Detection of the Offensive Language in Multilingual Communication Introduction Despite the high number of studies that deal with abuse on social media platforms, the vast majority of these studies focus on the English langu- age. With the increasing number of Arabs actively using social media, there is a paucity of studies on addressing the problems specific for the Arabic language. Conclusion and Future Work We observed diversity of Arabic dialects appearing in YouTube comments, therefore the annotators were chosen to be from three different countries. In our dataset we found that non-Arabic comments are 2%, and about 60% of them are in English. Interestingly, we found also 1% Arabic comments written in non-Arabic alphabet. As a next step we plan to analyse this dataset with a range of machine learning algorithms and a variety of pre-processing methods. Figure 1 illustrates the supervised learning process we follow and plan to complete. Azalden Alakrot, Nikola S. Nikolov Department of Computer Science and Information Systems, University of Limerick E-mail: Azalden.Alakrot@ul.ie, Nikola.Nikolov@ul.ie Research Goals The goal of this study is to develop techniques based on Natural Language Processing (NLP) for detection of offensive language on a social media platform. This work aims at identifying the character- ristics of the language generated on social media by Arabs from multiple countries in the Arab region, and finding proper solutions based on machine learning techniques. A secondary goal is to build a dictionary of offensive words that can be used in both this and future studies of offensive language in online communication. Methodology • Data collection and labelling. • Data pre-processing: data cleaning, feature selection and extraction, transformation of text data into forms ready for text miming, data normalisation, tokenization, filtering and stemming. • Feature selections and vector space modelling. • Employment of supervised machine learning techniques for building predictive models. • Statistical analysis of the accuracy of the models. References • Al-garadi, M.A., Varathan, K.D. and Ravana, S.D., 2016. Cybercrime detection in online communications: The experimental case of cyberbullying detection in the Twitter network. Computers in Human Behavior, 63, pp.433- 443. • Chen, Y., Zhou, Y., Zhu, S. and Xu, H., 2012, September. Detecting offensive language in social media to protect adolescent online safety. In Privacy, Security, Risk and Trust (PASSAT), 2012 International Conference on and 2012 International Confernece on Social Computing (SocialCom) (pp. 71-80). IEEE. • Huang, Q., Singh, V.K. and Atrey, P.K., 2014, November. Cyber bullying detection using social and textual analysis. In Proceedings of the 3rd International Workshop on Socially-Aware Multimedia (pp. 3-6). ACM. • Nobata, C., Tetreault, J., Thomas, A., Mehdad, Y. and Chang, Y., 2016, April. Abusive language detection in online user content. In Proceedings of the 25th International Conference on World Wide Web (pp. 145-153). International World Wide Web Conferences Steering Committee. • Warner, W. and Hirschberg, J., 2012, June. Detecting hate speech on the world wide web. In Proceedings of the Second Workshop on Language in Social Media (pp. 19-26). Association for Computational Linguistics. Data Collection and Labelling We collected and statistically analysed a large dataset of close to 16,000 YouTube comments in Arabic. Subsequently, our dataset was labelled by three annotators which is the standard methodology used in similar studies with English text corpora, see Warner et al. (2012), Huang et al. (2014) and Al-garadi et al. (2016). The three annotators were asked to label each YouTube comment in our dataset as either offensive or non- offensive. Scenario 1: Label as offensive comments on which all annotators agree. Scenario 2: Label as offensive comments on which at least two annotators agree. Table 1. Labelling Results Figure 1. Supervised Learning Process Dataset Labelling Machine Learning Algorithms Vector space model Unknown text data Predictive Model Predicted label Vector space model Training and Test Dataset Collecting Comments on which all annotators agree Percentage of matches 10715 comments 71 % Comments unlabelled by at lest one annotator At least one annotator disagree 848 comments 4335 comments Scenario 1 Comments labelled offensive by three annotators Percentage of offensive comments 3797 comments 25 % Scenario 2 Comments labelled offensive by at least two annotators Percentage of offensive comments 5817 comments 39%