Jost Zetzsche interviews Prof. Alan K. Melby

Jost: Alan, thanks for taking the time to answer some questions. In preparation for this conversation, I looked over your CV and list of publications as well as some things that others have said about you. And: wow! You are a true pioneer for many of the things that we translators today completely take for granted. If you had to choose just one of your contributions that you feel has had the greatest impact on the world of translation, what would it be?

AKM: The biggest impact so far? The translator workstation, which can be thought of as a combination of hardware (a personal computer) and software (what you first termed, and now many others call, a TEnT [Translation Environment Tool]).

Jost: In the article “The origins of the translator’s workstation,” [1] John Hutchins credits you with coming up with the concept of bilingual concordancing and using aligned texts for translation purposes as well as the idea of the “translator’s workstation” with remote access to termbases. All this at a time when there were no personal computers.

AKM: In the late 1970s, the idea of the translator workstation was in the air. Martin Kay was writing his influential paper “The Proper Place of Men and Machines in Translation” (which was not formally published until 1997 [2]). I was working with colleagues at Brigham Young University on ideas that were later used at ALPS as the basis for the first commercial translation memory system. So although I was perhaps the first person to use the term “translator workstation”, I was not the only person in the 1970s working on devices to help human translators become more productive.

It is true that there were no IBM personal computers in the late 1970s. The first IBM PC was sold in 1981. But before that I was working on a Z-89 microcomputer that was in some ways a precursor to the IBM PC, and I used it as a workstation in a distributed processing accounting system I designed.

Why do some consider me to be the father of the translator workstation? Perhaps it is because I kept at it despite public criticism. For example, I remember talking with Serge Pershke (director of the Eurotra machine translation project) in July 1984 on the campus of Stanford University at a computational linguistics conference. Serge said to me something like this, “Professor Melby, you are wasting your time developing translator workstations. Within five years there won’t be any more human translators.”

Also in 1984, Peter Wheeler made fun of me in a presentation at Aslib TC6, the 6th edition of the Aslib Translating and the Computer conference [3], mocking the Melby translator workstation as having 96 windows, each the size of a postage stamp.

A 1985 report on Eurotra [4] was still quite optimistic, but I stood my ground and kept predicting that human translators would be around for a very long time and that effort should go into translator workstation development as well as machine translation. By 1991 [5] expectations for Eurotra were much more realistic, and, amazingly (if you bought into the early Eurotra optimism), human translators were still around.

Jost: Looking at today’s translation tool landscape, do you feel that your vision has become reality?

AKM: Only partially. Back in 1982, during a presentation in Prague [6], I introduced the notion of three levels of assistance that a translator workstation could provide to a human translator. Two of those three levels have become a reality. The first level was assistance that can be provided without the source text being available in machine-readable form. Nearly all translators now use technology at this level, such as word processing and dictionary or termbase lookup, as well as research using Internet search engines (although the Web did not become available to the general public until the early 1990s, so I obviously did not specifically mention Web search engines in 1982).

The second level of assistance adds what can be done when the source text is in machine readable form. This level has blossomed into the myriad of translation-memory tools currently on the market.

The third level brings in machine translation, but with a twist. In 1982 I foresaw and rejected the kind of integration of machine translation and translation memory that we now have: if a particular segment is not found in a translation-memory database, a machine translation of that segment is provided. I predicted back then that for a very long time it would often be a waste of time to try to fix a bad machine translation. Jost, you pointed out in your Big Wave interview that now, thirty years of MT research later, this is still the case. What I asked machine translation researchers to focus on was to develop sophisticated automatic procedures for quality estimation, so that the human translator would only see a machine translation of a segment when it was likely to be a good translation. Fortunately, the question of automatic self-assessment of machine translation quality is now getting some attention. There was a 2012 workshop emphasizing this topic. Much of my original vision of the translator workstation will have been fulfilled when quality estimation is accurate and implemented in translator tools.

However, there will still be two essential missing elements that were in my early 1980s vision: (a) widely implemented standards for interoperability and (b) linguistic processing. There is essentially no language-specific linguistic processing in today’s tools. For example, most tools do not even know the morphology of the languages they deal with. They treat all languages the same. They can’t even find the base form of an inflected word. This is technology that has been around for a long time that needs to be implemented. Martin Kay’s 1980 pamphlet about translators and technology mentioned morphological processing. That was over thirty years ago.

Also, today’s tools are not even close to being sufficiently interoperable.

Jost: On the topic of interoperability, you also have been instrumental in shaping the terminology exchange standard TBX (ISO 30042). What are the areas in which terminology exchange and terminology maintenance have to improve for the average translator?

AKM: Despite all these years of working on TBX, terminology exchange and maintenance are far from where they need to be. In the next version of TBX (which we are actively working on), there will likely be an emphasis on some very well-defined dialects of TBX that are easier to implement than the full standard. Once we have widespread implementation of at least TBX-Basic or a subset of it, along with morphological processing in translation tools, terminology management will become more important yet less visible. Adding a new concept entry to a termbase will be nearly effortless for a human translator and the containers that will hold everything to do with a translation project (see Linport below) will contain termbases that are automatically consulted during initial translation and quality assurance.

On a philosophical note, terminology within a well-defined technical domain acts as if there were language-independent concepts because domain experts around the world agree to ignore culture and create such concepts. This does not, however, imply that there are universal concepts underlying general language.

Jost: Back in the seventies you were deeply involved with machine translation at BYU. This came at a time when you “believed in the existence of one universal set of language-independent concepts underlying all human languages” (according to your CV).

In 1978, you “experienced an intellectual crisis; rejected universal language-independent units of meaning (superficial ambiguity) and accepted a fundamental distinction between dynamic general language (fundamental ambiguity) and frozen domain-specific language (well-defined concepts defined by convention within a particular domain).”

Did this new intellectual re-orientation make you approach machine translation differently?

AKM: Yes. This drastic change in my view of language and meaning changed everything relative to my work in machine translation. The BYU MT project was based on the assumption that static, language-independent, culture-independent sememes actually exist and can be identified and associated with the words of a sentence. The grammatical theory developed by Eldon Lytle that we were using was wonderful, but it could not compensate for the non-existence of universal sememes. Interaction between the computer and a human during the analysis phase made sense so long as you could store the interlingual representation and bring it out of storage years later to be the basis for automatic generation of yet another target-language text. The lack of a stable, language-independent interlingua meant the whole approach had to be re-thought.

Various members of the BYU team went off in different directions from that dead-end.

Steve Richardson, a colleague on the team, eventually made his way to Microsoft and developed what has become Bing Translate (which is a hybrid of rule-based and statistical machine translation). For more on Steve’s journey, see a 2003 interview in Machine Translation News International, issue 33, and his paper at the 2013 AMTA conference.

I, on the other hand, focused on tools for human translators, such as MTX, the first PC-based structured-text terminology manager, and on standards for interoperability. For example, I was co-editor with Yves Savourel on the first version of TMX. More recently, I have been involved in the Linport project to define standard containers for translation projects and tasks. You can download my Aslib TC34 Linport slides for more information.

By the way, I believe that statistical machine translation research and development is based, consciously or not, on the assumption that universal sememes do not exist.

Jost: Where do you see machine translation today? Have recent quality advances in MT changed your way of thinking regarding MT?

AKM: You can’t talk seriously about advances in quality without defining quality. Much of my recent thinking and practical work have been on the question of translation quality. In August 2012, I gave the presidential lecture at LACUS on the topic of translation quality, whether it is definable and achievable. Hopefully, the written version of that lecture will be published by the end of 2013 (I am still tinkering with it). For now, let me suggest that I am proposing a dynamic approach to translation quality based on three main factors: the usual two (source and target text) along with a third: structured translation specifications. If readers want to know about structured translation specifications, they can consult ISO/TS 11669 (published in May 2012) or look at my webpage on the topic. Structured specifications are also part of Linport.

Given a particular set of specifications, statistical machine translation systems are indeed getting better at massively parallel plagiarism of human translations. Isn’t that what the Moses SMT engine is all about?

The question is whether SMT will continue to improve at the rate it has, or whether it will plateau. I suspect this will depend on whether it is necessary to understand a source text in order to translate it. I certainly think so, and I believe you do, too, but not everyone does, and there is the unresolved question of what constitutes understanding.

Interested readers might want to look at the annotated slides from the debate between me and Daniel Marcu at the 2006 meeting of AMTA. Of note are the claims by the SMT community that machine translation will become as good as or better than professional human translation and that it is not necessary to understand the source text in order to produce a good translation of it.

Jost: You’re also a teacher at BYU. What role does teaching play for you? What are your observations about younger generations of linguists and translators? What do they think of the future of their profession?

AKM: I like my current crop of students. The future of human translation is bright.

For many years, we in the translator training business have been trying to say that it takes more than a knowledge of two languages in order to translate. In a sense, this is clearly false. If you know two languages, you can produce a translation, though perhaps not a professional translation. The emphasis should be on what it takes to be a professional, certified translator — certified by an accredited program. I hope to spend the next decade or two helping establish a world-wide system to accredit translator certification programs based on one uniform set of requirements.

Also on the future of translation, my LACUS paper will unveil the new Turing Test: the Translation Turing Test, which consists of comparing how humans and computers perform in a translation task, relative to various source texts and structured specifications.

The big recent change in my attitude toward machine translation is in why humans need not fear what will happen if a machine eventually does well on the Translation Turing Test.

I used to say we translators need not fear machine translation because it will never work very well, and proceeded to list many reasons why it will not work. My new position is that instead of throwing up obstacles we should encourage researchers to build machine translation systems that can pass the Translation Turing Test on demanding texts and specifications. Why this about-face in my position on machine translation? Because if computers pass the highest levels of the Translation Turing Test, the impact on life in general will far overshadow any impact on translators. What I am saying is that translation is so intellectually complex that if machines can do it really well, they will also be able to accomplish all other intellectual tasks that humans can perform. We will have truly achieved Artificial Intelligence in a sense that is now only science fiction.

I want to be around to experience the adventure of interacting with truly intelligent machines, just as I have fully experienced the computer revolution so far (having been born the year they built the first computer). If, on the other hand, computers are not able to pass advanced levels of the Translation Turing Test, then perhaps along the way, we will learn a lot about what it means to be human translator in particular and a human in general.

References

1. Machine Translation, vol.13, no.4 (1998), p. 287-307

2. Machine Translation, Volume 12 Issue 1/2, 1997, Pages 3 – 23, Kluwer Academic Publishers

3. Peter Wheeler “Update on machine translation” Translating and the Computer 6. 20-21 November 1984, ed. Catriona Picken (London: Aslib, 1985)]

4. “Eurotra: A Multilingual System under Development” Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985

5. “Eurotra Continues” by Andrew Joscelyne, Language Industry Monitor 4, July-August 1991

“As time passed, expectations became tempered; illusory ideals became replaced by more pragmatic goals. Luxembourg conceded that “Fully Automatic High Quality Translation” was not a reasonably attainable goal, and the true character of EUROTRA was eventually acknowledged to be in fact pre-competitive research rather than prototype development.”

6. Melby, A.K.: 1982, ‘Multi-level translation aids in a distributed system’, in Coling 82. Proceedings of the Ninth International Conference on Computational Linguistics, Prague, July 5-10, 1982, ed.J.Horecký, North-Holland, Amsterdam, pp. 215-220.