A few pists for discussions

​Over the years, FRMG has shown the validity of the approach of meta-grammars, being

  • a large coverage grammatical resource for French, while staying manageable and extendable (sometimes difficult to tell the truth!)
  • competitive with statistical parsers in terms of accuracy
  • efficient enough to be used on large corpora (such as French Wikipedia) for large scale applications (information extraction, terminology extractions, knowledge acquisition, ...)

FRMG has also shown the possibility to hybrid symbolic approaches with data-driven ones by

  • using annotated tree-banks to tune its disambiguation phase
  • stacking with DyALog-SR, a statistical transition-based parser (with best results for French in 2014)

Several aspects of meta-grammars would still be interesting to discuss:

  • even better ways to develop and maintain a large coverage meta-grammar
  • how to start a meta-grammar for a new language
  • re-using meta-grammar components (hierarchy, classes)
  • exploring new target formalisms and/or richer ones (semantic info)
  • exploring meta-grammar extraction or parametrizing from a Tree Bank

Also:

  • using collaborative approaches (such as wiki) for designing linguistic resources
    (mgwiki used for Sequoia Annotation Guidelines) UD guidelines