How Many Words Do You Need to Understand a Text?

You'll often see the claim that you need to know 98% of the words in a text to understand it. The number is real, and it comes from actual research, but it's usually quoted without the details that tell you how far to trust it. Here's where it came from, what later studies found, how many words it implies, and how to check the coverage of any text you're thinking of reading.

On this page

What "coverage" means

Lexical coverage is the percentage of running words in a text that you know. Every occurrence counts, so if an unknown word appears four times, that's four unknown tokens. The percentages sound close together, but they describe very different reading experiences:

• 98% coverage: 1 unknown word in 50, about 6 on a 300-word page
• 95% coverage: 1 in 20, about 15 per page
• 90% coverage: 1 in 10, about 30 per page

Going from 95% to 98% cuts the unknown words on each page from 15 to 6, a 60% drop. That's why the difference between the two figures matters so much in the research.

Where 98% comes from: Hu and Nation (2000)

The source is a study by Marcella Hu Hsueh-chao and Paul Nation in Reading in a Foreign Language. Sixty-six adult learners on a pre-university English course in an English-speaking country (chosen as the most proficient on their course) read a 673-word short story. The researchers made four versions by replacing low-frequency words with nonsense words, so the text was 100%, 95%, 90% or 80% familiar. They defined adequate comprehension as at least 12 correct out of 14 multiple-choice questions.

The results:

• 100% version: 15 of 17 readers reached adequate comprehension
• 95% version: 6 of 17
• 90% version: 4 of 16
• 80% version: 0 of 16

Nobody read a 98% version. The 98% figure came from a simple regression on those results: the authors concluded that "around 98% coverage of vocabulary is needed for learners to gain unassisted comprehension of a fiction text." That's a reasonable estimate, but it rests on one story, one question format, and about 16 or 17 people per condition.

What later studies found

No sharp threshold. Schmitt, Jiang and Grabe (2011) tested 661 participants from 8 countries. They measured how many words in two texts each reader knew, then tested comprehension. They found "a relatively linear relationship between the percentage of vocabulary known and the degree of reading comprehension" and "no indication of a vocabulary 'threshold'". Every extra percentage point helps a bit. There's no cliff at 98%. The authors still concluded that 98% is "a more reasonable coverage target for readers of academic texts."

Two useful benchmarks. Laufer and Ravenhorst-Kalovski (2010) proposed two thresholds instead of one: an optimal one of 98% coverage, which they linked to knowing about 8,000 word families, and a minimal one of 95% coverage, linked to 4,000 to 5,000 word families. Both coverage figures count proper nouns as known words.

A replication that didn't fully replicate. In 2023, Kremmel and colleagues repeated Hu and Nation's design with 104 adult learners in Sri Lanka, outside an academic setting, and added the 98% version the original lacked. Scores rose as the density of unknown words fell, as you'd expect, but in the authors' words, "the results of the original study could not be fully replicated." In particular, they did not find that 98% coverage was a threshold that guaranteed adequate comprehension. Treat 98% as a sensible rule of thumb, not a law.

Coverage is necessary, not sufficient

Hu and Nation were careful about this themselves. They wrote that their conclusion "must not be interpreted as saying that with 98% coverage of the vocabulary no other skills or knowledge are needed." Their readers were literate adults who already knew a good deal of English grammar, had experience reading English and brought background knowledge to the story. Those things explain why a minority of readers still did well on the 95% and 90% versions.

It works in the other direction too. You can know every word in a sentence and still lose the thread if the grammar is dense, the topic is unfamiliar or the writer leans on cultural references. And the story they used was, in their words, not a "difficult" text: a fiction text with "a strong chronological story line". The authors expected newspapers and academic texts to "place greater demands on the reader."

They also tied the 98% figure to one particular kind of reading: extensive reading for language growth, where you read without a dictionary and pick up some new words along the way. They didn't set a figure for intensive reading, where a teacher, dictionary or glossary helps you through, though they expected learners with a glossary to understand much more. For reading purely to build fluency, they described material with no unknown language features at all.

How many words that takes

Paul Nation turned the coverage figures into vocabulary targets for English. In his 2006 analysis, 98% coverage needs a vocabulary of 8,000 to 9,000 word families for written text and 6,000 to 7,000 for spoken text. The specification for his Vocabulary Size Test breaks this down further: about 9,000 families for novels, 8,000 for newspapers, 7,000 for spoken English and 6,000 for children's films.

A word family is a base word plus its inflections and common derivations, so 8,000 families is a lot more than 8,000 dictionary entries. These are also English figures. A language with richer morphology, or one that shares many cognates with your first language, will produce different numbers, so use them as orders of magnitude rather than targets to copy.

Listening needs less than reading

Spoken language works differently. Van Zeeland and Schmitt (2013) manipulated coverage in four informal spoken narratives and tested 36 native and 40 non-native listeners. Most native and non-native participants "could adequately comprehend the spoken texts with only 90 per cent coverage", although the non-native listeners varied a lot at that level. At 95%, the non-native listeners showed relatively good comprehension with much less variation. By their estimate, 95% coverage for this kind of speech means knowing 2,000 to 3,000 word families, compared with the 6,000 to 7,000 that a 98% target implies.

The caveat is the material. Informal narratives are easier than lectures, news or fast dialogue between native speakers. Still, it explains a common experience: people often follow a casual podcast well before they can read a novel comfortably.

How to check the coverage of a text

You don't need software. Take a passage of exactly 200 words from the book, article or transcript you're considering. Read it and count every token you don't know, including repeats. Then:

• 4 or fewer unknown tokens: 98% or above. Good for reading without a dictionary.
• 5 to 10: between 95% and 98%. Readable with effort, and with occasional look-ups.
• 11 to 20: between 90% and 95%. Fine for intensive study with a dictionary, tiring for pleasure reading.
• More than 20: below 90%. In Hu and Nation's study, only 4 of 16 readers reached adequate comprehension at 90%, and none did at 80%.

Check two or three passages from different parts of the text, since the opening chapter of a novel is often harder than the rest. If you're deciding between books, this five-minute check beats the back-cover blurb.

What to do with this

The research suggests a two-track approach. For extensive reading, the high-volume reading where you build fluency, pick material at roughly 98% coverage or above, which for most learners means graded readers or easy native material. For intensive reading, where you're deliberately studying a text, 90 to 95% is workable as long as you expect to look words up.

Then keep an eye on your time per book or show. If a novel you rated at 95% is taking three times as long per chapter as your last one, coverage is probably the reason. In LangTrack, the media library on Pro records time per book or series, which makes that comparison easy to spot.

For a stage-by-stage plan for reading, see reading in a foreign language. To find out roughly how many words you know in the first place, see how to test your vocabulary size. For logging graded readers specifically, see tracking graded reader progress, and for the input side more generally, tracking comprehensible input.