Friday, March 12, 2010


After reading Miguel Helft's article called "Google's Computing Power Refines Translation Tool" which appeared on the front age of the New York Times (March 9, 2010), I decided that an article on his article and on the tool itself seemed to be in order. For those of you who are unfamiliar with the author, Miguel Helft covers Internet companies including Google and Yahoo for the Business Desk of the New York Times. More than 150 readers, including some translators, posted comments on the website and at least one translation agency sent a letter to the editor criticizing the journalist for not having consulted a single human translator.

As you might guess from the title, the article is about Google's use of human translations which are aligned to source-language texts in much the same way as translation memory software aligns two texts, "creating," what Dr. Te Taka Keegan, in The Official Google Blog, calls, "a virtuous cycle that benefits both human translators and machine translation". Now, I have absolutely no idea what he means by a "virtuous cycle". Perhaps he meant to say "virtual" in the sense that it was "simulated, or carried on by means of a computer network". Or, perhaps Dr. Keegan originally wrote in a language other than English and the blog was translated by Google Translate!

It goes without saying that there is definitely a place for this tool, just as there is a place for translation memory software, but I was appalled at the widespread use of Google Translate on jobs that really require a human translator who can use analysis, synthesis, association, memory, logic and, yes, imagination to recreate the text in the target language. This is especially true when dealing with documents that are not particularly well written and which contain a great deal of technical language, indeed, the bulk of the industrial translator's work. It is the human translator's knowledge of the subject area that enables him/her to understand the meaning of the source-language text and to accurately interpret it into the target language.

Interestingly, there is a sidebar to the article which features a passage from Le petit prince [The Little Prince] by Antoine de Saint-Exupéry and a passage from Cien años de soledad [One Hundred Years of Solitude] by Gabriel García Marquez. Both appear first in the original language followed by a human translation, then a Google Translate translation and one by each of two competitors. I thought it odd that only the names of the publishers of the human translations were given rather than the names of the human translators. This is an egregious omission on the part of the journalist.

In spite of the fact that many consider literary translation to be outside the domain of both machine translation and translation memory software, works of fiction were chosen to "put the tools to the test". When I first looked at the two passages, my initial thought was that both the original passages are well written in straightforward language, which makes a difference in any machine translation. Then, like a bolt out of the blue, it occurred to me that there are already human translations of those texts online and that that is exactly how Google works--by matching human-translated text with similar segments in the source language. Taking that into consideration, the fact that Google Translate's translation resembles the human translation so closely is no surprise at all.

Anyway, I decided that I would see what Google Translate could do with a couple of passages from real-life jobs. I chose one that has some of the characteristics of a literary work. The source-language text is first, followed by Google Translate's version:

Además del manejo de los lápices consta que hizo sus pinitos con los pinceles y se veía capacitado para emplear términos como "claroscuro", "efecto", "menudo", juzgar el mérito de una artista o llamar "bestia" nada menos que a Francisco Bayeu por su cuadro para el altar mayor de la iglesia de San Francisco el Grande, del cual conserva un boceto el X Museum, y a su arquitecto Juan de Villanueva por parecerle "bien".

Besides handling the pencils know that tried his hand with the brush and looked able to use terms such as "chiaroscuro", "effect", "often", judging the merit of an artist or call "beast"nothing less than Francisco Bayeu for his painting for the high altar of the church of San Francisco el Grande, which keeps a sketch of the X Museum, and his architect Juan de Villanueva because it seemed "fine."

So, I leave it to you to decide for yourself whether or not you agree with Alon Lavie, an associate research profesor in the Language Technologies Institute at Carnegie Mellon University when he says: "What you see on Google Translate is state of the art in computer translations that are not limited to a particular subject area".

Monday, March 1, 2010


I recently received an email from my friend and colleague Bernie Bierman about an internet posting made by a translation agency asking "experienced Portuguese Translators and Proofreaders" to submit their rates for "no match, repetitions, and 75-99% match". He also reminded me of then-ATA president Jiri Stejskal's article "On Statistics and Competition" in the February 2008 ATA Chronicle and suggested that I take another look at it.

Basically, Mr. Stejskal's article is what we call a "feel-good" piece, as are most Chronicle articles, telling translators that they need not fear competition from abroad, or for that matter, any competition at all. The Stejskal article states that "inexpensive translation from developing countries can be viewed as a threat...only insofar as translation is perceived as a commodity that can be produced regardless of location and supplied without qualitative differentiation across a given market." (my emphasis) He goes on to tell us that to "make sure translation is not traded or perceived as a commodity...we need to specialize in order to differentiate our translation or interpreting work qualitatively."

At the time, I wrote a commentary in the Gotham Translator about the absurdity of thinking of specialization as being a way to prevent the commoditization of translation, but I was thinking about outsourcing then. I was not thinking about the role that translation memory software would play, indeed, has played in the commoditization of translation. So, what exactly does it mean to "commoditize" something? Investopedia defines it as follows:

"The act of making a process, good or service easy to obtain by making it as uniform, plentiful and affordable as possible. Something becomes commoditized when one offering is nearly indistinguishable from another. As a result of technological innovation, broad-based education and frequent iteration, goods and services become commoditized and, therefore, widely accessible." (my emphasis)

That definition seems to be a fairly accurate description of what translation software does, and I think that we could go so far as to say that, more than outsourcing, the imposition of translation memory software on translators (with Trados appearing to be the most required software product for more and more jobs) is leading inexorably to the commoditization of translation. This trend is accelerated by the fact that translators have been seduced by sales pitches such as you will never have to translate the same sentence again". What is more troubling and worrisome is that this translation memory software has been used not just as a tool of efficiency, but as a kind of combined rotary saw and sledgehammer on translation rates and, therefore, translation earnings. The brutal fact of the matter is that today's independent or freelance translator who must adhere to the non-negotiable requirement that he or she use a CAT tool, quickly finds out that there is absolutely no payment or a pittance of a payment for the words in that sentence that are viewed by the program as "repeats" or "quasi-repeats" or "semi-quasi-repeats", say nothing of the fact that perhaps the "same" sentence actually should be translated differently in a different context. Although translation memories can, under specific circumstances, provide greater consistency, the consistency just may be all wrong. Inaccurate translations become part of translation memories and, as such, are perpetuated ad infinitum. It indeed seems to me that if we use Investopedia's definition, industrial (as opposed to literary) translation is well on its way to being commoditized.

To support his proposition that "specialization" will somehow prevent translators from becoming victims of the commoditization process, Stejskal gives former ATA president Marian S. Greenfield as an example of a successful translator who specializes in financial translation. In reality, she may be an excellent example of a translator who has successfully capitalized on the commoditization of the translation industry. Earlier this year many translators received an email from SDL-Trados which contained the following testimonial:

"I just completed [sic] a 34,501 word project in 10 hours thanks to AutoSuggest, Context Match and the other nifty time-saving features within SDL Trados Studio 2009 SP1. That's without having much of anything in the pre-existing TM!" --Marian S. Greenfield, Translator and Trainer.

When other translators expressed some doubt as to whether what she had done could actually be called "translation", it was revealed that she was working with an Excel file with many repetitions--in essence, the sort of document that translation memory software handles quite well. Nevertheless, thinking about this "feat" from a financial point of view, 34,501 words in ten hours at $0.10 per word (which, believe it or not, was on the low side a scant ten years ago) would indeed make her a financially successful translator! However, one might ask how much she actually was paid for all those repetitions, if she was paid at all for them. Also, one might wonder about the "qualitative differentiation across a given market" that Stejskal says is so important. It looks a little like Ms. Greenfield's "specialization" may have been the promoting of Trados software.

In closing, I do not want to leave readers with the impression that I am against translation memory software, for this is definitely not the case. It certainly has its place in today's technological scheme of things, for in the case where a translator is asked to translate a second document (a shop manual, for instance) from the same end client, the TM software might be quite useful. I do have doubts about its being used for all types of documents as some companies are now doing. Under those circumstances, no amount of "specialization" can prevent the commoditization of translation which, in turn, is bound to change the face of the translation industry itself, if it hasn't changed it already. But setting aside the discussion as to when or where translation memory software is or is not useful, the bigger question remains whether TM has become the instrument by which to drive down translator income to the point where translation practitioners become nothing more than extension of the clerical furniture, translating a "new" word or phrase here and there, and receiving a token payment for what is, in effect (at least in the eyes of the translation buyer), a token effort. These are things that all translators should be thinking about and discussing if we are to be prepared for the challenges of the 21st century.