Submodular Selection of Orthographic Correspondences Bounds the Script Component of the Sorani-Kurmanji Lexical Divide

Abstract

Background. Central Kurdish and Northern Kurdish are written here in different alphabets, so any comparison of their vocabularies depends on a conversion step that is a choice.

Methods. Correspondence rules are chosen after transliteration to maximise covered token mass under a rule budget: a monotone submodular problem, NP-hard in general, solved greedily and certified against the exact optimum.

Results. Transliteration alone bridges 0.586 of token mass onto forms attested at least five times; twenty rules raise it to 0.653, the ground set’s best twenty. Transliteration returns 0.311 against a Zazaki outgroup and 0.953 against the other half of the source corpus. On article titles held out as pairs the sixty-rule selection bridges 167 tokens, against 28.0 for jumbled-target rules.

Conclusions. Kurdish lexical overlap is not a well-defined single figure but a curve and a position on it; whether the varieties are one language is not decided here.