Duolingo – Learning the language of AI
Perhaps the most common new year resolution (behind losing weight) is learning something new- maybe a new language. As with losing weight, the resolve to learn a new language rarely outlives January. Duolingo, backed by AI, is here to make sure that you don’t require the earth to complete a revolution around the sun to engage in learning a new language or require as much effort!
Duolingo
After creating CAPTCHA and making sure that we are not bots, renowned Carnegie Mellon University computer scientist- Luis von Ahn– turned his attention to language learning. With a mission to “make education free and accessible to everyone in the world”, Luis von Ahn created Duolingo. Duolingo is a cross-platform product primarily for language-learning. Duolingo offers its 300 million users 94 different language courses- including English, Hindi, Spanish, Mandarin and even High Valyrian.
With the hire of Burr Settles, Duolingo embraced AI. With “interests at the intersection of language, AI in tech, and cognitive science” Settles was the perfect fit for Duolingo. Soon after joining, Settles and his team, undertook the challenge of bringing AI into Duolingo.
How Duolingo Creates Value with AI
Placement Test
When a new learner comes to a platform, they are usually faced with the question of where should they start. Some folks might have a basic understanding of the language, some might have none at all and most usually fall somewhere in between- but where exactly?
Duolingo’s AI-driven adaptive placement test probes users with questions and determines exactly where they should start.
Spaced Repetition and Lag Effect
Spaced Repetition is where you repeat short lessons at various intervals rather than repeating the same information in a short time frame. Spaced repetition has been proven effective if one could determine if and how often to repeat a particular lesson. Lag Effect tells us that users can improve more if the gap between practice sessions is gradually increased. However, we need to determine how much lag is required for each user for each lesson.
Duolingo employed AI fed by the tremendous amount of data they had about their users to tackle these questions. For each word in the curriculum, Duolingo’s AI keeps track of how many times you’ve seen it, how many times you’ve gotten it correct, the modes under which you got it correct, and how long it’s been since you’ve practiced it. The AI uses this information coupled with its learnings from the 300 million other users on Duolingo to determine how much space and lag is warranted for each lesson for each user.
Chat Bots
One of the greatest challenges to online conversational language learning is the value of offline conversations- to learn to converse in a language, you need to converse in the language. How can Duolingo aim to provide conversational grasp of a language without conversation? How can Duolingo offer conversations at scale to 300 millions users in 94 different languages? Enter AI enabled Chat Bots.
Duolingo created chat bot characters, designed to respond differently to a range of possible prompts. If the user ever got stuck, they could hit the “help me reply”. Duolingo’s chatbots are Renèe the Driver, Chef Roberto, and Officer Ada, with the promise of more characters coming to the app soon. Learners can use them to practice French, Spanish, and German, respectively.
Challenges and Opportunities
One of the foremost challenges and opportunities with AI is time. The cross network effect is strong in every AI system. The more users engage with an AI product, the more it learns and the better it gets. This is a huge opportunity for Duolingo. However, to get more users, the AI product would have to be accurate and effective. AI is never as good today at it will be tomorrow and therein lies the challenge. Duolingo will need to incentivize users initially to engage with the AI so that the it can become more reliable.
Duolingo also faces a unique challenge in having to provide similar features for a variety of languages. To provide the same offerings for its ever expanding list of languages is daunting. However, every language they leave behind is a potential for competition. While every language is different, there might be underlying structural similarities Duolingo can identify to decide which languages can be quickly AI enabled. Prioritizing based on demand alone might not necessarily reap the most value.
Conclusion
With travel becoming more desirable and affordable, more people are more exposed to more languages than every before. Duolingo has wisely leveraged AI to make their offerings more effective and efficient.If Duolingo is able to continue on this trajectory, I believe that Duolingo is poised to thrive in these times of wanderlust.
References
- https://venturebeat.com/2019/07/05/how-duolingo-is-using-ai-to-humanize-virtual-language-lessons/
- https://www.wired.com/brandlab/2018/12/ai-helps-duolingo-personalize-language-learning/
- https://www.zdnet.com/article/how-duolingo-uses-ai-to-disrupt-the-language-learning-market/
- https://www.theverge.com/2016/10/6/13188326/duolingo-language-tutor-chatbor-ai-announced
- https://www.pcmag.com/news/nervous-to-practice-a-new-language-try-ai-duolingo-says
The unique use of chat bots is a great way to use AI to create value for your users. I think this is a tool that really differentiates Duolingo from other free educational tools that are available.
Indeed! I think it really adds a unique value but I can see it being copied by competitors. However, unless the competition has great differentiators, I suspect people will stick around purely out of inertia and brand recognition.
Thank you for this article!
I’m a Duolingo user and I was not aware of the placement test. I believe I started using the app before this feature was rolled out. I always wished that Duolingo should develop this test – a form of competency based testing – given that it has the data required to do so. I will check how I can take this test – thank you 🙂
Right now, each category has 5 levels that you can test out of. I feel the next step for Duolingo would be to venture more into personalized learning plans for each individual based on their placement test, the speed at which they’ve been learning, the number of times they’ve failed certain levels, etc. This could also be complemented by a survey at the beginning that asks users what they wish to focus on (work conversations, etc.)
Me neither! I was really impressed by how they are focused on making things better for their users.
Great post that helps be better understand the experience that my fiancee had with Duolingo! While she was impressed with placement test’s capabilities to analyze her Korean (she can speak some but doesn’t know how to read/write), she found the product frustrating in a couple of ways that could perhaps be improved through further AI applications. The first is that there was that the platform relies too much on multiple choice (probably creates easier data for AI to use), when free response (spoken and written) is a better indication of learning. Perhaps Duolingo needs to improve its Natural Language Processing capabilities to make this possible. Additionally, the owl was annoying when it was trying to get her to reengage with the platform – I wonder if Duolingo could use AI for better customer engagement tactics?
Congratulations on the engagement!
I wonder if the annoyance your fiancee felt with multiple choice questions had been one of the influences behind going towards the chat bots.
Very interesting article, Krish. I really like that you picked education technology as the theme for the AI application. At my online education company, Upgrad, in India, we also played around a lot with AI to figure out how to create a better learning experience for users. I think we’re still in the early stages of AI application in education – I think it will become very interesting as soon as we’re able to create educational content in a faster manner – the main issue that we’re currently seeing RE individualized, AI-powered learning is that it requires up to 10x of content (given that there are so many different learning paths for users). I look forward to neural networks that can create assignments and lectures themselves basis the learner’s answers and inputs into the model.
You are at upGrad in India ?! So cool ! I am a graduating student from the Ed school doing my masters in EdTech so upGrad has been in my radar ! Would love to talk to you about it !
Thanks for sharing this – Duolingo is such a cool product, and this post explored me to product features I wasn’t yet aware of!
I’m sometimes surprised by the new languages that Duolingo is able to launch; my expectation would have been that it would be hard to take many of the learnings from one language model and transition it to the next, and therefore the new language units, until they generate enough of their “own” student data, could deliver experiences sub-par to what people have come to expect from what they have experienced in the more established Duolingon modules. However, the Lemonade insurance protagonist said that there are actually a lot of NLP learnings that can transfer from one language to the next (in reference to using the models they had trained on English for German- and Dutch-speaking customers), so I wonder if Duolingo is using similar models and technology to get a jump-start on training new modules!
Im sure your concerns do exist. Creating new language lessons is one of their big investments. The chatbot is lemonade is easier because it is designed to be knowledgable about one particular narrow area and is the mover in most conversations. For general conversations, the difficulty would be higher. However, it would not be as high as starting from scratch. It is interesting to see ideas- like chatbots- creating value across industries in different ways.
Thanks for a great article! Really interesting company. I wonder what the limits are for its expansion and applications of Duolingo’s learning algorithms. Could they be used to teach people how to play instruments? Math and coding lessons? Safety trainings for employees in dangerous workplaces? Plays for sports teams? It seems to me that a lot of the techniques leveraged by Duolingo apply to other subjects as well. A partnership with educational platforms could go a long way — especially with remote and distance learning becoming increasingly relevant.
Additionally, the chatbots may be useful in supporting customers of other businesses across languages and may offer an alternative revenue stream for the company. I also wonder if there is a speech to text learning that can be incorporated into the platform in a similar way. It will be interesting to see how competitors like Rosetta Stone respond to Duolingo’s increased market penetration.
Duolingo is a language learning platform but I can see it becoming a learning platform in the long run. The question is really to explore or exploit. They have room for growth in either direction.
This is really interesting–thank you for sharing! Given this kind of precision in placing people based on their true capabilities, I wonder if this technology can be applied in other contexts to shift our education system away from standardized tests, which have drawn fire over the years for inaccurately representing people’s skill level. I am also curious to see whether Duolingo has normalized and compared the performance of those who were placed using this technology versus those without.
Duolingo has already started to replace standardized english language proficiency tests like IELTS and TOEFL. These are mandatory for international students for admission and super expensive (~250$ each). They are disrupting this space with their own alternative (for 49$) which is presently accepted at over 300 schools across USA- like NYU and Yale.
Given this move, I wouldnt it put it beyond the space!