Padyam2Gadyam: Translating classical Telugu poetry with LLMs shows big gaps
600 ancient Telugu poems now have modern prose translations—but AI still struggles.
Researchers Kranti Chalamalasetti and Sowmya Vajjala present Padyam2Gadyam, a novel dataset for translating classical Telugu poetry (13th–17th century) into modern prose. The dataset contains 600 poems, each paired with human-verified translations into both contemporary Telugu and English. This bridges a gap in low-resource language translation and cultural preservation.
They evaluated five contemporary LLMs on the poem-to-prose task. Performance varied by model but overall left much to be desired in both Telugu and English outputs. Their qualitative analysis reveals that current MT evaluation metrics struggle to capture poetic nuances like meter, metaphor, and cultural context—underscoring a pressing need for better approaches in literary translation.
- Dataset Padyam2Gadyam includes 600 Telugu poems from 13th–17th century with verified prose translations.
- 5 LLMs tested; all show significant room for improvement in both Telugu and English.
- Current MT evaluation methods fail to assess poetic qualities like meter and metaphor.
Why It Matters
Preserving ancient literature via AI translation is still nascent—LLMs can't yet handle poetic nuance reliably.