A few months back I posted about a hilariously off Google translation from English to Japanese of a status I’d originally posted to Facebook. After reading this massive article about how neural networks have dramatically improved the service, I got curious and figured I’d plug the exchange back in and see what it gave me now.
The original English:
Adult & three little kids walking down my street:
Kid: What’s that?
Dad(?): It’s a rooster.
Kid: What’s a rooster?
Dad: It goes *does an admirable impression of a rooster crowing*
Kid 2: Cock-a-doodle-do!
Youngest Kid: *attempts to crow, ends up shrieking instead*
This made me smile as I sat in my window reading. lol
And now the new Google Translate result:
This is most definitely miles above the engine’s September attempt. Unlike that try, here the first sentence is completely correct, the sentences which used *description of action* are now at least grammatically if not contextually correct, and the very last bit of the last sentence has been corrected from an ultra-literal translation of “made me smile” in which “made” was interpreted in the sense of “create something” rather than “give rise to an action.” There are only two major problems with this translation: tone, and the engine’s inability to recognize misspelled words.
What I described was an informal exchange between a dad (I assume) and his three children. The use of the words “Dad” and “Kid” rather than “Father” and “Son” or “Child” partly indicated that in English. However, the tone of the Japanese translation is inconsistent. It used “Kiddo” for “kid” which feels informal, as does my narration which was translated in plain form. But the speech was translated into polite form, and “Dad” was translated to “Otou-san” (honorific way of saying “Father” when addressing your own father or when talking about another person’s father) when “Papa” would’ve been better for this scenario.
One thing I hadn’t realized when I made my first post about this translation was that I had misspelled the onomatopoeia for a rooster’s crow: “cock-a-doodle-doo” ends with two o’s, not one as I wrote it. Google Translate’s inability to interpret not what I wrote but what I meant to write created much of what made its first translation particularly funny. This time, instead of bringing a “huge cock” into the mix, it translated my misspelled sound effect as “Ojii-chan,” which is a somewhat affectionate-polite way to address your grandfather. How it got “Grandpa” from “cock-a-doodle-do” is beyond me.
I’ve actually come across a few similar cases doing translations from Japanese to English. For example, one customer wrote 通風 (“tsuufuu” meaning “ventilation”) in a list of diseases/conditions hot springs baths help to alleviate. Within that context, I figured this was a typo, and that the intended word was 痛風: also pronounced “tsuufuu,” but meaning “gout.” More amusing was 賛成用化粧品 (“sanseiyou keshouhin” meaning “cosmetics for [use by] agreements”) which appeared to be the product of the client typing in romaji input mode and hitting the “s” instead of the “d,” which would have gotten the correct 男性用化粧品 (“danseiyou keshouhin” meaning “cosmetics for men”).
Given the description of neural nets in the article, however, it’s only a matter of time before
we can make Mr. Data even machines can spot things that don’t make sense and correct human errors in typing to produce translations of what was intended rather than what was typed. That said, I think there might be an unfortunate trend that would hasten the death of human translators somewhat unrelated to how good machines get at doing it, and that’s people accepting (paying for) sub-par translations and people not wanting to pay for things at all.
There is hardly anything in the world that someone cannot make a little worse and sell a little cheaper, and the people who consider price alone are that person’s lawful prey.
—Common Law of Business Balance
I remember seeing a post somewhere (probably Tumblr) a few months ago about an otome game with an absolutely dreadful translation, which the OP rightfully pointed out. But then they said they hoped that wouldn’t turn anyone off from the game, because really it had a good story, and I wondered how they could gauge that when the writing was so horrible, and if they realized that companies would have no reason to invest in proper translations if they could make money off fans with crappy ones. I think the post was about the game Ephemeral by Hunex. I can’t find the precise post I’m thinking of but this one has some screencaps, as does this review. It is bad. Anyone who doesn’t think so either has a poor grasp of the English language, or has only ever been exposed to bad writing.
In a similar vein, I couldn’t help but recall all the reviews for Voltage games in the App Store where people were complaining that it wasn’t “fair” to charge for the games, how they loved the characters and everything but that the games would be better if they were free, etc. If so many people in your customer base like what you make but repeatedly tell you it’s not worth paying for, would you feel motivated to improve your product, a move which would cost you money and necessitate raising prices to remain profitable? Probably not.
While the examples I’ve given here are about a relatively niche market, I think it holds true even when taking a wider view. Services like One Hour Translation and Gengo wouldn’t exist if people weren’t okay with getting inferior products. Even if computers never become capable of producing 100% human native speaker-level translations, the skill itself has already been deprofessionalized enough to lower rates for translators working in the lower ends of the industry. Simply being bilingual and a good translator probably won’t get you very far without a STEM or law background.
Well, ultimately, what can I say? We’re probably living in the Eternal Sphere. What humans are doing now with AI is perhaps what someone else did to make us and we just think we’re thinking. Je pense, mais je n’en sais rien.