Syntactic movement in statistical translation

Statistical machine translation (SMT) provides a framework to learn automatic translation systems from parallel texts. The statistical approach recasts translation and its subproblems as optimization problems, where one must find the best output as scored by an empirically-derived model. One complication shared by most translation subproblems is movement. While sentence order usually remains the same during translation, the order of the concepts within those sentences can vary drastically from language to language. If one assumes that concepts can move with complete freedom during translation, then the set of possible outputs can become very large. One can reduce this complexity by assuming that sentences have a context-free syntactic tree-structure, which explains all movement. This decomposes a sentence into subtrees, which define syntactic phrases that must exist in both the source and its translation. This thesis employs two syntactic constraints: an inversion transduction grammar (ITG) constraint that considers all possible binary trees, and a cohesion constraint that considers only a single tree, which is provided for one of the two languages.;We develop three distinct methods that use syntactic movement constraints to improve either the efficiency or accuracy of existing, non-syntactic solutions to SMT subproblems, such as alignment and decoding. The first is a phrasal ITG, which introduces an ITG constraint in order to gain polynomial-time algorithms for phrasal translation modeling. The resulting syntactic system improves performance over a comparable, flat-string model. The second project compares and combines ITG and cohesion constraints, as they are applied to bilingual word alignment. We present two combined alignment spaces, and show that a combination of ITG and cohesion constraints improves upon a comparable, bipartite-matching aligner. We also present a method to discriminatively train bitext parsers, allowing us to incorporate a powerful soft cohesion constraint into discriminative word-alignment. Our third project defines cohesion on the translations output by a phrase-based decoder, given a source-side dependency tree. The resulting cohesive, phrase-based decoder is shown to produce translations that are preferred over non-cohesive output by both human evaluators and automatic metrics.