target-stopwords

Per-language function-word lists, induced from that language's own Bible text — frequency + dispersion (the classic corpus-linguistics stopword-induction recipe), then RESCUED against the language's own alignment output + a source-anchored content signal so genuinely frequent CONTENT words ("God", "Lord") are never dropped.

A candidate word is rescued out of the list (judged a real content word, not a function word) only when all four hold — see the lexeme-aligner's target_stopwords.py:

1. it carries at least 25 aligned occurrences — a share read off one or two observations is noise, and function words are systematically under-aligned, so thin aligned mass is itself evidence;
2. its dominant Hebrew/Greek lexeme holds ≥40% of its aligned mass — a true function word instead scatters thinly across dozens or hundreds of distinct lexemes;
3. that lexeme is marked content in the source-side prior pack;
4. that lexeme is not itself semantically light (config/light_lexemes.json). Copulas, have, quantifiers, negators, possessives, modals and the Hebrew nouns grammaticalised into prepositions (panim → "before", yad → "by") are all correctly rendered by target FUNCTION words, so aligning to one proves nothing about content-hood — without this the lists lost is/are, all, one, no, can, before.

Many of the covered languages have no existing curated stopword list anywhere — this is a reusable resource for search, IR, topic modeling, or any NLP task needing one in these languages.

Languages written in scripts without whitespace word separation (Han, Japanese, Myanmar) are segmented by rule — Han per character, Japanese at script boundaries, Myanmar per syllable — with no segmenter model or download, so the method still runs on any language that has a Bible and nothing else.

CC0-1.0 — derived word-frequency statistics, no source text redistributed. See manifest.json for per-language stats + content hashes.