Integrating Syntax and Word Alignment in Syntax-Based Machine Translation

Training a string-to-tree syntax-based statistical machine translation
system to translate from a source language (e.g. Chinese or Arabic)
into a target language (e.g. English) requires the following
resources: a parallel corpus (a large set of example sentences in the
source language that have been translated into the target language by
a human); a word alignment (a word-to-word correspondence between each
source-target sentence pair); and a parse tree (a syntactic
representation) of each sentence in the target language. From these
training examples, the system learns to translate source-language
sequences of words into target-language trees. In order to ensure
broad coverage, the parallel corpus of training examples must be
sufficiently large (on the order of millions of sentence pairs).
Manually annotating such large corpora would be prohibitively
time-consuming. Instead, these corpora must be word-aligned and
parsed automatically.

There are two problems with existing approaches to automatic word
alignment and parsing for syntax-based machine translation. First,
these processes are noisy and introduce errors which impact
translation quality. Second, these processes are typically performed
independently of one another. Since each process produces constraints
that can be used to guide the other, by more closely integrating them,
we can expect to improve the accuracy of each process. In this
thesis, we address these two problems as follows: first, we improve
upon the accuracy of a state-of-the-art parser; second, we use word
alignments to improve parse accuracy; third, we use parses to improve
word alignment accuracy; and fourth, we optimize parses and word
alignments simultaneously. We examine the impact of each of these
methods upon parse quality, alignment quality, and translation quality
in a downstream syntax-based machine translation system. Our results
demonstrate that more closely integrating word alignment and syntactic
parsing can indeed improve the accuracy of each process, and in some
cases leads to an improvement in translation quality relative to a
state-of-the-art syntax-based statistical machine translation system.