In preparation for the GALA conference in Prague, I queried members of the American Translators Association about their views and concerns regarding machine translation. The discussion, which took place on ATA LinkedIn group, is copied below. Condensed version in PowerPoint format is available here.
The views expressed in the discussion do not necessarily represent my own views; rather, I was simply providing a forum for ATA members to express theirs. The debate, titled “The Risks and Rewards of Machine Translation” and hosted by Don DePalma of Common Sense Advisory, will take place on Tuesday, May 11, from 9:00 to 10:30 (Prague time). Follow the event on Twitter at #GALAconf!
Jiri Stejskal, CETRA Language Solutions
Jiri Stejskal, President & CEO of CETRA Language Solutions
Input needed from freelancers on machine translation
Jiri, ATA’s Past President, will be presenting at the GALA conference on a panel titled “Risks and Rewards of Machine Translation,” representing translators and discussing shortcomings and risks associated with the use of MT. Please offer your insights here so that he can present your views and concerns in addition to his own observations. Thank you!
1. Tess Whitty, Swedish Translation Services
Hi Jiri! This paragraph below sums up my opinion. Machine translation can be good for repetitive, simple translations, for example a basic software interface, but it would still need editing. Some texts will never be fit for machine translations, for example legal, art, poetry, literature etc.
Computers can read it, but they just don’t get it. Regardless of advances in technology, computers cannot replace translators. Obviously, computers are capable of compiling extremely large databases and providing a translation for a given word or expression within a split second and that is a lot faster than a human brain. However, computers cannot read between the lines and interpret the shades of meaning.
I discussed this a bit in the post Why you should hire a professional translator: http://swedishtranslationservices.blogspot.com/2010/03/why-you-should-hire-professional.html
2. James Kirchner, Independent Translation and Localization Professional
From a theoretical linguist’s point of view, the risks of machine translation are relatively obvious.
1. Machine translation still cannot deal with structurally ambiguous sentences, such as these:
– “He saw that gasoline can explode.”
– “Teacher strikes idle kids” (a headline)
These are just two famous examples. Because the number of possible sentences that are structurally ambiguous is infinite, and the sentences are unpredictable, machine translation will never be able to deal with most of them. You can get one translation of the sentence that reflects how the machine has parsed it, but it may not be the appropriate one. If MT learns to offer the choice, a human will still have to make the decision.
2. MT is bad at context. I have just repeated an experiment I’ve done before in which I tried to get major MT engines online (usually Google and Babblefish) to distinguish between “bank” in the sense of financial institution and “bank” as the shore of a river. They generally retain the word for financial institution regardless of what is in the rest of the sentence. I have tried it from English to Czech and from English to German, and in both cases, I not only have people walking to their financial institution, but also fish swimming to financial institutions and rivers overflowing their financial institutions. Sensitivity can be programmed in for limited, predictable contexts, but it could still be thrown haywire by sentences like, “If you go near that bank, you’ll be in hot water,” or, “I went to the bank today and now I’m sunk.”
MT can’t determine whether or not something doesn’t make sense. I have just been working on a translation with a copy-and-paste disaster that would have stumped machine translation.
Furthermore, there can be copyright problems with the way some MT engines now approach their task. Google Translate, for example, often arrives at its translation by drawing on existing bilingual material gleaned from electronic sources. (This is why, for example, its translations of texts like “The Little Prince” are so impeccable — they are archived material translated by humans. However, I would assume that much of the material Google draws on is copyrighted, which could lead to lawsuits if incorporated into another text and published. In the Czech-to-English pair, I have often detected and had to redo fraudulent translations (not merely inaccurate or incompetent, but faked or plagiarized), some of which are nowadays drawn from material found on the web. This is considered unethical for a human to do, but it is a routine method within machine translation, and the legal implications are pretty nasty.
3. Joan Wallace, Translation Professional
I, for one, am not interested in editing machine translation. I find it very strange that anyone thinks it should be easier for a translator to understand and clean up a garbled machine translation than to work directly from the source text. And then, of course, the client expects to pay less because “most of the work is done”!
4. Eileen Hennessy, Translator-Writer-Editor, self-employed
Machine Translation is useful for providing a “rough idea” of what a text is about. Sometimes that’s all a client needs. Sometimes it’s what a client needs in order to make a decision about which documents in the pile on his desk need to be translated “for real” and which ones can be ignored. There’s a place for MT and a place for human translators…and a place for human editors of MT output, a type of work that calls for a very different skill set, and one that translators may not be very good at. (It’s a truism in the publishing industry that writers don’t make good editors.). By the way, is anyone working on the problem of, uh, payment of MT editors? of translator-editors who work with the output of CAT tools? Is it true that the livin’ is easier with a job at Pizza Hut or Taco Bell or McDonalds?
5. Jill R. Sommer, German→English Translator, Sommer Translation & Net Services
I agree with Joan. The only job/project I have quit mid-project was editing machine translation. It was too frustrating. It was a case of “garbage in, garbage out.” The client had had translators translate strings with no context and then once we saw the MT translations with the context it was quite apparent that the initial translations were not helpful at all and in many cases absolutely wrong. But since it was in the memory already the client insisted it stay as it was. It was a total nightmare.
6. Eileen Hennessy, Translator-Writer-Editor, self-employed
Re Jill’s comment Most appalling case of “client ignorance” I’ve heard of to date. It would be interesting to know why, notwithstanding the defects of the translation in memory, the client nevertheless insisted on letting it stand. I hope the client wasn’t a translation company!
7. James Kirchner, Independent Translation and Localization Professional
One little piece of evidence that MT doesn’t work well, except as a bargaining tool:
A small agency I do work for often has technical specs come in. The client will try to squeeze the price down below human subsistence level. When they didn’t like her very lowest price, they told her they were going to machine translate it. Then they came back with the next job, wanting the price reduced to practically nothing. The agency owner balked, so the client told her they want the price that low because the machine translation they’d done on the last job was “perfectly adequate”. If so, then why are they back asking for a bid on a human translation? But the pattern continues. They return again and again asking for human translation and bluffing that MT had done the job just fine when they tried it, and that therefore human translation should cost the same as MT.
8. Eileen Hennessy, Translator-Writer-Editor, self-employed
I’m caught between laughing and crying over this story (James’s). … This client has certainly crossed the line between using and abusing a tool.
9. Jiri Stejskal, President & CEO of CETRA Language Solutions
Thank you all for contributing, great stuff. Keep it coming! Jiri
10. James Kirchner, Independent Translation and Localization Professional
A reward of machine translation:
One of my students, a Russian-speaking Armenian, was introduced by her brother in the US to one of his coworkers. He didn’t know Russian or Armenian, and she didn’t know English. They used MT to get acquainted, and she says (broadly grinning), “It doesn’t do a good job, but it’s enough to get the general idea.”
After a year or two this MT-mediated correspondence culminated in a very happy marriage.
11. Olga Kellen, Research Professional, Translator, Citizen of the World
In business machine translation is okay to get a ‘rough idea’ of an email, text, website, but can’t be used as a marketing tool in a foreign language…
I wish clients could understand that clearly;-)
Here is my article on the subject with some ridiculous examples of machine translations in real estate business:
Anybody has my permission to copy the article with a link to the original – we have to educate clients on the proper use of machine translation: where one can use it and where it’s just harmful for one’s business…
12. Joan Wallace, Translation Professional
It is not only marketing where machine translation falls down. The emphasis should be on ‘rough idea’ (and even then, it may be the wrong idea). The only place I can see it is the case Eileen mentions, where they might need to sift through large amounts of material to see what actually needs to be translated.
I see the assertion, often supported by the ATA, that MT is fine as long as it’s for information only. As an into-English translator most of what I do, at least for US clients, is for information only. I generally assume that my clients need the right information and that they need it in correct and readable English. Let’s get rid of the idea that the only thing MT can’t handle is literature, or poetry, or marketing.
13. Joan Wallace, Translation Professional
I do like James’ MT love story, though. Love really does conquer all!
14. Rosana Wolochwianski, en RW Traducciones
Hi Jiri! You can read my views on MT at:
“What is really at stake with machine translation? An overview of its impact on the different stakeholders,” in The ATA Chronicle, A Publication of the American Translators Association, Vol. XXXVIII, N° 4, April 2009, pages 26-31.
“English>Spanish Translation in an MT environment”, in Multilingual Computing & Technology. September 2008 issue, pages 38-44.
Something else is coming up soon on the LTD Newsletter.
Good luck at GALA!
EN>ES Accredited Translator (ATA)
Licenciada en Traducción (UNR)
Rosario – Argentina
Director of CEIT (www.traductoresrosario.org,ar)
15. David Russi, Translator at UCAR
A major risk I see is that these tools are in the hands of people who do not understand them. Many people out there just click on a button and get a “translation,” without any idea just how bad the result truly is. The web is littered with “multilingual” sites that would be better off just staying in the source language.
I also wonder to what degree language will slowly erode. Leveraging bad translations and machine translations together will undoubtedly pollute any language… I’ve done very little post-editing, but I wonder, maybe you’ll question that strange use of a preposition the first time or twenty, but how long will it be before you stop wondering whether it actually is an error? After all, you’ve seen it enough times that it may come to seem normal.
16. Eileen Hennessy, Translator-Writer-Editor, self-employed
The bad news, David, is that the language erosion process is already well under way, for reasons having nothing to do with MT and everything to do with the fact that the Gutenberg Era has ended and we are now well into the Age of the Moving Image, which began with the invention of the “movie” in the early years of the twentieth century and reached critical mass in the closing years of that century. We read less and less, and as a result correct usage is no longer burned into our minds in the way that it used to be when reading was the only way of conveying and obtaining information.
In the Gutenberg Era, learning to meet standards of correct writing was considered important. Standards of behavior in general, including writing standards, began to become eroded in the 1970s, and this trend continues unabated today. In terms of language usage, we now see strange uses of prepositions and errors of syntax in publications that used to be models of correct use of language (I’ll mention specifically a publication that nevertheless still fancies itself to be “the newspaper of record” — The New York Times). As David notes, we may question that strange use of a preposition the first time or twenty, but how long will it be before we stop wondering whether it actually is an error. Sorry, David, that time is already upon us.
The applicable standard is now, “Do I more or less get the message that this text is trying to convey?” As long as the answer to that question continues to be “Yes,” there will be no reason to concern ourselves with the “outmoded niceties” of correct language usage (that concern is for old fogies of a bygone era). This situation will not be reversed until the erosion of language reaches the point at which the message is so badly garbled that deaths and lawsuits are the result.
In the bad old days before the advent of CAT tools and MT, terminology and turns of phrase were burned into the brains of us translators because we had to type them over and over again into our typewriters. Now that we are being converted into editors of canned translations, how will the would-be translators of the future learn to translate (assuming “real” translators will still be needed)? Directors of university translator training programs, are you thinking about this?
17. Susan Welsh, Freelance Translation and Editing Professional
A year ago, I agreed with the views of many freelancers that MT was an abomination. Now, I don’t. I think translators are doing themselves a disservice if they stick their heads in the sand about this matter, because, like it or not, MT is getting much better, very fast. That doesn’t mean that what they quaintly call HT (human translators) will become obsolete, but we humans do have to become more savvy and more closely attuned to what humans, and humans alone, can do in the domain of translating.
I have learned quite a bit more about MT in the past year, and I use it as an assist in translating. It works within my CAT tool, OmegaT, so you just hit Ctrl-M for a given segment and it inserts what Google Translate comes up with. Sometimes it is extremely helpful, sometimes it is absolutely ridiculous. If you don’t want any of what is offered, delete it. The quality depends on the language pairs, what sorts of documents exist in the corpus of bilingual translations, how much development work has gone into developing that particular language pair. If you are translating between Mongolian and Portuguese, I dare say the results will be extremely poor.
Everybody should know the distinction between rule-based MT and statistics-based MT. The former is the only one cheap enough for a freelancer to buy, and I have tried the Promt system on a 30-day trial and found it much worse than the free online Google Translate–and that’s for Russian-English, which you’d think it would be good at, given that it is a Russian company. The higher-priced models allow you to edit the dictionaries and create your own, but that is so time-consuming that it is really a full-time job in itself, not something I’d want to do. From what I’ve read, rule-based MT has gotten about as good as it is going to get. Statistics-based MT, however, is improving by leaps and bounds, with Google at the forefront. Whether Google is going to take over the world or not could be debated, but the fact is, their MT system is getting better and better.
OF COURSE machines will never understand metaphor, ambiguity, irony, all those things that make literature precious and the human mind unique. Who would ever think they would? The idea is that a human has to edit what is produced by a machine, unless it is used only for what they call “gisting.” It’s the same as with a TM–you are free to accept, reject, or edit what is offered.
Take a look at this 50-minute video from Google’s chief scientist. It’s worth your time. What you don’t know, can hurt you.
18. Andrew Bell, Translator – Scandinavian languages
Hi Jiri. This subject has been and continues to be much discussed on Watercooler. This isn’t a plug for the network, but this link will take you to the discussion: http://watercoolernetwork.com/forum/topics/machine-translation-angel-or
I have to agree with most the contributors so far; I think MT will play an increasing role in translation, particularly of repetitive texts such as manuals etc., but I can’t see it replacing humans in the near or distant future. But, and it’s a big but, translators need to raise their game – by which I mean raising professional standards, educating clients and new translators, adding value, going the extra mile etc., etc. The challenge is going to be finding a forum or venue by which experienced translators can pass on some of our acquired knowledge. Education is key.