New AI model reads text, images and social networks to spot cyberbullying

Researchers in Kazakhstan have built an artificial intelligence system that detects cyberbullying on social media by combining five different kinds of information at once: the words in a post, the surrounding conversation thread, the network of interactions among users, attached images or memes, and behavioral metadata such as posting times and engagement statistics. The model, called HMTF-Net, short for Hierarchical Multimodal Transformer Fusion Network, was described in a paper published in Frontiers in Artificial Intelligence on 29 October 2026. In benchmark tests on a publicly available multimodal cyberbullying dataset, the system reached 97.6 percent accuracy, 97.0 percent precision, 96.2 percent recall, a 96.6 percent F1-score, and a ROC-AUC of 0.986, outperforming a series of comparison methods that relied on narrower slices of the available data.

The motivation behind the work lies in a persistent weakness of most existing detection systems. Text-only classifiers, even those built on powerful transformer language models, often miss abuse that is implicit, sarcastic, or spread across multiple messages. A single reply can look harmless in isolation but become threatening in the context of a hostile thread. The authors of the new study argue that understanding whether a message constitutes cyberbullying frequently requires considering previous replies, relationships among users, attached memes or images, and temporal posting behavior. Coordinated harassment campaigns, in particular, involve complex interaction networks that cannot be effectively modeled through isolated text representations alone.

HMTF-Net addresses this by processing each modality through a dedicated representation-learning branch before fusing the results. The textual branch uses an embedding layer followed by one-dimensional convolutional layers and bidirectional gated recurrent units, a choice the authors justify on efficiency grounds: the encoder is reused across every conversation turn alongside four other modalities, and the convolution-plus-recurrence design captures local patterns such as obfuscated insults, hashtags, and short abusive phrases at lower memory and parameter cost than fine-tuning a separate pre-trained transformer for every post and reply. BERT was retained as a text-only comparator, and it is notably larger, with 109.5 million parameters and 28.9 milliseconds of inference time per sample, compared with 47.5 million parameters and 23.7 milliseconds for the complete HMTF-Net.

Conversation context is handled by a hierarchical encoder that treats each thread as an ordered sequence of utterances. Every utterance is independently encoded, and a hierarchical attention mechanism then assigns importance weights to individual utterances, allowing the model to focus on the most informative exchanges while suppressing irrelevant replies. This design reflects the observation that isolated messages may appear harmless without considering preceding discussions, and that sustained harassment differs qualitatively from a single offensive expression.

Structure among users is captured by a heterogeneous graph encoder. For each target post, the system builds a graph containing three entity types: user nodes, post nodes, and reply or comment nodes, connected by typed edges encoding author-of, reply-to, mention, and follow relationships. Direction is preserved for authorship, replies, and mentions, and inverse edges are added during message passing so evidence can flow between content and users. Stacked Graph Attention Network layers then aggregate neighborhood information with adaptive importance weights, and global mean pooling produces a graph-level representation. According to the authors, this construction helps the model identify repeated targeting, dense reply patterns, mentions of the same victim, and coordinated participation, structural evidence unavailable from an isolated message.

Visual content is extracted from attached images and memes using a convolutional neural backbone with global average pooling, while metadata and temporal attributes are projected into an embedding space and processed through a bidirectional recurrent encoder. The metadata branch draws on fields available across source platforms, including like or reaction counts, reply counts, share counts, account age, follower and following counts, total public-post counts, verification status, and temporal variables derived from timestamps such as hour of day, day of week, and elapsed time from the first message in a thread. A binary availability mask identifies fields absent on a given platform, and missing values are imputed with training-set statistics, an approach intended to keep the system usable across heterogeneous platforms.

Fusion of the five modality-specific representations is performed by an adaptive cross-modal attention mechanism. Each modality generates query, key, and value matrices, and scaled dot-product attention across multiple heads learns interactions among the heterogeneous inputs. Because different modalities contribute unequally to detection, learned gating coefficients weight the fused representation, and layer normalization, dropout, and feed-forward projection refine it before classification. The fused representation drives multiple prediction heads simultaneously: rather than performing only binary cyberbullying detection, the framework also predicts toxicity level, offensive language, target category, severity, emotion, and platform identity.

To improve generalization across platforms, the training process incorporates a domain-adversarial strategy. A gradient reversal layer encourages the encoder to learn platform-independent representations while a domain discriminator attempts to predict platform labels, forcing features that do not betray which site a message came from. The authors note that these transfer results were computed as unweighted macro-averages of per-platform test scores, so larger target corpora could not dominate the reported values.

The comparison experiments placed HMTF-Net against five representative baselines trained on identical data partitions: CNN plus text, BiLSTM, Text-BERT, GraphSAGE, and MM-Transformer. The strongest competitor, MM-Transformer, which combines textual and visual features, reached 92.8 percent accuracy and a 90.7 percent F1-score, well below the proposed system. An ablation study traced the contribution of each component: adding the hierarchical conversation encoder to a text-only baseline raised the F1-score from 74.3 to 82.3 percent, the heterogeneous graph encoder lifted it to 87.1 percent, cross-modal attention fusion brought it to 90.7 percent, and the complete architecture reached 96.6 percent. Robustness tests under noisy conditions, including random word deletion, synonym replacement, injected noise, and platform-specific slang, produced only minor degradation for HMTF-Net while baselines fell substantially further.

A distinctive feature of the system is its explainability module, which highlights the evidence behind each prediction: influential words such as the abusive terms and hashtags that dominated the text attention maps, the conversation turns carrying the most weight, the graph nodes, particularly the author and directly connected users, that drove structural inference, and the image regions that contributed to the visual judgment. The authors argue that such interpretable output matters for content moderation, where incorrect predictions directly affect users and moderators need transparent reasoning to validate automated decisions. They also caution that opportunities remain, including richer behavioral history, more advanced vision-language models for complex memes and sarcasm, continual learning for evolving vocabulary, and privacy-preserving techniques such as federated learning and differential privacy. The dataset used in the study, MultiBully, is publicly available, allowing independent verification of the reported results.

Subject of Research: A multimodal transformer-based deep learning framework for explainable cyberbullying detection on social media platforms

Article Title: A hierarchical multimodal transformer fusion network for cyberbullying detection across social media platforms

Article References: Zhidebayeva, A., Toktarova, A., Abdrakhmanov, R., Berkimbayev, K., Bekbolat, M., & Tolep, A. (2026). A hierarchical multimodal transformer fusion network for cyberbullying detection across social media platforms. Frontiers in Artificial Intelligence, 9, 1926150. https://doi.org/10.3389/frai.2026.1926150

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1926150

Keywords: cyberbullying, artificial intelligence, natural language processing, multimodal learning, graph neural networks, transformers, social media, content moderation, explainable AI, domain adaptation, deep learning, online harassment