<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3228">
  <Title>Dependencies vs. Constituents for Tree-Based Alignment</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Statistical approaches to machine translation, pioneered by Brown et al. (1990), estimate parameters for a probabilistic model of word-to-word correspondences and word re-orderings directly from large corpora of parallel bilingual text. In recent years, a number of syntactically motivated approaches to statistical machine translation have been proposed. These approaches assign a parallel tree structure to the two sides of each sentence pair, and model the translation process with reordering operations defined on the tree structure. The tree-based approach allows us to represent the fact that syntactic constituents tend to move as unit, as well as systematic differences in word order in the grammars of the two languages. Furthermore, the tree structure allows us to make probabilistic independence assumptions that result in polynomial time algorithms for estimating a translation model from parallel training data, and for finding the highest probability translation given a new sentence.</Paragraph>
    <Paragraph position="1"> Wu (1997) modeled the reordering process with binary branching trees, where each production could be either in the same or in reverse order going from source to target language. The trees of Wu's Inversion Transduction Grammar were derived by synchronously parsing a parallel corpus, using a grammar with lexical translation probabilities at the leaves and a simple grammar with a single nonterminal providing the tree structure. While this grammar did not represent traditional syntactic categories such as verb phrases and noun phrases, it served to restrict the word-level alignments considered by the system to those allowable by reordering operations on binary trees.</Paragraph>
    <Paragraph position="2"> Yamada and Knight (2001) present an algorithm for estimating probabilistic parameters for a similar model which represents translation as a sequence of re-ordering operations over children of nodes in a syntactic tree, using automatic parser output for the initial tree structures. This gives the translation model more information about the structure of the source language, and further constrains the reorderings to match not just a possible bracketing as in Wu (1997), but the specific bracketing of the parse tree provided.</Paragraph>
    <Paragraph position="3"> Recent models of alignment have attempted to exploit syntactic information from both languages by aligning a pair of parse trees for the same sentence in either language node by node. Eisner (2003) presented such a system for transforming semantic-level dependecy trees into syntactic-level dependency trees for text generation. Gildea (2003) trained a system on parallel constituent trees from the Korean-English Treebank, evaluating agreement with hand-annotated word alignments. Ding and Palmer (2004) align parallel dependency trees with a divide and conquer strategy, choosing a highly likely word-pair as a splitting point in each tree. In addition to providing a deeper level of representation for the transformations of the translation model to work with, tree-to-tree models have the advantage that they are much less computationally costly to train than models which must induce tree structure on one or both sides of the translation pair. Because Expectation Maximization for tree-to-tree models iterates over pairs of nodes in the two trees, it is O(n2) in the sentence length, rather than O(n6) for Wu's Inversion Transduction Grammar or O(n4) for the Yamada and Knight tree-to-string model.</Paragraph>
    <Paragraph position="4"> In this paper, we make a comparison of two tree-to-tree models, one trained on the trees produced by automatic parsers for both our English and Chinese corpora, and one trained on the same parser output converted to a dependency representation. The trees are converted using a set of deterministic head rules for each language. The dependency representation equalizes some differences in the annotation style between the English and Chinese treebanks. However, the dependency representation makes the assumption that not only the bracketing structure, but also the head word choices, will correspond in the two trees. Our evaluation is in terms of agreement with word-level alignments created by bilingual human annotators. Our model of alignment is that of Gildea (2003), reviewed in Section 2 and extended to dependency trees in Section 3. We describe our data and experiments in Section 4, and discuss results in Section 5.</Paragraph>
  </Section>
class="xml-element"></Paper>