Just for Fun: Bidirectional Multi-passes of Google Translate

[ainsoph9]

I wonder why on earth Google replaces "死" with "市" in this test.

死はそれがないと言っていますよ!!

Same sound?

[EusthEnoptEron]

I wonder why on earth Google replaces "死" with "市" in this test.

死はそれがないと言っていますよ!!

Same sound?

Yeah, but from a technical perspective it makes no sense to use another kanji if you're given it, as they presumably search for the meaning by the kanji and not its sound—unless there is some kind of auto-correction function, which I consider rather useless at the current level of quality.

[ainsoph9]

From what I can tell, the program runs on an iteration basis, which can have one of two effects.

1. It improves the translation.
2. It compounds the mistakes of the first mis-translation.

[larethian]

there may be a neural network or other AI system behind that not just parses blindly. it might take into account that the user makes a typo or is unsure of the actual kanji, and figured that the sentence will not make sense if the word 'death' is used.

[Mystes]

there may be a neural network or other AI system behind that not just parses blindly. it might take into account that the user makes a typo or is unsure of the actual kanji, and figured that the sentence will not make sense if the word 'death' is used.

That's more what I think.

[rpapo]

there may be a neural network or other AI system behind that not just parses blindly. it might take into account that the user makes a typo or is unsure of the actual kanji, and figured that the sentence will not make sense if the word 'death' is used.

That's more what I think.

From an e-mail conversation with Mr. Jim Breen (the guy behind WWWJDICT), I was told the following:

Back in the 80s parsing Japanese was considered to be too hard for computers.
Then Nagao showed with Juman that it could be done fairly reliably if you use a bit of science (Juman uses the Hidden-Markov-Model AI technique.) Things have moved on a bit with its successors - MeCab uses a combination of HMMs and CRFs (Complex Random Fields), the latter being one of the latest bits of AI applied to machine-learning.

Google used to use a commercial parser (http://www.basistech.com/base-linguistics/) but moved to an in-house one for Japanese. Taku Kudo, who wrote MeCab, now works for Google }

My parser contains the Juman code (version 5.1), but I use it to generate an alternate parsing of the text. The interesting thing I find is that where Juman misses, so does Google. That leads me to believe that the current Google code incorporates the Juman code.

My parser uses the WWWJDICT file EDICT as the basic dictionary, so it has a much larger vocabulary than Juman, but it is nowhere near as smart. It uses a "hit it with a bigger hammer" approach. Both approaches have their pluses and their minuses.

[ainsoph9]

So, it might be somewhat more ideal to see a cross between the two.

[rpapo]

So, it might be somewhat more ideal to see a cross between the two.

Perhaps. I am not an ivory tower computer person, but rather a practical one. I can imagine a program with a large dictionary and set of rules, and a collection of "hints" to help it out of hard spots. But I have relatively limited free time, and I've chosen, for the time being, to work on translating the hard way to improve my own skills. It rather helps to know how to do something yourself before you try and teach a computer to do it . . .

[Mystes]

I noticed that the auromatic trasnlators place the time clause in the fornt.

[ainsoph9]

So, it might be somewhat more ideal to see a cross between the two.

Perhaps. I am not an ivory tower computer person, but rather a practical one. I can imagine a program with a large dictionary and set of rules, and a collection of "hints" to help it out of hard spots. But I have relatively limited free time, and I've chosen, for the time being, to work on translating the hard way to improve my own skills. It rather helps to know how to do something yourself before you try and teach a computer to do it . . .

That is quite true. Sometimes, I do get the feeling that translation programs among other things are not done by the foremost experts in the field.

[Mystes]

Though if automatic translaion was perfect, translators would lose thier jobs.

[ainsoph9]

No, they would become computer programmers.

[Mystes]

Not my cousins though.