A Maximum Entropy Word Aligner for Arabic-English Machine Translation

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 89–96, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics A Maximum Entropy Word Aligner for Arabic-English Machine Translation Abraham Ittycheriah and Salim Roukos IBM T.J. Watson Research Center 1101 Kitchawan Road Yorktown Heights, NY 10598 {abei,roukos}@us.ibm.com Abstract This paper presents a maximum entropy word alignment algorithm for Arabic- English based on supervised training data. We demonstrate that it is feasible to cre- ate training material for problems in ma- chine translation and that a mixture of su- pervised and unsupervised methods yields superior performance. The probabilistic model used in the alignment directly mod- els the link decisions. Significant improve- ment over traditional word alignment tech- niques is shown as well as improvement on several machine translation tests. Perfor- mance of the algorithm is contrasted with human annotation performance. 1 Introduction Machine translation takes a source sequence, S = [s1 s2 . . . sK] and generates a target sequence, T = [t1 t2 . . . tM] that renders the meaning of the source sequence into the target sequence. Typically, algorithms operate on sentences. In the most general setup, one or more source words can generate 0, 1 or more target words. Current state of the art machine translation systems (Och, 2003) use phrasal (n-gram) features extracted automatically from parallel corpora. These phrases are extracted using word alignment algorithms that are trained on parallel corpora. Phrases, or phrasal features, represent a mapping of source sequences into a target sequences which are typically a few words long. In this paper, we investigate the feasibility of train- ing alignment algorithms based on supervised align- ment data. Although there is a modest cost associ- ated with annotating data, we show that a reduction of 40% relative in alignment error (AER) is possible over the GIZA++ aligner (Och and Ney, 2003). Although there are a number of other applications for word alignment, for example in creating bilingual dictionaries, the primary application continues to be as a component in a machine translation system. We test our aligner on several machine translation tests and show encouraging improvements. 2 Related Work Most of the prior work on word alignments has been done on parallel corpora where the alignment at the sentence level is also done automatically. The IBM models 1-5 (Brown et al., 1993) produce word align- ments with increasing algorithmic complexity and performance. These IBM models and more recent refinements (Moore, 2004) as well as algorithms that bootstrap from these models like the HMM algo- rithm described in (Vogel et al., 1996) are unsuper- vised algorithms. The relative success of these automatic techniques together with the human annotation cost has delayed the collection of supervised word-aligned corpora for more than a decade. (Cherry and Lin, 2003) recently proposed a di- rect alignment formulation and state that it would be straightforward to estimate the parameters given a supervised alignment corpus. In this paper, we ex- tend their work and show …

Content truncated to 500 words