The reCAPTCHA service is used to digitise old texts and, more recently, even old radio shows
The question of human versus machine intelligence has been debated quite extensively over the last several years. Despite its computational limitations, the human brain, owing to its learning and genetic capabilities, can perform certain tasks with carefree nonchalance, while a computer program may struggle with the same.
The most exciting things in the future are going to be achieved by optimising the man-machine balance. Fields such as machine learning and crowdsourcing aim to achieve the same.
While machine learning tries to narrow down the man-machine divide, crowdsourcing utilises the numbers in crowds to simplify large tasks.
In Terry Tempest Williams’ book Finding Beauty in a Broken World, she says, “Men inherently want to work and be part of something larger than themselves.”
To give them that opportunity, crowdsourcing breaks down a job that requires many man-hours into chunks, disguises these chunks as a game or as an easy task and redistributes these chunks among innumerable people. The purpose of the task takes a backseat while the task, at the forefront, is solved with the natural intuition that human beings have.
What is CAPTCHA?
If you have used a mail client or a popular website, you may have often encountered CAPTCHAs, an automated test that can tell if the reader of the web page is a human or a computer bot. CAPTCHAs prevent spammers from using a computer program to send an unnaturally high number of requests to a website in a short span of time and from retrieving sensitive information from banking and trading websites and so on.
A CAPTCHA contains about seven to eight characters that are distorted and then displayed to the user. It can successfully differentiate between man and machine because Optical Character Recognition systems are often unsuccessful at identifying distorted characters. As CAPTCHAs have achieved popularity, the distortions have become better understood and more predictable. This has prompted spammers to train a program to understand the possible distortions of a character and identify it, at times. Further, the creators of CAPTCHA were bothered by the number of man hours used to type out the characters in a CAPTCHAs (at the rate of over 100 million per day) in online forums and other places, only serving the small purpose of authenticating the viewer of the CAPTCHA.
reCAPTCHA and digitisation
To overcome all these failings, the reCAPTCHA service was created. The reCAPTCHA contains two distorted words instead of a sequence of unrelated characters. These words are scanned from old texts to utilise the man hours used in decoding a reCAPTCHA for the purpose of digitising written works. reCAPTCHA words have traditional distortions found in CAPTCHAs and are naturally faded as they are scanned from old texts, this makes them more difficult to be decoded programmatically.
Digitising a text
The process of digitising a text is as follows:
Firstly, the book is digitised by two OCR programs. In the resulting digitised text, if a particular word has been recognised differently by each OCR or if it is not a legal word in the dictionary, it is marked as suspicious. Suspicious words are then presented, along with a non-suspicious word (called a control word) in reCAPTCHA images. Fourthly, if the control word is identified correctly by the human, it is assumed that the suspicious word is identified correctly too. Lastly, the same reCAPTCHA image is presented to more than one human to verify the human entry for the suspicious word. In case of conflicts between OCR and human, the various entries received are voted for, with each OCR getting 0.5 points for a vote and each human getting 1 point for his or her vote. The option with at least 3 votes is chosen to be the right one.
It is astonishing to see that works digitised using reCAPTCHA have had an accuracy of about 99.1 per cent, which is equivalent to that of professional transcribers. reCAPTCHA was acquired by Google in 2009. Since then, Google has extended reCAPTCHA to include distorted sounds to digitise old radio shows as well.
Other crowdsourcing applications
Other crowdsourcing projects include games such as Foldit, which turns the process of obtaining the possible structures of a protein into a competitive game. Understanding the protein structure is crucial to understanding how the protein can be targeted with medicines and drugs. Foldit helps its players to competitively fold the best structure of the protein.