Pre-Reordering for Machine Translation Using Transition-Based Walks on Dependency Parse Trees

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 164–169, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Pre-reordering for machine translation using transition-based walks on dependency parse trees Antonio Valerio Miceli-Barone Dipartimento di Informatica Largo B. Pontecorvo, 3 56127 Pisa, Italy miceli@di.unipi.it Giuseppe Attardi Dipartimento di Informatica Largo B. Pontecorvo, 3 56127 Pisa, Italy attardi@di.unipi.it Abstract We propose a pre-reordering scheme to improve the quality of machine translation by permuting the words of a source sen- tence to a target-like order. This is accom- plished as a transition-based system that walks on the dependency parse tree of the sentence and emits words in target-like or- der, driven by a classifier trained on a par- allel corpus. Our system is capable of gen- erating arbitrary permutations up to flexi- ble constraints determined by the choice of the classifier algorithm and input features. 1 Introduction The dominant paradigm in statistical machine translation consists mainly of phrase-based sys- tem such as Moses (Koehn et.al.,2007). Differ- ent languages, however, often express the same concepts in different idiomatic word orders, and while phrase-based system can deal to some ex- tent with short-distance word swaps that are cap- tured by short segments, they typically perform poorly on long-distance (more than four or five words apart) reordering. In fact, according to (Birch et.al., 2008), the amount of reordering be- tween two languages is the most predictive feature of phrase-based translation accuracy. A number of approaches to deal with long- distance reordering have been proposed. Since an extuasive search of the permutation space is un- feasible, these approaches typically constrain the search space by leveraging syntactical structure of natural languages. In this work we consider approaches which in- volve reordering the words of a source sentence in a target-like order as a preprocessing step, be- fore feeding it to a phrase-based decoder which has itself been trained with a reordered training set. These methods also try to leverage syntax, typically by applying hand-coded or automatically induced reordering rules to a constituency or de- pendency parse of the source sentence. (Gal- ley and Manning, 2008; Xu et.al., 2009; Genzel, 2010; Isozaki et.al., 2010) or by treating reorder- ing as a global optimization problem (Tromble and Eisner, 2009; Visweswariah et.al., 2011). In or- der to keep the training and execution processes tractable, these methods impose hard constrains on the class of permutations they can generate. We propose a pre-reordering method based on a walk on the dependency parse tree of the source sentence driven by a classifier trained on a parallel corpus. In principle, our system is capable of generat- ing arbitrary permutations of the source sentence. Practical implementations will necessarily limit the available permutations, but these constraints are not intrinsic to the model, rather they depend on the specific choice of the classifier algorithm, its hyper-parameters and input features. 2 Reordering as a walk on a dependency tree 2.1 Dependency parse trees Let a sentence be a list of words s (w1, w2, . . …

Content truncated to 500 words