Spaced Repetition for Chinese: One Card Per Word Is Not Enough

Spaced repetition has a reputation problem among Chinese learners. Half of them credit it with everything they know; the other half abandoned it after three months buried under a review queue they could not clear. Both experiences come from the same software. The difference is almost always in how the cards were built and how the new-card tap was managed.

What the scheduler is actually doing

Two ideas do all the work.

The first is retrieval practice. Recognising a word when you see the answer feels like studying but changes very little. The memory strengthens when you try to produce the answer before seeing it, including when you fail. This is why a flashcard is not a list: the gap between the prompt and the reveal is the entire mechanism.

The second is expanding intervals. A memory decays over time, quickly at first and then more slowly. If you review too soon, the retrieval is easy and teaches your brain little. If you review too late, you have forgotten and must relearn from scratch. The scheduler is trying to catch you just before the memory would have failed — the point where retrieval is effortful but still succeeds.

Everything else in the software is bookkeeping. When you grade a card as easy, the next interval gets a larger multiplier; when you fail it, the interval collapses and the card re-enters short-term rotation. The algorithm has no idea what Chinese is. It only knows whether you pressed "good" or "again". This matters more than any algorithm choice: a well-designed card scheduled by a mediocre algorithm beats a badly designed card scheduled perfectly.

Why one card per word fails in Chinese

In a language written alphabetically, seeing a word largely gives you its pronunciation, and hearing it largely gives you its spelling. One card can plausibly cover both directions. Chinese decouples these skills far more aggressively, and a single card silently trains only one of them.

Take 银行 (yínháng, bank). There are at least five separable things you might mean by "I know this word":

1. Character to meaning. You see 银行 and think "bank". This is the layer most decks train, and the one that transfers least.
2. Character to pronunciation. You see 银行 and can say yínháng — not yínxíng, which is the trap, because 行 is read xíng in most other contexts.
3. Sound to meaning. Someone says yínháng at conversational speed and you understand without seeing anything. If you have never trained this, you will not have it, and it is the only version that helps you in a listening test or a real conversation.
4. Meaning to production. You need to say "bank" and 银行 arrives fast enough to be useful mid-sentence.
5. Handwriting. You can write it from memory. This is expensive and worth it only if you actually need to handwrite.

Layer 1 does not produce the others. Learners regularly discover that they can read a page comfortably and understand almost nothing when the same page is read aloud to them.

A workable minimum for each word is three cards, in this order of priority:

• Audio to meaning. Front: the recording, ideally inside a short sentence. Back: the meaning and the characters. This is the card most decks omit and the one that fixes listening.
• Characters to meaning and pronunciation together. Front: 银行. Back: yínháng, bank. Grade yourself wrong if the reading was wrong, even if the meaning was right — otherwise you are training a silent vocabulary.
• Meaning to production. Front: a short English, French or Arabic prompt plus a context cue. Back: the Chinese with pinyin. Say the answer aloud before flipping. If you only think it, you are not training production.

Add a handwriting card only for characters you genuinely need to write.

Card design rules that matter

One fact per card. A card asking for the characters, the tone, the meaning and an example sentence at once will be failed for any one of them, which corrupts the scheduling for all four.

Put words in sentences. Isolated words are ambiguous and forgettable. 还 on its own is unanswerable: it is hái (still, yet) and huán (to return something). In 我还没吃饭 (Wǒ hái méi chī fàn, I have not eaten yet) it is unambiguous, and the sentence gives the memory something to hang on. The same applies to 长 (cháng, long / zhǎng in 长大 zhǎngdà, to grow up), 乐 (lè in 快乐 kuàilè, happy / yuè in 音乐 yīnyuè, music) and 教 (jiāo, to teach / jiào in 教室 jiàoshì, classroom).

Make the prompt answerable. A production card fronted with only "book" is unfair — 书 or 本 or something else? Front it with a full cue: "the book is interesting" leading to 这本书很有意思 (Zhè běn shū hěn yǒuyìsi).

Include the tone in the grading standard. Decide once and for all that a card with the right meaning and the wrong tone is a failed card. Learners who let tone errors slide end up with a large vocabulary they cannot be understood in.

Do not put explanations on cards. Grammar is understood, not memorised by flash. Read the explanation in a lesson; use cards for the example sentences that instantiate it.

Preventing the backlog

The backlog is the reason most people quit, and it is almost entirely predictable. Every new card you add today generates a chain of future reviews — tomorrow, in three days, in a week, in three weeks. Adding cards feels free on the day you add them and is paid for over the following month. A learner who adds thirty new items a day for two weeks because they are enthusiastic will meet an unmanageable daily queue exactly when the enthusiasm fades.

Five rules keep it under control.

Reviews before new cards, always. If you only have fifteen minutes, spend all fifteen on reviews and add nothing. Reviews are debts already owed; new cards are debts you are choosing to take on.

Set the new-card limit for your worst day, not your best. Pick a number you could sustain on a day with an exam, a headache and a late train. If that number is five, it is five. A steady five beats an erratic thirty.

Freeze intake when the queue grows. Decide a personal ceiling — say, twice your normal daily review count — and when you cross it, new cards go to zero until the queue is back to normal. No exceptions, because the queue is a leading indicator of quitting.

Kill your leeches. A leech is a card you have failed many times. It is not a memory problem; it is a card problem. Suspend it, then meet the word again in reading or listening, and rebuild it from a sentence you have actually encountered. 差不多 (chàbuduō, more or less) and 舒服 (shūfu, comfortable) are classic leeches as bare glosses and become easy inside a remembered sentence.

Plan around known interruptions. Two weeks before travel or an exam period, cut new cards to zero and let the queue drain. You will come back to a manageable deck instead of a wall.

If you are already buried, you have two honest options: suspend everything older than a chosen date and restart the intake discipline, or declare bankruptcy, delete the deck and rebuild it from words you meet in real reading. Both are better than opening the app, seeing four hundred due cards, and closing it again — which is the third option, and the one everyone actually takes.

What spaced repetition will not do for you

It will not teach you grammar, it will not build fluency, and it will not turn recognition into speech on its own. It is a maintenance system: it keeps material you have already understood from decaying while you go and use it. The learners who get the most from it are the ones who add words after meeting them in a story, a dialogue or a conversation — so every card is a reminder of something, not a stranger with a translation attached.