Duolingo, a free English learning tool for free, predicts optimum learning timing and enhances learning effect by users in machine learning

Free application "Duolingo"Is a service that was devised to learn languages ​​that are not mother tongue fun,In 2013, Apple will also be selected as the iPhone App of the Year, the annual best app selected by AppleWe have earned high praise. In Duolingo, it seems that how to review in order to do learning effectively is judged individually for each user by memorizing the forgetting pace of memory using machine learning, and it is improving the learning effect surely.

How we learn how you learn | Making Duolingo Blog
http://making.duolingo.com/how-we-learn-how-you-learn

The Duolingo application has an index called "skill strength meter". A mark like the orange antenna displayed under each icon on the application screen below is a skill strength meter and when it becomes 2 or less in 4 steps, the role advising the user that it is time to learn the field soon To be holding.


In addition, Duolingo adopts a mechanism that makes it easy to iterate by displaying words and phrases that are weak points in card format with reference to the user's learning history. I can overcome my difficult point by thoroughly repetitively, so I can train my vocabulary skill effectively.


About Duolingo's how to calculate the optimal learning schedule for each of these users, Duolingo has blasted behind its back side in the official blog.

Even though we have memorized 100% of what we learned immediately after learning, memory gradually fades over time. The figure below is a graph of the rate of memory loss, in this example "half-life" that forgets 50% of what you learned is after 1 day, the oblivion starts abruptly immediately after learning, After that, it is shown that the pace of forgetting becomes gentle. According to Duolingo, learning again at a timing close to the half-life which forgets at a rate of about 50% is very effective for establishing memory.


Nonetheless, it is very difficult to figure out the exact time of half-life, since when the item learned by the user will have half-life depends on individual user and what you learned. So Duolingo devised a statistical model that averages the half-life period by analyzing statistical information of millions of users obtained from the Duolingo application by machine learning.

As a result, Duolingo has derived the following HLR function to predict the half-life "HLR" which is a learning point and minimize the rate of forgetting. As for this research, Apart from Duolingopaper(PDF file)It is announced as.


In addition, Duolingo used the German psychologist Hermann Evbinghouse to devise a learning schedule that can enhance the learning effect "Oblivion curveI am referring to it. In the 1880's Ebbinghaus memorized a series of meaningless alphabets to the subjects and measured the time required to memorize that alphabet again when the memory was lost, the time required for re-learning over time It was discovered that it will be long. Graphing this is an oblivion curve, which means that it is desirable to repeat at shorter intervals for fast paced, ie difficult to learn items whose memory is lost according to oblivity curves.


So, Duolingo has changed the pace until the next review depending on whether it was correct or incorrect when I learn iteratively. In other words, while increasing the interval until the correct answer appears again, if it was incorrect it means that learning is effective by shortening the interval until next time.


In addition, Duolingo scoring easy words that can be acquired relatively quickly and difficult words that are not easily fixed in memory, by analyzing information obtained from users. The learning difficulty of this word is used for the user's learning schedule by using it as a variable of HLR function.


HLR function system has the lowest error rate than other algorithms. The current HLR function system seems to succeed in lowering the error rate to about half compared with the Leitner system which was used in the initial version of Duolingo.


According to Duolingo, the introduction of the HLR of the half-rate prediction system which increases learning effect, the user's exercise retention ratio increased by 9.5%, the number of exercises increased by 1.7%, the total usage increased by 12% . Duolingo seems to have succeeded in attracting users by constructing a system that brings high learning effect backed by data.