Offensive Language Detection from Arabic Texts

Detecting offensive language (OL) in social media platforms has become an important task for researchers in natural language processing and understanding. Despite emerging research and efforts to address this problem in many languages, more efforts are still needed to improve the performance of OL detection in Arabic-language contexts. This work investigates the state of the art for both the English and Arabic languages, studies the research communities’ different proposed approaches, and compares their performances. We present new approaches to the use of word-embedding models, where each word has two representations: the first representation describes the target word’s context in offensive texts, while the second represents the context of the word in non-offensive texts. The primary results are promising, with a precision reaching an average of 65% of the detection of OL and around 71% for the identification of non-offensive language.