Improving Offensive and Hate Speech (OHS) classifiers’ performances requires a large, confidently labeled textual training dataset. Our study devises a semi-supervised classification approach with self-training to leverage the abundant social media content and develop a robust OHS classifier. The classifier is self-trained iteratively using the most confidently predicted labels obtained from an unlabeled Twitter corpus of 5 million tweets. Hence, we produce the largest supervised Arabic OHS dataset. To this end, we first select the best classifier to conduct the semi-supervised learning by assessing multiple heterogeneous pairs of text vectorization algorithms (such as N-Grams, World2Vec Skip-Gram, AraBert and DistilBert) and machine learning algorithms (such as SVM, CNN and BiLSTM). Then, based on the best text classifier, we perform six groups of experiments to demonstrate our approach’s feasibility and efficacy based on several self-training iterations.

Online auctions have become one of the most convenient ways to commit fraud due to a large amount of money being traded every day. Shill bidding is the predominant form of auction fraud, and it is also the most difficult to detect because it so closely resembles normal bidding behavior. Furthermore, shill bidding does not leave behind any apparent evidence, and it is relatively easy to use to cheat innocent buyers. Our goal is to develop a classification model that is capable of efficiently differentiating between legitimate bidders and shill bidders. For our study, we employ an actual training dataset, but the data are unlabeled. First, we properly label the shill bidding samples by combining a robust hierarchical clustering technique and a semi-automated labeling approach. Since shill bidding datasets are imbalanced, we assess advanced over-sampling, under-sampling and hybrid-sampling methods and compare their performances based on several classification algorithms. The optimal shill bidding classifier displays high detection and low misclassification rates of fraudulent activities.

Our study explores offensive and hate speech detection for the Arabic language, as previous studies are minimal. Based on two-class, three-class, and six-class Arabic-Twitter datasets, we develop single and ensemble CNN and BiLSTM classifiers that we train with non-contextual (Fasttext-SkipGram) and contextual (Multilingual Bert and AraBert) word-embedding models. For each hate/offensive classification task, we conduct a battery of experiments to evaluate the performance of single and ensemble classifiers on testing datasets. The average-based ensemble approach was found to be the best performing, as it returned F-scores of 91%, 84%, and 80% for two-class, three-class and six-class prediction tasks, respectively. We also perform an error analysis of the best ensemble model for each task.