The key contributions of this research are as follows:
• We construct a labeled Bangla politeness classification
dataset of 12,300 comments (three-annotator majority
voting), derived from a larger public Bangla comment
corpus.
• We conduct a systematic comparison of five deep
learning architectures and six transformer variants
for Bangla politeness classification, showing that
BanglaBERT achieves the best accuracy, precision,
recall, and F1-score (all 84%) among all models tested.
• We construct, to our knowledge, one of the first manu-
ally curated Bangla impoliteness lexicons (410 entries)
explicitly designed to bridge the gap between offensive-
language detection and paraphrase-based rewriting.
• We propose and evaluate a complete sentence-
composer pipeline classify, substitute, paraphrase that
converts impolite Bangla sentences into polite equiva-
lents while aiming to preserve their original meaning.
