The SALSA experience: semantic role annotation

1
The SALSA experience semantic role annotation

• Katrin Erk
• University of Texas at Austin

2
Semantic role annotation in SALSA

• SALSA The Saarbrücken Lexical Semantics
  Annotation and Analysis project
• Manual annotation of the German TIGER corpus with
  lexical semantic information
• Basis The Berkeley FrameNet database
• Verbs annotated with their Frame ( sense),plus
  semantic roles
• TIGER corpus
• 1.5 million words / 80 K sentences of German
  newspaper text (Frankfurter Rundschau)
• Stuttgart/Potsdam/Saarbrücken
• Phrase types and grammatical functions

3
Annotation Scheme
(They didnt want to pay the move back because
the employee had quit.)

• Semantics
• Independent frames
• Trees of depth one
• One edge points to target, others to frame
  elements
• Sem. roles point to syn. constituents

• TIGER Syntax
• Node labels constituents
• Edge labels gramm. functions
• Crossing edges
• POS

4
Experiences with the semantic role annotation in
Salsa

• Frame ( sense) assignment more difficult than
  role assignment
• Multiple tags possible, at frame level and at
  role level
• Limited compositionality phenomena, each with
  separate annotation format in Salsa
• Light verbs, metaphor, idioms
• Distinction often difficult metaphor vs idiom,
  bleaching
• If I did this again, one format, multiple tags
  possible
• Annotation beyond the sentence boundary
• Message role in Communication frames
• Annotation below the word boundary German noun
  compounds
• Mietrechtsdiskussion discussion of tenant law

5
Encoding sem. role annotation TIGER XML as a
great basis

• TIGER XML
• each constituent is an XML element with a
  globally unique ID
• Syn. edges explicitly encodedltedgegt elements
  links two nodes, referring to their IDs
• Models discontinuous constituents
• Salsa/Tiger XML
• Sem. annotation by adding a modular ltsemgt block
  to the XML structure of a sentence
• Semantics points to syn. constituents using their
  IDs
• Annotation beyond sentence boundary possible
  globally unique syn. IDs

6
Extracting a lexicon need for a deeper, richer
syntax

• Extracting syntax/semantics mapping
• needs to identify gramm. functions filled by
  sem. roles
• Problems
• Constituent structure rather thandependencies
  subjects hard to retrieve
• TIGER does not mark voice
• Shallow format for PPs determining heads is hard
• Coordination is a pain