Source reordering using MaxEnt classifiers and supertags

Source reordering using MaxEnt classifiers and supertags Maxim Khalilov and Khalil Sima’an Institute for Logic, Language and Computation University of Amsterdam P.O. Box 94242 1090 GE Amsterdam, The Netherlands {m.khalilov,k.simaan}@uva.nl Abstract Source language reordering can be seen as the preprocessing task of permuting the or- der of the source words in such a way that the resulting permutation allows as mono- tone a translation process as possible. We explore a simple but effective source re- ordering algorithm that works as a cascade of source string transforms, each consist- ing of swapping the positions of a single pair of adjacent words in order to unfold a candidate pair of crossing alignments. The decision to swap a pair of words is mod- elled as a binary classification task formu- lated as a log-linear model and trained un- der maximum entropy (MaxEnt). We ex- periment with features that consist of the local neighborhood of both words as well as lexico-syntactic representations known as supertags. Our experiments on the English-to-Dutch EuroParl translation task show that the cascaded alignment unfold- ing slightly improves the performance of a state-of-the-art phrase translation system that uses distance-based and lexicalized block-oriented reordering. 1 Introduction The word-order divergence (also called distortion) between source-target sentence pairs is a major research topic in statistical machine translation (SMT). The problem of reordering in SMT has been attacked from different angles. Standard phrase-based translation models (Och and Ney, 2004) search for the best reordering option dur- ing decoding within a limited distortion space de- c⃝ 2010 European Association for Machine Translation. fined using a local window of phrases, e.g., (Till- man, 2004). Hierarchical approaches, based on In- version Transduction Grammar (ITG), e.g., (Wu and Wong, 1998; Chiang, 2005), explore yet a wider range of reorderings defined by the choice of swapping the order of sibling subtrees under each node in a binary parse-tree of the source/target sentence. Finally, source sentence reordering is a preprocessing task that aims at finding a permuta- tion of the words of the source sentence that con- tains the least number of crossing alignments be- tween source words and their target sentence coun- terparts. This paper is concerned with the task of source language reordering as a preprocessing task. Figure 1 depicts the translation from source string S to target string T with alignment a (solid Sline) and the alternative of source reordering S into ′ followed by the translation S ′ → T with align- ment a ′ (in dashed lines). Source reordering of S is as successfull as much as it will yield a permu- tation S ′ that has as monotone an alignment a ′ as possible with T. Figure 1: Translation schemes with and without a reordering step. This problem can be seen as the task of learn- ing from a word-aligned parallel corpus a model of source permutation from S to S ′, where the lat- ter has monotone alignment with T. The learning task aims at learning a model that minimizes the number of …

Content truncated to 500 words