A Multifaceted Evaluation of Neural versus Phrase-Based Machine Translation for 9 Language Directions

TL;DR

This paper compares neural and phrase-based machine translation across nine language pairs, revealing neural methods produce more fluent, accurate, and diverse translations but struggle with very long sentences.

Contribution

It provides a comprehensive, multi-dimensional evaluation of neural versus phrase-based translation systems across multiple languages and metrics.

Findings

Neural translation outputs are more fluent and accurate in word order.

Neural systems better handle inflected forms.

Neural translation struggles with very long sentences.

Abstract

We aim to shed light on the strengths and weaknesses of the newly introduced neural machine translation paradigm. To that end, we conduct a multifaceted evaluation in which we compare outputs produced by state-of-the-art neural machine translation and phrase-based machine translation systems for 9 language directions across a number of dimensions. Specifically, we measure the similarity of the outputs, their fluency and amount of reordering, the effect of sentence length and performance across different error categories. We find out that translations produced by neural machine translation systems are considerably different, more fluent and more accurate in terms of word order compared to those produced by phrase-based systems. Neural machine translation systems are also more accurate at producing inflected forms, but they perform poorly when translating very long sentences.