Skip to content

3.9.0

Latest

Choose a tag to compare

@vthorsteinsson vthorsteinsson released this 02 Sep 18:41
· 1 commit to master since this release

Changes

  • Wider parse coverage, driven by a corpus study of previously unparsed news sentences: verbs with clitic subjects (ertu, viltu), verb particles before objects (bjó til hús), reflexive verb frames, hið + adjective nominalizations, quoted questions/exclamations with attribution („Hvenær?" spyr Jón), attribution continuations (… segir Jón og bætir við: „…"), sagður/sögð hafa + participle, þess í stað, and date periods (dagana 13. október til 14. nóvember 2017).
  • Deterministic parse-tree selection: ties between equally scored parse alternatives are now broken by grammar order and constituent spans instead of internal enumeration order, so identical input always yields the identical tree across runs, processes and platforms.
  • Grammar hygiene: a usage census over 10,000 parsed sentences drove the removal of redundant and unreachable rules (the compiled grammar shrinks by ~4%), repaired several constructions that could never win a parse, and fixed a typo that had left the wrong-case apposition fallback unreachable.
  • Article 68 of the Icelandic constitution now parses: „Nauðungarvinnu skal engum gert að leysa af hendi" — object topicalization out of the gera e-m að gera e-ð construction.
  • All changes verified against fixed 1,000-sentence regression baselines with no lost parses.