Translating Translationese: A Two-Step Approach to Unsupervised Machine Translation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3057–3062 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3057 Translating Translationese: A Two-Step Approach to Unsupervised Machine Translation Nima Pourdamghani♣ Nada Aldarrab♠ Marjan Ghazvininejad♦ Kevin Knight♥ Jonathan May♠ ♣ Amazon ♠ USC Information Sciences Institute ♦ Facebook AI Research ♥ DiDi Labs nimpourd@amazon.com aldarrab@isi.edu ghazvini@fb.com kevinknight@didiglobal.com jonmay@isi.edu Abstract Given a rough, word-by-word gloss of a source language sentence, target language natives can uncover the latent, fully-fluent rendering of the translation. In this work we explore this in- tuition by breaking translation into a two step process: generating a rough gloss by means of a dictionary and then ‘translating’ the resulting pseudo-translation, or ‘Translationese’ into a fully fluent translation. We build our Trans- lationese decoder once from a mish-mash of parallel data that has the target language in common and then can build dictionaries on de- mand using unsupervised techniques, resulting in rapidly generated unsupervised neural MT systems for many source languages. We ap- ply this process to 14 test languages, obtain- ing better or comparable translation results on high-resource languages than previously pub- lished unsupervised MT studies, and obtain- ing good quality results for low-resource lan- guages that have never been used in an unsu- pervised MT scenario. 1 Introduction Quality of machine translation, especially neural MT, highly depends on the amount of available parallel data. For a handful of languages, where parallel data is abundant, MT quality has reached quite good performance (Wu et al., 2016; Hassan et al., 2018). However, the quality of translation rapidly deteriorates as the amount of parallel data decreases (Koehn and Knowles, 2017). Unfortu- nately, many languages have close to zero parallel texts. Translating texts from these languages re- quires new techniques. Hermjakob et al. (2018) presented a hybrid human/machine translation tool that uses lexical translation tables to gloss a translation and relies on human language and world models to propa- gate glosses into fluent translations. Inspired by that work, this work investigates the following question: Can we replace the human in the loop with more technology? We provide the following two-step solution to unsupervised neural machine translation: 1. Use a bilingual dictionary to gloss the input into a pseudo-translation or ‘Translationese’. 2. Translate the Translationese into target lan- guage, using a model built in advance from various parallel data, with the source side converted into Translationese using Step 1. The notion of separating adequacy from fluency components into a pipeline of operations dates back to the early days of MT and NLP research, where the inadequacy of word-by-word MT was first observed (Yngve, 1955; Oswald, 1952). A subfield of MT research that seeks to improve flu- ency given disfluent but adequate first-pass trans- lation is automatic post-editing (APE) pioneered by Knight and Chander (1994). Much of the cur- rent APE work targets correction of black-box MT systems, which are presumed to be supervised. Early approaches to unsupervised machine translation include decipherment methods (Nuhn et …

Content truncated to 500 words