GOOGLE GETS LOST IN trans­la­tion

As Google in­cor­po­rates African lan­guages, the con­ver­sions ring true to vary­ing de­grees as neu­ral tech­niques im­prove

CityPress - - Front Page -

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

Newspapers in English

Newspapers from South Africa

© PressReader. All rights reserved.