First Meeting

Grand Themes in Machine Translation

Machine translation is a testing ground for computational models of language. Sometimes we can get away with surface level methods, but for syntactically divergent and semantically distant language, much heavy lifting is required.

• Ambiguity
• Sparse data
• Balancing sparse specific evidence against broad general evidence
• How much "supervision"? E.g., manual word alignments, morphological analyzers, syntactic parsers

Some Major Current Research Directions

• Syntactic / semantic machine translation: a slow climb uphill
• Machine learning / parameter estimation: so many features, so little time
• Scaling up to huge data sets (trillion of words of monolingual, billions of parallel data)
• Low resource challenges (e.g., exploiting comparable data)
• Integration with speech and information extraction
• Collaboration with human translators