Lantra-l file library
Lantra-l home page

Can humans help MT?

 

How might computational linguists apply the techniques of professional translators and interpreters to a computer? What advice would translators and interpreters give a translation machine? A message thread from LANTRA-L (continuously updated).


Date: Fri, 14 Nov 1997 17:01:36 -0500
From: Daniel Loehr <loehrd@GUSUN.GEORGETOWN.EDU>
Subject: [LANTRA-L] Can humans help MT?

Hello,

I write this on behalf of fellow students and faculty in Machine Translation (MT) at Georgetown University. We have searched the LANTRA archives, but have been unable to find any discussion related to our question below.

It is well known that Machine Translation produces nowhere near the quality of translations as humans do. Perhaps this is because researchers in MT have looked at the issue as a computer science problem, applied to translation. What if we were to look at it the other way around - as a human translation problem, applied to computer science?

Have MT researchers ever consulted professional interpreters and translators to find out how they do what they do, and then applied those techniques to their algorithms? If so, we'd be interested in their work.

We suspect that the answer is no. This, then, is the purpose of this posting. If you are a professional interpreter or translator, how _do_ you do what you do?

Are there at times fairly mechanical word or phrase mappings that you make from one language to the other? Is it more intuitive and abstract than this? Do you rely heavily on context? Are we even asking the right questions? (And for simultaneous interpreters, how do you translate a word you don't know, or supply a verb in English before you've heard it in German?)

Finally, how might we apply your techniques to a computer? To put it simply, what advice would you give a translation machine?

We realize our question is very open-ended. Comments are welcome.

Dan Loehr

------------------------------------------------------------------
Dan Loehr, PhD Candidate
Georgetown University
loehrd@gusun.georgetown.edu
www.georgetown.edu/users/loehrd
------------------------------------------------------------------


Date: Fri, 14 Nov 1997 15:48:33 -0700
From: Chris DeSantis <CCDeSantis@CHQ.BYU.EDU>
Subject: [LANTRA-L] Can humans help MT? -Reply

> what advice would you give a translation machine?

1. Work from 2:00 AM to 6:00 AM
2. Smoke like a bat
3. Read Lantra daily
4. Protect your hamsters
5. Find excuses to surf the net
6. Pay no attention to the man behind the curtain...

No, seriously, folks...

You are asking an honest question, but I don't think its the right one. The nature of translation is not something that can be distilled down into a set of procedures, or an attitude, or anything quantifiable. Chomsky, Lamb, Lytle, Makkai, Pei, Sapir... not even these have been able to figure out what language _is_, let alone how to get from one to another algorithmically except at the most restrictive levels.

If an MT algorithm has 15 steps, you will take 7 years to figure out that each step needs to be performed before all the others in order to produce acceptable output. Transfer and synthesis are pretty brainless operations, and with sufficient refinement time can cover a large percentage of the type of language most encountered in commercial translation. It's the analysis that's the killer, and because language is so *spiritual* in nature (not meant in the metaphysical sense, but in the sense of being inquantifiable) no machine can hope to do what a human does.

We translate the way we do because of what we have learned, and how we learned it, and whether we were eating madeleines with tea when we learned it, and what the morning dew rising off the fresh-cut hay in the north forty of grandpa's farm when we were visiting him when we were eight years old smelled like when we learned it, and so on. Every translator will give you a different answer, each as unencodable as the mercuriality of language itself.

Now, I am not implying that you should give up in your quest. I spent years of my life writing MT code, and have seen systems that are of great value _within certain limited spheres_. These programs have much room for improvement and will doubtless become more valuable as time goes on. But I don't think humans can help MT in the way that you hope.

Respectfully,

----------------------------------------------------

I bhfad uainn go le/ir an drochrud...

Christopher C. DeSantis
ccdesantis@chq.byu.edu


Date: Fri, 14 Nov 1997 23:48:43 GMT
From: Helge Niska <Helge.Niska@TOLK.SU.SE>
Subject: Re: [LANTRA-L] Can humans help MT?

Daniel, you raised a lot of good questions, and I'll comment on some of them from my back-ground as an interpreter. I hope you get lots of answers from other people too.

>It is well known that Machine Translation produces nowhere near the
>quality of translations as humans do. Perhaps this is because
<researchers in MT have looked at the issue as a computer science
>problem, applied to translation. What if we were to look at it the
>other way around - as a human translation problem, applied to computer
>science?

Are we talking about the same problem? If it is a question of quality, then we should define what we mean by quality. I am quite sure that in many circumstances even a bad translation is better than no translation at all.

And fully-automatic MT is largely used, as far as I know, to produce "draft" translations (Japan, EU /Systran/). This may be quite satisfactory for a decision-maker or a researcher who needs the information _now_, not in a couple of days. And this applies to humans as well as machines: a simultaneous interpreter doesn't have time to ponder about terminology or subtle nuances; if they don't come up at once, he'll have to do with something less perfect. NB. that a human translator's "drafts" would in most cases not look the same as a machine's drafts; the "errors" are different.

A great problem here, I think, is that we as translators and you as computational linguists do not share the same understanding of 1. translation 2. computing. But I guess that if the translators could be made to just forget what they "know" about computers and MT, _and vice versa_, we could have a fruitful discussion. One of the obstacles to a sound theory of translation has been the erroneous idea of communication as the sending and reception of "messages", sometimes via some conduit in the form of a translator. I wonder if computational linguists still believe it works like that ;-)

>Are there at times fairly mechanical word or phrase mappings that you
>make from one language to the other? Is it more intuitive and abstract
>than this? Do you rely heavily on context? Are we even asking the right
>questions? (And for simultaneous interpreters, how do you translate a
>word you don't know, or supply a verb in English before you've heard it
>in German?)

Educated guesses :-) Russian interpreter/researcher Gelly Chernov talks about the necessary "redundancy", without which we cannot interpret. Redundancy in this context means having back-ground knowledge about the speaker, the situation and setting, the discourse styles or genres used, etc. but also the redundancy in the discourse of the speaker.

Interpreting is always done in a unique situation; i.e. context is crucial for the strategy and the techniques, linguistic solutions etc. you use when you interpret. This means, e.g. that a term or a phrase that you use in one situation may not be suitable at all in other circumstances. It is not even certain that you do the same things in different situations: when you interpret in the court, you have to be careful with the words and stick to a "linguistic" interpreting style; interpreting in a hospital will often entail more--like showing empathy, sitting and talking with the patient, helping them to get a taxi etc. So this is another distinction we have to make: MT will only take care of the "linguistic" part of translation. Including the lingustic "redundancy" to use Chernov's term.

Helge Niska
Institute for Interpretation and Translation Studies, Stockholm University, Sweden


Date: Fri, 14 Nov 1997 23:45:46 +0000
From: Michael Benis <michaelb@PAVILION.CO.UK>
Subject: Re: [LANTRA-L] Can humans help MT?

At 17:01 14/11/97 -0500, you wrote:

>Hello,

>I write this on behalf of fellow students and faculty in Machine
>Translation (MT) at Georgetown University. We have searched the LANTRA
>archives, but have been unable to find any discussion related to our
>question below.
>
>It is well known that Machine Translation produces nowhere near the
>quality of translations as humans do. Perhaps this is because
>researchers in MT have looked at the issue as a computer science
>problem, applied to translation. What if we were to look at it the
>other way around - as a human translation problem, applied to computer
>science?
>
>Have MT researchers ever consulted professional interpreters and
>translators to find out how they do what they do, and then applied those
>techniques to their algorithms? If so, we'd be interested in their
>work.
>
>We suspect that the answer is no.

I'm afraid that they're really quite professional and the answer is in fact, yes (whatever you think of the results).
Unfortunately you've just missed their annual get-together.

cheers

Michael

Michael Benis BA (Hons)
MITI Approved Police Interpreter
Freelance Copywriter, Interpreter and Translator French / Italian / US -> UK English michaelb@pavilion.co.uk
http://www.pavilion.co.uk/users/michaelb
http://ourworld.compuserve.com/homepages/michaelbenis
Tel: + 44 (0)1273 562118
Tel/Fax: + 44 (0)1273 299664
Mobile: + 44 (0)370 670798


Date: Sat, 15 Nov 1997 00:55:06 +0100
From: Roger Hughes <roger.hughes@SKYNET.BE>
Subject: Re: [LANTRA-L] Can humans help MT?

At 17:01 14/11/97 -0500, Daniel Loehr wrote:

>Hello,
>
>I write this on behalf of fellow students and faculty in Machine
>Translation (MT) at Georgetown University. We have searched the LANTRA
>archives, but have been unable to find any discussion related to our
>question below.
>
>It is well known that Machine Translation produces nowhere near the
>quality of translations as humans do. Perhaps this is because
>researchers in MT have looked at the issue as a computer science
>problem, applied to translation. What if we were to look at it the
>other way around - as a human translation problem, applied to computer
>science?
>
>Have MT researchers ever consulted professional interpreters and
>translators to find out how they do what they do, and then applied those
>techniques to their algorithms? If so, we'd be interested in their
>work.

This seemed to be the general thrust of a number of people's work when I was studying in the Centre for Computational Linguistics (now the Department of Language Engineering) at UMIST in the UK a few years ago.

(The much-loathed Systran, for example, was generally condemned as being a computer scientists's attempt to solve a linguist's problem). I can't quote anything particularly reliable, unfortunately. Possibly the thrust was more towards linguisticists rather than translators per se setting the problems (the senior translation theorist at the centre seemed to me to take too abstract an approach to provide any practical input.) Contacts there (which you probably already have) can be found through http://www.ccl.umist.ac.uk/ (They won't remember me; I sort of failed a masters there a while back.)

>We suspect that the answer is no. This, then, is the purpose of this
>posting. If you are a professional interpreter or translator, how _do_
>you do what you do?
>
>Are there at times fairly mechanical word or phrase mappings that you
>make from one language to the other?

Inevitably, particularly when I do an initial translation, and particularly when working as a staff translator (ie lots of similar documents with some formulaic sections). They may well get changed on a second pass, of course.

> Is it more intuitive and abstract than this?

Often.

> Do you rely heavily on context?

Certainly, although "context" covers a wide range of concepts, from simple resolution of anaphora to highly familiar subject matter which makes it feasible for me to carry out some effective translation from SLs I wouldn't think of working from in other circumstances; for example simple factual knowledge can enable one to identify multi-word lexical items in a semi-unfamiliar language which one might otherwise attempt to render word by word. (Not a practice I'd recommend without a friendly revisor handy, mind you, and I'm only talking about e.g. working from Spanish where my usual SLs are Italian, French and Portuguese ...)

Are we even asking the right
>questions?

Not my problem ;-)

>Finally, how might we apply your techniques to a computer? To put it
>simply, what advice would you give a translation machine?

My personal feeling - purely a feeling - is that the breakthrough in MT to some sort of real general purpose functionality will basically be a matter of being able to crunch a big enough corpus fast enough - essentially an extension of the translation memory paradigm; although I'm very fuzzy about the processes that one would apply to extract useful information from a corpus (throw it at a neural network and see what sticks?). The only problrm with such a system as far as I can see is the impossibility of formally proving its accuracy; important translations would still require human assessment before they could be considered authoritative. More pertinently to you, I'm sure that attempting to follow a human model exactly would be a futile exercise, but there must be some useful information to be gleaned if you can find some sufficiently self-aware translators (I'm not even sure I could simultaneously translate and observe how I was doing it, particularly without the observation changing the (reasonably successful) modus operandi. Maybe I'll try, though).

Roger
-

Roger Hughes, Brussels
phone +32 2 673 66 48 - fax +32 2 675 77 47
<roger.hughes@skynet.be>
<http://users.skynet.be/rohughes>


Date: Sat, 15 Nov 1997 02:09:46 +0100
From: Werner Richter <wernr@XPOINT.AT>
Subject: Re: [LANTRA-L] Can humans help MT?

>If you are a professional interpreter or translator, how _do_
>you do what you do?
>
>Are there at times fairly mechanical word or phrase mappings that you
>make from one language to the other? Is it more intuitive and abstract
>than this? Do you rely heavily on context? Are we even asking the right
>questions? (And for simultaneous interpreters, how do you translate a
>word you don't know, or supply a verb in English before you've heard it
>in German?)
>
>Finally, how might we apply your techniques to a computer? To put it
>simply, what advice would you give a translation machine?
>
>We realize our question is very open-ended. Comments are welcome.

Dear Dan Loehr --

Good questions, most of them, but still not the right ones I fear. They are good because they point to the problem. But they are wrong because they don't begin to solve it.

As Chris de Santis pointed out, the formation of language utterances and also their "mapping" in another language are highly idiosyncratic processes.

Consider the following fact: practically any given meaningful sentence of more than 4-5 words can be translated in several, sometimes many, different ways without changing significantly in meaning. On the other hand, there are countless way to get it all wrong though on some superficial level the sentence may even sound "correct".

The solution will be a serious upgrade of memory capacity _plus_ some way to short-input those 20+ years of acquiring cultural (and other) knowledge, in short becoming a functional brain. World knowledge, street wisdom, eejit erudition, life experience, IQ-type intelligence, instant learning, news perception, raw data awareness, you name it.

I appreciate your efforts but to say the truth, this is only because I believe they are doomed to failure. Because in the final count you're tapping our brains of course. And the world being the one it is, these brains are our livelihood. So please include an algorithm for automatic cash transfer to each living translator in your MT program: my bank account #: 917 048 282.

You see, we'll all end up making fun of it. But not all the time, so you'll just have to bear with us :)

neuro-logically

WernR

------------> Xlations Eng>Ger mailto:wernr@xpoint.at --------------->


Date: Sat, 15 Nov 1997 14:40:50 +0100
From: "Meertens, Rene (HDS-TRE)" <REM@WHO.DK>
Subject: Re: [LANTRA-L] Can humans help MT?

In a text, there are expressions which have to be translated in one particular way. Then you translate the other words according to context.

Example:

The European Network of Health Promoting Scools (ENHPS) was set up to establish, throughout the European Region of the World Health Organization, a group of model schools that would demonstrate the impact of health promotion in the school setting.

Set expressions:

European Network of Health Promoting Scools: Reseau europeen d'ecoles-sante
European Region: Region europeenne
World Health Organization: Organisation mondiale de la sante
Words translated according to context:

set up - possible translations:
1 installer, monter, mettre en place, eriger. 2 mettre en place. 3 constituer, creer, instituer, creer, monter, fonde, constituer. 4 s'etablir. 5 declencher, provoquer. 6 ouvrir, diligenter. 7 invoquer.

establish: possible translations:
1 creer, instituer, fonder, constituer, edifier, eriger, etablir, mettre en place, implanter, asseoir, mettre sur pied, organiser. 2 etablir, instituer, poser, fixer, arreter, instaurer, introduire; (that) stipuler que, disposer que. 3 etablir; constater; montrer, demontrer, prouver, attester, justifier de. 4 titulariser.

demonstrate: possible translations:
1 demontrer. 2 faire preuve de. 3 faire la demonstration de. 4 prouver l'existence de, mettre en evidence. 5 manifester.

impact: possible translations:
1 (force d') impact, choc. 2 impact. 3 effet(s), consequences, repercussions, incidences, contrecoup, retombees, impact, poids, influence, efficacite, retentissement, effets, resultats (des activites), portee.

setting: possible translations:
1 cadre, environnement; lieu, situation. 2 fixation. 3 groupement selon les aptitudes. 4 configuration; reglages.

Rough draft:
Le Reseau europeen d'ecoles-sante a ete cree en vue de mettre en place, dans l'ensemble de la Region europeenne de l'Organisation mondiale de la sante, un groupe d'ecoles modeles qui demontreraient les effets de la promotion de la sante dans le cadre scolaire.

Edited draft:
Le Reseau europeen d'ecoles-sante a ete cree en vue de mettre en place, dans l'ensemble de la Region europeenne de l'Organisation mondiale de la sante, des ecoles modeles qui demontreraient les effets d'actions de promotion de la sante dans le cadre scolaire.

Your questions:

Are there at times fairly mechanical word or phrase mappings that you make from one language to the other?

Yes

Is it more intuitive and abstract than this?

Most of the time, yes

Do you rely heavily on context?

Yes, heavily.

Are we even asking the right questions?

...

Finally, how might we apply your techniques to a computer? To put it simply, what advice would you give a translation machine?

Before you translate, you should learn to write.

Regards

Rene Meertens
rem@who.dk
http://www.angelfire.com/ak/Guideenfr/


Date: Sat, 15 Nov 1997 22:53:38 +0900
From: Rene von Rentzell <rrr@TWICS.COM>
Subject: Re: [LANTRA-L] Can humans help MT?

Dan Loehr:

>Are there at times fairly mechanical word or phrase mappings that you
>make from one language to the other? Is it more intuitive and abstract
>than this? Do you rely heavily on context? Are we even asking the right
>questions?

Right questions, I'd say, but still the wrong conversation. MT is simply not able to deal with the unfamiliar and with input errors the way humans are. And as for repetitive and and standardized text, where MT can function, it is used already.

>(And for simultaneous interpreters, how do you translate a
>word you don't know, or supply a verb in English before you've heard it
>in German?)

A famous headache. Ideally, your time lag is long enough to catch the German verb, or at least to get so close to it that you can guess it.

If it isn't, you have to fudge something vague and hope it isn't all too wrong. And if it turns out to be the exact opposite, make some hasty correction and duck for cover.

>Finally, how might we apply your techniques to a computer? To put it
>simply, what advice would you give a translation machine?

Considering that if you actually managed to create perfect MT we'd all be unemployed practically over night, I am not sure how enthusiastic the response is you'll get for asking this question in a translator's list, but let's see :-)

-- Rene <rrr@twics.co.jp>

-- Support the anti-spam amendment. Join at http://www.cauce.org/ --


Date: Sat, 15 Nov 1997 09:38:12 -0500
From: "Asselin, Michelle" <masselin@ROCLER.QC.CA>
Subject: Re: [LANTRA-L] Can humans help MT?

At 17:01 14/11/1997 -0500, Dan Loehr wrote:

>Finally, how might we apply your techniques to a computer? To put it
>simply, what advice would you give a translation machine?

Get yourself a good human interface.

MT can be of great help for some types of texts -- mostly technical and repetitive -- and speed the translation process, but there is still the need for a flesh-and-blood translator to sit in front of the computer and "educate" the program, and later to edit the output. And just like humans continue to learn throughout their life...

Just my $0.02 addition to the comments others have already given you.

=>=>=>=>=>=>=>=>=>=>=>=>=>=>=>=>=>=>=>=><=<=<=<=<=<=<=<=<=<=<=<=<=<=<=<=<=<= Pour communiquer efficacement... Des mots qui traduisent vos idees.

Michelle Asselin, trad. a. / Membre de l'OTIAQ Montreal (Quebec) T: +514-767-7626 F: +514-767-2111 http://www.rocler.qc.ca/masselin/masselin.htm E-mail : masselin@rocler.qc.ca


Date: Sat, 15 Nov 1997 09:29:40 -0500
From: John Lejderman <prems@TOTAL.NET>
Subject: Re: [LANTRA-L] Can humans help MT?

Very good questions, and difficult to provide a short answer. However the use of context is very important. Often an isolated sentence makes no sense at all. Only by understanding the whole context, which may go as far as understanding the content of a 200 page book, is it then possible to "make up" a meaning for this sentence, guessing at what the author may have meant. Computers will have to achieve human level intelligence before they can achieve this. It is not a machine translation problem, but a machine intelligence problem. Good luck in your work! Don't be too much in a hurry.

Solving this problem could take a little while. I sympathize with you. I did my Master's degree in artificial intelligence.

- John Lejderman

French to English translation, specializing in information technology. English web site : www.total.net/~prems Site Web en français : www.total.net/~prems/traduction.htm E-mail : prems@total.net
14 chemin Sonny St-Sauveur-des-Monts, Québec Canada,J0R 1R2 (514) 227-6795


Date: Sat, 15 Nov 1997 11:05:32 -0600
From: Alex Lane <alexlane@EARTHLING.NET>
Subject: Re: [LANTRA-L] Can humans help MT?

At 05:01 PM 11/14/97 -0500, you wrote:

>Have MT researchers ever consulted professional interpreters and
>translators to find out how they do what they do, and then applied those
>techniques to their algorithms? If so, we'd be interested in their
>work.

I do not know if this has been done, but it's the classic baselining strategy for building an expert system. However, expert systems have only enjoyed success (and rarely at that) in cases where the domain of the expertise being developed is rather narrow and where experts can explain why they do the things they do.

>We suspect that the answer is no. This, then, is the purpose of this
>posting. If you are a professional interpreter or translator, how _do_
>you do what you do?

The problem here (as it is with most attempts to build expert systems) is that domain expertise generally has become so ingrained that any attempt to "explain" it is futile: any attempt to analyze the process kills the process, while any attempt to explain a completed process often cannot be explained by the expert (or worse, the expert feels obligated to fudge an explanation).

Personally, I read (or hear) stuff, grok it, and type out (or speak) what I grokked, with the awareness that the output is in a different language. I personally feel that many years of typing stuff out for publication (and speaking in public) in my native language, i.e., way before I undertook translation and interpretation professionally, as well as having a high command of my "non-native" language (to the point where I only rarely have to consciously dope out the grammatical sructure of an utterance), has been of inestimable help. So perhaps the key to your investigation is the answer to the following question: "How do people decide to use the words they do to express the thoughts they develop?" The language aspect is, I think, secondary.

>Finally, how might we apply your techniques to a computer? To put it
>simply, what advice would you give a translation machine?

The key question to ask is what advice CAN one give a computer? Recall that MYCIN (one of the first expert systems) was actually quite competent at diagnosing and prescribing treatments for bacterial infections of the blood.

If one *could* formulate a query to MYCIN whose point was to fix a broken bicycle, the program would start by asking for the results of the blood work.

Similarly, IBM's "Deep Blue" chess computer, although capable of defeating human chess grandmasters, is completely incapable of understanding tic-tac- toe. Develop a program that is capable of integrating new information into a well-connected knoweldgebase, and then we can readdress your question.

Cheers...

mailto:alexlane@earthling.net ...speaking only for myself
Seabrook, Texas http://www.ghg.net/alexlane


Date: Sat, 15 Nov 1997 15:02:08 -0500
From: Robert Paquin <jarp@AEI.CA>
Subject: Re: [LANTRA-L] Can humans help MT?

Hello to all you fleshy translators... and hi to the lurking machines,

I find that even humans need a context to translate adequately. For example. I'm presently adapting a German series for dubbing into French.

For this, I'm supplied with a first translation made from the German script. The translator has not seen the episode, she translates from a written text.

So, at one point, there are three characters discussing in an office. One of them is speaking, and there's a knock at the door. He says "Bitte".

Another character comes in. The translator missed this, because she thought the word "bitte" was somehow linked with what the fellow was saying. So, she translated it as an emphatic "S'il vous plaît" (please), which was sort of weird in the conversation. I corrected and wrote "Entrez" (Come in), which is what a person would say in this situation. Now a machine would not assess this, because the information is not even in the text.

Mechanically yours,

Robert (author of _La Conspiration des machines_, an unfinished choreographical poem)

"Fear is the darkroom in which all of your negatives are developed."


Date: Sun, 16 Nov 1997 13:29:22 +0100
From: AUSTIN communication services <austin@LOGIN.CZ>
Subject: Re: [LANTRA-L] Can humans help MT?

>Only by understanding the whole context, which may go as far as
>understanding the content of a 200 page book,

That's the real problem, isn't it? The closer you look at "context" the bigger it grows, finally encompassing the whole of human experience, synchronic and diachronic, for at least one language community. To really grasp, to understand in an organic way what an author "means" by a particular word, sentence, paragraph, book.... you have to be familiar with the entire discourse in which that text is situated. Not really the sort of thing computers are good at.
The major fallacy plaguing both MT and most laypeople's conception of translating is that language is simply a "code" for independently existing "messages" which can be "decoded" and transfered to another code. It just ain't so - the language is an inextricable part of any utterance, and the translator's job is not to extract meaning and recast it, but to create a new, functional text in the target language. The function of the translation need not even be the same as that of the original source text, in fact it often isn't. Even in the most "technical" of texts, things like, say, the social status of a given occupation in the language community or the conventions of expression for a given text genre (e.g. assembly instructions) can have an enormous effect on the composition of a document. These are things that a translator needs to be aware of (but not follow slavishly), and that require constant use of such hard-to-pin-down human faculties as intuitive judgment. I think there is a real danger that MT *will* become successful, at least commercially, and that we will increasingly be exposed to computer-generated translated texts. Over time we will come to accept machine output as normal, just as we have come to accept so many synthetic products in all facets of life, and will start writing, and maybe thinking, like machines ourselves.

Michael


Date: Sun, 16 Nov 1997 22:03:33 +0100
From: Aquitaine Traduction <Aquitaine.Traduction@AQUITAINE-TRADUCTION.COM> Subject: Re: [LANTRA-L] Can humans help MT?

Michael Austin wrote:

> I think there is a real danger that MT *will* become successful, at least
> commercially, and that we will increasingly be exposed to
> computer-generated translated texts. Over time we will come to accept
> machine output as normal, just as we have come to accept so many synthetic
> products in all facets of life, and will start writing, and maybe thinking,
> like machines ourselves.

I believe that two new skills will need to develop:

- dictating translations
- rewriting machine generated translations

As to the first, this will surely take much getting used to, and may even be impractical for many types of translation. But my money is on it, and I'm going to take the plunge with the new Via Voice. Sure, I'll have to change my work methods, but I'm not necessarily against that.

As to the second, machine translation *might* take a lot of the repetitive, monotonous and bullshitty side out of many kinds of translation. At the rate improvements are made, I think that most translators will be getting into it in short order. There may even be one good outgrowth of MT: it could just eliminate a lot of the people who provide black market translations. You never know...

As for making us more artificial and machine-like, I really don't see it. I see the primary revolution as follows:
At the present time, most translations involve:
- the translator
- his text (printed and/or electronic)

Translations of the future will be done using:
- the translator (we're here to stay, guys)
- the source text
- the MT text to correct and complete

I think that once people slip into the newer method, that it will be no more inhuman than a "straight" translation. It will just require a different sort of know-how.

In the future, instead of quoting for "1 257 words in Word", translation professionals will make bids for "1 257 words in Word put through the ABC machine translation program".

As opposed to MicroSoft* software, there will be competition for MT programs, and their usefulness will rely both on their intrinsic merits, and translators' mastery of them.

Best regards, Alex R.

* Yes, I *know* that MicroShit bought 20 or 30% of Trados, but they can't corner that market too - Can they ?


Mon, 17 Nov 1997 12:19:26 +0100
From: AUSTIN communication services <austin@LOGIN.CZ>
Subject: Re: [LANTRA-L] Can humans help MT?

>I think that once people slip into the newer method,
>that it will be no more inhuman than a "straight"
>translation. It will just require a different sort of
>know-how.

My comments about the danger of our thinking becoming more machine-like as we come to accept machine-generated translations (and maybe texts other than translations? I already shunt different pseudo-customized sections into our sales letters according to various data on the prospect) was not only intended in reference to translators, but to everyone who reads - maybe especially to those who read mostly the marketing & advertising texts that even "non-readers" can hardly avoid unless they're actually illiterate. Just as adspeak has wormed its way into the vocabulary and thinking style of people who should know better, machinespeak will start to influence those who don't pay much attention to words and language and percolate its way up (or down, or sideways, or whatever) to those of us who do.

Of course none of this affects your argument that machine translation or computer-assisted translation is potentially a valuable tool for flesh-and-blood translators.

If MT products achieve commercial success, one thing that will greatly influence their effect on human translators' business is whether the output will be abysmally bad enough to draw attention to translation quality issues that all too many clients are only vaguely aware of.

Michael


Lantra-l file library
Lantra-l home page

1