This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3, Slow and Steady.
I'm still a beginner in Japanese, and I've always been amazed watching more advanced learners go through flashcards full of words they've never seen before and still guess the reading and meaning. To me, Japanese text basically looked like hieroglyphics.
Then, after a while of doing my own vocab cards, it started happening to me too. Every now and then I'd see a word for the first time, take a guess, flip the card, and be exactly right.
So I got curious and tried to actually measure how often that's possible, by breaking words into the pieces they're built from and checking whether those pieces give the word away. That led me down a rabbit hole of greedy algorithms and graph theory. The idea is that if you study the right building blocks first, a lot of new words become guessable without having to memorize them individually.
The result is GuessWithAnki, an Anki add-on that reorders your deck so you can guess the meaning, and often the pronunciation, of a lot of cards on your first try.
What I Built
If you've already learned the pieces that make up a word, you can often figure out what it means without ever having studied the word itself. For example, suppose you've already studied
- 電車 (でんしゃ), train
- 天気 (てんき), weather
You've now encountered 電 pronounced as でん and 気 pronounced as き. GuessWithAnki recognizes this and moves 電気 (でんき) (electricity) forward in your deck, since you already know how to read its kanji in this word and have some clues about its meaning. Instead of memorizing 電気 entirely from scratch, you have a good chance of guessing it correctly the first time you see it.
A card you get right the first time you see it is what I call a free card, since you never had to learn it the hard way. GuessWithAnki looks for the words you're likely to figure out on your own, then puts them right after the cards that teach their pieces.
This builds on GuessWithKanji, my first add-on, which only worked for Japanese. This time I wanted to see how far the idea goes beyond Japanese, so it now handles five kinds of decks.
| Deck | What counts as a piece | When a word is guessable |
|---|---|---|
| Japanese | A kanji paired with its reading in that word (電 read as でん) | Both the meaning and reading follow from its pieces |
| Chinese | A character paired with its pinyin syllable and tone (开 read as kāi) | Both the meaning and pinyin follow |
| Korean | A syllable with the Chinese character it comes from (학 from 學, study) | The meaning follows from its parts |
| English | A prefix, root, or suffix (un-, employ, -ed) | The meaning follows from its parts |
| Medical | A Greek or Latin root or affix (hepat, -itis) | The meaning follows from its parts |
The goal is to reward pattern recognition over memorizing.
GuessWithAnki encourages this by minimizing the amount of memorization needed to grasp an underlying pattern in a vocabulary deck.
This has the added effect of improving your memory, and giving you access to skills that are usually reserved for more advanced language learners.
Demo
The video at the top walks through the different performance profiles of decks across languages. Below you'll see the potential gains in Chinese vocabulary decks, which are the best suited for this technique.
With GuessWithAnki you are able to learn the equivalent of 500 Chinese words with less than 200 words. Just by changing the order you study them in.
Code
- BlaiseLabs/GuessWithAnki has the add-on, measurement tools, experimental results, and video source files, including Manim scenes, narration, and build scripts.
- BlaiseLabs/GuessWithKanji is the original Japanese-only version.
How I Built It
1. How guessable is a word?
I started by looking at how much overlap in kanji there was across words. The more unique kanji there were, the more difficult it would be to guess a word's meaning.
Luckily for me that's not how Japanese works, and as it turns out there was a large overlap between the words and the kanji they shared.
The reading
This part was surprisingly straightforward.
Two free dictionaries, KANJIDIC2 and JMdict, provide kanji readings and tens of thousands of vocabulary entries.
I wrote a small Python function to match each word's reading to its individual kanji. Compared against JmdictFurigana, it agreed on 99% of 17,854 common words.
Once you have those matches, you can count how frequently each reading occurs.
For example, 気 is read as き in 130 dictionary words, but as け in only 14.
So if you've learned 気 as き, a new word using that same reading becomes much easier to guess.
The meaning
Figuring out whether a word's meaning follows from its pieces is trickier.
I tried two approaches, both using open models.
- Embeddings. EmbeddingGemma represents each meaning as a point in a vector space. If a word's meaning is close to the meanings of its kanji, there's a good chance you can infer it. After tuning on labelled words, this approach achieved an AUC of 0.83, compared with 0.73 for simple word overlap.
- Asking a model directly. Gemma 4 runs locally through Ollama and labels words in batches. It judges whether the reading and meaning are predictable, partly predictable, or not predictable, using a JSON schema that specifies the exact words in each batch.
My own deck
I ran the analysis on my own deck of 2,038 words containing kanji.
Here's what I found.
For 67% of the words, both the reading and meaning were predictable.
In other words, roughly two out of every three words I was about to study were, at least in theory, words I could have guessed on first sight.
That was a much bigger share than I'd expected.
2. Finding the right order with a greedy algorithm
There's a catch, though. You can only guess a word if you've already encountered the pieces it's made of. For example, you can't guess 天気 (てんき) from its kanji if you've never learned 天 read as てん.
So this basically becomes a max/min problem. How do I maximize the number of guessable Anki cards while minimizing the amount of memorization needed?
Trying every possible order is out of the question. For 2,000 cards there are about 3.3 × 105735 possible orders.
This basically turns into something a lot like the traveling salesman problem, and one of the simplest ways to approach this is using a greedy algorithm. A greedy algorithm builds a solution one step at a time. At each step, it chooses whatever looks most promising right now, without going back to reconsider earlier decisions.
It doesn't guarantee the perfect solution, but it's fast, and can get pretty close.
At every step, the algorithm asks which card should come next.
- If a word is already guessable, it gets priority, especially if you've just encountered the pieces that make it up.
- If nothing is guessable yet, the algorithm looks for an opener, a card that introduces just one new piece, preferably one that will help unlock other words.
For example, you might study 取る (toru, to take), followed by 戻る (modoru, to return).
Once you've learned those pieces, 取り戻す (torimodosu, to take back) becomes guessable.
The algorithm tries to create as many of these little chains as possible.
On my own deck, this changed the number of guessable words among the next 500 cards from 66 to 307.
That's nearly five times as many words you might be able to figure out without first memorizing them.
3. Does it actually help?
Of course, a word being theoretically guessable doesn't mean you'll actually guess it correctly.
And even if you do, will it help you remember more over time?
To explore that, I built a simulation of someone studying with Anki for 100 days.
The simulated learner studies 20 new cards per day and completes every scheduled review. Reviews follow FSRS-6, Anki's scheduling algorithm, which I reimplemented and checked against py-fsrs.
Each piece of a word has its own memory state. That memory gradually fades, but gets reinforced whenever another word containing the same piece appears.
A word can only be guessed if the learner still remembers its pieces.
On a JLPT vocabulary deck, with perfect guessing, a learner would need to learn just 16 cards from scratch to know their first 100 words. The other 84 could be guessed.
100% guessing isn't realistic though. This isn't magic, and as learners we're still going to get some words wrong.
I checked my own Anki review history and found that when a word was guessable and I'd seen its kanji recently, I answered correctly about 73% of the time.
So I also tested a more realistic 80% guessing rate.
At that rate, reaching the first 100 words required learning just 35 cards from scratch, rather than 100.
And there was another benefit I hadn't expected. Less forgetting.
Every new word reinforces pieces you've already learned. So you're not just accumulating vocabulary. You're repeatedly practicing the patterns behind it.
In the Core 2k simulation, forgotten reviews by day 30 fell from 189 to 72.
The same ordering that made words easier to guess also helped keep their building blocks fresh in memory.
4. Does it work beyond Japanese?
Japanese turned out to be a great fit for this idea. But I wanted to know whether it would work for other subjects too.
Then I ran the same experiment on two decks each for Chinese learners, English learners, and medical students, and three decks for Korean learners.
The underlying algorithm stays the same. The only thing that changes is what counts as a reusable piece.
For speed, I used a cloud model to label the 21,624 words across these decks in parallel, processing about 2.6 million tokens. The add-on itself uses Gemma locally, processing one batch at a time.
Here are the results at an 80% guessing rate.
| Deck | Guessable | Cards to reach 500 words | Speed-up |
|---|---|---|---|
| New HSK 3.0, levels 1–4 (Chinese) | 71% | 392 → 200 | 2.0× |
| HSK 2012, levels 1–5 (Chinese) | 66% | 473 → 256 | 1.8× |
| Core 2k (Japanese) | 60% | 481 → 265 | 1.8× |
| Kaishi 1.5k (Japanese) | 64% | 474 → 282 | 1.7× |
| Oxford 3000 (English) | 27% | 494 → 303 | 1.6× |
| Korean Vocabulary by Evita (Korean) | 39% | 492 → 310 | 1.6× |
| 2k Essential Beginner Vocab (Korean) | 31% | 486 → 350 | 1.4× |
| 4000 Essential English Words | 20% | 494 → 356 | 1.4× |
| Medical Terminology with Etymology | 27% | 482 → 474 | 1.0× |
| TTMIK My First 500 Words (Korean) | 25% | 495 → 492 | 1.0× |
Chinese did even better than Japanese
Chinese turned out to be an even better fit than I'd expected.
A character usually has one reading, so the pinyin of about 96% of words can be predicted from their characters.
And 90% of HSK words share every character with at least one other word.
That gives the algorithm plenty of reusable pieces to work with.
The advantage also lasts. Even beyond a thousand words, the optimized order maintains a 1.6–1.8× lead.
Korean lands in the middle
Korean was a late addition, after someone asked for it.
Most Korean vocabulary is Sino-Korean, built from syllables that each come from a Chinese character (Hanja). So once you know 학생 (student) and 교실 (classroom), 학교 (school) gives itself away.
The catch is that native Korean words and loanwords don't split into pieces like that, so only about a third of a beginner deck is guessable.
Evita's deck still reached 1.6× at 500 words, and kept 1.5× past a thousand. The small TTMIK deck barely moved, since its pieces rarely come back.
English is more modest
Most everyday English words don't break neatly into meaningful pieces. Knowing the spelling of beach, for example, doesn't tell you what it means.
But word families offer a useful shortcut.
Once you know employ, words like employer, employment, and unemployed become much easier to figure out.
The gains aren't as dramatic as in Chinese or Japanese, but they're still meaningful.
Medicine was the surprise
I initially thought medical terminology would be the perfect use case.
So many medical terms are built from Greek and Latin roots. So I figured they would especially benefit from greedy based sorting.
But the medical terminology deck barely improved on its own.
The problem was the deck itself. It was relatively small and mixed together technical terms, abbreviations, drug names, and everyday vocabulary. Many roots simply didn't appear often enough to be reused.
Medical students are often advised to learn the roots first. But with a roots deck placed ahead of the terminology deck, the learner has to get through 982 root cards before seeing the first medical term, around day 50.
GuessWithAnki takes a different approach.
Instead of teaching every root up front, it introduces only the roots needed for the next terms, right before those terms appear.
That means the learner can start guessing actual medical terminology within the first two weeks.
So even when the raw speed-up looks small, changing the order can make the learning experience very different.
Across all five profiles, the pattern became clear. What matters is how often meaningful pieces get reused, and how reliably they predict the words built from them.
5. Making the video with open models
I also wanted to explain the project visually, so I made a 3Blue1Brown-style video using Manim.
It features a little mascot named Gaku (学, meaning "study"), who reacts to the explanations as I talk.
The narration uses a clone of my own voice, and the whole production pipeline runs locally on a Mac mini.
Here's how I put it together.
- The voice. I used Qwen3-TTS (1.7B, open weights) through MLX, cloning my voice from just 5.7 seconds of recorded speech. I generate whole paragraphs at once to keep the delivery natural. Whisper then detects the pauses between sentences so I can cut the audio and align each animation with its narration.
- Checking the narration. Every generated take is transcribed with Whisper and compared word for word against the script. If something comes out wrong, the pipeline automatically generates another take.
- Fixing Japanese pronunciation. The cloned voice kept pronouncing とる (toru) as "toh-ruh", and no amount of creative respelling helped. In the end, I recorded that paragraph and the intro myself, then used CosyVoice2 for voice conversion. That preserved my actual pronunciation and timing, including the laughs and stumbles, while giving it the cloned voice's sound.
- Catching layout bugs. Animation overlaps are surprisingly hard to spot by eye. So I wrote a checker that runs after every animation and flags overlapping text, text escaping its box, anything covering Gaku, and objects falling outside the frame. It caught 30 problems, plus one particularly funny bug where Gaku's eyebrows collapsed into a single dot in the middle of the screen.
Why Does Open Innovation Matter?
None of this would have been possible without open-source work.
The language data comes from open resources, KANJIDIC2 and JMdict from EDRDG, plus CC-CEDICT for Chinese.
The scheduling algorithm is open too. It's FSRS, developed by the open-spaced-repetition project.
And the AI models run on my own machines. GuessWithAnki uses Gemma 4 and EmbeddingGemma, while the video pipeline uses Qwen3-TTS, CosyVoice2, and Whisper.
Put together, these tools let one language learner go from wondering, "How do advanced learners do that?" to actually measuring the phenomenon and building something useful from it.
Just as importantly, other people can check the results and build on the work.
The add-on doesn't need a server. Even its built-in heuristics, without a language model, judged 68% of a 4,995-word HSK deck to be guessable, close to the model's estimate.
Guessability labels are just data. You can share them, bundle them with a deck, or improve them together as a community.
That means the idea doesn't have to stop with my experiments.
The Lazy Payoff
With an 80% guessing rate, the simulation suggests you could reach your first 500 Chinese words while learning only about half as many cards from scratch. Japanese comes close.
You get more words from less memorization, reinforce the pieces you already know, and spend more of your time recognizing patterns instead of treating every word as something completely new.
And that's what brings me back to those advanced Japanese learners who seemed to know words they'd never studied.
Now I understand they were able to do this by learning enough pieces to start seeing the patterns in kanji across words.
GuessWithAnki is my attempt to help the rest of us get there sooner.
Prize Categories
Best Use of Gemma
Gemma 4 labels word guessability locally through Ollama, processing batches of 20 words with a JSON schema that specifies the exact batch contents and a daily processing cap.
EmbeddingGemma provides a second way to estimate guessability by measuring the semantic similarity between a word's meaning and the meanings of its pieces.
On 1,839 words from my own deck, Gemma 4 12B agreed with a much larger cloud model on 86% of reading labels and 79% of meaning labels.
Sentry Agent Tracing
Every labelling job is recorded in a self-hosted Sentry instance using Sentry's AI agent tracing format.
Each batch gets its own transaction, recording token usage, timing, and unreadable words. Individual embedding calls are tracked as spans.
This wasn't just instrumentation for its own sake. The traces helped me catch several real bugs.
-
Empty responses with thinking enabled. Gemma sometimes returned nothing at all. Setting
reasoning_efforttononefixed it. - Context overflow. Batches of 40 words exceeded Ollama's default 4,096-token context and crashed the parser. Reducing the batch size to 20 fixed the problem.
- Incomplete but valid JSON. Gemma 4 E4B once returned a perfectly valid JSON response containing just 1 of the 20 requested words. I updated the schema to enforce the exact words and their count.
- Missing traces on long jobs. Some long-running jobs disappeared from the tracer. Splitting the work into separate transactions per batch solved that.
After those fixes, Gemma 12B successfully labelled all 1,839 words in 2.3 hours, with no failures.
Top comments (0)