Uppsala Group on Future of UD
Aitziber, Anders, Filip, Jan, Jenna, Tommi, Veronika, Zeljko
This group discussed possible extensions towards annotation of semantics.
Discussion covered the following topics:
- semantic roles: which ones, extent, representation
- word senses: which “ontology”, difficulties with IAA
- propbanking/valency
- conjunct propagation
- Named entities (wrt to existing annotation guidelines in UD)
- Grounding
- co-reference, information structure, discourse
- ellipsis / NULL element addition
- in view of preceding group discussions, especially MWE group, LVCs and idioms
- principles, minimal set of annotation to be called “semantic annotation”
Group-internal consensus:
- no change to current format and structure - if anything, must be “add-on” (only addition to current scheme and format)
- basis features to include in “UD Semantics 1.0”:
- exlicitly identified content words
- graph structure over content words, with conjunct propagated (“effective heads”)
- labeled relations on every “semantic” edge (~30 types: not semantic roles, but rather arguments as in propbank, plus non-argument relations as in PDT)
- LVCs and idioms annotated
- Named entities, labeled at least on “top level” hierarchy (~5-7 classes)
- NULL words: only positively defined
- WSD with caution (maybe not in UD Semantics 1.0 - which ontology??)
- something should be out by end of next year (process: survey of treebanks, guidelines, annotation)
- minimal set for a treebank to be considered as having semantic annotation: content words, all relations (with effective heads resolved), labels at least for verbs and their arguments
Report and discussion with all
- several concerns raised
- do “enhanced” things first
- named entities, LVCs, idioms perhaps the most pressing in terms of need?
- several non-group people expressed interest to contribute to the process