Surfaces and Essences

by Douglas Hofstadter

  Despite the complexity that we’ve just described, translation seems like a relatively mechanical activity to some people. And indeed, the idea of mechanical translation was first proposed three score and several years ago, at which time it seemed reasonable and relatively straightforward. For example, one of the founders of the field, the noted mathematician Warren Weaver, wrote the following: “When I look at an article in Russian, I say, ‘This is really written in English, but it has been coded in some strange symbols. I will now proceed to decode.’ ” Weaver’s humorous statement expresses the credo underlying machine translation, which is that translation is an act of “decoding” essentially analogous to using a substitution cipher, in which, in order to encode or decode a message, all one needs to do is to replace the message’s symbols, one by one, by other symbols, according to a fixed table of correspondences. To be concrete, one might encode a message by replacing every letter in it by its alphabetic successor, which will in general yield an illegible piece of text, such as “Gpvs tdpsf boe tfwfo zfbst bhp”; then, in order to decode such an encrypted message, one would do the reverse transformation, replacing every letter by its alphabetic predecessor, in this case yielding Abraham Lincoln’s famous opening gambit, “Four score and seven years ago”.

  In the early days of machine translation, translation between languages was seen as this same process but just on a larger scale, in the sense that it operated not on letters but on words, and the correspondence table wasn’t just 26 items long but was a huge bilingual dictionary, giving, for each word, the matching word in the other language. And lastly, it was presumed that, in order to fix up awkwardnesses due to discrepancies between the grammars of the two languages involved, the scrambled word order that would almost surely result could be cleaned up by a complex but straightforward mechanical rearrangement process that took both grammars into account.

  This substitution-and-rearrangement process remains the most common philosophy underpinning mechanical-translation efforts even today, except that the units in the semantic correspondence table are often taken at a somewhat higher level than that of individual words — they can include idiomatic phrases and other large chunks (for example, “to be under the weather” might be rendered in French, as an indivisible chunk, by “ne pas être dans son assiette”, literally meaning “not to be in one’s dish”). Indeed, bilingual data bases containing many millions of corresponding phrases are thoroughly scoured in ultra-rapid fashion in order to allow the computer to find “the best” (in some sense) alignments that will yield the “decoding” into Language B of a phrase, sentence, or any passage originally written in Language A.

  We might test the efficacy of this strategy by seeing how the world’s most readily available (and perhaps also its most sophisticated) machine-translation “engine” performs on our little Lincolnian phrase “Four score and seven years ago”. (The term “engine” is a friendly tip of the hat to Charles Babbage, the great nineteenth-century British computing pioneer, who invented an important predecessor of today’s computers, which he called the “Analytical Engine”.) We asked Google’s engine to translate this phrase into French, and in but milliseconds it shot back at us the words “Quatre points et il ya sept ans”. (Strangely enough, “ya” is not a French word, but that’s nonetheless what the engine came up with.) To test the program’s understanding of its own output, we fed this phrase (with “ya” unchanged) into the engine and threw it into reverse gear; instantly, out popped “Four points and seven years ago” (the clever engine didn’t stumble in the least over the non-word “ya”). Anyone can imagine how the translation engine might have gotten “four points” out of the words “four score”, but, despite the plausibility of that guess, the engine’s output, whether in French or in English, doesn’t make any sense.

  We carried this experiment out not just once but many times, and discovered that the French output was not stable at all. Sometimes it was better, sometimes not; in any case, the first output we quoted reappeared on more than one occasion. When we asked the same translation engine to render Lincoln’s phrase in German, it returned to us “Vier der Gäste und vor sieben Jahren” — “Four of the guests and seven years ago”. We have no clue as to how, beginning with “score”, it came up with “guests”. In Spanish, it gave us “La puntuación de cuatro y siete años atrás” — “The grade of four and seven years back” (“grade” meaning “school grade”).

  The translation engine has no notion of meaning. It is not trying to understand its input, but simply to operate on the marks making it up. In that regard, the operations carried out by the Google translation engine are, indeed, much like the operations of encoding or decoding using a substitution cipher, which don’t involve meaning in any way at all. This is exactly the view that Warren Weaver proposed sixty years earlier, his vision having been of an unthinking, meaning-ignoring, mechanical substitution process that would allow one to toggle back and forth between any two languages.

  It is difficult to fathom how such a simplistic view of translation could possibly have been put forth by the same person who wrote Alice in Many Tongues, a delightful short study devoted, with clear love, to a discussion of how some of the trickier passages in Lewis Carroll’s wordplay-filled Alice in Wonderland had been rendered by highly creative translators into French, Italian, German, Danish, Swedish, Spanish, Russian, Polish, Hungarian, Hebrew, Swahili, Pidgin English, Japanese, and Chinese. In this book, Warren Weaver pays careful attention to how superb translators handled such complex challenges as parodied verse, wordplay, nonsense passages, and other ways in which form and content are deeply entangled with each other, and with reverence he describes the artistry with which some of the translators found brilliant solutions to these challenges. Over and over again he points out that high-quality translation is anything but mechanical, and in so doing, he implicitly demonstrates as many times that translation in no way resembles a “mechanical decoding” process, but that it requires continual discovery of ingenious new analogies. Indeed, his book’s message is that translation depends crucially on drawing deeply from the well of one’s life experiences and mental resources in order to come up with appropriate analogues in Language B for strings of characters written in Language A.

  Only a few years after the dream of machine translation was hatched, it was already starting to run into profound problems. These problems were articulated by a number of skeptics, of whom perhaps the most vociferous was the logician Yehoshua Bar-Hillel, who had once been one of the field’s earliest and most enthusiastic researchers. In the mid-1950’s, then, there was already deep skepticism over the idea of translation as a “mechanical” or “algorithmic” process, and such skepticism is still warranted today, as we have just seen.

  Good Analogies Make Good Translations

  How, then, should one translate Lincoln’s famous speech-opener? Suppose we wished to translate it into French, for instance. It helps to know what it means. To begin with, then, since a score consists of twenty items, it would follow that “four score”

  means “eighty”, and that “four score and seven” means “eighty-seven”. Can we say this in French? Yes, fortunately, the French, who have a strong mathematical tradition, do have a way of saying “eighty-seven” — namely, “quatre-vingt-sept”. To say “years”, say “ans”, and to say “ago”, just put “il y a” at the very front. So putting the little pieces together, we get “Il y a quatre-vingt-sept ans”. All done!

  There is a problem, though, which is that Lincoln’s phrase “four score and seven” is not the usual way of saying “eighty-seven” in English, but a special way of saying it, based on a somewhat odd way of looking at the number and which, more importantly, has a clear poetic resonance. These are both key qualities of Lincoln’s turn of phrase, so it would be crucial to preserve both of them in our translation.

  Now, by a remarkable coincidence, the French word for “eighty” happens not to be based on “huit” (the French word for “eight”),
but on the words “quatre” and “vingt” (“four”, “twenty”) — so that “quatre-vingts”, the standard French word for “eighty”, actually means “four twenties”! It would seem that the ideal solution — “Il y a quatre-vingt-sept ans” (“Four-twenty-seven years ago”) — fell right out of the sky. What luck! It certainly would not have worked if the Gettysburg Address had been delivered in 1843 (an event that would have been confusing to more than one historian), since “three score and seven” is expressed in French not as “trois-vingt-sept” (that is, not as “three-twenty-seven”) but as “soixante-sept” (that is, as “sixty-seven”, very literally).

  It might at first blush seem that thanks to this stroke of luck, we are in fact done, but this is a hasty conclusion. French speakers almost never notice the “quatre” and the “vingt” inside “quatre-vingts” — no more than English speakers hear the words “for” and “tea” inside “forty”, or, for that matter, an allusion to the terror of Mongol invaders inside “hundred”. Even if for English speakers, “quatre-vingts” seems to brim with the proper arithmetical meaning and also to exude an archaic, poetic flavor, it does nothing of the sort for French speakers; to them, it is just an ordinary, pedestrian word for the number halfway between 70 and 90. By contrast, for English speakers, Lincoln’s phrase requires a tad of conscious calculation and resounds with poetry and nobility. Therefore, if we hope to capture its essence in another language, then we have to avoid, at all costs, the mundane — and this means that what we at first took for a great stroke of luck turns out to have been a cursing in disguise! And so, it’s back to the drawing board.

  Our goal is to be high-quality copycats, which is to say, to “do what Lincoln did” — and what Lincoln did is to replace the standard English eight-tens view of eighty by an exotic, attention-grabbing four-twenties view. At this point, it may seem obvious to some that the solution is simply to use a reversal analogy — namely, to replace French’s standard four-twenties view of eighty by an exotic, attention-grabbing eight-tens view. Indeed, this is even reminiscent of that elegant Copycat flip of perspective whereby xyz is seen as the mirror image of abc. And as a matter of fact, the word “huitante” (meaning essentially “eight-ty”), though exotic in France, is commonly used in Switzerland in lieu of “quatre-vingts”, and in some French-speaking locales the even more exotic word “octante” is also a dialectal way of saying “eighty”, though today it has almost fallen out of currency. So what about using one or the other of these unusual French words for “eighty”? Well, unfortunately, the phrases “Il y a huitante-sept ans” and “Il y a octante-sept ans”, rather than sounding poetic and uplifting, come across to most French speakers as simply quaint. As would-be translations of Lincoln’s noble phrase, they are both extremely wanting. In this case, the often clever idea of a reversal analogy doesn’t pay off.

  Clearly, then, we will need to search around for a deeper French analogue to Lincoln’s phrase. What might be the French counterpart of the poetic phrase “four score and seven years ago”? Without recounting all our failed forays, we can simply say that we began by exploring the idea of “seven dozen and three years ago”, which uses the common French word “douzaine”. This, however, did not seem analogous enough to the original because, among other things, it involved too much conscious calculation, although it did at least activate the closely related thought of using the French word “dizaine” (which is analogous to “douzaine”, but features only ten items). And once again, although “il y a huit dizaines et sept ans” sounded silly to our ears, it reminded us of the word “trentaine” (which also is like a dozen, but involves thirty items), so that “eighty-seven” could be expressed as “trois trentaines moins trois” — “three thirties minus three”. At this point, however, our translation was becoming too much like an arithmetic problem, or even a mild tongue-twister. Hardly our goal!

  Eventually, it dawned on us that French has the word “vingtaine” as well — based on “vingt” and meaning a collection of twenty items (in other words, a score). Thus we finally hit upon the translation “Il y a quatre vingtaines d’années, plus sept ans”. This was far more promising, but still, its flavor was not sufficiently analogous to that of the original, since the explicit mention of the arithmetical operation “plus” was too heavy-handed, as if Lincoln had said “four score plus seven years ago”. In the end, though, we were able to raise the loftiness of the tone by making some minor adjustments, as follows: “Voici quatre vingtaines d’années, et encore sept ans…” (“Four score years ago, and yet seven more…”). We were quite proud of our collaborative find.

  Another idea was suggested to us by a translator friend who took advantage of the poetic word “lustre”. Although most native speakers of French think of it as simply meaning “a long time”, it can also mean a five-year chunk of time. This fact allowed our friend to render the president’s lustrous words by “Voici seize lustres et encore sept ans…” (“Sixteen lustres and seven years ago…”), which exudes a rather grand and lofty flavor. (Of course he could have said “Seventeen lusters and two years ago…”, but it’s not clear that that’s an improvement.)

  Altogether, then, through a slow process of carefully honed analogy-making and analogy-judging, we eventually managed to recreate some of the high-sounding flavor of Abraham Lincoln’s immortal phrase, while sidestepping various superficially enticing traps along the way.

  Potential Progress in Machine Translation

  The preceding anecdote confirms the pervasive thesis of Warren Weaver’s book Alice in Many Tongues, which is that to translate well, the use of analogies is crucial. In order to come up with possible analogies and then to judge their appropriateness, one must carefully exploit one’s full inventory of mental resources, including one’s storehouse of life experiences.

  Could machine translation possibly do anything of the sort? Is it conceivable that one day, computer programs will be able to carry out translation at a high level of artistry? A couple of decades ago, some machine-translation researchers, spurred by the low quality of what had then been achieved in their field, began to question the methods on which the field had been built (mostly word-matching and grammatical rules), and started exploring other avenues. What emerged with considerable vigor was the idea of statistical translation, which today has become a very important strategy used in tackling the challenge of machine translation.

  This approach is based on the use of statistically-based educated guesswork, where the data base in which all guesses are rooted consists of an enormous storehouse of bilingual texts, all of which have been carefully translated by human experts. A typical example of such a data base is the proceedings, over several decades, of the Canadian Parliament, which are legally required to be made available in both English and French. Such a data base is a marvelous treasurehouse of linguistic information, if only one can figure out how to exploit it.

  The basic idea of statistical machine translation is to choose among the many possible meanings of a “chunk” (that is, a word or several-word segment) in a piece of input text (i.e., a text to be translated) by exploiting the context in which the chunk appears in the given passage. Suppose, for instance, that the engine is translating from English to French. The English chunk to translate may appear in many thousands of diverse contexts in the English side of the bilingual data base, but only a small number of those thousands of contexts (say, two dozen) are likely to be found sufficiently “similar” to the original context (where “similarity” is judged by a complex statistical calculation). This phase of narrowing-down on the basis of statistical similarity is the crux of the matter. In the human-translated bilingual data base, each of these relatively few English-language contexts comes aligned with a corresponding French-language context. The translation problem would thus seem to have been reduced to simply zeroing in on the corresponding chunk in these few French contexts. Unfortunately, though, this vision is too optimistic. In general there won’t be just one precisely corresp
onding chunk in the French contexts; there may be quite a few rival candidate French chunks, and so, to get a good candidate, educated guesswork (i.e., further statistical calculations, the details of which we will skip) is called for. The long and the short of it is that in this computationally intensive fashion, which takes advantage of vast amounts of human-translated text, the French chunk that is “most probably equivalent” to the English chunk is pinpointed and is inserted into the outgoing stream of French words.

  One way of describing the translation algorithm that we’ve just sketched is that, through a sophisticated and highly efficient set of computations, it repeatedly makes analogies between pieces of text in the two languages involved. This sounds nothing if not promising, but the proof of the pudding is in the eating, and so we will now proceed to sample the pudding. In order to do so, we’ll take a careful look at a short piece of French text in order to see how two extremely different machine-translation programs dealt with it — one using the old strategy, and one using the new strategy. The passage we will examine is taken from an obituary of the novelist Françoise Sagan, written by the literary critic Bertrand Poirot-Delpech, and which appeared in the highly respected national newspaper Le Monde in September of 2004. The paragraph we selected is written in elegant and evocative but standard French, readily understood by any literate native speaker. We did not choose it for its difficulty; indeed, its density of “traps” for a translator is no higher than that of any typical article in Le Monde.