Treebanks as Training Data for Parsers

1
Treebanks as Training Data for Parsers

• Joakim Nivre
• Växjö University and Uppsala University
• E-mail nivre_at_msi.vxu.se

2
Q1 What do you really care about when youre
building a parser?

• For parsing unrestricted text, I care about the
  joint optimization of
• Robustness
• Disambiguation
• Accuracy
• Efficiency
• Requirement on syntactic annotation
• Balance between expressivity and complexity

3
Example Mildly Non-Projective Dependency
Structures

• Dependency structure in two treebanks
• Strictly projective (efficiently parsable)
• PDT 75
• DDT 85
• Unrestricted non-projective (often intractable)
• PDT 100
• DDT 100
• Well-nested, gap degree 1
• PDT 99.5
• DDT 99.7
• Design choice in treebank annotation?

4
Q2 What works, what doesnt?

• Anything works?
• Top systems in CoNLL 2006 shared task
• MSTParser Global, exhaustive, graph-based
• MaltParser Local, greedy, stack-based
• Features more important than parsers?
• But not for all languages?
• Results from CoNLL 2007 shared task
• Configurational languages 85 LAS(Catalan,
  Chinese, English, Italian)
• Richly inflected languages 75 LAS(Arabic,
  Basque, Czech, Greek, Hungarian, Turkish)
• Treebank problem or parser problem?

5
Q3 What information is useful, what is not?

• Word level
• Morphological analysis (lemma, derivation,
  inflection)
• Hierarchical parts-of-speech (incl. features)
• Sentence level
• Complete structural annotation (phrases, heads)
• Complete functional annotation (syntactic
  relations)
• Deep/non-local dependencies
• Integrated morpho-syntactic annotation
• The key to parsing richly inflected languages?

6
Skipping a few questions

• Q4 How does grammar writing interact with
  treebanking?
• No idea. Not my cup of tea.
• Q5 What methodological lessons can be drawn
  for treebanking?
• Q6 What are advantages and disadvantages of
  preprocessing the data to be treebanked with an
  automatic parser?
• Dont know. Never got funding to build a real
  treebank.

7
Q7 Advantages of a phrase structure and/or a
dependency treebank?

• Obvious answer
• Phrase structure is good for phrase structure
  parsing.
• Dependency is good for dependency parsing.
• Methodological point
• Parsing lossy conversions can be questionable.
• Remedy
• Make annotations (just) rich enough to support
  both.
• Annotation scheme
• Minimal source annotation
• Well-defined conversions to target annotations