GOOGLE GETS LOST IN translation
As Google incorporates African languages, the conversions ring true to varying degrees as neural techniques improve
Google. Aguera y Arcas steers a research team of 300 who are exploring the possibilities of emerging artificial intelligences.
Using what Google calls LSTMs, or long short-term memories, machines are now being taught to teach themselves, increasing the virtual capabilities of machine-based recognition techniques. The search engine is now using its own “neural network”, designed by engineers and linguistic scientists, to mimic the way our brains work, picking up similarities between sentences in two different languages. This is opposed to earlier technologies which placed definitions alongside one another in strict grammatical order, resulting in illogical sentences.
The new Google Translate promises more logical, not purely literal, English translations, for longer vernacular sentences. “With neural translation we believe that the translations are now significantly better than they were before, particularly with long sentences,” says Laura Scott, manager of Reputation Communications and Public Affairs at Google.
“It is of course very complicated to see improvements in languages where there’s not such a body of reference material, like in isiZulu, isiXhosa and Kiswahili – the improvements won’t be as huge as the Chinese improvements we saw last year, for instance.
“So that might mean that some of the short word translations are not as good as you might expect, but we think that the trade-off is good, and we’ve seen enough of an improvement to think that those are worth being let into the world.”
Aguera y Arcas says that with more use, the system will improve.
“The headroom for these neural techniques improves as more people use them,” he says. “Machine learning is hard, because it’s not based on rules like traditional computing would be. The things that these conceptual spaces give you that ordinary statistics of languages or the more old-fashioned techniques cannot give you are, for example, metaphor.”
But anyone who has tried the Zulu translate function in the past knows that with highly idiomatic languages like isiZulu, Google Translate can have some pretty hilarious results.
We tried the app’s new improvements in the office this week, speaking to isiZulu speaker and City Press sports editor S’Busisu Mseleku, and isiXhosa speaker and education reporter Msindisi Fengu, who each took the app for a spin. We used some of Isolezwe newspapers’ posters, known for their elegant use of idiomatic isiZulu and isiXhosa.
TALK TO US What are some of the funniest translations you have received from Google’s translation apps?
SMS us on 35697 using the keyword LOST and tell us what you think. Please include your name and province. SMSes cost R1.50