Treebanks: Layering the Annotation

1
Treebanks Layering the Annotation

• Jan Hajic
• Institute of Formal and Applied Linguistics
• School of Computer Science
• Faculty of Mathematics and Physics
• Charles University, Prague
• Czech Republic

2
Layering the PDT 2.0

• 4 (5) stand-off layers
• Deep structure (t)
• Syntax semnatics
• Dependecy non-dep. links
• Surface structure (a)
• Dependency, function
• Morphology (m)
• Lemma, tag (detailed)
• Word (token) (w)
• Audio/auto transcript (z)

• PML Scheme (XML based)

z-layer
3
The Links

• Within t-layer
• Co-reference links
• Pronoun to antecedent, (future full coref
  chains)
• Complement to 2nd governor, etc.
• Lexicon links
• Verbs, nouns, adjectives, adverbs to dictionary
  entry
• Word sense disambiguated, valency/frame-based
• t-layer to a-layer
• Which a-node the t-node comes from
• No restrictions (crossing, many-to-many, )

4
The Questions I

• Influence of choices made in the underlying
  annotation influenced upper layer choices?
• Minimal or none
• thanks to stand-off annotation style, and
  many-to-many references/links allowed (XML IDs)
• Added annotation (over surface syntax)
• Node order (information structure), deep
  dependencies, 30 node labels (time, modalities,
  semantic POS, number, pronoun classes, ),
  co-reference, valency dictionary ( frame
  files) links (word sense annotation), empty
  nodes (args),

5
The Questions II

• Hard to circumvent syntactic choices?
• Not really (again, thanks to XML stand-off)
• Only 1 label at surface syntactic level
  (function)
• Dependency(-only) no problem (no need to refer to
  phrases all represented by subtrees)
• but there will be a problem with the t-layer
• When referring from some higher (logic)
  layer
• (Probably) need to refer to labels (attributes)
• Solution
• Add IDs to attributes (should be easy, in fact
  XML ID)

6
The Questions III

• Desirable characteristics for adding layers
• Stand-off annotation
• Proper IDs for in-, between-layer reference
• In advance, if possible, but usually can be added
  later
• Quality Control
• !! Easier with layers - cross-layer constraints
• Invisible to annotators -gt catch random errors
• Links (between-layer type) can be pre-annotated
• PS vs. dep. impact on additional annotation
• Not observed