amrisi / amr-guidelines

240 stars 86 forks source link

Unification Documentation #151

Open timjogorman opened 8 years ago

timjogorman commented 8 years ago

Hi all, Here's a provisional set of statements of "what's new about these unification frames". I can also provide a more traditional guide for how to read the frame files, if we need one. Please suggest any changes or clarifications you might want.

(This is more rigid than we may need to have when we move into MWEs)

The new Unified (AMR-style) PropBank rolesets:

Prior PropBank frame files, and the rolesets (individual senses) within them, were subtyped by part of speech tags, and gave various senses for a single lemma. Earlier AMR releases utilized the verbal subset of those frames, and had annotators generalize across parts of speech to the nearest verbal sense. The unified PropBank rolesets, in order to add power and coverage to the Abstract Meaning Representation project, essentially formalize that process by lumping etymologically related lemmas with the same sense into the same roleset. These rolesets, in AMR style, are now closer to a representation of an actual concept, and represent the different forms that the roleset might take with a new "alias" element that expresses its possible realizations.

Alias field conventions and assumptions:

nschneid commented 8 years ago

Thanks @timjogorman! Depending on the intended audience, it might make sense to walk through an example of related rolesets and their aliases.