SYNTAX_STATUS.md

PLAN

This branch changes the pipeline to insert the syntax CG right after disam and capstag (before tagger), which gives us syntax @-tags on all lexical units. Then, after bidix and lexical selection, the new step refsyn.t1x removes the @-tags, but also attends to the last seen @subj and puts it in the ref-field (of the following words). Then t1x reads the ref-field and uses that to compute the gender/number of pp's created from passives.

• See refsyn.t1x which stores cur_subj and places it in the ref field (as well as remove syntactic function tags so tNx doesn't have to deal with them – using syntax tags in transfer will have to be a later extension)

We wait with apertium-anaphora for now. What we want is to put the governed @subj – which is typically the nearest – into for transfer to use. It's syntactic, not anaphoric.

Pretty much done:

• lrx needs to deal with @-tags (typically does, but some rules might end in )

• lsx needs to deal with @-tags

  • Explicit in lsx needs to be turned into (or ) which can skip the function tags; try xmllint --xpath '//l/s[@n="aa"]/../../l/d/../..' apertium-nno-nob.nob-nno.lsx

• passive gender/number now uses the ref field, via refsyn.t1x

• In some cases, we need to use a subject that's to the right, refsyn.t1x needs some rules for that

• Syntax CG now before tagger, using syntax for disambiguation.

• Most lrx/lsx regressions fixed.

• Use the subj→ref method for participles as well as passives. The old method was to "disambiguate" participles based on preceding subject, but that fails when subject changes gender in bidix, since we disambiguate based on nob gender. OTOH it's the target language subject that gets stored in the ref field, which will have the right gender. T1X now uses the ref field if it's set, falling back to the input gender/number (given by the old method) if unset.

• Subjects of relative clauses and subclauses tagged as @xubj

  • de<@subj> nevner mannen<@xubj> i 50-åra som …
  • et antall<@subj> av deres krigere<@xubj> …

• Regular (non-pp) adjectives

  • restricted to the cases where the subject is plural, or it changes in translation between nt and non-nt genders

Main work remaining:

• Remaining regressions in passive genders (missing refsyn.t1x patterns, bad syntax disambiguation?)

• We should not remove marked plurals! Our tagging doesn't show if e.g. utelatt was the ambiguous form "utelatt" or the plural-only "utelatte", so T1X might turn that into mf.sg.ind to match the @subj, but if it actually was unambiguously plural in nob we should keep it that way (e.g. when adj is used as noun in "rapportering av utelatte").

• Sometimes the subject is a whole clause – should give nt:

  • [At disse ble solgt] er fint<nt!>
  • [Hva de mistenkte skal ha gjort], er ikke kjent<nt!>. May require "outer" ref, flip on clause boundary. But we can't always trust that som+comma ends the clause:
  • Han stadfestar at alle tenestemennene som avfyrte skot, no er avhøyrde

• Coordination should give plural

  • To politivakter og en vakt fra et privat vaktselskap ble drept<pl!>
  • Et fransk kjærestepar sier til VG at det var de som fant norske Maren Ueland og danske Louisa Vesterager Jespersen drept i Marokko mandag morgen. But sometimes hard to know if it's the same person (sg) or different (pl):
  • Programleder og tidligere gjengleder skutt og drept i København.

• Missing relative pronouns are difficult:

  • Jeg er her på grunn av beslutninger tatt på europeiske møter
  • Brasil er hardt truffet med vel 55.000 dødsfall relatert til pandemien
  • … stemte med mannens fingeravtrykk funnet hos … Difficult, since e.g.: ble det ved flere anledninger tatt opp i Stortinget
  • Maybe we can use ngram frequencies f(ta beslutninger) >>> f(ta anledninger)
  • Current solution: We treat sequences n.ind pp.@o-pred as agreeing. So a participle tagged as object predicate will agree in this manner, but this fails in e.g. «med dødsfall relatert til» where «dødsfall» is tagged @←p-utfyll and so «relatert» is just tagged @adv (maybe syn.rlx could have an @a-pred, but this is in general a difficult attachment problem)

• adj.sg.nt.@adv should be adv_movable

  • domineres sterkt → er sterkt dominert

• Crossing subject are difficult:

  1. Samtidig skal vi ha respekt for den politiske plattformen som de fire partiene fremforhandlet på Granavollen, og som ble godkjent av partienes organer
  2. Antallet mennesker som døde i etterkant, var kraftig underrapportert.

  We could use commas as a signal to switch to a previous subject, which would fix the above, but but commas after "non-optional" relatives should not lead to a switch:

  3. Han opplyser at alle flyvinger som var planlagt tirsdag, er innstilt.
  4. Regjeringa stadfestar no at lisenskontoret i Mo i Rana, som har 106 arbeidsplassar, legges ned.
  5. Tidligere i februar ble det kjent at seks politifolk som jobbet for å avdekke narkotikakriminalitet, var pågrepet for selv å ha smuglet narkotika.

  (Though in those cases, we have «at.cnjcoo.clb» on which to empty prev-subj.) But it's hard to consistently tag these commas as clb; e.g. below we have an "optional" relative sentence where we do want to switch subjects on the comma:

  6. Den ene såkalte svarte boksen fra flyet som styrtet i Etiopia, er funnet, melder etiopiske medier.

  (Note: In example 1. we have xubj-subj-ref:subj-ref:xubj, whereas in 6. we have subj-xubj-ref:xubj-ref:subj, ie. we can't always expect the relative subject to be the inner one.)

  Possible idea: split @clb-tag into clause-with-subject vs clause-without-subject (ellipsis)?

• Uncommon, but referent may be far to the right:

  • Desto lenger ut i kampen vi er, desto mer sliten blir muskulaturen.
  • Mange har reist langveisfra, spesielt populær kan det virke som om Foyle er i Australia.