Goal:
Translating Udru poetry to Shakespearean English
Methodology:
- Trained existing Urdu to English Helsinki model on Bible and Quran Urdu-English parallel to translate Urdu to Modern English (Pivot language).
- After zero-shot translation of Urdu poetry to Modern English, fine-tuned Shakespeare GPT-2 model to get output in Shakespearean English.
Resources:
Used University of Michigan, Ann Abor computing cloud GPU service to train the models. The trained models are here- https://drive.google.com/drive/folders/1Nd9KFbgLXVmNgcG5tUmOVOPvLGarNo1V?usp=sharing
Replicate:
To try other Urdu poems or text just download the two trained models from the Google Drive and the notebook on your machine and hit run.
Scope of Improvement:
- The Urdu-to-English translation model was trained on the Bible and Quran Parallel corpus, so the poetic translation can be improved by using a parallel corpus of Urdu poetry to Modern English or vice-versa.
- Due to memory constraints on the machine we could not experiment with models like mBert50 but anticipate the BLEU score will be higher for it.