Text and image fused online comment toxicity detection and filtering method and system

Abstract

The invention discloses a text and image fused online comment toxicity detection and filtering method and system, and relates to the technical field of natural language processing, and the method comprises the steps: extracting a semantic vector of a comment text through a RoBERTa-Marge model; the visual features of the image are extracted through an OfficientNet-V2 model; aligning heterogeneous modal features by adopting a double-flow contrast loss function; self-adaptive decision making of culture sensitivity: loading a regional sensitive rule table according to a user IP address, and dynamically adjusting symbolic semantics; calculating an intimacy correction factor based on the social relationship between the publisher and the receiver; and performing context weighted toxicity scoring, calculating a user historical behavior weight, and outputting a final toxicity probability. According to the method, a deep dynamic mapping mechanism of text and visual features is constructed, heterogeneous features are extracted through RoBERTa-Large and OfficientNet-V2 double-flow architectures, semantic space alignment is forced by utilizing comparative learning, and the problem of image-text splitting detection in the traditional technology is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to an online review toxicity detection and filtering method and system fusing text and images. BACKGROUND

[0002] The popularity of the Internet has made online reviews an important way for people to communicate and express their opinions, but it has also been accompanied by a large amount of toxic content, such as insults, discrimination, and malicious attacks, which not only affect user experience, but also can cause harm to individuals and society, so effective detection and filtering means are needed.

[0003] Toxic information in online reviews is often not only reflected in text, but images can also convey harmful or undesirable intentions. For example, some memes, pictures, and other content containing offensive or discriminatory content. In order to more comprehensively and accurately detect toxicity, text analysis and image understanding techniques need to be combined to integrate multi-modal information for comprehensive judgment.

[0004] Advances in the fields of natural language processing and computer vision have provided powerful tools for text and image analysis. For example, pre-trained language models have made significant progress in text semantic understanding, while convolutional neural networks and others have performed well in image feature extraction and classification, making it possible to integrate information from both for toxicity detection.

[0005] However, there are still problems to be solved in existing technology: multi-modal semantic fragmentation, heterogeneous representation spaces for text and images, information loss due to simple concatenation and average pooling; context dependence and ambiguity, different interpretations of the same content in different cultures and groups; lack of context such as conversation history and user profile, making it difficult to distinguish between jokes and attacks. SUMMARY

[0006] To solve the above technical problems, an online review toxicity detection and filtering method and system fusing text and images are provided, which solve the problems of multi-modal semantic fragmentation and context dependence and ambiguity.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is:

[0008] An online review toxicity detection and filtering method fusing text and images, comprising:

[0009] S1, extracting the semantic vector of the review text through the RoBERTa-Large model; extracting the visual features of the image through the EfficientNet-V2 model; aligning the heterogeneous modal features using a dual-stream contrastive loss function;

[0010] S2, cultural sensitivity adaptive decision, load regional sensitive rule table according to user IP address, dynamically adjust symbol semantics; based on the social relationship between the publisher and the receiver, calculate the intimacy correction factor;

[0011] S3, context weighted toxicity score, calculate user historical behavior weight, output final toxicity probability.

[0012] Preferably, the S1 specifically includes:

[0013] Text feature extraction, using Transformer encoder, hidden layer dimension, maximum input sequence length 512 tokens;

[0014] For super-long comments, implement head and tail double-end truncation strategy to retain key semantic information; dialect normalization preprocessing, built-in dialect dictionary to realize real-time translation;

[0015] Image feature extraction, based on EfficientNet-V2-L backbone network, uniform scaling resolution for input image, output global feature vector;

[0016] Anti-interference enhancement mechanism, against disturbance detection, inject Gaussian noise into input image, calculate the Euclidean distance of features before and after disturbance, if the Euclidean distance is greater than the given threshold, it is determined as an adversarial sample and triggers an alarm;

[0017] Sensitive area enhancement, generate class activation heat map through Grad-CAM, and apply multiple feature weights to the detected text slogan area.

[0018] Preferably, the S1 specifically includes:

[0019] Feature projection layer, text projection path, map text features to a unified space through a learnable matrix, and calculate the activation function; image projection path, align visual features through a learnable matrix;

[0020] Contrastive loss function, using improved InfoNCE loss function;

[0021] Dynamic hard example mining, compare the current text with all images in the batch, calculate the similarity matrix, shield positive sample pairs, and select the negative sample index with the maximum similarity in each row; preferentially select the interference image most similar to the current text as the negative sample, so that the model learns the difference between malicious semantic combinations;

[0022] Multi-modal parallel processing pipeline, CPU asynchronously loads data, GPU parallelly executes feature extraction, synchronously flows and calculates projection hardware accelerated similarity calculation.

[0023] Preferably, the S2 specifically includes:

[0024] Hierarchical symbol semantic parsing, symbol feature encoding, input symbol, mapped to 128-bit hash code, preloaded three-dimensional parsing tree retrieved through distributed cache, level one symbol type, level two geographic and cultural region, level three real-time risk index;

[0025] IP parsing pipeline, put user IP into GeoIP2 database, make cultural region judgment, load halal taboo library in Islamic region, load network meme semantic library in East Asia region, load hate symbol database in Europe and America region;

[0026] Dynamic weight calculation:

[0027] w symbol =α·base_weight+β·Δ risk

[0028] In the formula, w symbol is the weight of the symbol; α is the basic rule weight, α = 0.7; base_weight is; β is the real-time public opinion factor, β = 0.3; Δ risk is the negative report increment of the symbol in the target region in the past 1 hour;

[0029] Cultural conflict resolution, ambiguous symbol double verification, when the weight difference of the symbol in the sending and receiving cultural circles is greater than the specified threshold, trigger the sender intent analysis, check the historical use record; The proportion of animal symbols involved in the receiving party's report record is scanned for receiving party sensitivity; The maximum weight takes the weighted average of the cultural threshold of both parties.

[0030] Preferably, the S2 specifically comprises:

[0031] Intimate correction factor calculation model, social relationship quantitative index, including private message interaction intensity, community correlation degree, historical interaction emotion;

[0032] Intimate correction factor formula:

[0033]

[0034] In the formula, γ is the intimate correction factor, used to adjust the influence of the message, the value range is between 0.2 and 1.0, when the receiving party is in the special attention list of the sending party, γ is additionally multiplied by a factor of 1.2; f msg is the message sending frequency factor, which measures the relationship between the frequency of sending messages and the intimate correction factor; g group is the community activity or similarity, the value range is between 0 and 1, the larger the value, the higher the activity or similarity of the community; s sentiment is the sentiment tendency value, indicating the sentiment tendency degree of the message, the value range is between -1 and 1;

[0035] Real-time relationship graph calculation, event-triggered local graph refresh, periodic decay operation; graph database acceleration, using Neo4j+SparkGraphX dual engine;

[0036] Rule table hot loading, dynamic protocol update, when regional cultural policy changes, configuration center pushes Zookeeper node change event; service node receives incremental rules through gRPC streaming;

[0037] De-identification processing, social relationship calculation uses differential privacy technology to add Laplace noise; special processing for EU GDPR, disable close factor calculation for European users;

[0038] Regional rule cache, deploy lightweight rule engine in each CDN edge node, American node, preload US and European, Latin American rule library; Dubai node, preload Islamic culture rule library; Singapore node, preload East Asia, Southeast Asia rule library.

[0039] Preferably, the S3 specifically comprises:

[0040] User historical behavior weight dynamic calculation, three-dimensional behavior feature extraction, toxic comment time series distribution, define time decay function, calculate user recent behavior:

[0041]

[0042] In the formula, ρ toxic is the weighted toxic comment ratio, indicating the proportion of recent toxic comments by the user, considering the time decay factor; N is the number of recent toxic comments; M is the number of all recent comments; k is an index from 1 to N, indicating the serial number of toxic comments; j is an index from 1 to M, indicating the serial number of all comments; I is an indicator function, which is 1 when the toxicity probability is greater than 0.7, otherwise 0;

[0043] Benign behavior gain factor, introduce positive behavior offset mechanism; social trust evaluation, based on graph neural network to calculate the trustworthiness of users in the social graph;

[0044] Multi-source signal fusion toxic probability calculation, basic signal input, signal types include text toxicity probability, image risk probability, context conflict score; dynamic weighted fusion, add cultural correction term;

[0045] Three-level processing pipeline, threshold judgment is performed on the original probability, if the original probability is less than the predetermined threshold, it is directly released; if the original probability is in the predetermined threshold interval, it is added to the artificial review queue; if the original probability is greater than the predetermined threshold, it is automatically intercepted and punished, and the final review result is fed back to the model.

[0046] Further, the online review toxicity detection and filtering system fusing text and image includes:

[0047] Heterogeneous feature extraction module: text encoder generates review text semantic vector, image encoder generates image visual feature;

[0048] Dynamic feature projection alignment module: including text projection matrix and image projection matrix, calculating double-flow contrast loss function; difficult example mining unit, dynamically selecting the most similar negative sample in batch based on cosine similarity;

[0049] Culture-sensitive decision engine: IP geolocation unit, accessing MaxMindGeoIP2 database, with symbolic semantic three-level parser; setting cross-cultural conflict arbitration protocol;

[0050] User context modeling module: based on behavior decay factor, calculating weighted toxicity history; calculating intimacy correction factor; homomorphic encryption probability fusion under federated learning framework.

[0051] Optionally, the dynamic feature projection alignment module specifically includes:

[0052] Text and image projection matrix: nonlinear mapping, text projection path, input text vector, after linear transformation by learnable matrix, output projection vector through GELU activation function, retaining negative value information; image projection path, input EfficientNet-V2 image feature, mapped to the same dimension through independent projection matrix, and orthogonal constraint is applied, forcing the image projection direction to be decoupled from the text space; dynamically identifying the output dimension of the upstream encoder, and automatically adjusting the output channel of the projection matrix to a unified high-dimensional space;

[0053] Double-flow contrast loss function: hybrid similarity calculation, defining an improved similarity measure function, dynamic temperature coefficient, automatically adjusting according to the discrete degree of similarity distribution within the batch; gradient equalization back propagation, monitoring the ratio of text projection gradient norm and image gradient norm;

[0054] Difficult example mining unit: dynamic negative sample screening algorithm is adopted, similarity matrix construction, calculating the cosine similarity matrix of all text-image pairs within the batch, diagonal line shielding, Top-K difficult negative sample extraction, adversarial enhancement, adding directional disturbance to the screened negative samples; semantic confusion enhancement, text level, randomly replacing keywords to generate semantic conflict negative samples; image level, using StyleGAN2 to locally tamper with negative sample images; mixed enhancement, exchanging the text and image components of positive sample pairs.

[0055] Optionally, the culture-sensitive decision engine specifically includes:

[0056] IP geolocation unit: three-layer positioning architecture, including a commercial database layer, accessing MaxMindGeoIP2 coordinates mapping accurate to the city level; a physical base station layer, self-built base station fingerprint library; a network topology layer, through CDN node backtracking and AS number attribution analysis; VPN penetration technology, real-time comparison of mainstream VPN service IP segments, VPN triggering, starting browser time zone detection and system language package analysis, outputting cultural circle confidence;

[0057] Symbol semantic three-level parser: hierarchical parsing mechanism, symbol type classification, YOLOv7 multi-label model based on physical attribute identification, output basic classification label and confidence; cultural circle rule matching, calling preloaded rule library; real-time risk correction;

[0058] Cross-cultural conflict arbitration agreement: dual-track arbitration flow, conflict detection, when the difference between the sender cultural weight and the receiver weight exceeds the predetermined threshold, arbitration is started; intention depth analysis, sender history analysis, statistics of the benign proportion of the same symbol used by the user; receiver sensitive scanning, calculating the same symbol proportion in the report record;

[0059] Intimate relationship correction, social intimacy factor from user portrait is divided into intimate relationship and stranger according to threshold, and the weight ratio is adjusted.

[0060] Optionally, the user context modeling module specifically comprises:

[0061] Behavior attenuation factor: dynamic toxicity history calculation, based on time attenuation weight function, dividing behavior type weighting, behavior type including reported and established, automatically intercepting content, manual review violation, withdrawing report / complaint success; comprehensive toxicity integral algorithm, combined with context correction coefficient;

[0062] Intimate correction factor: relationship-driven toxicity tolerance, multi-dimensional relationship quantification, including communication intensity factor and social graph embedding; dynamic correction is performed;

[0063] Privacy fusion under the federated learning framework: homomorphic encryption probability fusion, calculating aggregation function; attack-resistant security, gradient noise injection, Byzantine node detection, using Krum algorithm to remove malicious parameters deviating from the median by more than 3σ.

[0064] Compared with the prior art, the beneficial effects of the present application are that:

[0065] The application constructs a deep dynamic mapping mechanism of text and visual features. Through the RoBERTa-Large and EfficientNet-V2 dual-flow architecture, heterogeneous features are extracted, and the semantic space is aligned by using contrast learning, which solves the problem of text and image separation detection in traditional technology. When the user sends content with implicit conflicts, the system automatically captures cross-modal semantic contradictions, and the dynamic outlier mining module targets maliciously constructed semantic confusion samples, and real-time strengthens the model anti-interference ability, so that the hidden toxic combination has no place to hide.

[0066] Three-dimensional dynamic rule engine, symbolic physical property identification, YOLOv7 precise positioning sensitive area; regional culture adaptation, IP analysis engine loads multi-country rule library in seconds; real-time public opinion correction, automatically increases the risk weight of related symbols in sudden religious controversial events; establish a dynamic balance between respecting cultural differences and maintaining community safety, and completely end the cross-cultural misjudgment caused by the "one-size-fits-all" interception of traditional systems. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 Flowchart of the online review toxicity detection and filtering method fusing text and images;

[0068] Figure 2 Internal framework diagram of the online review toxicity detection and filtering system fusing text and images. DETAILED DESCRIPTION

[0069] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only as examples, and other obvious modifications can be thought of by those skilled in the art.

[0070] REFERENCE Figure 1 As shown in the figure, the online review toxicity detection and filtering method fusing text and images includes:

[0071] S1, extract the semantic vector of the review text through the RoBERTa-Large model; extract the visual features of the image through the EfficientNet-V2 model; use a dual-flow contrast loss function to align heterogeneous modal features;

[0072] S2, culture sensitivity adaptive decision, load regional sensitive rule table according to user IP address, dynamically adjust symbolic semantics; based on the social relationship between the publisher and the receiver, calculate the intimacy correction factor;

[0073] S3, context weighted toxicity score, calculate the user historical behavior weight, output the final toxicity probability.

[0074] It should be noted that the dual-flow projection space optimization, the text (1024 dimensions) and the image (1280 dimensions) features are independently projected to a unified 1536-dimensional space through a projection layer, and an orthogonal regularization constraint (||W t W i Τ || F Eliminate modal bias (p < 0.01);

[0075] Dynamic contrast loss function:

[0076] In the formula, L align is the dynamic contrast loss function, which is used to measure the alignment loss in contrast learning; s is the similarity score of the positive sample pair; is the similarity score of the kth hard example sample; τ is the temperature coefficient, which is used to adjust the distribution temperature coefficient of the similarity score; K is the number of hard example samples selected within the batch;

[0077] Hard example mining: select the top 3% samples with cosine similarity within the batch as negative examples,

[0078] Adaptive temperature: τ = 0.07 + 0.05 × sigmoid(batch_std) balance hard and easy sample gradient.

[0079] Engineering-level acceleration, quantization distillation, teacher model (FP32) guides EfficientNet-V2 student model (INT8), precision loss <0.4%; inference speed is improved by 3.2 times (V100 GPU: 42ms → 13ms).

[0080] Asynchronous pipeline: text input → CPU preprocessing → RoBERTa encoding → projection layer → image input → GPU parallel → EfficientNet-V2 → projection layer; projection layer → loss calculation → gradient backpropagation;

[0081] Culture-sensitive decision engine, region-sensitive rule table:

[0082] IP location region Rule version Effective field Middle East ISLAMIC_v4 Animal symbols / religious signs / limb language East Asia CONFU_v3 Historical symbols / political metaphors Western Europe WEST_v5 Hate gestures / discriminatory language

[0083] Hot update protocol: transmit incremental update package (average 78KB / time) through gRPC streaming.

[0084] The S1 specifically comprises:

[0085] Text feature extraction, using Transformer encoder, hidden layer dimension, maximum input sequence length 512 tokens;

[0086] For ultra-long reviews, implement a double-end truncation strategy to retain key semantic information; dialect normalization preprocessing, built-in dialect dictionary for real-time translation;

[0087] Image feature extraction based on EfficientNet-V2-L backbone network, uniform scaling of input image resolution, output global feature vector;

[0088] Anti-interference enhancement mechanism, against disturbance detection, inject Gaussian noise into input image, calculate Euclidean distance of features before and after disturbance, if Euclidean distance is greater than a certain threshold, it is determined as an adversarial sample and an alarm is triggered;

[0089] Sensitive area reinforcement, generate class activation heat map through Grad-CAM, apply multiple feature weights to detected text slogan area;

[0090] Feature projection layer, text projection path, map text features to a unified space through a learnable matrix, calculate the activation function; image projection path, align visual features through a learnable matrix;

[0091] Contrastive loss function, using an improved InfoNCE loss function;

[0092] Dynamic hard example mining, compare the current text with all images in the batch, calculate the similarity matrix, shield positive sample pairs, and select the negative sample index with the maximum similarity in each row; prefer to select the most similar interference image as the negative sample to make the model learn the difference between malicious semantic combinations;

[0093] Multi-modal parallel processing pipeline, CPU asynchronous data loading, GPU parallel feature extraction, synchronous stream and hardware accelerated similarity calculation.

[0094] It should be noted that the Transformer encoder is strengthened, the dynamic dimension is expanded, and the RoBERTa-Large architecture (hidden layer 1024 dimensions) is adopted. For ultra-long reviews, implement a double-end truncation strategy: retain the first 128 tokens (title / first key sentence) and the last 382 tokens (conclusion / core argument), and extract 2% key sentences in the middle segment through semantic importance sampling (based on TF-IDF weighting) to ensure 98% information retention rate.

[0095] Dialect real-time translation engine, built-in three-level dialect dictionary:

[0096] Original review → dialect detection → Cantonese → Cantonese to Mandarin model;

[0097] Dialect detection → Minnan → Minnan word segmentation;

[0098] Dialect detection → Other → ISO639-3 standard mapping;

[0099] Cantonese-to-Mandarin model / minnan wordpiece tokenizer / ISO 639-3 standard mapping→normalized text output.

[0100] OOV processing, semantic complementation through Pinyin embedding layer.

[0101] Image feature extraction, adversarial sample detection, noise injection protocol: inject Gaussian noise with mean 0 and standard deviation 0.05 into the input image; calculate the Euclidean distance of the feature vectors before and after perturbation: d = ||V orig -V noised ||2; dynamic threshold decision: if d > 0.7 x median(V orig ), trigger an alarm, discard the sample and record the attack fingerprint; median() is the median of the original feature vector.

[0102] where d is the Euclidean distance of the feature vectors before and after perturbation; V orig is the feature vector of the original image; V noised is the feature vector of the image after injecting noise;

[0103] Sensitive area reinforcement strategy, Grad-CAM heat map generation, gradient of the last layer of convolution output of EfficientNet-V2, heat map with resolution 14x14 is generated;

[0104] Text slogan area detection, use CRNN text detection model to locate the Chinese text area of the image, apply 3 times feature weight to the heat area with overlapping rate > 60%.

[0105] Dual-modal projection path:

[0106] Text projection path: T' = GELU(W t T + b t ), W t ∈R 1024×1536 , initialized with Kaiming orthogonal matrix; gradient clipping, scaled to unit norm when ;

[0107] where T' is the projected feature vector of the text; W t is the weight matrix of text projection, with size 1024x1536, initialized with Kaiming orthogonal matrix; T is the original text feature vector (1024 dimensions); b t is the bias vector of text projection; is the gradient of weight matrix W t ;

[0108] Image projection path: Introduce spectral normalization constraint W iLipschitz constant ≤ 1.5; dynamic dimension reduction, automatically compressing background class features to 512 dimensions (PCA identifies dimensions with a contribution rate < 5%);

[0109] where I' is the projected feature vector of the image; W i is the weight matrix of the image projection; I is the original image feature vector (1280 dimensions);

[0110] Improved InfoNCE loss function:

[0111]

[0112] where T' is the projected feature vector of the text; I' is the projected feature vector of the image; T k is the projected feature vector of the text corresponding to the kth negative sample; I' k is the projected feature vector of the image corresponding to the kth negative sample; s(T', I') is the similarity score between the text feature vector T' and the image feature vector I'; τ is the temperature coefficient, used to adjust the distribution of the similarity score; λ is the weight coefficient for balancing the loss of positive and negative samples; N is the number of negative samples in the batch; k is the index variable for indexing the negative samples in the batch;

[0113] Bidirectional negative sample mining, simultaneously constraining the bidirectional similarity of image-text;

[0114] Temperature coefficient adaptation: τ = 0.05 + 0.1 × sigmoid(batch_variance / 5), dynamically adjusted by the formula; τ is the temperature coefficient, with an initial value of 0.05.

[0115] Dynamic hard example mining, malicious semantic combination detection, similarity matrix calculation:

[0116] S ∈ R B×B = cosine_sim(T', I')

[0117] where S is the similarity matrix, storing the similarity scores between the text feature vector and the image feature vector; B is the batch size, defaulting to 256; T' is the feature vector of the text, the high-dimensional representation after projection; I' is the feature vector of the image, the high-dimensional representation after projection.

[0118] Negative sample screening process: shielding positive sample pairs, selecting the most similar image index for each row, selecting the most similar text index for each column, and merging the indexes;

[0119] Adversarial enhancement: adding directional disturbance to the screened negative samples

[0120]

[0121] where Ineg image feature vector representing a negative sample; gradient of the loss function L with respect to the image feature vector I, guiding the direction of the perturbation.

[0122] The S2 specifically comprises:

[0123] hierarchical symbol semantic parsing, symbol feature encoding, input symbol, mapped to 128-bit hash code, preloaded three-dimensional parsing tree retrieved through distributed cache, level one symbol type, level two geographic and cultural region, level three real-time risk index;

[0124] IP parsing pipeline, put user IP into GeoIP2 database, make cultural region judgment, load halal taboo library in Islamic region, load network meme semantic library in East Asia region, load hate symbol database in Europe and America region;

[0125] Dynamic weight calculation:

[0126] w symbol = a base_weight + b Delta risk

[0127] In the formula, w symbol is the weight of the symbol; a is the basic rule weight, a = 0.7; base_weight is; b is the real-time public opinion factor, b = 0.3; Delta risk is the negative report increment of the symbol in the target region within the last 1 hour;

[0128] Cultural conflict resolution, ambiguous symbol double verification, when the weight difference between the sender and receiver cultural circles is greater than the specified threshold, trigger the sender intent analysis, check the historical use record; receiver sensitivity scanning, query the proportion of animal symbols involved in the receiver's report record; The maximum weight takes the weighted average of the cultural threshold of both parties;

[0129] Intimacy correction factor calculation model, social relationship quantitative indicators, including private message interaction intensity, social group correlation, historical interaction emotion;

[0130] Intimacy correction factor formula:

[0131]

[0132] In the formula, gamma is the intimacy correction factor, used to adjust the influence of the message, the value range is between 0.2 and 1.0, when the receiver is in the special attention list of the sender, gamma is additionally multiplied by a factor of 1.2; f msg is the message sending frequency factor, which measures the relationship between the frequency of sending messages and the intimacy correction factor; g groupCommunity activity or similarity, value range between 0 and 1, the larger the value, the higher the activity or similarity of the community; sentiment Sentiment tendency value, representing the degree of sentiment tendency of the message, value range between -1 and 1;

[0133] Real-time relationship graph calculation, event-triggered local graph refresh, periodic execution of decay operation; graph database acceleration, using Neo4j+SparkGraphX dual engine;

[0134] Rule table hot loading, dynamic update protocol, when the regional cultural policy changes, the configuration center pushes the Zookeeper node change event; the service node receives the incremental rules through gRPC streaming; memory double-buffering switching;

[0135] De-identification processing, social relationship calculation using differential privacy technology to add Laplace noise; special processing for EU GDPR, disabling close factor calculation for European users;

[0136] Regional rule cache, deploying lightweight rule engine in each CDN edge node, American node, preloading European and American, Latin American rule library; Dubai node, preloading Islamic culture rule library; Singapore node, preloading East Asia, Southeast Asia rule library.

[0137] It should be noted that event-triggered local graph refresh, real-time update trigger:

[0138] Private message event→update communication strength; group interaction→refresh community correlation; sentiment fluctuation>0.3→recalculate sentiment tendency.

[0139] GDPR strict mode implementation, differential privacy injection, Laplace noise ensures individual untraceability:

[0140]

[0141] In the formula, γ GDPR is the correction factor after applying differential privacy in the GDPR strict mode; γ is the original correction factor, the value when differential privacy is not applied; is the Laplace noise; ∈ is the privacy budget parameter, used to control the strength of privacy protection, smaller ∈ value provides stronger privacy protection;

[0142] Special processing logic: European users automatically skip close factor calculation, relationship data storage for up to 7 days.

[0143] The S3 specifically includes:

[0144] User historical behavior weight dynamic calculation, three-dimensional behavior feature extraction, toxic comment time series distribution, define time decay function, calculate user recent behavior:

[0145]

[0146] wherein ρ toxic is the proportion of weighted toxic comments, indicating the proportion of recent toxic comments of the user, considering the time decay factor; N is the number of recent toxic comments; M is the number of all recent comments; k is an index from 1 to N, indicating the serial number of the toxic comment; j is an index from 1 to M, indicating the serial number of all comments; I is an indicator function, which is 1 when the toxic probability is greater than 0.7, and 0 otherwise;

[0147] a benign behavior gain factor, introducing a positive behavior offset mechanism; social trust assessment, calculating the trustworthiness of the user in the social graph based on a graph neural network;

[0148] multi-source signal fusion toxic probability calculation, basic signal input, signal types include text toxicity probability, image risk probability, and context conflict score; dynamic weighted fusion, adding a cultural correction term;

[0149] a three-level processing pipeline, threshold judgment is performed on the original probability, if the original probability is less than a predetermined threshold, it is directly released; if the original probability is in a predetermined threshold interval, it is added to the artificial review queue; if the original probability is greater than the predetermined threshold, it is automatically intercepted and punished, and finally the review result is fed back to the model.

[0150] It should be noted that the benign behavior gain factor:

[0151]

[0152] wherein BenefitFactor is the benign behavior gain factor, used to measure the gain brought by positive comments and effective reports; N positive is the number of positive comments; N report_valid is the number of effective reports;

[0153] signal source reliability verification, text model confidence detection, when the standard deviation of RoBERTa output probability > 0.15, trigger adversarial training to enhance reasoning; image-text conflict processing, if |p text -p image |>0.4, start cross-modal alignment verification; ptext is the confidence of the text model, p image is the confidence of the image model.

[0154] artificial review acceleration strategy, LTR optimization queue: ReviewPriority=p final ×0.7+UserRisk×0.3

[0155] In the formula, ReviewPriority represents the review priority, which is used to determine the content that needs to be reviewed in priority; p final represents the final prediction probability, reflecting the prediction confidence of the model on the sample; UserRisk represents the user risk score, reflecting the risk level of user behavior or content.

[0156] Audit result feedback loop, misjudgment cases, enhance the corresponding cultural circle rule weight; missed cases, inject countermeasures into training data.

[0157] Referring to Figure 2 The online review toxicity detection and filtering system that fuses text and image, as shown in the figure, includes:

[0158] Heterogeneous feature extraction module: text encoder generates comment text semantic vector, image encoder generates image visual feature;

[0159] Dynamic feature projection alignment module: includes text projection matrix and image projection matrix, calculates double-flow contrast loss function; difficult case mining unit, based on cosine similarity to dynamically select the most similar negative sample within the batch;

[0160] Culturally sensitive decision engine: IP geolocation unit, access MaxMindGeoIP2 database, with symbolic semantic three-level parser; set up cross-cultural conflict arbitration agreement;

[0161] User context modeling module: based on behavior decay factor, calculate weighted toxicity history; calculate intimacy correction factor; homomorphic encryption probability fusion under federated learning framework.

[0162] It should be noted that heterogeneous feature complementary enhancement, cross-modal attention bridge, text semantic vector (RoBERTa output) and image visual feature (EfficientNet-V2 output) realize information complementation through bidirectional cross-attention mechanism:

[0163]

[0164] In the formula, A t2i represents the text-to-image attention weight matrix; A t2t represents the image-to-text attention weight matrix; T is the text feature vector (RoBERTa output); I is the image feature vector (EfficientNet-V2 output); W q is the query weight matrix, used to project the query vector into the feature space; W k is the key weight matrix, used to project the key vector into the feature space; d is the dimension of the feature vector, used to scale the attention score;

[0165] Through the bidirectional cross-attention mechanism, the text and image features can complementarily enhance each other's information, thereby improving the representation ability and performance of the model.

[0166] Culture adaptation, dynamic analysis of symbolic semantics, real-time risk propagation algorithm, and diffusion of symbolic risk value in social networks; geographical heat map driven, regional sensitivity heat map generated based on the real-time reporting of symbolic usage frequency of CDN nodes;

[0167] Conflict arbitration dual-channel verification, sender intention analysis, extraction of context features (emotional tendency, collocation keywords) of the symbol in historical behavior; receiver defense mechanism, detection of the frequency of similar symbols in the receiver's shielding list; arbitration logic, when the intention analysis confidence is <0.4, the receiver's rules are preferred.

[0168] The dynamic feature projection alignment module specifically comprises:

[0169] Text and image projection matrix: nonlinear mapping, text projection path, input text vector, linear transformation through a learnable matrix, output projection vector through a GELU activation function, and negative value information is retained; image projection path, input EfficientNet-V2 image features, mapped to the same dimension through an independent projection matrix, and orthogonal constraint is applied to force the image projection direction to be decoupled from the text space; dynamically identify the output dimension of the upstream encoder, and automatically adjust the output channel of the projection matrix to a unified high-dimensional space;

[0170] Dual-flow contrast loss function: hybrid similarity calculation, definition of improved similarity measurement function, dynamic temperature coefficient, automatic adjustment according to the discrete degree of similarity distribution within the batch; gradient equalization back propagation, monitoring the ratio of the text projection gradient norm and the image gradient norm;

[0171] Hard example mining unit: dynamic negative sample screening algorithm, similarity matrix construction, calculation of the cosine similarity matrix of all image-text pairs within the batch, diagonal line shielding, Top-K difficult negative sample extraction, adversarial enhancement, addition of directional disturbance to the screened negative samples; semantic confusion enhancement, text level, random replacement of keywords to generate semantic conflict negative samples; image level, local tampering of negative sample images using StyleGAN2; mixed enhancement, exchange of image-text components of positive sample pairs.

[0172] It should be noted that the nonlinear mapping, the differential equation of the text projection path, and the input text feature vector T ∈ R dt The projection process can be modeled as:

[0173]

[0174] In the formula, H represents the hidden state, which is usually the state of the text feature vector after projection; t represents the time step or the number of iterations; σ represents the activation function; Wt represents the text projection weight matrix, which is used to map the input text feature vector to the hidden space; T represents the input text feature vector; bt represents the text projection bias vector; α represents the decay coefficient, which is used to control the decay rate of the hidden state;

[0175] Analytical solution of GELU activation:

[0176] In the formula, x is the input value, usually an element of a feature vector; β and γ are parameters of the GELU function, used to control the shape of the function;

[0177] Negative value preservation mechanism, by introducing a complex function mapping f(z) = z + 0.3i·Im(z), preserving negative phase information in complex space;

[0178] In the formula, z is a complex number, representing a feature vector; i is the imaginary unit; Im(z) represents the imaginary part of the complex number z.

[0179] Multi-modal adversarial enhancement:

[0180]

[0181] The culture-sensitive decision engine specifically includes:

[0182] IP geolocation unit: three-layer positioning architecture, including a commercial database layer, accessing MaxMindGeoIP2 coordinates mapping accurate to the city level; a physical base station layer, self-built base station fingerprint library; a network topology layer, through CDN node backtracking and AS number attribution analysis; VPN penetration technology, real-time comparison of mainstream VPN service provider IP segments, VPN triggering starting browser time zone detection and system language package analysis, outputting cultural circle confidence;

[0183] Symbol semantic three-level parser: hierarchical parsing mechanism, symbol type classification, identifying symbol physical properties based on YOLOv7 multi-label model, outputting basic classification labels and confidence; cultural circle rule matching, calling preloaded rule library; real-time risk correction;

[0184] Cross-cultural conflict arbitration agreement: dual-track arbitration flow, conflict detection, when the difference between the sender's cultural weight and the receiver's weight exceeds a certain threshold, arbitration is started; intention depth analysis, sender history analysis, statistics of the benign proportion of the same symbol used by this user; receiver sensitive scanning, calculating the proportion of the same symbol in its report records;

[0185] Intimate relationship correction, social intimacy factors from user portraits are divided into intimate relationships and strangers according to a threshold, and the weight ratio is adjusted.

[0186] Note that the three-layer positioning architecture improves accuracy through heterogeneous data fusion. The commercial database layer calls MaxMindGeoIP2's City-level coordinate mapping and combines AS number attribution analysis to achieve city-level positioning. The physical base station layer relies on a self-built base station fingerprint library to construct a spatial grid through radio frequency signal characteristics (RSRP / RSSI / TA) and achieve 50-meter-level positioning on mobile devices. The network topology layer is based on CDN node backtracking and BGP route analysis to identify cross-border proxy traffic paths.

[0187] VPN penetration technology adopts a dynamic countermeasures strategy: real-time comparison with 187 VPN service provider IP segment libraries (updated every 30 seconds); trigger dual-factor verification: browser time zone and IP time zone deviation > 2 hours, or system language package and location country official language mismatch;

[0188] Output three-dimensional cultural confidence vector: [religious weight (0-1), language weight (0-1), political sensitivity (0-1)], such as the Islamic cultural circle output [0.92, 0.15, 0.67].

[0189] Symbolic semantic three-level parser, hierarchical parsing process realizes dynamic quantification of symbolic risk:

[0190] Physical property identification, YOLOv7-MultiLabel model is used to detect 38 types of basic symbols (religion / politics / sexual hints, etc.); output multi-label confidence, such as "Nazi symbol: 0.93, historical symbol: 0.21"

[0191] Cultural rule matching, preloaded rule library contains 12 cultural circles and 195 country-specific taboos (such as the East Asian region avoiding the number 4); real-time access to public opinion API: when the symbol spreads on Twitter / Facebook within 1 hour and the spread volume increases by 300%, trigger risk reevaluation;

[0192] Rule hot update protocol ensures timeliness, deploys Zookeeper to monitor cultural policy changes in 195 countries, incrementally updates through gRPC stream push to edge nodes, and takes effect globally within 8 seconds.

[0193] Cross-cultural conflict arbitration protocol, dual-track arbitration flow realizes two-way benefit balance:

[0194] Conflict detection, start when the sender and receiver cultural weight difference Δw > 0.4 (such as an American user Δw = 0.8 sending a thumbs-up gesture to an Iranian user)

[0195] Intention deep analysis, sender history analysis, calculate the proportion of benign use of the same symbol (recent behavior accounts for 82% of the weight); receiver sensitive scanning, quantify the proportion of the same symbol in its complaint records.

[0196] The user context modeling module specifically includes:

[0197] Behavior decay factor: dynamic toxicity history calculation, based on time decay weight function, behavior type division weighted, behavior types include reported and established, automatically intercepted content, manual review violations, withdrawal of reports / complaints successful; comprehensive toxicity score algorithm, combined with context correction coefficient;

[0198] Intimate correction factor: relationship-driven toxicity tolerance, multi-dimensional relationship quantification, including communication intensity factor and social graph embedding; dynamic correction is performed;

[0199] Privacy fusion under the federated learning framework: homomorphic encryption probability fusion, computing aggregation function; attack-resistant security, gradient noise injection, Byzantine node detection, using Krum algorithm to remove malicious parameters deviating from the median by more than 3σ.

[0200] It should be noted that the behavior type differentiation processing, the complaint success negative weight hedging, when the report is withdrawn or the complaint is established, a negative decay pulse is generated: ΔT = -0.8 × original integral × e -0.2t0

[0201] Instantaneous offset history record impact (e.g. original violation score 1.2 → offset to 0.24)

[0202] Context correction coefficient: political sensitive period (election / diplomatic event), violation weight × 1.5; entertainment scene emoticon, low-risk violation weight × 0.3;

[0203] Comprehensive toxicity score algorithm:

[0204]

[0205] In the formula, ToxScore represents the comprehensive toxicity score, which is used to measure the toxicity degree of text or content; TypeWeight(k) represents the weight of the kth type of toxic behavior, and the type weight is configured as reported and established = 1.0, automatically intercepted = 0.8, and manually reviewed = 0.6; w(t_k represents the time weight of the kth type of toxic behavior, which is usually related to the time when the toxic behavior occurs, and can be used to reflect its timeliness or influence; α represents the context coefficient, which ranges from 0.2 to 1.8, and is dynamically adjusted by the public opinion monitoring system, which is used to adjust the toxicity score according to the context environment; CtxFactor represents the context factor, which reflects the influence of the current context environment on the toxicity score, and can contain user historical behavior, content type, etc. information; N represents the total number or type of toxic behaviors; k represents the index variable of toxic behavior, from 1 to N.

[0206] Type weight configuration: complaint upheld = 1.0, automatic blocking = 0.8, manual review = 0.6; context coefficient a range 0.2-1.8 (dynamically adjusted by public opinion monitoring system).

[0207] Embodiment: Cross-border e-commerce community platform

[0208] Event: Saudi Arabian user (publisher) sends product review to American user (recipient)

[0209] Step 1: Text semantic analysis, input text in Arabic, processed by RoBERTa-Large model, identify sensitive words → output political conflict semantic vector, detect implicit provocation intent (confidence 0.92);

[0210] Step 2: Image feature extraction, upload product image, blue six-point star logo (visual approximation of Israeli flag), EfficientNet-V2 model processing, locate logo area (YOLOv7 output: religious symbol / confidence 0.89); adversarial filtering, confirm non-AI tampering (true image probability > 95%);

[0211] Step 3: Dual-flow feature alignment, contrastive learning forces text and image features to be associated, text vector: "flag" concept → mapped to image six-point star area; conflict detection, image-text combination strengthens political sensitive attributes;

[0212] Step 4: Load regional rules, publisher IP located in Jeddah, Saudi Arabia → trigger Middle East sensitive rule library, prohibit public mention of Israeli sovereign symbols (risk weight +0.6), automatically load halal taboo word list;

[0213] Step 5: Recipient cultural adaptation, recipient IP located in New York, USA → cultural weight analysis; USA, high tolerance for political symbols (base weight 0.3), Saudi Arabia, high sensitivity to religious symbols (base weight 0.9); cross-cultural conflict value Δw = 0.6 (exceeds threshold 0.4);

[0214] Step 6: Intimacy correction factor calculation, query relationship between both parties, 12 private messages in 7 days (communication intensity medium), join 3 shopping groups together (social graph correlation 0.65); output correction factor is 0.45 (partially reduce sensitivity);

[0215] Step 7: Publisher historical behavior analysis, record 2 violations in 30 days (1 automatic blocking + 1 complaint upheld), 1 successful appeal (negative weight offset); decay calculation, latest violation (5 days ago) retains 40% weight, successful appeal reduces historical toxicity by 80%, final user toxicity baseline is 0.38 (medium risk);

[0216] Step 8: Real-time risk synthesis, original risk of cultural conflict is 0.85 (Saudi rule trigger); weighted adjustment, intimacy correction factor 0.45 → risk reduced to 0.85x0.55≈0.47, superimposed user toxicity baseline 0.38 → comprehensive score 0.71;

[0217] Step 9: Hierarchical treatment, score 0.71 → medium risk interval (0.6-0.85); the action includes automatic transfer to manual review (priority P1), the receiver interface displays "content under review", and the publisher receives "send success" but limited display;

[0218] Step 10: Artificial review results, decision is violation (violation of Middle East political sensitive policy), treatment is to delete the content and notify the publisher, and the publisher's account deducts 15 points of credit score;

[0219] Step 11: Model incremental learning, case added to federal training set, update Saudi rule library, add "six-point star-Israel" mapping rule;

[0220] Step 12: User complaint processing, publisher appeals "unintentional offense" → system verifies no malicious record in history, returns 10 points of credit score, and retains 5 points of warning.

[0221] In summary, the advantages of the present application are:

[0222] Breakthrough in image-text segmentation detection bottleneck, extract heterogeneous features through RoBERTa-EfficientNet dual-stream architecture, and realize dynamic alignment of semantic space through contrast learning. When the user sends content with implicit conflict, the system accurately captures cross-modal contradictions, and combines with difficult example mining mechanism to strengthen anti-interference ability, solving the semantic blind area of traditional single-mode detection.

[0223] Three-dimensional dynamic rule engine, symbolic physics identification, YOLOv7 positioning sensitive area; regional rule second loading, Middle East IP automatic activation of halal taboo library; real-time public opinion correction, religious controversial event minute-level adjustment; social graph driven intimacy arbitration mechanism, realize the dynamic balance of cross-cultural tolerance and safety control, end "one-size-fits-all" misjudgment.

[0224] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection required by the present application is defined by the appended claims and their equivalents.

Claims

1. An online review toxicity detection and filtering method that fuses text and image, characterized in that, The method comprises the steps of: S1, extracting the semantic vector of the comment text through the RoBERTa-Large model, and extracting the visual feature of the image through the EfficientNet-V2 model; Aligning heterogeneous modal features by using a dual-flow contrast loss function; S2, culture-sensitive adaptive decision-making, loading regional sensitive rule table according to user IP address to dynamically adjust symbolic semantics; calculating the intimacy correction factor based on the social relationship between the publisher and the receiver; S3, context-weighted toxicity score, calculating user historical behavior weight, and outputting the final toxicity probability. 2.The method of claim 1, wherein, The S1 specifically comprises: Text feature extraction, using a Transformer encoder, a hidden layer dimension, and a maximum input sequence length of 512 tokens; For super-long comments, implement a double-end truncation strategy to retain key semantic information; dialect normalization preprocessing, built-in dialect dictionary for real-time translation; Image feature extraction based on the EfficientNet-V2-L backbone network, uniform scaling of the input image resolution, and output of the global feature vector; Anti-interference enhancement mechanism, against disturbance detection, injecting Gaussian noise into the input image, calculating the Euclidean distance of the features before and after the disturbance, and if the Euclidean distance is greater than a certain threshold, it is determined as an adversarial sample and an alarm is triggered; Sensitive area reinforcement, generating class activation heat maps through Grad-CAM, and applying multiple feature weights to the detected text slogan area. 3.The online review toxicity detection and filtering method of fusing text and image according to claim 2, characterized in that, The S1 specifically comprises: Feature projection layer, text projection path, mapping text features to a unified space through a learnable matrix, and calculating an activation function; image projection path, aligning visual features through a learnable matrix; Contrast loss function, using an improved InfoNCE loss function; Dynamic hard example mining, comparing the current text with all images in the batch, calculating a similarity matrix, shielding positive sample pairs, and selecting the negative sample index with the maximum similarity in each row; preferentially selecting the most similar interference image to the current text as the negative sample, so that the model learns the difference in malicious semantic combination; Multi-modal parallel processing pipeline, CPU asynchronous data loading, GPU parallel feature extraction, and synchronous stream and hardware accelerated similarity calculation. 4.The method of claim 3, wherein, The S2 specifically comprises: Hierarchical symbolic semantic analysis, symbolic feature encoding, inputting symbols, mapping to 128-bit hash codes, retrieving preloaded three-dimensional analysis trees through distributed caching, level one being symbol type, level two being geographical and cultural region, and level three being real-time risk index; IP analysis pipeline, putting user IP into GeoIP2 database for cultural region determination, loading the Islamic region taboo library in the Islamic region, loading the network meme semantic library in East Asia, and loading the hate symbol database in Europe and America; Dynamic weight calculation: w symbol = a base weight + β Δ risk In the formula, w symbol is the weight of the symbol; a is the basic rule weight, a = 0.7; base_weight is; β is the real-time public opinion factor, β = 0.3; Δ risk is the negative report increment of the symbol in the target area in the last 1 hour; Culture conflict resolution, ambiguous symbol double verification, when the weight difference of a symbol in the sender and receiver cultural region is greater than a certain threshold, triggering sender intent analysis to check historical usage records; receiver sensitivity scanning, querying the proportion of animal symbols involved in the receiver's report records; the final weight is the weighted average of the cultural threshold of both parties.

5. The method of online review toxicity detection and filtering that fuses text and image as claimed in claim 4, wherein, The S2 specifically comprises: Intimacy correction factor calculation model, social relationship quantitative indicators include private message interaction intensity, community correlation, historical interaction sentiment; Intimacy correction factor formula: wherein γ is an affinity modifier factor to adjust the influence of the message, with a value range between 0.2 and 1.0, and γ is additionally multiplied by a factor of 1.2 when the receiver is in the special attention list of the sender; f msg is a message sending frequency factor to measure the relationship between the frequency of sending the message and the affinity modifier factor; g group is a community activity or similarity, with a value range between 0 and 1, and the larger the value, the higher the activity or similarity of the community; s sentiment is an emotional tendency value to represent the degree of emotional tendency of the message, with a value range between -1 and 1. Real-time relationship graph calculation, event-triggered local graph refresh, periodic decay operation; graph database acceleration, using Neo4j+SparkGraphX dual engine; Rule table hot loading, dynamic update protocol, when regional cultural policy changes, configuration center pushes Zookeeper node change event; service node receives incremental rules through gRPC streaming; memory double-buffering switching; De-identification processing, social relationship calculation uses differential privacy technology to add Laplace noise; special processing for EU GDPR, disable intimacy factor calculation for European users; Regional rule cache, deploy lightweight rule engine at each CDN edge node, American node, preload North America, South America rule library; Dubai node, preload Islamic culture rule library; Singapore node, preload East Asia, Southeast Asia rule library.

6. The method of fusing text and image based online review toxicity detection and filtering according to claim 5, wherein, The S3 specifically includes: User historical behavior weight dynamic calculation, three-dimensional behavior feature extraction, toxic comment time series distribution, define time decay function, calculate user recent behavior: where ρ toxic is the proportion of weighted toxic comments, representing the proportion of recent toxic comments of the user, considering the time decay factor; N is the number of recent toxic comments; M is the number of all recent comments; k is an index from 1 to N, representing the serial number of toxic comments; j is an index from 1 to M, representing the serial number of all comments; I is an indicator function, which is 1 when the toxicity probability is greater than 0.7, and 0 otherwise. Benign behavior gain factor, introduce positive behavior offset mechanism; social trust evaluation, calculate user trustworthiness in social graph based on graph neural network; Toxicity probability calculation with multi-source signal fusion, basic signal input, signal types include text toxicity probability, image risk probability, context conflict score; dynamic weighted fusion, add cultural correction term; Three-level processing pipeline, threshold judgment on original probability, if the original probability is less than the specified threshold, directly pass; if the original probability is in the specified threshold interval, add to the manual review queue; if the original probability is greater than the specified threshold, automatically intercept and punish, and finally the review result is fed back to the model.

7. An online review toxicity detection and filtering system that fuses text and image for implementing the online review toxicity detection and filtering method as claimed in claims 1-6, characterized in that, Including: Heterogeneous feature extraction module: text encoder generates comment text semantic vector, image encoder generates image visual feature; Dynamic feature projection alignment module: includes text projection matrix and image projection matrix, calculates double-flow contrast loss function; difficult example mining unit, dynamically selects the most similar negative sample in batch based on cosine similarity; Cultural sensitive decision engine: IP geolocation unit, access MaxMindGeoIP2 database, with symbolic semantic three-level parser; Set cross-cultural conflict arbitration protocol; User context modeling module: based on behavior decay factor, calculate weighted toxicity history; Calculate intimacy correction factor; Homomorphic encryption probability fusion under federated learning framework.

8. The system for fusing text and image based online review toxicity detection and filtering as claimed in claim 7, wherein, The dynamic feature projection alignment module specifically includes: Text and image projection matrix: nonlinear mapping, text projection path, input text vector, after linear transformation by learnable matrix, output projection vector through GELU activation function, negative value information is preserved; image projection path, input EfficientNet-V2 image feature, mapped to the same dimension through an independent projection matrix, and orthogonal constraint is applied to force the image projection direction to be decoupled from the text space; dynamically identify the output dimension of the upstream encoder, automatically adjust the output channel of the projection matrix to a unified high-dimensional space; Dual-flow contrastive loss function: hybrid similarity calculation, define improved similarity metric function, dynamic temperature coefficient, automatically adjust according to the dispersion degree of similarity distribution within the batch; gradient equalization back propagation, monitor the ratio of text projection gradient norm and image gradient norm; Difficult case mining unit: dynamic negative sample screening algorithm, similarity matrix construction, calculate the cosine similarity matrix of all text-image pairs within the batch, diagonal line shielding, Top-K difficult negative sample extraction, adversarial enhancement, add directional disturbance to the screened negative samples; semantic confusion enhancement, text level, randomly replace keywords to generate semantic conflict negative samples; image level, use StyleGAN2 to locally tamper with negative sample images; mixed enhancement, exchange the text and image components of the positive sample pair.

9. The system for fusing text and image based online review toxicity detection and filtering as claimed in claim 8, wherein, The culture-sensitive decision engine specifically includes: IP geolocation unit: three-layer positioning architecture, including a commercial database layer, accessing MaxMindGeoIP2 coordinates mapping accurate to the city level; a physical base station layer, self-built base station fingerprint library; a network topology layer, analyzing CDN node backtracking and AS number attribution through VPN penetration technology, real-time comparison of mainstream VPN service IP segments, VPN triggering, starting browser time zone detection and system language package analysis, and outputting cultural circle confidence; Symbol semantic three-level parser: hierarchical parsing mechanism, symbol type classification, identifying symbol physical attributes based on a YOLOv7 multi-label model, and outputting basic classification labels and confidence; cultural circle rule matching, calling a preloaded rule library; real-time risk correction; Cross-cultural conflict arbitration agreement: dual-track arbitration flow, conflict detection, when the difference between the sender's cultural weight and the receiver's weight exceeds a certain threshold, arbitration is started; intent depth analysis, sender history analysis, statistics of the benign proportion of the same symbol used by the user; receiver sensitivity scanning, calculating the proportion of the same symbol in its report records; close relationship correction, dividing social closeness factors from user portraits into close relationships and strangers according to a threshold, and adjusting the weight ratio.

10. The system for fusing text and image based online review toxicity detection and filtering of claim 9, wherein, The user context modeling module specifically includes: Behavior decay factor: dynamic toxicity history calculation, based on a time decay weight function, dividing behavior types for weighting, including reported and established, automatically intercepted content, manually reviewed violations, and successful withdrawal of reports / complaints; comprehensive toxicity score algorithm, combined with context correction coefficients; Close correction factor: relationship-driven toxicity tolerance, multi-dimensional relationship quantification, including communication intensity factor and social graph embedding; dynamic correction; Privacy fusion under the federated learning framework: homomorphic encryption probability fusion, calculating aggregation functions; attack-resistant security, gradient noise injection, Byzantine node detection, using Krum algorithm to remove malicious parameters deviating from the median by more than 3σ.

Images