Lodestar.Text 0.4.0. 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.

CountVectorizerOptions

Everything that decides what counts as a term, before anything is counted.

public sealed record CountVectorizerOptions

Properties — Lowercase (default true) folds case before tokenizing, so Apple and apple are one term. TokenPattern (default \b\w\w+\b) is the regular expression a token must match — note the two \w, which is why single-letter words are dropped. Analyzer (default AnalyzerKind.Word) chooses words or character n-grams. NgramRange (default (1, 1)) is the inclusive range of n-gram lengths. StopWords (default none) is a set removed after tokenizing. StripAccents (default false) folds accented characters to their base. MinDf and MaxDf (defaults 1 and 1.0) drop terms appearing in too few or too many documents. Binary (default false) records presence as 1 rather than the count.

Example — the two defaults that surprise people, made visible.

using Lodestar.Text.Vectorization;

// "a" is a single letter, so the default token pattern never sees it as a term.
var cv = new CountVectorizer();
int features = cv.FitTransform(["a cat eats"]).ColumnCount;  // => 2

Remarks — every default here is scikit-learn's, and the properties answer to lowercase, token_pattern, analyzer, ngram_range, stop_words, strip_accents, min_df, max_df and binary. Reproducing \b\w\w+\b rather than choosing something more obvious is the single decision that keeps a ported pipeline giving the same columns, and it is also the one that makes "I" and "a" vanish from a corpus without saying so.

MinDf is read as a count when integral and as a proportion when fractional — MinDf = 2 means two documents, MinDf = 0.5 means half of them. MaxDf does not follow that rule at its default: MaxDf = 1.0 is a proportion meaning "in up to all of them", which is why the default drops nothing. Measured, over two documents sharing the, MaxDf = 1.0 keeps all three terms. Both properties are double, so writing 1 rather than 1.0 changes nothing.

This is a record, so two options objects with the same settings are equal. StopWords is compared as a set rather than as a sequence, which is why Equals and GetHashCode are written by hand rather than synthesised.

Applies to — net10.0, netstandard2.0.

See also — CountVectorizer, AnalyzerKind, StopWords, the Python equivalence table.

Members

Member	What it does
CountVectorizerOptions.Equals	Value equality, comparing stop words as a set.
CountVectorizerOptions.GetHashCode	A hash consistent with that equality, in O(1).