How Many English Words Do You Actually Need?
要点まとめ
- The first 2,000 words of English cover about 80% of the running words in a general written corpus, and about 90% of a short novel, which is why early vocabulary work feels so productive.
- Reaching the 98% coverage associated with comfortable unassisted reading of novels and newspapers takes roughly 8,000 to 9,000 word families, by the estimate Nation published in 2006. Spoken text needs 6,000 to 7,000.
- The numbers depend entirely on what is being counted. A word family, a lemma and a word type give very different totals for the same knowledge.
- Educated adult first-language speakers of English are estimated at around 17,000 base words by Goulden, Nation and Read, and at around 20,000 word families in the figure Nation and Waring quote. Method drives the difference.
- Vocabulary size is a symptom of reading volume, not a substitute for it. The number is useful for diagnosis and useless as a goal on its own.
"How many words do I need?" has a real answer, but it depends on two things people rarely specify: what you are counting, and what you want to do with the language.
What is actually being counted
Three different units get called "words", and they produce wildly different totals for the same person.
| Unit | Definition | Example, one item |
|---|---|---|
| Word type | Every distinct written form | run, runs, running, ran = 4 items |
| Lemma | A headword plus its inflections | run, runs, running, ran = 1 item; runner = separate |
| Word family | A base word plus inflections and regular derivations | run, runs, running, ran, runner, rerun = 1 item |
Most vocabulary research uses word families, so that is the unit in the rest of this post. It matters: a learner who "knows 3,000 word families" knows well over 6,000 distinct written forms. If you see two vocabulary claims that disagree by a factor of two, this is usually why.
The frequency curve, and why early progress feels so fast
English vocabulary is distributed very unevenly. A small number of words do an enormous amount of the work.
Here is Nation and Waring's Table 1, cumulative coverage of the Brown corpus — about a million running words across 500 texts:
| Vocabulary size | Text coverage |
|---|---|
| 1,000 | 72.0% |
| 2,000 | 79.7% |
| 3,000 | 84.0% |
| 4,000 | 86.8% |
| 5,000 | 88.7% |
| 6,000 | 89.9% |
| 15,851 | 97.8% |
Three caveats the authors attach themselves, and which most reproductions of this table drop. The figures are lemmas, not word families — word families would give fractionally higher coverage. They come from a deliberately diverse corpus, which makes them a conservative estimate. And they assume you learn the high-frequency words before the low-frequency ones, which is idealised.
Coverage also depends heavily on what you are reading. The same first 2,000 words of the General Service List cover 90.6% of short novels but only 78.4% of academic text, and around 96% of informal spoken English.
The shape is the point. The first 2,000 words buy you roughly 80% of a general page. The next 4,000 buy you ten points more. That is why beginners feel like they are flying and intermediate learners feel like they have stopped — the return per word learned collapses by an order of magnitude, and nothing about your effort has changed.
Why 98% is the number that matters
Coverage sounds abstract until you connect it to comprehension. Marcella Hu and Paul Nation's 2000 study manipulated known-word density in a text and measured what learners actually understood. The finding: most learners reached adequate unassisted comprehension at around 98% coverage. At 95%, only a minority did.
Ninety-five per cent sounds close to 98. It is not. Ninety-five per cent means one unknown word in twenty — five per hundred, one every line or so. Ninety-eight per cent means one every two or three lines. That difference is the difference between reading and decoding.
One honest caveat, because 98% is repeated everywhere as though it were a physical constant. The regression that produced it was run on 66 university students in New Zealand, and a 2023 replication in Language Learning by Kremmel and colleagues revisited exactly that limitation. Treat 98% as a well-supported working threshold with a thin empirical base, not as a law.
Nation's 2006 estimate for what it takes to reach 98% coverage is 8,000 to 9,000 word families for written text, and 6,000 to 7,000 for spoken English, which uses a narrower vocabulary. Those figures come from trialling fourteen 1,000-word-family lists built from the British National Corpus.
So the honest answer to "how many words do I need" is: 2,000 to survive, 3,000 to 5,000 to have a working base for comprehension — Nation and Waring's own figure — and 8,000 to 9,000 to read whatever you want without help. The last of those is a multi-year project for almost everyone.
Thresholds worth knowing about
| Size (word families) | What it gets you |
|---|---|
| 1,000 | Basic survival: signs, menus, simple exchanges. |
| 2,000-3,000 | The high-frequency core. Graded readers, simple conversation, most spoken English. This band deserves direct study. |
| 2,000 + the Academic Word List | The AWL is 570 word families identified by Averil Coxhead in 2000. It covers about 10.0% of the tokens in academic text — against just 1.4% of a fiction collection of the same size — sitting on top of the general high-frequency words. Very high return for university study. |
| 5,000 | Comfortable with most non-fiction on familiar topics. Newspapers still resist. |
| 8,000-9,000 | Unassisted reading of novels and newspapers. |
| 17,000-20,000 | Educated adult first-language speakers. Goulden, Nation and Read give around 17,000 base words; Nation and Waring quote around 20,000 word families. |
The single most useful entry there for anyone studying or working in English is the Academic Word List. It is small, well defined, published free, and disproportionately productive — 570 families is a few months of deliberate work for a step change in how academic text feels.
How CEFR levels line up
The CEFR attaches no word count to any level, so every table you find pairing the two is somebody working backwards. The best-sourced attempt is the one James Milton reproduces from Meara and Milton's own work, built by testing learners who had taken and passed Cambridge exams at every CEFR level:
| CEFR | Milton (2010), X_Lex estimate |
|---|---|
| A1 | under 1,500 |
| A2 | 1,500-2,500 |
| B1 | 2,750-3,250 |
| B2 | 3,250-3,750 |
| C1 | 3,750-4,500 |
| C2 | 4,500-5,000 |
These are lemmas, not word families, and the test scores out of the most frequent 5,000 words only — so the C2 row is bounded by the instrument rather than by the learners. That is why the numbers look small next to the 8,000-9,000 figure above: they are counting a different unit with a different ceiling. Do not put the two tables side by side and subtract.
Do not read this table as a route. Vocabulary size correlates with level; it does not cause it. Someone who has drilled 5,000 word families on flashcards and never read a long text in English is not B2 — they are a B1 reader with a large passive word list, and they will discover this the first time they try a real article. What the levels mean in practice is set out in CEFR levels explained.
What your test result actually measures
Vocabulary size tests estimate by sampling: they draw words from across the frequency bands, ask whether you know them, and extrapolate.
Two limitations are worth holding on to.
They measure recognition, not use. Saying yes to "ubiquitous" is not the same as being able to use it. Recognition vocabulary typically runs well ahead of production vocabulary, so your test result describes the ceiling of your reading, not of your writing.
Learners over-report. This is why well-built tests include invented words that look like real English. If you claim to know several of those, the test scales your result down. A test without that correction will flatter you, sometimes by thousands.
Given both, treat a single result as a band rather than a number. The useful signal is the direction over six months, and — more actionably — the frequency band where your accuracy falls off. That tells you exactly where to aim. The vocabulary size test reports the band as well as the estimate for this reason.
What to do with the number
Very little, on its own. It is a diagnostic, not a target.
If you are below 3,000: work directly on high-frequency vocabulary. This band is small enough to study deliberately and pays back immediately.
If you are between 3,000 and 8,000: stop optimising flashcards. This is the mid-frequency range, and it is too large to memorise item by item. It is acquired by reading a great deal at a level where you meet these words often enough to keep them. How to learn vocabulary from reading covers the filter that decides which of them are worth a card anyway.
If you are above 8,000: vocabulary is no longer your bottleneck. Reading speed, academic register or listening probably is.
In every case the mechanism is the same: meet words repeatedly in text you can mostly understand. Which brings the question back to text selection — how to choose reading material at your level — and to having a steady supply of it. LingoRead generates passages at a chosen CEFR level and builds the vocabulary list from the passage itself, so the words you review are ones you have actually met rather than a list downloaded from somewhere.
Sources
- Nation, I. S. P. and Waring, R. (1997). Vocabulary size, text coverage and word lists. In Schmitt, N. and McCarthy, M. (eds), Vocabulary: Description, Acquisition and Pedagogy, 6-19. Cambridge University Press. Source for the Brown corpus coverage table, the General Service List coverage figures, the 3,000-5,000 comprehension base and the 20,000 word-family native-speaker figure.
- Hu, M. and Nation, P. (2000). Unknown vocabulary density and reading comprehension. Reading in a Foreign Language, 13(1), 403-430. Source for the 98% threshold.
- Kremmel, B. et al. (2023). Unknown vocabulary density and reading comprehension: replicating Hu and Nation (2000). Language Learning. The replication, and the source for the caveat about the original sample size.
- Nation, I. S. P. (2006). How large a vocabulary is needed for reading and listening? The Canadian Modern Language Review, 63(1), 59-82. Source for 8,000-9,000 written and 6,000-7,000 spoken.
- Goulden, R., Nation, P. and Read, J. (1990). How large can a receptive vocabulary be? Applied Linguistics, 11(4), 341-363. Source for the ~17,000 base-word native-speaker estimate.
- Coxhead, A. (2000). A new Academic Word List. TESOL Quarterly, 34(2), 213-238. Source for 570 word families, 10.0% of academic tokens and 1.4% of fiction. The list itself is free.
- Milton, J. (2010). The development of vocabulary breadth across the CEFR levels. EUROSLA Monographs 1, 211-232. Source for the CEFR vocabulary table, given there as Table 5, adapted from Meara and Milton (2003), and for the X_Lex 5,000-word ceiling.