Perfect is the enemy of good

Dans ses écrits, un sàge Italien dit que le mieux est l’ennemi du bien.
François-Marie Arouet de Voltaire

A Lego Turing MachineLe Roi est mort, vive le Roi! (The King is dead. Long live the King.) was first declared upon the accession to the French throne of Charles VII after the death of his father, Charles VI, in 1422, and it means that there would be never a time without a king. Today, this phrase crops up regularly on themes of succession or replacement.

On May 30, Yahoo! retired Babel Fish, the first online machine translation engine launched in 1999, and replaced it with Microsoft Bing Translator.

Founded and developed by the Canadian entrepreneur and technology specialist Oscar A Jofre, Babel Fish was powered by a rule-based machine translation engine provided by SYSTRAN.

Rule-based machine translation engines run on linguistic information about source and target languages retrieved from bilingual dictionaries and grammars. The engine generates the translation of input sentences on the basis of morphological, syntactic, and semantic analysis and the regularities of both the source and the target languages.

Today, statistical machine translation (SMT) is the dominant approach in machine translation. Essentially, SMT uses a corpus-based approach founded on the use of parallel corpora, following two distinct and separate processes: training and decoding.

During the training process, a statistical model of translation is derived from a parallel corpus, and a statistical model of the target language is derived from a monolingual corpus. These models are used in the decoding process (i.e. translation), where translation is treated as a search problem: given the sentence to be translated, all the possible translations according to the translation and the language model are searched over: the translation with the highest overall probability is chosen.

RBMT was giving good results, but nowadays SMT has been improved considerably and is being used in a totally new way, benefiting from the typical task of a computer: brute force computing.

In 1950, Alan Turing, the father of artificial intelligence, wrote a single essay, asking the question “Can machines think?” Since then, computers have not evolved quite as Turing expected them to. On the other hand, Turing had actually no very convincing arguments of a positive nature to support his views; and yet, he put his trust in the future development of “learning machines,” even though he knew it would not be possible to apply to a machine exactly the same teaching process that is applied with a normal child, as education requires that communication in both directions (between teacher and pupil) can take place by some means or other. For example, a “learning machine” should be so constructed that those events which shortly precede a punishment signal are unlikely to be repeated, whereas a reward signal increases the probability of repetition of the events which led up to it.

Nonetheless, with respect to machine translation, we often talk of “training” a machine translation system.

To underline the progress made since Turing’s question, in celebrating Turing’s birth centennial, David Levy said that in certain areas of medical diagnostics it has been proven that computers can be more accurate than doctors. He poses then an intriguing logical question: “Would you rather be diagnosed by a human doctor who’s 80% right or a computer doctor that’s 95% right?”

A similar question could arise with machine translation: “Without knowing the language, would you by a human translator who’s 80% right or by a computer that’s 95% right?”

In unveiling Duolingo, Luis von Ahn, the CAPTCHA/reCAPTCHA man, said that machine translation still makes too many mistakes, and that one can’t know whether one could trust it or not. At the same time, von Ahn said that translating Wikipedia into Spanish would cost about US $ 15 billion, even applying a rate lower than the “standard” US $ 0.20 per word. Too much.

The premise behind Duolingo is that users who don’t know a language can learn it while simultaneously translating content, in a cooperative, social endeavor. This idea may sound crazy to a translator, who is supposed to have spent his best years learning a language, and yet it should be no surprise that “the public doesn’t understand translation, by machines or by humans,” as Jost Zetsche wrote in the 210th edition of his newsletter.

However, this is clearly due to the conviction that translation is “only” a linguistic task, rather than challenging problem solving task.

On the other hand, the arrogance in approaching those problems is typical of linguists and translators. One example for all: the case of EU’s DGT and Karl-Johan Lönnroth with Juvenes Translatores. Lönnroth has always – wrongly – boasted a deep knowledge of Latin. If he really knew Latin, he would also know that “translator” does not exist in the language of Cicero.

In fact, the word “traducere” instead of “interpretare” is due to a misinterpretation by Leonardo Bruni.

Lönnroth is in good company, as even the distinguished Gisèle Sapiro fell over a sad “translatio” in her essay on the translation market (of course strictly literary) published in France.

It’s no wonder, then, that, in her article commemorating Alan Turing, Somini Sengupta hammers on what she deems is MT’s Archilles heel, i.e. the trouble algorithms supposedly have with satire and jokes, irony, wordplay, and cultural context. Ms. Sengupta seems to “forget” that a very tiny fraction of the content currently translated relies on irony and sarcasm to convey information, while wordplay and cultural context can be “re-enacted” by a SMT engine during the decoding process.

In Ms. Sengupta’s article, Alon Lavie, a Carnegie Mellon professor cofounder of a machine translation company called Safaba says that human and machine translation can work in different scenarios, and when businesses need to translate large amounts of text into multiple languages, machine translation can be more useful, particularly if business confidentiality is at stake.

According to Prof. Lavie, Duolingo’s crowdsourcing makes a lot of sense in these scenarios where a small translation job has to be done quickly and cheaply, and the translation needs to come out at a ‘human’ quality level — i.e. the quality has to be similar to the one generated by a human translator or ‘bilingual’ speaker.

Who should be blamed, then, if “the public doesn’t understand translation?”

Nataly Kelly’s article in the Huffington Post is not really what might justify the relevance of our profession.

For years many such articles  have not even scratched the surface of that sound and solid jumble of indifference and misconceptions that surround translation. Maybe these articles can have an effect on the public in the U.S., over a very limited period of time, but in any case the authors do not seem to think the public is “ready”.

How many readers, for example, in the remotest areas of the U.S. read the Huffington Post? And how many in the suburban areas are aware of the importance of translation? And finally, how many readers in the major metropolitan areas are willing to read carefully such an article and give it some thought?

It’s still better than nothing, of course,  even though, once you found space on the #1 in Technorati ranking, you could use it first to tell why “translation has an impact on virtually every aspect of society, politics, and economics,” i.e. why translation is and should be important even for an outsider like those portrayed in certain movies.

Until a few years ago, Renato Beninatto strongly advocated the idea that professional bodies should develop communication and public relations strategies, possibly in common, on behalf of the whole industry. During his presidency, however, ELIA has remained an irrelevant club, like many other industry bodies, showing that maybe the lack of interest in translation is not entirely unjustified, and that if industry events do not see many, if any, customers, the fault is not the customer’s.

Perhaps, then, it is not true that the public does not understand.

Maybe not realizing that the world keeps turning, some simply reiterate the same old, obsolete patterns, even though they once were pioneers.

It is the case of Alan Melby and what it looks like an endowment for Multi-Languages Corporation. The only thing that sounds sensational – although we should keep in mind that we are in 2012 – is the claim that “a quality translation is one that follows appropriate specifications,” and that “the job of a TSP is to make sure specifications are appropriate and then follow them exactly.”

Professor Melby is late. After having wrote and sponsored the ASTM F2575 – 06 standard, against the European standard EN 15038, he now promotes the Canadian standard CAN CGSB 131.10-2008, which is based on the latter. As for specifications, it might be worth giving a (second) look at an old article published in 2008 by Clientside News.

Professor Melby is definitely right in stating that, in all cases, project specifications should be agreed on before the production phase begins; on the other hand, he is definitely wrong when he states that if project specifications turn out to be inappropriate for the target audience or otherwise faulty, the project manager should be involved in modifying them, in cooperation with the requester and in seeking the approval of the client. In fact, these are not the project manager’s responsibilities. And thus another piece is added to the puzzle of ignorance of project management.

It is true that a project manager is responsible for finalizing specifications, selecting translators and other team members, and coordinating the execution of the project. But this almost never occurs, especially in the case of larger projects involving large LSPs.

Moreover, project managers are not necessarily required to understand the translation process, while they should be able to read a specification of requirements, know end-users, and understand the business processes of both the TSP and its customer.

However, the real pearls of “Almost everything you ever wanted to know about Translation” are the recommended hiring requirements for translators including, among others:

  • University degree in the target language (general);
  • University degree in linguistics and/or translation;
  • Certification by a professional association;
  • International certifications evaluated on an individual basis;
  • Work experience;
  • Participation in professional development activities;
  • References.

Obviously, a translator with such a profile is perfectly right in demanding a conspicuous fee, which would be actually very far away even from that Mr. von Ahn claims as unjustified.

This is not the way to promote awareness and appreciation of translation practice, not only because it’s far  from market reality, where, by the way, very few professional translators make full and savvy usage of TEnTs, while more and more translators, – certainly more than in 2004 – are actually using the large online machine translation systems for a first “dirty” translation pass if no translation memory hit is found.

These translators recognize the validity of the “good enough” approach they straightforwardly deny with colleagues and acquaintances.

Perhaps, this is why Nathaly Kelly’s article has been so widely appreciated in the translation community, where in spite of everything “quality” is still the most debated issue, even though it’s once again too late (as in Sharon O’Brien article on dynamic quality evaluation model for translation).

The transition from the 20th century to the 21st century has been marked by radical changes under the banner of globalization and massification of information.

The transformation of translation products into consumer goods is also the result of a shift  in the paradigm of the production and circulation of translated goods, toward  a purely commercial market approach. Following this massive transformation of translation services into a commodity, there are more people consuming translated materials than ever before, thanks to an increase in the sales of consumer goods.

The trend calls for the application of business models that are superseded in their original fields. These models do not account for the training received by translators (those who should implement them), a training which is far from the rules of mass production, starting with quality control.

We all have bought a mass production item to find a quality control tag in the package with a reference to the employee who inspected the item. A product that has passed the quality check is considered “perfect” within predefined thresholds. In other words, it is “adequate,” “good enough” because perfect is the enemy of good, and no process can be perfect even though it can be improved to ensure nearly perfect results.

In buyer-led commerce, like translation, the buyer dictates the terms and scope of the commerce. In seller-led commerce, the seller makes — often unsolicited — offers to the buyer on the seller terms.

Consider the pharmaceutical companies’ aggressive marketing of prescription drugs directly to consumers, a practice now prevalent in the U.S. but prohibited in most other countries. If you’ve ever seen the television commercials on the evening news, you could be forgiven for thinking that the greatest health crisis in the world are not AIDS, malaria, river blindness or sleeping sickness but erectile dysfunctions.

Technology does not change human behavior: it only allows certain problems to be solved more quickly and easily. This explains why technology is often used to address erectile dysfunctions, and even AIDS, but not so often for onchocerciasis and trypanosomiasis while in these cases it might be at least as effective as it was in the fight against malaria.

The challenge of new technologies must be addressed with courage, not to let it sink in worthlessness, be outcast, and, sooner or later, become obsolete and, eventually, junk forgotten in a cellar.