WSD-Based Bangla Cyberbullying Detection Using Transform Learning

This research inquires into the core of Bengali online bullying detection, employing Word Sense Disambiguation (WSD) in conjunction with Transfer Learning. A dataset of 5000 Bengali comments, each annotated with corresponding labels, was meticulously curated. Ambiguity removal algorithms have been developed and used in this dataset. Machine learning algorithms, utilizing TF-IDF features for unigrams, bigrams, and trigrams, were compared with deep learning algorithms leveraging word embeddings. Results revealed an 81.9% accuracy with Multinomial Naive Bayes (MNB) for machine learning, and a striking 96% accuracy with Convolutional Neural Network (CNN) for deep learning.