What’s in a translation rule?

What’s in a translation rule? Michel Galley Dept. of Computer Science Columbia University New York, NY 10027 galley@cs.columbia.edu Mark Hopkins Dept. of Computer Science University of California Los Angeles, CA 90024 mhopkins@cs.ucla.edu Kevin Knight and Daniel Marcu Information Sciences Institute University of Southern California Marina Del Rey, CA 90292 {knight,marcu}@isi.edu Abstract We propose a theory that gives formal seman- tics to word-level alignments defined over par- allel corpora. We use our theory to introduce a linear algorithm that can be used to derive from word-aligned, parallel corpora the minimal set of syntactically motivated transformation rules that explain human translation data. 1 Introduction In a very interesting study of syntax in statistical machine translation, Fox (2002) looks at how well proposed trans- lation models fit actual translation data. One such model embodies a restricted, linguistically-motivated notion of word re-ordering. Given an English parse tree, children at any node may be reordered prior to translation. Nodes are processed independently. Previous to Fox (2002), it had been observed that this model would prohibit certain re-orderings in certain language pairs (such as subject- VP(verb-object) into verb-subject-object), but Fox car- ried out the first careful empirical study, showing that many other common translation patterns fall outside the scope of the child-reordering model. This is true even for languages as similar as English and French. For example, English adverbs tend to move outside the lo- cal parent/children in environment. The English word “not” translates to the discontiguous pair “ne ... pas.” English parsing errors also cause trouble, as a normally well-behaved re-ordering environment can be disrupted by wrong phrase attachment. For other language pairs, the divergence is expected to be greater. In the face of these problems, we may choose among several alternatives. The first is to abandon syntax in statistical machine translation, on the grounds that syn- tactic models are a poor fit for the data. On this view, adding syntax yields no improvement over robust phrase- substitution models, and the only question is how much does syntax hurt performance. Along this line, (Koehn et al., 2003) present convincing evidence that restricting phrasal translation to syntactic constituents yields poor translation performance – the ability to translate non- constituent phrases (such as “there are”, “note that”, and “according to”) turns out to be critical and pervasive. Another direction is to abandon conventional English syntax and move to more robust grammars that adapt to the parallel training corpus. One approach here is that of Wu (1997), in which word-movementis modeled by rota- tions at unlabeled, binary-branching nodes. At each sen- tence pair, the parse adapts to explain the translation pat- tern. If the same unambiguous English sentence were to appear twice in the corpus, with different Chinese trans- lations, then it could have different learned parses. A third direction is to maintain English syntax and investigate alternate transformation models. After all, many conventional translation systems are indeed based on syntactic transformations far more expressive than what has been proposed in syntax-based statistical MT. We take this approach …

Content truncated to 500 words