Sorry, but (Japanese) translation A.I. is still asleep

By | December 16, 2016

Recently someone told me of an article in the New York Times titled “The Great A.I. Awakening”, which talks about how Google Translate has switched over to using an A.I. based system, and the supposed improvements.

The beginning of the article talks about Japanese to English translation, and includes several English translations of a passage from Hemingway’s “The Snows of Kilimanjaro,” created by feeding in a Japanese translation (of the original text) by professor Jun Rekimoto. While there is one obvious grammatical error in the “new A.I.” version, the claim is that the translation is much better than the old system.

But, after skimming this article (it was extremely long and I didn’t read the entire thing), I was doubtful how good the A.I. really was, since I had done an experiment before of using Google Translate to convert Japanese to English, and that was a real train wreck.

So I decided to do another quick experiment, using just a single sentence from a novel I am currently translating (first sentence of Chapter 5 of “Japan: A New Age”, which I am in the final editing stages of and will be releasing soon.)

===

Original sentence:

その夕食から1日おいて、山戸と牧山は、順平の母の洋子に連絡を取って、父の洋平が早めに帰って会えるこということで、午後19時に吉川家を訪問した。

Google’s Translate’s translation:

One day from that dinner, Yamado and Makiyama contacted Yokohe ‘s mother Yoko, and visited the Yoshikawa family at 19 pm by saying that her father’ s young woman could come home early and meet.

My translation:

Two days after their dinner meeting, Dr. Yamato and Dr. Makimura visited Junpei’s home at 7 pm, having previously contacted his mother Yoko to make plans for his father Yohei to come home early that day for the meeting.

===

A good translation requires a good mix of semantic correctness and style to give a certain mood. I’m not going to focus on style too much, since I am still working on that myself (and there is a subjective element to it), and will instead keep to verifying whether the meaning and basic grammar of Google Translate’s result is correct.

Here is a list of semantic problems with Google’s translation:

  1. The expression “から1日おいて is incorrectly translated as “one day (later)” instead of “two days (later)”. I double-checked and technically speaking, this expression can be interpreted as “one day later” by some people, but the consensus seems to be “two days” is correct. I also verified this with the author of the novel.
  2. Several names are spelled wrong (ex: 順平 should be Junpei, but it is rendered as Yokohe by Google). Japanese names written in Kanji typically can have more than one reading, so this has a subjective element to it. In my case, I verified with the author, though some of these names have a dominant reading. It is interesting to note here is Google actually translated  牧山 correctly as “Makiyama” that tipped me off to this typo by the author. When reading it originally, I just assumed it was the correct 牧村 and translated that as “Makimura”.
  3. Google translated 午後19時に as “19 pm” which is actually technically correct in a sense, however apart from the fact whether the author used a technically incorrect expression (I’ve seen other people use this as well), it is clear that “7 pm” is the natural way to translate this.
  4. Google put a mysterious space within “Yokohe ‘s”. This is an obvious grammatical error, not sure how this got through.
  5. For some reason 父の洋平 is translated as “her father’ s young woman” instead of “his father Yohei”. I can see how this is tricky because whose 父 (father) it is isn’t specified, a trend commonly seen in Japanese.
  6. The “〜帰って会えるこということで〜” is translated as “~by saying that…could come home early and meet~”. While the “こということで” doesn’t really have a direct parallel in English, it’s meaning is roughly parallel to “~with the thing~”. I used “having previously contacted” to represent this (and the “contacted” part is literally there early the sentence), which I think works well. But it is clear Google’s rendering is confusing. While the verb いう originally comes from 言う which means “to say”, in many cases this meaning is lost partially if not completely.

This isn’t a cherry picked example–it is the first thing I tried after hearing about Google Translate’s improved A.I. While I don’t doubt things are better than before, this is clearly not close to being at a stage where it can replace professional (or even amateur) translators. And I haven’t really started to look into stylistic things, where emotions can be invoked by subtle change in wording. This isn’t a good passage to discuss that anyway.

Some of these things, like checking with the author or even discovering mistakes in the original text, I think will be extremely difficult for computers to handle, even with enhanced A.I. In all fairness, some of the differences between my translation and Google’s rely on previous information (like how I used “Dr.” for two of the characters), though I don’t any of the ones I pointed out above fall into this category.

This is not to say this technology is useless, of course there are a great variety of uses, and in most cases I think having this would be much better than having nothing.

My personal opinion is that we won’t have real human-level translation (at least for relatively distant language pairs like Japanese and English) until we have human-level A.I., which can do nearly anything else a human can do. When that happens, I think we will have much bigger problems than worry about translators losing their jobs (:

(Visited 616 times, 1 visits today)

Leave a Reply

Your email address will not be published.