Zero-shot learning for low resource code-mixed Indic Languages

Team:

Arjun Manjunatha Rao (arjunman@andrew.cmu.edu) Aprameya Bharadwaj (ambharad@andrew.cmu.edu) Naveen Suresh (naveensu@andrew.cmu.edu)

Running Instructions

1. To train the model, please run fine-tune.py. The current file will run the training sequence for Urdu-English code-mixed language. To change the language of training, please comment out lines 205-226 and uncomment lines 229-246 (for Kannada-English code-mixed training) or uncomment lines 250-267 (for Tamil-English code-mixed training). Further, by default, this will train m-BERT. To train MuRIL, please comment out line 275 and uncomment line 276.
2. To get the evaluation metrics of the trained model, please run ModelTransferZero-shot.py. This file is set to evaluate the saved m-BERT model on Urdu-English code-mixed language. To change the model/language evaluation, please comment out lines 323-335 and uncomment the required section below line 335.
3. To Run the alignment, open the awesome-align folder and run the "run.sh" file with the required paralllel dataset. The dataset is in Dataset>aligned folder. Once the model is trained on this, follow steps 1 and 2 to replicate the results.