Political speech persuades long before it informs. BanglaRhet is the first systematic benchmark for detecting how that persuasion works in Bangla, a language spoken by over 240 million people yet still underserved by NLP research.
Built a manually annotated corpus of 30,289 Bangla political speech segments from Motamot, Prothom Alo, and Manab Zamin. Formulated two classification tasks: rhetorical technique detection (contrast, repetition, exaggeration, metaphor, rhetorical questions) and persuasion technique detection (blame assignment, call to action, unity call, moral, emotional and logical appeals). Four transformer models were benchmarked, with BanglaBERT setting the strongest baselines.
30,289
manually annotated speech segments
65.40%
macro-F1, rhetorical technique detection
66.46%
macro-F1, persuasion technique detection
4
transformer models benchmarked
Models benchmarked: BanglaBERT, Bangla-BERT-Base, SahajBERT, XLM-RoBERTa-Base.
Low-resource NLP, political discourse analysis, persuasion and rhetoric modeling, context-aware and multi-label classification.