Transfer Learning Based Youtube Toxic Comments Identification

Abstract

Online users are negatively affected by the spread of offensive content on social media sites. A fear, dislike, unease, or distrust of lesbian, gay, bisexual, or transgender persons is known as homophobia or transphobia. Homophobic/transphobic speech, which can be summed up as bigotry directed towards LGBT+ people, has grown to be a significant problem in recent years. The major social problem of online homopho- bia/transphobia threatens to eliminate equity, diversification, and acceptance while also making online places toxic and unwelcoming for LGBT+ people. It is found to be sensitive subject and untrained crowd sourced annotators have trouble in identifying homophobia due to cultural and other preconceptions. As a result, annotators had been educated and provided them with thorough annotation standards. 15,141 multilingual annotated comments make up the dataset. The proposed work identifies the best Machine Learning Classifier with BERT embedding model for the Code-Mixed Dravidian Languages in order to identify the toxic languages directed towards LGBTQ+ individuals. Adaboost classifier outperforms other three classifiers in terms of accuracy.