The science behind langkit

Langkit is a tool I built for myself when I first started learning Japanese. I wanted an app that I could use every day and get close to the optimal learning for that day. It builds on findings and ideas from the Anki community and the latest research in psychology and neurology.

You don't know how to learn

This 2008 study compared two study strategies: "massed practice" vs "spaced practice". 64 out of 72 participants judged massed practice to be the more effective strategy. Yet, when they were tested on the material they learned using each strategy, these were the results:

• massed practice: 35% correct
• spaced practice: 61% correct

This result is replicable, learners consistently judge the least effective strategies as the most effective.

Since we can't trust our intuition or the experience of previous learners we have to be extremely science and data based.

How do we learn?

Learning means rewiring your brain. The first day you see a new word, your brain will have trouble representing it. The knowledge will be encoded ineffectively over many neurons and it will take a lot of effort to retrieve it.

As you use the word more and more, your brain will grow new neurons or wire them together more effectively, essentially hardcoding the knowledge in your brain so that This principle applies to words, grammar, sounds, writing systems, and every other part of language.

Growing new neurons is expensive and your body doesn't like expensive things. As a rule, it will always take time and effort to improve at any skill. Any solution that is either "easy" or "fast" (Duolingo, comprehensible input, etc.) can never work.

Repetition, especially spaced repetition, is the best mechanism we have for strengthening connections in our brain. Good language learning apps like Anki use this as their primary driver, but there is a critical flaw!

Varied practice

A vocabulary flashcard looks like this: "出す = to take out / to submit / to produce". If you learn this card, you'll become really good at translating from "出す" to "to take out". But when you see it in real life, it will be in a sentence like "声を出した。". Even if your brain can decode 出す extremely quickly, it has never seen 出した before. Even worse, the natural translation for "声を出した。" is something like "I cried out" or "I made a sound". The literal translation "I let out my voice" just wouldn't be used in English. 声を出す is an extremely common collocation — a Japanese speaker can look at "声を_す" and will automatically be able to fill in the blank because their brain is hardwired to recognize it.

This is where two new principles of learning come into play: varied practice and interleaving.

There is a classic study that demonstrates this effect. In it, children who practiced throwing beanbags from two different distances (2 feet and 4 feet) outperformed children who only practiced from the test distance (3 feet) — even though the varied group had never once practiced from 3 feet.

Learning a language is not the same as throwing beanbags, and pop psychology often oversimplifies. But the same effect has been observed in relation to learning.

Varied practice leads to knowledge that is more flexible, more robust, and on top of all that, it speeds up learning. Anki completely misses this. The "fix" recommended by the online learning community is to consume content or to "sentence mine", but this is extremely inefficient and overwhelming for beginners.

Langkit instead will not quiz you on single words, but on sentences. A sentence usually contains a hard word and a bunch of easy words. Langkit stores a representation of how well you remember each word / conjugation / form / spelling in its database and uses that knowledge to generate or select completely new sentences that challenge you at exactly the right level. Under the hood it uses FSRS, the same system as Anki, but the knowledge it stores is a lot more granular.

I wrote a whole dedicated article about the most effective way to learn Kanji.

Langkit is accurate

If you've used a pop-up dictionary like Yomitan or sites like jpdb.io, you might have noticed that they often make mistakes when translating sentences. This is because they use tokenizers (engines that aim to split up sentences into words) like "MeCab" or "Kuromoji". Langkit uses a combination of Ichiran and AI (Claude Opus 4.6) to break down sentences, which together create extremely high-quality results.

A note on Comprehensible input

Comprehensible input or "Immersion learning" is largely pseudoscience. At least the majority of claims you can find on YouTube or Reddit about it are wrong. The original paper by Stephen Krashen is not a high-quality scientific work. Here is an actual quote from his 1989 paper:

One interesting case history was provided by Goodman and Goodman, who reported that their daughter Kay, at six-and-a-half years of age, could independently read materials written at the fifth-grade level, even though she had had no formal instruction in reading

Besides the anecdotal evidence, the quantitative studies he cites often have no control group, and when they do, then the input group usually doesn't outperform the control group. I don't dislike input and I always recommend learners to consume content in their target language, but it is not time effective at all and most importantly, input practice does not lead to output! Until I can prepare a full paper debunking this, I recommend the paper "Input is not a panacea".