Simultaneous Word-Morpheme Alignment for Statistical Machine Translation

Proceedings of NAACL-HLT 2013, pages 32–40, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Simultaneous Word-Morpheme Alignment for Statistical Machine Translation Elif Eyig¨oz Computer Science University of Rochester Rochester, NY 14627 Daniel Gildea Computer Science University of Rochester Rochester, NY 14627 Kemal Oflazer Computer Science Carnegie Mellon University PO Box 24866, Doha, Qatar Abstract Current word alignment models for statisti- cal machine translation do not address mor- phology beyond merely splitting words. We present a two-level alignment model that dis- tinguishes between words and morphemes, in which we embed an IBM Model 1 inside an HMM based word alignment model. The model jointly induces word and morpheme alignments using an EM algorithm. We eval- uated our model on Turkish-English parallel data. We obtained significant improvement of BLEU scores over IBM Model 4. Our results indicate that utilizing information from mor- phology improves the quality of word align- ments. 1 Introduction All current state-of-the-art approaches to SMT rely on an automatically word-aligned corpus. However, current alignment models do not take into account the morpheme, the smallest unit of syntax, beyond merely splitting words. Since morphology has not been addressed explicitly in word alignment models, researchers have resorted to tweaking SMT systems by manipulating the content and the form of what should be the so-called “word”. Since the word is the smallest unit of translation from the standpoint of word alignment models, the central focus of research on translating morphologi- cally rich languages has been decomposition of mor- phologically complex words into tokens of the right granularity and representation for machine transla- tion. Chung and Gildea (2009) and Naradowsky and Toutanova (2011) use unsupervised methods to find word segmentations that create a one-to-one map- ping of words in both languages. Al-Onaizan et al. (1999), ˇCmejrek et al. (2003), and Goldwater and McClosky (2005) manipulate morphologically rich languages by selective lemmatization. Lee (2004) attempts to learn the probability of deleting or merg- ing Arabic morphemes for Arabic to English trans- lation. Niessen and Ney (2000) split German com- pound nouns, and merge German phrases that cor- respond to a single English word. Alternatively, Yeniterzi and Oflazer (2010) manipulate words of the morphologically poor side of a language pair to mimic having a morphological structure similar to the richer side via exploiting syntactic structure, in order to improve the similarity of words on both sides of the translation. We present an alignment model that assumes in- ternal structure for words, and we can legitimately talk about words and their morphemes in line with the linguistic conception of these terms. Our model avoids the problem of collapsing words and mor- phemes into one single category. We adopt a two- level representation of alignment: the first level in- volves word alignment, the second level involves morpheme alignment in the scope of a given word alignment. The model jointly induces word and morpheme alignments using an EM algorithm. We develop our model in two stages. Our initial model is analogous to IBM Model 1: the …

Content truncated to 500 words