Machine Translation

3

IBM Model of Alignment

In Chapter 2, we discussed a basic framework for alignment. We started with noisy channel formulation of SMT:

e^=arg maxe(P(e|f))=arg maxe(P(e)⋅P(f|e))(3.1)

P(f|e) is the parameter called translation model whose computation involves alignment. The alignment task is tackled in two simple settings:

1. If the source and target languages are very close to each other, then we can take a part of speech (POS) tagging-like approach for translation. Typical examples of such close pairs are Hindi-Urdu, Spanish-Catalan, etc., where words and their translations almost always occupy identical positions on the two sides of the translation. The target language vocabulary V_E can be looked upon as the “tag” repository, and a hidden ...