杏吧原创

MACHINE translation can make French, Spanish or even Japanese accessible to English speakers. But it requires a wealth of documents with copies in each relevant language to learn how to translate. This works for widely spoken languages, but it can be a tall order for some of the world鈥檚 7000 or so tongues. Software that invents languages by mashing up words from existing languages could help translate foreign texts for speakers of little-used tongues.

and Vicent Alabau at the Polytechnic University of Valencia, Spain, were inspired by the film Blade Runner, in which characters use a street language called 鈥淐ityspeak鈥�, a mix of languages including Japanese, Spanish and German.

Their system, Culturally Influenced Interlanguage (CI²), exploits the similarity in words and grammar found in language families like the Romance family of French, Spanish and other southern European tongues.

CI² constructs translations for minority languages by borrowing words from languages in the same family. 鈥淭he idea is to pick words from languages for which there are machine translators available,鈥� says Leiva. A resulting phrase is unlikely to be grammatically correct, and may contain unusual spellings, but it should be understandable to a minority language speaker.

聯The idea is to pick words from languages for which there are machine translators available聰

For example, imagine that Spanish is a minority language, and a native speaker wants to read the English sentence 鈥淎nother label with the same name already exists鈥�. Machine translations between English and Italian, French and Portuguese are fairly accurate and Spanish shares many characteristics with those languages. By choosing the appropriate words from each language, it is possible to automatically construct the sentence 鈥�Un鈥檃ltra 茅tiquette con mesmo nome existe d茅j脿鈥�, which is not Spanish but should be reasonably comprehensible to a Spanish speaker.

Leiva and Alabau choose words by calculating the probability that a word from the major languages appears in the minority vocabulary by looking at the number of letter changes required to turn one word into another 鈥� so 鈥�con鈥�, which occurs in both Italian and Spanish, has a probability of 1, whereas the Italian 鈥�nome鈥� has a probability of 0.79 because the Spanish word is 鈥�nombre鈥�. Choosing the words with the highest probability means that the translated sentence has the best chance of being understood.

The pair tested CI² by asking 17 native Spanish speakers to read a selection of Swedish sentences translated into CI² 鈥淪panish鈥� via Italian, French and Portuguese, along with direct translations in all three languages. The volunteers found the CI² text easier to understand than the Italian and French translations, but the Portuguese was more familiar still, perhaps because of its closeness to Spanish. Leiva and Alabau plan to test the system with a minority language, even though it is often hard to find native speakers to take part in experiments. CI² will be presented at in Austin, Texas, in May.

鈥淚t is a technical solution to a cultural need, which is very exciting,鈥� says , a machine translation researcher at Johns Hopkins University in Baltimore, Maryland. He says the rise of the internet means that languages with less than a million speakers will struggle to survive. 鈥淭he teenagers of this planet will decide where this goes. They will decide what they want to speak and how important culture is to them.鈥�