Text keyword extraction - CyrilB1531/lodestar GitHub Wiki

Development build. This page describes main, not a released package. The latest published Lodestar.Text is 0.6.0 — read its documentation.

Home › Text

Keyword extraction

Lodestar.Text.Keywords holds two unsupervised extractors — Rake.Extract and TextRank.Extract — and neither needs a model: both read the co-occurrence of the document's own words. Together with Lodestar.Onnx and Mmr.Select they also compose into a KeyBERT-style pipeline, the last section below.

dotnet add package Lodestar.Text

RAKE: candidates from stop-word-delimited runs

RAKE never builds a graph. It splits the document into runs of words between stop words and punctuation, then scores each run by summing a per-word score — how often the word occurs, and how much company it keeps — over the run.

using Lodestar.Text.Keywords;

string abstractText =
    "Compatibility of systems of linear constraints over the set of natural numbers.";

var rake = new Rake();   // StopWords.English by default -- nothing downloaded
IReadOnlyList<KeywordMatch> hits = rake.Extract(abstractText);

foreach (KeywordMatch hit in hits.Take(3))
{
    Console.WriteLine($"{hit.Phrase}  ({hit.Score:F2})");
}

Nothing here is downloaded at run time: RakeOptions.StopWords defaults to StopWords.English, already in the assembly, and a caller who wants rake-nltk's own list passes it explicitly. See decision 0005 for why, and RakeOptions for every switch — Metric, the length bounds, and whether a repeated candidate is reported once or every time it occurs.

TextRank: candidates from a ranked graph

TextRank builds a co-occurrence graph over the document's stems — an edge between two words that stood within TextRankOptions.Window tokens of each other — and ranks it the way PageRank ranks a link graph: a power iteration to the dominant eigenvector. The highest-ranked stems are kept, and ones that stood adjacent in the source are glued back into a phrase.

using Lodestar.Text.Keywords;

string abstractText =
    "Compatibility of systems of linear constraints over the set of natural numbers. " +
    "Criteria of compatibility of a system of linear Diophantine equations.";

var textRank = new TextRank(new TextRankOptions { Words = 4 });
IReadOnlyList<KeywordMatch> ranked = textRank.Extract(abstractText);

string best = ranked[0].Phrase;

"Dominant" is not a formality. summa, the Python reference, reads scipy.linalg.eig's first column without checking it is the dominant one, and for a near-bipartite co-occurrence graph it often is not — measured on a real abstract, where that column belongs to an eigenvalue of −0.85 against a dominant 1.0. TextRank.Extract always ranks by the dominant eigenvector; two documents drafted for the frozen oracle corpus disagreed with it for exactly this reason and were removed by hand before the corpus was written. What is left is worse than a wrong pick: when the graph's transition matrix has a repeated eigenvalue — measured, two_sentences carries 0.85 at multiplicity 3 — which column eig returns first is not even the same from one machine's BLAS build to another's, so summa's own output is not reproducible. The oracle generator no longer trusts it: it selects the dominant left eigenvector itself, by eigenvalue rather than by column position, before calling summa at all — forced by reproducibility, not chosen. Decision 0005 has the measurement and TextRank's own Remarks have the rest of the divergence.

KeyBERT-style selection: RAKE candidates, embedded, spread out by MMR

keybert — the Python library the equivalence table compares Mmr.Select against — is not one algorithm but a composition of three: RAKE-shaped candidate generation, a sentence embedding of the document and of each candidate, and Maximal Marginal Relevance to pick a spread-out subset instead of the n most relevant (which are usually near-duplicates of each other). Nothing in Lodestar bundles that composition into one call — each of the three pieces already exists as its own package, and writing it out is a dozen lines:

using Lodestar.Embeddings.Search;
using Lodestar.Onnx;
using Lodestar.Text.Keywords;

// 1. Candidates: every RAKE run, deduplicated, is a reasonable keyword shape.
IReadOnlyList<KeywordMatch> candidates = new Rake(new RakeOptions { IncludeRepeatedPhrases = false })
    .Extract(document);
string[] phrases = candidates.Select(c => c.Phrase).ToArray();

// 2. Vectors: the document and every candidate, in one batch.
using var embedder = new OnnxTextEmbedder("model.onnx", wp);
string[] batch = [document, .. phrases];
float[][] vectors = embedder.EmbedBatch(batch);
float[] documentVector = vectors[0];
float[][] candidateVectors = vectors[1..];

// 3. Selection: spread the top picks out instead of returning near-duplicates.
int[] chosen = Mmr.Select(documentVector, candidateVectors, count: 5, lambda: 0.7);
string[] keywords = chosen.Select(i => phrases[i]).ToArray();

Three divergences from keybert itself, all recorded in decision 0005 and in the equivalence table's Mmr.Select row: keybert parameterises diversity = 1 − λ rather than taking λ directly, it rounds its scores to four decimals, and it returns its picks sorted by relevance rather than in selection order — Mmr.Select returns selection order, so a caller matching keybert's output ordering re-sorts by score afterwards. The oracle corpus behind Mmr.Select's parity claim therefore compares the selected set, not the sequence.

See also