saphes

Readability and lexical diversity — two metrics, done carefully, with the parameters other implementations hardcode.

saphes — σαφής, "clear, plain, distinct". Aristotle makes clarity the chief virtue of λέξις (style); the other classical axis is ποικιλία, variety. The two metrics here are exactly those axes: LIX measures clarity, TTR measures variety.

Why this exists

textstat, textdescriptives, lexicalrichness and taaled already cover this ground. Two reasons to still build it:

1. The LIX long-word threshold is hardcoded at 6 everywhere. That 6 comes from Björnsson's Swedish original, and it is wrong for the languages we work on. Hungarian is agglutinative and Ancient Greek heavily inflected, so at threshold 6 nearly every token counts as "long" and the index saturates into a flat line. Measured over the full Hungarian Webcorpus, 44.5% of running tokens are "long" at threshold 6, against 25.7% in Swedish. On real Hungarian prose that pushes LIX to 60.4 — "very difficult" — where the calibrated threshold gives 43.4. Parameterising the threshold is the whole point.
2. Implementations disagree. They count words, sentences and long words differently, so they rank the same texts differently. So: expose the counts, document every choice, make results auditable.

Non-goal: becoming another kitchen-sink readability library.

Installation

uv add saphes

The core has no dependencies — plain Python and the standard library. Two optional extras stay out of the core path: saphes[snowball] for Hungarian stemming and saphes[punkt] for the NLTK sentence splitter.

Quickstart

from saphes import lix, lexical_diversity

# LIX takes SURFACE FORMS — word length is the signal.
result = lix("The cat sat on it. Complicated sentences generally frighten us.")
result.score                                       # 45.0
result.words, result.sentences, result.long_words  # (10, 2, 4)  -> A, B, C
result.band                                        # 'standard'

# Diversity takes LEMMAS, and `unit` is required — there is no default.
lexical_diversity(lemmas, unit="lemma")

# Comparing texts of different lengths? Use MATTR, not TTR.
lexical_diversity(lemmas, unit="lemma", window=100).mattr

Hungarian needs three things English does not — a retuned threshold, a letter count, and a stand-in for a lemmatiser:

from saphes import hungarian_letter_count, hungarian_stems, lix, recommended_threshold

# `sz` is one letter, not two characters. The threshold is calibrated per policy.
hu = recommended_threshold("hu-letters")          # 8, not Björnsson's Swedish 6
lix(text, length_policy=hungarian_letter_count, long_word_threshold=int(hu))

# No lemmatiser? Snowball is the fallback, declared as its own unit.
lexical_diversity(hungarian_stems(tokens), unit="stem")

The data contract

The two metrics require opposite token streams.

Feed the same list to both and exactly one is silently wrong — no error, no NaN, just a plausible number. unit is required, the parameter names differ, a raw string is refused where it could only be wrong, and every result records what it measured.

saphes consumes lemmas; it does not produce them. Lemmatisation is language-specific and heavy — CLTK or a treebank for Greek, huspacy for Hungarian. The caller lemmatises; saphes measures.

The one exception is optional and honest about itself: saphes[snowball] gives Hungarian Snowball stemming for callers with no lemmatiser. Its output is declared as unit="stem", a third stream with its own name, because a stemmer both over- and under-merges and a stem-based number is comparable only to another from the same stemmer.

Documentation

saphes.readthedocs.io, organised by Diátaxis:

• Tutorial — new here? Measure your first text in about ten minutes, with nothing to download.
• How-to guides — measure Hungarian text, supply a sentence count, count letters rather than characters, stem without a lemmatiser, calibrate a threshold.
• Reference — every function, its contract and its failure modes, plus the calibration data.
• Explanation — why the two metrics need opposite input, why the threshold has to move, and why implementations disagree.

Every code block in the docs is executed by CI, so nothing there can drift.

Roadmap

• LIX with a parameterised long-word threshold
• TTR and MATTR with a required, recorded token unit
• RIX (long words per sentence)
• Empirically calibrated per-language thresholds, from token-weighted word-length distributions — Hungarian ships as recommended_threshold("hu")
• Phonotactically aware Hungarian letter counting — sz is one letter, ssz is two, and morpheme boundaries are handled by rule plus an attested table
• Optional Snowball stemming for callers with no lemmatiser, as a declared third unit
• Mean dependency distance and mean hierarchical distance, from any parser's output
• Zero-dependency adapters for HuSpaCy/spaCy and CoNLL-U/emtsv
• Loan-word ratio (idegenszó-arány) against a caller-supplied lexicon
• A public-domain Hungarian loan-word lexicon to ship with it
• The same study for Ancient Greek, for the Homer project
• POS-filtered diversity, once lemmas carry tags
• MTLD, HD-D, vocd-D, Maas

Research

• Diachronic analysis: test MDD and loan word density changes on the ParlaMonitor corpus across time
• Data diversity benchmarking: benchmark MDD and idegenszó-arány across diverse Hungarian registers (Webcorpus, legal texts, literary prose, social media)
• Multilingual scaling: test the agnostic MDD calculation engine on Universal Dependencies corpora across multiple languages
• Dependency motifs: chain motif identification for advanced cognitive load mapping, extending the MDD/MHD pair (Jing & Liu 2017)

Maintenance

• Logo and README banner
• Mutation-testing baseline

Explicitly out of scope: Flesch, Kincaid, SMOG and relatives. They need syllabification, which is language-specific and a different project. LIX was chosen precisely because it needs only word length and sentence count, so it travels across languages.

Made by

saphes is made by Crow Intelligence.

License

MIT