nx_keyness.nx

nx_keyness.nx -- substrate-native keyword extraction. "Keyness" answers the canonical corpus-linguistics question: what words are distinctively COMMON in target corpus vs reference? Standard metric is Dunning's log-likelihood applied to a 2x2 contingency table comparing token frequencies between corpora. This is identical in shape to the collocation log-likelihood -- the difference is the meaning of the cells: Collocation: word_a vs word_b in same corpus, n_ab co-occurrences Keyness: same word in target vs reference corpus Composes: nx_collocation.nx -- log-likelihood already there; we reuse it Sealed verdict bands (Sketch Engine convention, see Kilgarriff 2009): POSITIVE_KEYWORD_STRONG -- strongly over-represented in target POSITIVE_KEYWORD_WEAK -- modestly over-represented NEUTRAL -- about the same in both NEGATIVE_KEYWORD -- under-represented in target genealogy_id: dunning_1993_log_likelihood + scott_1997_wordsmith_keywords + kilgarriff_2009_simple_maths lineage_id: keyness_extraction_v1