Interlingua Annotation

1
Interlingua Annotation

• Owen Rambow
• Advaith Siddharthan
• Kathleen McKeown
• rambow_at_cs.columbia.edu

2
Goal

• Determine feasible deep semantic,
  language-independent annotation (interlingua)
  for text
• Different from PropBank, FrameNet, WordNet these
  projects are language-dependent

3
Expected Results

• Annotation guidelines (methodology, manual) for
  annotating language-independent meaning
  representation on texts in 7 languages
• Methodology for porting to new languages
• Annotated corpora

4
Methodology

• Use source-language texts and multiple
  translations into English
• Develop successively more language-independent
  levels of representation
• (deep syntax)
• language-specific lexical disambiguation and
  thematic structure (agent, theme, )
• language-independent representation

5
Methodology (2)

• Six sites (CMU, Columbia, ISI, Mitre, NMSU, UMd)
• Each site has one language Columbia Hindi
• Closer cooperation Columbia-UMd on Arabic and
  Hindi
• Division of tasks and expertise among sites

6
Methodology (3)

• Use annotators from beginning to test
  inter-annotator agreement
• Columbia have hired a native Hindi annotator
  (near-native English) and an English-language
  annotator

7
Research Issues

• Research develop annotation scheme(s),
  methodology, manuals
• Levels (reminder)
• (deep syntax)
• language-specific lexical disambiguation
• language-independent representation
• Questions
• Which levels do we annotate explicitly?
• What is included where?
• How do we annotate? Using which tools?

8
Timeline

• January develop language-specific disambiguation
• February-March annotate, measure
• April-June develop language-independent
  annotation
• July-August annotate, measure
• Year 2 review results, adjust annotation scheme
• Year 3 annotate

9
Arabic Dialects

• Owen Rambow
• (Nizar Habbash)
• rambow_at_cs.columbia.edu

10
Goal

• Investigate representation of linguistic
  resources for closely related languages/dialects
• Example Arabic
• Automatically derive NLP tools for cross-dialect
  MT

11
Note on Arabic

• Interest Only one written dialect Modern
  Standard Arabic (MSA), rarely spoken
  spontaneously
• Many spoken dialects, almost never written
• Dialects function of geography, urban/rural,
  Bedouin/sedentary, sex, religion,
• Code switching (mainly dialect-MSA) several
  linguistic systems in same sentence
• Challenge for traditional NLP approaches!

12
Expected Results

• Representation of phonology, lexicon, morphology,
  and syntax for Modern Standard Arabic and
  Egyptian Colloquial Arabic
• Tools for converting between MSA and ECA
• Demonstration of tools in several domains (ECA
  speech recognition, ECA -gt English translation)

13
Methodology

• Use existing scholarly resources to compile
  sound-change rules, morphological
  representations, syntactic representations
• Use native speakers to validate, and augment
  lexicon
• Develop representation
• Develop automatic compilation of NLP tools

14
Timeline

• Sep-Dec start compiling sound change rules,
  morphological rules, syntax
• Jan-April develop representations for sound
  change rules, morphology
• Jan-Apr develop conversion rules
• May-August work on ECA speech recognition
  application
• Note also working on MSA syntax
• Year 2 extend to syntax, extend to second
  dialect (Palestinian? Iraqi?)