Text stemming

Stemming — Lodestar.Text.Stemming

ran, running and runs are three spellings of one idea, and an index that stores them as three terms cannot match a query that uses the fourth. A stemmer cuts each word down to a shared key so they collide on purpose.

Lodestar.Text.Stemming holds seven stemmers, one static class each, all with the same single method: a string in, a string out.

Which stemmer?

flowchart TD
    A["What language is the text?"] --> B["English"]
    A --> C["French, German, Italian,<br/>Spanish, Portuguese"]
    B --> D{"Matching an existing<br/>Porter index?"}
    D -->|"no — this is a new index"| E["EnglishSnowballStemmer"]
    D -->|"yes"| F["PorterStemmer"]
    C --> G["the Snowball stemmer<br/>for that language"]

Loading

Language picks the stemmer, and for six of the seven that is the whole decision. A stemmer is built from one language's suffix rules and has nothing sensible to say about another's: GermanSnowballStemmer.Stem applied to French returns something, and that something is noise.

The one real choice: Porter or Porter2

English has two, and they are the same algorithm a generation apart. PorterStemmer.Stem is Martin Porter's 1980 original; EnglishSnowballStemmer.Stem is his own later revision, published as Snowball and universally called Porter2.

Porter2 is the one to reach for. It is Porter's own later revision, so where the two disagree it is the original being corrected. Over the 86 words both are pinned on, they disagree six times, and Porter2 trims less in five of the six:

word	Porter	Porter2
ties	ti	tie
fairly	fairli	fair
communism	commun	communism
generalization	gener	general
formative	form	format
homologou	homolog	homologou

generalization is the clearest of them: Porter trims all the way to gener, a fragment short enough that several unrelated words reach it. Porter2 stops at general, which still names something.

The reason to choose the original anyway is compatibility, not quality. An index built by Porter has to be queried by Porter, and a corpus already stemmed one way cannot be searched the other. That is the whole of the case for it.

using Lodestar.Text.Stemming;

string old = PorterStemmer.Stem("generalization");    // => gener
string current = EnglishSnowballStemmer.Stem("generalization");  // => general

What a stem is not

A stem is a key, not a word. musico, musica, musicos and musicas all stem to music, and Spanish cantar stems to cant — which is not Spanish. The output is meant to be compared against other output, never shown to a reader.

This is what separates stemming from lemmatisation, which returns the dictionary form and needs a dictionary to do it. Nothing here carries one: these are rule engines, small and fast, and they are wrong on irregular words by construction. Lodestar ships no lemmatiser.

The consequence for a search index is that both sides must be stemmed by the same stemmer — the documents when they are indexed, the query when it arrives. Stem one and not the other and the keys never meet.

What all seven share

• Input is lowercased first. The algorithms are defined on lowercase, so Running and running give the same stem, and the result is always lowercase.
• A null word is refused, with ArgumentNullException. An empty string is not: it comes back empty.
• Each is a static class with no state, so all seven are safe to call from any number of threads at once.
• Each is checked word for word against nltk, and the corpora are in tests/oracles. Where a stem looks wrong, it is wrong in the same way the reference is.

Types

Type	What it is
EnglishSnowballStemmer	English Porter2 — the one to use for new English text.
FrenchSnowballStemmer	French Snowball.
GermanSnowballStemmer	German Snowball.
ItalianSnowballStemmer	Italian Snowball.
PorterStemmer	English Porter (1980), for compatibility with an existing index.
PortugueseSnowballStemmer	Portuguese Snowball.
SpanishSnowballStemmer	Spanish Snowball.

See also

• Python → C# equivalence — the nltk call each of these replaces.
• From string to vector — where a stemmer sits in a pipeline.
• decisions/0008 — the one place a stemmer here follows nltk over the published algorithm.