MT Challenges

1
MT Challenges

• Ed Kenschaft
• University of Maryland
• kensch at umd

2
My Perspective

• Software engineer
• Linguistics student at UMD
• Researcher in NLP group
• Studied with SIL translators
• Analyst with SIL software development

3
SIL International

• Faith-based Christian organization
• Partner with speakers of languages that have
  never been written down
• Purposes
• preserve the language and culture
• document the language for study
• translate the Bible and community development
  materials
• Documented 1400 languages in 70 countries

4
Challenges of Bible Translation

• Ultra-low-density languages
• Nearly endless variety of target languages
• 2000-3000 remaining
• Exceedingly rich domain of discourse
• approximates all of natural language
• Demand for 100 accuracy/fluency
• Cultural variation
• Intelligibility ? Fidelity

5
Cultural Context

• Cleanse me with hyssop, and I will be
  clean wash me, and I will be whiter than
  snow.(Psalm 517, NIV)
• What is hyssop?
• What is snow?
• What does it mean to be white?
• Cleanse me with a plant indigenous to the lands
  of the ancient Near East, used in Jewish
  religious ceremonies, and I will be whiter than
  the precipitation that falls like rain when the
  weather is very cold, which indicates a state of
  moral purity.

6
Intelligibility ? Fidelity

• Where there is no vision, the people perish.
  (Proverbs 2918a, KJ21)
• When people do not accept divine guidance, they
  run wild. (Pr 2918a, NLT)

7
Waste of Time?

• Can a computer replace a translator?
• Limited domains only
• What can it do?
• Word-processing
• Data storage analysis
• First draft?

8
General Approach

• CAT vs. MT
• Linguistically informed systems
• Supervised learning
• Exploit all available resources
• SL resources
• Existing TL data

9
Data Representation

• Text encoding
• Unicode
• Fonts
• Graphite
• Interlinear text
• LinguaLinks, Toolbox, FieldWorks

10
Elicitation Analysis

• Elicit syntactic morphological data
• AVENUE, EXPEDITION
• Elicit word lists for language survey
• WordSurv

11
SL Resources

• Related language adaptation
• CARLA
• Projection across word alignment
• GIZA, Multi-Align, Parser Projection

12
NLG

• Rich interlingua
• TBTA (Tod Allman)
• Statistical fluency enhancement
• (Sebastian Varges)

13
The Limits of NLP

• Who knows?
• TMI-2004