Machine Translation with Scarce Resources

1
Machine Translation with Scarce Resources

• The Avenue Project

2
Scarce Resources

• Not much text in electronic form.
• Very few linguists who can write computational
  rules.
• No standard orthography
• Kudaw, kusaw (work) (Mapudungun, Chile)
• Not even sure of pronunciation
• EH-nvelope, AH-nvelope (envelope) (English, US,
  not a language with scarce resources)

3
Our Approach

• Learn rules from a controlled corpus.
• Corpus is elicited from bilingual speakers.
• The informant only needs to translate and align
  words.

4
AVENUE Project

• New Ideas
• Use machine learning to learn translation rules
  from native speakers who are not trained in
  linguistics or computer science.
• Multi-Engine translation architecture can
  flexibly take advantage of whatever resources are
  available.
• Research partnerships with indigenous communities
  in Latin America and Alaska (Mapudungun (Chile),
  Siona (Colombia), Inupiaq (Alaska))

Interface for data elicitation

• Impact
• Rapid and low-cost development of machine
  translation for languages with scarce resources.
• Policy makers can get input from indigenous
  people.
• Indigenous people can participate in government
  and internet.

Schedule
Year 1 Seeded Version Space learning first
version Year 2 Example-Based Machine
Translation of Mapudungun (Chile). Year 3
Multi-Engine Mapudungun system (EBMT and
partially learned transfer rules)
Carnegie Mellon University, Language Technologies
Institute L. Levin, J. Carbonell, A. Lavie, R.
Brown
5
Elicitation Interface
6
Elicitation Corpus example

• English I fell.
• Spanish Caí
• Mapudungun Tranün
•  
• English I am falling.
• Spanish Estoy cayendo
• Mapudungun Tranmeken

7
Elicitation Corpus example

• English You (John) fell.
• Spanish Tu (Juan) caiste
• Mapudungun Eymi tranimi (Kuan)
• English You (Mary) fell.
• Spanish Tu (María) caiste
• Mapudungun Eymi tranimi (Maria)
• English The rock fell.
• Spanish La piedra cayó
• Mapudungun Trani chi kura

8
(No Transcript)