An exploration of novel identification techniques for mitigating multi-label online toxicity on social media platforms

Social media have emerged as some of the most mainstream media communication platforms yet they are also the platforms where different forms of online toxicity exist; hate speech, harassment, cyberbullying, threats, and obscene content. Such content is difficult to detect because of overlapping labels, ambiguity of contexts and class imbalance in a multi-label environment. This paper introduces a hybrid architecture, which is called Identification of Multi-Label Toxicity Classification (IMLTC), to fine-grained multi-label toxicity detection in texts. The method uses contextual embeddings obtained through transformer models ALBERT, RoBERTa and DistilBERT in order to identify subtle semantic relations in comments. Rather than using standard end to end transformer classification heads, which share the same output of sigmoids, the suggested technique divides the representation learning and decision modelling. The obtained embeddings are then used to classify them with an XGBoost algorithm in a one-vs-rest setup, which allows the label-specific nonlinear decision boundary and offers better performance on overlapping and imbalanced groups of toxicity. The dataset on Google Jigsaw Toxic Comment Classification Challenge was experimented. The experimental design involves comparative analysis against various baseline models, including BERT with Multinomial Naïve Bayes (MNB), FastText with Random Forest (RF), and GloVe with Support Vector Machine (SVM), and makes sure that there is a balanced benchmarking of the classical, hybrid, and transformer-based paradigm. As standard multi-label measures, accuracy and F1-score were used to determine performance. The suggested IMLTC framework had a competitive and strong score of 95.34 percent and 95.39 percent accuracy and F1-score respectively. These findings are shown by the fact that the designed combination of contextual embeddings and gradient boosted label specific classifiers provide an effective and scalable solution to intelligent content moderation in the online setting.