Is That a Fish in Your Ear?

TWENTY-THREE The Adventure of Automated Language-Translation Machines

The reluctance of European peoples to retain Latin or to adopt some other transmission language—such as Esperanto—for the dissemination of important information has created a costly and difficult set of translation tasks, carried out under time pressures unimaginable in earlier ages. Now that nearly all other aspects of news transmission are carried out not by couriers but by electronic devices, it seems natural to ask why the core activity itself cannot be handled likewise, by automatic translation machines.

Although it is still in its infancy, machine translation has had an eventful and uneven history. It first arose in dramatic historical circumstances and in response to an overriding political need. It wasn’t initiated by an explicit act of political will, like the language rules of the European Union, but its launching ground was the climate of terror at the start of the Cold War. The United States had developed and used the atomic bomb. For the time being it had a monopoly on this terrible weapon. How long would the monopoly last? When would the Soviet Union catch up? One way of guessing the answer was to comb through all the research journals being published in the U.S.S.R., looking for clues as to the state of knowledge in the relevant disciplines.[145] The journals were in Russian. The United States needed either to train up a veritable army of Russian–English scientific translators—or to invent a machine that would do the job for them.

But it takes a long time to constitute a large group of translators from a language not widely known. There was no obvious source of English-educated, scientifically literate Russian translators in 1945, and so the authorities began to look toward machines. There were good reasons to think they could help with the urgent task of tracking the Soviets’ ability to design an atomic bomb.

The Second World War had fostered great advances in cryptography, the making and breaking of secret codes. Statistical techniques had been developed for decoding messages even when the language that had been encoded was not known. The astounding successes of the code breakers at the Bletchley Park site in England prompted some thinkers to wonder whether language itself could not be treated as a code. In a famous memorandum written in July 1949, Warren Weaver, then a senior official with the Rockefeller Foundation, found it “very tempting to say that a book written in Chinese is simply a book in English which was coded into the ‘Chinese code.’ If we have useful methods for solving almost any cryptographic problem, may it not be that with proper interpretation we already have useful methods for translation?”[146]

Weaver was aware of the pioneering work of Claude Shannon and others in the nascent disciplines of information theory and cybernetics and could see that if language could be treated as a code, then there would be huge development contracts available for mathematicians, logicians, and engineers working on the new and exciting number-crunching devices that had only just acquired their modern name of “computers.” But the temptation to see “language as code” comes from much deeper sources than just an intuition that it would create interesting jobs for very smart boys.

A code, or cipher, is a way of representing a piece of information in a way that is receivable only if the (secret) key to the code is available. However sophisticated the key, however complicated the algorithm that turns the “source” into “code,” there is always a discoverable relationship between the expression in code and the encoded expression. If a language itself is a code of that kind, what does it encode? There’s only one possible answer in the long Western tradition of thinking about language since the time of the Greeks, and that answer is: meaning (sometimes called “thought”). A translation machine would need to strip away from the actual expression in language A all that is “code,” so as to access the real thing that it encodes, namely, the actual, irreducible, plain, and basic meaning of the expression. It’s really no more than a rehearsal of the ancient idea that language is the dress of thought. Weaver himself proposed the following analogy:

That dream of “easy and useful communication” with all our fellow humans in the “great open basement” that is the common foundation of human life expresses an ancient and primarily religious view of language and meaning that has proved very hard to escape, despite its manifestly hypothetical nature. For what language would humans use to communicate with one another in the “great open basement”? The language of pure meaning. At later stages in the adventure of machine translation and modern linguistics, it came to be called “interlingua” or “the invariant core” of meaning and thought that a communication in any language encodes.

The task that machine-translation pioneers set themselves was therefore almost identical to the task of the translator as expressed by many modern theorists and philosophers: to discover and implement the purely hypothetical language that all people really speak in the great open basement of their souls.

How was that to be done by machines? Plenty of intellectual machinery already existed that seemed designed for the purpose. Ever since the Romans started teaching their young to read and write Greek, language learners in Western tongues have always been told they have two basic tasks: to acquire vocabulary in the foreign tongue, and to learn its grammar. That’s why we have bilingual dictionaries separate from our grammar books, which give the set of rules by which the “words” in the vocabulary may be combined into acceptable strings. That’s what a language is, in our ancient but undimmed language theology: a Meccano set, made up of one part nuts, bolts, girders, beams, cogwheels, and perforated bars (let’s say, prepositions, verbs, nouns, adjectives, particles, and determiners) and, for the other part, rules about how to fix them together. A nut goes on a bolt but not on a cogwheel, just as a verb clicks on a subject before and an object after …

It was theoretically possible at the start of the machine-translation adventure (and it soon became practically possible as well) to store a set of words on a computer, divided into the grammatical classes the Greeks and Romans devised. It was equally possible to store two sets of words, one for Russian, one for English, and to tell the computer which English word matched which Russian one. More dubious was the proposition implicit in Weaver’s fable that you could bring people down from their separate towers to the common basement—that’s to say, tell a computer what to do to unwrap the meaning of a sentence from the form of the sentence itself. To do that, the computer would first need to know the entire grammar of a language. It would have to be told what that consists of. But who knows the entire grammar of English? Every language learner quickly realizes that systematic regularities are frequently overruled by exceptions of arbitrary kinds. Every speaker of a native language knows that she can (and frequently does) break the “rules” of grammar. A complete linguistic description of any language remains an aspiration, not a reality. That is one of the two reasons why the first great phase of machine translation hit the skids. The second is that even humans, who can be assumed to be in full possession of the grammar of their language, still need a heap of knowledge about the world in order to fix the meaning of any expression—and nobody has figured out how to get a computer to know what a sentence is about. A classic conundrum that computers could not solve is to attribute the correct meanings to the words in the following two sentences: “The pen is in the box” and “The box is in the pen.” Understanding them calls on knowledge of the relative sizes of things in the real world (of a pen-size box and a sheep pen, respectively) that can’t be resolved by dictionary meanings and syntactic rules. In 1960, the eminent logician Yehoshua Bar-Hillel, who had been hired by MIT specifically to develop “fully automated high-quality translation,” or FAHQT, sounded a testy retreat:

That pretty much put an end to easy money from the grant-giving foundations. But the establishment of the European Union in 1957 provided a new political impetus—and a new funding source—for the development of the tools that Bar-Hillel thought impossible. Ambitions were scaled down from FAHQT to more feasible tasks. As computers grew in power and shrank in size, they could more easily be relied upon for tasks that humans find tiresome, such as checking that a given term has been translated the same way each time it occurs in a long document. They could be used for compiling and storing dictionaries not just of technical terms but of whole phrases and expressions. The era not of fully automatic translation but of CAT—computer-aided translation—began. Private companies started developing proprietary systems, for although the big demand came from transnational entities such as the EU, there was a real need for such tools among major companies producing aircraft, automobiles, and other goods to be sold all over the world.

It is easier to achieve good results from CAT when the input conforms not to a natural language in its raw and living state but to a restricted code, a delimited subspecies of a language. In an aircraft-maintenance manual you find only a subset of the full range of expressions possible in English. To produce the hundred or so language versions of the manual that are needed through an automatic-translation device, you do not need to make the device capable of handling restaurant menus, song lyrics, or party chitchat—just aircraft-maintenance language. One way of doing this is to pre-edit the input text into a regularized form that the computer program can handle, and to have proficient translators post-edit the output to make sure it makes sense (and the right sense) in the target tongue. Another way of doing it is to teach the drafters of the maintenance manuals a special, restricted language—Boeinglish, so to speak—designed to eliminate ambiguities and pitfalls within the field of aircraft maintenance. This is now a worldwide practice. Most companies that have global sales have house styles designed to help computers translate their material. From computers helping humans to translate we have advanced to having humans help computers out. It is just one of the truths about translation that shows that a language is really not like a Meccano set at all. Languages can always be squeezed and shaped to fit the needs that humans have, even when that means squeezing them into computer-friendly shapes.

Computer-aided human translation and human-aided computer translation are both substantial achievements, and without them the global flows of trade and information of the past few decades would not have been nearly so smooth. Until recently, they remained the preserve of language professionals. What they also did, of course, was to put huge quantities of translation products (translated texts paired with their source texts) in machine-readable form. The invention and the explosive growth of the Internet since the 1990s has made this huge corpus available for free to everyone with a terminal. And then Google stepped in.

Using software built on mathematical frameworks originally developed in the 1980s by researchers at IBM, Google has created an automatic-translation tool that is unlike all others. It is not based on the intellectual presuppositions of Weaver, and it has no truck with interlingua or invariant cores. It doesn’t deal with meaning at all. Instead of taking a linguistic expression as something that requires decoding, Google Translate (GT) takes it as something that has probably been said before. It uses vast computing power to scour the Internet in the blink of an eye looking for the expression in some text that exists alongside its paired translation. The corpus it can scan includes all the paper put out since 1957 by the EU in two dozen languages, everything the UN and its agencies have ever done in writing in six official languages, and huge amounts of other material, from the records of international tribunals to company reports and all the articles and books in bilingual form that have been put up on the Web by individuals, libraries, booksellers, authors, and academic departments. Drawing on the already established patterns of matches between these millions of paired documents, GT uses statistical methods to pick out the most probable acceptable version of what’s been submitted to it. Much of the time, it works. It’s quite stunning. And it is largely responsible for the new mood of optimism about the prospects for FAHQT, Weaver’s original pie in the sky.

GT could not work without a very large preexisting corpus of translations. It is built upon the millions of hours of labor of human translators who produced the texts that GT scours. Google’s own promotional video doesn’t dwell on this at all. At present it offers two-way translation between fifty-eight languages, that is to say, 3,306 separate translation services, more than have ever existed in all human history to date. Most of these translation relations—IcelandicFarsi, YiddishVietnamese, and dozens more—are the newborn offspring of GT: there is no history of translation between them, and therefore no paired texts, on the Web or anywhere else. Google’s presentation of its service points out that given the huge variations among languages in the amount of material its program can scan to find solutions, translation quality varies according to the language pair involved.[149] What it does not highlight is that GT is as much the prisoner of global flows in translation as we all are. Its admirably smart probabilistic computational system can offer 3,306 translation directions only by using the same device as has always assisted intercultural communication: pivots, or intermediary languages. It’s not because Google is based in California that English is the main pivot. If you use statistical methods to compute the most likely match between languages that have never been matched directly before, you must use the pivot that can provide matches with both target and source.

The service that Google offers appears to flatten and diversify interlanguage relations beyond the wildest dreams of even the EU’s most enthusiastic language-parity proponents. But it is able to do so only by exploiting, confirming, and increasing the central role played by the most widely translated language in the world’s electronic databank of translated texts, which can only be the most consistently translated language in all other media, too.

A good number of English-language detective novels, for example, have probably been translated into both Icelandic and Farsi. They thus provide ample material for finding matches between sentences in the two foreign languages; whereas Persian classics translated into Icelandic are surely far fewer, even including those works that have themselves made the journey by way of a pivot such as French or German. This means that John Grisham makes a bigger contribution to the quality of GT’s Icelandic–Farsi translation device than Halldór Laxness or Rumi ever will. And the real wizardry of Harry Potter may well lie in his hidden power to support translation from Hebrew into Chinese.

GT-generated translations themselves go up on the Web and become part of the corpus that GT scans, producing a feedback loop that reinforces the probability that the original GT translation was acceptable. But it also feeds on human translators, since it always asks users to suggest a better translation than the one it provides—a loop pulling in the opposite direction, toward greater refinement. It’s an extraordinarily clever device. I’ve used it myself to check I had understood a Swedish sentence more or less correctly, for example, and it is used automatically as a Webpage translator whenever you use a search engine. Of course, it may also produce nonsense. However, the kind of nonsense a translation machine produces is usually less dangerous than human-sourced bloopers. You can usually see instantly when GT has failed to get it right, because the output makes no sense, and so you disregard it. (This is why you should never use GT to translate into a language you do not know very well. Use it only to translate into a language in which you are sure you can recognize nonsense.) Human translators, on the other hand, produce characteristically fluent and meaningful output, and you really can’t tell if they are wrong unless you also understand the source—in which case you don’t need the translation at all.

If you remain attached to the idea that a language really does consist of words and rules and that meaning has a computable relationship to them (a fantasy to which many philosophers still cling), then GT is not a translation device. It’s just a trick performed by an electronic bulldozer allowed to steal other people’s work. But if you have a more open mind, GT suggests something else.

Conference interpreters can often guess ahead of what a speaker is saying because speakers at international conferences repeatedly use the same formulaic expressions. Similarly, an experienced translator working in a familiar domain knows without thinking that certain chunks of text have standard translations that she can slot in. At an even more basic level, any translator knows that there are some regular transpositions between the two languages she is working with—the French impersonal pronoun on, for example, will almost always require the English sentence to be in the passive; adjectives following a noun in French will need to be put in front of the equivalent English noun; and so on. These automatisms come from practice and experience. Translators don’t reinvent hot water every day, and they don’t recalculate the transformation “French on English passive construction” each time it occurs. They behave more like GT—scanning their own memories in double-quick time for the most probable solution to the issue at hand. GT’s basic mode of operation is much more like professional translation than is the slow descent into the “great basement” of pure meaning that early machine-translation developers imagined.

GT is also a splendidly cheeky response to one of the great myths of modern language studies. It was claimed, and for decades it was barely disputed, that what was so special about a natural language was that its underlying structure allowed an infinite number of different sentences to be generated by a finite set of words and rules. A few wits pointed out that this was no different from a British motorcar plant capable of producing an infinite number of vehicles each one of which had something different wrong with it—but the objection didn’t make much impact outside Oxford. GT deals with translation on the basis not that every sentence is different but that anything submitted to it has probably been said before. Whatever a language may be in principle, in practice it is used most commonly to say the same things over and over again. There is a good reason for that. In the great basement that is the foundation of all human activities, including language behavior, we find not anything as abstract as “pure meaning” but common human needs and desires. All languages serve those same needs, and serve them equally well. If we do say the same things over and over again, it is because we encounter the same needs and feel the same fears, desires, and sensations at every turn. The skills of translators and the basic design of GT are, in their different ways, parallel reflections of our common humanity.

In September 2009, the new administration in the White House issued a science policy road map, titled A Strategy for American Innovation. The last section of this document calls for science and technology to be harnessed to address the “‘Grand Challenges’ of the 21st Century,” of which it gives half a dozen examples, such as solar cells “as cheap as paint” and intelligent prosthetics. The last line of the whole strategy puts among these long-range targets for national science policy the development of “automatic, highly accurate and real-time translation between the major languages of the world—greatly lowering the barriers to international commerce and collaboration.”[150] Not every science policy target is achieved, but with serious backing from the U.S. administration now in place for the first time since 1960, machine translation is likely to advance far beyond the state in which we currently know it.