search

richardbaihe / paperreading

NLP papers
MIT License
2 stars 0 forks source link

Arxiv 2020 | Pre-training via Paraphrasing #39

Closed richardbaihe closed 4 years ago

richardbaihe commented 4 years ago
image

https://arxiv.org/pdf/2006.15020.pdf

richardbaihe commented 4 years ago

The key idea is similar to Pegasus, but is shifted to the document level. By retrieving similar docs for a target doc, the pretraining objective is to reconstruct the target doc with all retrieved docs.

Tips: The docs are multilingual.