Individual vs. Group Violent Threats Classification in Online Discussions
Noman Ashraf, Rabia Mustafa, Grigori Sidorov, Alexander Gelbukh
Abstract
Violent threat is a serious crime affecting the targeted individuals or groups. It is essential for media providers to block the users that post such threats. In this paper, we focused on detection of violent threat language in YouTube comments. We categorized the threatening comments into those targeting an individual or a group. We started from an existing dataset with violent threat language identified, but without any categorization into comments targeting individuals or groups. We adopted a binary classification approach for the prediction of individual- vs. group-targeting threats. We compared two text representations: bag of words (BOW) and pre-trained word embedding such as GloVe and fastText. We used deep-learning classifiers such as 1D-CNN, LSTM, and bidirectional LSTM (BiLSTM). GloVe embedding showed the worst results, fastText performed much better, and BiLSTM on BOW with term frequency-inverse document frequency (TF-IDF) weighting scheme gave the best results, achieving 0.94% ROC-AUC and Macro-F1 score of 0.85%.