Syntax-Based Alignment: Supervised or Unsupervised?

Syntax-Based Alignment: Supervised or Unsupervised? Hao Zhang and Daniel Gildea Computer Science Department University of Rochester Rochester, NY 14627 Abstract Tree-based approaches to alignment model translation as a sequence of probabilistic op- erations transforming the syntactic parse tree of a sentence in one language into that of the other. The trees may be learned directly from parallel corpora (Wu, 1997), or provided by a parser trained on hand-annotated treebanks (Ya- mada and Knight, 2001). In this paper, we compare these approaches on Chinese-English and French-English datasets, and find that au- tomatically derived trees result in better agree- ment with human-annotated word-level align- ments for unseen test data. 1 Introduction Statistical approaches to machine translation, pio- neered by Brown et al. (1990), estimate parame- ters for a probabilistic model of word-to-word cor- respondences and word re-orderings directly from large corpora of parallel bilingual text. In re- cent years, a number of syntactically motivated ap- proaches 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 oper- ations defined on the tree structure. The tree-based approach allows us to represent the fact that syn- tactic constituents tend to move as unit, as well as systematic differences in word order in the gram- mars of the two languages. Furthermore, the tree structure allows us to make probabilistic indepen- dence 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. 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 nonter- minal providing the tree structure. While this gram- mar 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 opera- tions on binary trees. This restriction corresponds to intuitions about the alignments that could be pro- duced by systematic differences between the two language’s grammars, and allows for a polynomial time algorithm for finding the highest-probability alignment, and for re-estimation of the lexical trans- lation and grammar probabilities using the Expecta- tion Maximization algorithm. Yamada and Knight (2001) present an algorithm for estimating probabilistic parameters for a simi- lar 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 reorder- ings to match not just a possible bracketing as in Wu (1997), but the specific bracketing of the parse tree provided. In …

Content truncated to 500 words