Vocabulary test: method

What it measures

Receptive (recognition) vocabulary size in lemmas, by self-report corrected for false alarms to pseudowords. Not definition knowledge, production or language proficiency.

Procedure

English (method en-adaptive-1, list en-1): step 1 is a grid of 40 items: 5 random words from each of 7 frequency bands plus 5 pseudowords, shuffled; the player ticks the words they know ('could explain at least one meaning'). Step 2 is a grid of 60 new items: 4 fresh words per band, 25 extra words placed around the band where the corrected known share in step 1 is closest to 50% (weights halve with each band away), and 7 fresh pseudowords. No word repeats. Untimed; ticks can be changed before each step is submitted. Optional questions after the result: age group, native speaker or learner, years of learning, reading frequency. German, Spanish, French, Italian, Polish, Portuguese and Russian (lists *-1): one word at a time, 10 random words from each of 6 frequency bands (60 real words) plus 30 pseudowords, shuffled (90 items); the player answers 'I know it' or 'Made up'. 2 practice real words and 2 practice pseudowords with feedback first (skippable). Untimed; leaving the tab does not end the run.

Parameters

Scoring

Per band b: h = share of the band's real words marked known (English: step-2 words only, so the adaptive choice does not bias the estimate); f = share of pseudowords marked known (English: all 12 from both steps). Known share p = (h − f) / (1 − f), kept between 0 and 1 (0 if f = 1). Estimate = sum over bands of band size x p, rounded to 100 and limited to the list size; higher means a larger vocabulary. Range, English: central 90% of 4,000 simulated draws (Jeffreys Beta posteriors for each band's hit rate, one shared Beta draw for the false-alarm rate, and Beta uncertainty on each band's size from the original 70-candidate review), widened to include the estimate; at f = 1 the range is the whole list. Range, other languages: plus or minus 1.645 binomial standard errors summed over bands. Other languages also get a LexTALE-style accuracy = mean of the percentage of real words accepted and the percentage of pseudowords rejected. English percentiles: share of eligible players strictly below the estimate (current method and list, reliable results, not likely automated, first eligible attempt per player), shown only for groups of at least 100; otherwise published research figures are shown as context.

Valid-run rules

The server rebuilds the items from the run's seed (which items are real never reaches the page, except in practice). Every item must be answered; the English two-step format needs exactly 40 and 60 answers and cannot be mixed with the one-step format. Results with too many pseudowords marked known are marked shaky and get no percentile. The server rebuilds every run from the raw log (and the run's seed where there is one) and computes the score itself; anything the page says about right or wrong is ignored. A log that breaks the game's rules is rejected, and nothing is stored. Stored runs can carry quality flags that are never shown and never block a result: 10 or more runs from one IP address, a missing visitor id, tab switches, and automation signals (automated browser, impossible speed, no input events, machine-regular timing).

Differences from the classic task

Yes/no tests with pseudowords go back to Meara and Buxton (1987). LexTALE (Lemhöfer & Broersma 2012) is a fixed 60-item test with a validated percentage score; Brysbaert et al. (2016) estimated vocabulary size from a very large online word-knowledge dataset with a different dictionary and a looser idea of knowing a word. This estimator, its lists and its two-step design are new and have not been independently calibrated. The unit is distinct written lemmas, not word families as in Nation's Vocabulary Size Test.

Limitations

Self-report: people differ in how readily they mark a word as known, and pseudowords correct this only partly. Band sizes come from reviewed samples, not a full dictionary count. Lists and band sizes differ between languages, so estimates are not comparable across languages. Specialist interests and familiar word parts change the result. Remembering words from an earlier run changes later results. It does not measure intelligence, fluency or health. Comparison players are site visitors, not population norms.

Languages and review status

English (two-step) plus German, Spanish, French, Italian, Polish, Portuguese and Russian (one-step); other site languages show the English test. Real words: Wiktionary lemmas (via kaikki.org) ranked by word frequency (wordfreq 3.1.1), with a hand-reviewed sample per band; shared under CC BY-SA 4.0 as the licence requires. Pseudowords were made for each list from its letter patterns and reviewed by hand.