Adventures in Machine Translation

October 7, 2009 | 2 books mentioned 1 4 min read

coverIt’s not uncommon for a website based in Russia or Italy or Venezuela to link to The Millions. Keeping up with these mentions and trying to figure out how somebody in Milan or Caracas is reacting to an essay or review of ours has made me a frequent user of Google Translate, which lets you drop in a block of text and press “translate.” In ever magical Google fashion, a passable English translation appears.

What’s interesting to me is that over the last few years the translation seems to have become more passable and it’s now easier than ever for me to glean meaning and intent from the product of Google’s machinations.

If one assumes that the improvement in quality of these translations might continue in a linear fashion, then it follows that I might be reading a machine translated book one of these days. It’s a liberating notion. I have no affinity for languages but I have often wished I could dig into to the untranslated oeuvres of favorites like Alvaro Mutis or Ryszard Kapuscinski or read their translated books in their original forms.

But then again, the idea might inspire fear that some essentially human quality of the literature would, literally, be lost in translation. And certainly for translators, who would be replaced by stacks of processors in a climate-controlled warehouse somewhere, such a development would be devastating.

Even if computers never approach the craftsmanship of Natasha Wimmer and Edith Grossman, Google or something like might get good enough at doing the heavy lifting and letting the reader clean up the language here and there.

And indeed that might be fine for some applications even today, but using Google to create a passable translation of the blog posts of a Spanish or German blogger is one thing, using it to translate a work of literature is quite another. A translated novel needs to be perfect and Google’s success in completing undemanding translation tasks was no guarantee that it would be able to manage the nuanced language of a literary master.

An experiment was in order. In the interest of seeing how close we are to this brave new world of machine translation, I decided to give a recent work of fiction, written originally in Spanish, the Google test. I chose Roberto Bolaño’s 2666 because I haven’t read it and because I was able to find the same excerpt in both English and Spanish.

I began with the Spanish:

La primera vez que Jean-Claude Pelletier leyó a Benno von Archimboldi fue en la Navidad de 1980, en París, en donde cursaba estudios universitarios de literatura alemana, a la edad de diecinueve años. El libro en cuestión era D’Arsonval. El joven Pelletier ignoraba entonces que esa novela era parte de una trilogía (compuesta por El jardín, de tema inglés, La máscara de cuero, de tema polaco, así como D’Arsonval era, evidentemente, de tema francés), pero esa ignorancia o ese vacío o esa dejadez bibliográfica, que sólo podía ser achacada a su extrema juventud, no restó un ápice del deslumbramiento y de la admiración que le produjo la novela.

I plugged that passage into Google and it spit out:

The first time Jean-Claude Pelletier to read Benno von Archimboldi was at Christmas 1980 in Paris, where a university student of German literature at the age of nineteen. The book in question was D’Arsonval. The young Pelletier knew then that this novel was part of a trilogy (consisting of the garden, full English, leather mask, Polish theme and D’Arsonval was obviously French theme), but that ignorance or the vacuum or the neglect literature, which could only be blamed on his extreme youth, it detracts from glare and admiration that led to the novel.

The result is hardly poetry, but it seemed surprisingly decipherable. There are some issues with sentence structure, and trying to figure out the antecedents of the various pronouns is difficult. So, and this wasn’t an entirely unpleasant exercise, I jumped in and attempted to clean it up myself, knowing that I might be skewing the meaning of the passage badly, but interested in at least applying a certain degree of polish:

The first time Jean-Claude Pelletier read to Benno von Archimboldi was Christmas 1980 in Paris, where he was a nineteen-year-old university student of German literature. The book in question was D’Arsonval. The young Pelletier knew then that this novel was part of a trilogy (consisting of the garden, full English, leather mask, Polish themes and, obviously, French themes as well), but beyond that his ignorance or a vacuum or neglect for literature, which could only be blamed on his extreme youth, detracted from glare and admiration that he would have for the novel.

I decided that Pelletier is the nineteen year old and that Google’s muddled translation was trying to tell me that the young Pelletier is reading to this Archimboldi and though Pelletier had some rote understanding of the book D’Arsonval, he was too immature to appreciate it as he one day would.

Then I looked at Natasha Wimmer’s translation, and I saw what Google and I got right and what we got very wrong:

The first time that Jean-Claude Pelletier read Benno von Archimboldi was Christmas 1980, in Paris, when he was nineteen years old and studying German literature. The book in question was D’Arsonval. The young Pelletier didn’t realize at the time that the novel was part of a trilogy (made up of the English-themed The Garden and the Polish-themed The Leather Mask, together with the clearly French-themed D’Arsonval), but this ignorance or lapse or bibliographical lacuna, attributable only to his extreme youth, did nothing to diminish the wonder and admiration that the novel stirred in him.

Pelletier is indeed the youth here, but he didn’t read to Archimboldi, he read a book by Archimboldi. Worse, Google and I totally misread Pelletier’s reaction to the book. We find that Pelletier, despite his youth, indeed appreciated D’Arsonval on a gut level, but did not yet appreciate its literary context, essentially the reverse of what the machine translation came up with. And Google and I totally flubbed the idea that the parenthetical list was a list of titles and not descriptors for D’Arsonval.

Natasha Wimmer, your job is safe.

Despite my failed experiment, machine translation might one day be able to figure out how to properly align those pronouns and antecedents and it might make short work of that complicated list of book titles, but would a machine ever, as Wimmer has, be able to convey the urgency in Pelletier’s literary discovery?

If a machine could one day do that, we might no longer think of it as a machine. It would have passed the Turing test (which tells us a machine has demonstrated intelligence when you can no longer distinguish a machine’s actions from those of a human). J.M. Cohen, a prolific translator whose efforts included Don Quixote and Rousseau’s The Confessions, put it another way: “It is impossible… to imagine a literary-translation machine less complex than the human brain itself.”

[Image credit: Jared Tarbell]

created The Millions and is its publisher. He and his family live in New Jersey.