Representation-Aware Prompting for Zero-Shot Marathi Text Classification: IPA, Romanization, Repetition

AbstractLarge language models (LLMs) often underperform in zero-shot text classification for low-resource, non-Latin languages due to script and tokenization mismatches. We propose representation-aware prompting for Marathi that augments the original script with International Phonetic Alphabet (IPA) transcriptions, romanization, or a repetition-based fallback when external converters are unavailable. Experiments with two instruction-tuned LLMs on Marathi sentiment analysis and hate detection show consistent gains over script-only prompting (up to +2.6 accuracy points). We further find that the most effective augmentation is model-dependent, and that combining all variants is not consistently beneficial, suggesting that concise, targeted cues are preferable in zero-shot settings.