***
But why? Poetic language in almost every language has traditionally involved prosody, figuration, rhetorical devices, rhyme and other sonic devices, allusions and symbolic registers rooted in the language and culture in which it was produced, and the overall and often intricate interplay between all of these elements, in part because it arose out of orality, for which all of these aspects of poetry are required, and while computers have been increasingly able to perform extraordinary complex intellectual tasks, including readable, often idiomatic translation of prose, poetry and poetic language entails many more potentially insurmountable hurdles. Even the idea of paraphrasing poetry, whether in translation or not, can present difficulties; what, for example, is the paraphrase--or, to put it another way, a précis or simple meaning rendered in prose--of Stéphane Mallarmé's famous poem, "Ses purs ongles très haut....," a sonnet most likely remembered for its dazzling use of the teleuton "-yx"?
Google software engineer Dmitriy Genzel and his team presented a paper at the Empirical Methods in Natural Language Processing (EMNLP) conference at MIT this past October, in which they focused on the "purely technical challenges around generating translations with fixed rhyme and meter schemes." Part of the team's debate has centered on the importance of preserving form and meter in translating poetry, and in his blog post Genzel cites Vladimir Nabokov arguments about the impossibility of maintaining such features, while approvingly noting computer scientist Douglas Hofstadter's arguments on behalf of trying to do so. As anyone who has read my many poetry translations on here or elsewhere knows, I agree wholeheartedly with Hofstadter.
Genzel continues in his Google post:
One interesting thing to consider here is their belief in a baseline fidelity in terms of the "translation's accuracy"; one thing most translators of poetry in particular recognize, following in the wake of theorists like Walter Benjamin, is that the possiblity of perfect accuracy is an impossibility, that we can never completely capture all the nuances of the source language or recapture an Ursprache in which both languages would be equal. Something is always lost and something else is gained in the process of carrying something across. For poetic language, this raises a host of questions and issues which to which some eminent scholars and translators like Lawrence Venuti, for example, or my colleague Reg Gibbons, have devoted careers, but I will just say that in the case of some poems "translation accuracy," which is to say, semantic accuracy and fidelity, may in some cases be less important that other elements of the poem, such as rhythm, feeling, figuration, and so forth.
A Statistical Machine Translation system, like Google Translate, typically performs translations by searching through a multitude of possible translations, guided by a statistical model of accuracy. However, to translate poetry, we not only considered translation accuracy, but meter and rhyming schemes as well. In our paper we describe in more detail how we altered our translation model, but in general we chose to sacrifice a little of the translation’s accuracy to get the poetic form right.
One last point about the Google poetry translator that will prove a useful tool, I imagine, for poets and others interested in digital and electronic poetries and natural language processing:
As a pleasant side-effect, the system is also able to translate anything into poetry, allowing us to specify the genre (say, limericks or haikus), or letting the system pick the one it thinks fits best.This entire blog entry could thus become a poem, and in Urdu or Chinese, with the click of a few buttons. In a year or two, that is.
