AI often mangles African languages. A network of thousands of coders and researchers is working to develop translation tools that understand their native languages
Imagine joyfully announcing to your Facebook friends that your wife gave birth, and having Facebook automatically translate your words to “my prostitute gave birth.” Shamsuddeen Hassan Muhammad, a...
Imagine joyfully announcing to your Facebook friends that your wife gave birth, and having Facebook automatically translate your words to “my prostitute gave birth.” Shamsuddeen Hassan Muhammad, a computer science Ph.D. student at the University of Porto, says that’s what happened to a friend when Facebook’s English translation mangled the nativity news he shared in his native language, Hausa.
Such errors in artificial intelligence (AI) translation are common with African languages. AI may be increasingly ubiquitous, but if you’re from the Global South, it probably doesn’t speak your language.
Makes sense, given where AI was developed and our current society, I imagine these were developed mostly by English speaking, white folk (probably majority males). Same reason AI has trouble...
Makes sense, given where AI was developed and our current society, I imagine these were developed mostly by English speaking, white folk (probably majority males). Same reason AI has trouble recognizing darker skinned individuals, because of inherent biases in the development process. (I am not claiming that any of this was malicious, only a product of the people who made it)
Biases in the training data, as the article confirms. The standard process itself is neutral, but there is comparatively little Hausa on the internet to train language models with.
Biases in the training data, as the article confirms. The standard process itself is neutral, but there is comparatively little Hausa on the internet to train language models with.
Imagine joyfully announcing to your Facebook friends that your wife gave birth, and having Facebook automatically translate your words to “my prostitute gave birth.” Shamsuddeen Hassan Muhammad, a computer science Ph.D. student at the University of Porto, says that’s what happened to a friend when Facebook’s English translation mangled the nativity news he shared in his native language, Hausa.
Such errors in artificial intelligence (AI) translation are common with African languages. AI may be increasingly ubiquitous, but if you’re from the Global South, it probably doesn’t speak your language.
Makes sense, given where AI was developed and our current society, I imagine these were developed mostly by English speaking, white folk (probably majority males). Same reason AI has trouble recognizing darker skinned individuals, because of inherent biases in the development process. (I am not claiming that any of this was malicious, only a product of the people who made it)
Biases in the training data, as the article confirms. The standard process itself is neutral, but there is comparatively little Hausa on the internet to train language models with.