Abstract

Generating natural language requires conveying content in an appropriatestyle. We explore two related tasks on generating text of varying formality:monolingual formality transfer and formality-sensitive machine translation. Wepropose to solve these tasks jointly using multi-task learning, and show thatour models achieve state-of-the-art performance for formality transfer and areable to perform formality-sensitive translation without being explicitlytrained on style-annotated translation examples.

 

Quick Read (beta)