Surfaces and Essences: Analogy as the Fuel and Fire of Thinking (90 page)

But now let us fast-forward to the spring of the year 2009. At that point, Google’s translation-engine developers had radically switched strategies in favor of the new idea of statistical machine translation, so their new engine had little in common, other than its name, with its former incarnation. Given the inadequacy of the old method, which we have just witnessed, this would seem like a wise decision. To cut to the chase, here is the output text that the totally revamped translation engine yielded:

This is English, in some sense, but it is still English that makes no sense. (What on earth is
backhand casino
, for instance? And why does “the proper number” not necessarily
spell
? And why is, or was, it “nice” to think very hard? And how happy can
the sort of taunt
get, anyway?) In sum, there is little visible improvement here over what was produced in 2004.

To round out this section, we give one last translation of the paragraph in the Sagan obituary in
Le Monde.
This human-produced translation comes from our joint pen, and in fact it is what we would have offered had we been requested to anglicize Poirot-Delpech’s obituary of Sagan for publication in an American newspaper.

On reading this hand-done translation along with the two machine translations given earlier, one sees that humans and translation engines are not playing on the same turf — in fact, they are not even playing at the same game. Any decent human translator has a rich storehouse of first-person and vicarious experiences in all areas of life. Human translators can imagine in detail a scene that was sketched with just a few brushstrokes and can pick up on subtle allusions, they have an excellent mastery of the grammar of both languages, and they are past masters at expressing themselves fluently and idiomatically. Just like the original passage, this translation paints a vivid picture of the complicated strategems a certain human being had for dealing with life’s setbacks.

How could one possibly translate without making use of one’s knowledge? For example, in the very limited context of weather reports, it will in general be a safe strategy to translate “ciel” as “sky”, but in the wide world of human affairs, the different possible meanings of “ciel” include (but certainly are not limited to) the sky, heaven, the heavens, the air, mid-air, climes, blueness, a vault, a canopy, a ceiling, the atmosphere, clouds, the firmament, the stars, space, the cosmos, the universe, Providence, God, divine will, destiny, predestination, and on and on. If we know for sure that we are dealing only with meteorological phenomena in the most mundane of contexts, then blindly replacing “ciel” by “sky” will probably do (although even in a weather forecast, someone could always throw a curve ball), but in a completely open-textured domain where any possible idea might come up or be evoked in some standard or brand-new metaphorical manner, then all bets are off. In our translation of the paragraph taken from the Sagan obituary, we happened to opt for “divine fate”, but in some other context we might have chosen “heavens to Betsy”, and in yet other contexts perhaps “the orange glow” or “the stars above” or “the celestial vault” or “light blue” or just plain old “azure”. And the word “ciel” is by no means exceptional; it is quite typical.

We’ve argued, in this book, that choices of words and longer lexical items are carried out by an unconscious process of analogy, and this holds in translation between two different languages no less than in monolingual speech production. But the analogies involved in carrying out human translation are not of the sort exploited by current machine-translation techniques, which make surface-level analogies between linguistic “chunks” based on statistical information that can be extracted, using intense calculation, from human-translated bilingual data bases, as we described above, rather than on understanding the
ideas
that are being talked about, which of course are the entire reason that pieces of text are created.

In a human translator’s mind, the evocation of words and phrases in the target language will take place much as when they summon up words and phrases of their native language in a conversation without any kind of input from a foreign language. This evocation process consists of analogy-making once again — the type of analogy-making that gives rise to the most mundane acts of labeling, the type of analogy-making that zeroes in on the
mot juste
or the
locution juste
or the
proverbe juste
, the type of analogy-making that makes someone say “probab-
lee
!” as opposed to just “
prob
-ably”, the type of analogy-making that makes someone call a certain room a “study” or an “office”, based on their prior life experiences, the type of analogy-making that makes Cheryl repeatedly call her husband “Chuck”, which is her brother’s name, the type of analogy-making that makes phrases like “sour grapes” and “the left hand doesn’t know what the right hand is doing” pop instantly to mind in certain types of circumstances, and so forth and so on.

In the specific case of the obituary paragraph, when we humans read “les revers de casino”, we do not think of
backhand strokes
in games with rackets, but of
setbacks in gambling
, because we are familiar with the kinds of things that go on in casinos. When we humans read “les caprices de la banque”, knowing that the context is one of casinos, we don’t think about
financial institutions
where we save our money but about the casino’s own
storehouse of money
, because we understand what gambling is about. And when we humans read about “les caprices du ciel”, we don’t think of the
whims of the atmosphere
, nor about whims of the
starry sky
, nor of the
color blue
; we think of the
mysterious heavenly forces
that someone might imagine as governing the rolls of dice and roulette wheels.

And this is because all the foregoing French phrases trigger
ideas
in our human minds, rather than merely triggering counterpart English phrases, and the evoked ideas fit together into large patterns that trigger large-scale
memories
, which, only at that stage of the game, evoke English
words.
The process of translation depends crucially on the intermediate phase in which memories and concepts are triggered — an unavoidable phase usually called “understanding”. And this process involves putting together all the pieces of a sentence in a carefully coordinated manner, which means exploiting all the indications that grammar gives us about how the ideas fit together in a sensible pattern. No translation worthy of the term can afford to ignore the meaning of the text to be translated, and meaning can be grasped only if complex grammatical constructions are taken into account, which means making a precise linguistic analysis of the text, which today’s translation engines are unable to do.

Consider just one example from our obituary paragraph. The last sentence opens as follows: “Qui ne l’a pas vue « récupérer » en quelques quarts d’heure les pertes de toute une nuit…”. There arises the question of why an “e” has been tacked onto the past participle “vu” of the verb “voir” (“to see, to watch”). This is a question of French grammar, and its answer is that whenever a past participle is preceded by its direct object, it must agree in gender and in number with that object. That little “e” is therefore telling us that the direct object of the verb “voir” is feminine and singular. It carries a crucial meaning! It tells us that a thing or a person of feminine gender is being watched (or rather, since the verb is negated, that this thing or person is
not
being
watched). And indeed, the text is talking about failing to see an “entity” (possibly an animate one) that was quickly recovering from the losses of an all-night gambling session. We instantly realize that this “entity” is in fact none other than Françoise Sagan. The past participle’s feminine singular ending has clued
us
in, but both the 2004 and 2009 engines were clueless. To ignore grammar is to invite disaster.

Could a good mastery of French grammar conceivably have helped the translation engines? That is a rhetorical question, whose answer is “yes”, of course. Translation depends on understanding, and understanding depends on grammar, because grammar tells us how the smaller pieces fit together to make a large, coherent structure.

The preceding section’s purpose was not to deride efforts at machine translation. Indeed, the blinding speed with which virtually anyone today can, entirely for free, get a glimpse of what is going on in a piece of text written in a language of which they don’t know a single word, or perhaps even a single symbol, is very impressive. Thanks to the efforts of many researchers over many decades, anyone today can, in but seconds, get a feel for what a Website written in, say, Arabic, Estonian, Hindi, Icelandic, Korean, Malay, Maltese, Romanian, Swahili, Thai, Turkish, or Vietnamese is talking about, and sometimes the one-second miracle is coherent for fairly long stretches. This accomplishment is both astonishing and humbling. But for all its impressiveness, it is not to be confused with what human translation is about.

Genuine
translation — translation that merits the label “translation” — is indeed about analogy-making on all levels imaginable, from the most minuscule grammatical ending of a word to the entire overarching cultural context in which the text and the events and notions of which it speaks are embedded. In further support of this thesis, we will now briefly touch on some of the issues in analogy-making that have arisen in the translation of this book, since it is appearing simultaneously in both French and English, and in order to produce these two versions, we, its authors, have had to grapple with innumerable nontrivial translation challenges.

Take the second paragraph of
Chapter 1
, for instance. In the French version (which was the source for that particular paragraph), a picture is painted of what a rider of the Paris métro is taking in, and this picture is familiar to every French reader. We could have converted that passage into English in a non-transculturating manner, in which case the corresponding two sentences of
Chapter 1
would have run like this:

But we chose not to translate it in that fashion, not wanting to disorient our American readers from the first moment. Of course our American readers are likely to know, either personally or indirectly, the Paris métro. Nonetheless, it struck us that it would seem distinctly
foreign
if the book started out in a foreign country, and we didn’t want to give our book that tone. Indeed, that is why we started out the
French
version of our book in the Paris métro — we wanted it to feel familiar to our French-speaking readers. Naturally, we were shooting for “exactly the same” effect in English; and especially since our book is about analogies, it was self-evident that we had to find an
analogous
way to start our book in English.

The most obvious analogy would have been to transplant the scene to the New York subway system, in which case most of what we wrote above would have worked just fine, although we might have changed the accordion to another instrument. The problem we felt with that idea was that riding the New York subway is not a universally shared experience of all Americans. It feels “New-Yorkocentric” in a way that the French version did not feel Paris-centric. Paris plays such a central role in France that scenes taking place in Paris feel generic to all French readers, whereas to many American readers, scenes in New York City exude an unfamiliar feel. Moreover, riding the subway is a rather unusual, perhaps even exotic, experience for many Americans. In that case, what in the life of a typical American corresponds to taking the Paris métro? It could of course be driving a car somewhere, but we preferred to keep the image a little closer to the original, thus in the domain of public transport, and so we chose the idea of sitting in an airport somewhere — not a specific airport, but just a random, generic one. Accordingly, here is what we came up with: