The Absurdity of Duolingo Sentences

I’m currently trying to learn Spanish, and since language courses are unfortunately extremely expensive I decided to use an app for now. In comes Duolingo. I’m guessing that many of you have used it at one point or another, but I’ll give you a quick introduction to it anyways.

Duolingo is a language learning app founded in the US and was released in 2012. Once downloaded, you can currently choose between 43 different languages, of which 39 are available for English speakers, and all this of free of charge (although there is of course a premium version, in case you feel the need to access more functions and want to get rid of the ads). They constantly work on expanding and improving the courses they offer, and have about 49.5 million monthly active users.

Now I have to say, the idea is great, and apparently the courses are all based on proper research. But somehow, Duolingo never fully convinces me. I’ve used Babbel before, which worked much better for me, but unfortunately at this point that is almost as cost-intensive as an in-person course. One of the reasons I never quite get along with this threatening green owl is the seemingly random absurdity and uselessness of the sentences it tries to teach me. If you want to have a laugh, I advise you to watch the following video, which is basically a collection of the most hilarious sentences this app has to offer:

Matt Rose, Duolingo Sentences are Madness

For a long time, I was convinced that these ridiculous sentences are just created by AI, there is actually a strategy behind it. 

In fact, Duolingo works with a combination of human experts and Artificial Intelligence in a process that is divided in four stages: curriculum design, raw content creation, exercise creation, and lesson personalization. While the first stages are still mostly designed by humans, the further along they go in the process, the more the AI takes over the work.

The first stage is carefully planned out by professional curriculum designers who decide what will be taught and at what point, following the CEFR standard while taking into consideration the individual language background of the user. The reason why some of the sentences truly are random is the mere fact that a specific set of words is meant to be taught at a particular time, but some grammatical concepts that would generally make sense of them is only gonna be included later in the course. So, while they so look at vocabulary, grammatical structures and individual concepts, they do not match them well together at this point.

All this to say, a curriculum designer is looking at the individual concepts, vocabulary, and structures that make up a sentence, but they’re not stringing them together… yet.”

Bozena Pajak and Klinton Bicknell, Duolingo

In the second stage the ‘raw’ content for each lesson is created by human experts, out of which specific exercises will be stitched together at a later point. In order to guarantee that this step will be carried out more efficiently, they are assisted by AI. This content can range from paragraphs, to sentences or dialogues that exemplify new concepts and vocabulary. The AI helps to create a variety of possible translations of the sentences for cases in which several answers could be correct. At this point the example sentences will be written, some of which are standard ones (‘I like sports.’), while others would certainly raise an eyebrow (‘My horse likes TV.’) This step is apparently mostly about having a wide variety of sentences and phrases, and not so much about the actual meanings thereof.

Now that the ‘raw’ content is created, as pool of interactive exercises needs to be created from it, which is mainly done using algorithms that automatically combine the content to exercises. These could ask you to, for example, complete a sentence, fill in a missing word or translate a sentence out of the provided words. 

exemplary exercises: complete the sentence, form the sentence, and fill in the missing word

Finally, the lessons need to be personalized to the user’s specific abilities and needs. For this, several AI models work alongside each other, mostly using the so-called Birdbrain model, which figures out which course content needs to be revised, among other things.

But why are some of these sentences so absurd?

Sentences like “I am eating bread and crying on the floor.” do generally make sense, and currently express my mood pretty well. However, these are not the kind of sentences you would actually need to learn or encounter in any other language courses (that I know of).

Don’t we all?

The reason for that is quite simple: Duolingo wants to engage its users, and apparently these quirky sentences are more memorable than our normal day-to-day language, functioning as anchors for the newly learned grammatical structures, and enabling you to use them more flexible. And in the end, it’s all about the fun.

Quirky sentences also have a hidden superpower: they are memorable! They work as a grammatical “anchor,” helping you remember key examples of essential grammar concepts. 

Duolingo

References

James Leow, Jacqueline Bialostozky and Cindy Blanco, “How Silly Sentences can help you learn,” Duolingo Blog, 3 March 2021.

Bozena Pajak and Klinton Bicknell, “At Duolingo, humans and AI work together to create a high-quality learning experience,” Duolingo Blog, 14 September 2022.

duolingo announces 51% bookings growth and accelerating user growth in second quarter 2022; raises full year guidance,” Duolingo investor relations, 04 August 2022.

“the complete list of every duolingo language,” duoplanet, 17 June 2021.