[Cannikin]

Maybe only marginally related to Pentax cameras, but I thought this was pretty amusing. I was looking at a French review of the 645D as translated by Google Translate. I stumbled across this line:

French:
Modèle : PENTAX 645D

Google Translate into English:
Model: Canon 645D

This is a perfect example (among countless others) of why I think Google's purely statistical (probabilistic) approach to machine translation is ridiculous.

---

Last edited by Cannikin; 10-20-2011 at 05:15 AM.

[GabrielFFontes]

Canon fanboys!
Hahah just kidding

[average-guy]

If this is more than a one-time translation fluke, I'd be fun to see the fallout at some of the canon forums after a number of people start asking about the new "Canon 645D".

[Eruditass]

This is a perfect example (among countless others) of why I think Google's purely statistical (probabilistic) approach to machine translation is ridiculous.

It is a good example. The alternative, however, hiring at least 64! / (2!*(64 - 2)!) translators to work countless days translating phrases, combinations, context, spot checking each other's work, quality control, categorical sequence labeling and sentence parsing, is far more ridiculous.

There just needs to be more people giving feedback to the simple proper noun misidentification and other more important mistranslations. It's ridiculously easy to do, given that when you hover an area of text, it shows the original and a button to suggest a better translation.

FWIW, it looks like it may have already been fixed, at least on the french webpages I just looked at.

[Cannikin]

It is a good example. The alternative, however, hiring at least 64! / (2!*(64 - 2)!) translators to work countless days translating phrases, combinations, context, spot checking each other's work, quality control, categorical sequence labeling and sentence parsing, is far more ridiculous.

There just needs to be more people giving feedback to the simple proper noun misidentification and other more important mistranslations. It's ridiculously easy to do, given that when you hover an area of text, it shows the original and a button to suggest a better translation.

FWIW, it looks like it may have already been fixed, at least on the french webpages I just looked at.

(Getting even more off topic now):

It's not fixed. Enter this exact phrase: "Modèle : PENTAX 645D" into Google Translate. The fact that this is completely different from entering just "Pentax 645D" is exactly the reason why Google's approach is just plain bad.

Plenty of linguistically informed statistical systems (at the very least able to recognize basic syntax) have demonstrated significant improvements, yet Google is famous for totally ignoring linguistics, and relying on 100% purely statistical models driven by brute force computing and ridiculous amounts of data (for the rate of return). As a result, unless a 100% exactly identical phrase is found thousands of times in their corpus, the translation will fail utterly.

Here's another example where they can't even get basic numbers right (Japanese on left, Google translation on right):

50万 -> "500,000" (correct: 500,000)
50万年 -> "50 years" (correct: 500,000 years)
50万年前 -> "50 million years ago" (correct: 500,000 years ago)

Notice the numbers are exactly the same, yet Google gives a different one for each (a few months ago, they even got the first one wrong). Complete inability to recognize even the most basic language structures, and internal inconsistency. Will someone have to submit a "correct" translation for every single number in every single combination in existence (infinite)? So many more awful examples I can come up with (dedicated part of a paper I wrote on this).

As a side note, rule based systems are not as impractical as they would like you to believe, as people (including the computational linguistics graduate department I'm currently in) are demonstrating first hand. You just need to break away from the mentality of "A -> B point-to-point" translation, and approach it from a novel and entirely different angle which allows scalability on a large scale. But that's a whole other topic that I can go into way too much detail about.

[Eruditass]

(Getting even more off topic now):

It's not fixed. Enter this exact phrase: "Modèle : PENTAX 645D" into Google Translate. The fact that this is completely different from entering just "Pentax 645D" is exactly the reason why Google's approach is just plain bad.

The whole thread was never really on topic and belongs in General Talk.

Why aren't you using the website translator to translate a website? Or rather, why are you using it to translate proper nouns that you already know the correct translation for? That isn't its purpose.

Go to any french website with 645D and click translate, it works fine.

As a side note, rule based systems are not as impractical as they would like you to believe, as people (including the computational linguistics graduate department I'm currently in) are demonstrating first hand. You just need to break away from the mentality of "A -> B point-to-point" translation, and approach it from a very novel and entirely different angle which allows scalability on a large scale. But that's a whole other topic that I can go into way too much detail about.

If you read what I wrote, I am in favor of syntax based systems integrating with the numerous advantages of probabilistic machine learning. Your criticism of it is something I agree with, but to call the entire thing rubbish and throw it away is what is ridiculous.

[Cannikin]

Why aren't you using the website translator to translate a website? Go to any french website with 645D and click translate, it works fine.

Google Translate

Check the captions of every single photo on this page.

Your criticism of it is something I agree with, but to call the entire thing rubbish and throw it away is what is ridiculous

Didn't say that. Read what I wrote on "linguistically informed statistical systems".

The whole thread was never really on topic and belongs in General Talk.

Agree. My bad. Moderators, feel free to move this thread.