mirror of
https://github.com/GrammaticalFramework/gf-rgl.git
synced 2026-08-28 06:59:51 -06:00
aarneranta
344342012d
declensionNounForms only tested hardConsonant (d t g h k n r), so the neutral consonants (b f l m p s v), which take the hard endings, matched nothing and fell through to the soft declSTROJ fallback: graf, atom, profil, přístup were all declined as if soft. Same gap in guessNounForms. Add hardishConsonant (hard + neutral + the foreign z, x) and use it in both. The genitive passed to the three-argument mkN was only used to pick a paradigm, then thrown away, so the oblique stem was re-derived from the nominative by dropFleetingE. That guesses wrong in both directions: it fires on člen -> člnu, and fails to fire on uzel -> uzelu, doplněk -> doplněku, výpočet -> výpočetu. Give declHRAD, declHRADA and declMUZ stem-taking variants and let declensionNounForms pass the stem it was given. The one-argument guessNounForms keeps the old dropFleetingE behaviour. New paradigms, all previously falling back to declSTROJ: declLATINUS algoritmus - algoritmu, kosmos - kosmu (also -os) declLATINUSA the same, masculine animate declLATINUM kontinuum - kontinua declGREEKMA schéma - schématu declHRADA les - lesa, zákon - zákona declADJF proměnná - proměnné declADJM nultý - nultého declINVAR indeclinable loans: bombé, tamari, software Also match feminine consonant stems by genitive (-i -> kost, -e/-ě -> píseň), masculine -e/-ě genitives (kužel - kužele, král - krále), and neuter -ě (těžiště), none of which were reachable before. Measured on the 1177 nouns of informath's WikidataWordsCze: 225 hit the fallback before, 1 after (the acronym ADE). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>