What's going on?
I've read that they're using our responses to refine their Optical Character Recognition software.
Perce
Yes, I read something like that too. That's for the hard one. It seems fair to me.
But the lab where I have medical blood tests done, I just noticed an online way to get my test results emailed to me. (I've gotten them in the past but I don't remember how.)
And the capcha looks so easy. Typed numbers, impossible to misread. The only question is whether I should duplicate the spaces between the numbers so I tried both methods, and nothing worked.
So I emailed them and that didnt' work, so I called them and she asked what browser. Firefox. "We've had trouble with that." Just for me to type in letters and numbers? I don't think I believe her. While I was in my first and only sentence that the webpage gave no warning not to use Firefox, she hung up on me. ;-)
I believe you are referring to the reCAPTCHA project started by a guy from Carnegie Mellon University.
"Beyond its obvious use for foiling bot attacks and would-be spammers, the reCAPTCHA Project has another, more altruistic purpose. Several years after introducing the world to CAPTCHA technology, von Ahn realized that, despite taking just a few seconds to type a CAPTCHA, humans were spending hundreds of thousands of hours each day typing in more than 100 million CAPTCHAs. reCAPTCHA technology was developed not merely with an eye toward improving cyber security, but also as a way to harness and reuse the collective human time and mental energy spent solving and typing CAPTCHAs--a concept von Ahn has dubbed "human computation." By constructing CAPTCHAs using words tagged as unreadable in the digitizing of books and other printed material, millions and millions of cyber users play a part every day in the digitization and preservation of human knowledge by transcribing words. Tests have shown that reCAPTCHA textual images are deciphered and transcribed with 99.1% accuracy, a rate comparable to the best human professional transcription services. In just the first year after launching reCAPTCHA, humans correctly deciphered and transcribed more than 440 million words, roughly the equivalent of 17,600 books."
...and...
"Google says reCAPTCHA's technology can help it with some of its high-profile initiatives, like scanning books and newspapers to create searchable archives. As users type in the words, they help teach computers to read scanned text, improving computer accuracy when converting scanned images into plain text, a process known as optical character recognition. 'Having the text version of documents is important because plain text can be searched, easily rendered on mobile devices and displayed to visually impaired users,' Google said in a blog post about the deal. Wall Street Journal, 9-16-09"
For the record CAPTCHA is an acronym for "Completely Automated Public Turing test to tell Computers and Humans Apart.
I.
don't.
do.
captchas!
No, they're using you to do the OCR on words that stump their software. You're doing the work that computers won't. But be happy: this means we won't be terminated by our new machine overlords.
Re-read my earlier post. It's used for both.
First, it is used to force users to decipher words that the OCR software had trouble with and then the results are used to refine the accuracy of the OCR software.
Google said it. It must be true.
But how do they know what we type in is right?
When they use the 2 image reCAPTCHA system, the first image is one the control, the second Image is the one they want to figure out.
If you get the control image right, you are in. For the second image, they keep track of what users enter and use that data to refine their software and/or use the answers in their digitization of the document. It's explained in more detail here:
You raise the question, This might be great for interpreting a portion of text, which, when combined with other text that is readable, or which was also interpreted using captcha users, makes sense.
But what about cases where a plurality or even a vast majority of users give the same answer, what it looks like, but the answer is wrong? I don't know. I haven't read the wikip page or anything else, but I suspect they don't have an answer. There are a lot of cases, many times in archaeology for example, where competent people say, "It appears that...." Nothing wrong with that. Unfortunately, in the case of archaeology, often others will read those cautious, proper statements and repeat what was thought likely as if it were fact. Then people think the archaeologists said it that way, and it makes them look bad.
Perhaps you should read the page. I posted the link for a reason.
Think about it. If they can examine all the RAW VIDEO coming in from their roving drones to *isolate* the house number, do you really think they can't figure out what the number *is*? How do they know where the number is located on the property? On the house? Mailbox? Curb? etc. Do they show you a picture of a HOUSE and tell YOU to find the number -- and tell them what it is??
Recognizing digits is relatively trivial. The post office recognizes HAND PRINTED zip codes on mail FLYING past a camera at a high rate of speed. Surely, recognizing STORE BOUGHT digits (that THEY have already "located") is a piece of cake! Especially for a firm that has THOUSANDS of computers available to do that work!
Note that they also have some clues as to what valid responses (for THEIR recognizer algorithms) are likely to be. I suspect they already have a database that tells them the street numbers at each cross street. So, they know the upper and lower limits on the numbers between any two intersections.
Additionally, they have *some* idea of the number for the house immediately "before" and "after" the house in question. If the previous is "5" and the next is "15", then it's probably a safe bet to assume the current house is NOT "23" but, rather "13"! And, as numbers tend to be odd/even (alternate sides of the street), you can bet it's "13" and not "12"!
A more interesting challenge would be to see how "13-1/2" would be handled. Or, "27B".
When Kurzweil created the "Data Entry System", an "operator" monitored an image of the text that was being scanned. So, if the algorithms were unsure of their results, they could highlight the portion of the image and ask the operator for clarification. This may be how (relatively RARE) exceptions like "13-1/2" are handled!
When I read Don's reply, I was reminded that originally you were asking about house numbers. I agree they can verify what house numbers are, and so I find it hard to believe that house numbers are the unknown half of a reCaptcha. Regardless, your shorter post above didn't specifically refer to house numbers and my reply was about words and numbers of all kinds that might be used.
I expected you to reply. You know, you're a very competitive guy.
I don't mean that as a bad thing, just an observation.
Okay, I read the page. This is the closest it comes to addressing what I said: "If the human types the control word correctly, then the response to the questionable word is accepted as probably valid. If enough users were to correctly type the control word, but incorrectly type the 2nd word which OCR had failed to recognize, then the digital version of documents could end up containing the incorrect word. The identification performed by each OCR program is given a value of 0.5 points, and each interpretation by a human is given a full point. Once a given identification hits 2.5 points, the word is considered valid. Those words that are consistently given a single identity by human judges are later recycled as control words.[14] If the first three guesses match each other but do not match either of the OCRs, they are considered a correct answer, and the word becomes a control word.[15] When six users reject a word before any correct spelling is chosen, the word is discarded as unreadable.[15]"
It says that if users agreed on a meaning but they were wrong, their opinion is accepted as correct. That doesn't answer Ken's question "how do they know what we type in is right?" They don't know, and they don't addreess that they don't know.
If you re-read my response to Ken's question, you will see that I began the post with:
"When they use the 2 image reCAPTCHA system, the first image is the control, the second image is the one they want to figure out."
When they use 2 images, they do in fact know that "what we type is right" for the first image. The first image - the "control" - is used as the bot screener, the original purpose of CAPTCHA. The second image was added later once von Ahn realized that there was a lot "human computation" power available for free.
Ken's question is a valid question when the only thing you are told is "they're using you to do the OCR on words that stump their software." The correct answer is they *always* know when users are correct for the control image, either by itself (CAPTCHA) or when used as the first word in the reCAPTCHA system. It is true that they do not know if the users are correct for the second image, but you can be sure that they tested the system in order to see if it was a viable method for helping out with the digitization of documents when the OCR software struggles.
From:
"Tests have shown that reCAPTCHA textual images are deciphered and transcribed with 99.1% accuracy, a rate comparable to the best human professional transcription services. In just the first year after launching reCAPTCHA, humans correctly deciphered and transcribed more than 440 million words, roughly the equivalent of 17,600 books."
My main concern is that they don't have a lot of time in that they let you in right away, if you get it right.
But ... how do they have enough TIME (as in split second time) to make that decision?
I don't think I'm being clear so, let me try an example:
If Person A is the first to hit Enter, they have nobody to compare answers with.
Of course, they can just automatically say "no" to the first 10 people but all this has to happen in something like 10 seconds, right?
Did you see my response(s)?
Here is a clear example I saw today of the *control* word, I think:
And, here is another example I saw today of the *easy* street numbers:
However, what perplexes me are these types of *impossible* cyphers:
Where is the *control* word in this captcha I encountered moments ago?
Of those last two examples, I can't imagine that *either* of them portrayed a word in an OCR document, unless they are OCR'ing documents after sending them through a washing machine first!
So, what's the OCR-corrective sense in the latter two captcha's?
Here are links of actual captchas that I ran into today.
None of these are cipherable by me, an avowed human.
I can't imagine that anyone can get these within reasonable accuracy.
Can you find any OCR use for these washing-machined words?
Here are some captchas I found today of actual mailbox numbers:
Here are curb numbers:
And here are house numbers:
And even numbers attached to the roofline!
I wonder how they even FIND these numbers, automatically anyway?
HomeOwnersHub website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.