AI Struggles to Translate Ancient Poetry to Modern Prose

Researchers at an unnamed institution introduced Padyam2Gadyam, a new dataset of 600 Telugu poems from the 13th to 17th centuries, along with human-created contemporary Telugu and English prose translations. The team evaluated five large language models on their ability to translate the classical poetry into modern prose in both languages.

Results reveal that while the models vary in performance, none currently excel at accurately capturing the meaning and nuance of the original poems. Qualitative analysis suggests current machine translation evaluation methods are also inadequate for assessing this type of complex translation. The work highlights the challenges of adapting AI to handle highly stylized and culturally specific texts, and further research is needed to improve performance.