Bangla Radical Text Categorization Using Lightweight Convolutional-LSTM Framework

In the era of the fourth industrial revolution, with the advent of internet technology, human life is becoming more and more dependent on online media for social interaction and making important decisions in their lives. Opinions expressed on social media can either be positive or negative which can trigger radicalization in society. This paper mainly focuses on detecting radical text in Bangla from the opinions of people conveyed on social media. Radical Long Short Term Memory (RAD-LSTM) Network, a novel technique, is proposed in this paper to solve the issue. Our technique consists of a one-dimensional convolutional layer of 64 units and only 10 units of LSTM cells. We carried out our experiments using Bangla-Aggressive-Text, a publicly available Bangla dataset. We compared the performance of our architecture with many state-of-the-art techniques and found satisfactory performance in detecting radical speech in Bangla. Despite having the lightweight characteristics, an adequate performance provided by our proposed technique by acquiring 93% accuracy.