search

thompsonb / prism

MT Evaluation in Many Languages via Zero-Shot Paraphrasing
Other
102 stars 23 forks source link

Potential extension of PRISM ? (collaboration?) #9

Closed PierreColombo closed 3 years ago

PierreColombo commented 3 years ago

Hello, Thanks for your work ! I hope you are doing well :-) . We read your papers and would like to build on your work.

I would like to better understand your tokenizer choice. Why did you choose the sentencepiece ? Another question I had, how costly is it to retrain your translation model ? Are you currently working on extensions of this work ? Cheers, Pierre

thompsonb commented 3 years ago

The main reason we use subwords is to make the vocabulary size reasonable. For the advantages of sentencepiece, I will refer you to their paper: https://aclanthology.org/D18-2012

We describe the cost of training Prism on AWS in our paper (https://aclanthology.org/2020.emnlp-main.8/) - see Appendix C.

I am not currently working on any extensions but Google had success training a larger model: https://aclanthology.org/2021.naacl-main.91/

PierreColombo commented 3 years ago

Many thanks for the response :)