Original and Supplementary Questions
Submitted to the MT Summit III Conference,
Here are the original questions for this panel as submitted to the speakers:
1. At the last MT Summit, Martin Kay stated that there should be "greater attention to empirical studies of translation so that computational linguists will have a better idea of what really goes on in translation and develop tools that will be more useful for the end user." Does this mean that there has been insufficient input into MT processes by translators interested in MT? Does it mean that MT developers have failed to study what translating actually entails and how translators go about their task? If either of these is true, then to what extent and why? New answers and insights for the MT profession could arise from hearing what human translators with an interest in the development of MT have to say about these matters. It may well turn out that translators are the very people best qualified to determine what form their tools should take, since they are the end users.
2. Is there a specifically "human" component in the translation process which MT experts have overlooked? Is it reasonable for theoreticians to envision setting up predictable and generic vocabularies of clearly defined terms, or could they be overlooking a deep-seated human tendency towards some degree of ambiguity—indeed, in those many cases where not all the facts are known, an inescapably human reliance on it? Are there any viable MT approaches to duplicate what human translators can provide in such cases, namely the ability to bridge this ambiguity gap and improvise personalized, customized case-specific subtleties of vocabulary, depending on client or purpose? Could this in fact be a major element of the entire translation process? Alternately, are there some more boring "machine-like" aspects of translation where the computer can help the translator, such as style and consistency checking?
3. How can the knowledge of practicing translators best be integrated into current MT research and working systems? Is it to be assumed that they are best employed as prospective end-users working out the bugs in the system, or is there also a place for them during the initial planning phases of such systems? Can they perhaps as users be the primary developers of the system?
4. Many human translators, when told of the quest to have machines take over all aspects of translation, immediately reply that this is impossible and start providing specific instances which they claim a machine system could never handle. Are such reactions merely the final nerve spasms of a doomed class of technicians awaiting superannuation, or are these translators in fact enunciating specific instances of a general law as yet not fully articulated?
Since we now hear claims suggesting that FAHQT is creeping in again through the back door, it seems important to ask whether there has in fact ever been sufficient basic mathematical research, much less algorithmic underpinnings, by the MT Community to determine whether FAHQT, or anything close to it, can be achieved by any combination of electronic stratagems (transfer, AI, neural nets, Markov models, etc.).
Must translators forever stand exposed on the firing line and present their minds and bodies to a broadside of claims that the next round of computer advances will annihilate them as a profession? Is this problem truly solvable in logical terms, or is it in fact an intractable, undecidable, or provably unsolvable question in terms of "Computable Numbers" as set out by Turing, based on the work of Hilbert and Goedel? A reasonable answer to this question could save boards of directors and/or government agencies a great deal of time and money.
It was also envisioned that a list of Supplemental Questions would be prepared and distributed not only to the speakers but everyone attending our panel, even though not all of these questions could be raised during the session, so as to deepen our discussion and provide a lasting record of these issues.
FAHQT: Pro and Con
Consider the following observation on FAHQT: "The ideal notion of fully automatic high quality translation (FAHQT) is still lurking behind the machine translation paradigm: it is something that MT projects want to reach." (1) Is this a true or a false observation?
Is FAHQT merely a matter of time and continued research, a direct and inevitable result of a perfectly asymptotic process?
Will FAHQT ever be available on a held-held calculator-sized computer? If not, then why not?
To what extent is the belief in the feasibility of FAHQT a form of religion or perhaps akin to a belief that a perpetual motion device can be invented?
Technical Linguistic Questions
Let us suppose a writer has chosen to use Word C in a source text because s/he did not wish to use Word A or Word B, even though all three are shown as "synonyms." It turns out that all three of these words overlap and semantically interrelate quite differently in the target language. How can MT handle such an instance, fairly frequently found in legal and diplomatic usage?
Virtually all research in both conventional and computational linguistics has proceeded from the premise that language can be represented and mapped as a linear entity and is therefore eminently computable. What if it turns out that language in fact occupies a virtual space as a multi-dimensional construct, including several fractal dimensions, involving all manner of non-linear turbulence, chaos, and Butterfly Effects?
Post-Editors and Puppeteers
Let's assume you saw an ad for an Automatic Electronic Puppeteer that guaranteed to create and produce endless puppet plays in your own living room. There would be no need for a puppeteer to run the puppets and no need for you even to script the plays, though you would have the freedom to intervene in the action and change the plot as you wished. Since the price was acceptable, you ordered this system, but when it arrived, you found that it required endless installation work and calls to the manufacturers to get it working. But even then, you discovered that the number of plays provided was in fact quite limited, your plot change options even more so, and that the movements of the puppets were jerky and unnatural. When you complained, you were referred to fine print in the docs telling you that to make the program work better, you would have to do one of two things: 1) master an extremely complex programming language or 2) hire a specially trained puppeteer to help you out with your special needs and to be on hand during your productions to make the puppets move more naturally. Does this description bear any resemblance to the way MT has functioned and been promoted in recent years?
A Practical Example
Despite many presentations on linguistic, electronic and philosophical aspects of MT at this conference, one side of translation has nonetheless gone unexplored. It has to do with how larger translation projects actually arise and are handled by the profession. The following story shows the world of human translation at close to its worst, and it might be imagined at first glance that MT could easily do a much better job and simply take over in such situations, which are far from atypical in the world of translation. But, as we shall see, such appearances may be deceptive. To our story:
A French electrical firm was recently involved in a hostile take-over bid and law suit with its American counterpart. Large numbers of boxes and drawers full of documents all had to be translated into English by an almost impossible deadline. Supervision of this work was entrusted to a paralegal assistant in the French company's New York law firm. This person had no previous knowledge of translation. The documents ran the gamut from highly technical electrical texts and patents, records of previous law suits, company correspondence, advertisements, product documentation, speeches by the Company's directors, etc.
Almost every French-to-English translator in the NYC area was asked to take part. All translators were required to work at the law firm's offices so as to preserve confidentiality. Mere translation students worked side by side with newly accredited professionals and journeymen with long years of experience. The more able quickly became aware that much of the material was far too difficult for their less experienced colleagues. No consistent attempt was made to create or distribute glossaries. Wildly differing wages were paid to translators, with little connection to their ability. Several translation agencies were caught up in a feverish battle to handle most of the work and desperately competed to find translators.
No one knows the quality of the final product, but it cannot have been routinely high. Some translators and agencies have still not been fully paid. As the deadline drew closer, more and more boxes of documents appeared. And as the final blow, the opposing company's law firm also came onto the scene with boxes of its own documents that needed translation. But these newcomers imposed one nearly impossible condition, also for reasons of confidentiality: no one who had translated for the first law firm would be permitted to translate for them.
Now let us consider this true-life tale, which occurred just three months ago, and see how—or whether—MT could have handled things better, as is sometimes claimed. Let's be generous and remove one enormous obstacle at the start by assuming that all these cases of documents were in fact in machine-readable form (which, of course, they weren't). Even if we accord MT this ample handicap, there are still a number of problems it would have had trouble coping with:
1. How could a sufficient number of competent post-editors be found or trained before the deadline?
2. How could a sufficiently large and accurate MT dictionary be compiled before the deadline? Doesn't creating such a dictionary require finishing the job first and then saving it for the next job, in the hope that it will be similar ?
3. The simpler Mom & Pop store & smaller agency structure of the human translation world was nonetheless able to field at least some response to this challenge because of its large slack capacity. Would an enormously powerful and expensive mainframe computer have the same slack capacity, i.e., could it be kept inactive for long periods of time until such emergencies occurred? If so, how would this be reflected in the prices charged for its services?
4. How would MT companies have dealt with the secrecy requirement, that translation must be done in the law firm's office?
5. How would an MT Company comply with the demand of the second law firm, that the same post-editors not be used, and still land the job?
6. Supposing the job proved so enormous that two MT firms had to be hired—assuming they used different systems, different glossaries, different post-editors, how could they have collaborated without creating even more work and confusion?
Larger Philosophical Questions
Is it in any final sense a reasonable assumption, as many believe, that progress in MT can be gradual and cumulative in scope until it finally comes to a complete mastery of the problem? In other words, is there a numerical process by which one first masters 3% of all knowledge and vocabulary building processes with 85% accuracy, then 5% with 90% accuracy, and so on until one reaches 99% with 99% accuracy? Is this the whole story of the relationship between knowledge and language, or are there possibly other factors involved, making it possible for reality to manifest itself from several unexpected angles at once. In other words, are we dealing with language as a linear entity when it is in fact a multi-dimensional one?
Einstein maintained that he didn't believe God was playing dice with the universe. Is it possible that by using AI rule-firing techniques with their built-in certainty and confidence values, computational linguists are playing dice with the meaning of the that universe?
It would be possible to design a set of "Turing Tests" to gauge the performance of various MT systems as compared with human translation skills. The point of such a process, as with all Turing Tests, would be to determine if human referees could tell the difference between human and machine output. All necessary safeguards, handicaps, alternate referees, and double blind procedures could be devised, provided the will to take part in such tests actually existed. True definitions for cost, speed, accuracy, and post-editing needs might all have at least a chance of being estimated as a result of such tests. What are the chances of their taking place some time in the near future?
"Computerization is the first stage of the industrial revolution that hasn't made work simpler." Does this statement, paraphrased from a book by a Harvard Business School professor, (2) have any relevance for MT? Is it correct to state that several current MT systems actually add one or more levels of difficulty to the translation process before making it any easier?
While translators may not be able to articulate precisely what kind of interface for translation they most desire, they can certainly state with great certainty what they do NOT want. What they do not want is an interface that is any of the following:
harder to learn and use than conventional translation;
more likely to make mistakes than the above;
lending less prestige than the above;
less well paid than the above.
Are these also concerns for MT developers?
What real work has been done in the AI field in terms of treating translation as a Knowledge Domain and translators as Domain Experts and pairing them off with Knowledge Engineers? What qualifications were sought in either the DE's or the KE's?
Are MT developers using the words "asymptote" and "asymptotic" in their correct mathematical sense, or are they rather using them as buzzwords to impart a false air of mathematical precision to their work? Is the curve their would-be asymptote steadily approaching a representation of FAHQT or something reasonably similar, or could it just turn out to be the edge of a semanto-linguistic Butterfly Effect drawing them inexorably into what Shannon and Weaver recognized as entropy, perhaps even into true Chaos?
Must not all translation, including MT, be recognized as a subset of two far larger sets, namely writing and human mediation? In the first case, does it not therefore become pointless to maintain that there are no accepted standards for what constitutes a "good translation," when of course there are also no accepted standards for what constitutes "good writing?" Or for that matter, no accepted standards for what constitutes "correct writing practices," since all major publications and publishing houses have their own in-house style manuals, with no two in total agreement, either here or in England. And is not translation also a specialized subset of a more generalized form of "mediation," merely employing two natural languages instead of one? In which case, may it belong to the same superset which includes "explaining company rules to new employees," public relations and advertising, or choosing exactly the right time to tell Uncle Louis you're marrying someone he disapproves of?
Are not the only real differences between foreign language translation and such upscale mediation that two languages are involved and the context is usually more limited? In either case (or in both together), what happens if all the complexities that can arise from superset activities descend into the subset and also become "translation problems?" at any time? How does MT deal with either of these cases?
Does the following reflection by Wittgenstein apply to MT: "A sentence is given me in code together with the key. Then of course in one way everything required for understanding the sentence has been given me. And yet I should answer the question `Do you understand this sentence?': No, not yet; I must first decode it. And only when e.g. I had translated it into English would I say `Now I understand it.'
"If now we raise the question `At what moment of translating do I understand the sentence? we shall get a glimpse into the nature of what is called `understanding.'" To take Wittgenstein's example one step further, if MT is used, at what moment of translation does what person or entity understand the sentence? When does the system understand it? How about the hasty post-editor? And what about the translation's target audience, the client? Can we be sure that understanding has taken place at any of these moments? And if understanding has not taken place, has translation?
Practical Suggestions for the Future
1. The process of consultation and cooperation between working translators and MT specialists which has begun here today should be extended into the future through the appointment of Translators in Residence in university and corporate settings, continued lectures and workshops dealing with these themes on a national and international basis, and greater consultation between them in all matters of mutual concern.
2. In the past, many legislative titles for training and coordinating workers have gone unused during each Congressional session in the Department of Labor, HEW, and Commerce. If there truly is a need for retraining translators to use MT and CAT products, it behooves system developers—and might even benefit them financially—to find out if such funding titles can be used to help train translators in the use of truly viable MT systems.
3. It should be the role of an organization such as MT Summit III to launch a campaign aimed at helping people everywhere to understand what human translation and machine translation can and cannot do so as to counter a growing trend towards fast-word language consumption and use.
4. Concomitantly, those present at this Conference should make their will known on an international scale that there is no place in the MT Community for those who falsify the facts about the capabilities of either MT or human translators. The fact that foreign language courses, both live and recorded, have been deceitfully marketed for decades should not be used as an excuse to do the same with MT. I have appended a brief Code of Ethics document for discussion of this matter.
5. Since AI and expert systems are on the lips of many as the next direction for MT, a useful first step in this direction might be the creation of a simple expert system which prospective clients might use to determine if their translation needs are best met by MT, human translation, or some combination of both. I would be pleased to take part in the design of such a program.
DRAFT CODE OF ETHICS:
1. No claims about existing or pending MT products should be made which indicate that MT can reduce the number of human translators or the total cost of translation work unless all costs for the MT project have been scrupulously revealed, including the total price for the system, fees or salaries for those running it, training costs for such workers, training costs for additional pre-editors or post-editors including those who fail at this task, and total costs of amortization over the full period of introducing such a system.
2. No claims should be made for any MT system in terms of "percentage of accuracy," unless this figure is also spelled out in terms of number of errors per page. Any unwillingness to recognize errors as errors shall be considered a violation of this condition, except in those cases where totally error-free work is not required or requested.
3. No claim should be made that any MT system produces "better-quality output" than human translators unless such a claim has been thoroughly quantified to the satisfaction of all parties. Any such claim should be regarded as merely anecdotal until proved otherwise.
4. Researchers and developers should devote serious study to the issue of whether their products might generate less sales resistance, public confusion, and resentment from translators if the name of the entire field were to be changed from "machine translation" or "computer translation" to "computer assisted language conversion."
5. The computer translation industry should bear the cost of setting up an equitably balanced committee of MT workers and translators to oversee the functioning of this Code of Ethics.
6. Since translation is an intrinsically international industry, this Code of Ethics must also be international in its scope, and any company violating its tenets on the premise that they are not valid in its country shall be considered in violation of this Code. Measures shall be taken to expose and punish habitual offenders.
Respectfully Submitted by
Alex Gross, Co-Director
Cross-Cultural Research Projects
(1) Kimmo Kettunen, in a letter to Computational Linguistics, vol. 12, No. 1, January-March, 1986
(2) (2) Shoshana Zuboff: In the Age of the Smart Machine: The Future of Work and Power, Basic Books, 1991.