NICE: Native language Interpretation and Communication Environment

1
NICE Native language Interpretation and
Communication Environment

• Lori Levin, Jaime Carbonell, Alon Lavie, Ralf
  Brown
• Carnegie Mellon University

2
Low Cost MT for Minor Languages

• Speakers of electronically underrepresented
  languages can participate in the information age.
• Policy makers can access ideas, viewpoints, and
  information from the developing world.
• MT for unforseen translation needs e.g.,
  humanitarian aid.
• Documentation and preservation of endangered
  languages.

3
NICE Languages

• Spanish and two or three indigenous languages of
  Latin America.
• Part of a larger program of Western Hemisphere
  collaboration.
• Planning joint meetings of indigenous people,
  language technologies people, government and
  funding agencies.

4
Our Approach to Low Cost MT

• Apply new machine learning techniques that
  require
• less development data
• shorter development time
• less specialized human expertise

5
Our Approach to Low Cost MT

• Use multi-engine architecture to
• flexibly make use of whatever resources are
  available
• data Generalized EBMT
• human consultants automatic elicitation
• human experts knowledge-based MT
• Tap the strengths of different approaches to MT.

6
Grammar Acquisition Tool
7
A Noun Phrase Learning Instance and Transfer Rule

• Learning Instance
• English the big boy Hebrew ha-yeled ha-gadol
• Acquired Transfer Rule
• Hebrew NP N ADJ ltgt English NP the ADJ
  N
• where (HebrewN ltgt English N)
• (HebrewADJ ltgt EnglishADJ)
• (HebrewN has ((def )))
• (HebrewADJ has ((def )))

8
Version Space Abstraction Lattice
9
Scientific Challenges of NICE

• Automatic linguistic elicitation from native
  speakers
• bottom up from small phrases to larger
  constructions
• interrogation intermingled with confirmation of
  induced grammatical rules

10
Scientific Challenges of NICE

• Transfer rule induction
• version space learning hone in on the right
  level of generalization
• take advantage of expected constituents, e.g.,
  ART-N but not V-ART
• only local optimality is guaranteed
• converges faster no worst case exponential
  problem

11
iKBMT Instructable Knowledge Based MT
12
Grammar Acquisition 2
13
A Sentence Learning Instance and Transfer Rule

• Learning Instance
• English I saw the big boy
• Hebrew Raiti et ha-yeled ha-gadol
• Acquired Transfer Rule
• Hebrew S V et NP ltgt English S I NP V NP
• where (HebrewV ltgt English V)
• (HebrewNP ltgt English NP)
• (HebrewV has ((agreement 1sg)))
• (HebrewNP has ((def )))