Machine and Human Translation: COMAL

Machine translation (MT) is currently booming. I can’t recall its getting so much optimistic press since the landmark Georgetown-IBM Experiment in 1954. What’s more, it's available to everyone and, to a limited extent, it’s free. A few days ago, the BBC got in on the trend by organizing SuperPower Nation Day, an experiment in multilingual debate and discussion (see photo):

“By using a specially created website, users from around the world could post and reply to each other's messages, even if they did not share the same language… comments online were translated using software created by Google, allowing users to write in their own language before seeing it translated into six others instantaneously.”

I could say a lot about the recent development, because I spent six formative years working in a major MT project in Canada. However, this blog isn’t the place for it.

Except for one aspect. Even the most enthusiastic promoters of MT, its designers and vendors, concede that it provides at most “80-90% effectiveness.” (In fact, the most important part of the BBC debate was translated by humans.) So it still has problems, and from the difficulties posed by those problems we can learn a lot of things about human translation. That's why my title embodies a reversal of the title of Alexander Ljudskanov’s book: see December 6 post. Many years ago, a young researcher named Martin Kay - he’s now the aging chairperson of the International Committee on Computational Linguistics - said one of the wisest and pithiest things ever on the topic: “The trouble with research on machine translation is that we don’t know enough about translation.” In this post, I want to point to one of the things we can learn.

When humans translate, be they Natural or Expert Translators, they’re constantly monitoring what they’re producing. How do we know this? Because we can observe them correcting themselves. There are several aspects to their corrections, but the most important is correction of the meaning. In other words, translators are constantly asking themselves, “Does my translation mean the same as the original?” If not, and if they don’t give up altogether by skipping a segment, they try again.

In order to answer the above question, translators must be able to compare the meaning of the translation with the meaning of the original. This activity forms part of what psycholinguists call metalinguistic awareness: a speaker-writer’s consciousness of what he or she is doing and of how. Despite the advanced sounding term for it, it’s something that bilingual children develop early on. Here’s an example that was reported by psycholinguist Walburga von Raffler Engel:

S lived in Italy until he was five years old. He learnt English from his English-speaking father and Italian from his Italian mother and the rest of his entourage. Deliberately, no one in the household asked him to translate, because they didn’t want him to mix his languages. So, at three years old, he was still a very natural Natural Translator when the following incident occurred.

The family was at table. Father tasted the soup and said in English that it was well flavoured but it was slightly oversalted. Whereupon S spontaneously ran to the kitchen and told their Italian cook, in Italian, “Father says the soup is too salty but it tastes good, I mean it has good ingredients.” Then, with reference to his translations “well flavoured" and "has good ingredients,” he added, “È una cosa que somiglia, ma non è uguale.” (It’s something similar, but it’s not the same.)

Bianca Sherwood and I, in our review of reports of child translators, called this competence COMAL, comparison of meaning across languages, which we defined as “a discriminatory judgement.”

However, what a three-year-old child can do, MT systems can’t. They don’t have COMAL discrimination. So to judge whether the translation is correct as to meaning, it’s still necessary to have recourse to bilingual humans. And since the systems can’t by themselves know when they’ve made a translation mistake, they don’t correct themselves. Therefore, unless the MT output undergoes human revision, any errors of meaning are served up willy-nilly to users for them to cope with as best they can.

To be continued.


REFERENCES
W. J. Hutchins. Machine Translation: Past, Present and Future. Chichester: Ellis Horwood / New York: Halsted, 1986.

Dave Lee. BBC debate demonstrates power of machine translation. BBC News, March 18, 2010. http://news.bbc.co.uk/2/hi/technology/8575526.stm.

Google Translate. http://translate.google.com/#.

Candace Séguinot. The translation process: an experimental study. In The Translation Process, Toronto: H. G. Publications, 1989, pp. 21-53.

Walburga von Raffler Engel. The concept of sets in a bilingual child. In Actes du Xe Congrès International des Linguistes, vol. III, Bucharest, Romanian Academy Press, 1970, pp. 181-184.

Brian Harris and Bianca Sherwood. Translating as an innate skill. In Language Interpretation and Communication, ed. D. Gerver and W. H. Sinaiko, Oxford and New York, Plenum, 1978, pp. 155-170. Digitized copy available without charge from translatology@yahoo.com.

Photo: BBC News