Escaping Duolingoland
Toucan, Fluent, and context-first language learning
Hello and welcome to the first installment of Naturally Processed. You’ll note that there is a lot here. In the interest of time, I have ignored my generous editor’s advice to split this up into several posts. Future newsletters will be shorter.
Can we learn a language by surfing the Web?
The browser is an underestimated platform for educational applications. It’s easy to forget the thing is even there, mediating and rendering the entire Internet. I think that’s in part because it just works. Most of us barely notice the ease with which it supports our interactions with a bewildering quantity of information. We take its dutiful recovery of our open tabs after a restart for granted. We stay blissfully ignorant of the ways it enables surveillance capitalism. But for those interested in building applications for learning, web browsers are extensible in exciting and under-appreciated ways.
Their extensibility lends them well to “meta” applications composed from the content they render. Think tools that extract, annotate, explain, translate, and even gamify underlying information. Browsers have built-in speech-to-text and text-to-speech capabilities, which enable audio and voice driven features. They also manage local databases, which extensions and web clients use to store non-trivial amounts of structured data on our machines. That statefulness enables things like analytics, progress tracking, and mnemonic techniques - all without a need to phone home to a centralized web service or API.
Toucan and Fluent are browser extensions that leverage these features to help users acquire vocabulary in a foreign language. As tools for learning and as pieces of software, they have an enviable elegance. They work by simply replacing words with a target language translation. Load a website, and suddenly a handful of words are transformed, say, from English to Spanish, in a satisfying cascade down the page.
Pleasant pastel highlights draw the eye to the chosen words’ new incarnations. Hover over a translated word, and a little popup card appears, offering more features: mildly creepy but still useful incantations in robot voice (both apps), often slightly off-the-mark dictionary definitions (Toucan), handy inline recall quizzes in “test me mode" (Fluent), games and exercises (Toucan), error reporting (both), and ads (Toucan).
What logic dictates which words are transformed, and which are left alone? It’s not entirely clear. While both apps have settings for controlling the relative quantity of translated words on a page, neither is explicit about their selection criteria. Toucan seems to do mostly nouns and noun phrases, while Fluent seems to stick to adjectives and adverbs. Neither seem to have a robust way of ascertaining the user’s level of lexical fluency in the target language - a hard problem, no doubt - but Toucan’s pop-up has a little button that lets you inform it when you already know a word.
Toucan reserves its games and highest word density setting for premium subscribers ($4.99 per month with a year-long commitment, $8.99 monthly); Fluent is still figuring out what a Pro subscription will entail, according to its website. Chances are good it will involve something akin to Toucan’s cute games and exercises. These make use of a technique called spaced repetition, which is a system for scheduling practice that research has demonstrated helps optimize how well we remember what we’re trying to learn given the time we spend practicing.
Most language learning apps make use of spaced repetition systems (SRS), and they are an active research area in the cognitive and computer sciences. Classical approaches to SRS, such as the Leitner and Pimsleur methods, used algorithms with fixed parameters for scheduling. In 2016, researchers from Duolingo published a clever paper that generalized these approaches, casting them as cases of half-life regression, a model with learnable parameters. The paper states that the company deployed HLR for all users; it attributes at least some of its impressive growth to the algorithm. A popular open-source flashcard app called Anki also implements SRS, and has a cult following that includes medical students and forward-thinking luminaries like Michael Nielsen, whose essay on Anki and SRS is worth a read.
I haven’t dropped on Toucan’s games yet. I don’t know what flavor of SRS they implement, or if it is as sophisticated as half-life regression. I do know the games are cute, because Toucan teases freemium users with a few minutes of free game play before rate-limiting them. My current strategy is to keep both extensions on at once. The resulting text is something syntactically English, and lexically Spanish-esque. An occasional effect of using the apps simultaneously is that some words disappear entirely - lost, assumably, to some competition between the two. But because the apps currently handle different parts of speech, combining them works pretty well.
That’s assuming I’m in the mood. There are times when I don’t feel like stopping and puzzling over the periwinkle blue Spanish spans, like when I’m doing deeply focused or time-sensitive tasks. Fortunately, each app lets you turn it off. Fluent makes this easy, with one click at the top right of the main extension modal. Toucan’s off button defaults to a period of 5 minutes, cheekily hiding Indefinitely off under a slider, a click, a dropdown, and another click.
I’m excited about these apps, for reasons I will expound on later in this essay. That said, it’s still early days for these products, and users should approach them with a healthy amount of skepticism. While Fluent sticks somewhat safely to adjectives and adverbs, Toucan’s more ambitious translations and definitions leave room for improvement. Let’s walk through some common issues you’ll have to contend with if you’re using Toucan.
Missed polywords
Toucan consistently fumbles polywords. These are lexical chunks consisting of more than one word, but which function as a unit. They often have a certain arbitrariness to them, although they are not necessarily figurative. For example, we say “washing machine,” or “washer,” but not “machine washer,” or “mechanical clothes washer,” or “automatic garment cleaner.” While these other options do communicate the idea of a mechanical object which performs a cleaning operation on clothing - and being able to compose phrases like “automatic garment cleaner" in a foreign language is a useful and necessary skill - there is nonetheless a specific, fixed, lexical chunk or two which represent the idea in the language over all the others. Learning these chunks is important for fluency.
In Toucanese Spanglish, “natural language processing” in English becomes “natural idioma processing,” rather than the more fulsome "procesamiento de lenguajes naturales.” The same goes for “machine learning,” which in Toucanese Spanglish is rendered “machine aprendizaje,” rather than one of a number of the available idiomatic options, such as “aprendizaje automatico” or “aprendizaje de maquinas.”
Personally, if I’m going to expend energy using a tool to learn vocabulary, I want one that can map whole lexical chunks in English to whole lexical chunks in Spanish. These should just be learned and practiced as whole items. Indeed there is a whole pedagogical school of thought developed around this basic idea, and the importance of lexical chunks. I also want to be able to trust that the tool gets it right, most of the time, without having to consult references like Wikipedia and Google Translate, or phoning a friend.
While we’re here, my wish list for the pop-up is long… a toggleable Spanish definition would be nice. Some other example sentences would be classic. Synonyms, antonyms, and related words would all be handy. And I'd love to know sociolinguistic and pragmatic details. Does the Spanish-speaking machine learning community actually use “aprendizaje de maquinas?” Maybe they tend to use an acronym, like the English-speaking ML community commonly does? Maybe they often just say the English phrase? There is no reason an app couldn’t one day supply such information about given lexical chunks, assuming it could source it reliably.
Word sense disambiguation errors
Toucan struggles to pick the right sense of a selected word in English before translating it. This is the most frequent error I have observed the app make, and probably the biggest opportunity for improvement. Consider the following example, drawn from the quotidian surfing activities of yours truly:
I don’t actually know Spanish, so I can’t make assertions about the correctness of this translation without referring to some authoritative source. We see that the definition offered in the pop-up doesn’t make much sense for this technical tutorial. We might also note that “solicitud” shares a root with the English words “solicit” and “solicitation.” Neither of these are commonly used terms in the software context, but a related term, “request,” is quite common. We can compare this to what Google Translate gives us for the passage:
If we trust Google here, then we can say that Toucan has failed to distinguish between different senses of the English noun “application,” which, in the context of software, maps to the Spanish cognate “aplicación.”
A similar word sense error occurs with the English verb “to run,” which in the software context means “to execute.” We see that Google has landed on “ejecutar,” which we might recognize is cognate with “execute.” But Toucan gives us “corrida,” which is both the wrong word sense as well as the wrong part of speech:
Other errors in this example are left as an exercise for the reader. There are a couple.
If I were a Toucan engineer, I might think about this problem like a machine learning task, even if the current implementation does not use ML. I would find or formulate an evaluation metric for this particular flavor of word sense disambiguation. Jurafsky and Martin have a chapter on this which is a great starting point. Then I would find or construct a labelled training set of texts, and measure the performance of the current implementation against that training set. At that point we would have a baseline sense of errors, and could start to explore algorithmic improvements.
Escaping Duolingoland
These issues notwithstanding, what Toucan and Fluent do feels new to me, in ways that inspire both skepticism and excitement. The somewhat dubious textual artifacts they produce resist easy naming. They are a kind of artificial Spanglish. Or, more academically, a form of algorithmic code switching. What resemblance they bear to human code switching - as if that in itself were easy to pin down - seems worth investigating. I recently posed this question in a Toucan-sponsored Clubhouse session called “El rol de Spanglish." The room, led by a professional Spanish language educator, seemed to agree that these texts are not Spanglish, but something else: "a way to practice vocabulary.” Other open questions are whether similar hybrid texts have been used for effective vocabulary practice before, and whether any research has tried to evaluate their effectiveness against other methods.
But even if there is a precedent for textual code switching of this sort, perhaps lodged in a textbook series somewhere, these browser-based implementations just hit different. There is a fundamental pedagogical shift happening here, made possible by the combination of the internet, the extensibility of the browser, and natural language processing. What Toucan and Fluent do is not possible with traditional classroom instruction, where teachers and curricula must generally direct the attention of the entire class to the same content. It’s also a fundamental departure from language learning apps like Duolingo, Babbel, Speakly, Memrise, Chatterbug, Drops, Falou, Busuu, Pimsleur, Lingvist, and Rosetta Stone.
What all of these methods share is a basis in canned context. They create content and situations ahead of time, then lay them out before the learner, usually in a strict linear fashion, according to a pedagogical strategy based on successive mastery. They dictate to the user what should be relevant for them, and when it should be relevant.
Anyone who has stared down a long gauntlet of Duolingo skills knows what I’m talking about - you scroll endlessly through grayed-out skill modules that you can’t yet access, forced to imagine that one day, you might make it all the way to the 8th cartoon castle icon and finally master the secret skills that At Home 6 and Paris 2 and Past Tense 6 hold for you. Anyone who has ever been forced to act out the mortifying little skits and role plays or to follow the sagas of recurring fictional characters featured in rapidly aging textbooks commonly used in classroom settings knows well the depths of cringe that come with canned attempts to create life-like situations as context for language learning.
My shorthand for these brittle, disenchanting pedagogical universes is Duolingoland. Many of us share the common belief that Duolingoland is a good place to learn and maintain languages. I have certainly spent my fair share of hours in Duolingoland. And it's definitely the case that it is - for some learners, some languages, and some stages of learning - a helpful and enabling environment. After all, some amount of structure and linear progress in a language learning journey is probably necessary, whether it is imposed by the learner or by the teaching system.
Not all corners of Duolingoland are equally divorced from real life. What little I have seen of Memrise's content has been vivid, differentiated, and real-seeming; it’s the best I’ve encountered so far among the litany of options in the App store. Same goes for Duolingo’s podcasts, which are engaging, often uplifting stories of real, interesting people, narrated in a mix of English and the target language. And of course, a human teacher can ground language learning experiences in reality, although doing so for many students synchronously, day in and day out, is an enormous task.
But with Toucan and Fluent, we can begin to imagine an escape route from Duolingoland. We see that there may be another world in which we could be learning a foreign language: our daily lives on the internet.
These apps are context-first, in that they take as a starting point the learner’s real-life interactions with content where it lies in the world. They represent an innovative break from traditional approaches to teaching vocabulary. They don’t subject us to an endless gauntlet of canned context. They don’t assume a user should be motivated to learn a comprehensive list of French words for fruit one week, and then a comprehensive list of French words for the parts of the body the next, then a comprehensive list of French words for sporting activities the next, etc. They don’t assume the user can’t yet handle being exposed to words for concepts like confidence, science, and washing machine.
Instead, they turn Being Online into serendipitous moments of pedagogical opportunity. They harness content that is relevant to the learner, in the moment at which it is relevant. This is a powerful and elegant idea, made possible by this killer combo of the internet, the extensible web browser, and natural language processing. What’s most exciting is that these apps are just scratching the surface of what is possible with context-first language learning. I will be writing and thinking about more ways of escaping Duolingoland in the weeks and months to come.
Words for Sale
When I first started using Toucan, its pop-ups Called Me to Act on ads to some seriously niche content. First was a travel blog called Yi Fly, dedicated to "Travel ✈️ Hotel🏨 Miles🗺 Flight records💺 overseas life📝.” The site is in Mandarin, which, unlike some of my badass friends, I’m disappointed to report I have never studied and don’t yet know how to read. Another led to the LinkedIn page of a Brazilian software engineer - a seemingly nice enough fellow with Ruby on Rails experience. Why would the app have served me these links? Perhaps a majority of Toucan’s 40k plus users are Mandarin or Portuguese speakers, I thought. Perhaps they’re into traveling throughout Asia and monolithic web apps (looking at you, digital nomads).
Being an eager early adopter, I offered Toucan my puzzled reactions via bug reports. A week later, and the nimble young startup had rolled out a fresh approach. Brands, organizations, and individuals could now “own” words. Browsing the site of a German data annotation startup, I learned that the Girl Scouts had laid claim to “confidence.” Shortly thereafter, Taylor Nieman, the company’s CEO and one of its three co-founders, whom I had bugged with pedantic questions about the app’s pedagogical efficacy, gifted me "a word to own" within Toucan. “We’ve learned that giving people a piece of the dictionary helps make learning new vocabulary even more fun,” she wrote.
To understand what being the proud new owner of an expansive concept like “confidence” might feel like, I thought I should at least try it out. Toucan word owners are not owners in the strict sense of the term; rather, they rent a word’s pop-up’s Call To Action banner for a subscription fee, currently priced at .99 USD per week. I set my sights high - this was clearly a word renter’s market. For the price of a cappuccino, pending approval from the Toucan team, I could own “learning” for an entire month. That meant I could drive traffic from “learning” across all the app’s languages to a URL of my choosing, and upload an image to help encourage users to convert.
What image would work best for me? Should I use my face? Unlike my namesake, I am a nobody. I needed something fun and fresh to represent “learning." To my dismay, in the minutes I spent faffing around with these questions, looking for a good galaxy brain picture, an Ed Tech VC firm called Reach Capital had laid its claim to “learning." That, or they had owned it all along, and a buggy availability search in Toucan’s checkout flow had filled me with false optimism.
Word-level marketing gets right to the point. In some cases, it distills and condenses a brand’s more oblique messaging, like the effects being a Girl Scout may purportedly have on one’s sense of self efficacy (confianza: confidence: Girl Scouts). In other cases, its use is rather on the nose (peliculas: movies: Netflix). Toucan’s current system lends a strong first-mover advantage to early adopters. We might say there is an "early bird gets the word" principle at play that - for now, at least - puts seemingly random, normal people on the internet on an equal footing with streaming platforms worth > 200 billion in market cap.
I realized that the link to a Brazilian programmer’s LinkedIn profile I had been served a week ago belongs to one such early bird (programación: programming: Cesar Alves). Before, due perhaps to a bug or a release quirk, the ad had simply lacked the framing and copy of word ownership. Case closed - Cesar was just the proud owner of programming.
I’m at the Pizza Hut. I’m at the Memory Palace. I’m at the combination Pizza Hut / Memory Palace.
After using it for a few weeks, and despite its flaws, I’m convinced Toucan is useful for acquiring and maintaining lexical chunks of at least some languages. There’s a lot to like in the company's cute branding, fun vibes, and clear business ingenuity. In lean startup fashion, it has shipped a usable chassis for iterating towards a more trustworthy, more accurate product. Reaching out to early users like myself across different social platforms and channels, they are doing proper Customer Development. They engage users on Slack, Twitter, email, and Clubhouse, where Toucan-sponsored language teachers have been leading free, fun Spanish and French classes several nights a week. They’re doing all the things a young venture-backed startup should.
That said, I hope we can all agree that the idea of word ownership is absurd. Concepts and ideas can’t - and shouldn’t - be owned. No one has exclusive use over the notion of confidence, or learning, or washing machine. The shift I observed, from niche ads delivered to the user without a reason, to niche ads delivered because something or someone "owns the word," illustrates that this framing is a thinly veiled pretense for monetizing users’ attention. Word ownership is not fun; it’s downright dystopian.
As an educational entity, Toucan has made a Faustian bargain by serving ads. While ads let them grow and improve a free product, they also subvert the educational pretenses and fun vibe of the tool. And they put Toucan in an awkward and somewhat compromised position as a provider of information as generic and accessible as word definitions and translations. Imagine if Wikipedia sold exclusive, early-bird-gets-the-word style ad real estate. Imagine having to see Bill Nye, or Pfizer, or some totally random guy with a boring name like "Bill Roberts" above the entry for Science.
I also wonder whether word ownership marks a new kind of incursion by the market, one that hijacks a key pedagogical moment. Despite Toucan’s website copy suggesting otherwise, learning and recalling lexical items in a foreign language does take some effort and attention on the part of the user. The use of a word’s pop-up as a marketing channel is based on this very fact: the user will hover over a word’s pop-up and actively attend to it, attempting to make or re-enforce a connection across the languages in their memory.
Toucan enables the market to insert itself into this granular, lexeme-level moment of cognitive effort. Whether we use the method of loci or not, it seems an apt metaphor for this attempt to create a billboard at the site of every word we learn in a new language; a colonization of our personal palaces of the mind; a combination Pizza Hut / Memory Palace.