EOT-Net: An Emoji-Aware Obfuscated Toxic Chat Detection and Moderation Framework for Streaming Platforms

Objectives: This research introduces a system that detects toxic content in a hidden manner in real time communication services by using an intelligent system that detects toxic content in a conversation. The system would try to solve the problems that the current moderation approach had, and recognize the toxic phrases in the message, substitute them with symbols and characters (emoji/symbols) and leave the automatic moderation function unchanged. Method: This study uses an EOT-Net (Emoji-Aware Obfuscated Toxic Chat Detection and Moderation Framework), which involves ORL (Obfuscation Recovery Layer), EAE (Emoji-Aware Semantic Encoder), a module of the Contextual Feature Extraction based on the RoBERTa network, a BiLSTM network and a Multi-Head Attention mechanism. Adaptive Moderation Engine provides warnings, action suppression and action blocking due to toxicity. This framework was tested with an integrated Kaggle dataset comprising of roughly 435,000 chat messages that are toxic and non-toxic. Findings: The experiments have demonstrated the usefulness of the framework proposed for detection of different types of toxic communications that already exist. The accuracy, precision, recall, f1 score and AUC-ROC of EOT-Net are 97.14%, 96.52%, 96.03%, 96.27%, and 0.984 respectively, outperforming the basic models (CNN, BiLSTM, BERT, and RoBERTa). The results demonstrate the usefulness of obfuscation recovery and the knowledge of the meaning of emojis to improve the effectiveness of the hidden toxicity detection and moderation. Novelty: The novelty of this research is the obfuscation recovery, emoji-aware semantic encoding, contextual transformer learning, attention-based toxicity extraction and adaptive moderation. By using emojis to detect hidden abusive expressions and threats, EOT-Net outcompetes the existing methods which are only able to detect the explicit abusive expressions, so that the automated real-time moderation of next generation content moderation systems is possible. Keywords: Toxic Chat Detection, Emoji-Aware Learning, Obfuscated Text Recovery, Deep Learning, RoBERTa, BiLSTM, Multi-Head Attention, Content Moderation, Hate Speech Detection, Streaming Platforms