matra
on this page
matra reports the structure of text and measurements over it. It parses a document into a typed tree of sections, paragraphs, sentences, and tokens, measures that tree, and ranks what is in it. Every value it returns is structure or a number over structure. What any of it means for your purpose is your code's decision, which is what makes the output reusable across purposes.
Four tiers of output
- Structure. Sections, paragraphs, sentences, and full CoNLL-U tokens, plus the structural primitives read off each dependency tree: negations, modals, clausal complements, root adverbials, Hearst pairs, and the root clause's mood.
- Measures. Readability, lexical density, and compression ratio per paragraph; counts, sentence length, vocabulary TTR, nominalization ratio, and passive ratio per document. Every document also records where it came from: the parser and its model, and a digest of the analysed text.
- Extraction. Two extractive summarizers (TF-IDF, TextRank) and two keyphrase extractors (RAKE, YAKE).
- Semantic clusters. Sentence embeddings grouped by cosine similarity, behind the
model2vecfeature, carrying the identity of the model that produced them.
End to end
import matra
engine = matra.Engine()
text = (
"The committee approved the proposal without debate. "
"Three amendments were submitted by the working group."
)
doc = engine.analyze(text)
print(doc["metrics"]["passive_ratio"])
print(doc["sections"][0]["paragraphs"][0]["metrics"]["readability_grade"])
The same analysis from Rust, and from the command line:
use matra::domain::Format;
use matra::{Engine, Input};
let engine = Engine::new()?;
let text = "The committee approved the proposal without debate. \
Three amendments were submitted by the working group.";
let doc = engine.analyze(Input::text(text, Format::PlainText))?;
println!("{:?}", doc.metrics.passive_ratio);
$ matra analyze essay.md
essay.md
sentences 35
words 332
mean sentence len 9.5
sentence len sd 7.4
passive ratio 0.057
All three return the same shape: a struct in Rust, a read-only Document that reads as a dict in Python, a table or --json at the command line. The second sentence in the example is passive, and amendments carries the relation nsubj:pass. matra records that and stops there.
Where to go next
Understand the model. Concepts covers the document tree, the four tiers, and what a structural primitive is. Situation model covers what the output is for and what it deliberately withholds.
Write against it. Programming model is the surface: Input, Engine, ports, bounds, and what Python exposes. Pragmatics answers the choices you have to make.
Get it running. Installation, then Rust, Python, or CLI.
Look something up. What matra gives you lists every field and function the pipeline returns.
This page, measured by matra 0.3.0 with english-ewt-ud-2.5-191206: 7 paragraphs, in the margin. the data