Application of a transformer based MuRIL with hybrid Deep Learning Model for effective Multilingual text sentiment analysis

Abstract

Social media sites like YouTube produce multilingual user-generated content in great quantities, particularly in countries like India, where users frequently combine Hindi, English, and other regional languages in the same sentence. Traditional sentiment analysis models designed for a single language often underperform on this kind of noisy and mixed data. In this paper, the transformer-based MuRIL and hybrid deep learning architectures are combined for effective multilingual sentiment analysis. Experiments are conducted on the Hindi-English (Hinglish) and Bangla-English (Romanized Bangla) datasets. The study evaluates transformer models such as MuRIL, mBERT, and XLM-R, as well as hybrid CNN-BiLSTM models with MuRIL. Among the models evaluated, the MuRIL-CNN-BiLSTM model achieves the highest test accuracy  87.55% and 87.54% Macro-F1 score. Thus, the proposed model of MuRIL along with CNN- BiLSTM  layers displays maximum accuracy in comparison to the performance of other transformer models.