MS-UEdin Submission to the WMT2018 APE Shared Task: Dual-Source Transformer for Automatic Post-Editing

TL;DR

This paper presents a dual-source Transformer model for automatic post-editing that achieved state-of-the-art results in the WMT2018 shared task, highlighting the potential limitations of neural-on-neural approaches.

Contribution

Introduces a dual-source Transformer architecture with data selection techniques for APE, achieving new state-of-the-art results in SMT post-editing.

Findings

Achieved state-of-the-art in SMT post-editing.

Close second in NMT sub-task, indicating limited benefit.

Neural-on-neural APE may have limited usefulness.

Abstract

This paper describes the Microsoft and University of Edinburgh submission to the Automatic Post-editing shared task at WMT2018. Based on training data and systems from the WMT2017 shared task, we re-implement our own models from the last shared task and introduce improvements based on extensive parameter sharing. Next we experiment with our implementation of dual-source transformer models and data selection for the IT domain. Our submissions decisively wins the SMT post-editing sub-task establishing the new state-of-the-art and is a very close second (or equal, 16.46 vs 16.50 TER) in the NMT sub-task. Based on the rather weak results in the NMT sub-task, we hypothesize that neural-on-neural APE might not be actually useful.