Bengali Abusive Speech Classification: A Transfer Learning Approach Using VGG-16

Abstract

Swear words used in speech to abuse someone is frowned upon in every society. Abusive speeches can destroy the victim's morale, mental strength, and the will to live. Abusing others through social media, video streaming sites, and over voice calls are becoming a common problem. There are laws to punish the offenders. However, without proper surveillance, stopping abusive speech is tough. Machine learning can help to create surveillance methods by detecting abusive speech from human conversation. There have been a few works in the relevant field to detect abusive speech. However, detecting abusive speech in the Bengali language remains an unexplored area. This paper aims at providing an approach towards the classification of abusive and non-abusive Bengali speech. The authors collected 960 voice recordings of native Bengali speakers. The authors used Transfer Learning for extracting features from the data. Then, the authors used different methods for classification. The proposed approach achieves high accuracy (98.61%) in classifying abusive and non-abusive Bengali speech.