Comprehensible input, and how to tell whether something is at your level

The idea, minus the ideology

Comprehensible input is the observation that you acquire language from material you can almost understand. Not material you understand completely, which teaches nothing new, and not material you cannot follow, which teaches nothing at all because you never work out what anything means.

The theory around it gets argued about at length. The practical core does not really need defending: you learn a word by meeting it somewhere you could work out what it meant. That requires the rest of the sentence to be within reach.

The window, in numbers

Reading research puts the useful band somewhere around 95 to 98 per cent known words for independent reading, and about 90 per cent as the floor below which comprehension collapses. For listening, where you cannot stop and re-read, the numbers sit a little higher.

Translated into something usable: at 98 per cent you meet roughly one unknown word every two lines, which is a pleasant read that teaches you a handful of words an hour. At 90 per cent it is one word in ten — a sentence with two or three holes in it — and comprehension is doing more guessing than it can support.

Note what this does not say. It does not say hard material is useless. It says hard material is expensive, and you should know you are paying.

Why difficulty as it feels is a bad guide

Two things distort the feeling badly. The first is topic familiarity: an article about something you know well feels easy even when a quarter of its vocabulary is new, because you are reconstructing the argument from prior knowledge rather than from the text. You finish it feeling fluent and having learned nothing.

The second is that comprehension is lumpy. You can follow 95 per cent of a paragraph and miss the one clause that reverses its meaning. Feeling is measuring the 95; the 5 is where the information was.

This is the argument for counting rather than judging. A count is about your vocabulary, not an average learner’s, and it is not fooled by a familiar topic.

A two-minute test

Before committing forty minutes to something, spend two on this.

1. Take 200 words of it — the opening paragraphs, or the first two minutes of dialogue.
2. Count the words you would have to look up to be sure of the meaning. Not the ones you can guess; the ones you would actually check.
3. Under 4: comfortable. Good for volume and for listening, weak for learning new words.
4. Between 4 and 10: the useful window. This is what you want most of the time.
5. Over 20: you are decoding. Fine in short bursts with a specific goal, ruinous as a habit.

Automating the count

Doing this by hand works and costs two minutes. Doing it continuously is better, because the decision comes up several times a day.

Any tool that knows which words you have already saved can mark the ones you have not. In NoaLingua, Alt+Shift+W highlights every word on a page that your deck has never met, and the transcript panel puts a difficulty estimate on a YouTube video before you start it. On Netflix it cannot, for the reason explained in the Netflix piece: the platform never hands over the whole subtitle file.

The estimate is not the point. Being able to reject the wrong material in five seconds is.