Language Scaling for Universal Suggested Replies Model

NAACL 2021 Industrial Track |

Published by ACL

We consider the problem of scaling automated suggested replies for Outlook email system to multiple languages. Faced with increased compute requirements and low resources for language expansion, we build a single universal
model for improving the quality and reducing run-time costs of our production system. However, restricted data movement across regional centers prevents joint training across languages. To this end, we propose a multitask continual learning framework, with auxiliary tasks and language adapters to learn universal language representation across regions. The experimental results show positive crosslingual transfer across languages while reducing catastrophic forgetting across regions. Our online results on real user traffic show significant gains in CTR and characters saved, as well as 65% training cost reduction compared with per-language models. As a consequence, we have scaled the feature in multiple languages including low-resource markets