0090 — Finnish falls back to the genitive where nltk does, and nowhere else

Status: accepted · Date: 2026-09-09

Context

The published Finnish description gives step 3 as one search:

Search for the longest among the following suffixes in _R_1, and perform the action indicated.

Snowball's reading of that is settled: the longest alternative is selected, its action runs, and if the action's condition fails the search is over. Nothing shorter is tried. Two constructed forms confirm nltk reads it that way too:

word	ending selected	condition	nltk 3.10.1
kotimaahon	hon	fails — a before it, not o	kotimaahon, whole
kissatta	tta	fails — a before it, not e	kissat, and not kis

kissatta is the sharper of the two: ta is in the same list and would have matched, and it is not used.

Four endings do not behave that way. When their condition fails, the final n is removed anyway — the outcome the step's own shortest alternative would have produced:

word	ending selected	condition	nltk 3.10.1	under the description
ihmisiin	siin	fails — mi before it, not Vi	ihmis	ihmisiin, whole
ihmisden	den	fails	ihmisd	ihmisden, whole
ihmistten	tten	fails	ihmist	ihmistten, whole
ihmiseseen	seen	fails — no long vowel before it	ihmises	ihmiseseen, whole

siin, den, tten and seen are exactly the four endings in the step that are a longer spelling of the genitive n. hön and the rest of the hXn group end in n as well and do not fall back, so the shared final letter is not the rule; being a spelling of the genitive is.

Decision

Match nltk. FinnishSnowballStemmer ends the search on a failed condition, except for siin, den, tten and seen, where it strips the genitive n — including that rule's own "delete the last vowel if a long vowel or ie is uncovered".

This is the same call as 0008 on Italian enza/enze, 0087 on Danish's apostrophe and 0005 on jellyfish. 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

• ihmisiin stems to ihmis rather than coming back whole. It is the one real Finnish word in the 217-word corpus that meets this rule, and it was the only failure of the first implementation, which had been written from the description and got the other 211 right.
• One real word is thin evidence for a rule, so the four constructed forms above are in tests/oracles/snowball_fi.json beside it, together with kotimaahon and kissatta for the side that does not fall back. The boundary is pinned from both directions rather than inferred from a single case.
• Where the two readings agree — every other word in the corpus — this decision changes nothing, which is why it took a constructed vocabulary to find at all.
• If nltk ever makes the step uniform, the corpus will drift and the Oracles are reproducible job will catch it. At that point this decision should be revisited rather than the corpus quietly regenerated.