Threatening Language Detection and Target Identification in Urdu Tweets
Maaz Amjad, Noman Ashraf, Alisa Zhila, Grigori Sidorov, Arkaitz Zubiaga, Alexander Gelbukh
Abstract
Automatic threatening language detection is an important task and most of the existing studies relied on English. However, threatening language detection in poor-resource language remains briefly addressed. In this paper, we introduce a new publicly available dataset for threatening language detection in Urdu tweets to fill the scientific gap, particularly, in the Urdu language. The proposed dataset contains 3,564 tweets manually annotated by human experts with two labels: threatening and non-threatening. The threatening tweets are further classified into two classes: threatening to an individual person or threatening to a group. This research follows a two-step approach: (i) classify a given tweet as threatening or non-threatening and (ii) classify whether a threatening tweet is used to threaten an individual or a group. We compare three forms of text representation: two count-based, where the text is represented using either character n-gram counts or word n-gram counts as feature vectors and the third text representation is based on fastText pre-trained word embeddings for Urdu. We perform several experiments using machine learning and deep learning classifiers and our study shows that MLP classifier with the combination of word n-gram features outperformed other classifiers in detecting threatening tweets. Whereas, SVM using fastText pre-trained word embedding obtained the best results for the target identification task.