Bhasha/MT

This website consolidates resources, source-code and publications associated with research in machine translation undertaken as part of a larger group which works across the three modalities of vision, speech and language in Center for Visual Information Technology (CVIT).

Our MT systems cover 11 languages at the moment: [en, bn, gu, hi, ml, mr, pa, or, ta, te, ur].

If you use our datasets or models, please cite the relevant publication accordingly.

Code

  • jerinphilip/fairseq-ilmt: fairseq fork used to train multilingual models efficiently.
  • jerinphilip/ilmulti: Tokenizers, wrappers around fairseq for models trained on Indian languages. (Pretrained models are available; instructions in repository)
  • shashanksiripragada/pib-crawl: Crawling code for Press Information Bureau website. The database and models with a front-end to inspect is provided and you can mine better sentences if you have superior methods.
  • jerinphilip/mann-ki-baat: Crawling code for the Prime-minister's speeches.

Resources

Papers and Pre-Prints

License

The datasets and pretrained models provided here are licensed under Creative Commons Attribution-ShareAlike 4.0 International License.