AlexU-BackTranslation-TL at SemEval-2020 Task 12: Improving Offensive Language Detection Using Data Augmentation and Transfer Learning

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1881–1890 Barcelona, Spain (Online), December 12, 2020. 1881 AlexU-BackTranslation-TL at SemEval-2020 Task 12: Improving Offensive Language Detection using Data Augmentation and Transfer Learning Mai Ibrahim, Marwan Torki and Nagwa El-Makky Computer and Systems Engineering Department Alexandria University Alexandria, Egypt eng-mai.ibrahim, mtorki, nagwa.elmakky@alexu.edu.eg Abstract Social media platforms, online news commenting spaces, and many other public forums have become widely known for issues of abusive behavior such as cyber-bullying and personal attacks. In this paper, we use the annotated tweets of Offensive Language Identification Dataset (OLID) to train three levels of deep learning classifiers to solve the three sub-tasks associated with the dataset. Sub-task A is to determine if the tweet is toxic or not. Then, for offensive tweets, sub-task B requires determining whether the toxicity is targeted. Finally, for sub-task C, we predict the target of the offense; i.e. a group, individual or other entity. In our solution, we tackle the problem of class imbalance in the dataset by using back translation for data augmentation and utilizing fine-tuned BERT model in an ensemble of deep learning classifiers. We used this solution to participate in the three English sub-tasks of SemEval-2020 task 12. The proposed solution achieved 0.91393, 0.6300 and 0.57607 macro F1-average in sub-tasks A, B and C respectively. We achieved the 8th, 14th and 21st places for sub-tasks A, B and C respectively. 1 Introduction Nowadays, public online communities such as blogs, forums and social networks have become an integral part of many people’s lives. These platforms enable their users to express their opinions and discuss things they care about as well as immediately react to and comment on other users posts and stories. Knowing the importance of online commenting, many online news providers nowadays have also established commenting services that enable their users to exchange their thoughts and opinions regarding the published news (Cho and Acquisti, 2013). Despite all its benefits, unfortunately without proper moderation, these platforms can easily be used for online abuse and harassment that can have serious effects on its victims. Therefore offensive content detection has become a great concern for online communities and social media platforms. The online offensive content can vary in different aspects, such as the toxicity type, the target and whether the abuse is implicit or explicit. SemEval-2020 task 12 (Zampieri et al., 2020) addresses the problem of offensive language detection in social media at three levels that map to three sub-tasks. The task dataset consists of tweets labeled for offensive content using three-level hierarchical annotation scheme. Each tweet in the dataset has three labels. The first label shows if the tweet is offensive or not (sub-task A). For offensive tweets, the second label divides them to targeted and untargeted (just profane) ones (sub-task B). Finally, the third label categorizes the targeted offensive tweets based on their target that can be an individual, a group of people or other entity or organization (sub-task C). The main challenge in this task is that …

Content truncated to 500 words