Re:AWS Public Datasets
by pavlov on Monday March 26, 2018 @06:06AM
(#16677807)
When Google Translate was new in my language Finnish, their translation quality was atrocious.
I tried this sentence: "Hei äiti, puhun suomea". The expected translation would be "Hi mom, I speak Finnish".
Instead Google's result was: "July's mother, I speak English".
Obviously the engine had been trained on unvetted data sets where the word "English" occurred in translations in a position where the original had the word "Finnish", and no context was provided to avoid this kind of mistake.
The word "July" came about because "hei" is also used as an abbreviation for "heinäkuu" (July). It was sobering that a supposed world-class AI couldn't distinguish between these two usages. Machine learning needs a lot of old-fashioned handtuned human-made heuristics.
I tried this sentence: "Hei äiti, puhun suomea". The expected translation would be "Hi mom, I speak Finnish".
Instead Google's result was: "July's mother, I speak English".
Obviously the engine had been trained on unvetted data sets where the word "English" occurred in translations in a position where the original had the word "Finnish", and no context was provided to avoid this kind of mistake.
The word "July" came about because "hei" is also used as an abbreviation for "heinäkuu" (July). It was sobering that a supposed world-class AI couldn't distinguish between these two usages. Machine learning needs a lot of old-fashioned handtuned human-made heuristics.