The goal of machine translation is to develop algorithms that
produce human-quality translations of natural language sentences.
The evaluation of machine translation quality is split broadly into
two aspects: adequacy and fluency. Adequacy measures
how faithfully the meaning of the original sentence is preserved,
whereas fluency measures whether this meaning is expressed in valid
sentences in the target language. While both of these criteria are
difficult to meet, fluency is a much more difficult goal.
Generally, this likely has something to do with the asymmetrical
nature of producing and understanding sentences; although humans are
quite robust at inferring the meaning of text even in the presence
of lots of noise and error, the rules that govern grammatical
utterances are exacting, subtle, and elusive. To produce
understandable text, we can rely on this robust processing hardware,
but to produce grammatical text, we have to understand how it works.
This dissertation attempts to improve the fluency of machine
translation output by explicitly incorporating models of the target
language structure into machine translation systems. It is
organized into three parts. First, we propose a framework for
decoding that decouples the structures of the sentences of the
source and target languages, and evaluate it with existing
grammatical models as language models for machine translation.
Next, we apply lessons from that task to the learning of grammars
more suitable to the demands of the machine translation. We then
incorporate these grammars, called Tree Substitution Grammars, into
our decoding framework.