An AI-based live content review method and system
By constructing a cognitive intelligent review framework and utilizing multi-dimensional cognitive feature extraction and deep learning technology, the problems of deep intent understanding and advanced violation identification in live content review are solved, enabling proactive defense and precise blocking, and providing transparent review decision support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHENGZHOU PENGXIN NETWORK TECHNOLOGY CO LTD
- Filing Date
- 2026-03-06
- Publication Date
- 2026-06-05
Smart Images
Figure CN122153484A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of online live streaming content supervision technology, and in particular relates to an AI-based live streaming content review method and system. Background Technology
[0002] With the rapid development of the online live streaming industry, live streaming platforms are facing increasingly severe challenges in content moderation. Existing live streaming content moderation technologies mainly employ single-dimensional methods such as keyword filtering, image recognition, and audio detection, which have the following technical shortcomings: First, existing technology can only identify surface-level violations and cannot understand the deeper intentions and implicit meanings of the streamers, resulting in a large number of disguised and suggestive violations escaping review. Second, traditional review methods lack a holistic cognitive analysis of the streamer's behavior patterns and are unable to identify advanced violations such as cognitive deception and psychological manipulation. Third, the existing system adopts a passive detection mechanism, which can only identify violations after they have already occurred, lacking predictive and proactive defense capabilities. Fourth, current technology cannot effectively analyze the spread and influence mechanisms of illegal content among audiences, making it difficult to achieve precise blocking.
[0003] Therefore, there is an urgent need for an intelligent review technology solution that can deeply understand the cognitive level of live streaming content and has proactive defense capabilities. Summary of the Invention
[0004] To address the aforementioned technical issues, this invention provides an AI-based live streaming content review method and system. By constructing a cognitive intelligent review framework, it achieves a deep understanding of live streaming content and proactive defense.
[0005] In a first aspect, the present invention provides an AI-based method for reviewing live streaming content, comprising the following steps: S1. Multidimensional Cognitive Feature Extraction: Acquire live stream data and simultaneously extract the anchor's voice features, facial expression features, body movement features, and scene environment features through a cognitive consistency analysis model to construct a multidimensional cognitive feature vector; by calculating the deviation between the current feature correlation matrix and the normal behavior benchmark matrix, detect abnormal imbalances between feature dimensions and generate cognitive imbalance indicators. S2. Micro-expression intention recognition: High-frequency sampling analysis of facial expression features is performed. Micro-expression change sequences are captured through optical flow tracing algorithm. Temporal features of activation intensity of facial muscle movement are extracted. The mapping relationship between activation intensity and implicit intention is learned based on long short-term memory network to infer intention categories including deception, tension, anger, fear and disgust. S3. Semantic Evolution Trajectory Analysis: Real-time semantic encoding of live content, constructing word vector trajectories in semantic space, calculating the distance between the current semantic point and the cluster center of sensitive topics, and generating early warning signals when a drifting trend toward sensitive areas is detected by monitoring drift speed, acceleration and directional persistence. Specifically, in S3, a semantic trajectory is constructed by calculating the distance between the current semantic vector and the cluster center of sensitive topics; the drift speed, acceleration and directional persistence of the semantic trajectory are monitored, and when the trajectory is determined to show a trend of drifting towards the sensitive area based on the drift speed, acceleration and directional persistence, a semantic drift warning signal is generated. S4. Emotional propagation link modeling: Collect anchor's emotional state indicators and real-time audience feedback data, construct an emotion propagation model based on graph neural networks, evaluate the intensity of emotion propagation by calculating the spectral radius of the propagation matrix, and identify inflammatory and manipulative emotion propagation patterns; S5. Reverse Causal Reasoning: Monitor indicators such as sudden changes in the audience's barrage emotions, abnormal gift giving, and a surge in the exit rate. Construct a Bayesian network containing an observation layer, a latent variable layer, and a causal layer. Infer the potential violations of the streamer by calculating the posterior probability. S6. Cognitive Load Assessment: Analyze live broadcast content from three dimensions: information density, switching frequency, and multimodal conflict degree. Calculate the cognitive processing load index of the audience. When the load index exceeds the normal range and sensitive information is detected, it is determined that there is a behavior of using information overload to cover up violations. S7. Blocking the spread of illegal content: Model the spread of illegal content as a directed graph network, identify key nodes by calculating the betweenness centrality, degree centrality and eigenvector centrality of nodes, use a greedy algorithm to select the optimal blocking set, and implement hierarchical blocking such as content delay, fuzzing or link disconnection. S8. Comprehensive Decision Output: Integrate the analysis results of each step to construct a comprehensive feature vector, output the probability distribution of violation types through a multi-layer neural network, calculate the comprehensive violation level and generate disposal suggestions, and use the integrated gradient method to generate an interpretable report.
[0006] Furthermore, the cognitive consistency analysis model adopts a multi-branch deep neural network architecture, integrating multimodal features through a feature fusion layer; the semantic space constructs a sensitive topic distribution through hierarchical clustering of historical violation cases; the emotion propagation model uses the GraphSAGE architecture to learn node representations; and the system includes an incremental learning mechanism.
[0007] Secondly, the present invention provides an AI-based live streaming content review system, comprising: The cognitive feature analysis module is used to extract multi-dimensional cognitive features from the live stream. It analyzes the consistency relationship between voice, facial expressions, actions, and scenes through deep learning networks to detect cognitive dissonance. The micro-expression recognition module is used to capture micro-expression changes through optical flow tracing algorithm, extract the temporal features of activation intensity of facial muscle movements, and infer the implicit intention using a long short-term memory network; The semantic trajectory monitoring module is used to encode live semantic content in real time, calculate the distance from semantic vectors to sensitive topic clusters, and generate early warnings by analyzing drift dynamic indicators; The emotion propagation analysis module is used to construct a dynamic graph structure of the audience group, analyze the emotion propagation path through graph neural networks, and calculate the propagation intensity. The reverse reasoning module is used to monitor abnormal audience behavior indicators and calculate the posterior probability of violations through a Bayesian network. The cognitive load calculation module is used to assess cognitive load based on information density, switching frequency, and multimodal conflict degree. The propagation blocking module is used to calculate the centrality index of nodes in the propagation network and select the optimal blocking set. The decision output module is used to integrate the outputs of various modules to generate violation determinations and interpretable reports.
[0008] Furthermore, the modules work together through an event-driven mechanism: when the cognitive feature analysis module detects cognitive dissonance, it triggers the micro-expression recognition module to perform high-frequency analysis; when the semantic trajectory monitoring module detects drift risk, it generates an early warning; when the emotion propagation analysis module detects abnormal propagation, it activates reverse reasoning; and the outputs of each module converge to the decision output module to generate the final judgment. The system adopts a streaming processing architecture to ensure real-time performance.
[0009] Beneficial effects
[0010] 1. Deep cognitive understanding ability: Through cognitive consistency analysis and micro-expression recognition, this invention can break through the limitations of surface content, deeply understand the true intentions and implicit semantics of the anchor, and effectively identify disguised and suggestive illegal content.
[0011] 2. Proactive Prediction and Defense Mechanism: Through semantic trajectory monitoring and sentiment propagation analysis, this invention can identify risk trends before the illegal content is fully presented, achieving a technological upgrade from passive detection to proactive defense.
[0012] 3. Capability to detect concealed violations: Through reverse causal reasoning and cognitive load assessment, this invention can detect concealed violations that are difficult to detect by traditional methods, including advanced violations such as psychological manipulation and cognitive deception.
[0013] 4. Precise dissemination blocking effect: By constructing a network model for the spread of illegal content and a causal intervention mechanism, this invention can accurately locate key dissemination nodes and effectively block the spread and impact of illegal content.
[0014] 5. Comprehensive Review Coverage: Through a multi-dimensional and multi-level comprehensive analysis framework, this invention achieves comprehensive review coverage from the surface level of content to the depth of cognition, and from individual behavior to group influence.
[0015] 6. Explainable decision support: By generating detailed reasoning paths and evidence descriptions, this invention provides a transparent and traceable review decision process, which facilitates manual review and appeal handling. Attached Figure Description
[0016] Figure 1 A flowchart illustrating the steps of the method described in this invention is shown. Figure 2 A schematic diagram of the system architecture described in this invention is shown. Detailed Implementation
[0017] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0018] Combination Figure 1 Firstly, this invention provides an AI-based method for reviewing live streaming content. This method constructs a cognitive intelligent review framework to achieve a deep understanding and proactive defense of live streaming content. Specifically, it includes the following steps: S1. Multidimensional Cognitive Feature Extraction: This step acquires live stream data, which includes three components: video stream, audio stream, and text stream. The video stream uses a sampling rate of 30 frames per second to extract image sequences, the audio stream uses a sampling rate of 16kHz to acquire audio signals, and the text stream captures real-time records of bullet comments, comments, and gift-giving.
[0019] The cognitive consistency analysis model employs a multi-branch deep neural network architecture, including branches for speech feature extraction, facial expression feature extraction, action feature extraction, and scene feature extraction. The speech feature extraction branch converts the audio signal into a Mel-ray spectrogram, which is then input into a convolutional neural network to extract speech prosodic feature vectors. This feature vector can characterize the speaker's intonation changes, rhythm patterns, and emotional tendencies. The facial expression feature extraction branch locates facial regions using a face detection algorithm and extracts facial expression feature vectors using a ResNet-50 backbone network. This process captures the overall movement patterns and facial expressions of the face. The motion feature extraction branch uses the OpenPose algorithm to extract the coordinates of 25 key points on the human body, and generates motion feature vectors through a temporal coding network. This reflects the anchor's body language and posture changes. The scene feature extraction branch uses a DeepLabV3+ semantic segmentation network to analyze the background environment and extract scene semantic feature vectors. It can identify object categories, spatial layout, and scene context in a live streaming environment.
[0020] The feature vectors extracted from each branch are integrated through a feature fusion layer to construct a unified cognitive feature representation. The semicolon indicates a vector concatenation operation. Cognitive dissonance detection is based on feature correlation analysis, constructing a feature correlation matrix. Matrix elements Indicates the first The feature branch and the first Correlation coefficients between feature branches: in: The length of the time window; For the first Each feature branch at time... eigenvectors; For the first The average eigenvector of each feature branch within the time window.
[0021] Under normal behavior patterns, stable correlations exist between various feature dimensions. The system constructs a baseline correlation matrix through statistical learning on a large number of normal live streaming samples. The cognitive dissonance index is obtained by calculating the degree of deviation between the current correlation matrix and the baseline matrix: in: As an indicator of cognitive dissonance; Denotes the Frobenius norm; As the reference matrix The elements in.
[0022] The determination of cognitive dissonance uses a dynamic threshold strategy, with the threshold... Adaptively adjusts based on the type of live streaming scenario and historical data. At this point, the system determines that cognitive dissonance exists, indicating inconsistencies in the anchor's multimodal performance and potential faking or deception. This detection result serves as a trigger signal, activating subsequent micro-expression analysis and semantic trajectory monitoring modules for in-depth analysis.
[0023] S2. Micro-expression Intent Recognition: Based on the cognitive dissonance signals detected in step S1, this step performs refined analysis of the facial region for the corresponding time period. Specifically, in response to the cognitive dissonance signals detected in step S1 exceeding a preset threshold, this step triggers refined analysis of the facial region for the corresponding time period. A high-frequency sampling rate of 200Hz is used to densely sample the facial image sequence to ensure the capture of transient micro-expression changes. An optical flow tracing algorithm is used to calculate the pixel displacement field between adjacent frames. in: For position At any moment The optical flow vector; For the image at position time grayscale value; For calculation window; For regularization parameters; This is the gradient operator.
[0024] Preferably, for the time period marked in step S1, the video frames decoded from the live stream are subjected to frame interpolation or super-resolution processing to obtain a denser image sequence in the time dimension. Then, the pixel displacement field between adjacent frames is calculated by optical flow tracing algorithm to capture the micro-expression change sequence, ensuring that transient micro-expression changes can be captured.
[0025] Optical flow calculation employs the Lucas-Kanade pyramid algorithm, constructing a four-layer image pyramid with a scaling factor of 0.5 for each layer. Sub-pixel accuracy motion estimation is obtained through a coarse-to-fine iterative optimization strategy. The facial region is divided into 17 action units according to the Facial Action Coding System (FACS) standard, specifically including: AU1 corresponding to inner eyebrow elevation controlled by the medial frontalis muscle; AU2 corresponding to outer eyebrow elevation controlled by the lateral frontalis muscle; AU4 corresponding to eyebrow descent controlled by the depressor supercilii and corrugator supercilii muscles; AU5 corresponding to upper eyelid elevation controlled by the levator palpebrae superioris muscle; AU6 corresponding to cheek elevation controlled by the orbicularis oculi muscle in the orbital region; AU7 corresponding to eyelid tightening controlled by the orbicularis oculi muscle in the palpebral region; AU9 corresponding to nasolabial folds controlled by the levator nasolabial muscle; AU10... AU12 corresponds to the levator labii superioris muscle controlling upper lip lifting, AU14 corresponds to the zygomaticus major muscle controlling corner lifting, AU15 corresponds to the buccinator muscle controlling dimples, AU16 corresponds to the depressor anguli oris muscle controlling corner pulling down, AU17 corresponds to the depressor labii inferioris muscle controlling lower lip drooping, AU18 corresponds to the mentalis muscle controlling chin lifting, AU20 corresponds to the orbicularis oris muscle controlling lip stretching, AU23 corresponds to the orbicularis oris muscle controlling lip tightening, AU25 corresponds to the orbicularis oris muscle controlling lip separation, and AU26 corresponds to the depressor mandibular muscle controlling chin drooping.
[0026] The activation intensity of each action unit is obtained by calculating the average optical flow amplitude within its coverage area: in: For the first Each action unit at time The activation intensity; For the first The spatial area of an action unit; This represents the number of pixels within the region.
[0027] The temporal patterns of action units are constructed by building activation sequences. It means that among them For a moment The activation vector is used. This sequence is input into a Long Short-Term Memory (LSTM) network for temporal modeling. The LSTM network consists of two layers: the first layer contains 128 hidden units, and the second layer contains 64 hidden units. Dropout is used to prevent overfitting, and the dropout rate is set to 0.3.
[0028] The output of the LSTM is weighted and aggregated using an attention mechanism:
[0029] in: This is the global representation vector of micro-expressions; For LSTM at time The hidden state output; For a moment The attention weights are calculated as follows: in: , , These are learnable attention parameters.
[0030] Global representation vector By mapping to the intent space through a fully connected layer, the probability distributions of five implicit intents are output:
[0031] in: ; and These are the weights and bias parameters for the fully connected layer.
[0032] The results of micro-expression analysis corroborated the cognitive dissonance indicators, indicating a higher probability of detecting deceptive intent. or probability of tension And simultaneously satisfy When this occurs, the system marks the time period as a high-risk period, triggering subsequent in-depth analysis processes.
[0033] S3. Semantic Evolution Trajectory Analysis: After receiving the high-risk period marker, this step performs in-depth semantic analysis on the live broadcast content of that period and the 30 seconds before and after it. The live audio is converted into text by an automatic speech recognition system. The ASR system adopts an end-to-end Transformer architecture, with the encoder and decoder containing 12 and 6 self-attention layers, respectively. The model is pre-trained on multi-domain speech datasets and supports the recognition of various dialects and accents.
[0034] The transformed text is segmented into 10-second time windows, and the text within each window is semantically encoded using a pre-trained BERT-Base model. The BERT model contains a 12-layer Transformer encoder with 768 hidden dimensions, using the output of the last layer's [CLS] tags as the semantic representation of the sentence to generate semantic vectors. .
[0035] The semantic space is constructed based on deep learning of historical violation cases. The system maintains a knowledge base containing eight categories of sensitive topics: Sensitive Public Opinion (including content related to sensitive topics and events); Pornographic and Vulgar Content (including content with sexual innuendo, indecent actions, and vulgar language); Violent and Bloody Content (including content with violent behavior, bloody images, and terrorist threats); Gambling and Fraud Content (including gambling inducements, financial fraud, and fake prize winnings); Illegal and Irregular Content (including content related to the trading and sale of prohibited items and illegal services); False Advertising Content (including false advertising, exaggerated claims, and misleading consumer content); Infringing Content Content (including content related to the dissemination of pirated content, trademark infringement, and privacy leaks); and Content Promoting Negative Values, Spreading Negative Emotions, and Demonstrating Inappropriate Behavior.
[0036] Each sensitive topic category is analyzed by clustering corresponding violation cases to form multiple feature clusters. The clustering algorithm uses a hierarchical clustering method, and the distance metric used is cosine similarity. The first of the sensitive topics The cluster center vector is denoted as The radius of the cluster is denoted as , which represents the average distance from the sample within the cluster to the center.
[0037] The distance from the current semantic point to each sensitive topic is calculated in the following way: in: For a moment semantic vector to the first Normalized minimum distance for sensitive topics; Denotes the Euclidean norm; This indicates taking the minimum value over all clusters; For the first Category of topics The radius of each cluster.
[0038] The evolution of semantic trajectories is characterized by several dynamic metrics. Instantaneous drift velocity: in: To the first The drift speed of sensitive topics; The distance at the previous moment; The distance at the current moment; The time interval is indicated by a negative value, while a positive value indicates a distance from sensitive topics.
[0039] Drift acceleration:
[0040] in: This is drift acceleration; The value represents the speed at the previous moment; a negative value indicates an accelerated approach to a sensitive topic.
[0041] Directional persistence is assessed by evaluating the consistency of drift direction within a statistically continuous time window: in: This is a directional persistence indicator; the closer the value is to 1, the more likely it is to drift continuously in the same direction. This represents the number of sampling points within the observation window; This is a sign function; it returns 1 when the input is positive, -1 when the input is negative, and 0 when the input is zero. For the first Each sampling time.
[0042] The overall drift risk score is calculated using a weighted combination: in: For comprehensive drift risk scoring; To correct the linear unit function, the output is the original value when the input is positive and zero when the input is negative; , , For the weighting coefficients, satisfying ; This indicates that the maximum value is taken for all sensitive topic categories.
[0043] Semantic trajectory prediction employs an autoregressive model, using historical trajectories to predict future semantic evolution: in: This is the predicted semantic vector for the next time step; The order of autoregression; For the first The autoregressive coefficient matrix of order 1; for The semantic vector at any given time; This is the prediction error vector.
[0044] when And the predicted trajectory shows that in the future When entering a sensitive area within a specified time, the system generates a semantic drift warning, in which... As a risk threshold, For the prediction time window.
[0045] S4. Emotion Propagation Link Modeling: Combining semantic drift warning and micro-expression intent recognition results, this step constructs an emotion propagation network between the broadcaster and the audience to analyze the diffusion mechanism of emotions in the group.
[0046] The broadcaster's emotional state is obtained through multimodal feature fusion. Voice emotion feature extraction employs a deep spectrogram convolutional network. The network input is an 80-dimensional Mel spectrogram, which is processed through six convolutional layers to extract acoustic features. The number of filters in the convolutional layers are 64, 128, 256, 256, 512, and 512 respectively, with a 3×3 kernel size and a stride of 1. Batch normalization and ReLU activation functions are used. The output of the convolutional layers is then subjected to global average pooling and mapped to a 256-dimensional emotion vector through a three-layer fully connected network. .
[0047] Facial expression sentiment analysis reuses the action unit activation sequences extracted in step S2, transforming them into a Valence-Arousal two-dimensional continuous sentiment space via a sentiment mapping network. The mapping relationships are learned through a neural network. in: Valence is the dimension representing the positive or negative nature of emotion; To awaken the dimension, indicating the intensity of emotion; Activation vector for action units; , This is the weight matrix; , It is the bias vector; It is the hyperbolic tangent activation function; It is the sigmoid activation function.
[0048] The text sentiment analysis uses a fine-tuned RoBERTa model, fine-tuned on six basic sentiment datasets, to output sentiment probability vectors: in For the probability of joy, For the probability of anger, For the probability of sadness, For the probability of fear, For the probability of surprise, This represents the probability of aversion.
[0049] The emotional features of the three modalities are integrated through an adaptive fusion mechanism: in: The anchor's overall emotional vector; Indicates the modal type, For voice, For the face, For text; For modality The transformation matrix maps emotional features of different dimensions to a unified space; For modality The emotional vector; Modal weights are dynamically calculated using a gating mechanism: ;
[0050] in: For modality The gate value; This is the gating weight vector; This is the gated weight matrix; This is the gated bias vector; It is an exponential function; To modify the activation function of the linear unit.
[0051] Audience group modeled as a dynamic graph ,in for The collection of audience nodes at any given moment This is the set of edges. Nodes eigenvectors It consists of three parts: emotional characteristics of bullet comments The text in the bullet comments is extracted using a BiLSTM-CRF model, which performs bidirectional encoding on the bullet comment text. in: For the first Embedding vectors of each word; This represents the hidden state of the feedforward LSTM network; This represents the hidden state of the backward LSTM network; This is a concatenation of two hidden states; and These are forward and backward long short-term memory network units, respectively.
[0052] The CRF layer performs global optimization on the sentiment tag sequence: in: Given an input sequence Time tag sequence The conditional probability; Normalization factor; It is the characteristic function; and Labels for adjacent positions; The sequence length; For location index.
[0053] Gift-giving behavior characteristics Encode information on gift type, value, and frequency: in: The one-hot encoded vector for the gift type; For the first The type of gift for each audience member; The value of the gift; Frequency of gifting; It is the natural logarithm function.
[0054] Activity characteristics Reflecting audience participation:
[0055] in: Number of times spoken; For the number of likes; This refers to the viewing duration.
[0056] marginal weight among audiences Determined based on interaction relationships: in: For nodes To the node Edge weights; For nodes Reply Node The number of times; For nodes Mentioning nodes The number of times; The cosine similarity function; , , For the weight parameters, satisfying .
[0057] The propagation of emotions among the audience is modeled using graph neural networks. Node representation learning is performed using the GraphSAGE architecture, which updates node representations through neighbor sampling and aggregation mechanisms. in: For nodes In the The layer's representation vector; For nodes In the The layer's representation vector; For nodes The set of neighbors; The aggregation result represents the neighboring nodes; This is a vector concatenation operation; For the first Layer weight matrix; This is the activation function.
[0058] Neighbor aggregation uses mean pooling:
[0059] in: The number of neighboring nodes; For neighboring nodes; For neighboring nodes In the Layer representation.
[0060] The emotional influence of broadcasters on viewers is achieved through a communication matrix. Modeling, matrix elements Indicates from node To the node The intensity of emotional transmission:
[0061] in: For the intensity of propagation; and They are nodes and The representation vector; This is the attention weight matrix; The dimension representing the vector; The set of all nodes; It is an exponential function; This is a scaling factor to prevent numerical instability.
[0062] The intensity of emotion propagation is assessed by calculating the spectral radius of the propagation matrix:
[0063] in: The intensity of emotional transmission; For matrix spectral radius; For matrix The One eigenvalue; Represents the modulus of a complex number.
[0064] when At that time, the system determined that there was an abnormal pattern of emotion transmission, in which This is the preset propagation intensity threshold.
[0065] S5. Reverse Causal Reasoning: After detecting abnormal emotional transmission patterns, this step uses the analysis of abnormal behavioral reactions of the audience to infer the potential violations of the anchor.
[0066] The audience abnormal behavior indicators include three dimensions: the bullet screen sentiment change index. Abnormal Gift Giving Index and Exit Rate Surge Index .
[0067] The method for calculating the bullet screen sentiment mutation index is as follows:
[0068] in: The index of sudden changes in the sentiment of bullet comments; For a moment The average sentiment of the bullet comments; For a moment The average sentiment of the bullet comments; For a moment The standard deviation of the sentiment of the bullet comments; Use small positive numbers to prevent division by zero errors.
[0069] Gift-giving anomaly index:
[0070] in: Abnormal index for gift-giving; For a moment The total value of the gifts given; The historical average value of gifts; The standard deviation of the value of historical gifts.
[0071] Exit Rate Surge Index:
[0072] in: The index for the surge in exit rates; For a moment Audience dropout rate; This represents the historical average exit rate.
[0073] A Bayesian inference network is constructed for causal analysis. The network consists of three layers: an observation layer, a latent variable layer, and a cause layer. Observation layer nodes... Corresponding to the three abnormal indicators, the hidden variable layer Representing potential intermediate states, cause layer Indicates the possible types of violations.
[0074] The posterior probability of the violation is calculated using Bayes' theorem:
[0075] in: For a given observation Violations The posterior probability; For a given violation Abnormal reactions were observed at that time. The likelihood probability; For violations The prior probability; This represents the total number of types of violations.
[0076] Likelihood probability Calculations are made by marginalizing latent variables: in: Given latent variables Observed at time The probability of; For a given violation Latent variables The probability of.
[0077] The inference process employs a joint tree algorithm for precise reasoning, propagating probability information across the factor graph via a message-passing mechanism. This involves determining the posterior probability of a given violation. At that time, the system generates a corresponding violation warning.
[0078] S6. Cognitive Load Assessment: After detecting potential violations, this step analyzes the cognitive processing burden of the live stream content on the audience and identifies behavioral patterns that use information overload to cover up violations.
[0079] Cognitive load is quantified using three dimensions: information density. Switching frequency and multimodal conflict degree .
[0080] Information density is calculated using information entropy per unit time: in: Information density; The length of the time window; The number of types of information elements; For a moment No. The probability of a certain information element appearing; It is the natural logarithm function.
[0081] Information element recognition is achieved through multimodal content analysis. The visual modality uses an object detection network to identify object categories, the audio modality uses sound event detection to identify sound types, and the text modality uses entity recognition to extract key information.
[0082] Switching frequency reflects the rate of change of content themes: in: For frequency switching; For a moment Theme categories; This is an indicator function that returns 1 if the condition is true, and 0 otherwise.
[0083] Topic recognition employs the Hidden Dirichlet Allocation (LDA) model, mapping content to 20 predefined topic spaces.
[0084] Multimodal conflict measures the degree of inconsistency between different modal information: in: Multimodal conflict degree; The number of modes; and The first and the Feature vectors of each modality; This is the similarity function.
[0085] The comprehensive cognitive load index is calculated through a weighted combination: in: The comprehensive cognitive load index; , , For the weighting coefficients, satisfying .
[0086] Sensitive information fragments are detected by combining keyword density and semantic similarity within a sliding window.
[0087] in: For a moment Sensitivity score; Sensitive keyword density; This is the current semantic vector; For the first A cluster of sensitive topics; For balancing parameters.
[0088] when and When both conditions are met, the system determines that there is an act of using cognitive overload to cover up illegal content. The cognitive load threshold, This is the sensitivity threshold.
[0089] S7. Blocking the Spread of Illegal Content: Based on the analysis results of the preceding steps, this step constructs the spread network of illegal content and implements precise blocking. The spread network is modeled as a directed graph. ,in For a set of nodes, It is a set of directed edges.
[0090] Node set It includes three types of nodes: content source nodes. This refers to livestreamers and content producers, and active dissemination nodes. Viewers who actively participate in the interaction, potentially affected nodes This refers to viewers who passively receive information. Node classification is determined using the following indicators: in: For nodes The degree of departure; For nodes in-degree; , , This is the classification threshold.
[0091] The edge weights represent the propagation probability, calculated using historical interaction data and current activity levels: in: For the node To the node The probability of propagation; The intensity of historical interaction is obtained by counting the number of interactions within a historical time window. For nodes At any moment Activity level; For nodes and The strength of the relationship between them; It is the sigmoid function; , , These are the weight parameters.
[0092] Key nodes are evaluated using a combination of three centrality metrics. Betweenness centrality: in: For nodes Betweenness centrality; For the node To the node The total number of shortest paths; For the nodes from arrive Find the shortest path count; sum the results by traversing all distinct source nodes. and target node right.
[0093] Degree centrality:
[0094] in: For nodes Degree centrality; For nodes The degree; This represents the total number of nodes.
[0095] The eigenvector centrality is obtained by solving the characteristic equation:
[0096] in: It is an adjacency matrix; Let be an eigenvector, and let its th eigenvector be... Each component That is, a node eigenvector centrality; It is the largest eigenvalue.
[0097] Overall centrality score:
[0098] in: For nodes The overall centrality score; For eigenvector centrality; , , For the weighting coefficients, satisfying .
[0099] The extent of the spread was estimated using an independent cascade model. Let the initial set of infected nodes be... The propagation process is iterative:
[0100] in: For a moment The set of infected nodes; for Uniform random numbers in an interval; when At that time, node The node Infect.
[0101] Run through Monte Carlo simulation Estimated average impact range for each propagation process: in: For the initial set The average range of influence; For the first The final infected set in this simulation; The cardinality of a set.
[0102] The blocking set is selected using a greedy algorithm, with the goal of minimizing the propagation range after blocking. in: The optimal blocking set; This represents the maximum number of nodes that can be blocked. This indicates that the blocking node is removed from the initial infected set.
[0103] Greedy algorithm iteratively selects the node with the highest marginal benefit: in: The optimal node selected in the current iteration; the marginal benefit is the number of nodes added. The amount of reduction in the spread range afterward.
[0104] Blocking measures are implemented in stages based on node type and severity of violation: in: The comprehensive violation risk score is obtained by weighting the outputs of the aforementioned modules.
[0105] S8. Comprehensive Decision Output: Integrate the analysis results of the above steps and generate the final review conclusion through a weighted decision network.
[0106] Constructing a comprehensive feature vector It includes the key outputs of each module: in: As an indicator of cognitive dissonance; This is a probability vector of micro-expression intentions; Score the risk of semantic drift; The intensity of emotional transmission; This is the violation probability vector obtained through reverse reasoning; Cognitive load index; This represents the maximum node centrality.
[0107] The decision network adopts a three-layer fully connected architecture. Layer 1:
[0108] in: Output for the first hidden layer; This is the first layer weight matrix. For input dimensions; This is the first layer bias vector; To modify the activation function of the linear unit.
[0109] The second layer introduces batch normalization: in: The output is a linear transformation. and These are the batch mean and variance, respectively. It is the stability constant; and These are learnable scaling and offset parameters; This is the output of the second hidden layer.
[0110] The probability distribution of violation types generated in the third layer:
[0111] in: The probability distribution for 12 types of violations; It is a normalized exponential function.
[0112] Violations include: violations related to public opinion, pornography and vulgarity, violence and terrorism, illegal and criminal activities, negative publicity, false advertising, copyright infringement, malicious marketing, misleading minors, spreading rumors, personal attacks, and privacy breaches, with corresponding probabilities of [not specified]. arrive .
[0113] The violation level is calculated by weighted summation:
[0114] in: The overall level of violation; For the first The severity weight of each type of violation, with a value range of [value range missing]. .
[0115] Violation level classification:
[0116] Action recommendations are generated based on the level of violation: Interpretability analysis employs an ensemble gradient method to calculate the contribution of each input feature to the final decision: in: For the first The integrated gradient of each feature; For the first One input feature; Baseline value; For decision network functions; These are the interpolation parameters.
[0117] In actual calculations, the Riemann summation approximation is used: in: This represents the number of interpolation steps.
[0118] The system sorts the data by contribution, selects the top 5 most important features as key evidence, and generates a structured review report that includes the time period of the violation, the type of content involved in the violation, a list of key evidence, and a visualization of the decision-making reasoning path.
[0119] On the other hand, the present invention provides an AI-based live streaming content review system for implementing the review method described in Embodiment 1 above; the system includes a cognitive feature analysis module, a micro-expression recognition module, a semantic trajectory monitoring module, an emotion propagation analysis module, a reverse reasoning module, a cognitive load calculation module, a propagation blocking module, and a decision output module.
[0120] The cognitive feature analysis module extracts multi-dimensional cognitive features from the live stream. It analyzes the consistency relationships between speech, facial expressions, actions, and scenes using deep learning networks to detect cognitive dissonance and identify spoofed and deceptive content. This module employs a multi-branch parallel processing architecture. The speech branch uses a convolutional neural network to process Mel spectrograms and extract prosodic features; the facial expression branch extracts facial expression features based on a ResNet-50 backbone network; the action branch extracts human pose features using the OpenPose algorithm; and the scene branch uses a DeepLabV3+ semantic segmentation network to extract environmental context features. The feature vectors extracted from each branch are integrated through a feature fusion layer to construct a unified cognitive feature representation. Cognitive dissonance detection is achieved by calculating the Frobenius norm difference between the current feature correlation matrix and the normal behavior baseline matrix. When abnormal dissonance is detected, a cognitive dissonance index is generated. .
[0121] The micro-expression recognition module is used to capture and analyze the micro-expression changes of the anchor. It extracts subtle expression features through a facial motion coding system and infers the anchor's true intentions and emotional state by combining them with a deep learning model. This module adopts a 200Hz high-frequency sampling strategy and calculates the pixel displacement field using the pyramid Lucas-Kanade optical flow algorithm. The facial region is divided into 17 motion units according to the FACS standard, including AU1 medial frontalis muscle, AU2 lateral frontalis muscle, AU4 depressor supercilii muscle, AU5 levator palpebrae superioris muscle, AU6 orbicularis oculi muscle, AU7 eyelid, AU9 levator nasolabial fold, AU10 levator labii superioris muscle, AU12 zygomaticus major muscle, AU14 buccinator muscle, AU15 depressor anguli oris muscle, AU16 depressor labii inferioris muscle, AU17 mentalis muscle, AU20 orbicularis oris muscle, AU23 orbicularis oris muscle contraction, AU25 orbicularis oris muscle dissociation, and AU26 depressor mandibular muscle. The activation intensity is obtained by statistically analyzing the optical flow amplitude within each motion unit region. The temporal activation patterns of action units are analyzed using a two-layer LSTM network, outputting the probability distributions of five intention categories: deception, tension, anger, fear, and disgust. .
[0122] The semantic trajectory monitoring module tracks the semantic evolution path of live content in real time, detecting the drift trend of semantics towards sensitive areas through word vector space analysis technology, thus enabling early warning of illegal content. This module maintains a knowledge base containing eight categories of sensitive topics: public opinion sensitivity, pornography and vulgarity, violence and gore, gambling and fraud, illegal and irregular activities, false advertising, infringing content, and negative guidance. Each topic category is hierarchically clustered into multiple feature clusters. After the live audio is converted to text by the ASR system, it is encoded into 768-dimensional semantic vectors using the BERT-Base model. By calculating the normalized distance from the semantic vectors to each sensitive topic cluster, the drift speed of the semantic trajectory is monitored. acceleration and directional continuity Comprehensive assessment of drift risk .
[0123] The emotion propagation analysis module is used to model the emotion contagion mechanism between the broadcaster and the audience. It analyzes the diffusion pattern of emotions within the group using a graph neural network to identify inflammatory and manipulative violations. This module obtains the broadcaster's emotional state through multimodal fusion, integrating three dimensions: vocal emotion, facial expression, and text emotion, and calculates modal weights using an adaptive gating mechanism. The audience group is modeled as a dynamic graph structure, with node features including three parts: bullet screen emotion, gift behavior, and activity level. Bullet screen emotion is extracted using a BiLSTM-CRF model, and gift behavior and activity information are encoded. A GraphSAGE architecture is used to learn node representations, aggregating neighbor information through mean pooling. The intensity of emotion propagation is calculated by determining the spectral radius of the propagation matrix. An assessment will be conducted.
[0124] The reverse reasoning module is used to infer the streamer's violations from abnormal audience reactions. It constructs a causal reasoning chain using a probabilistic graphical model to discover hidden violations. This module monitors the bullet screen sentiment abrupt change index. Abnormal Gift Giving Index and Exit Rate Surge Index Three dimensions of anomaly indicators. A three-layer Bayesian network is constructed, including an observation layer, a latent variable layer, and a causal layer. The posterior probability of the violation is calculated using Bayes' theorem. The reasoning process employs a joint tree algorithm for precise inference, propagating probability information across the factor graph through a message passing mechanism.
[0125] The cognitive load calculation module is used to assess the cognitive processing burden of live content on viewers and identify behavioral patterns that use information overload to cover up illegal content. This module calculates information density... Switching frequency and multimodal conflict degree Cognitive load is quantified using three dimensions. Information density is assessed by calculating information entropy per unit time; switching frequency is detected by using LDA topic modeling to detect content changes; and multimodal conflict is obtained by calculating the semantic similarity between different modalities. These three dimensions are weighted and combined to obtain a comprehensive cognitive load index. .
[0126] The propagation blocking module analyzes the propagation network structure of illegal content, identifies and blocks key propagation nodes through causal intervention strategies, and prevents the spread of the illegal impact. This module models the propagation network as a directed graph, with nodes categorized into three types: content source nodes, active propagation nodes, and potentially affected nodes. Betweenness centrality is calculated. Degree centrality and eigenvector centrality The importance of nodes is comprehensively assessed. An independent cascade model is used to simulate the propagation process, and the impact range is estimated using the Monte Carlo method. A greedy algorithm is used to select the optimal blocking set and implement tiered blocking measures.
[0127] The decision output module integrates the analysis results from various modules, generates a final review conclusion through multi-source information fusion and weighted decision-making, and provides interpretable explanations. This module constructs a comprehensive feature vector. The algorithm includes indicators of cognitive dissonance, micro-expression intent probability, semantic drift risk, emotion propagation intensity, backward reasoning probability, cognitive load index, and maximum node centrality. It employs a three-layer fully connected neural network with 256, 128, and 64 neurons, using ReLU activation and batch normalization. It outputs the probability distribution of 12 violation categories and calculates the overall violation level. This generates four levels of treatment recommendations. The interpretability analysis employs an ensemble gradient method to calculate feature contribution. The first five key features were selected as evidence.
[0128] Data exchange and collaborative work between modules are achieved through standardized interfaces. The cognitive feature analysis module processes the live stream at a frequency of 30fps. The micro-expression recognition module is triggered to perform high-frequency analysis. The semantic trajectory monitoring module updates the semantic trajectory every 10 seconds. A warning is generated in real time. The emotion propagation analysis module updates the audience emotion map at a frequency of 1Hz. An anomaly alarm is triggered at certain times. The reverse inference module uses an event-driven mechanism, initiating inference when an abnormal indicator exceeds a corresponding threshold. The cognitive load calculation module evaluates every 10 seconds, and when... and To enhance the sensitivity of the audit process, the propagation blocking module maintains a real-time network snapshot and makes a blocking decision within 100ms when a high-risk violation is detected. The decision output module generates a final judgment within 50ms after receiving input from each module.
[0129] The system employs a streaming processing architecture to ensure real-time performance, and improves processing efficiency through parallel computing and asynchronous processing strategies. Processing latency on the critical path is controlled through algorithm optimization and hardware acceleration to ensure that the total latency from the appearance of infringing content to the output of the review result meets real-time review requirements. The system also includes an adaptive learning mechanism that continuously updates its ability to identify new violation patterns through incremental learning, maintaining the timeliness of its review capabilities.
[0130] This invention constructs a cognitive intelligent review framework, enabling a deep understanding and proactive defense of live streaming content. Compared to traditional review methods, this system can identify complex violation patterns such as cognitive deception, psychological manipulation, and hidden violations, and effectively control the scope of violation impact through a propagation blocking mechanism, providing reliable technical protection for the content ecosystem security of live streaming platforms.
[0131] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An AI-based method for reviewing live streaming content, characterized in that, Includes the following steps: S1. Multidimensional cognitive feature extraction: Acquire live stream data and simultaneously extract the anchor's voice features, facial expression features, body movement features, and scene environment features through a cognitive consistency analysis model to construct a multidimensional cognitive feature vector; calculate the deviation between the current feature correlation matrix and the normal behavior benchmark matrix; when an abnormal imbalance is detected between feature dimensions, generate a cognitive imbalance index. S2. Micro-expression intention recognition: High-frequency sampling analysis is performed on the facial expression features, and the pixel displacement field between adjacent frames is calculated through optical flow tracing algorithm to capture the micro-expression change sequence; the activation intensity temporal features of facial muscle movement are extracted, and the mapping relationship between the activation intensity temporal features and the implicit intention is analyzed through a long short-term memory network to generate an intention category probability distribution; S3. Semantic Evolution Trajectory Analysis: Perform real-time semantic encoding on live content and calculate the distance between the current semantic vector and the cluster center of sensitive topics; By monitoring the drift speed, acceleration, and directional persistence of semantic trajectories, a semantic drift warning signal is generated when a tendency for the semantic trajectory to drift toward a sensitive area is detected. S4. Emotional propagation link modeling: Collect anchor's emotional state indicators and real-time audience feedback data, and model the audience group as a dynamic graph structure; analyze the propagation path of anchor's emotions in the audience group through graph neural networks, calculate the spectral radius of the propagation matrix as the emotional propagation intensity, and identify inflammatory and manipulative emotional propagation patterns; S5. Reverse Causal Reasoning: Monitor the audience's sudden changes in barrage emotions, abnormal gift giving, and surge in exit rate indicators. Construct a Bayesian network containing an observation layer, a latent variable layer, and a causal layer. Infer the potential violations of the streamer by calculating the posterior probability of the violation. S6. Cognitive Load Assessment: The cognitive load index of the audience is calculated from three dimensions: information density, switching frequency, and multimodal conflict degree. When the cognitive load index exceeds the normal range and sensitive information fragments are detected, it is determined that there is a behavior of using information overload to cover up illegal content. S7. Blocking the spread of illegal content: Model the spread of illegal content as a directed graph network, identify key propagation nodes by calculating the centrality index of nodes, and use a greedy algorithm to select the optimal blocking set to implement hierarchical blocking; S8. Comprehensive Decision Output: Integrate the analysis results of the above steps to construct a comprehensive feature vector, output the probability distribution of violation types through a multi-layer neural network, calculate the comprehensive violation level and generate disposal suggestions, and use the integrated gradient method to calculate the feature contribution and generate an interpretable report.
2. The method according to claim 1, characterized in that, The cognitive dissonance index in step S1 is obtained by calculating the Frobenius norm difference between the current feature correlation matrix and the benchmark matrix, which is pre-constructed through statistical learning of normal live streaming samples.
3. The method according to claim 1, characterized in that, In step S2, facial muscle group regions are located based on facial key points, and the optical flow amplitude within each muscle group region is calculated as the activation intensity. A temporal sequence of activation intensity is then constructed and input into a long short-term memory network.
4. The method according to claim 1, characterized in that, The sensitive topic clustering in step S3 is constructed by hierarchically clustering historical violation cases, and the semantic encoding uses a pre-trained language model to convert the text into a high-dimensional vector representation.
5. The method according to claim 1, characterized in that, The anchor's emotional state in step S4 is obtained by fusing voice emotion, facial expression emotion, and text emotion, and modal weights are dynamically calculated using an adaptive gating mechanism; the audience node features include bullet screen emotion, gift behavior, and activity information.
6. The method according to claim 1, characterized in that, The centrality index in step S7 includes a weighted combination of betweenness centrality, degree centrality, and eigenvector centrality; an independent cascade model is used to simulate the propagation process of illegal content, and the scope of propagation impact is estimated using the Monte Carlo method.
7. The method according to any one of claims 1 to 6, characterized in that, Each step works in concert through a threshold triggering mechanism: when the cognitive dissonance index exceeds a preset threshold, high-frequency analysis of micro-expression recognition is triggered; when the semantic drift risk exceeds a threshold, an early warning is generated; when the intensity of emotion propagation exceeds a threshold, reverse reasoning is activated; and the outputs of each step converge into the comprehensive decision-making step to generate the final judgment.
8. An AI-based live streaming content review system, characterized in that, include: The cognitive feature analysis module is used to extract multi-dimensional features of voice, facial expressions, actions, and scenes from the live stream, and detect cognitive dissonance by calculating the degree of deviation between the feature correlation matrix and the baseline matrix. The micro-expression recognition module is used to calculate the pixel displacement field of the facial region through optical flow tracing algorithm, extract the temporal features of activation intensity of facial muscle movement, and use a long short-term memory network to learn the mapping relationship between activation intensity features and implicit intentions. The semantic trajectory monitoring module is used to encode live semantic content in real time, calculate the distance from semantic vectors to sensitive topic clusters, and generate early warnings by analyzing drift speed, acceleration, and directional persistence. The emotion propagation analysis module is used to construct a dynamic graph structure of the audience, analyze the emotion propagation path through graph neural networks, and calculate the propagation matrix spectral radius to assess the propagation intensity. The reverse reasoning module is used to monitor abnormal audience behavior indicators and calculate the posterior probability of violations through a Bayesian network. The cognitive load calculation module is used to assess cognitive load based on information density, switching frequency, and multimodal conflict degree. The propagation blocking module is used to calculate the centrality index of nodes in the propagation network and select the optimal blocking set using a greedy algorithm. The decision output module is used to integrate the outputs of various modules and generate violation judgments and interpretable reports through neural networks.
9. The system according to claim 8, characterized in that, The system adopts a streaming processing architecture, with each module working collaboratively through an event-driven mechanism; it includes an adaptive learning mechanism that incrementally learns and updates the ability to identify new violation patterns.
10. The system according to claim 8 or 9, characterized in that, The decision output module uses a three-layer fully connected neural network to calculate the contribution of each feature to the violation determination through an integrated gradient method, and selects the feature with the highest contribution as key evidence to generate an interpretability report.