Hacker News Reader

Hacker News

Comment by solutionyogi | Hacker News Reader

Parent Full thread

solutionyogi

Wow. Just Wow.

I had to put the App to test.

Printed some simple English words and tried it out.

Original paper:

http://imgur.com/EYsko

What the App shows in REALTIME:

http://imgur.com/FhaSW

Comparison with Google Translate:

http://imgur.com/GOd94.jpg

[EDIT: Note, Translating Muerto Fin to English in Google Translate, does result in 'Dead End'. Can any Spanish reader clarify why the original Google Translate chose to translate 'DEAD END' to 'callejón sin sali'.]

[EDIT: So let me get this straight. Their program, running on iPhone [256MB RAM, 600 MHz ARM CPU], can take a live image, perform OCR, translate, create another image with the translated words. And all of this happens real time? Wow.]

I think it's really hard to get a 'Wow' reaction from HN crowd. And you have hit a home run! As someone else said in this thread, you are going to make boatload of money. Congrats!

View on HN

(Native Spanish speaker here.)

First: This app is amazing!

But I think that this app is doing almost a word by word translations, without a list of usual expressions or grammatical analysis.

For example, the correct translation of 'Dead End' is 'callejón sin salida', that means literally something like 'street without exit'. Translating word by word 'Dead End' you obtain 'Muerto Fin' that is unintelligible.

Another example from the video is the translation of

  'Lengua boliviana con una salsa picante de anchoas'.

The app translates this as

  'Tongue Bolivian with a sauce spicy of anchovies',

But the correct translations is something like

  'Bolivian Tongue with a spicy sauce of anchovies',

because in a translation between English and Spanish you must reverse the order of the adjectives and nouns.

yep, this is a word-for-word translation, because it's fast and it gets the point across. we're working to improve translation quality and finesse, but it's a much harder problem to understand grammar. so, we hope it gives the general meaning, and you can learn to piece it apart.

Word for word is far, far better than nothing. And no network connection required, so awesome. Roaming charges when travelling are brutal, so very important feature.

Real time OCR + word for word translation....maybe its not particularly technically interesting on paper, but to watch it on a handheld device in front of your eyes is breathtaking.

Do you have a timeline for other languages? (Japanese)

Yeah, no network connection is the killer app here, since you're probably going to need this in another country where you probably don't have network access.

That's great too. Text detection + character recognition + 3d position estimation + translation, each one of the problems is difficult but state-of-art method usually works quite well. The combination, however, is super hard for a small team because each one of them requires special knowledge in that particular manner. To pick and choose which one to focus on and can still get awesome result is a rare skill.

I think that most of the time a word-for-word translation is enough. I use Google translator a lot, and most of the time the translation is not perfect but it is good enough. So I think that is app is very useful.

But I think that the real problem are the idioms and phrases, like "dead end", that have a completely different translation.

On the other hand, I think that you have more processing time that 1/10 sec. Most of the time the user will point the app to the same text for 10 or 15 seconds. I think that it is possible to show first a very fast translation almost instantly and a few seconds later show an improved version.

Some times, but not always. Not even sure if it's "most" of the times. "Right turn" to "correcto turno" is so wrong, anyone following directions with that would be lost pretty easily.

Still, amazing technology, specially since it's all client-side. Hopefully they'll have some contextual analysis in the future or enable Google Translation API use.

Google Translate is not word-to-word though.

I know that Google Translate is not word-to-word, but it is not perfect.

For example, a few years ago, some students of my wife give to her a homework about clocks and gears. When she read it, she was annoyed because the redaction was incredible horrible. But later we realized that the students didn't write it, the "homework" was a web page translated with Google Translate.

Another time, I need an example of the differences between JPEG and JPEG2K for a internal talk. I found a photograph with a zoom of Lena's eye with a legend like "JPEG2K 1% vs JPG 1%", but the webpage was in Japanese, and I can’t reed it. So I use Google Translate to be sure that it was a comparation of the two methods with the same compression level.

So Google Translate is very good to get an idea of what some web page means, but it is not good enough to make a final version of the translation.

This app have some additional difficulties that GT doesn't need to solve: they need to OCR the text in the wild, they have less computational power and they have to do it in real time. It is almost incredible that they can solve these things. With a word to word translation, I think that you can get a good enough translation 90%, 95% or even 99% of the times, but the corner cases can be really unintelligible. The translation of "Tongue Bolivian ..." is fine, and the user can understand what it means, so it is useful. The translation of "Dead End" is something that they should improve in next version.

Google Translate is definitely far from perfect.

I once saw this forum post that was in German, that Google translated "A: Nyet." to "A: Yes."

Since it was a question and answer post regarding information for an upcoming game, there was quite of bit of fuss over it before this translation error was discovered.

It surprised me so much I took a screenshoot of it:

http://dl.dropbox.com/u/5815494/GW-Deutch.jpg

I cannot understand what heuristic would translate the complete German sentence "Nein." into "Yes.".

Theres a huge market for Indians going back home and who cant read their native tongue.

Do you have an API or something? I would love to work on a tamil to english dictionary.

Can you please email me? My email is in my profile. Thanks.

John, outstanding work. Word-for-word is IMO more interesting too, because you understand enough to get the point across - but you still get the flavour of the language and enjoy the idioms. Huge, huge potential. You win the Internet today.

Thanks for the positive feedback (I'm one of the developers). We never imagined we'd win the internets!!! :P

About the word for word thing, that is mostly correct to say it's word-for-word. We have just a few short phrases in there, like "por favor". If that didn't translate correctly, it would have been teh lame.

>We never imagined we'd win the internets!!!

Oh please, I mean, it's only one of a kind app that millions could make use of :)

I have some experience working with machine translation during my thesis(using Latent Semantic indexing, LSI), and after watching the video I was very impress with the speed of the translation, specially knowing how resource demanding are other strategies like probabilistic translation or LSI.

Are you hiring? ;) I am a native Spanish speaker btw.

"dead end" isn't grammar thought, it's a metaphor / idiom, and the only way to handle cases like that is a database of special cases.

Frickin awesome app tho'

John, still, this is amazing for the first iteration. I wish you could licence 'grammar logic' from Google Translate and incorporate in your App.

Do it for German and French, and I'd pay $9.99.

And in the last case, from English to Spanish, it puts direction as a masculine substantive, while in Spanish is feminine.

Because "dead end" means "callejon sin salida". That's the correct translation (callejon = road, sin salida = w/out exit).

On the other hand, "muerto fin" has NO meaning in spanish, (and that's what you get when you translate word for word without context).

Thank you. I think context is extremely important and very difficult to get right. I remember reading one article by Google engineer where they discuss how they keep improving their analysis of search queries.

E.g.

[Search Term] -> [Interpretation]

New York -> New York city

New York Times -> The news paper

New York Times Square -> Famous tourist spot in NYC.

There were other examples in the article. I will try to find that article and link it up.

I think Google should buy these guys and provide them with their knowledge of 'context'. Google Translate team and these guys should talk right away.

I used to work in machine translation, about 15 years ago. At the time, one of the big things they were working on was an improvement to the contextual inferences. The prior versions were able to consider context within a single sentence, and input documents could be tagged with a subject area, so the system would know to resolve ambiguities in favor of, say, medical terms.

The best examples of this, I thought, were the following passages:

Yesterday I went to the symphony. The sound was beautiful.

versus

Yesterday I went to Long Island. The sound was beautiful.

In the first passage, sound refers to an auditory experience; in the second, it refers to a body of water. Within the sentence "The sound was beautiful", there's no way to know which to use, so one must look to surrounding information to make a guess.

But I've given a short and clear example; anyone can see the right solution. But in real-life usage it's frequently unclear.

Was it this one? http://www.wired.com/magazine/2010/02/ff_google_algorithm/al...

context is extremely important and very difficult to get right

Yep... solving "context" 100% is equivalent to solving A.I.!

"Muerto fin" means "an ending that has died", no?

No, that would be "fin muerto"; "muerto fin" is gibberish.

Ah, thanks.

(2 years of classroom Spanish here.)

I like how "Make right turn" becomes "Make correct turn".

It becomes "make correct shift" (as in "I work in the night shift"). The word in Spanish would be "giro". The natural translation would be "gira a la derecha".

"turno" can mean either "shift" as in "I work the night shift" or "turn" as in "it's my turn now" (but not the turning sort of "turn").

"gira a la derecha" is correct but not idiomatic in every country; in Mexico we say "da vuelta a la derecha".

"make right turn" is apparently hard! Word Lens translates it to "make correct turn" but where "turn" is in the sense of taking turns not in the sense of turning. Google Translate gives "to the right which in turn".

Reply on news.ycombinator.com