Machine and Human Translation 4: Usable Translations

A contributor commented on one of my earlier posts as follows:

I don’t think we are looking for a 100% accuracy when we use MT software. If you reduce the effectiveness, not as low as 80%, but not so high as a human translator, I think the MT is useful enough for performing interesting tasks.

It depends what the translations are used for. You would certainly want better than, say, 90% accuracy (i.e., 10% wrong) in a legally binding contract, or an aircraft maintenance manual, or a doctor’s prescription or in court interpreting.

Talking of medical prescriptions, consider the following.

Sharif and Tse, who is with Dartmouth College in Hanover, New Hampshire, surveyed 286 pharmacies in the Bronx, New York — where 44 percent of the population speaks Spanish — about whether they provided medicine labels in Spanish to their customers who needed them. About three-quarters did so. Among these pharmacies, nearly 90 percent used computers to translate labels from English into Spanish, 11 percent used staff members, and 3 percent used professional interpreters.
Sharif and Tse then looked at 76 medicine labels they had generated using 13 of the 14 computer programs pharmacists reported using for translation.
They found that half of all the labels contained serious mistakes. Thirty-two of the labels included incomplete translations and six contained major spelling or grammatical errors.
Computer translation programs can clearly be improved, Sharif said, but this doesn't mean a human being shouldn't be checking the computer's work. Ideally, she added, pharmacies should have professional translators on staff to ensure that labels are being translated properly. “Figuring out how to pay for this,” Sharif said, “is probably something that belongs within the health reform conversations."

There Sharif has hit the nail on the head. Who’s going to pay for it? Professional Translators with pharmaceutical expertise are specialists, and they expect to be paid accordingly.

However, there’s the other side to MT. A Chinese student who came to Spain wrote to me recently that although she knows no Spanish, she had been able to obtain the information she needed for her trip and make her local bookings thanks to the MT translations of web pages.

So let’s take it for granted that MT has its uses. The question remains: how good does the translating need to be for it to be used at all? Or to put it the other way round, at what level does it become so bad that it’s useless? It’s a question that concerns not only MT. Many human translations are far from perfect, and senior Expert Translators who correct and improve translations (they’re called Revisers or Editors in professional circles) will tell you that some are so bad, it would take them less time to do a fresh translation from the original than to patch up the bad one. It’s a subject that lacks proper study. Almost all the literature I’ve read is concerned with good translation and how to achieve it, and preaches openly or by implication that the bad ones should be consigned to the rubbish basket. I once thought of writing a counter-article with as title In Praise of Bad Translations, but I haven’t gotten around to it.

The usefulness of imperfect translation is a very practical matter, not least because there’s so much of it. As noted above, perfect or near-perfect translation is expensive. It’s also very labour intensive. There comes a point of diminishing return beyond which more time and money spent on a translation may not be worth it.

And last but not least, it’s a matter for concern in research on Natural Translators, since we expect that a significant part of their production will contain mistakes.

So let’s admit another standard besides the traditional good translation / bad translation one. It’s the usable translation / unusable translation threshold. To which must always be added the qualifiers, usable when, where, for whom and for what?

More to come.

REFERENCES
Anne Harding. Drug label accuracy getting lost in translation. Reuters U.S. Edition. New York, April 9, 2010. http://www.reuters.com/article/idUSTRE63853K20100409.

Iman Sharif and Julia Tse. Accuracy of computer-generated, Spanish-language medicine labels. Pediatrics, April 5, 2010. http://pediatrics.aappublications.org/cgi/content/abstract/peds.2009-2530v1.