A Study of Contrastive Learning Methods for Bengali Social Analysis

Across various domains, Contrastive Learning (CL) has already proven to be a powerful technique but using the Bengali language in the domain of Natural Language Processing (NLP) its' application is still unexplored. In this research, we introduce the implementation of CL in Bengali NLP by presenting a comprehensive benchmark study on two distinct datasets: Bengali Hate Speech, and Rokomari Book Review. The efficiency of the supervised contrastive techniques is emphasized by our methodology. We have implemented a contrastive learning technique through Bangla Bert. The superiority of supervised contrastive learning techniques over traditional Cross Entropy (CE) methods has been showcased by our result. Detailed experiments reveal performance variations across datasets, models, and hyperparameters but ensure the superiority of CL techniques over CE. Our findings show the significance of contrastive learning methods in low-resource settings, contributing to the advancement of Bengali NLP research.