Alignment symmetrisation optimization targeting phrase pivot statistical machine translation

Alignment Symmetrization Optimization Targeting Phrase Pivot Statistical Machine Translation Ahmed El Kholy and Nizar Habash Center for Computational Learning Systems, Columbia University 475 Riverside Drive New York, NY 10115 {akholy,habash}@ccls.columbia.edu Abstract An important step in mainstream statisti- cal machine translation (SMT) is combin- ing bidirectional alignments into one align- ment model. This process is called sym- metrization. Most of the symmetrization heuristics and models are focused on di- rect translation (source-to-target). In this paper, we present symmetrization heuristic relaxation to improve the quality of phrase- pivot SMT (source-[pivot]-target). We show positive results (1.2 BLEU points) on Hebrew-to-Arabic SMT pivoting on En- glish. 1 Introduction One of the main issues in statistical machine trans- lation (SMT) is the scarcity of parallel data for many language pairs especially when the source and target languages are morphologically rich. A common SMT solution to the lack of parallel data is to pivot the translation through a third language (called pivot or bridge language) for which there exist abundant parallel corpora with the source and target languages. The literature covers many pivoting techniques. One of the best performing techniques, phrase pivoting (Utiyama and Isahara, 2007), builds an induced new phrase table between the source and target. Our effort in this paper is based on phrase piv- oting. We focus on word alignment to improve translation quality. Word alignment is an essen- tial step in building an SMT system. The most commonly used alignment models, such as IBM Model serial (Brown et al., 1993) and HMM (Och and Ney, 2003), all assume one-to-many align- ments. However, the target is to produce a many- to-many word alignment model. A common prac- tice solution in most state-of-the-art MT systems c⃝ 2014 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. is to create two sets of one-to-many word align- ments (bidirectional alignments), source-to-target and target-to-source, and then combine the two sets to produce the final many-to-many word align- ment model. This combination process is called “Symmetrization”. In this paper, we propose a symmetrization re- laxation method targeting phrase-pivot SMT. Un- like the typical symmetrization methods, the pro- cess is carried out as an optimization for phrase- pivot SMT and eventually increase the matching on the pivot phrases. We show positive results (1.2 BLEU points) on Hebrew-Arabic phrase-pivot SMT (pivoting through English). Next, we briefly present some background in- formation on symmetrization (Section 2) and dis- cuss previous related work in Section 3. This is followed by our symmetrization approach in Sec- tion 4. We present our experimental results in Sec- tion 5. 2 Background In this section, we briefly describe different sym- metrization heuristics. We then explain how sym- metrization affects phrase extraction and discuss the motivation for our approach. 2.1 Symmetrization Heuristics The simplest approach is to merge the two di- rectional alignment functions using a symmetriza- tion heuristic to produce a many-to-many align- ment matrix (Och et al., 1999; Och and Ney, 2003; Koehn et al., 2003). …

Content truncated to 500 words