Faster constrained decoding and better sentence rewriting

Constrained decoding, which we used in ParaBank, is not quite fast enough to be useful. We came up with an algorithmic improvement to boost its performance by 5-fold, and fixed a couple edge cases along the way. Armed with this improved framework, we showed that simply rewriting existing dataset and using it for data augmentation could give non-trivial improvement to an ELMo baseline.

The paper is titled “Improved Lexically Constrained Decoding for Translation and Monolingual Rewriting,” which I presented at NAACL 2019 as a poster.

Check it out on ACL Anthology: or here.