Skip to content

Latest commit

 

History

History
21 lines (17 loc) · 886 Bytes

WT5.md

File metadata and controls

21 lines (17 loc) · 886 Bytes

"WT5?! Training Text-to-Text Models to Explain their Predictions."

Narang, Sharan, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan.

arXiv preprint arXiv:2004.14546 (2020).[arXiv]

Whats Unique Authors present the extension of T5 model to generate explnations as well. And, also demonstrate how it is transferable across datasets, and to some extent across tasks.

How It Works

Source: Author

  • During training WT5 model, (acronym for Why? T5), "explain" is prepend

  • Explnations are generated followed by prediction, and after the token "explanation:"

  • Following figure shows the result fo WT5 model:

Source: Author