Quantifying Linguistic Variation in Bangla through Dialect-to-Dialect Translation Metrics

This research addresses Bangla dialect-to-dialect translation, a largely unexplored area compared to standard language translation. We focus on five major Bangla dialects — Barishal, Chittagong, Mymensingh, Noakhali, and Sylhet — and propose a retrieval-based approach optimized for low-resource settings. Our method uses Term Frequency-Inverse Document Frequency (TF-IDF) vectorization and cosine similarity for sentence-level translation. We evaluate two models: word-level TF-IDF and character-level TF-IDF using 2–5-character n-grams. The character-level model achieves superior results, with a maximum BLEU score of 0.313 and ChrF score of 0.362 for the Mymensingh to Barishal pair, compared to 0.193 BLEU and 0.323 ChrF with word-level TF-IDF. Across all 20 dialect pairs, character-level modeling consistently outperforms, boosting BLEU by around 62%. Our evaluation, including BLEU, ChrF, and heatmap analyses, confirms that character-based features effectively capture dialectal variation. The proposed system offers a practical, lightweight baseline for real-time Bangla dialect translation in low-resource environments.