Combating Online Harassment

Abstract

Online harassment such as trolling, threatening, and bullying becomes a serious issue regarding our country's present situation. Millions of people from this country are connected through the internet. Regarding some major social platforms like Facebook, Youtube and tiktlak, abusive content creates enormous public issues including fear. Numerous research has been conducted considering this issues in different languages. however, there were very few research has published on abusive Bangla text detection. In our experiment, we tried to extract abusive comments from Bangla text using several classification algorithms such as (Logistic Regression, Multinomial Naive Bayes,Decision Tree, Support vector Machine ,Random Forest,KNN). The data, we used in this research, was collected from different online social media, forum, Bangla slang books and direct speech. The data set consist 2999 sentences in which 1710 sentences are abusive and the rest of the data are non abusive. In the pre processing part, we have categorized the data into two polarities abusive noted as 1 and non abusive as 0. The data was labeled manually. All the special characters and symbols was removed from the raw data. TF-IDF was used to extract feature from the data. After applying all the algorithms,the experiment has shown that SVM and LR both performed well. they both attained 84% accuracy in which SVM achieved 0.84 precision along with 0.91 recall.