Incremental Decoding and Training Methods for Simultaneous Translation in Neural Machine Translation

TL;DR

This paper introduces incremental decoding and training techniques for neural machine translation, enabling more efficient simultaneous translation with improved segmentation strategies and lower latency.

Contribution

It presents a tunable agent for segmentation in simultaneous translation and data-driven training modifications to better align with incremental decoding.

Findings

Agent outperforms previous methods on BLEU with lower latency

Proposed training changes improve incremental decoding performance

Demonstrates effective dynamic segmentation strategies

Abstract

We address the problem of simultaneous translation by modifying the Neural MT decoder to operate with dynamically built encoder and attention. We propose a tunable agent which decides the best segmentation strategy for a user-defined BLEU loss and Average Proportion (AP) constraint. Our agent outperforms previously proposed Wait-if-diff and Wait-if-worse agents (Cho and Esipova, 2016) on BLEU with a lower latency. Secondly we proposed data-driven changes to Neural MT training to better match the incremental decoding framework.