Lodestar.Text 0.6.0. 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.

Bm25Idf

Which inverse document frequency a Bm25Index weights terms by.

public enum Bm25Idf

Values — RobertsonFloored is log((N - n + 0.5) / (n + 0.5)) with negatives floored; Lucene is log(1 + (N - n + 0.5) / (n + 0.5)).

Example — the two are different numbers, not a rescaling of one another.

using Lodestar.Abstractions;
using Lodestar.Text.Search;
using Lodestar.Text.Vectorization;

var vectorizer = new CountVectorizer();
CsrMatrix counts = vectorizer.FitTransform(
    ["the cat sat", "the dog sat sat", "a bird flew far away today"]);
int sat = vectorizer.GetFeatureNames().ToList().IndexOf("sat");

double floored = new Bm25Index(counts).Score([sat])[1];  // => 0.1013…
double lucene = new Bm25Index(counts, new Bm25Options(Idf: Bm25Idf.Lucene)).Score([sat])[1];

bool higher = lucene > floored;  // => True

Remarks — the choice is stated rather than inherited because the three published forms give different numbers, and the ordering of two documents can differ when a query mixes a common and a rare term.

RobertsonFloored is what this package is checked against. Robertson's raw form goes negative once a term appears in more than half the corpus, which would let a match subtract from a score. rank_bm25 replaces every negative with Epsilon × (mean of the raw values), and this reproduces that. Two details matter and are easy to get wrong:

• the mean is taken over the raw values, negatives included — flooring first and averaging after gives a larger floor;
• the floor is itself negative when that mean is below zero, so scores in such a corpus are negative and a longer document scores higher. That is the reference's behaviour, not a defect.

Lucene is non-negative by construction, so it needs no floor and ignores Epsilon. It is the variant to pick when negative scores would be awkward downstream — a fusion step, for instance.

BM25L and BM25+, the other lower-bounding variants, are not here.

Applies to — net10.0, netstandard2.0.

See also — Bm25Options, Bm25Index.Score.