Lodestar.Text 0.5.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

AnalyzerKind

Whether a document is cut into words or into runs of characters.

public enum AnalyzerKind { Word, Char, CharWordBoundary }

Members — Word splits on the token pattern and makes each token a feature, which is the default and what most corpora want. Char slides an n-gram window over the whole document, spaces included, so a feature can straddle two words. CharWordBoundary slides the same window but never crosses a word boundary, padding short words instead.

Example — the same document, cut three ways.

using Lodestar.Text.Vectorization;

string[] docs = ["the cat eats"];

var words = new CountVectorizer(new CountVectorizerOptions { Analyzer = AnalyzerKind.Word });
int wordFeatures = words.FitTransform(docs).ColumnCount;  // => 3

var chars = new CountVectorizer(new CountVectorizerOptions
{
    Analyzer = AnalyzerKind.Char,
    NgramRange = (3, 3),
});
int charFeatures = chars.FitTransform(docs).ColumnCount;

Remarks — character n-grams are what to reach for when word boundaries are unreliable: misspellings, agglutinative languages, code, or text with no spaces at all. They cost columns — a character trigram vocabulary is far larger than a word vocabulary over the same corpus — and they are robust in exactly the cases a word analyzer breaks.

CharWordBoundary is the middle position: it keeps the robustness to spelling inside a word while refusing to invent features that span two of them. scikit-learn calls the three analyzer='word', 'char' and 'char_wb', and the behaviour is the same.

NgramRange applies to whichever of the three is chosen, and means something different in each: a range of (1, 2) under Word is unigrams and bigrams of words, and under Char is one- and two-character runs.

Applies to — net10.0, netstandard2.0.

See also — CountVectorizerOptions, CountVectorizer, the Python equivalence table.