February 15, 2006 machine translation interlinguas machine learning for MT learn from word alignments learn from sentence alignments bag-of-words model (ICA 10) why MT is hard lexical reasons (ambiguity, hypernyms, hyponyms, lexical gap, light verb, idioms, other idiosyncracies) syntactic/sematic reasons (ambiguity, word order) pragmatic reasons (context affects interpretation)