Hebrew Morphological Preprocessing for Statistical Machine Translation

Hebrew Morphological Preprocessing for Statistical Machine Translation Nimesh Singh and Nizar Habash Center for Computational Learning Systems Columbia University nks2118@columbia.edu habash@ccls.columbia.edu Abstract This paper presents a range of preprocess- ing solutions for Hebrew-English statisti- cal machine translation. Our best system, using a morphological analyzer, increases 3.5 BLEU points over a no-tokenization baseline on a blind test set. The next best system uses Morfessor, an unsupervised morphological segmenter, and obtains al- most 3.0 BLEU points over the baseline. 1 Introduction Much research in statistical machine translation (SMT) has shown the importance of morpholog- ical preprocessing (aka, tokenization, segmenta- tion) on translation quality. The common wisdom in the field is that such preprocessing helps, es- pecially for morphologically rich languages, such as Arabic, Spanish or Finnish, because it reduces model sparsity and increases source-target sym- metry (particularly when the target is morpholog- ically poor, as in English). However, the value of preprocessing generally decreases with added training data, and is highly dependent on the lan- guage pair and particular preprocessing approach (Popovi´c and Ney, 2004; Lee, 2004; Goldwater and McClosky, 2005; Habash and Sadat, 2006; Fishel and Kirik, 2010; Al-Haj and Lavie, 2012). In this paper, we present results from a set of experiments to determine an optimal preprocess- ing method for Hebrew-English SMT, a language pair with limited previously published work (Lavie et al., 2004; Lembersky et al., 2011). We report on three types of preprocessing techniques using de- terministic regular-expressions, unsupervised mor- phology learning, and morphological analysis and c⃝ 2012 European Association for Machine Translation. disambiguation. Our results show that using a mor- phological analyzer helps translation quality the most, followed by using an unsupervised morpho- logical segmenter. The paper is structured as follows: Section 2 presents relevant related work. Section 3 discusses the linguistic challenges of translating Hebrew to English. Section 4 describes the different pre- processing techniques we study. And Section 5 presents our evaluation results. 2 Related Work A wide range of preprocessing techniques have been studied for a variety of language pairs requir- ing different treatments. Nießen and Ney (2004) studied the impact of various types of morpho- syntactic restructuring on German-English SMT and Popovi´c and Ney (2004) studied the effect of splitting words into stems and suffixes on SMT into English from Spanish, Catalan and Serbian. Their results show significant error reduction when stemming is used. Koehn and Knight (2003) com- pared different methods for compound splitting when translating from German to English. All of their methods improve SMT quality over a no- splitting baseline; however, the methods with the highest accuracy are not the best SMT performers. Lee (2004) investigated the use of automatic align- ment of POS tagged English and affix-stem seg- mented Arabic to determine whether affixes should be kept separate, deleted or reattached to stems. Her results show that morphological preprocess- ing helps, but only for the smaller corpora sizes she investigated. As size increases, the benefits di- minish. Goldwater and McClosky (2005) showed that incorporating various methods for specifying morphological …

Content truncated to 500 words