An Empirical Study of Generation Order for Machine Translation

TL;DR

This study empirically investigates how different generation orders affect machine translation quality, revealing that traditional left-to-right order is not always optimal and that language pair characteristics influence order sensitivity.

Contribution

Introduces a soft order-reward framework for training models with arbitrary generation orders and systematically explores their impact on translation quality.

Findings

Order has little impact on English-German translation quality.

Unconventional orders like alphabetical match standard Transformer performance.

Translation for less aligned language pairs shows greater order sensitivity.

Abstract

In this work, we present an empirical study of generation order for machine translation. Building on recent advances in insertion-based modeling, we first introduce a soft order-reward framework that enables us to train models to follow arbitrary oracle generation policies. We then make use of this framework to explore a large variety of generation orders, including uninformed orders, location-based orders, frequency-based orders, content-based orders, and model-based orders. Curiously, we find that for the WMT'14 English $\to$ German translation task, order does not have a substantial impact on output quality, with unintuitive orderings such as alphabetical and shortest-first matching the performance of a standard Transformer. This demonstrates that traditional left-to-right generation is not strictly necessary to achieve high performance. On the other hand, results on the WMT'18 English $\to$ Chinese task tend to vary more widely, suggesting that translation for less well-aligned language pairs may be more sensitive to generation order.