SAAD: Sentiment Analysis and Abuse Detection for Arabizi Language on Social Media

Abstract

Bullying and abusive behaviors are becoming more common as social media becomes more widely used. Users on social media are subjected to violence on a daily basis, which has long-term psychological consequences on people, resulting in a global crisis across communities. Abusive language detection systems have been widely investigated on social media. However, they are limited to official and formal languages. The Arabic Chat Alphabet (also known as Arabizi) was developed in response to the growth of the internet and social media to facilitate communication in Latin-based characters. Arabizi is an informal language representation, it is the method of transliterating Arabic with Latin letters and numerals e.g. "mar7aba, kifak?" means "Hello, how are you?". This language is now being popular in text chatting and on social media platforms to Latinize communication between Arabic-speaking people. So far, no models proposed to automatically detect abuse in Arabizi. In this paper, we propose a framework that uses supervised learning to assign sentiment labels to comments obtained from various social media sites and then classifies comments using a polarity classification score to abuse text. Our proposed approach has shown an accuracy of 80% for the K-nearest neighbor model.