Sentence mining: learn sentences, not words

A word on its own is a fact. A word inside a sentence is a fact with instructions for use: which particle follows it, which measure word it takes, whether it is something you say to your boss or only to your cat. Sentence mining is the habit of collecting those sentences from things you actually read and watch, and reviewing them so the word and its instructions stay together.

It is also easy to overdo. The method works as a supplement to reading and listening, and fails when it tries to replace them.

What sentence mining is

You mine a sentence when you meet it in real content, understand everything in it except one word or pattern, look that one thing up, and save the sentence as a flashcard. The card then goes through spaced repetition like any other.

Three things separate mining from simply memorising phrases:

• The sentence comes from your own input. You have already met it in a scene or a chapter, so it arrives with a memory attached.
• It contains one unknown. Learners call this a one-target sentence. If two or three things are new, the card tests too much at once and you stop knowing why you failed it.
• You review for understanding. The task is to read or hear the sentence and know what it means. You are not required to recite it.

The idea reached most learners through Antimoon, which recommended putting sentences into an SRS, and then through Khatzumoto's blog All Japanese All The Time. A critical look at the first of these is in the Antimoon method review.

Is it better to learn sentences or words

Neither wins outright. They are good at different things.

	Sentence cards	Single-word cards
Context and usage	Built in	Missing unless you add it
Grammar exposure	Comes along for free	None
Time to make	Longer	Short
Time per review	Longer	A few seconds
Risk	Recognising the sentence instead of knowing the word	Knowing a translation you cannot use
Best for	Verbs, particles, set phrases, anything with tricky usage	Concrete nouns, characters, exam word lists

A sensible default: use word cards for the first thousand or so common words, where a picture or a short gloss is enough, and move towards sentence cards as the words you meet become more abstract. The Japanese word for "apple" does not need a sentence. The word for "to manage somehow" does.

Whichever you use, the card is only a reminder. The broader case for where words should come from is on the vocabulary page.

Learn words in the context of sentences, and sentences in the context of paragraphs.

How to pick sentences worth mining

Most sentences you meet should not be mined. Apply four tests.

1. One target. Exactly one unknown word or structure.
2. Clear once looked up. After checking the target, the whole sentence makes sense without further digging.
3. Short. Under about fifteen words, or one line of subtitles. Trim a long sentence to the clause that holds the target.
4. Worth knowing. The target is something you expect to meet again. A rare herb from a historical novel can be looked up and released back into the wild.

Two examples of good one-target sentences, assuming the bold word is the only new one:

Language	Sentence	Reading	Meaning	Target
Chinese	他終於答應了。(他终于答应了。)	Tā zhōngyú dāying le.	He finally agreed.	答應 (答应) dāying, to agree, to promise
Japanese	締め切りに間に合った。	Shimekiri ni maniatta.	I made the deadline.	間に合う maniau, to be in time

The Japanese card also teaches that 間に合う takes the particle に, which a word card would leave out. That is the whole argument for sentences in one line.

The best mine is your own life

Shows and books are good sources, but the best mine is your own life. Write down or record the phrases you hear in your everyday life: what the barista asks, what a colleague says when the meeting runs late, what your landlord texts. These sentences come with a place, a face and a reason to remember them, and you will need them again within days. Mine those first and you will make progress in no time.

How to do sentence mining with Anki

Anki is the usual tool because it lets you design the card yourself. A workable layout:

Front: the sentence in the target language, with the target word highlighted. No reading aids, no translation.

Back: the meaning of the target word, its reading (Pinyin, furigana or romanisation), audio of the sentence if you have it, and optionally a translation of the whole sentence.

Then the routine:

1. While reading or watching, mark candidate sentences without stopping for long. A screenshot or a highlight is enough.
2. After the session, go through the candidates and discard at least half. The ones that looked essential at 11 pm rarely all survive.
3. Make cards from the rest, with a short definition. A definition in the target language is ideal once you can read one; until then, English is fine.
4. Add audio where possible. For Chinese and Japanese in particular, a sentence you have only seen is half a sentence.
5. When reviewing, read the front, check that you understand the sentence and can recall the target's meaning and reading, then grade honestly. Do not grade on whether you could have recited it.

Keep new sentence cards to a number you can sustain, often five to ten a day. They take longer to review than word cards, and the pile grows quietly.

Tools that speed up mining

Making cards by hand is slow enough that many learners give up. A few tools shorten it:

• Pop-up dictionaries and subtitle tools. Migaku and similar browser extensions create a card with sentence, audio and screenshot from Netflix or YouTube in one click.
• Reading apps with export. Several e-readers and reading apps for Chinese and Japanese let you save a sentence with its lookup and export it later.
• Dictionary example sentences. A fallback when your own sentence is too long or messy: take a clean example for the same word from a good dictionary.

One caution. One-click mining removes the friction that used to limit how many cards you made. The limit now has to come from you.

Where sentence mining goes wrong

Treating the collection as the method. The best-known case is the "10,000 sentences" approach from All Japanese All The Time. Khatzumoto, who reported reaching fluency in Japanese in about 18 months, mined his sentences out of enormous amounts of listening and reading. Many imitators kept the sentences and skipped the listening and reading, with predictable results. He later declared the 10,000-sentence approach dead himself and moved to shorter cloze-style cards.

Memorising instead of understanding. If you can recite a card but cannot follow a new sentence with the same word, the card has taught you a recording. Shorten it, or replace it with a second sentence for the same word.

Recognising the card, not the word. After a few reviews you may know the answer from the shape of the sentence. A leech like this should be rewritten with a different sentence, not drilled harder.

Mining at the wrong level. In a beginner's first months, almost every sentence has three unknowns, and mining is frustrating. At an advanced level, most new words are rare, and a quick lookup while reading serves better than a card. Mining pays best in the long middle.

Letting it eat the input time. If card-making takes longer than the reading or watching that produced the cards, the proportions are upside down.

A first week of sentence mining

• Pick one source you would follow anyway: a drama, a manga, a podcast with a transcript.
• Mine at most five sentences per session, all one-target.
• Build cards in the layout above and review them daily.
• At the end of the week, delete any card that annoys you. A deck you dislike is a deck you will abandon.
• If the reviews feel useful after a month, raise the daily number slightly. If they feel like homework, lower it and read more instead.