Methodology

on this page
grade13.2density0.42compression0.56

What each number matra reports is computed from, how to cite the work it comes from, and what it does not license you to conclude. Every formula on this page is transcribed from the implementation, not from the publication it descends from; where matra departs from the published method, the departure is stated.

grade10.4density0.61compressionnone

Related pages: Domain types for the fields these values live in, Errors for the caps each algorithm enforces.

The parse layer

Model identity

gradenonedensity0.57compressionnone

The English model is pinned in src/nlp/udpipe.rs.

PropertyValue
Fileenglish-ewt-ud-2.5-191206.udpipe
Size16,309,608 bytes
SHA-256784bd0fa85e3d831fd02a55290d0acfd05c953159dc38cc33d52e1b28add9957
TreebankEnglish Web Treebank, Universal Dependencies 2.5, release 191206
DistributorLINDAT/CLARIAH-CZ repository, handle 11234/1-3131
grade11.2density0.49compression0.51

The engine's constructors, Engine::new and Engine::from_config, download the file into the model directory when it is absent, checks the size, computes the SHA-256, and loads the same bytes it hashed. There is no second read from disk between verification and load. A cached file that fails verification is not loaded: the model is downloaded, and a download that also fails the digest is retried once, so a call makes at most two downloads. The file is replaced only by bytes that verify. If both downloads fail the digest the call returns Error::ModelInvalid, nothing is loaded, and the file that was there is still there.

grade13.2density0.62compressionnone

That path is the only way a parsing model arrives: no surface loads another model file, so every result names the pinned model it came from.

grade12.1density0.64compressionnone

Changing the model means editing the pinned constants, which makes the change visible in version control and in any caller that pins a matra version.

What the parse produces

grade13.4density0.70compressionnone

One Token per token with all ten CoNLL-U columns, grouped into sentences. Two properties of the shipped adapter matter when you report results:

  • deps (CoNLL-U column 9) is always _. The udpipe-rs binding does not surface enhanced dependencies.
  • Sentence::text is built from token surface forms, with a space inserted unless the preceding token carries SpaceAfter=No in misc. It is a reconstruction, not a substring of your input.

Which text is parsed

grade9.1density0.55compressionnone

The route through the library changes what gets segmented into sentences.

grade11.7density0.57compression0.57

Every pipeline route parses each paragraph that is not a quote separately. Quotes are not parsed, so they contribute nothing to any metric. Calling a provider's parse directly parses the whole string in one call with no decomposition, and its segmentation can differ from the pipeline's at paragraph boundaries.

grade7.6density0.50compressionnone

Input::file picks the format from the file's bytes and then its extension, so the same text segments differently under .md and under .txt; Input::text takes the format as an argument.

grade11.4density0.66compressionnone

Paragraph-at-a-time parsing means a sentence can never be attributed to the wrong paragraph. It also means sentence segmentation cannot run across a paragraph boundary. Report which format your numbers were produced under.

Citation

grade5.3density0.66compressionnone

Straka, M., & Straková, J. (2017). Tokenizing, POS Tagging, Lemmatizing and Parsing UD 2.0 with UDPipe. Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, 88-99.

grade5.1density0.76compressionnone

Silveira, N., Dozat, T., de Marneffe, M.-C., Bowman, S., Connor, M., Bauer, J., & Manning, C.D. (2014). A Gold Standard Dependency Corpus for English. Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC 2014).

grade11.7density0.79compressionnone

Zeman, D., et al. (2019). Universal Dependencies 2.5. LINDAT/CLARIAH-CZ digital library, Institute of Formal and Applied Linguistics, Charles University.

Limitations

grade10.8density0.60compressionnone

UDPipe is a statistical syntactic parser. It performs no named entity recognition, no coreference resolution, and no semantic role labeling. Tags and relations are predictions, not ground truth, and error rates on text unlike the training corpus are higher than on text like it. The shipped configuration is English only.

Paragraph metrics

grade9.7density0.61compressionnone

Two different counts appear below, and confusing them changes the answer. Paragraph::word_count counts non-punctuation tokens from the parse, and every applicability gate is stated in those terms. The formulas themselves split the paragraph text on whitespace. A gate stated in word_count terms is not a gate on whitespace words.

Flesch-Kincaid grade level

grade = 0.39 * (words / sentences) + 11.8 * (syllables / words) - 15.59
InputAs computed
wordsCount of whitespace-separated pieces of the paragraph text
sentencesCount of ., !, and ? characters in the paragraph text, floored at 1
syllablesSum over words of the syllable estimate below. A word with no alphabetic characters contributes 0
grade22.6density0.54compression0.59

The syllable estimate for one word: keep only alphabetic characters, lowercase them; if the result is three bytes or shorter, return 1; otherwise count runs of the letters a, e, i, o, u, y, where adjacent vowels count once; subtract 1 if the word ends in e and the count is above 1; return at least 1.

grade10.0density0.50compression0.57

Applied to paragraphs that are not quotes and whose word_count is above 10. Stored in ParagraphMetrics::readability_grade, the paragraph's metrics.readability_grade. The value is unbounded and can be negative on short, simple text. The formula is computed inside the metric suite only; matra publishes no function that computes it on a string.

grade7.2density0.82compressionnone

Citation. Kincaid, J.P., Fishburne, R.P., Rogers, R.L., & Chisholm, B.S. (1975). Derivation of New Readability Formulas for Navy Enlisted Personnel. Research Branch Report 8-75, Naval Technical Training Command.

grade14.3density0.66compression0.53

Limitations. The formula was fitted to Navy training material in 1975 and predicts nothing about conceptual difficulty, argument quality, or required background knowledge. Two paragraphs with the same syllable and length profile score the same whatever they assert. The original publication specifies no syllable-counting algorithm, so grade values from different tools are not comparable; the heuristic above miscounts many words, and non-ASCII letters register as consonants. Counting sentence-ending characters treats abbreviations, decimals, and ellipses as sentence boundaries.

Lexical density

lexical_density = content_words / total_words
InputAs computed
total_wordsCount of whitespace-separated pieces of the paragraph text
content_wordsThose pieces whose alphabetic characters, lowercased, form a non-empty string that is not in the stop word list
grade10.7density0.47compressionnone

A piece made only of punctuation counts in the denominator and never in the numerator.

grade7.7density0.57compressionnone

Applied to paragraphs that are not quotes, whose word_count is not zero, and whose text yields at least one whitespace-separated piece. Stored in ParagraphMetrics::lexical_density. Range 0.0 to 1.0.

grade8.3density0.77compressionnone

Citation. Ure, J. (1971). Lexical density and register differentiation. In G. Perren & J.L.M. Trim (Eds.), Applications of Linguistics. Cambridge University Press.

grade8.8density0.64compression0.60

Limitations. The value depends entirely on the stop word list, which is matra's own list of 105 English words, not a standard one. Ure's definition counts lexical items by part of speech; matra approximates that with a stop word list over whitespace pieces. Density says nothing about whether the content words are used well, or at all correctly.

Compression ratio

compression_ratio = compressed_bytes / original_bytes
InputAs computed
original_bytesUTF-8 byte length of the paragraph text
compressed_bytesByte length of the brotli-compressed text at quality 6 with a window parameter of 18, which is a 256 KiB window
grade7.8density0.58compression0.59

Applied to paragraphs that are not quotes, whose word_count is above 50, and whose text is at most 262,144 bytes. Paragraphs above that byte ceiling are skipped and keep None, which bounds worst-case compression time. The slot also stays None if the encoder fails to consume the text. Stored in ParagraphMetrics::compression_ratio. A lower ratio means the text compressed further, which means more surface repetition.

grade7.7density0.82compressionnone

Citation. Alakuijala, J., & Szabadka, Z. (2016). Brotli Compressed Data Format. RFC 7932, Internet Engineering Task Force.

grade12.7density0.64compression0.58

Limitations. This is a redundancy proxy over bytes, not a measure of meaning, quality, or novelty. Precise technical prose that reuses terminology compresses much like repetitive filler. The number is a property of one encoder at two specific parameter settings: a different compressor, or a different version of this one, gives different values.

grade12.0density0.63compressionnone

The three paragraph measures, computed for every paragraph of one longer text, The Federalist No. 10:

Paragraph measures across one document

Mean readability
18.20
Vocabulary TTR
0.25
Nominalization ratio
0.04
Passive ratio
0.42
Paragraph measures across the documentReadability gradeFlesch-Kincaid; higher reads harder010203040mean 18.2Lexical densitycontent words over all words0.300.400.500.60Compression ratiocompressed over original bytes; lower repeats more0.300.400.500.600.701357911131517192123
show the data
#Opens withWordsGradeDensityCompression
1AMONG the numerous advantages promised by a well…38420.720.470.46
2By a faction, I understand a number of…5124.890.430.56
3There are two methods of curing the mischiefs…2210.690.45none
4There are again two methods of removing the…4018.890.40none
5It could never be more truly said than…7912.900.460.52
6The second expedient is as impracticable as the…15714.590.440.45
7The latent causes of faction are thus sown…26318.680.530.42
8No man is allowed to be a judge…26712.920.470.46
9It is in vain to say that enlightened…8013.540.470.55
10The inference to which we are brought is,…3212.870.35none
11If a faction consists of less than a…17216.100.430.45
12By what means is this object attainable? Evidently…12113.740.470.50
13From this view of the subject it may…17921.570.450.48
14A republic, by which I mean a government…6113.910.440.58
15The two great points of difference between a…5023.030.48none
16The effect of the first difference is, on…17520.220.490.48
17In the first place, it is to be…12326.530.450.42
18In the next place, as each representative will…7632.060.420.55
19It must be confessed that in this, as…10317.280.460.52
20The other point of difference is, the greater…21624.700.440.45
21Hence, it clearly appears, that the same advantage…16914.140.490.48
22The influence of factious leaders may kindle a…12819.860.480.51
23In the extent and proper structure of the…5214.750.480.54

One panel per paragraph measure, paragraphs in document order along the bottom. A cross marks a paragraph matra declined to measure: the field is None because the paragraph falls outside the measure's applicability conditions above, and the line breaks there rather than inventing a value. The dashed line is the document's mean readability: the mean of the grades of the paragraphs matra measured. matra computes no document-level lexical density or compression ratio, so those panels have no line.

Input: The Federalist No. 10, the essay body, James Madison (as Publius). Public domain (first published 1787). Produced by matra 0.3.0 with the UDPipe model english-ewt-ud-2.5-191206 (SHA-256 784bd0fa85e3). Data: federalist-10/metrics.json.

Document metrics

grade12.8density0.53compressionnone

Every field of DocumentMetrics is computed over the sentences attached to the document's paragraphs, which excludes quotes. There is exactly one sentence set, and it is derivable from the tree in one way, so the numbers cannot disagree with the structure they describe.

Vocabulary type-token ratio

vocabulary_ttr = distinct_lemmas / total_lemmas
InputAs computed
total_lemmasCount of non-punctuation tokens across the sentence slice
distinct_lemmasCount of distinct lemma strings in that same set, compared exactly, without case folding
grade10.2density0.61compressionnone

Applied to the whole slice. Stored in DocumentMetrics::vocabulary_ttr, which stays None when the slice contains no non-punctuation token.

grade6.3density0.63compressionnone

Citation. Johnson, W. (1944). Studies in language behavior: A program of research. Psychological Monographs, 56(2), 1-15.

grade13.0density0.74compressionnone

Limitations. Type-token ratio falls as documents get longer, because repetition accumulates. Values from documents of different lengths are not comparable without a length correction, and matra applies none. Because lemmas are compared exactly, a provider that emits case-varying lemmas inflates the distinct count.

Nominalization ratio

nominalization_ratio = nominalizing_nouns / total_lemmas
InputAs computed
total_lemmasThe same denominator as vocabulary type-token ratio: non-punctuation tokens across the slice
nominalizing_nounsTokens whose pos is NOUN and whose text, lowercased, ends in tion, ment, ness, ity, ence, or ance
grade12.6density0.58compressionnone

The suffix test reads the surface form, not the lemma, so a plural nominalization such as conditions or measurements is missed: the plural -s falls after the suffix, so "conditions".ends_with("tion") is false even though the singular condition matches.

grade12.2density0.61compressionnone

Applied to the whole slice. Stored in DocumentMetrics::nominalization_ratio, which stays None when the slice contains no non-punctuation token.

grade10.4density0.57compression0.55

Limitations. Six suffixes are a heuristic, not a morphological analysis. Words that end in those letters without being nominalizations count as false positives, and nominalizations formed otherwise, such as growth or failure, are missed. The test depends on the POS tagger having assigned NOUN, so tagging errors propagate. The denominator counts every running word rather than every noun, so the value is a share of the text, not a share of the nouns in it.

Passive ratio

passive_ratio = passive_sentences / total_sentences
grade10.5density0.56compressionnone

A sentence counts as passive when any of its tokens carries a dep of nsubj:pass, nsubjpass, or aux:pass. The first two are the Universal Dependencies and older Stanford spellings of a passive subject; the third is a passive auxiliary.

grade11.8density0.48compressionnone

Stored in DocumentMetrics::passive_ratio when the document is measured, and per sentence in Sentence::passive, which Sentence::new sets; both cross to Python as fields. The field is None when there are no sentences.

grade5.4density0.78compressionnone

Citation. de Marneffe, M.-C., Manning, C.D., Nivre, J., & Zeman, D. (2021). Universal Dependencies. Computational Linguistics, 47(2), 255-308.

grade8.3density0.61compressionnone

Limitations. This counts sentences that contain a passive construction, not passive clauses, so a sentence with three passives counts once. Detection is only as good as the parser's relation labels. Nothing here judges whether a passive is appropriate.

Sentence length statistics

grade14.5density0.56compression0.49

DocumentMetrics::mean_sentence_length is total_words / total_sentences, where total_words counts non-punctuation tokens. DocumentMetrics::sentence_length_std is the sample standard deviation of per-sentence non-punctuation token counts, with denominator n minus 1. Both are None when undefined: no sentences for the mean, fewer than two for the deviation. Measuring the document stores both, with the counts total_words, total_sentences and paragraph_count, so all of them cross to Python.

Corpus aggregates

grade11.6density0.52compression0.50

matra computes none. Every measure above is of one document, and how to combine documents is a decision about what the combination is for, so it is the caller's. The counts sum across documents. A ratio pools by summing its numerator and denominator rather than averaging the ratios, so the passive ratio of a set is its passive sentences over its sentences. vocabulary_ttr does not pool at all: it falls as text grows, so documents of different lengths are not comparable on it, alone or summed. A mean of paragraph readability grades averages paragraphs, not documents, so a long document contributes more terms than a short one; leave out the paragraphs whose slot is None.

Summarization

grade13.6density0.66compressionnone

Both summarizers take a Document and a count, score every sentence, keep the highest N, and return them in ascending index, the sentence's position among the document's sentences. Neither generates text: the output sentences are the input sentences.

Shared inputs

grade10.0density0.63compression0.56

A term is a token's lemma with punctuation tokens removed and stop words removed. The stop word test lowercases the lemma before looking it up; the term itself is the lemma as the provider produced it, without case folding. The stop word list holds 105 English words and is shared with the paragraph lexical density metric.

TF-IDF

grade11.2density0.55compressionnone

Each sentence is treated as a document when computing document frequency.

score(s) = ( sum over distinct terms t in s of tf(t, s) * idf(t) ) / distinct_terms(s)

tf(t, s)  = occurrences of t in s / total terms in s
idf(t)    = ln( total_sentences / df(t) )
df(t)     = number of sentences containing t
grade4.9density0.49compressionnone

A sentence with no terms scores 0.0. The mean divides by the number of distinct terms, not by the number of term occurrences. A term present in every sentence has an idf of 0 and contributes nothing.

grade5.1density0.45compressionnone

Cap. 2,000 sentences. Above it, Error::InputTooLarge with what set to "tfidf".

grade7.4density0.61compression0.60

Citation. Luhn, H.P. (1958). The Automatic Creation of Literature Abstracts. IBM Journal of Research and Development, 2(2), 159-165. Spärck Jones, K. (1972). A statistical interpretation of term specificity and its application in retrieval. Journal of Documentation, 28(1), 11-21. Salton, G., & Buckley, C. (1988). Term-weighting approaches in automatic text retrieval. Information Processing and Management, 24(5), 513-523.

grade11.7density0.59compression0.54

Limitations. Treating sentences as documents is an adaptation of a corpus-level weighting scheme, and the resulting weights are not the weights you would get from a real document collection. Scores measure term rarity within one text, not informativeness. Selected sentences are not guaranteed to read coherently together, and nothing checks for redundancy between them.

TextRank

similarity(a, b) = shared(a, b) / ( ln(1 + |a|) + ln(1 + |b|) )

shared(a, b) = sum over terms t in both of min(count_a(t), count_b(t))
|a|          = number of distinct terms in a
grade7.6density0.50compressionnone

Similarity is 0 when either sentence has no terms or they share none. The matrix is symmetric and its diagonal stays 0, so a sentence never reinforces itself.

score(i) at step k+1 = (1 - d) / N
                     + d * sum over j not equal to i of
                       ( similarity(j, i) / out_sum(j) ) * score(j) at step k

d          = 0.85
N          = number of sentences
out_sum(j) = sum of row j of the similarity matrix
grade4.4density0.55compressionnone

All scores start at 1 / N. Iteration stops after 50 passes, or earlier when the largest single-sentence change in a pass falls below 1e-6. Sentences whose out_sum is 0 send no weight.

grade5.9density0.52compressionnone

Cap. 2,000 sentences, which bounds the dense similarity matrix at roughly 32 MB of f64. Above it, Error::InputTooLarge with what set to "textrank".

grade5.4density0.69compression0.64

Citation. Mihalcea, R., & Tarau, P. (2004). TextRank: Bringing Order into Text. Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, 404-411. Page, L., Brin, S., Motwani, R., & Winograd, T. (1999). The PageRank Citation Ranking: Bringing Order to the Web. Stanford InfoLab Technical Report 1999-66.

grade11.0density0.56compression0.59

Limitations. Similarity here is lemma overlap. Two sentences expressing one idea in different words have similarity 0. The normalization by the logarithm of distinct-term counts is matra's, not the one in the 2004 paper, so scores are not comparable with other TextRank implementations. Reaching the iteration ceiling without converging is possible and is not reported.

Keyphrase extraction

grade9.2density0.66compressionnone

Both extractors take &[Sentence] and a maximum count, and return Keyphrase values in descending score order. Both cap input at 200,000 tokens counted across the slice with punctuation included.

RAKE

grade6.6density0.64compressionnone

Candidates. Within each sentence, a candidate is a maximal run of tokens whose pos is NOUN, ADJ, or PROPN and whose lowercased lemma is not a stop word. Punctuation, stop words, and any other part of speech end the run. Candidate words are lowercased lemmas.

for each candidate phrase p of length k, for each word w in p:
    freq(w)   += 1
    degree(w) += k

word_score(w)     = degree(w) / freq(w)
phrase_score(p)   = sum over w in p of word_score(w)
grade5.8density0.48compressionnone

When the same phrase string occurs more than once, the highest score is kept. Phrases are then sorted by score descending and truncated to the requested count.

grade4.1density0.45compressionnone

Cap. 200,000 tokens. Above it, Error::InputTooLarge with what set to "rake".

grade4.7density0.79compressionnone

Citation. Rose, S., Engel, D., Cramer, N., & Cowley, W. (2010). Automatic Keyword Extraction from Individual Documents. In M.W. Berry & J. Kogan (Eds.), Text Mining: Applications and Theory. John Wiley and Sons.

grade11.9density0.62compression0.57

Limitations. Published RAKE delimits candidates with stop words and punctuation over surface words. matra adds a part-of-speech filter and works over lemmas, so scores and candidate sets differ from a reference implementation. A high degree over frequency ratio means a word appears in long phrases relative to how often it appears at all; that is a structural property, not evidence of topical centrality.

YAKE

grade11.9density0.70compressionnone

Terms. Non-punctuation tokens whose lowercased lemma is not a stop word and is longer than one character. Position is a running index over every non-punctuation token, so stop words and single-character tokens occupy positions without becoming terms.

pos_score(t)         = ln( 1 + mean_position(t) / total_positions )
freq_score(t)        = occurrences(t) / total_positions
context_diversity(t) = distinct_neighbors(t) / neighbor_observations(t)

term_score(t) = ( pos_score(t) + context_diversity(t) ) / ( freq_score(t) + 1.0 )
grade8.4density0.51compression0.54

total_positions is the count of all non-punctuation tokens across the slice, floored at 1. A neighbor is the adjacent non-punctuation token on either side within the same sentence, recorded only when it is not a stop word; a term with no recorded neighbor gets a context diversity of 0.0. Lower term_score means more relevant at this stage.

ngram_score(p) = product over words w in p of term_score(w)
output_score(p) = 1.0 / ngram_score(p)
grade9.3density0.53compression0.57

Candidates are sliding windows of 1, 2, and 3 words over each sentence's term list, with stop words and single-character tokens already removed, so an n-gram can join words that were not adjacent in the source. A sentence with fewer than three terms yields correspondingly shorter windows. When a phrase string occurs more than once, the lowest ngram_score is kept.

grade10.9density0.47compression0.58

Candidates whose score is not strictly positive, or is not finite, are dropped before inversion. A term reaches a score of exactly 0 when it occurs only at position 0 and has no eligible neighbor, and any n-gram containing it is dropped with it. The remainder are sorted by output_score descending, so higher is more relevant in the returned Keyphrase.

grade4.1density0.45compressionnone

Cap. 200,000 tokens. Above it, Error::InputTooLarge with what set to "yake".

grade5.4density0.72compressionnone

Citation. Campos, R., Mangaravite, V., Pasquali, A., Jorge, A., Nunes, C., & Jatowt, A. (2020). YAKE! Keyword extraction from single documents using multiple local features. Information Sciences, 509, 257-289.

grade13.4density0.61compression0.53

Limitations. Published YAKE combines five features, including casing and sentence spread, and applies deduplication over similar candidates. matra implements three features, no deduplication, and a different combination formula. Scores are not comparable with the reference implementation, and the two will not agree on rankings. The inversion means output scores are unbounded above and have no interpretable unit; use them for ordering, not as magnitudes.

Determinism and reproducibility

What is fixed

grade7.4density0.54compression0.62

matra draws no random numbers, sets no seeds, samples nothing, and runs the analysis path on one thread. The model is pinned by SHA-256, and the bytes that pass verification are the bytes that get loaded. Given the same model file and the same input, the parse is the same.

grade18.2density0.68compression0.58

Given the same parse, these are reproducible bit for bit on one machine: every paragraph metric, every document metric, every derived method on Document, TextRank scores and selection, RAKE word and phrase scores, YAKE term and n-gram scores, and the order of keyphrases with equal scores. TextRank in particular is order-stable because its shared-term counts are integer sums and its iteration walks sentence indices in order.

What is not bit-stable

gradenonedensity0.80compressionnone

One place carries hash map iteration order into the result.

  1. TF-IDF sentence scores. The per-sentence score sums over a hash map whose iteration order is not fixed between runs. Floating-point addition is not associative, so the low bits of a score can differ between runs.

    Selection changes only when two sentences differ by less than that rounding difference.

grade20.6density0.52compression0.61

RAKE and YAKE collect their candidates in the order each phrase is first met and sort them by score with a stable sort, so phrases with exactly equal scores come out in the same order on every run, and a tie straddling the truncation point is decided the same way each time.

grade8.8density0.67compressionnone

Two further sources of variation are worth recording in a methods section.

  • Natural logarithms come from the platform's math library. Values can differ by an ulp across operating systems and toolchains, so compare cross-platform results with a tolerance rather than for equality.
  • The compression ratio depends on the brotli encoder. A change in the encoder version can change the value for unchanged text.

What to record

grade16.3density0.58compression0.57

To let someone reproduce a result, report: the matra version, the model file name and its SHA-256, which format the document was analyzed under, the method, and the parameter values it ran with, such as the requested sentence count or maximum phrase count. A summary, keyphrases or clusters carry every one of these in their provenance: matra, parser, format, method and parameters.

matra <crate version> (https://github.com/mox-labs/matra)
parse: UDPipe, english-ewt-ud-2.5-191206.udpipe
       SHA-256 784bd0fa85e3d831fd02a55290d0acfd05c953159dc38cc33d52e1b28add9957
format: markdown
method: tfidf
parameters: n = 5
grade15.6density0.57compressionnone

Cite the algorithm publications listed above for the methods themselves, and the UDPipe and Universal Dependencies references for the parse layer.

grade9.6density0.49compression0.47

Cite matra separately from the methods. The CITATION.cff file at the repository root carries the software entry, and it identifies the implementation and its version. The publications identify the method. Because matra departs from several published methods in the ways recorded on this page, a reader who has only the method citation does not know which values you produced, and a reader who has only the software citation does not know what the values are meant to measure. Both are needed.

General limitations

gradenonedensity0.50compressionnone

These hold for every number on this page.

  • matra reports structure and stops. It measures; your application decides what a measurement means.
  • No output is a judgment of quality, correctness, originality, or authorship. Nothing here detects machine-generated text.
  • No output carries meaning. Passive ratio counts a syntactic pattern, compression ratio counts bytes, keyphrase scores count co-occurrence. None of them read.
  • Metrics that share a name with published measures are matra's implementations of them, with the departures listed per metric. Do not compare matra's values with another tool's values for a like-named metric.
  • Every value downstream of the parse inherits the parser's errors. On text unlike the English Web Treebank, that inheritance is larger.

This page, measured by matra 0.3.0 with english-ewt-ud-2.5-191206: 75 paragraphs, in the margin. the data