Exploring Deep Learning for Chittagonian Slang Detection in Social Media Texts
Sultana Rokeya Naher, Sayfun Sultana, Tanjim Mahmud, Mohammad Tarek Aziz, Mohammad Shahadat Hossain, Karl Andersson
Abstract
Detecting and understanding the meaning of sentences has emerged as a crucial area of research in natural language processing, driven by the prevalence of social media platforms where users freely express their thoughts and emotions. Unfortunately, the anonymity provided by these platforms has led to the proliferation of slang, vulgar, and abusive language, contributing to negative social and community impacts. While extensive research has been conducted in English, relatively few studies focus on low-resource languages. To address this gap, we introduce a dataset in the Chittagonian language, primarily spoken in Chittagong, Bangladesh. Our dataset comprises 2,100 comments, evenly distributed between slang and non-slang expressions. Leveraging machine learning and deep learning classifiers such as SVM, RF, LR, DT, NB, simple RNN, LSTM, CNN, GRU, and Bi-directional LSTM, coupled with feature extraction techniques like TF -IDF Vectorizer and Word2vec, we identify and classify slang comments on social media. Our results show that LSTM or CNN with Word2vec achieves the highest accuracy of 76%, demonstrating promising prospects for detecting slang in low-resource languages.