Facebook Guang Guang Guang Guang translate loop
« previous post | next post »
From Jeff DeMarco:
I hit the translation button for this Facebook post and this is what I got!
Here's the same announcement as rendered by Google Translate:
Yóuyú shòu yìqíng yǐngxiǎng, xiānggǎng wénhuà zhōngxīn guānbì zhì 3 yuè 2 rì, hóngguāng guó yuètuán 2 yuè 29 rì “bàwáng yǔ guìfēi-mínyuè xìqǔ jiāo huīyìng” yīnyuè huì qǔxiāo. Qǐng guānzhòng bǎoliú ménpiào, liúyì wénhuà zhōngxīn jí hóngguāng guó yuètuán facebook de zīxùn, yǒuguān tuìpiào huò bǔ yǎn ānpái róng hòu gōngbù.
由於受疫情影響, 香港文化中心關閉至3月2日, 宏光國樂團2月29日《霸王與貴妃－民樂戲曲交輝映》
Due to the impact of the epidemic, the Hong Kong Cultural Centre was closed until March 2nd, and the Hong Kong National Orchestra's February 29th concert of "The King and the Concubine-The Folk Music and Opera" was cancelled. Visitors are advised to keep their tickets, pay attention to the Facebook information of the Cultural Center and Hongguang Chinese Orchestra, and the details of refunds or make-up arrangements will be announced later.
If this sort of thing happens fairly often with Facebook's translation feature, maybe they should just partner with Google for translation purposes.
Reading:
- "Facebook = you must die" (6/9/12)
- "Facebook Absolutely Must Die" (5/22/10)
- "President Shithole" (1/19/20)
mg said,
February 22, 2020 @ 3:54 pm
FB Translate is terrible. I guess they don't want to pay to license Google translate, but they are by far the worst.
Several years ago, they insisted on translating the Hebrew "chag sameach" (happy holiday – all purpose phrase for all Jewish holidays) as "Merry Christmas"!
Thomas Rees said,
February 22, 2020 @ 7:33 pm
Did it get confused because the Wang Kwong Chinese Orchestra is Hóngguāng in pinyin?
MJ said,
February 23, 2020 @ 1:43 am
This is actually super common in neural MT systems (and neural language models as well) that aren't trained on enough data or on out-of-domain data. Best one I've seen was a research NMT we had trained on news and then tried to translate a patent for some kind of medical device and it generated a reasonable sentence or two before churning out a page worth of "and her blood and her blood and her blood". A little too "The Shining" for me
Chas Belov said,
February 24, 2020 @ 4:16 am
@MJ: For clarification, would that be:
that aren't trained on enough data or aren't trained on out-of-domain data
or
that aren't trained on enough data or are trained on out-of-domain data
?
MJ said,
February 24, 2020 @ 9:50 pm
Sorry, I meant the latter: not enough data or data that is not in the domain