Context is King: Inductive Language Learning with Anki

A lot of people use spaced repetition (i.e. smart flash cards) to study vocabulary when learning a language.

This makes sense, because you need to learn approximately one metric ton of words to reach intermediate competency in a language, and flash cards are a great strategy for learning a large number of things. Anki is the most popular spaced repetition tool, and “use Anki for vocab!” is common language learning advice.

Some of my first-generation Anki cards for Spanish. Even arbitrary images make cards more memorable, and synonyms can be resolved by adding them to the front side of the card (as a sort of “taboo list”).

Spaced repetition is a fantastic tool. I maintain that it’s qualitatively different from any other study strategy, and cannot be approximated via note-taking, massed practice, or whatever else you do to learn. If you aren’t using Anki or another SRS tool, you probably should be (note: as a rule, popular apps like Quizlet and Brainscape are not spaced repetition).

But using spaced repetition really well is a “virtuoso skill,” as Michael Nielsen puts it in his legendary essay. Some card designs are better than others.

In this post I’ll explain how I use an inductive strategy based on sentence fragments to learn complex grammar. This strategy has proved to be a simple and effective approach that works unmodified across the four target languages that I have been working on for the past few years (Spanish, Korean, Ancient Greek, and Old English).

The Problem with Isolation

The problem with making flash cards for individual vocab words is that, well, this isn’t really how learning works.

• Isolated words can be really hard to remember if you don’t have lots of real-world experience with them.
• Knowing words isn’t enough to be good at using them in sentences, or understanding them when you hear them in a sentence — to say nothing of “nuance” across synonyms.
• Most grammar is very difficult to practice in isolation: prepositions, conjugations, and declensions are virtually impossible to learn well by practicing them by themselves.
• This goes for writing too: trying to learn an alphabet — or, Lord almighty, Chinese characters — one symbol at a time can be a very frustrating and error-prone experience.

For a while I tried to learn conjugations one-at-a-time via a strategy much like Andy “Functools’s” Ultimate Italian Conjugation deck, which is a tour de force in expressing conjugations as short, atomic translation prompts. He’s now also got an Ultimate Spanish Conjugation deck following the same principles.

In the past, I’ve used a similar strategy for languages with declension (like Ancient Greek), applying standardized prompts to cue for the accusative case, dative case, etc.:

A couple of my older designs for learning declensions in Greek. I’ve abandoned this in favor of short sentence fragments.

This felt effective at first, but as time went on (and spaced repetition intervals became large), some of these cards became very difficult. I obviously can’t learn Ancient Greek by immersion, and with my busy schedule I can’t spend hours a day reading Greek texts like scholars do. I have time for Anki — but things weren’t sticking as well as I’d like, nor were my reading skills advancing very quickly.

This is a general principle I’ve observed in my Anki experience: I can remember anything for a few weeks via rote memorization (even arbitrary digits of π!), but for long-term knowledge retention, card design matters.

For these reasons, as soon as I got the hang of using Anki for individual words, I started looking for ways to build up higher-order skills. Things like the ability to form sentences, or to recognize chunks of speech, or to get comfortable with different tenses and conjugations/declensions.

(Audio) Sentences to the Rescue

After a few years and several experiments, I’ve settled on a very flexible approach that I find effective for any language. It has two parts:

1. Audio cards: I find audio clips much more pleasant and natural to review, and more useful than image-based cards. They also force me to translate in working memory, rather than while reading out one word at a time; I believe this is beneficial for learning to chunk things in memory. Bonus: with iOS Voice Commands, I can review hands free while walking, doing dishes, etc.!
2. Sentences and sentence fragments: basically everything — from declensions and idioms to complex characters in written Chinese — is easier to learn with a little bit of context.

Where to get audio clips? I record my own voice. Sure, my pronunciation isn’t perfect, but it affords me maximum flexibility to decide what cards to create. I’ll use native-speaker clips if they are convenient.

UPDATE: I recorded my own voice because I didn’t like the quality of popular text-to-speech (TTS) engines. But with the advent of TTS systems that use deep learning, AI-based speech synthesis has finally come into its own and sounds natural enough to make high-quality Anki cards. The difference is night and day. I highly recommend the AwesomeTTS plugin for Anki—but be sure to set up a Google Cloud API key so that you can access the state-of-the-art neural network voices (which all have WaveNet in the name).

Many people combine sentence cards with movies to harvest audio clips from the wild. If you want to try this strategy, you might have a look at tools like Subs2SRS or (the newer) KnowClip.

Conjugations: A Spanish Example

For example, let’s say I’m a Spanish student, and I come across this sentence while reading Harry Potter y la Piedra Filosofal:

Eran las últimas personas que se esperaría encontrar relacionadas con algo extraño o misterioso, porque no estaban para tales tonterías.

I use two-sided Anki cards for my language clips. Production (native → target language) is far more important for skill-building than recognition (target → native), however. I often press “easy” on recognition cards, so I end up reviewing them less often.

I would never Ankify this sentence directly, because it is far too long to be used with spaced repetition.

What I will do is take smaller fragments for anything that I think I need practice with or help remembering. It doesn’t matter if I technically know all the words already: if something makes me uncomfortable, then I figure I need practice with it. As an advanced beginner, I might make fully 7 or so cards from this single sentence:

• Eran las últimas personas ← → They were the last people
• que se esperaría ← → that one would expect
• Eran las últimas personas que se esperaría ← → They were the last people that one would expect [This is an example of combining two chunks]
• encontrar relacionadas con algo ← → to find related to something
• algo extraño o misterioso ← → something strange or mysterious
• tales tonterías ← → such nonsense
• no estaban para tales tonterías ← → they were not for such nonsense

These cards give me some much-needed practice with the imperfect tense in Spanish. Since Spanish has two past tenses (imperfect and preterite) that often translate the same into English, sometimes I’ll insert “used to” into the English side of the card. This is only one possible translation of the imperfect tense, but it’s my personal standard marker to disambiguate it from the preterite:

• Eran las últimas personas ← → They [used to be] the last people
• no estaban para tales tonterías ← → they [didn’t used to be] for such nonsense

In practice, though, I don’t find disambiguating tenses (or synonyms, for that matter) in this way to be all that important. The point is to get practice with the language, and I’ll often remember the tense I’m supposed to use from context when answering a card.

Keep it Small

One core principle guides my audio fragments:

• The ideal length for a sentence fragment is 3 or 4 words.

Any shorter than this, and we run into the isolation problems described above. For example, two-word fragments are a tempting way to practice adjective-noun agreement (mujer hermosa ← → “pretty woman”). But these cards can be quite difficult once intervals become large.

Any longer, and the sentence becomes overly complex, difficult to fit in working memory, and hard (if not impossible) for the spaced repetition to schedule correctly. While people are often tempted to create long cards while sentence mining (Un momento más tarde aparecieron los postres ← → A moment later the desserts appeared), these can make for unpleasant and frustrating cards.

Complex cards might also encourage you to translate one word at a time in isolation, instead of processing the whole sentence at once like you would when listening in the real world.

The one exception I make is if I’ve already created short cards for two different clauses: sometimes I’ll make a third card that combines the two chunks into a longer sentence. I may sometimes not count articles or other helper words (like all the little nonsense words that fill Ancient Greek sentences). But in general I try to severely limit the number of cards I create with 5 or greater words.

Declensions: A Greek Example

Here’s a sentence from one of the practice readings in the venerable Athenaze:

ἐπεὶ δὲ τῷ αὺλίῳ προσχωροῦσιν ὁ τε Φίλιππος καὶ ὁ πάππος, πολὺν ψόφον ὰκούουσιν.

And here’s how I might break it up into spaced repetition cards:

• ἐπεὶ δὲ ← → but when
• τῷ αὺλίῳ προσχωροῦσιν ← → they approach the sheepfold
• ἐπεὶ δὲ προσχωροῦσιν ← → and when they approach
• ὁ τε Φίλιππος καὶ ὁ πάππος ← → (both) Phillip and his grandfather
• πολὺν ψόφον ὰκούουσιν ← → they hear a lot of noise

Notice that this example gives me practice with several declensions: τῷ αὺλίῳ is in the dative case, πολὺν ψόφον is in the accusative case, etc.

Notice also that I deleted “τῷ αὺλίῳ” from the longer phrase “ἐπεὶ δὲ τῷ αὺλίῳ προσχωροῦσιν.” I’ll often do this to create shorter sentences out of native sentences. Dropping subjects and adjectives is also a great way to shorten sentences: i.e. “ψόφον ὰκούουσιν” instead of “ὁ τε Φίλιππος καὶ ὁ πάππος πολὺν ψόφον ὰκούουσιν.”

With this strategy, I arguably don’t need to do anything special to practice declensions. Skill with case inflections falls naturally out of numerous example sentences. Regular patterns are fairly easy to master in this way — and I can hit irregular nouns more often by creating extra sentences that use them.

Bonus: using audio cards for Attic Greek gives you an easy way to learn the oft-ignored Attic pitch accent. This makes spelling with those pesky diacritics easier, and makes parts of the language easier to understand.

Prepositions: an Old English Example

In languages with case systems, prepositions take certain noun cases, and their meaning can change depending on the form of the noun that is used. This is the case (ha) with modern German, and Old English had a largely identical matching of prepositions to cases.

I pay special attention to examples from exercises in my textbooks that use prepositions in Old English. Here are some of the sentences and fragments that I’ve captured to help practice them:

• tō þǣm hāliġe stede ← → to the holy place
• mid him ← → with him
• fram ðǣm beorgum ← → from the mountains
• Hīe comon fram ðǣm beorgum ← → they came from the mountains
• samod hēr on Witanċeastre ← → also here in Winchester
• Iċ eom of Frisum ← → I am from Frisia

Old English is not a pro-drop language, so we can’t just drop subjects to create conjugation exercises. Instead, I’ll often replace a complex subject with a pronoun to simplify the sentence card.

For example, instead of

• Ðāra biscopa prēostas fērdon mid him ← → The bishops’ priests went with him

I might substitute

• Hīe fērdon mid him ← → They went with him

This allows me some flexibility, without devolving into me just making up sentences out of whole cloth (which is very dangerous, since as a beginner I’ll make all kinds of mistakes in inventing sentences!).

Particles: a Korean Example

The beauty of the inductive (sentence-based) strategy is that it generalizes to almost any language. Korean is an agglutinative language that is quite different from the Indo-European languages above (making it famously capable of complex, one-word sentences). But sentence fragments make short work of its various particles (some of which act much like case endings):

현대 백화점에서 만나요
“We meet at the Hyundai department store.”

The inductive strategy works for almost any language.

• 현대 백화점 ← → the Hyundai department store
• 백화점에서요 ← →At the department store.
• 백화점에서 만나요 ← → [We] meet at the department store.

Other particles can be learned in a similar fashion:

• 여덟시 ← →eight o’clock
• 여덟시에요 ← →at eight o’clock
• 학교에 있어요← →[I] am at school
• 학교가 있어요 ← →[I] have a school

Aside, levels of politeness and honorifics are tricky to translate into English. I largely get around this by inserting the word “sir” or “hey!” under my breath in the English prompt, to indicate a respectful or casual speech level, respectively.

If there is no modifier in the English prompt, I assume the standard “polite” level (요):

• 사랑해요← → I love you
• 사랑해 ← → [Hey!] I love you
• 맛있어요 ← →It’s delicious
• 맛있어 ← →[Hey!] It’s delicious
• 어디입니까? ← →[Sir,] where is it?
• 고향이 어디예요? ← →Where is your hometown?
• 연필입니다 ← → [Sir,] it’s a pencil.

I also use this system to differentiate between the several different words Korean uses for conjunctions like “and”:

• 남자하고 여자 ← →a man and a woman
• 남자와 여자 ← →[Sir,] a man and a woman

Why Context Helps

The spaced repetition community has various folk explanations for why good cards tend to be ones that are neither very isolated nor overly complex.

In general, whether we’re studying languages, mathematics, medicine, history, or anything else, the best flash cards tend to have at least two properties:

1. they are simple and fairly atomic (known as the minimum information principle), and
2. they are coherent, in the sense that the cards form a network of knowledge that is collectively easier to remember than the parts would be in isolation

A coherent vs. diffuse network of flash cards. Coherent knowledge is far quicker to access, easier to remember for long intervals, and easy to reason about. Borrowed from SuperMemo.guru.

Coherence is especially evident in well-designed cards for complex areas, such as geography, or algorithms —where we need to build mental landmarks to understand what’s significant, what’s not, and how each piece relates to that larger picture.

Something like this also seems to be at work in language, though perhaps not anything as obvious as a network of theorems.