How to Test Your Vocabulary Size in Any Language

"How many words do I know?" is one of the few progress questions in language learning that has a reasonably objective answer. Not an exact one, but an estimate with a known margin of error, which is more than most of us get from gut feel. This post covers the established tests, a do-it-yourself method that works for any language with a frequency list, and the maths that tells you how much to trust the result.

On this page

What a vocabulary size test actually measures

Nobody tests you on 20,000 words. Size tests work by sampling: pick a small number of words from each frequency band, check how many you know, and scale up. If you know 7 of 10 words sampled from the third thousand most common words (ranks 2,001 to 3,000), the estimate for that band is 700.

Three details change what the final number means:

• Receptive, not productive. Almost every size test asks whether you recognise a word's meaning, not whether you can use it. Your productive vocabulary will be smaller.
• The counting unit. Research on English usually counts word families, where use, used, useful and useless can count as one. Frequency lists for other languages are often lemmas (usar covers all its verb forms but not útil). The same learner gets a bigger number in lemmas than in families, so never compare figures across units.
• Your threshold for "know". Recognising a word in a multiple-choice line-up is easier than producing its meaning from a blank. Stricter criteria give lower numbers.

The established tests

The Vocabulary Size Test (English). Paul Nation and David Beglar's test is the best-known size measure for English learners. According to Nation's own specification document, the original version has 140 multiple-choice items, 10 from each 1,000-word-family level up to 14,000, and your score is multiplied by 100. Newer versions have 100 items covering 20,000 families, with the score multiplied by 200. It takes around 40 minutes for the 140-item test and around 30 for the 100-item versions, and bilingual versions exist for a number of first languages.

The same document is candid about the limits. There's deliberately no "I don't know" option, because Nation wants test-takers to make informed guesses, and no correction for guessing. Because of that, and because the multiple-choice format gives credit for partial knowledge, he describes the result as "a slightly generous estimate of vocabulary size." It also warns that the score depends on how seriously people sit the test. For some learners, a one-to-one administration, where the tester can pronounce unfamiliar words, encourage them and give feedback, "can double the score that they got on a group-administered test." If you take it, take it properly.

LexTALE (English, Dutch, German). LexTALE, developed by Kristin Lemhöfer and Mirjam Broersma (2012), is a quick yes/no test: after three practice items, you see 60 letter strings, 40 real words and 20 made-up ones, and say which are real. The score averages your accuracy on words and on non-words, so saying "yes" to everything doesn't pay. It doesn't give you a word count. It's a proficiency proxy: in the original study with Dutch and Korean speakers of English, it predicted scores on a translation-based vocabulary test well and correlated substantially with a general English proficiency test. It's useful for a fast snapshot, less useful for tracking growth in word terms.

Other languages have their own versions and adaptations, but coverage is patchy. For many languages, the practical option is to build your own.

A do-it-yourself size test for any language

You need a frequency list for your target language, ideally based on a large corpus of everyday language such as film subtitles, with at least 10,000 entries. Then:

1. Split the list into 1,000-word bands: 1–1,000, 1,001–2,000, and so on, up to 10,000.
2. Pick 10 words at random from each band. Use a random number generator, not your eye, or you'll unconsciously pick words you know. That's 100 words.
3. Skip proper nouns and obvious cognates you'd only "know" through your first language, or at least mark them, since they inflate the score.
4. For each word, write a meaning before you check. A translation, a synonym or a usage example all count. "Seen it before" does not.
5. Add five or six fake words that look plausible in the language. If you find yourself claiming to know one, your criterion is too loose. This is the same logic LexTALE uses with its non-words.
6. Score each band as known ÷ 10 × 1,000, and add the bands up.

A worked example. Say your results by band are 10, 9, 8, 6, 5, 4, 3, 2, 1 and 0 out of 10. That's 48 words known out of 100, so the estimate is 4,800 words (in whatever unit your list uses). Done carefully, it's a 20- to 30-minute job.

How much to trust the number

Ten words per band is a small sample, and small samples are noisy. You can put a number on the noise with basic binomial statistics. For a band where you know a fraction p of the words, the standard error of the band estimate is 1,000 × √(p(1−p) ÷ 10). At p = 0.5 that's about 158 words for that band alone.

Combining the bands from the example above gives a standard error of about 380 words for the total. A rough 95% range is twice that either side, so the 4,800 estimate really means "somewhere between about 4,050 and 5,550". If you sample 20 words per band instead, the range narrows to about ±540.

That matters most when you retest. The noise in the difference between two tests is larger than the noise in either one. With 10 words per band each time, a change smaller than roughly 1,050 words can't be told apart from luck. With 20 words per band, the threshold is roughly 740. So if you test in January, test again in March and see 4,800 become 5,100, you haven't learned anything about your progress. That's the sampling error doing what sampling error does.

Three ways to get a clearer signal:

• Use more words per band. Doubling the sample doesn't halve the noise (it cuts it by about 30%), but it helps.
• Test less often. Every three to six months gives real growth time to outrun the noise. Monthly tests mostly measure randomness.
• Draw a fresh random sample each time but keep the method identical: same list, same bands, same criterion for "know". If you change the method, the trend line breaks.

What to do with the result

The band-by-band pattern is often more useful than the total. Most learners show a clear drop-off somewhere, the point where bands go from mostly known to mostly unknown. That edge is where your vocabulary study gets the best return. If you're shaky in the 2,000–3,000 band, those words appear constantly in anything you read or hear. Learning rare words from the 9,000 band while that gap exists is poor value.

It's also a useful reality check on tools that report "words learned". An app's count is the number of items you've been shown and passed in that app. A sampled size test measures what you can recognise from the language as a whole. The two can differ a lot in either direction, and only the second tells you how much of a real text you'll understand.

Pair it with hours

A size estimate every few months gives you an outcome. Logged study time gives you the input that produced it. Together they tell you something neither can alone, such as roughly how many hours each extra thousand words is costing you, and whether that's changing as you move to rarer words. Keeping the time side honest is the easy part if logging is quick: in LangTrack you log each session by activity and minutes, and per-language stats keep the totals separate if you're learning more than one language.

For the broader trade-offs between counting words and counting time, see tracking vocabulary vs tracking time. For what your number means in practice, see how many words you need to understand a text, which turns coverage research into vocabulary targets. For other ways to measure ability, from recordings to can-do checklists, see how to measure your language progress, and for why depth of knowledge matters as much as breadth, vocabulary acquisition: quality over quantity.