A Hybrid Deep Learning Approach for Dialect Specific Sentiment Analysis of Bangla Regional Linguistics

Bangladesh is enriched with distinct regional languages. Almost ten dialects have been classified in Bangladesh, each with its own unique phonological, grammatical, and lexical features. Despite substantial breakthroughs in sentiment analysis for mainstream Bangla, the investigation of regional dialects remains limited, offering hurdles in communication and sentiment identification. This study tries to overcome these issues by focusing on dialect detection and sentiment analysis of three Bangla regional dialects: Chittagong, Noakhali, and Sylhet. Using a dialect dataset, which comprises sentences from these languages, the dialect detection work was accomplished using the Multinomial Naive Bayes model with the highest F1-score of 0.95 for the Chittagong-Bangla dialect. Machine learning models such as Logistic Regression and Multinomial Naive Bayes are used for sentiment analysis, and deep learning approaches such as a Hybrid LSTM & CNN model and Bangla-base-bert. The hybrid LSTM & CNN model attained the highest accuracy of 94.69% for sentimental analysis of the Chittagong-Bangla dialect, outperforming previous models and existing studies. This work will help and improve natural language processing (NLP) technologies for improved communication and sentiment recognition. Our findings have larger implications for social media surveillance and customer service centers, where people of different language backgrounds communicate.