26.09.2013

present:

• Sjur
• Linda

grammar checker project plan

0 intro

• working definition: errors that cannot be resolved by the spellchecker
• Excluding real word errors by default

1 done until now:

• error type classification
  • lexical errors (&lex-majuscule)
  • morphosyntactic errors (&msyn-inf_not_actio)
  • syntactic errors (&syn-case_congruence)
  • real-word errors (&real-vuosttaš)
  • correct tags (&corr-not-compound)
• additional error types
  • punctuation errors
  • number formatting errors

  • capitalisation errors

  • specific syntactic grammar for the grammar checker philosophy: sme-gramdis.rle, rules are marked (REMOVE:GramPo)
  • grammarchecker grammar: sme-gramchk.rle
  • publication: Constraint Grammar based Correction of Grammatical Errors for North Sámi LREC 2012

2 todo:

• practical things:
  • move SME (and GC) from old to new infrastructure
  • meetings with Francis
• maintenance:
  • add/change/update semantic/syntactic tags
• work on things started:
  • Duommá’s 250 word list (compounds that lead to real word errors) - excluding real word errors by default
  • rules for valency example sentences collected in gramchkcorpus.txt
• errors:
  • find out which types of errors are most frequent
  • error corpus - size?? other sources??
    • $GTFREE/goldstandard/orig/sme (xserve)
    • main/gt/sme/src/gramchk/gramchkcorpus.txt

• possible classes?

• presentation:
  • sponsor-demonstrations
  • release early/often (Open Source principles)
  • we cannot make a Microsoft Office grammar checker - prohibited by MS - users can protest by writing to them ;) (we can only deliver to LibreOffice)
  • look at a graphic grammarchecker (voikko - Finnish)
  • http://wiki.apertium.org/wiki/Spellchecking
• rules:
  • for real word errors: which semantic tags can be combined? - dálkkádat + rap + poarta
  • bigrams and statistics for compounds?
  • fix/annotate grammatical errors (compounds) already in preprocessing/tokenization/morphological analysis (i.e. treat space as compound border for relevant POS’s) (other ideas - Eckhard?)
  • hfst-proc må truleg oppdaterast for å gje alle analyser av potensielle samansetjingsfeil

Samansetjingsfeil - særskriving:

[N Nom]         [N ...] ===== kasusfeil (Gen not Nom) / sammensettingsfeil
[N Nom/N Gen]   [N ...] =====
[N Gen]         [N ...] =====
[N Nom+VR]      [N ...] ===== med vokalreduksjon (VR) - alltid feil
[N Nom/N Gen+VR][N ...] ===== --"--
[N Gen+VR]      [N ...] ===== --"--

VR = Vokalreduksjon

what is one word?

• stavekontroll - space before and after
• tokenizer:
  • space as a possible sign in a compound (in the case: [[N Nom] [[N …] the error tag can get annotated right away)
  • CG needs to clean up - disambiguate
• tools to be used:
  • dependencies
  • valencies
  • semantic roles

  • semantic prototypes

  • evaluation:
    • precision and recall
    • how much has been resolved