Discriminative Phrase-based Lexicalized Reordering Models using Weighted Reordering Graphs

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 47–55, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Discriminative Phrase-based Lexicalized Reordering Models using Weighted Reordering Graphs Wang Ling, Jo˜ao Grac¸a, David Martins de Matos, Isabel Trancoso, Alan Black L2F Spoken Systems Lab, INESC-ID, Lisboa, Portugal Language Technologies Institute, Carnegie Mellon University, Pittsburgh, PA, USA {wang.ling,joao.graca,isabel.trancoso,david.matos}@inesc-id.pt awb@cs.cmu.edu Abstract Lexicalized reordering models play a cen- tral role in phrase-based statistical ma- chine translation systems. Starting from the distance-based reordering model, im- provements have been made by consider- ing adjacent words in word-based mod- els, adjacent phrases pairs in phrase- based models, and finally, all phrases pairs in a sentence pair in the reordering graphs. However, reordering graphs treat all phrase pairs equally and fail to weight the relationships between phrase pairs. In this work, we propose an extension to the reordering models, named weighted re- ordering models, that allows discrimina- tive behavior to be defined in the estima- tion of the reordering model orientations. We apply our extension using the weighted alignment matrices to weight phrase pairs, based on the consistency of their align- ments, and define a distance metric to weight relationships between phrase pairs, based on their distance in the sentence. Experiments on the IWSLT 2010 evalu- ation dataset for for the Chinese-English language pair yields an improvement of 0.38 (2%) and 0.94 (3.7%) BLEU points over the state-of-the-art work’s results us- ing weighted alignment matrices. 1 Introduction Reordering in Machine Translation (MT) is the task of word-order redistribution of translated words. An early reordering paradigm uses a sim- ple distance based reordering model, which penal- izes words that diverge from their original posi- tion after being translated (Koehn et al., 2003). This works moderately well for language pairs where reordering distances are small, but per- forms poorly for language pairs such as Chinese- English, where the opposite occurs. One of many approaches to implement improved reorder- ing models is to use the lexical information dur- ing the phrase extraction algorithm to predict re- ordering orientations, using word-aligned bilin- gual sentences. However, the fact that spurious word alignments might occur leads to the use of alternative representations for word alignments that allow multiple alignment hypotheses, rather than the 1-best alignment (Venugopal et al., 2009; Mi et al., 2008; Christopher Dyer et al., 2008). More recently, a more efficient representation of multiple alignments was proposed in (Liu et al., 2009) named weighted alignment matrices, which represents the alignment probability distribution over the words of each parallel sentence. The method for building a word-based lexicalized re- ordering model using these matrices is proposed in (Ling et al., 2011). However, phrase-based re- ordering models have been shown to perform bet- ter than word-based models for several language pairs (Tillmann, 2004; Su et al., 2010; Galley and Manning, 2008), such as Chinese-English and Arabic-English. In this work, we propose an extension to the phrase-based lexicalized model approach using re- ordering graphs presented in (Su et al., 2010), which …

Content truncated to 500 words