0092 — Romanian keeps the cedilla alphabet the description writes, and reads its regions off the word

Status: accepted · Date: 2026-09-09

Context

Romanian is the only Romance language of #176's nine, so it is the one that reaches RomanceSnowballWorker rather than only SnowballWorkerBase. Two things had to be settled before it could be written, and neither is visible from the algorithm alone.

1. Two Unicode spellings of the same two letters

Romanian writes ș and ț with a comma below (U+0219, U+021B). Before Unicode 3.0 had those code points, the same letters were typed with a cedilla — ş U+015F and ţ U+0163 — and both spellings are still in circulation.

The published Snowball description writes the cedilla, and so does nltk. Its four suffix tables carry ş/ţ and nothing else, which the tables themselves say:

step 1  aţiune  iţiune          step 2  işti  ităţi
step 3  aţi  eţi  iţi  âţi  seserăţi  eşti  eşte  aşi  uşi  işi  âşi …

A word typed the modern way therefore matches none of them, and is stemmed only by the endings that happen to have no ș/ț in them:

word	cedilla	comma below
informaţie / informație	inform	informaţ → informaț
româneşte / românește	român	româneșt
ştiinţă / știință	ştiinţ	științ
mulţumesc / mulțumesc	mulţum	mulțum (agree: esc has neither letter)

2. nltk measures its regions against the word it started with

nltk keeps R1, R2 and RV as strings, sliced from the word before step 0, and updates them by hand — but only in some branches. Step 1's at and it replacements re-slice R2; its abil, ibil, iv and ic replacements do not, and step 0's aţie and aua branches never touch RV. Step 3 then asks where a suffix sits with rv.index(suffix), which finds its first occurrence rather than the one that ends the word.

Swept over 30 252 words — 47 stems crossed with every ending of every table, and with a second ending appended — the two readings part on 35, in exactly two shapes:

shape	example	nltk	this
two step 1 replacements chained, so step 2 measures R2 against a word that no longer exists	dormativitate	dorm	dormat
a step 3 ending that occurs twice in RV	posândând	posândând	posând

Both need an ending repeated or a derivational chain Romanian does not build. Over the 349 real words of tests/oracles/snowball_ro.json the two readings agree everywhere, and so do the 27 real -itate nouns probed separately (activitate, responsabilitate, electricitate, …), which are the shape the first row would have to come from.

Decision

Keep the cedilla alphabet, and measure the regions against the word as it stands.

RomanianSnowballStemmer folds nothing: ş and ș are two letters, as they are to the description and to nltk, and the corpus freezes both spellings so the behaviour is visible rather than implied. A caller whose text is modern Romanian normalises it before calling — one Replace per letter — and that is a decision about their text, not one this library can make for them without changing what Stem returns for everyone else.

On the regions, this is 0086's line again, in a second language: 0008's parity is about the reference's rules, and a region that has gone stale in one branch of one step is its bookkeeping. Following it would mean carrying R1, R2 and RV as mutable strings updated branch by branch, which is a port of nltk's implementation rather than of the published algorithm — what 0003 and this lot's brief both refuse — and RomanceSnowballWorker, which Spanish, Portuguese and Italian already share, does not work that way.

Consequences

• The library is exact against nltk on every Romanian word this corpus and the probes could reach. The two divergent shapes are named above and pinned by RomanianSnowballStemmerTests, since the corpus is compared exactly and cannot carry them.
• A user with modern Romanian text gets a stemmer that mostly does not fire. That is the reference's behaviour and the description's, but it is the one thing in this lot most likely to be read as a bug, so it is said plainly on the reference page rather than left to the corpus.
• If a later lot wants the two spellings to collide, it is a new decision with a new public shape — an option, or a documented normaliser — and not a quiet fold inside Stem.