0087 — Danish's apostrophe rule: follow nltk, not the published description

Status: accepted · Date: 2026-09-09

Context

Danish is the one Snowball language whose published description gives R1 two definitions. Every other language sets it the standard way; Danish sets it in three steps, because the apostrophe separates a Danish suffix from a word of foreign origin:

1. If the word contains an apostrophe, R1 is set to start after the first apostrophe.
2. Otherwise, R1 is set in the standard way.
3. In either case, R1 is then adjusted so that the region before it contains at least 3 characters.

nltk implements step 2 and step 3 and not step 1: its Scandinavian stemmers share one region routine, and that routine has no apostrophe in it.

The two readings part on the words the rule exists for. pc'er has its first vowel at index 3, so the standard R1 lands at 5 — past the end of a five-letter word — and no suffix can lie in it. Under the published rule R1 starts after the apostrophe, at 3, so er is inside it:

word	published description	nltk 3.10.1
pc'er	pc'	pc'er
cd'er	cd'	cd'er
tv'et	tv'	tv'et

The rule is invisible on longer words. computer'en and bil'er reach er/en through the standard R1 as well, and stem to computer' and bil' either way, which is why a corpus of ordinary Danish never meets this.

Decision

Match nltk. DanishSnowballStemmer sets R1 the standard way and floors it at three characters, with no apostrophe case.

This is the same call as 0008, where the published Italian description and nltk part on enza/enze, and as 0005, where jellyfish diverges from the textbook definition. The contract this package offers is behavioural parity with the Python library a user is migrating from, and for the stemmers that library is nltk — it is what the oracle corpora are frozen from and what docs/equivalence.md names as the reference.

Consequences

• pc'er, cd'er and tv'et come back whole, which the published description says they should not. They are correct as parity, and that is the guarantee this library offers.
• The six apostrophe words are in tests/oracles/snowball_da.json rather than left out of it. A divergence nothing exercises is a divergence that gets quietly re-introduced by the next reader of the description; these pin it, and the Oracles are reproducible job is what would report nltk adopting the rule.
• Found before the implementation rather than by a failing case, which is why this decision reads as a choice and 0008 reads as a diagnosis: the published description states the apostrophe rule in the same paragraph as R1, so the question was unavoidable rather than latent.