Large model-based bullying behavior identification and real-time response method

Abstract

The invention belongs to the technical field of artificial intelligence application, and discloses a bullying behavior recognition and real-time response method based on a large model. According to the method, multi-modal data streams in an interaction scene are obtained, time sequence alignment is carried out on texts, voices and behavior data, semantic and emotional features are accurately extracted, and a complete bullying behavior evidence chain is constructed. And a role relation graph is introduced, and bullying behaviors caused by unequal power and the like are identified from the social relation dimension. Based on the overall strength and continuity density of the evidence chain, a hierarchical intervention strategy from observation to blocking is implemented, and model parameters are continuously optimized through closed-loop feedback. According to the method, the accuracy and the sensitivity of bullying behavior identification are improved, and a safer communication environment is provided for interaction participants.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence application technology, and more specifically, to a method for bullying behavior recognition and real-time response based on large models. Background Technology

[0002] Current technologies for identifying and responding to bullying suffer from multiple systemic flaws, resulting in limited effectiveness in practical applications. Traditional bullying detection systems rely excessively on explicit keyword filtering and surface semantic analysis, failing to effectively identify covert bullying statements employing puns, metaphors, or seemingly friendly but actually derogatory language. This allows numerous "joke-style bullying" attacks to evade detection on campus social platforms and in online learning environments. Furthermore, existing technologies generally suffer from modal fragmentation, focusing only on a single data dimension and failing to perform temporal alignment and comprehensive analysis of multimodal information such as text content, emotional changes in speech, and user behavior. Consequently, in video conferencing and voice-based social scenarios, when attackers deliberately use sarcastic speech paired with neutral text, the system struggles to capture cross-modal inconsistencies. More critically, existing methods severely neglect the analysis of social relationships and power structures among participants, failing to distinguish between normal teacher-student guidance and systemic derogatory behavior based on power inequality. Particularly in corporate internal communications and teacher-student interaction platforms, implicit power abuse based on status differences is almost entirely undetectable by existing systems. Furthermore, the lack of in-depth analysis of the temporal continuity of bullying behavior makes it difficult for the system to distinguish between occasional arguments and persistent harassment. This leads to low-intensity but high-frequency bullying behavior being mistakenly interpreted as normal communication in long-term group interactions. Regarding intervention mechanisms, existing technologies mostly employ a one-size-fits-all, fixed response model, lacking refined intervention strategies based on the severity of bullying, contextual characteristics, and historical patterns. This results in either over-intervention affecting normal discussions on educational platforms or insufficient intervention failing to truly protect vulnerable students. Finally, existing systems generally lack a self-optimizing closed-loop mechanism based on actual intervention effects. They cannot learn from intervention results and adjust model parameters, leading to a continuous degradation in system performance as bullies' strategies and language expressions evolve, failing to adaptively improve. This is particularly evident in long-running social platforms and collaborative environments. These multi-layered technical limitations collectively result in the severe inadequacy of existing bullying identification and intervention systems in complex real-world environments.

[0003] In view of this, the present invention proposes a bullying behavior recognition and real-time response method based on a large model to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a bullying behavior recognition and real-time response method based on a large model, comprising: Acquire multimodal data streams from interactive scenarios, including text dialogue sequences, speech prosody parameters, and user behavior temporal markers; Differential calibration is performed on the timestamps of each modality in the multimodal data stream to identify the synchronization anchor points of cross-modal interaction events, and the multimodal data stream is time-aligned based on the synchronization anchor points; Based on the time-aligned data, we extract the semantic feature vector and sentiment feature vector of the text, and calculate the differential evolution trajectory of the semantic feature vector and sentiment feature vector in the time dimension. Based on the abrupt change point of the slope of the differential evolution trajectory, the non-consistency interval between semantic expression and sentiment tendency is located, and the semantic deviation index within the non-consistency interval is calculated. Extract historical dialogue relationship data between interactive participants, construct a role relationship graph containing power gradient and interaction frequency, and label the weak and strong nodes in the role relationship graph; Based on semantic deviation index, duration of inconsistency interval and node power difference of role relationship graph, a chain of evidence for bullying behavior is generated. The confidence levels of each evidence node in the evidence chain are weighted and fused to calculate the overall strength value of the evidence chain. The intervention response level is determined based on the overall strength value and the continuity density of the evidence nodes. The corresponding real-time intervention strategy is triggered based on the intervention response level, and data on changes in the interaction status after the intervention are collected. Based on the interaction state change data and the preset intervention effect evaluation index, the intervention deviation value is calculated, and the intervention deviation value is backpropagated to the feature extraction layer of the large model to update the extraction weights of semantic feature vector and sentiment tendency feature vector.

[0005] The technical effects and advantages of the bullying behavior recognition and real-time response method based on a large model in this invention are as follows: In everyday educational settings, this invention's method can keenly detect covert bullying patterns in student interactions that appear friendly but actually contain underlying aggression, effectively avoiding the omissions of traditional systems that detect "joking bullying" and providing a safer digital learning space for teachers and students. On enterprise collaboration platforms, this invention effectively identifies power abuses based on job position differences, even when these behaviors are disguised as professional guidance, allowing for accurate identification and appropriate intervention, promoting a more equitable and healthy work communication environment. For online social networks, this invention demonstrates high sensitivity to persistent, low-intensity harassment, capable of identifying seemingly independent but actually systemic attack patterns of speech, providing comprehensive protection for vulnerable users. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the bullying behavior recognition and real-time response method based on a large model according to the present invention. Detailed Implementation

[0007] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0008] This application provides a method for identifying and responding to bullying behavior based on a large model. The execution entities of this method include, but are not limited to, intelligent dialogue systems, social media monitoring platforms, educational environment monitoring systems, and network security management tools, which can be considered as general computing nodes in this application. The intelligent dialogue system includes, but is not limited to, at least one of the following: a large-model chatbot, an affective computing system, and an interactive tutoring assistant.

[0009] In this embodiment of the invention, the detailed implementation steps of the bullying behavior recognition and real-time response method based on a large model include: The system acquires multimodal data streams from the interaction scenario, including text dialogue sequences, speech prosodic parameters, and user behavior time stamps. The text dialogue sequences record the text content exchanged between participants, including speaking time, speaker identification, and complete text. Speech prosodic parameters capture the prosodic features of speech expression, such as pitch, volume, speech rate, and pauses, and are extracted in real-time using specialized audio analysis algorithms. User behavior time stamps record nonverbal behaviors during the interaction, such as clicks, swipes, and facial expressions, along with corresponding timestamps. These multimodal data collectively constitute a comprehensive record of the interaction scenario, providing a rich source of information for subsequent analysis.

[0010] Differential calibration is performed on the timestamps of each modality in the multimodal data stream to identify synchronization anchors for cross-modal interaction events. The multimodal data stream is then time-aligned based on these synchronization anchors. Time alignment is fundamental to multimodal analysis, ensuring that data from different sources can be correlated and analyzed within a unified temporal reference frame. By accurately calibrating the temporal relationships between different modalities, the manifestations of the same interaction event in different modalities can be identified, establishing correspondences between modalities. The time-aligned data provides an accurate time base for subsequent feature extraction and pattern recognition, improving the accuracy of the analysis.

[0011] Based on time-aligned data, semantic feature vectors and sentiment feature vectors are extracted, and their differential evolution trajectories over time are calculated. The semantic feature vectors, extracted through the deep semantic understanding capabilities of a large model, capture the core meaning, theme, and contextual relationships of the text; the sentiment feature vectors reflect the emotional state, attitude, and intensity expressed in the text. The differential evolution trajectories of these two feature vectors over time demonstrate the dynamic relationship between semantic expression and sentiment, serving as a crucial basis for identifying bullying rhetoric. An abnormal separation between semantic and sentiment expression may indicate covert bullying behavior.

[0012] Based on the slope abrupt change points of the differential evolution trajectory, inconsistencies between semantic expression and emotional tendency are located, and semantic deviation indices are calculated within these inconsistencies. Slope abrupt change points reflect sharp changes in the relationship between semantics and emotion, typically corresponding to shifts in the nature of the interaction. Identifying and analyzing these abrupt change points allows for the accurate location of potential bullying scenarios. Within these inconsistencies, semantic deviation indices are further calculated to quantify the degree of inconsistency between semantic expression and emotional tendency, providing a quantitative basis for determining bullying behavior.

[0013] Historical dialogue data among interactive participants was extracted to construct a role relationship graph containing power gradients and interaction frequencies, marking weak and strong nodes within the graph. The role relationship graph is a crucial tool for understanding the social relationships of interactive participants, reflecting the power structure and interaction patterns among them. By analyzing behavioral characteristics in historical dialogues, dominant strong nodes and subordinate weak nodes can be identified. This role classification provides a social relationship dimension for determining bullying behavior, helping to distinguish between legitimate guidance and power abuse.

[0014] Based on semantic deviation indicators, the duration of inconsistency intervals, and the node power differences in the role relationship graph, a bullying behavior evidence chain is generated. This chain comprises multiple temporally related evidence nodes. The bullying behavior evidence chain is the core basis for determining bullying behavior, integrating evidence from multiple dimensions including semantics, emotion, and social relationships. The evidence chain is not an isolated judgment but rather continuous evidence based on temporal relationships, capable of capturing the persistence and pattern characteristics of bullying behavior. Each evidence node is accompanied by detailed time information, involved roles, and intensity assessments, collectively forming a complete bullying behavior evidence system.

[0015] The confidence levels of each evidence node in the chain of evidence are weighted and fused to calculate the overall strength value of the chain. The intervention response level is determined based on the overall strength value and the continuity density of the evidence nodes. The confidence assessment considers both the reliability and importance of the evidence nodes, and the overall assessment result is generated through weighted fusion. The overall strength value and the continuity density of the evidence nodes jointly determine the intervention response level, ensuring that the intervention measures match the severity of the bullying behavior. The intervention response levels are divided into different levels from low to high, corresponding to different intensities of intervention measures, ensuring that the response is neither excessive nor insufficient.

[0016] Real-time intervention strategies are triggered based on the intervention response level, and data on changes in interaction status after intervention are collected. Intervention strategies are actual responses to bullying behavior, including multi-level measures ranging from mild warnings to stern alerts and behavioral intervention. The selection of intervention strategies is based on the response level to ensure the appropriateness and effectiveness of the intervention. After intervention, changes in interaction status are continuously monitored, and data on participants' behavioral and emotional responses are collected to provide a basis for evaluating the intervention's effectiveness.

[0017] Based on interaction state change data and preset intervention effect evaluation indicators, an intervention bias value is calculated. This bias value is then backpropagated to the feature extraction layer of the large model to update the extraction weights of semantic feature vectors and sentiment tendency feature vectors. Intervention effect evaluation is a crucial step in self-optimization; by comparing the actual effect with the expected goal, the accuracy of the intervention can be quantified. The intervention bias value, as a loss signal, is backpropagated to the model to guide the update of the feature extraction layer weights, enabling continuous learning and optimization. This closed-loop feedback mechanism ensures continuous improvement through practical application, enhancing the accuracy of bullying identification and the effectiveness of interventions.

[0018] In this embodiment of the invention, the detailed implementation steps for differentially calibrating the timestamps of each modality in the multimodal data stream, identifying the synchronization anchor points of cross-modal interaction events, and performing time-series alignment of the multimodal data stream based on the synchronization anchor points include: The raw timestamp sequence for each modality in the multimodal data stream is extracted. This raw timestamp sequence represents the time points recorded during the acquisition of each modality's data and forms the foundation for time series analysis. The extraction process uses a data parser to uniformly process timestamps of different formats, converting them into a standard time representation. For text dialogue sequences, message sending timestamps are extracted; for speech prosody parameters, audio sampling timestamps are extracted; and for user behavior time series markers, behavior trigger timestamps are extracted. These raw timestamps typically vary in precision and benchmark, requiring further processing before they can be used for cross-modal analysis.

[0019] The interval difference between adjacent timestamps is calculated to construct a timestamp interval sequence for each modality. This sequence reflects the temporal regularity of data acquisition and is a crucial basis for detecting sampling anomalies. The calculation process uses the difference between adjacent timestamps: Interval Difference = Current Timestamp - Previous Timestamp, generating a complete interval sequence. These interval sequences reveal the sampling frequency and stability of data for each modality, providing a reference for subsequent identification of temporal anomalies. Under normal circumstances, the sampling interval for the same modality should be relatively stable; intervals significantly deviating from the normal range may indicate anomalies or significant events during data acquisition.

[0020] A sliding window statistical analysis is performed on the timestamp interval sequence to identify time periods with abnormally fluctuating intervals, denoted as intramodal time series jitter zones. Intramodal time series jitter zones represent periods of unstable sampling that may affect the temporal accuracy of the data. The analysis employs a sliding window technique to calculate the mean and standard deviation of the intervals within the window, identifying fluctuations outside the normal range. The judgment criterion typically uses the Z-score method, marking regions deviating from the mean by more than a preset threshold (usually 2.5 standard deviations) as jitter zones. These jitter zones reflect instability during data acquisition and require special processing during subsequent alignment to ensure temporal accuracy.

[0021] The process involves identifying the minimum time offset between event trigger times across different modalities. Cross-modal event pairs with time offsets less than a preset synchronization threshold are marked as candidate synchronization anchors. Candidate synchronization anchors are data point pairs in different modalities that may correspond to the same actual event, and are crucial for achieving cross-modal alignment. The search process identifies event points with distinct characteristics in the time series of different modalities, such as the start of text input, sudden changes in voice volume, and user operation triggers. The time differences of these events in different modalities are calculated, and event pairs with smaller time offsets are selected as candidates. The preset synchronization threshold is typically determined based on latency characteristics, generally within the range of 50-200 milliseconds, reflecting the maximum possible latency difference between different modal acquisition systems.

[0022] Semantic similarity is performed on the event content of candidate synchronization anchors to calculate the content relevance of cross-modal events. Content relevance assesses the semantic consistency of candidate synchronization anchors, verifying whether they truly correspond to the same event. The calculation employs a cross-modal semantic matching algorithm, extracting features and semantically mapping different modalities to generate comparable representations, and then calculating their similarity. For text and speech modalities, speech-to-text technology is used to extract common semantic content; for behavioral modalities, the contextual meaning or triggering reason is extracted. The similarity calculation formula is: Content relevance = cosine similarity between the event representation of modality 1 and the event representation of modality 2; The content relevance value ranges from [-1, 1]. The closer the value is to 1, the stronger the content relevance and the more reliable the cross-modal event correspondence.

[0023] Candidate synchronization anchors with content relevance greater than a preset relevance threshold are confirmed as synchronization anchors, and linear interpolation calibration is performed on intra-modal temporal jitter regions using these synchronization anchors as a benchmark. Synchronization anchors are the final confirmed cross-modal alignment reference points used for overall temporal alignment. The preset relevance threshold is typically set to 0.7-0.85 to ensure the selected synchronization anchors have sufficiently high reliability. After confirmation, these synchronization anchors are used as fixed reference points to correct the previously identified intra-modal temporal jitter regions. Calibration uses linear interpolation, assuming that the time drift within the jitter region is linear. Based on the alignment relationship between the preceding and following synchronization anchors, the calibration timestamp of each data point within the jitter region is recalculated. This method maintains the temporal alignment of key events while smoothly handling time deviations in intermediate regions, ensuring that multimodal data is analyzed subsequently under a unified time benchmark.

[0024] In this embodiment of the invention, the detailed implementation steps for locating the inconsistency interval between semantic expression and sentiment tendency based on the slope abrupt change point of the differential evolution trajectory, and calculating the semantic deviation index within the inconsistency interval, include: First-order differencing is performed on the differential evolution trajectory to obtain the instantaneous slope sequence. The instantaneous slope sequence reflects the rate of change in the relationship between semantic expression and sentiment tendency, serving as a direct indicator for identifying abrupt changes in the relationship. The processing employs adjacent-point differencing: instantaneous slope = (current difference value - previous difference value) / time interval, generating a complete slope sequence. These slope values ​​quantify the dynamic characteristics of the semantic-sentiment relationship, helping to accurately capture the turning points of relationship patterns. In normal interactions, semantics and sentiment typically maintain a relatively stable relationship, with small fluctuations in the instantaneous slope; however, in abnormal interactions such as bullying behavior, this relationship may change drastically, manifesting as significant fluctuations in the slope values.

[0025] Calculate the rate of change of adjacent slope values ​​in the instantaneous slope sequence, and identify time points where the absolute value of the rate of change exceeds a preset abrupt change threshold, denoted as slope abrupt change points. Slope abrupt change points signify a sharp shift in semantic-emotional relationships, typically corresponding to changes in interaction characteristics. The formula for calculating the rate of change is: Rate of change = (Current slope value - Previous slope value) / Absolute value of the previous slope value; This method of calculating relative change rates effectively captures proportional changes and avoids biases caused by differences in absolute values. The preset abrupt change threshold is typically set to 0.4-0.6, determined based on the statistical distribution of slope changes under normal interaction mode. Points where the absolute value of the change rate exceeds the threshold are marked as slope abrupt change points; these points are key boundary markers for locating inconsistencies.

[0026] Using the slope abrupt change point as the boundary, the differential evolution trajectory is divided into intervals. Intervals where the semantic feature vector and the sentiment feature vector change in opposite directions are selected and marked as inconsistent intervals. Inconsistent intervals represent periods where semantic and sentiment expressions are inconsistent, potentially implying bullying behavior. The segmentation process uses the identified slope abrupt change point as the boundary, dividing the differential evolution trajectory into multiple continuous intervals. Then, the changing trends of the semantic feature vector and the sentiment feature vector within each interval are analyzed, paying particular attention to intervals where the two change in opposite directions, such as positive semantic expression with negative sentiment, or negative semantic expression with exaggerated positive emotional response. This inconsistency is an important feature for identifying covert bullying rhetoric, because bullies often use seemingly neutral or positive language to convey negative intentions.

[0027] Within the inconsistency interval, the product of the polarity label of the semantic feature vector and the sentiment extremum of the sentiment tendency feature vector is calculated and denoted as the polarity deviation. The polarity deviation quantifies the degree of inconsistency between semantics and sentiment and is a key indicator for assessing inconsistency. The calculation process first extracts the polarity label from the semantic feature vector, representing the basic tendency of the text (positive, neutral, or negative); then, it extracts the sentiment extremum from the sentiment tendency feature vector, representing the peak value of sentiment intensity and its direction. The product of the two reflects the harmony between semantics and sentiment; a positive value indicates consistency, and a negative value indicates deviation. The closer the polarity deviation is to -1, the stronger the opposition between semantics and sentiment, and the more obvious the potential bullying tendency.

[0028] The percentage of samples with negative polarity deviation within inconsistent intervals is statistically analyzed. Combined with the duration of the inconsistent interval, a semantic deviation index is calculated. The semantic deviation index is determined by a weighted product of the sample percentage and the duration. The semantic deviation index is a comprehensive quantitative indicator for assessing bullying characteristics within inconsistent intervals, considering both the universality and persistence of the inconsistency. The calculation formula is: Semantic deviation index = percentage of negative samples × (1 + duration factor × duration of inconsistency interval); The duration factor is a weighted parameter determined based on the characteristics of the interaction scenario, reflecting the influence of duration on bullying determination. This indicator comprehensively considers the frequency and persistence of inconsistencies, effectively distinguishing between brief misunderstandings and systemic bullying behavior. The semantic deviation index typically ranges from [0,1], with values ​​closer to 1 indicating more pronounced potential bullying characteristics.

[0029] In this embodiment of the invention, the detailed implementation steps for constructing a role relationship graph containing power gradients and interaction frequencies, and labeling weak and strong nodes in the role relationship graph, include: This study extracts the number of dialogue initiations, response delays, and utterance lengths from historical dialogue relationship data to construct a participant behavioral feature matrix. The behavioral feature matrix is ​​a data structure that quantifies participant interaction patterns, reflecting role characteristics in social interactions. The extraction process focuses on each participant, statistically analyzing key indicators such as dialogue initiation frequency (initiative), average response time to others' messages (responsiveness), and typical utterance length (expressiveness). This data is organized into a multidimensional matrix, where rows represent participants, columns represent behavioral features, and cell values ​​are corresponding statistics. The behavioral feature matrix reveals differences in the behavioral patterns of interacting participants, providing a quantitative basis for role relationship analysis.

[0030] The initiative score for each participant in the behavioral feature matrix is ​​calculated. The initiative score is determined by a weighted sum of the number of conversations initiated and the average utterance length. The initiative score is a comprehensive indicator measuring a participant's dominance in the interaction, reflecting their influence and control. The calculation formula is as follows: Initiative score = α × number of standardized dialogue initiations + β × standard average utterance length; Here, α and β are weighting coefficients, determined based on the characteristics of the interaction scenario. Typically, α is slightly larger than β, reflecting a greater contribution of the initiating behavior to initiative. Standardization eliminates differences in units, making different indicators comparable. Participants with higher initiative scores are usually more proactive in controlling the topic and pace of the interaction, potentially corresponding to a dominant role.

[0031] The frequency of interactions between participants is statistically analyzed to construct an interaction frequency adjacency matrix. Each element in the matrix represents the number of interactions between two participants. This matrix describes the interaction density network among participants, reflecting the strength distribution of social relationships. Each element (i,j) in the matrix represents the number of interactions initiated by participant i to participant j, forming a typically asymmetric square matrix. This adjacency matrix structure can intuitively display the topological characteristics of the interaction network, such as interaction concentration, reciprocity, and preference. High interaction frequency indicates strong connections, but it requires comprehensive analysis in conjunction with directionality and power gradients to determine its significance in bullying contexts.

[0032] Based on the difference in initiative scores, the power gradient value between participant pairs is calculated. The power gradient value is determined by the ratio of the participant with the higher initiative score to the participant with the lower initiative score. The power gradient value quantifies the strength of the dominance-subordination relationship between participants and is an important indicator for identifying potential bullying relationships. For any participant pair (i,j), if initiative score i > initiative score j, then the power gradient value is calculated: The power gradient value (i,j) = initiative score i / initiative score j; This ratio reflects the relative dominance of one party over the other; a larger value indicates a more significant power imbalance. Relationships with a power gradient value greater than a certain threshold (usually 2.0) may indicate a power imbalance, requiring close attention to the quality of interaction to identify potential bullying patterns.

[0033] Multiply the power gradient value by the corresponding element value in the interaction frequency adjacency matrix to generate a weighted relation matrix. A role relationship graph is then constructed based on this weighted relation matrix. The weighted relation matrix integrates both power differences and interaction frequency dimensions, comprehensively describing the nature of the relationships between participants. The matrix elements are calculated as follows: Weighted relation value (i,j) = power gradient value (i,j) × interaction frequency (i,j); This weighted strategy emphasizes relationships with high-frequency interactions and significant power disparities, which represent high-risk scenarios for potential bullying. Based on the weighted relationship matrix, a visual role relationship graph is constructed, where nodes represent participants, edges represent relationships, and the thickness and color of the edges reflect the magnitude of the weighted relationship values, intuitively displaying the power distribution and interaction patterns in the interaction network.

[0034] Calculate the sum of the in-degree weights and the sum of the out-degree weights for each node in the role relationship graph. Nodes whose sum of in-degree weights is greater than the sum of their out-degree weights and whose ratio exceeds a preset weakness threshold are marked as weak nodes, while nodes whose ratio is less than a preset strength threshold are marked as strong nodes. The in-degree and out-degree weights reflect the degree to which a node is affected and exerts influence, respectively, and their ratio reveals the node's role in the network. The calculation formula is as follows: The sum of in-degree weights equals the sum of the weighted relation values ​​pointing to that node; Out-degree weight sum = the sum of the weighted relation values ​​of all other nodes pointed to by this node; Impact ratio = sum of in-degree weights / sum of out-degree weights; The preset vulnerability threshold is typically 1.5-2.0, and the preset dominance threshold is typically 0.5-0.7. Nodes with an influence ratio greater than the vulnerability threshold are marked as vulnerable nodes, indicating that the participant is primarily influenced by others; nodes with an influence ratio less than the dominance threshold are marked as dominant nodes, indicating that the participant primarily influences others. This role-classification method based on network flow analysis can accurately identify power structures in interactive networks, providing social relationship dimension support for determining bullying behavior.

[0035] In this embodiment of the invention, the detailed implementation steps for generating a chain of evidence for bullying behavior based on semantic deviation index, duration of inconsistency intervals, and node power differences in the role relationship graph include: Time periods with semantic deviation indices exceeding a preset deviation threshold are marked as suspected bullying periods. These suspected bullying periods represent the initial screening of time ranges potentially containing bullying behavior and form the basis for constructing the chain of evidence. The preset deviation threshold is typically set between 0.65 and 0.75, determined based on the statistical distribution of the semantic deviation index and expert-annotated samples. This initial screening process employs a sliding window technique to evaluate semantic deviation over a continuous time stream, marking consecutive time periods exceeding the threshold as candidate regions. Suspected bullying periods may overlap or be discontinuous; subsequent analysis will further refine the boundaries and merge related regions.

[0036] During the suspected bullying period, the participant node pairs that interacted are extracted, and the power difference between these pairs in the role relationship graph is calculated. The power difference reflects the difference in social status between the interacting parties and is an important basis for judging potential bullying relationships. The extraction process first identifies all interaction events and their participants during the suspected period, and then determines the role relationship between each pair of interacting participants based on the node labels in the role relationship graph. For each pair of participants (i,j), their power difference is calculated: Weight difference (i,j) = |out-degree weight sum i / in-degree weight sum i - out-degree weight sum j / in-degree weight sum j|; This calculation method takes into account the power position of participants in the overall network. The larger the power difference, the more significant the difference in social status between the two parties, and the higher the risk of bullying may be.

[0037] Pairs of participant nodes whose power difference exceeds a preset power difference threshold are designated as unequal interaction pairs. These pairs represent combinations of participants with significant power imbalances and are key targets for monitoring potential bullying relationships. The preset power difference threshold is typically set between 1.0 and 1.5, reflecting the minimum power gap that could lead to bullying behavior. This screening step further narrows the scope of analysis, focusing on interaction relationships with problematic power structures. The identification of unequal interaction pairs combines structural information from graph analysis with dynamic information from real-time interactions, accurately capturing abnormal interaction patterns under conditions of power inequality.

[0038] For each unequal interaction pair, the duration of the corresponding inconsistency interval is extracted, and the product of the duration and the semantic deviation index is calculated, denoted as the evidence node strength. Evidence node strength comprehensively considers the degree and persistence of semantic inconsistency and is a comprehensive indicator for evaluating the credibility of a single evidence point. The calculation formula is as follows: Evidence node strength = semantic deviation index × (1 + γ × log(duration / baseline duration)); Here, γ is the duration weighting factor, typically set to 0.3-0.5, and the baseline duration is a scene-related reference value. This calculation model uses logarithmic transformation to process the duration, considering both the impact of duration and avoiding excessive dominance of results by long time periods. Evidence node strength provides a weighted evaluation of single-point evidence, providing a basic unit for constructing the chain of evidence.

[0039] Evidence nodes with strength exceeding a preset strength threshold are arranged chronologically to construct an initial evidence sequence. This initial evidence sequence is a set of valid evidence arranged along the time dimension, reflecting the temporal distribution of suspected bullying behavior. The preset strength threshold is typically set to 0.4-0.6 to ensure the included evidence nodes have sufficient credibility. The chronological arrangement preserves the context of event development, aiding in the identification of systemic bullying patterns and progressive bullying behaviors. The initial evidence sequence may contain discrete evidence points; these will be merged and clustered based on temporal correlation in the next step.

[0040] The time interval between adjacent evidence nodes in the initial evidence sequence is calculated, and adjacent evidence nodes with time intervals less than a preset association time window are merged into evidence clusters. An evidence cluster is a set of temporally densely related evidence, representing a potential ongoing bullying event. The preset association time window is a threshold for judging the temporal relevance of evidence, typically set between 60 and 180 seconds, determined based on the rhythmic characteristics of the interaction scenario. The merging process uses a clustering algorithm to group evidence nodes with similar times into the same evidence cluster, while simultaneously calculating the cluster's start and end times, involved participants, and cumulative strength. The formation of evidence clusters strengthens the temporal continuity of the evidence, filters out isolated potential misjudgments, and enhances the reliability of the judgment.

[0041] Based on the temporal distribution of evidence clusters and the strength of evidence nodes, a bullying behavior evidence chain is generated. This chain includes the start and end times of each evidence cluster, the nodes involved, and their cumulative strength. The bullying behavior evidence chain serves as the final basis for bullying determination, integrating multi-dimensional and multi-time-period evidence information. The generation process organizes evidence clusters according to time sequence, constructs a complete evidence framework, and calculates the overall characteristics of the chain, such as total duration, average strength, and fluctuation trends. The construction of the evidence chain not only focuses on the strength of individual pieces of evidence but also considers the temporal correlation and pattern characteristics between evidence, effectively distinguishing between occasional conflicts and systemic bullying behavior. The complete evidence chain provides a comprehensive factual basis for intervention decisions, supporting accurate and reasonable responses.

[0042] In this embodiment of the invention, the detailed implementation steps for weighted fusion of the confidence levels of each evidence node in the evidence chain to calculate the overall strength value of the evidence chain, and for determining the intervention response level based on the overall strength value and the continuity density of the evidence nodes, include: For each evidence node in the evidence chain, a semantic aggression score is calculated based on the semantic understanding results of the corresponding interactive text by the large model. The semantic aggression score is a direct assessment of the bullying characteristics of the text content, reflecting the aggressiveness and harmfulness of the language expression. The scoring process utilizes the deep semantic understanding capabilities of the large model to analyze the implicit intentions, ironic elements, and potential harm of the text, generating a quantitative score within the range [0,1]. The scoring criteria consider various bullying language features, including derogatory remarks, threatening expressions, identity discrimination, and exclusionary language. This score provides supplementary evidence from the content dimension, complementing the previous feature vector-based analysis and enhancing the comprehensiveness and reliability of the judgment.

[0043] By combining semantic aggression scores and evidence node strength, the confidence level of each evidence node is calculated. The confidence level is determined by the geometric mean of the semantic aggression score and the evidence node strength. This integration of results from feature analysis and semantic understanding provides a more balanced assessment of evidence reliability. The calculation formula uses the geometric mean, ensuring that both factors reach a certain level to achieve a high confidence level, avoiding the dominance of a single factor. This fusion strategy is particularly suitable for tasks like bullying determination that require multi-dimensional evidence support, improving robustness and resistance to interference.

[0044] The confidence scores of all evidence nodes in the chain of evidence are weighted by time decay, with higher weights for evidence nodes closer to the current time. The sum of the weighted confidence scores is then calculated and denoted as the overall strength value. The overall strength value is a quantitative assessment of the overall credibility of the chain of evidence, taking into account the impact of time on the importance of evidence. The weighting formula is: Weighted confidence level = confidence level × exp(-λ × (current time - evidence time)); Here, λ is the time decay factor, typically set to 0.01-0.05 / minute. This time-weighted strategy emphasizes the importance of recent evidence, aligning with the real-time requirements of interactive systems, enabling timely responses to the latest bullying events, while not completely ignoring historical patterns. The overall strength value, serving as the overall score of the evidence chain, is the primary quantitative basis for intervention decisions.

[0045] The number of evidence nodes within a time span of the statistical evidence chain is counted, and the continuity density of these nodes is calculated. Continuity density is determined by the ratio of the number of evidence nodes to the time span. Continuity density reflects the persistence and intensity of bullying incidents and is an important dimension for assessing the severity of bullying. The calculation formula is as follows: Continuity density = Number of evidence nodes / Time span of the evidence chain; This indicator is measured in nodes per minute, with higher values ​​indicating more frequent and persistent bullying. Continuity density can distinguish between sudden conflicts and ongoing bullying; the former may have high node intensity but low density, while the latter typically presents as continuous evidence of moderate intensity. This indicator provides a quantitative description of bullying time patterns, supplementing the limitations of intensity assessment.

[0046] A two-dimensional assessment coordinate system is established, integrating overall intensity and continuous density. Based on the pre-defined zones where the coordinate points are located, intervention response levels are determined, including observation, alert, warning, and blocking levels. This two-dimensional assessment coordinate system integrates the two key dimensions of bullying behavior—intensity and persistence—providing a more comprehensive basis for grading. The coordinate system zones are determined based on expert experience and historical case analysis, typically dividing the plane into four quadrants: low-intensity, low-density zones correspond to the observation level, requiring continuous monitoring but not immediate intervention; low-intensity, high-density zones or high-intensity, low-density zones correspond to the alert level, providing mild intervention reminders; high-intensity, medium-density zones correspond to the warning level, issuing explicit warnings; and high-intensity, high-density zones correspond to the blocking level, implementing mandatory intervention measures. This multi-level response mechanism ensures the targeted and appropriate nature of interventions, effectively preventing severe bullying while avoiding excessive intervention in normal interactions.

[0047] In this embodiment of the invention, the detailed implementation steps for calculating the intervention deviation value based on interaction state change data and preset intervention effect evaluation indicators, and backpropagating the intervention deviation value to the feature extraction layer of the large model, include: Within a pre-defined observation window following the intervention, new interaction data between vulnerable and dominant participants is collected. This new interaction data reflects changes in participants' behavior and emotions after the intervention and serves as a direct basis for evaluating the intervention's effectiveness. The pre-defined observation window is typically 3-10 minutes, determined based on the characteristics of the interaction scenario and the intervention level. The data collection process continuously monitors all messages and behaviors in the interaction channels, recording complete time series and multimodal characteristics. Particular attention is paid to changes in the emotions of vulnerable participants and behavioral adjustments in dominant participants, as these are core indicators of intervention effectiveness. The new data is compared with baseline data before the intervention to provide a basis for effectiveness evaluation.

[0048] Sentiment feature vectors were extracted from newly added interaction data, and the change in sentiment tendency of vulnerable nodes before and after the intervention was calculated. The change in sentiment tendency is one of the main indicators for evaluating the intervention's effectiveness, reflecting the impact of the intervention on the psychological state of the vulnerable party. The extraction process used the same sentiment analysis methods as the previous steps to ensure data comparability. The method for calculating the change was as follows: Change in emotional tendency = Average emotional value of vulnerable nodes after intervention - Average emotional value of vulnerable nodes before intervention; Positive values ​​indicate improved emotional state, while negative values ​​indicate worsened emotional state. This indicator directly reflects the psychological effect of the intervention and is a key indicator for measuring its effectiveness. An ideal intervention should improve the psychological state of the disadvantaged party, manifested as significant positive emotional changes.

[0049] Based on the expected affective improvement value corresponding to the intervention response level, the difference between the change in affective tendency and the expected affective improvement value is calculated and denoted as the affective bias component. The affective bias component quantifies the degree of deviation between the intervention effect and the expected goal, and is an important basis for adjusting the intervention strategy. The expected affective improvement value is determined according to the intervention response level; the higher the level, the greater the expected improvement value, reflecting the difference in goals between interventions of different intensities. The formula for calculating the difference is: Affective bias component = Change in affective tendency - Expected improvement in affective value; Negative values ​​indicate that the intervention effect is less than expected, while positive values ​​indicate that the effect exceeds expectations. This bias indicator provides precise feedback signals to guide the adjustment of subsequent intervention strategies and model parameters.

[0050] The frequency of aggressive discourse by dominant individuals after the intervention was compared with the pre-set frequency reduction target to calculate the frequency deviation component. The frequency deviation component assesses the inhibitory effect of the intervention on the dominant individual's behavior and is another important dimension of the intervention's effectiveness. The statistical process continuously monitors the verbal expressions of dominant individuals, identifies the frequency of aggressive content, and compares it with the pre-intervention baseline and the pre-set target. The frequency reduction target is determined based on the intervention response level, reflecting the expected effects of interventions of different intensities. The deviation calculation formula is as follows: Frequency deviation component = (Frequency before intervention - Frequency after intervention) - Target value of frequency reduction; Negative values ​​indicate insufficient suppression, while positive values ​​indicate suppression exceeding expectations. This indicator directly reflects the effectiveness of the intervention in controlling the root causes of bullying behavior and is an important basis for evaluating the targeting and effectiveness of intervention measures.

[0051] The weighted sum of the affective bias component and the frequency bias component is used as the intervention bias value. The intervention bias value is a comprehensive indicator for evaluating the intervention effect, integrating the protective effect on the vulnerable party and the inhibitory effect on the dominant party. The calculation formula is: Intervention bias value = δ × emotional bias component + (1-δ) × frequency bias component; Here, δ is the weighting coefficient, typically set to 0.4-0.6, reflecting the relative importance of the two dimensions in the evaluation. A negative intervention bias value indicates that the overall intervention effect is lower than expected, requiring stronger intervention strategies or optimization of model parameters; a positive value indicates that the intervention effect is good, and the current strategy can be maintained or moderately adjusted. This comprehensive evaluation mechanism ensures a holistic assessment of the intervention effect, avoiding biased judgments based on a single dimension.

[0052] The intervention bias value is used as the loss gradient and backpropagated to the semantic feature vector and sentiment feature vector extraction layers in the larger model to update the attention weight matrix of the feature extraction layer. Backpropagation is the core mechanism of self-optimization, transforming effect evaluation into a key step in parameter updating. The propagation process treats the intervention bias value as the negative gradient of the loss function, calculates the gradient of each layer's parameters using the chain rule, and focuses on optimizing the attention weight matrix of the feature extraction layer. The update formula is: New weights = Old weights - Learning rate × Gradient of intervention bias; The learning rate is dynamically adjusted based on model size and data characteristics to ensure the stability and efficiency of the learning process. This parameter update mechanism, based on actual intervention effects, can continuously learn from experience, improving the accuracy of bullying identification and the effectiveness of intervention.

[0053] The model records the feature combination patterns of interactive samples that trigger intervention bias values. These feature combination patterns are then added to the difficult sample memory of the large model, increasing their sampling weight in subsequent training. The difficult sample memory is a crucial component of long-term optimization, storing features of cases that are difficult to identify or where intervention is ineffective. The recording process extracts key feature combinations from interactive samples, including semantic-sentiment patterns, role-relationship features, and temporal evolution characteristics, forming structured feature vectors. These feature vectors are assigned higher sampling weights and appear more frequently in subsequent training, allowing the model to concentrate resources on optimizing its ability to handle difficult cases. This memory-enhanced learning strategy significantly improves the ability to identify boundary cases and complex bullying patterns, enhancing the overall robustness and adaptability of the model's performance.

[0054] This invention achieves accurate identification and timely response to bullying behavior in interactive scenarios through multimodal data analysis, temporal feature extraction, social relationship graph construction, and adaptive intervention mechanisms. The closed-loop optimization feature of this invention continuously adjusts model parameters based on the intervention effect, improving adaptability and effectiveness over time, and providing a safe and healthy communication environment for participants in various interactive scenarios.

[0055] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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.

[0056] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0057] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for bullying behavior recognition and real-time response based on a large model, characterized in that, include: Acquire multimodal data streams in interactive scenarios, the multimodal data streams including text dialogue sequences, speech prosody parameters, and user behavior temporal markers; Differential calibration is performed on the timestamps of each modality in the multimodal data stream to identify the synchronization anchor points of cross-modal interaction events, and the multimodal data stream is time-aligned according to the synchronization anchor points. Based on the time-aligned data, extract the text semantic feature vector and the sentiment feature vector, and calculate the difference evolution trajectory of the semantic feature vector and the sentiment feature vector in the time dimension. Based on the slope abrupt change point of the differential evolution trajectory, locate the inconsistency interval between semantic expression and sentiment tendency, and calculate the semantic deviation index within the inconsistency interval; Extract historical dialogue relationship data between interactive participants, construct a role relationship graph containing power gradient and interaction frequency, and label the weak and strong nodes in the role relationship graph; Based on the semantic deviation index, the duration of the inconsistency interval, and the node power difference of the role relationship graph, a chain of evidence for bullying behavior is generated. The confidence levels of each evidence node in the evidence chain are weighted and fused to calculate the overall strength value of the evidence chain. The intervention response level is determined based on the overall strength value and the continuity density of the evidence nodes. The corresponding real-time intervention strategy is triggered based on the intervention response level, and data on changes in the interaction status after the intervention are collected. Based on the interaction state change data and the preset intervention effect evaluation index, the intervention deviation value is calculated, and the intervention deviation value is backpropagated to the feature extraction layer of the large model to update the extraction weights of the semantic feature vector and the sentiment tendency feature vector.

2. The method for bullying behavior recognition and real-time response based on a large model according to claim 1, characterized in that, The step of differentially calibrating the timestamps of each modality in the multimodal data stream to identify synchronization anchors for cross-modal interaction events includes: Extract the original timestamp sequence of each modality data in the multimodal data stream; Calculate the interval difference between adjacent timestamps and construct the timestamp interval sequence for each modality; Sliding window statistical analysis is performed on the timestamp interval sequence to identify time periods with abnormal interval fluctuations, which are denoted as intramodal time series jitter zones; Find the minimum time offset of the event triggering time between different modalities, and mark cross-modal event pairs with time offsets less than a preset synchronization threshold as candidate synchronization anchors; Semantic similarity verification is performed on the event content of the candidate synchronization anchors, and the content correlation degree of cross-modal events is calculated; Candidate synchronization anchors whose content relevance is greater than a preset relevance threshold are identified as synchronization anchors, and linear interpolation calibration is performed on the intramodal timing jitter region based on the synchronization anchors.

3. The method for bullying behavior recognition and real-time response based on a large model according to claim 1, characterized in that, The step of locating the inconsistency interval between semantic expression and sentiment tendency based on the slope abrupt change point of the differential evolution trajectory, and calculating the semantic deviation index within the inconsistency interval includes: The differential evolution trajectory is subjected to first-order difference processing to obtain the instantaneous slope sequence of the differential evolution trajectory; Calculate the rate of change of adjacent slope values ​​in the instantaneous slope sequence, identify the time point where the absolute value of the rate of change is greater than a preset abrupt change threshold, and record it as the slope abrupt change point; Using the slope abrupt change point as the boundary, the interval segments of the differential evolution trajectory are divided, and the interval segments in which the semantic feature vector and the sentiment tendency feature vector change in opposite directions are selected and marked as the non-consistent intervals. Within the inconsistency interval, the product of the polarity label of the semantic feature vector and the emotional extreme value of the sentiment tendency feature vector is calculated and denoted as the polarity deviation degree. The percentage of samples with negative polarity deviation within the inconsistency interval is statistically analyzed. Combined with the duration of the inconsistency interval, the semantic deviation index is calculated. The semantic deviation index is determined by the weighted product of the sample percentage and the duration.

4. The method for bullying behavior recognition and real-time response based on a large model according to claim 1, characterized in that, The construction of a role relationship graph including power gradient and interaction frequency, and the labeling of weak and strong nodes in the role relationship graph, includes: Extract the number of dialogue initiations, response delay time, and utterance length between each participant from the historical dialogue relationship data, and construct a participant behavior feature matrix; Calculate the initiative score for each participant in the behavioral feature matrix, which is determined by a weighted sum of the number of dialogues initiated and the average utterance length; The frequency of interactions between participants is counted, and an interaction frequency adjacency matrix is ​​constructed. The element values ​​of the interaction frequency adjacency matrix represent the number of interactions between two participants. Based on the difference in initiative scores, a power gradient value is calculated between participant pairs, which is determined by the ratio of the participant with a higher initiative score to the participant with a lower initiative score. Multiply the power gradient value with the corresponding element value in the interaction frequency adjacency matrix to generate a weighted relation matrix, and construct the role relationship graph based on the weighted relation matrix; Calculate the sum of in-degree weights and the sum of out-degree weights for each node in the role relationship graph. Nodes whose sum of in-degree weights is greater than the sum of out-degree weights and whose ratio exceeds a preset weakness threshold are marked as weak nodes, and nodes whose ratio is less than a preset strength threshold are marked as strong nodes.

5. The method for bullying behavior recognition and real-time response based on a large model according to claim 1, characterized in that, The process of generating a bullying behavior evidence chain based on the semantic deviation index, the duration of the inconsistency interval, and the node power difference in the role relationship graph includes: The time periods when the semantic deviation index is greater than a preset deviation threshold are marked as suspected bullying periods; During the suspected bullying period, extract the participant node pairs that interacted and calculate the power difference of the participant node pairs in the role relationship graph; Participant node pairs whose power difference is greater than a preset power difference threshold are selected and recorded as unequal interaction pairs. For each unequal interaction pair, the duration of the corresponding inconsistency interval is extracted, and the product of the duration and the semantic deviation index is calculated and recorded as the evidence node strength. Arrange the evidence nodes whose strength is greater than a preset strength threshold in chronological order to construct an initial evidence sequence; Calculate the time interval between adjacent evidence nodes in the initial evidence sequence, and merge adjacent evidence nodes with time intervals less than a preset association time window into evidence clusters; Based on the temporal distribution of the evidence clusters and the strength of the evidence nodes, the bullying behavior evidence chain is generated. The evidence chain includes the start and end times of each evidence cluster, the nodes involved, and the cumulative strength.

6. The method for bullying behavior recognition and real-time response based on a large model according to claim 1, characterized in that, The step of weightedly fusing the confidence levels of each evidence node in the evidence chain to calculate the overall strength value of the evidence chain, and determining the intervention response level based on the overall strength value and the continuity density of the evidence nodes, includes: For each evidence node in the chain of evidence, a semantic aggression score is calculated based on the semantic understanding results of the corresponding interactive text by the large model. The confidence level of the evidence node is calculated by combining the semantic attack score and the strength of the evidence node. The confidence level is determined by the geometric mean of the semantic attack score and the strength of the evidence node. The confidence scores of all evidence nodes in the chain of evidence are weighted by time decay, with higher weights for evidence nodes closer to the current time. The total confidence scores after weighting are calculated and denoted as the overall strength value. The number of evidence nodes within the time span of the evidence chain is counted, and the continuity density of the evidence nodes is calculated. The continuity density is determined by the ratio of the number of evidence nodes to the time span. A two-dimensional evaluation coordinate system is established for the overall intensity value and the continuity density. The intervention response level is determined according to the preset partition where the coordinate point is located. The intervention response level includes observation level, prompt level, warning level and blocking level.

7. The method for bullying behavior recognition and real-time response based on a large model according to claim 1, characterized in that, The step of calculating the intervention deviation value based on the interaction state change data and preset intervention effect evaluation indicators, and backpropagating the intervention deviation value to the feature extraction layer of the large model, includes: Within a preset observation window after the intervention is triggered, new interaction data between the weak node and the strong node is collected. Extract the sentiment tendency feature vector of the newly added interaction data, and calculate the change in sentiment tendency of the vulnerable node before and after the intervention; Based on the expected emotional improvement value corresponding to the intervention response level, the difference between the change in emotional tendency and the expected emotional improvement value is calculated and denoted as the emotional deviation component. The frequency of aggressive statements from the dominant nodes after statistical intervention is compared with the preset frequency reduction target value, and the frequency deviation component is calculated. The weighted sum of the emotional deviation component and the frequency deviation component is used as the intervention deviation value; The intervention deviation value is used as the loss gradient and backpropagated to the extraction layer of the semantic feature vector and the sentiment tendency feature vector in the large model to update the attention weight matrix of feature extraction. Record the feature combination patterns of the interactive samples that trigger the intervention deviation value, add the feature combination patterns to the hard sample memory of the large model, and increase the sampling weight of the feature combination patterns in subsequent training.

Images