PropBank-based SRL with RRG-based macro (core) roles and human-readable labels
Idea is to develop a vocabulary that builds on FrameNet, PropBank, NomBank, VerbNet and PrepWiki, but one that is closer to syntax , that uses a smaller and interpretable inventory of roles and a clear-cut (i.e., syntactically defined) distinction between core roles and modifiers.
- designed primarily for accusative languages
- the replacement for core roles are RRG ACTOR (
ARG0: proto-subject), UNDERGOER (ARG1: proto-object) and THEME (ARG2: third argument, indirect object) - all other roles are identified by human-readable labels, we do not distinguish core roles and modifiers, but macro roles and semantic role labels
- role labels are based on PropBank labels (optional in PropBank)
- abandon sense/predicate differences, instead, we assume that different senses require different roles. Underspecification is not sufficient to distinguish different senses.
- reduce SRL span annotations to head annotation
- seamless integration with CoNLL-U format (as
DEPSannotation)
Current SSRL derived corpus is in data/export
Make sure to install python requirements in requirements.txt
See Makefile, for further information on the conversion of annotated corpora to sentential semantics based SRL: see in sub-folders:
* `data/pmb/` manually corrected Gold subcorpus of the Parallel Meaning Bank (constructed examples) => sentential semantics based SRL edition, see there