Language learning has surged during the pandemic. Duolingo, which is synonymous with gamified language learning, saw its fastest growth period this March, with a 101% global increase in new users. From those who simply have more time on their hands to students trying to keep up during the pandemic school year, the app is a huge boon. All that extra data isn’t going to waste — because Duolingo invested early in AI, the app keeps getting better as it grows beyond the 30 million monthly active users reported in December 2019.

At Duolingo, we believe it’s important to use data to drive our decisions. This includes decisions about how best to teach languages. You probably aren’t thinking about it while you’re doing your daily Spanish lessons, but in the background, we have dedicated teams of volunteers and staff working on improving the courses we teach you and developing new versions of the courses, which we call “trees”. Once these trees are done, we release them as A/B experiments to a portion of learners and track the impact on engagement with the app. This lets us ensure that the changes we made would actually be beneficial to the learning experience on Duolingo.

However, one downside to this approach is that it is a very retrospective measure: it gives us information on how a tree performed after we have finished making changes. But what if we wanted to see how we’re doing on improving a course while we’re still working on it?

Have you ever wondered how Duolingo figures out whether your answer counts as right or wrong? Let’s say you’re learning Portuguese and get “Eu tenho gatos demais” (literally I have cats too many) so you type “I have too many dogs.” You'll get a red ribbon and a sad Duo. How does Duolingo know? Although this example is pretty simple, deciding which translations should be accepted turns out to be a complex problem!

In this post, we’ll talk about why translation grading is so hard, and we'll report on a recent research project we organized to develop sophisticated artificial intelligence solutions to this problem.

Daily practice is essential for language learning, so Duolingo helps learners stay on track by sending daily practice reminders. In fact, Duo's persistence is so well known that it's even become a popular internet meme. And let's be honest, most of us have probably swiped away one of these notifications...and probably felt a bit guilty in the process.

But, have you ever wondered how Duo decides what message to send? Well, last year Duolingo’s Machine Learning Engineers built a really neat AI system to find the perfect reminder to send each learner each day! We recently published this novel algorithm in a paper and short presentation at the Knowledge Discovery and Data Mining (KDD) Conference 2020. In this post we’ll take a peek at the AI behind these notorious notifications.

Duolingo is the world’s most popular language learning app with more than 150 million users (at the time of this writing). At the core of the Duolingo experience, users learn in bite-sized lessons which consists of several interactive exercises (internally, we call them “challenges”).

So, for any given lesson, which exercises should a user see and in what order? This is the responsibility of Session Generator, our backend module which gets data from one of our 88 language courses (and counting!) in the Duolingo Incubator, sprinkles some machine learning magic, and proceeds to serve a sequence of exercises tailored to the needs of each of our millions of users.

Duolingo works to improve the performance of our apps for all of our learners. As Duolingo has grown over time with new courses, features, and millions of additional users, several teams are working behind the scenes to keep things running smoothly. This post will focus on a backend feature called Request Tracing that has led to significant improvements in performance.

Duolingo is steadily moving from a monolithic architecture (a few large, tightly-coupled services) to a microservice architecture (many small, loosely-coupled services). This has its advantages, but also has tradeoffs, including operational complexity and observability.

At Duolingo, striving for excellence is one of our core operating principles. We strive for excellence in our mobile app by devoting ourselves to a high-quality experience for every learner. As part of our commitment to building for a globally diverse audience, we've dedicated our efforts across the plethora of languages, operating systems, and mobile devices that our learners use. We run hundreds of A/B tests simultaneously, and staying on top of all possible versions of the app a learner might see depending on their user interface (UI) language, course, and device can become overwhelming. Until last year, there was no easy way to visually compare the same learner experience across multiple languages and devices to ensure a smooth, enjoyable experience for everyone.

At Duolingo, our goal is to make language learning fun and effective. We think the best education should be full of play, so we're constantly developing new features that make learning new things — and practicing old things — feel like a game! At the same time, we're serious about taking a scientific, data-driven approach to all of our products, and about sharing what we learn with the world. In this post, we'll take a look at the science behind the Duolingo skill strength meter, which we published in an Association of Computational Linguistics article earlier this year....

我学会用荷兰语说的第一句话是 "我们用棍子打死他们"。这不是语言课第一周的标准内容，但这又不是普通的语言课。作为一名研究16、17世纪文献的历史学家，我上的是一门学习阅读当时荷兰语的专业课--因此，我们没有学习如何谈论兴趣爱好或询问去火车站的路线，而是直接跳进了所谓的荷兰黄金时代的文章。