2
0 Comments

GPT3 for low resource Indic languages results are impressive!

I played with GPT3 API over the weekend. Was quite surprised how effective it is for low resource languages. I tried to do Natural Language Processing tasks like translation and Named Entity Recognition (NER) on Hindi and Marathi (two of the most widely spoken languages in India).

The results were extremely surprising considering:

  • There's a small number of parallel datasets (which aren't that great)
  • Our previous attempts to do translations on above datasets yeilded really bad translations.

GPT3 with just a few datapoints was able to get to competent translations! I've put up a list of tasks that I did with it in this Twitter thread: https://twitter.com/prat0s/status/1300134930022760448?s=20

Has anyone got access and tried to play around/integrate it in their products?

on August 31, 2020