TakeTwo: A Word Aligner based on Self Learning

PACLIC 28 !282 TakeTwo: A Word Aligner based on Self Learning Jim Chang, Jian-Cheng Wu, Jason S. Chang Department of Computer Science National Tsing Hua University 101, Guangfu Road, Hsinchu, Taiwan {jim.chang.nthu, wujc86, jason.jschang}@gmail.com Abstract State of the art statistical machine translation systems are typically trained by symmetrizing word alignments in two translation directions. We introduce a new method that improves word alignment results, based on self learn- ing using the initial symmetrized word align- ments results. The method involves align- ing words and symmetrizing alignments, gen- erating labeled training data, and construct a classifier for predicting word-translation rela- tion in another alignment round. In the first alignment round, we use the original grow- diag-final-and procedure, while in the second round, we use the classifier and a modified GDFA procedure to validate and fill in align- ment links. We present a prototype system, TakeTwo, which applies the method to im- prove on GDFA. Preliminary experiments and evaluation on a hand-annotated dataset show that the method significantly increases the pre- cision rate by a wide margin (+16%) with comparable recall rate (-3%). 1 Introduction The first statistical machine translation (SMT) mod- els are the IBM models, based on statistics collected over a parallel corpus of translated text. These gen- erative IBM models break up the translation process into a number of steps. The most important step is word translation, which is modelled by the lex- ical translation probability, trained from a parallel corpus, typically with the Expectation Maximization (EM) algorithm (Dempster, Laird, and Rubin 1977). However, EM word aligners are data-hungry and produce noisy links due to data sparseness. Many researchers (e.g., Gale and Church 1992, Johnson et al., 2007) have pointed out that, even with a large parallel corpus, the EM algorithms running IBM models still produces noisy links for low frequency words and non-literal translations. Koehn, Och, and Marcu (2003) propose an im- proved word alignment method based on running IBM models in both translation directions for the two languages involved, and symmetrizing the re- sults using a so-called grow-diag-final-and (GDFA) procedure. In a nutshell, GDFA is a heuristic greedy algorithm that starts by accepting reliable links in the intersection of the two alignments. Then, GDFA attempts to add union links neighboring intersec- tion links. Finally, other non-neighboring links are added, subject to 1-1 alignment constraint. This pro- gressively expanding scheme substantially enhances word alignment accuracy. However, the GDFA pro- cedure still leaves much room for improvement, es- pecially for low-frequency translations, non-literal translations, and sentences with extraneous/deleted translations. Consider the following English sentence with Mandarin Chinese translation in a parallel corpus: (1) He made this remark after Heinonen arrived in Tehran. ÷ / ( w˛Á µT ∑—- ta shi zai hainuoning dida deheilan he is when Heinonen arrive Tehran å |h ⇡ ⇧ «q ⇥ hou fabiao zhe xiang tanhua . after deliver this MEASURE talk . See Figures 1(c) for examples of noisy and miss- ing links, produced by Giza++ with the GDFA sym- Copyright 2014 by …

Content truncated to 500 words