In an era where social media platform users are growing rapidly, there has been a marked increase in hateful content being generated; to combat this, automatic hate speech detection systems are a necessity. For this purpose, researchers have recently focused their efforts on developing datasets, however, the vast majority of them have been generated for the English language, with only a few available for low-resource languages such as Roman Urdu. Furthermore, what few are available have small number of samples that pertain to hateful classes and these lack variations in topics and content. Thus, deep learning models trained on such datasets perform poorly when deployed in the real world. To improve performance the option of collecting and annotating more data can be very costly and time consuming. Thus, data augmentation techniques need to be explored to exploit already available datasets to improve model generalizability. In this paper, we explore different data augmentation techniques for the improvement of hate speech detection in Roman Urdu. We evaluate these augmentation techniques on two datasets. We are able to improve performance in the primary metric of comparison (F1 and Macro F1) as well as in recall, which is impertinent for human-in-the-loop AI systems.

The task of automatic hate-speech and offensive language detection in social media content is of utmost importance due to its implications in unprejudiced society concerning race, gender, or religion. Existing research in this area, however, is mainly focused on the English language, limiting the applicability to particular demographics. Despite its prevalence, Roman Urdu (RU) lacks language resources, annotated datasets, and language models for this task. In this study, we: (1) Present a lexicon of hateful words in RU, (2) Develop an annotated dataset called RUHSOLD consisting of 10,012 tweets in RU with both coarse-grained and fine-grained labels of hate-speech and offensive language, (3) Explore the feasibility of transfer learning of five existing embedding models to RU, (4) Propose a novel deep learning architecture called CNN-gram for hate-speech and offensive language detection and compare its performance with seven current baseline approaches on RUHSOLD dataset, and (5) Train domain-specific embeddings on more than 4.7 million tweets and make them publicly available. We conclude that transfer learning is more beneficial as compared to training embedding from scratch and that the proposed model exhibits greater robustness as compared to the baselines.