Language Swaps: On the Reversibility of Translation

Launched in 2006, Google Translate quickly became the most popular automated translation system in the world and is now used by hundreds of millions of people every day. It is also the bane of language teachers, who fear that its conveniences will create an overwhelming temptation for students struggling to compose sentences in a foreign tongue. The service's computations were initially guided by its statistical modeling of parallel corpora, collections of texts, and their translations that it found by trawling billions of web pages. Transcripts of UN and EU meetings were a cornerstone of the database. Especially when translating from or into English, the results were decidedly superior to the existing rule-based computer translation systems, but the best was yet to come. In 2016, Google Translate was upgraded, becoming a neural machine translation engine. As the product lead explains: "At a high level, the neural system translates whole sentences at a time, rather than just piece by piece. It uses this broader context [the entire sentence] to help it figure out the most relevant translation. … [T]he system learns over time to create better, more natural translations" (Turovsky 2016).¹ A marked improvement in accuracy was evident even to casual users of the service, but human translators of novels or lyric poems need not fear for their livelihood. As a web search for "Google Translate fails" reveals, the engine continues to have problems with polysemy and idioms, and its command of diction is far from perfect.

If Google Translate's deficiencies are readily identifiable, there is more to be said about what it does well. One of its distinctive and arguably most helpful features is also the means by which its failings tend to be exposed, namely the "swap language" function that instantly reverses the prior operation and translates the translated text back into the original language. Our intuition is that if everything is working correctly, the translation of the translation should be identical to the original input.² It would be less than reassuring if dog became chien or Hund with one click of the mouse, only to see the next click yield greedy or beleaguer. In the same vein, "it's a nice day" should neatly become "c'est une belle journée," "es ist ein schöner Tag," or "es un lindo día," but all of these formulations should no less neatly revert to "it's a nice day."³

Is it really an indication of Google Translate's reliability if its translations of its own translations are indistinguishable from the original block of text? Would a human translator necessarily be able to pass this test, or want to, and what ideas about language are we committing ourselves to when we posit that such computations should be reversible? We certainly cannot proceed on the assumption of a perfect one-to-one correspondence between the expressive resources of different languages. As any reader of Ferdinand de Saussure, Walter Benjamin, or Roman Jakobson will be aware, the relationships between the noun dog and all of the other words, idioms, and grammatical and morphological conventions of English are not identical to the relationships between chien and the rest of the French language, leaving aside that in some contexts orthography may have crucial implications for meter, rhythm, or rhyme. If anything, we would expect the translation of a translation to bear indelible marks of the decisions that went into the two rounds of operations that produced it, decisions that will inevitably reveal the incommensurability of languages.

One might counter that the conviction that it should be possible to undo each of Google Translate's operations with a click of the swap languages button bespeaks an expectation that the system will be internally consistent more than a belief in interlinguistic symmetry. Even consistency, however, may be a misplaced value. Going from English to German, Google turns "it'sa nice day" into "es ist ein schöner Tag," an irreproachable translation to be sure, but reverse the operation and we read: "it'sa beautiful day."⁴ It would be a mistake...