabulary level.

[0004] Early social media platforms like Twitter and Weibo adopted the first paradigm for building cyberbullying detection systems. This involved first weighting words using sentiment dictionaries like NTUSD, and then setting thresholds to trigger violence flags. Cyberbullying detection systems built on this first paradigm exhibit poor adaptability to different social media platforms: the binary classifier in such systems requires tens of thousands of manually labeled samples for training. Once migrated to a new social media platform or a new trending event, performance drops drastically. For example, a classifier trained on platform A may see its accuracy drop by more than 30% on platform B. Furthermore, the first paradigm suffers from both false positives and false negatives, forcing platform operators to invest heavily in manual verification, resulting in high costs and a detrimental user experience.

[0005] The second paradigm is the direct application of large language models. Existing technical documents 1, 2, and 3 all point out that the zero-shot or few-shot capabilities of GPT / BERT can be used to identify violent content. For example, the S3 framework proposed in existing technical document 4 and RecAgent in existing technical document 3 both use large language models (LLMs) as the core for sentiment and behavior generation, with text-level F1 scores of approximately 0.60 on public datasets. However, this paradigm treats LLMs merely as "black box" classifiers, typically directly inputting user-related data into the LLM for analysis, lacking targeted and hierarchical analysis. Furthermore, while large language models possess semantic understanding capabilities, they exhibit a clear "moderation tendency" in violent scenarios. Experiments show that when large language models are asked to simulate the language of real perpetrators, their output is often overly rational, significantly differing from the intense language of real social media platforms. In addition, each inference by the LLM requires cloud computing power, and the cost accumulates rapidly when there is too much input data (e.g., in scenarios with tens of thousands of concurrent users).

[0006] Existing technologies either rely on literal matching or are trapped in a black box, resulting in an accuracy rate of around 60% in identifying cyberbullying in practical applications, a recall rate of less than 50%, and a persistently high false positive rate. Existing technologies for cyberbullying suffer from the following main shortcomings: ① Cyberbullying is typically highly personalized and contextualized, but existing technologies lack long-term memory of individual historical behavior, making it impossible to establish differentiated violent tendency profiles, hindering personalized early warning, and easily leading to numerous false positives and missed detections. ② As the forms of cyberbullying continue to evolve, the boundary between cyberbullying and non-cyberbullying is becoming increasingly blurred. New, covert forms of cyberbullying (such as "insulting slang") are emerging in an endless stream. Researchers have found, based on numerous online comment samples, that perpetrators of cyberbullying often operate in the shadows, and existing technologies (especially the first paradigm) struggle to accurately identify these covert behaviors, such as metaphors, irony, and homophonic variations. ③ Existing technologies are mainly limited to historical statements themselves. For example, the first paradigm relies on the surface lexical features of the text, and the second paradigm relies on the semantic features of the text. They do not consider the impact of multi-dimensional variables on users' cyberbullying behavior. ④ Existing technologies are all post-event identification, lacking an early warning window, and are prone to missing the golden opportunity for intervention.

[0007] Therefore, there is an urgent need for a method that can dynamically and accurately predict users' cyberbullying behavior based on personalized user profiles containing multi-dimensional features, thereby achieving a technological leap from "post-event handling" to "pre-event prevention" while improving the low accuracy of existing cyberbullying behavior identification.

[0008] Existing technical documents Existing technical literature 1: Wang L, Ma C, Feng X, et al. A survey on large language model based autonomous agents[J]. Frontiers of Computer Science, 2024, 18(6):186345. Prior art document 2: Gao C, Lan X, Li N, et al. Large language modelsempowered agent-based modeling and simulation: A survey and perspectives[J]. Humanities and Social Sciences Communications, 2024, 11(1): 1-24. Prior art document 3: Zhang A, Chen Y, Sheng L, et al. On generative agents in recommendation [C]. Proceedings of the 47th international ACM SIGIRconference on research and development in Information Retrieval. 2024: 1807-1817. Existing technical document 4: Gao C, Lan X, Lu Z, et al. S3: Social-network Simulation System with Large Language Model-Empowered Agents[J]. arXivpreprint arXiv:2307.14984, 2023. Summary of the Invention

[0009] In view of this, embodiments of the present invention provide a method and system for predicting cyberbullying behavior, which can not only accurately identify hidden cyberbullying behavior and improve the identification accuracy, but also construct user profiles based on multi-level analysis of user social data, thereby combining user profiles and large language models to achieve accurate prediction of cyberbullying behavior.

[0010] One aspect of the present invention provides a method for predicting cyberbullying behavior, the method comprising the following steps: Obtain the input data corresponding to the prediction task; wherein, the input data corresponding to the prediction task includes the user profile features obtained based on user social data, the user profile features include basic profile features, mechanism-centered psychological profile features and attribute-centered psychological profile features, and user social data includes user content footprints and user content interaction behaviors; The acquired input data, the topic and scenario information to be discussed, and the preset task request corresponding to the prediction task are input into the large language model, and the output is the probability of a user engaging in cyberbullying behavior under the topic and scenario information to be discussed.

[0011] In some embodiments of the present invention, the user content footprint includes the user's posting and commenting records, and the user content interaction behavior includes posting behavior, commenting behavior, liking behavior, forwarding behavior, and content blocking behavior; User social data also includes user profiles, and the corresponding user profile features also include account-side profile features determined based on the user profiles. Account-side profile features include account registration duration and / or account authentication status.

[0012] In some embodiments of the present invention, the base row portrait features are obtained in the following manner: Obtain the input data corresponding to the basic bank profile task; the input data for the basic bank profile task is obtained based on user social data; The input data corresponding to the basic profile task and the preset task request corresponding to the basic profile task are input into the large language model, and the basic profile features are output. The basic profile features include the user's demographic features and social interaction features. The demographic features include one or more of age group, gender, education level and occupation, and the social interaction features include one or more of social activity, interaction intensity, social influence and interaction behavior selection distribution. The preset task request corresponding to the basic profile task is used to indicate the information dimensions contained in the basic profile features output by the large language model.

[0013] In some embodiments of the present invention, the psychological profile features of the mechanism center are obtained through the following methods: The input data for the acquisition mechanism-psychological profiling task is obtained based on user social data and basic profile features. The acquired input data and the preset task request corresponding to the mechanism-mental profiling task are input into the large language model, and the output is the mechanism-centered mental profiling features. Among them, the psychological profile characteristics of the mechanism center include aggressive characteristics, role type characteristics of cyberbullying incidents, and psychological mechanism variables corresponding to cyberbullying behavior; the information dimension contained in the aggressive characteristics is determined based on the aggressiveness theory related to cyberbullying, and the psychological mechanism variables corresponding to cyberbullying behavior include risk factors and protective factors, social domination orientation characteristics, and negative personality tendencies characteristics. The information dimension contained in the negative personality tendencies characteristics is determined based on the dark personality theory.

[0014] In some embodiments of the present invention, the psychological profile features of the mechanism center also include moral foundation features, empathy features, and personality difference features; the information dimensions contained in the moral foundation features are determined based on moral foundation theory, the information dimensions contained in the empathy features are determined based on empathy theory, and the information dimensions contained in the personality difference features are determined based on the Big Five personality theory.

[0015] In some embodiments of the present invention, the psychological profile features of the attribute center are obtained in the following ways: Obtain the input data corresponding to the attribute-psychological profiling task; wherein, the input data corresponding to the attribute-psychological profiling task is obtained based on user social data and basic profile features; The acquired input data and the preset task request corresponding to the attribute-mental profiling task are input into the large language model, and the output is the attribute-centered mental profiling features. Among them, the psychological profile features of the attribute center include theme focus features, value factor features, cognitive style features, lifestyle-related indicator features, and language style features; theme focus features include one or more of the following: theme preference distribution, valence emotion tendency for each theme, and stance firmness; language style features include the frequency of use of words related to cyber violence, syntactic features, and the frequency of use of non-literal rhetoric.

[0016] In some embodiments of the present invention, the psychological profile features of the attribute center also include one or more of the process-function dimension features of emotions and the conflict handling style features. The process-function dimension features of emotions include information on the dimensions of emotional awareness, emotional intelligence, emotional resilience, emotional expression and / or emotional connection. The information dimension included in the conflict handling style features is determined based on the conflict handling model theory. Language style characteristics also include one or more of the following: the degree of expression of social emotions related to cyberbullying, the formality of language, the degree of humor, and the frequency of stereotypes.

[0017] In some embodiments of the present invention, the input data for the mechanism-psychological profiling task includes some or all of the social data and demographic features; the input data for the attribute-psychological profiling task includes some or all of the social data and basic profiling features.

[0018] In some embodiments of the present invention, after inputting the input data corresponding to the prediction task into a large language model, the method further includes: outputting the possible statements made by the user under the topic scenario information; The preset task request corresponding to the prediction task is used to instruct the large language model to analyze the user's vocabulary usage patterns based on the user's content footprint, and to predict the possibility of cyberbullying behavior and possible speech based on the analysis results of vocabulary usage patterns, the user's corresponding profile features, and the topic scenario information to be discussed; among them, the analysis results of vocabulary usage patterns include vocabulary usage frequency, usage of offensive words, usage of emotional expression words, and the user's personalized expression patterns. The topic context information includes one or more of the following: topic title, topic description, topic controversy level, and topic keywords.

[0019] Another aspect of the present invention provides a network behavior prediction system, including a processor, a memory, and a computer program / instructions stored in the memory, wherein the processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method described in any of the above embodiments.

[0020] The proposed method and system for predicting cyberbullying behavior can perform multi-level analysis based on user social data (mainly including users' historical statements) to obtain basic user profile features, mechanism-centered psychological profile features, and attribute-centered psychological profile features. These user profile features can not only distinguish individual differences among different users, but also effectively reduce existing false positive and false negative rates. Furthermore, by combining the natural semantic understanding capabilities of a large language model with user profile features, it is possible to accurately predict potential cyberbullying behavior by users when facing social events.

[0021] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0022] Those skilled in the art will understand that the objectives and advantages achievable with this invention are not limited to those specifically described above, and that the above and other objectives achievable with this invention will become clearer from the following detailed description. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a flowchart illustrating a method for predicting cyberbullying behavior in one embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the overall model framework of the method proposed in this application in one embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0026] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0027] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0028] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0029] Cyberbullying has long existed on major social media platforms, and its forms continue to evolve with trending events. Although there are many existing methods for detecting and identifying cyberbullying, most of them are built using sensitive word-feature engineering classification paradigms and direct application paradigms of large language models. The core contradictions in the development of existing technologies are as follows: ① Existing technologies based on the sensitive word-feature engineering classification paradigm and the direct application paradigm of large language models are limited to the text itself. For example, sensitive word matching technology relies on the surface lexical features of the text, and the application of large language models relies on the deep semantic features of the text. The reliance on single-dimensional modeling of the text leads to insufficient comprehensiveness and accuracy in the analysis of users' cyberbullying behavior; ② Slang and suggestive violent language are emerging in an endless stream. Traditional literal matching and statistical features are ineffective against such "invisible attacks," that is, it is difficult to identify hidden cyberbullying behaviors (such as metaphors, homophones, and irony); ③ Cyberbullying behavior presents highly personalized and contextualized characteristics, while existing cyberbullying identification technologies uniformly analyze user speech and cannot effectively distinguish the individual differences of different users, resulting in a large number of false positives and false negatives; ④ Existing technologies do not screen the input data of cyberbullying analysis models and only regard violence discrimination as an "end-to-end" black box classification, lacking targeted and hierarchical analysis.

[0030] To overcome the severe challenges faced by existing technologies in combating cyberbullying, this application proposes a method for predicting cyberbullying speech. It establishes a multi-dimensional display model based on a multi-layered analytical architecture, encompassing "user social data - basic user profile - psychological profile - contextual prediction." This addresses the shortcomings in comprehensiveness and accuracy caused by the single-dimensional text modeling of existing solutions. Furthermore, multi-dimensional user profile modeling effectively distinguishes individual differences among users, improving the accuracy and relevance of input data for network analysis models. The proposed method utilizes the semantic understanding capabilities of LLM and user profile features to determine the true meaning of user statements or behaviors, enabling more accurate identification of hidden violent expressions such as metaphors, irony, and homophones, thus improving the accuracy of cyberbullying identification and reducing false positive and false negative rates. In addition, this application leverages a large language model to predict users' cyberbullying behavior in response to social events, achieving a leap in thinking from "cyberbullying identification" to "cyberbullying prediction."

[0031] Furthermore, users may exhibit varying probabilities of participation and aggressive tendencies in response to different trending events. Therefore, this application proposes using a large language model (LLM) to predict user cyberbullying behavior across different topical contexts, thereby improving users' adaptability to dynamic topical situations and achieving dynamic behavior prediction. In addition, existing cyberbullying identification technologies generally remain at the level of "providing results," lacking explanations of the credibility of the results and the generation process, resulting in practical effects far below experimental metrics. Therefore, the method proposed in this application, when profiling users, can utilize LLM to provide clear evidence for each inference result (e.g., the user's representative historical statements and the confidence level of the inference), thus providing decision-making support for platform operators or regulators.

[0032] In some embodiments of this invention, this application proposes a method for predicting cyberbullying behavior based on multi-dimensional feature analysis and large language modeling. This method uses user social data analysis to perform basic profile and psychological profile features to build user profiles, and utilizes LLM to predict users' participation levels and violent behavior tendencies in specific topic contexts, in order to distinguish individual differences among users. Specifically, the core tasks of the proposed method include basic profile construction, mechanism-centered psychological profile construction, attribute-centered psychological profile construction, and prediction of cyberbullying behavior. Figure 1 As shown, the cyberbullying behavior prediction method proposed in this application specifically includes: Step S110, obtaining the input data corresponding to the prediction task. Step S120, inputting the obtained input data, the topic scenario information to be discussed, and the preset task request corresponding to the prediction task into a large language model, and outputting the probability of a user engaging in cyberbullying behavior under the topic scenario information to be discussed. The input data corresponding to the prediction task includes user profile features obtained based on user social data, and these user profile features include basic profile features, mechanism-centered psychological profile features, and attribute-centered psychological profile features.

[0033] As an example, the overall model framework of the method proposed in this application can be divided into three main stages: data acquisition, profiling and analysis, and prediction. Figure 2 As shown, each stage of the analysis module mines user social data from different perspectives, thereby providing accurate information support for the final prediction and intervention of violent behavior. Specifically, the input data stage receives user social data; the profiling analysis stage analyzes user social data across three levels: basic profile, mechanism-centered psychological profile, and attribute-centered psychological profile, obtaining corresponding user profile features and achieving in-depth processing of user social data; the prediction stage includes steps S110 and S120, which combine the three-layer user profile analysis to output prediction results such as user participation level, probability of violent behavior, and specific statements, and can further provide interpretability analysis.

[0034] In some embodiments of the present invention, user social data can be obtained based on raw user data from social platforms to ensure that subsequent analysis is based on real, accurate, and comprehensive data. The raw user data mentioned in this application may be publicly available data from the platform.

[0035] Specifically, user social data refers to all digital traces of "who a user interacts with, how they interact, what they say, and where they are on the network," encompassing all information that can characterize "user-other interaction relationships and content." The user social data mentioned in this application includes at least user content footprints, user content engagement behaviors, and user profiles. User content footprints describe all publicly available text or multimedia outputs by users on the platform (e.g., user posting and commenting records). User content engagement behaviors include passive feedback behaviors (e.g., liking and blocking content) and active output behaviors (e.g., posting, forwarding, and commenting), referring to all actions that connect users with platform content. User profiles can be user identification information that distinguishes different users on the social platform, including account-level information (account ID, registration time, registration IP, etc.), basic-level information (nickname, avatar, gender, phone number, and email, etc.), and extended-level information (personal signature, interest tags, and real-name information, etc.). In addition, user social data may also include relationship chain information containing fields such as following / follower lists, group tags, and intimacy scores, as well as network topology information containing fields such as bridging coefficients, influence rankings, and group tags. The user social data and user profiles used for user profiling in this application are open-source data and do not involve any undisclosed user privacy data.

[0036] As an example, the user profile features may include account-side profile features determined based on the user's personal information. Account-side profile features may include one or more of the following: account registration duration, account authentication status, and profile completeness. Furthermore, the account-side profile features (account_features) output by the large language model can be represented in the following format: { "account_features": { "account_age_days": 245, "verification_status": false, "profile_completeness": 0.75 } } Among them, the account registration duration (account_age_days) and profile completeness (profile_completeness) can be represented by numerical labels, while the account verification status (verification_status) can be represented by boolean labels. The format of the account profile features mentioned above is only an example, and the present invention is not limited thereto.

[0037] Furthermore, user social data can be obtained by efficiently crawling the original user data corresponding to each user on the social platform using web crawling tools, providing rich raw input for subsequent user profile construction and violent behavior prediction; user social data can also be directly obtained from the original user data corresponding to each user stored on the social platform through API interfaces. This application does not specifically limit the method of obtaining user social data. In addition, after obtaining data from the original user data on the social platform, the method further includes: cleaning and integrating the obtained data to obtain user social data. Among them, data cleaning includes removing redundant data and handling missing values, and data integration includes integrating multi-dimensional user social data into a unified standardized data format.

[0038] In some embodiments of the present invention, the user profile is divided into a basic profile, a mechanism-centered psychological profile, and an attribute-centered psychological profile. The user profile process does not involve the prediction of specific events and behaviors. The goal is to accumulate stable and interpretable user characteristics to provide reliable input for the prediction of cyberbullying behavior.

[0039] More specifically, statistical analysis and basic feature calculation based on user social data can yield base profile features containing multi-dimensional information. That is, base profile features can be obtained through the following methods: acquiring the input data corresponding to the base profile task; inputting the input data and the preset task request corresponding to the base profile task into a large language model, and outputting the base profile features. The input data for the base profile task is obtained based on user social data; the purpose of setting the preset task request is to instruct the large language model to perform semantic induction on the input data corresponding to the base profile task, return fields in a pre-defined format (such as structured JSON format), and map them to the corresponding dimensions of the base profile features. In other words, the preset task request for the base profile task instructs the large language model to analyze the input data corresponding to the base profile task and output the information dimensions (and the information format contained in the base profile features) contained in the output base profile features. The features of each dimension in the base profile features of this application can be represented using text labels or numerical labels (i.e., the features output by the large language model can be adaptively represented in text or numerical form). As an example, the input data for the basic profile task can be all social data, or it can be user content footprints and user content interaction behaviors from social data, or it can be a selected portion of representative user content footprints and user content interaction behaviors. That is, the input data for the basic profile task is selected from user social data. The preset task request for the basic profile task must include dimensions of information from the basic profile features. For example, the preset task request for the basic profile task could be: "Based on the input data for the basic profile task, the large language model outputs age groups in numerical form and outputs the distribution of interaction behavior selections in graphical form."

[0040] Furthermore, the core demographic profile can be called a unified core-behavior profile. The core profile describes the user's demographic characteristics, while the behavioral profile describes the user's social interaction characteristics. Therefore, core demographic profile features can include both demographic and social interaction features. Demographic characteristics include one or more of age group, gender, education level, and occupation. Social interaction features include one or more of social activity, interaction intensity, social influence, and the distribution of interaction behavior choices. In addition, demographic characteristics may also include features such as region and / or marital status, and social interaction features may include social volume features such as the number of followers and the number of people they follow.

[0041] Among them, demographic characteristics are the basic positioning elements, and they are mainly obtained by reasoning from the input data corresponding to the baseline profiling task using a large language model. To improve the reliability of the reasoning results, the large language model can output a small amount of representative user social data (such as users' historical posts) as explanatory evidence. At this time, the preset task request corresponding to the baseline profiling task can contain the instruction information of "providing evidence for the reasoning results". Social interaction features are used to provide influence and exposure correction parameters. Social activity can be measured by one or more of the following indicators: number of posts, number of comments, number of likes, number of private messages, number of @ mentions, and number of login days, to represent the frequency of user social interaction. Interaction intensity is used to represent the depth of user social interaction (the depth of the relationship between users or the dimension of user emotional investment). For example, it can be measured by one or more of the following indicators: two-way ratio (the ratio of the other party replying after commenting on someone), interaction speed, and average communication duration. Social influence can be represented by indicators such as centrality and community affiliation. Interaction behavior selection distribution represents the proportion and preference of users when interacting on social platforms using different interaction behaviors (likes, comments, reposts, following, or emojis, etc.). In addition, basic profile features may also include posting pattern features used to depict behavioral rhythm and observability, as well as preliminary language expression features. Posting pattern features may include posting frequency, posting time distribution, media preference, and post length distribution, etc. Preliminary language expression features may include the formality of language (formal or colloquial) and emotional tone features (gentle or sharp tone).

[0042] For example, in the preset task request corresponding to the base bank profiling task, the source code for the format of the base bank profiling features can be as follows: { "demographic_features": { "age_group":"youth", "gender":"female", "location":"Beijing", }, "social_features": { "follower_count": 1280, "following_count": 210, "social_influence_level":"medium", "social_activity_level":"High", }, } In this context, demographic features (age_group, gender, and location) can be represented using text labels; social features (social_features), including social activity level and social influence level, can be represented using text labels, while follower count and following count can be represented using numerical labels. The format of the base row profile features output by the LLM mentioned above is merely an example, and this invention is not limited to it.

[0043] For the multiple dimensions of features included in the basic profile, indicators that can be obtained directly from user social data (such as user posting frequency, and account registration time and account authentication status included in account-side features) can be obtained directly or through large language models. For features that are difficult to read directly or are missing, such as age group and occupation in demographic features, lightweight statistics and text understanding of user social data can be performed, supplemented by semantic induction of large language models, to generate interval-based or label-based inference results. Therefore, the methods for obtaining basic profile features in this application may include: ① obtaining preliminary statistical features based on user social data and personal information, obtaining input data corresponding to the basic profile task based on user social data, and inputting the preset task request corresponding to the basic profile features into the big oracle model, thereby combining the output results of the big language model (and the preliminary statistical features) to obtain the user's basic profile features; wherein, the preliminary statistical features may include account-side profile features, social volume (number of fans and number of followers), and posting pattern features, etc.; ② obtaining input data corresponding to the basic profile task based on user social data and personal information, inputting it into the LLM, having the LLM perform semantic analysis and statistics, and directly outputting the user's basic profile features according to the preset task request corresponding to the basic profile features.

[0044] As an example, LLM can record confidence levels and traceable evidence (including example posts and timestamps) for each inference result in the base profile features to improve the reliability of the inference results. Before obtaining the input data corresponding to the base profile task, the method also includes: performing standardization, time decay (for selecting data from more recent times), and exposure normalization on user social data to reduce the impact of traffic fluctuations. Furthermore, after obtaining demographic features, posting pattern features, preliminary language expression features, and social interaction features, user base profile features can be obtained through cross-domain weighting, which can be expressed by the following formula: ; Specifically, features of each dimension After normalization, the data is first synthesized within the domain, and then weighted across domains to obtain the user's data. Basic portrait features .in, express The corresponding weights It indicates demographic characteristics, social interaction characteristics, posting pattern characteristics, or preliminary language expression characteristics.

[0045] By analyzing objective indicators such as users' posting history, interaction patterns, and social activities using statistical methods, a basic behavioral profile (i.e., baseline profile) can be formed. The construction of this baseline profile can generate structured features around core dimensions such as demographics, social interaction, and posting patterns, outlining the user's basic profile. In other words, based on the user's baseline profile on social networks, an objective baseline of user behavior can be provided. This baseline profile helps understand users' social activities and interactive behaviors, providing a stable and interpretable data foundation for subsequent, more complex psychological characteristic analysis and prediction of violent behavior tendencies.

[0046] In some embodiments of this invention, the invention explicitly extracts psychological features (moral detachment, empathy deficit, the dark triad, and impulse control deficit, etc.) to ensure that both the mechanism-centered and attribute-centered psychological profile features are traceable to psychological theories, thus meeting interpretability compliance requirements. The core objective of this application's mechanism-centered profiling of users is to deeply analyze users' psychological motivations and emotional responses in social interactions, understand users' psychological characteristics, behavioral motivations, and risk tendencies, and provide a scientific basis for the platform to identify potential violent behavior risks. The core task of attribute-centered profiling is to deeply analyze users' personality traits, interests, values, and lifestyles to further understand users' behavioral patterns. Mechanism-centered profiling focuses on using psychological theories to explain the mechanisms behind user behavior, while attribute-centered profiling focuses on directly describing the attribute labels exhibited by users. The combined psychological profile provides a psychological baseline for subsequent behavior prediction. Finally, a complete user profile is constructed based on the baseline profile, the mechanism-centered profile, and the attribute-centered profile.

[0047] More specifically, the psychological profiling side of the mechanism center can perform a structured assessment of the psychological dimensions related to cyberbullying based on summaries of user social data and baseline profile features, outputting psychological profiling features of the mechanism center containing multi-dimensional information. The psychological profiling features of the mechanism center can be obtained through the following methods: acquiring the input data corresponding to the mechanism-psychological profiling task; inputting the acquired input data and the preset task request corresponding to the mechanism-psychological profiling task into a large language model, and outputting the psychological profiling features of the mechanism center.

[0048] The input data for the mechanism-psychological profiling task is derived from user social data and baseline profile features. For example, the input data can be all information contained in all user social data and baseline profile features, or it can be representative information contained in a portion of user social data and baseline profile features. That is, this application does not limit the specific information contained in the input data for the mechanism-psychological profiling task; the input data selected from user social data and baseline profile features is sufficient to accurately profile the user's mechanism-psychological behavior. For example, a sample of a user's historical posts (at least 5-10 representative posts) can be combined with a baseline profile summary (such as social activity, social influence, age group, and education level) to form structured prompts, which are then input into a large language model for semantic analysis.

[0049] Furthermore, the pre-defined task request corresponding to the mechanism-psychological profiling task instructs the large language model to analyze the input data corresponding to the mechanism-psychological profiling task and output the information dimensions contained in the mechanism-centered psychological profiling features (as well as the information format contained in the mechanism-centered psychological profiling features, such as JSON format). The features of each dimension in the mechanism-centered psychological profiling features of this application can be represented by text labels (i.e., the features output by the large language model are represented in text form) or by numerical labels (i.e., the large language model can identify relevant behaviors in user social data according to the pre-defined task request corresponding to the mechanism-psychological profiling task and output standardized scores for the corresponding feature dimensions).

[0050] Furthermore, mechanism-oriented psychological profiling uses psychological theoretical variables (personality, motivation, and cognitive biases, etc.) to explain the mechanisms and causal paths of user behavior. This application constructs a psychological framework based on existing psychological theories to conduct in-depth psychological characteristic assessments of users. The psychological theories used in mechanism-oriented psychological profiling include Moral Foundations Theory, Empathy Theory, Aggression Theory, Big Five Personality Theory, and Dark Personality Theory, etc., to assess users' moral foundations, empathy abilities, and aggressive tendencies.

[0051] In some embodiments of the present invention, the psychological profile features of the mechanism center in this application include moral foundation features, empathy features, aggression features, role types in cyberbullying incidents, personality difference features, and psychological mechanism variables corresponding to cyberbullying behavior.

[0052] The information dimensions included in the basic moral characteristics are determined based on the theory of moral foundations. The core tenet of this theory is that human moral judgment is driven by six "innate modules"—care / harm, fairness / deception, loyalty / betrayal, authority / subversion, sanctity / degradation, and freedom / oppression. Basic moral characteristics can be derived by analyzing these six driving dimensions. That is, basic moral characteristics may include features from one or more of the following dimensions: care / harm, fairness / deception, loyalty / betrayal, authority / subversion, sanctity / degradation, and freedom / oppression, used to determine a user's moral foundation.

[0053] The information dimensions contained in empathy features are determined based on empathy theory. For example, empathy features may include information from two dimensions: cognitive empathy (understanding others' emotions) and affective empathy (having an emotional response to others' emotions) (and may also include information from the sympathetic empathy dimension), which are used to indicate how users treat others.

[0054] The information dimensions included in the aggressive features are determined based on aggressive theories related to cyber violence. Aggressive theories related to cyber violence may include social learning theory, the two-dimensional model of aggressive behavior, and social information processing theory. Furthermore, based on aggressive theories, aggressive behavior can be classified into verbal attacks, relational attacks, and cyber attacks according to the form of attack, and into reactive attacks and instrumental attacks according to the motivation of attack. Therefore, this application uses aggressive features to indicate the likelihood of users engaging in different types of aggressive behavior.

[0055] To distinguish the essential differences in emotions, personality, and social relationships among perpetrators, victims, and bystanders, this application uses a large language model to infer the possible cyberbullying role type of a user in a cyberbullying incident and outputs the cyberbullying incident role type features. That is, this application uses the cyberbullying incident role type features to indicate the possibility of a user playing various role types in a cyberbullying incident. Therefore, the cyberbullying incident role type features output by the large language model can be the possibility of perpetrator, victim, bystander, defender, accomplice, or instigator.

[0056] The information dimensions included in personality difference characteristics are determined based on the Big Five personality theory. Personality difference characteristics can include information on openness, conscientiousness, extraversion, agreeableness, and neuroticism. For example, the Big Language Model can capture complex personality signals in the input data corresponding to the mechanism-mental profiling task, such as openness expressions (e.g., creative thinking, artistic interests), conscientiousness performance (e.g., planning, responsibility), extraversion characteristics (e.g., social activity, desire to express), agreeableness tendency (e.g., cooperativeness, empathy), and neuroticism performance (e.g., mood swings, anxiety expression), and output standardized scores for core dimensions such as openness, conscientiousness, extraversion, agreeableness, and neuroticism.

[0057] The psychological mechanism variables corresponding to cyberbullying behavior in the large language model are the psychological mechanism variables of violence against others, which may include the variable values ​​of risk factors and protective factors corresponding to cyberbullying behavior, social dominance orientation characteristics, and negative personality tendencies.

[0058] As an example, the risk factors and protective factors corresponding to cyberbullying can be psychosocial-behavioral risk and protective factors (hereinafter referred to as risk and protective factors). Among them, risk factors may include one or more of the following: seeking anonymity, moral detachment, social isolation, negative peer influence, and low self-control. Protective factors may include one or more of the following: empathy intensity, social support, moral reasoning, self-regulation, and positive peer influence.

[0059] The information dimensions included in the characteristics of social dominance orientation are determined based on the theory of social dominance orientation (SDO). SDO is a core personality-attitude variable in social dominance theory, used to measure an individual's preference for "group-based social hierarchy and inequality". This application uses indicators from the indicator pool or content domain of SDO as information dimensions included in the characteristics of social dominance orientation. Therefore, the characteristics of social dominance orientation include information on the dimensions of group dominance, anti-equality, hierarchy preference, power seeking, status consciousness, and competitive worldview.

[0060] The information dimensions contained in the negative personality tendency characteristics are determined based on the Dark Personality Theory. This application may use the indicators mentioned in the Dark Tetrad and Dark Dyad supplements of the Dark Personality Theory as information dimensions contained in the negative personality tendency characteristics. Therefore, the negative personality tendency characteristics in this application may include information on narcissistic tendencies, Machiavellianism, psychopathic traits, sadistic tendencies, impulsivity, and dominance needs.

[0061] Furthermore, the psychological mechanisms underlying cyberbullying can include dehumanization tendency, moral justification and victim blaming, group-think susceptibility and authority obedience, social identity dependence, and emotion regulation capacity (a comprehensive index of emotional awareness, control, tolerance, impulse control, and recovery), to further explain and assess users' psychological traits. These psychological mechanisms can all be scored using appropriate scales.

[0062] Therefore, when creating a profile of a user at the mechanism center, a large language model can be used as a psychological scale for intelligence to output the psychological profile features of the mechanism center.

[0063] As an example, all results in the psychological profiling features of the mechanism center are presented as standardized scores and labeled elements. All features are represented by standardized scores of 0 to 1, and default values ​​are used for fallback when parsing fails. Furthermore, to improve the reliability of the results, the large language model can retain and output corresponding traceable evidence (such as statements and timestamps) for interpretation, aligning with the dimension system already selected for this project while maintaining the readability, comparability, and traceability of the output, providing robust input for subsequent contextualized predictions. In addition, before obtaining the input data corresponding to the mechanism-psychological profiling task, the method further includes: performing standardization and time decay on user social data, preserving contextual semantics while selecting a small number of representative fragments as reading samples, and inputting the summary of the basic profiling features into the large language model.

[0064] For example, in the preset task request corresponding to the mechanism-mental profiling task, the source code for the format of the mental profiling features of the mechanism center can be as follows: { "aggression_tendencies": { "verbal_aggression": 0.73, "relational_aggression": 0.68, "cyber_aggression": 0.82, "reactive_aggression": 0.79, "proactive_aggression": 0.35}, "empathy_dimensions": { "cognitive_empathy": 0.42, "affective_empathy": 0.28, "compassionate_empathy": 0.31 }, } In this context, the verbal aggression, relational aggression, cyber aggression, reactive aggression, and proactive aggression dimensions within the aggression tensities can all be represented using numerical labels; similarly, the cognitive empathy, affective empathy, and compassionate empathy dimensions within the empathy dimensions can also be represented using numerical labels. In other words, each dimension of the mechanism-centered psychological profile features can be represented using numerical labels. The output format of the mechanism-centered psychological profile features mentioned above is merely an example, and this invention is not limited to it.

[0065] Furthermore, taking the psychological profile features of the mechanism center, which consist of moral foundation features, empathy features, and aggression features, as an example, after obtaining the moral foundation features, empathy features, and aggression features, the psychological profile features of the mechanism center can be obtained through cross-domain weighting. This process can be expressed by the following formula: ; in, Indicates user The corresponding psychological profile characteristics of the mechanism center Indicates user Corresponding basic moral characteristics Indicates user Corresponding empathy characteristics Indicates user Corresponding aggressive characteristics , and The weights are respectively for the basic moral characteristics, the empathy characteristics, and the aggression characteristics.

[0066] Attribute-centered psychological profiling describes users using directly observable or self-described attribute tags such as interests, values, emotional expression, lifestyle, and cognitive style. By comprehensively considering multiple dimensions of personality characteristics, it can more comprehensively assess users' behavioral patterns and potential risks, making it an indispensable part of predicting violent behavior tendencies. When conducting attribute-centered psychological profiling analysis on users, user social data and pre-generated baseline profile features can be used as input to structurally summarize users' interests, values, lifestyles, and cognitive styles. Specifically, attribute-centered psychological profile features can be obtained through the following methods: acquiring the input data corresponding to the attribute-psychological profiling task; inputting the acquired input data and the preset task request corresponding to the attribute-psychological profiling task into a large language model, and outputting the attribute-centered psychological profile features.

[0067] As an example, the input data for the attribute-mental profiling task is obtained based on user social data and baseline profile features. Similar to the mechanism-mental profiling task, this application does not limit the specific information contained in the input data for the attribute-mental profiling task; the input data selected from user social data and baseline profile features is sufficient to accurately profile the user's attributes. For example, a small number of representative passages can be selected as reading samples (at least 5 representative statements), and combined with baseline profile features to form structured prompts, which are then input into a large language model for semantic analysis. Furthermore, the preset task request for the attribute-mental profiling task instructs the large language model to analyze the input data for the attribute-mental profiling task and output the information dimensions (and the information format contained in the attribute-center profile features) contained in the attribute-center profile features. The features of each dimension in the attribute-center profile features of this application can be represented using text labels or numerical labels; that is, the features output by the large language model can be adaptively represented using either text or numerical forms.

[0068] More specifically, the psychological profile features of the attribute center may include theme focus features, value factor features, cognitive style features, lifestyle-related indicator features, emotional process-function dimension features, and conflict handling style features. Theme focus features include theme preference distribution (indicating the user's preference and participation in topics such as history, politics, and entertainment), valence-emotional tendency (positive, negative, or neutral) for each theme, and one or more of the following: stance firmness. Value factor features include features in dimensions such as morality, self-direction, family orientation, love, and compliance / work. Cognitive style features indicate the likelihood that a user's cognitive style is analytical, creative, critical, abstract, and / or pragmatic. Lifestyle-related indicators are behavioral-resource-attitude indicators reflecting lifestyle, and these indicators reflect the user's situation regarding balanced diet and daily exercise. Emotional process-function dimension features include information on emotional awareness, emotional intelligence, emotional resilience, emotional expression, and / or emotional connectivity. The information dimensions contained in the conflict handling style characteristics are determined based on the Thomas-Kilmann Conflict Mode Instrument (TKI). The conflict handling style characteristics are used to indicate the likelihood of a user's conflict handling style (including competitive, accommodating, avoidant, cooperative, and compromising).

[0069] Furthermore, the psychological profile features of the attribute center can also include language style features to analyze the user's cyberbullying language expression style. Language style features can include one or more of the following: frequency of use of words related to cyberbullying, syntactic features, frequency of use of non-literal rhetoric, degree of expression of social emotions related to cyberbullying, formality of language, degree of humor, and frequency of stereotype expression. Among these, words related to cyberbullying include offensive words (such as "garbage," "waste," and "disgusting"), dehumanizing words (such as expressions of "this kind of person" and "that kind of thing"), extreme words (such as "absolute," "must," and "deserves it"), and moral judgment words (such as "unworthy," "unqualified," and "shameful"). Syntactic features include the frequency of use of various sentence types (such as rhetorical questions, exclamatory sentences, and imperative sentences), sentence completeness, and sentence complexity. Non-literal rhetoric includes irony (such as irony) and hyperbole, and social emotions related to cyberbullying include anger, disgust, contempt, and a sense of moral superiority.

[0070] By analyzing the characteristics of cyberbullying language in online discourse from a linguistic perspective using a large language model, including lexical level (density of aggressive vocabulary, dehumanizing expressions, extreme adjectives and moral judgment vocabulary), syntactic level (frequency of rhetorical questions, density of exclamatory sentences, use of imperative sentences, sentence fragmentation), rhetorical level (irony, exaggeration, derogatory contrasts, labeling and stereotypes), and social emotions related to cyberbullying (intensity of anger, aversion markers, contempt indicators, and sense of moral superiority), multiple dimensions of language style characteristics can be obtained.

[0071] As an example, for the psychological profiling features at the core of the mechanism, the output of the large language model can retain a small amount of speaking evidence and timestamps for interpretation. This aligns with the large language model-driven implementation of this project while ensuring the readability, comparability, and traceability of the results. When using numerical labels to represent features in each dimension, standardized scores of 0-1 or structured labels can be used, returned strictly in JSON format, and default values ​​are used as a fallback in case of parsing failure. Furthermore, before obtaining the input data corresponding to the attribute-psychological profiling task, the method also includes performing standardization and time decay on the user's social data.

[0072] Furthermore, taking the psychological profile features of the attribute center, composed of thematic focus features, value factor features, and lifestyle-related indicator features, as an example, after obtaining the thematic focus features, value factor features, and lifestyle-related indicator features, the psychological profile features of the attribute center can be obtained through cross-domain weighting. During fusion, the results of each dimension are standardized and then synthesized according to their weights to form a compact representation of the psychological profile of the attribute center. The weights are configurable and auditable. This process can be expressed using the following formula: ; in, Indicates user The corresponding attribute center's psychological profile features Indicates user The corresponding lifestyle and its corresponding indicator characteristics Indicates user Corresponding topic focus characteristics, Indicates user Corresponding value factor characteristics, , and These are the weights corresponding to the indicator characteristics of the theme focus, value factors, and lifestyle.

[0073] Following in-depth analysis using a large language model, the system outputs elements related to interests and values, lifestyle and emotions, and cognition and language in a unified format. Simultaneously, enhancements from the project are incorporated to supplement the analysis of conflict handling styles and cyberbullying language styles, improving interpretability and discriminative power, and providing usable signals for subsequent risk analysis. This stage, by analyzing users' personality traits, interests, values, and lifestyles, provides a stable personality baseline for subsequent behavior prediction. The attribute-centered psychological profile features can better predict user behavior in different topical contexts, especially how users might express violent tendencies or emotional reactions on social media.

[0074] The core task of the prediction phase is to assess users' potential behavior in specific situations, especially their tendency toward violent behavior, by integrating information such as basic behavioral profiles, mechanism-centered psychological profiles, and attribute-centered psychological profiles. This provides the platform with in-depth insights into user behavior and helps to identify potential cyber violence risks in a timely manner.

[0075] Most existing models have fixed parameters, making it impossible to dynamically adjust prediction weights based on different event types such as "educational controversies" and "entertainment gossip" to achieve contextual adaptation. To address this challenge, this application, based on user profiling, introduces information about the topic to be discussed (including the topic's title, description, level of controversy (which can be determined through a large language model), and keywords) to generatively assess users' participation probability, risk of cyberbullying, and potential statements in specific topic contexts.

[0076] Specifically, in the prediction phase, the method proposed in this application can input the obtained input data corresponding to the prediction task, the topic scenario information to be discussed, and the preset task request corresponding to the prediction task into a large language model, and output the probability of a user engaging in cyberbullying behavior under the topic scenario information to be discussed. The content output by the large language model in the prediction phase is determined based on the preset task request corresponding to the prediction task. If the preset task request instructs the large language model to output the probability of a user engaging in cyberbullying behavior under the topic scenario information to be discussed, based on the input data corresponding to the prediction task, then the output of the large language model can include a cyberbullying behavior risk score; if the preset task request instructs the large language model to output the probability of a user engaging in cyberbullying behavior under the topic scenario information to be discussed and the possible statements a user might make under the topic scenario information, then the output of the large language model can include a cyberbullying behavior risk score and possible statements.

[0077] As an example, the input data for the prediction task may include some or all of the user's profile features, as well as some or all of the user's social data. For instance, the basic profile features, mechanism-centered psychological profile features, and attribute-centered psychological profile features can be compressed into a feature summary, accompanied by several representative statements as explanatory evidence, and combined with topic information. The combined and concatenated data is then input into a large language model to perform the prediction task. Moreover, before combining and concatenating, since the profile features in the input data for the prediction task include text labels and / or numerical labels, the features represented by numerical labels in the user's profile features can be converted into natural language descriptions to standardize the input of the large language model.

[0078] Furthermore, during the prediction phase, the large language model can output the probability of user participation, the likelihood of a user engaging in cyberbullying (which can be represented by a cyberbullying risk score, etc.), the probability of different types of cyberbullying (cyberbullying can be categorized as abusive, defamatory, disruptive, information bombardment / harassment, malicious marketing, and privacy breaches), the statements a user might make, the corresponding risks and protection factors for the user, and the confidence level for each result. This application does not limit the specific types of information input to the large language model; they can be set according to the preset task requests corresponding to the prediction task.

[0079] When calculating behavioral probabilities, large language models can predict the types of behavior a user might exhibit, such as aggressive language, sarcasm, and insults, based on the user's psychological and emotional state. For example, based on the user's psychological profile characteristics and historical behavioral data, the probability and depth of the user's participation in a discussion on a particular topic, as well as the risk of violent behavior, can be calculated. This process considers the user's emotional inclination, language style, and social interaction patterns. Furthermore, when calculating behavioral probabilities, large language models can also assess the risk level of a user engaging in cyberbullying in a given context, based on the nature of a specific topic (such as a controversial topic) and the user's behavioral patterns. Assessment indicators may include depth of participation, frequency of aggressive language use, and intensity of emotional fluctuations.

[0080] In some embodiments of the present invention, if the large language model is required to output the possible statements made by users under topic-specific context information, the preset task request corresponding to the prediction task may include a specific process for analyzing the user's vocabulary usage patterns based on the user's content footprint. For example, the large language model in this application can analyze vocabulary usage patterns from two perspectives: vocabulary usage and expression habits. The output analysis results may include vocabulary usage frequency (to statistically analyze high-frequency words), usage of aggressive words (or usage of words related to cyberbullying), usage of emotional expression words (including emotional expression words with positive and negative emotional tendencies, etc.), and the user's personalized expression patterns. The user's personalized expression patterns may include the usage of discourse markers (such as catchphrases) and the usage of modal particles (such as "ah," "ba," etc.). Vocabulary usage pattern analysis can further provide the large language model with supplementary features of the user at the language level.

[0081] As an example, when performing prediction tasks, large language models can also correlate the output results with profile features and parts of user social data, providing interpretability analysis. That is, large language models not only output prediction results but also interpretable analysis results, helping to understand the basis for predictions and providing decision support for subsequent interventions. Furthermore, the corresponding explanations accompanying user behavior predictions help the platform better understand which factors contributed to the prediction. Moreover, to avoid interference from traffic and short-term fluctuations, time-series and interaction signals undergo time decay and exposure normalization processing before obtaining the input data for the prediction task. Additionally, after obtaining the data, the risk of cyberbullying behavior can be obtained through cross-domain weighting, a process that can be represented by the following formula: ; in, Indicates user The risk of cyberbullying , and The weights are respectively the psychological profile features of the base line, the psychological profile features of the mechanism center, and the psychological profile features of the attribute center.

[0082] In some embodiments of the present invention, the input data for the baseline profiling task includes some or all of the social data; the input data for the mechanism-psychological profiling task includes some or all of the social data and demographic features; the input data for the attribute-psychological profiling task includes some or all of the social data and baseline profiling features; and the input data for the prediction task may include some or all of the user's corresponding profiling features. Furthermore, the social data in the input data for the mechanism-psychological profiling task and the attribute-psychological profiling task may also differ. That is, when obtaining the user's corresponding profiling features and predictions of cyberbullying behavior, certain features and data dimensions can be enabled or disabled as needed to adapt to different topic scenarios and compliance requirements.

[0083] The prediction results output by the large language model can be represented by standardized scores and structured labels, returned strictly in JSON format. In case of parsing failure, a default value is used as a fallback, and speaking evidence and timestamps are retained for interpretation, providing decision support for platform intervention. At this point, preset task requests corresponding to prediction tasks, baseline profiling tasks, mechanism-psychological profiling tasks, and attribute-psychological profiling tasks can all instruct the large language model to return conclusions and reasons, as well as the format of the returned conclusions. Within this framework, prediction results can be quantitatively compared and also subject to human review and verification through traceable evidence.

[0084] The cyberbullying prediction method proposed in this application has the following significant advantages: ① This application proposes an accurate method for predicting cyberbullying behavior by integrating multi-dimensional feature analysis and a large language model. It provides an objective behavioral baseline by analyzing user social data, while the large language model is responsible for identifying implicit features and deeply mining potential risks of violent behavior. The combination of the two ensures the comprehensiveness and accuracy of feature analysis, thereby accurately predicting users' behavioral tendencies on social platforms.

[0085] ② Multi-level and dynamic user profile construction: This method uses a multi-level analysis framework to construct user basic profiles, mechanism-centered psychological profiles, and attribute-centered psychological profiles, analyzes users' personalized performance in terms of emotions and psychology, and enables the large language model to identify hidden violent expressions and make personalized risk predictions.

[0086] ③ This application utilizes the natural semantic understanding capabilities of large language models to predict cyberbullying behavior, achieving superior accuracy and adaptability. Compared to traditional statistical analysis methods and rule-based analysis systems used to analyze user behavior and emotional fluctuations, LLM offers improved flexibility and accuracy, particularly in handling covert violent language and complex emotional fluctuations. While deep learning models and social network analysis have certain analytical advantages, they typically require large amounts of training data and computational resources, whereas LLM effectively reduces training costs.

[0087] ④ This application has a flexible contextualized prediction mechanism that can comprehensively consider behavioral tendencies in different contexts, and ultimately achieve efficient prediction and risk assessment of cyberbullying behavior.

[0088] ⑤ Explainability and Feedback Mechanism: The explainability analysis function provides clear evidence for each inference result. For example, the large language model will list in detail which features and factors led to the prediction of a user's violent behavior risk, helping the platform understand the motivations behind these behaviors. At the same time, user feedback will also be incorporated into the system to continuously optimize the prediction model and improve accuracy.

[0089] ⑥ Efficiently acquire information such as users' historical behavior, social interactions, and engagement levels from social platforms through web crawlers or APIs, thereby providing a foundation for user profiling and sentiment modeling. Furthermore, as user social data accumulates, the large language model can continuously update and improve the behavioral characteristics of each user, enhancing the model's predictive capabilities. In practical applications, this invention can process user social data in real time and generate interpretable prediction results, providing effective risk assessment for social platforms. Through dynamic modeling of behavioral tendencies, the platform can promptly detect potential violent behavior and take corresponding intervention measures, applicable to multiple fields such as social platforms and content moderation. Ultimately, this technical solution not only enhances the capabilities of social platforms in managing cyberbullying but also provides important support for creating a safer and more harmonious online environment.

[0090] Corresponding to the above method, the present invention also provides a cyberbullying behavior prediction system, the system including a computer device, the computer device including a processor and a memory, the memory storing computer programs / instructions, the processor being used to execute the computer programs / instructions stored in the memory, and when the computer programs / instructions are executed by the processor, the system implements the steps of the method described above.

[0091] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0092] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0093] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. 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 cyber violence prediction method, characterized by, The method comprises the following steps: Obtaining input data corresponding to a prediction task; wherein the input data corresponding to the prediction task comprises user corresponding portrait features obtained based on user social data, the user corresponding portrait features comprising basic portrait features, mechanism center psychological portrait features and attribute center psychological portrait features, and the user social data comprising user content footprints and user content interaction behaviors; Inputting the obtained input data, topic scene information to be discussed and a preset task request corresponding to the prediction task into a large language model to output a possibility of the user's online violence behavior under the topic scene information to be discussed.

2. The method of claim 1, wherein, The user content footprints comprise user post records and comment records, and the user content interaction behaviors comprise post behaviors, comment behaviors, like behaviors, forwarding behaviors and content shielding behaviors. The user social data further comprises a user profile, and the user corresponding portrait features further comprise account side portrait features determined based on the user profile, the account side portrait features comprising account registration duration and / or account authentication status.

3. The method of claim 1, wherein, The basic portrait features are obtained by the following method: Obtaining input data corresponding to a basic portrait task; wherein the input data corresponding to the basic portrait task is obtained based on user social data; Inputting the input data corresponding to the basic portrait task and a preset task request corresponding to the basic portrait task into a large language model to output basic portrait features; wherein the basic portrait features comprise user corresponding demographic features and social interaction features; the demographic features comprise one or more of age, gender, education level and occupation, and the social interaction features comprise one or more of social activity, interaction intensity, social influence and interaction behavior selection distribution; the preset task request corresponding to the basic portrait task is used to indicate information dimensions contained in the basic portrait features output by the large language model.

4. The method of claim 1, wherein, The mechanism center psychological portrait features are obtained by the following method: Obtaining input data corresponding to a mechanism-psychological portrait task; wherein the input data corresponding to the mechanism-psychological portrait task is obtained based on the user social data and the basic portrait features; Inputting the obtained input data and a preset task request corresponding to the mechanism-psychological portrait task into a large language model to output the mechanism center psychological portrait features; The mechanism center psychological portrait features comprise aggressive features, online violence event role type features and psychological mechanism variables corresponding to online violence behaviors; information dimensions contained in the aggressive features are determined based on aggression theories related to online violence, the psychological mechanism variables corresponding to the online violence behaviors comprise risk factors and protection factors corresponding to the online violence behaviors, social dominance orientation features and negative personality tendency features, and information dimensions contained in the negative personality tendency features are determined based on dark personality theory.

5. The method of claim 4, wherein, The mechanism-centered psychological portrait feature further includes a moral foundation feature, an empathy feature, and a personality difference feature; the information dimensions included in the moral foundation feature are determined based on a moral foundation theory, the information dimensions included in the empathy feature are determined based on an empathy theory, and the information dimensions included in the personality difference feature are determined based on a Big Five personality theory.

6. The method of claim 4, wherein, The attribute-centered psychological portrait feature is obtained by: obtaining input data corresponding to the attribute-psychological portrait task; wherein the input data corresponding to the attribute-psychological portrait task is obtained based on the user social data and the base psychological portrait feature; inputting the obtained input data and a preset task request corresponding to the attribute-psychological portrait task into a large language model to output the attribute-centered psychological portrait feature; The attribute-centered psychological portrait feature includes a topic focus feature, a value factor feature, a cognitive style feature, a lifestyle corresponding indicator feature, and a language style feature; the topic focus feature includes one or more of a topic preference distribution, a valence emotion tendency for each topic, and a stance firmness, and the language style feature includes a frequency of using words related to cyber violence, a syntax feature, and a frequency of using non-literal rhetoric.

7. The method of claim 6, wherein, The attribute-centered psychological portrait feature further includes one or more of an emotional process-function dimension feature and a conflict processing style feature, the emotional process-function dimension feature includes information in the dimensions of emotional awareness, emotional intelligence, emotional resilience, emotional expression, and / or emotional connection, and the information dimensions included in the conflict processing style feature are determined based on a conflict processing model theory; The language style feature further includes one or more of an expression degree of social emotion related to cyber violence, a degree of formality of language, a degree of humor expression, and a frequency of stereotype expression.

8. The method of claim 6, wherein, The mechanism-psychological portrait task corresponding input data includes part or all of the social data and the demographic characteristics; The attribute-psychological portrait task corresponding input data includes part or all of the social data and the base psychological portrait feature.

9. The method of claim 1, wherein, After inputting the input data corresponding to the prediction task into the large language model, the method further includes: outputting a possible statement published by the user under the topic scene information; The preset task request corresponding to the prediction task is used to instruct the large language model to analyze the user's vocabulary usage pattern based on the user content footprint, and to predict the possibility of cyber violence behavior and the possible statement based on the analysis result of the vocabulary usage pattern, the user's portrait feature, and the topic scene information to be discussed; wherein the analysis result of the vocabulary usage pattern includes vocabulary usage frequency, aggressive vocabulary usage, emotional expression vocabulary usage, and user's personalized expression pattern; The topic scene information includes one or more of a topic title, a topic description, a topic controversy degree, and a topic keyword.

10. A network behavior prediction system comprising a processor, a memory, and a computer program / instructions stored on the memory, wherein, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method of any one of claims 1 to 9.

Images