Lower Translation Rates: *ϗԇḨἭᾎ‡₩ion and the Idiocy of Averages

"...averages and extreme values (as abstractions and unrepresentative instances respectively) often provide only partial, if not downright misleading, views of a totality's behavior."

Stephen Jay Gould, Full House: The Spread of Excellence from Plato to Darwin

There seems to be a meme whereby any discussion of lower rates in the translation industry is immediately, or at least eventually, appended by the coda that, “of course, this is partly due to the increased usage of machine translation (MT).” The curious thing is that this is also accompanied by the meme that “the volume of material that needs to be translated has mushroomed exponentially.” Allegedly, we are facing a case in which demand for a service has expanded fantastically but the price per unit of that service has dropped. Surely, there can’t be any parallel for such a trend anywhere else in the history of the world economy since the first caveman traded a pelt for a sharp-edged stone with another caveman.

It is truly stupefying to note how even intelligent discussions of machine translation sooner or later crash into the lower rate/greater demand conundrum and come limping out of it with the “machine translation” crutch.

At first blush, this appears absurd. Supply and demand dictates that greater demand and stable supply breeds higher prices. If demand for translation services is rising but the price of translation a single word is going down, something odd must be happening. Our fuzzy thinker jumps to the conclusion that somehow an advance in productivity is making it cheaper to translate the average word despite the demand rise. Baboom! Machine translation comes to mind. And thus is born the questionable thesis that “greater adoption of machine translation is driving down the price per unit of a translated word.”

This way of (wrongheaded) thinking is so seductive that it is hard to disentangle the sleight of hand that our 21^st century way of thinking plays upon our own poor reptilian cortex.

Step back for a moment and think about this situation critically. Imagine for a moment a world in which all translation was done by post-editors. They charge a fraction of a penny for every word they edit or they charge by the hour. The crucial point is that the entirety of all the translation work carried out in this hypothetical world is first done by a computer and then processed by a human being who does post-editing.

In this hypothetical word, any improvement in the technology naturally feeds higher productivity. Any profit increase from this productivity improvement is swallowed by the creator or owner of the technology, albeit with perhaps a slight lag. A period of adjustment ensues, after which the post-editor drops her rate slightly to accommodate the greater productivity. Crucially, she still makes the same amount of money every day. But society benefits because a greater amount of work is done for the same amount of money. The same amount of money buys a greater amount of output. In this scenario, the average rate per word has dropped.

Return for a minute to the basic components of this model:

A.    Demand for translation/machine translation technology/population of post-editors/stable output of translation/x cost per unit

Let us tweak it slightly, by improving the quality of the technology:

B.    Demand for translation/BETTER MT technology/post-editors/GREATER output of translation/x cost per translated word DROPS

Let us tweak that slightly again by hypothesizing a concurrent increase in overall demand for translation in this world:

C.    HIGHER demand for translation/BETTER MT technology/post-editors/GREATER output/unit cost per word of translation STILL DROPS (if, and only if, the productivity boost outstrips demand growth, which is not necessarily the case)


In cases B and C, it is perfectly reasonable for the average price per word to drop. And also for the post-editor to charge less per word. Otherwise, there would be no incentive for the investor in the technology to fund the research necessary for improvement. If the post-editors brake a trend toward lower per word rates under conditions of higher productivity (say, through a worldwide union of some sort), society in general is harmed. While higher productivity is ensured, the incentive for better technology is hijacked by a hypothetically all-powerful Brotherhood of Teamsters and Post-Editors and no further improvements are likely to occur.

This is unconsciously the model we are thinking of when we say price per unit is going down because of machine translation.

The problem is we still live in a hybrid world of both post-editors and translators, or, to put it another way, translators who process MT output and translators who don’t. As we proceed, let us take as a given the premise that the price per unit of non-post-editing (i.e., human translation) is going down.

I hope that by now it should be clear to the careful reader that a massive cognitive conflation is occurring here. Leaving aside borderline cases and grey areas, post-editing and translation still remain two distinct activities. To say that lower cost of post-editing (since that is the proper name for the activity, not machine translation) is driving down the cost of translation is tantamount to claiming that the invention and massive adoption of frozen dinners in the United States of the 1960s drove down the average salaries of chefs in Paris during the same decade.

We’re not talking apples and oranges here. We’re talking apples and moon pebbles.

What machine translation might be doing is driving down the average cost of a word that is translated (or, perhaps more accurately, machine translated).

But “translated” here has to be rendered here with an asterisk. Perhaps also a skull and crossbones, but my word processor doesn’t have that emoticon… yet! (as machine translation enthusiasts like to exclaim so mischievously).

So let me restate the thesis thusly:

Machine translation might be driving down the average cost of a word that is *ϗԇḨἭᾎ‡₩ed.

Wherein “*ϗԇḨἭᾎ‡₩ion” refers to the activity of creating sequences of words in one language that are conceptually and stylistically equivalent to other sequences of words in another language. This is done either through the processing by so-called post-editors of linguistic output produced by a computer or via the more traditional process of “translation” (now often referred to as "human translation"). While the two methods of *ϗԇḨἭᾎ‡₩ion bear some distant family resemblance, they in fact belong to two widely divergent genera.

Without this neologism (the “‡” is silent, by the way), the idea that machine translation can have any sort of impact on "human" translation is quite frankly questionable. Although the cost of servicing a mortgage in many countries has gone down drastically and some commodities such as oil have plummeted from their dizzying heights of 2008, the cost of translating texts (essentially equal to feeding and clothing one human translator and his or her family plus a little extra cash on the side for incidentals) is essentially the same as in 2008. Ditto for 2005 and 2000. Perhaps even 1885 after adjusting for inflation.

So are translation costs really plummeting? Is this really what is happening?

Completely unnecessary picture of Sofía Vergara

As Sofia Vergara would say: “I don’t think so, papi.”

Or is it perhaps something else? Something perhaps more complicated… or perhaps simpler?

Miguel Llorens is a freelance financial translator based in Madrid who works from Spanish into English. He is specialized in equity research, economics, accounting, and investment strategy. He has worked as a translator for Goldman Sachs, the US Government's Open Source Center and H.B.O. International, as well as many small-and-medium-sized brokerages and asset management companies operating in Spain. To contact him, visit his website and write to the address listed there. Feel free to join his LinkedIn network or to follow him on Twitter.