The receipts
How the word list was built
Every vocabulary app claims its list is the important words. Here is exactly where this one came from, including the parts that didn't survive.
Where it starts
Seven sources, checked against each other
The list starts from 7 independent Danish frequency lists: a 3,000-word appendix from an academic paper on written Danish, a subtitle frequency list, and five teaching lists of various sizes.
The important step is what came next. Rather than trusting any of them, each source was measured against the others two ways — do they order words the same (rank correlation), and are they secretly the same list (set overlap)?
That test changed the answer.
What didn't survive
What we threw out
One "source" was a copy of another
A 500-word teaching list correlated +0.98 with the subtitle frequency list — it was a truncated copy of it. Counting both would have let one body of evidence vote twice. They were merged into a single spoken-language voice before anything was allowed to vote.
Two lists had no frequency signal at all
A 100-word phrasebook correlated −0.14 against the academic word list — statistically, nothing. Another turned out to be ordered by the length of the English translation, not by Danish frequency in any form.
Neither can say anything about which word is more common. But a human did choose to teach those words, so they still contribute a small membership bonus — enough to break a tie, never enough to reorder the top of the list.
The subtitle list was cut off at 5,000
Past that point it degrades into proper nouns and OCR damage — real
entries in the raw data include giad, ioyaie
and geoiog, where the letter l was misread as
i. Depth is not the same as quality.
The merge
How the surviving sources were combined
The ranked sources are fused with Reciprocal Rank Fusion,
scoring each word as the sum of weight / (60 + rank) across the
lists that contain it.
The alternative — averaging each word's rank — fails badly here, because the lists have very different depths. A word missing from a 1,000-word list would be indistinguishable from a word that list judged rare, when in fact the list simply stopped early.
Written sources vote at full weight; the spoken voice at half. Nothing is weighted by feel — the weights follow the correlation evidence above.
Inside a row
What each entry carries
An entry is one learnable item: a word paired with a part of speech. Have the verb and have the noun are two different things to learn, so they are two rows — which is why 2,000 entries cover 1,757 distinct spellings.
Every row also records how much corroboration it has
(confidence), whether it appears in written Danish, spoken
Danish or both (register), where each individual source ranked
it, and how much those sources disagreed. None of that is hidden from you —
it's all in the published CSV.
How it sounds
Pronunciation
1,229 entries carry IPA, and the transcription is per form, not per headword — være, er, var and været each get their own, because a learner who only ever sees the infinitive's pronunciation is being short-changed.
Transcriptions come from Wiktionary's Danish entries. Where a form is not documented directly but is a known stem plus a regular Danish ending, it is derived by a named rule and labelled as derived rather than presented as if a dictionary had said it. Anything the rules can't reach is left blank instead of guessed.
Where it falls short
Known limitations
- Frequency data ages. The written source is an academic appendix, not a live corpus. Word frequencies drift.
- Subtitles aren't conversation. Film dialogue is written to be spoken, which is not the same as speech.
- 2,000 minus 1,610 entries rest on a single ranked source. They're marked as such rather than quietly mixed in.
- Part-of-speech tagging is imperfect. Tags come partly from automatic tagging of words in isolation, with no sentence for context. Sources also disagree with each other about labels.
The full pipeline — every script, every intermediate file, and the reasoning behind each decision — is open, and each stage documents not just what it does but which alternatives were tried and failed.
That's the list. Now use it.
1,639 of those entries are ready to study today, in the order this page just explained.
Start learning