Embedded Files
The Role of Artificial Intelligence in Self-study: A Case Study of Duolingo
Joshua Kruger Haasbroek
2023/08/05
Introduction
Recent years have seen the rapid advancement of artificial intelligence (AI) and it has had an impact on all aspects of modern society. It is evident that the education sector and AI will continue to converge. With this in mind, AI provides endless possibilities of uses in aiding students. One of the many possibilities is that AI could identify the areas in which students need improvement and adjust their learning material to suit their level of understanding and educational needs.
An app that claims to support the learning process of acquiring a second language or practicing language with the help of AI, is Duolingo. Duolingo is an American educational company that has developed a language learning app and website with a variety of languages to select from, including English, Spanish, Korean, Japanese, etc. The app provides easy bite-sized, gamified lessons with interactive animated exercises and well recorded audio that matches the intonation of native speakers. Moreover, the app takes pride in having over 300 million users worldwide taking advantage of the free personalized lessons offered by the service. Duolingo’s engaging design makes it ideal as a tool for language learners seeking to improve their skills outside of the classroom.
This paper aims to explore the mechanics that power Duolingo, and more importantly, assess how effective it is for the educational needs of students seeking additional support in their studies, especially students studying a second language at a tertiary level. The paper will first discuss the inner workings of Duolingo and AI, and then analyze data collected through questionnaires, before finally reaching a conclusion.
Duolingo and AI
The prospect of using AI to create high-quality learning experiences is an exciting feature. Duolingo makes use of a combination of human expertise, algorithms, machine based learning (ML) and AI. Before diving deeper into the subject at hand, let's identify the differences between algorithms, ML, and AI. Algorithms are automated programming that functions with if + code to execute an action. The complexity of the algorithms depends on the amount of functions inserted in the code. Although ML and AI are often used interchangeably, they differ in how they operate despite being built upon algorithms. Basically, ML is algorithms and statistical models that operate with structured data to complete tasks without the input from humans, whereas AI can be given unstructured data and give structured output. In essence, AI imitates human cognitive abilities to produce coherent decision making. However, it should be noted that ML is a subdivision of various AI technologies and enables AI to improve its performance.
Furthermore, AI is divided into three main types: rule based AI, learning based AI and limited memory AI. Rule-based AI uses predefined rules to make decisions or find solutions. An example of rule based AI is a chatbot that answers frequently asked questions. Learning based AI is AI that improves its performance overtime using ML. Learning based AI could be further divided into supervised learning and unsupervised learning. Supervised learning uses labeled data to make predictions or classifications, for example image recognition, whereas unsupervised learning uses unlabeled data to find patterns in the data, for example customer segmentation. Unsupervised learning systems track and analyze the behavior of a cluster of customers. Lastly, Limited memory AI has the ability to recall past experiences to make decisions in the future. ChatGPT is an example of limited memory AI. Grasping the core mechanics of AI technology is vital in dissecting Duolingos operational features
As mentioned before, Duolingo uses a combination of human expertise, MI and AI. According to the Head of AI Duolingo, Klinton Bicknell, and Bozena Pajak, VP of Learning & Curriculum at Duolingo, the materials of the lessons are based on four stages, namely the curriculum design stage, raw content, exercise creation stage and personalization stage. The authors explain that the first three stages are put together by human experts with AI helping staff to design the course. It is within the final stage, the personalization stage, that the MI and AI technologies truly shine.
Users are unknowingly introduced to Duolingo's AI system when they are taking a placement test after launching the app for the first time. The app starts quizzing students’ knowledge of the selected language and places students in a level that matches their abilities. In this way, students with prior experience of the language can skip over basic material and are placed at a level that does not feel like a waste of time, but also does not frustrate users with new complex materials. This simple introduction remains just the tip of the AI iceberg.
The most significant use of AI in Duolingo comes in the lesson personalisation stage, where a variety of scientific fields, including Natural Language Processing, ML, Cognitive Science and Computer Sciences, are utilized. It should be mentioned that the sources use the terms AI and MI interchangeably. To avoid confusion, it would be preferable to refer to the discussed technology as MI, unless there are clear references to Limited Memory AI that is comparable to ChatGPT.
Duolingo refers to its MI technology as “birdbrain.” The key feature of birdbrain is its ability to personalize learning to each individual’s level. The MI is used to actively monitor learners’ comprehension, and identify areas where they struggle. Birdbrain analyzes the user's data log to predict the chances of users answering correctly and proposes reinforcing exercises to learners accordingly. Bicknell notes that birdbrain gives learners the ability to automatically skip language material that they are already familiar with and skip to areas they find more challenging. As students use the app more frequently, birdbrain is able to make stronger connections with students' needs and make exercise recommendations accordingly. Bicknell also added that students are more likely to return regularly thanks to the personalized lessons.
Another fascinating aspect of Duolingo’s MI is the combination of Machine learning and psycholinguistic theory which led to the development of the half-life regression model. To put it simply, psycholinguistic theory describes how humans obtain, generate and receive language. Duolingo incorporated two known established learning methodologies, spaced repetition (studying small amounts of workloads spread over a period of time to improve long term retention) and lag effect (information is gradually better stored when there are time gaps between lessons). However, these methodologies, developed in the 1960s and 1980s, are inadequate when used with MI and mass consumers, as the binary nature of MI models neglects the individual's learning needs. To address this, Burr Settles and Brendan Meeder developed the half-life regression model (HLR), which they describe as:
“a novel model for spaced repetition practice with applications to second language acquisition. HLR combines psycholinguistic theory with modern machine learning techniques, indirectly estimating the “halflife” of a word or concept in a student’s long-term memory”.
HLR theorizes that learners tend to forget half of the information they acquired if they are not often exposed to the material, and Duolingo employs HLR with its MI technology to revisit and adjust targeted material according to students' needs. By combining the HLR model and birdbrain, patterns would look different for students who are more familiar with a language compared to a student who is less comfortable with the same language. Two students of different levels would see different lessons when using this approach. Research conducted by Settles and Meeder shows that the HLR used through Duolingo has reduced errors by 45% when compared to other models, and that HLR improved daily user activity by 12%. The MI and HLR model lay the foundation of Duolingo’s personalized lessons, but what more could we expect from Duolingo and the advancement of AI?
Duolingo recently introduced Duolingo Max, a language learning platform that makes use of limited memory AI. Duolingo Max incorporates OpenAI’s GPT-4 technology, allowing users to engage with an interactive AI chatbot when they are practicing natural conversations,such as making vacation plans or ordering food at a restaurant. This feature enables learners to request corrective feedback and clarification on mistakes, thus enhancing their language learning experience. Unfortunately, the Max version is only available in select countries outside of South Korea, and currently supports only English for Spanish and French learners. Additionally, the new product comes with a hefty price tag of $29.99 per month or &167.99 per year.
In short, combining Duolingo’s current MI technology and HLR model with their own specialized limited memory AI can produce new and innovative ways of language learning that could greatly benefit students committed to self study. However, the high price range of the app may make the app inaccessible to university students who have not yet obtained a high-paying job. Moreover, the recent introduction of Max makes it challenging to measure the effect of limited memory AI when combined with self study at a tertiary level. Nevertheless, the paper still aims to evaluate the effectiveness of Duolingo’s MI technology, and in the next section will discuss how students perceive the app.
Duolingo in the field
Duolingo sounds great in theory, but the true question remains: how effective is it for students of Gyeongsangnam National University? To answer this question, ten students studying accredited courses and attending nightly conversation classes (CVE) at GNU volunteered for the study. After a trial period of two weeks, the students were presented with a questionnaire to assess the usefulness of the app.
Although the study attempts to provide accurate data, it should be noted that two factors could potentially obscure the findings of the paper. First, students had limited time to experiment with the app since the trial period coincided with the midterm tests. Secondly, the questionnaire was presented in English rather than the native language of the students. In order to address this, the questions were formulated to be short and straightforward. The questions consisted of two sections using closed-ended questions and open-ended questions. The open-ended questions were ranked on a scale from 1 to 5, where 1 represents a very negative response and 5 a very positive one. The open-ended questions asked the students to express their opinions in complete answers.
The data yielded interesting results. The students stated that they wanted to use the app to improve their English skills, citing grammar, speaking, and listening as significant areas of improvement. Ninety percent of the students responded positively, expressing enjoyment while using Duolingo and recommending it to friends. The usage time was varied but the average was twice a day. The majority of students used the app while at home, while a few used it during transit. Additionally, the majority of the students were highly positive that the app improved their English skills. They felt satisfied with the learning experience and indicated future use of the app after completion of the questionnaire. Students mentioned that they enjoyed the listening feature of the app, which provides a good way of listening to pronunciation. They also reported experiencing improvement when prompted to answer incorrectly answered questions from before. Moreover, students pointed out that the storytelling combined with cute characters and an interactive community made learning a fun experience.
In contrast to these positive results, the data also reveals many downsides of Duolingo. Birdbrain, the MI technology that revolutionized the app, fell flat when students were asked if the lessons and exercises were personalized to their levels. Although students' answers were mixed, it is clear that the app only received an average score when it came to the aspect of personalized learning. Students commented that it is helpful for beginners, but they wanted more advanced lessons and more exercises for high-level students. This data shows clear discontent between the app and its ability to elevate students to their appropriate levels.
Students also noted other criticisms that hampered their enjoyment of the app. One of these is the heart system, which is required to do exercises and progress through the lessons. Whenever users make a mistake, they lose a heart. On average, students may have four to five hearts a day, and additional hearts can be purchased with in-game currencies or by subscribing to the premium service. Another general criticism is the amount of advertisements that appear while using the app. Minor criticisms include the app’s lack of comfort and need for more systemic learning context. Although the lessons do follow a structured approach, they are mostly based around exercises without much study material.
From the data it is clear that Duolingo's birdbrain MI helps students to practice language in areas where they struggle and it is an enjoyable gamified app, but there are serious cons when it comes to personalized lessons.
Conclusion
The paper first sets out to examine the AI mechanic and structure of Duolingo. Duolingo touts its state-of-the-art AI technology as revolutionary with the process of acquiring new languages. However, the term AI has become an umbrella term for various technology companies without specifying the type of AI being used. There are clear differences between types of AI and the purposes and functions they serve. Without clear differentiation between types of AI, it is easy to market and oversell products as limited memory AI, while in reality, when they are actually using existing MI technologies that have been around for several years. It appears that the term AI has become an overused hype word.
Regarding its MI, Duolingo keeps track of students' mistakes, and with the aid of the HLR model, helps them to practice their errors until they have mastered the specific language focus. However, it could be argued that birdbrain improves on the format of language learning apps rather than revolutionizing the learning process.
The introduction of Duolingo paid limited memory AI, Duolingo Max, is a missed opportunity for the company to excel its services. Max is considered high priced for students who just want a casual experience from the app and would only cater for users who are seriously using the app for language acquisition. It could be argued that ChatGPT delivers a similar experience for free at this time.
Duolingo is a great and enjoyable language learning app that appeals to learners who are interested in a new language thanks to its easy-to-use gamified design. Despite this, there are several significant criticisms of the app. It fails adjusting to learners' levels, despite clearly marketing the opening test as a benchmark for adjusting students to their intended level. This highlights some serious flaws with the MI technology, and birdbrain needs to take more external factors in consideration when assessing the level of students. This could be addressed by giving learners more options when it comes to their level rather than relying solely on AI assessment.
Is Duolingo a good choice for GNU students? I recommend it for students who need to improve their basic English skills. Instructors could suggest the app to students who are struggling to keep up with the rest of the class. Duolingo provides a fun learning experience for students while refreshing their English skills, making it a perfect tool to use for when they feel bored. The MI technology takes care of the rest, helping students to improve their English abilities over time. With consistent use, the app may even lead to improved test performances and assignment outcomes.
Sources
Malik, A. 2023.03.15, Duelingo launches new subscription tier with access to AI tutor powered by GPT-4, TechCrunch,
Marr, B. 2020.10.06, The amazing ways Duolingo is using artificial intelligence to deliver free language learning, Forbes, https://www.forbes.com/sites/bernardmarr/2020/10/16/the-amazing-ways-duolingo-is-using-artificial-intelligence-to-deliver-free-language-learning/?sh=6e6f64bf5511
Quinyx, 2023, What’s the difference between AI, ML, and algorithms, https://www.quinyx.com/blog/difference-between-ai-ml-algorithms
Pajak, B & K. Bicknell, 2022.9.14, At Duolingo, humans and AI work together to create a high-quality learning experience, Duolingo Blog, https://blog.duolingo.com/how-duolingo-experts-work-with-ai/
Perenandam, C. 2018, AI helps Duolingo personalize language learning, Wired, https://www.wired.com/brandlab/2018/12/ai-helps-duolingo-personalize-language-learning/
Protalinski, E. 2020.08.18, How Duolingo uses AI in every part of its app, VentureBeat, https://venturebeat.com/ai/how-duolingo-uses-ai-in-every-part-of-its-app/
Sawers, P. 2019.07.05, How Duolingo is using AI to humanize virtual learning lessons, VentureBeat, https://venturebeat.com/ai/how-duolingo-is-using-ai-to-humanize-virtual-language-lessons/
Settles, B. & B. Meeder, 2016, ‘A trainable spaced repetition model for language learning’, Proceeding of the Association for Computational Linguistics, Vol. 1: Long Papers, pp. 1848-1858, https://research.duolingo.com/papers/settles.acl16.pdf
World Intelligence Property Organization, Duolingo: learning a new language with an AI owl, https://www.wipo.int/wipo_magazine/en/ip-at-work/2022/duolingo.html
Page updated
Report abuse