A Comparison of Syntactic Reordering Methods for English-German Machine Translation

Proceedings of COLING 2012: Technical Papers, pages 2043–2058, COLING 2012, Mumbai, December 2012. A Comparison of Syntactic Reordering Methods for English-German Machine Translation Jiˇrí Navrátil Karthik Visweswariah Ananthakrishnan Ramanathan IBM Research jiri@us.ibm.com, v-karthik@in.ibm.com, aramana2@in.ibm.com ABSTRACT We describe two methods for syntactic source reordering developed for English-German SMT. Both methods learn from bilingual data accompanied by automatic word alignments to reorder the source such that it resembles that of the target. While the first method is an extension of a parse-based algorithm and accommodates contextual triggers in the parse, the second method uses a linear feature-based cost model along with a Traveling Salesman Problem (TSP) solver to perform the reordering. Our results indicate that both methods lead to improvements in BLEU scores in both directions, English→German and German→English. Significant gains in human translation quality assessment are observed for German→English, however, no signifi- cant changes are observed in the human assessment for English→German. KEYWORDS: Parse-based reordering, TSP, English-German SMT. 2043 1 Introduction Language-specific word order differences have presented a long-standing challenge in the development of statistical machine translation (SMT) systems. Most mainstream SMT ap- proaches do incorporate word order information, either implicitly (e.g., as part of word phrases), or explicitly, e.g., by means of lexicalized distortion models. However, challenges remain, particularly in modeling long-range word movement. Furthermore, certain language pairs, more than others, exhibit particularly demanding reordering patterns. One such pair is English-German - the focus of this work. With its long-range verb movement and its use of separated verb prefixes, German and English word ordering can induce movements spanning an entire sentence. Section 1.1 describes some of these phenomena more in detail. Recent improvements in phrase-based SMT algorithms have included source word reorder- ing so as to resemble its target word order. A reordering typically uses rules obtained either manually, e.g., (Niessen and Ney, 2001; Collins et al., 2005), or derived from data by auto- matic means, e.g., (Xia and McCord, 2004; Rottmann and Vogel, 2007; Zhang et al., 2007; Crego and Habash, 2008; Niehues and Kolss, 2009). In this paper we further develop two methods first introduced in (Visweswariah et al., 2010, 2011) with the aim to address specific issues arising in English↔German SMT. In partic- ular, we extend the parse-based reordering introduced for a variety of language pairs in (Visweswariah et al., 2010) by creating rules capable of capturing essential contextual triggers in the parse tree hierarchy. For the second method we describe refinements of the feature- based reordering (Visweswariah et al., 2011), which corresponds to solving a Traveling Sales- man Problem (TSP). While the method described in (Visweswariah et al., 2011) carries over and performs well for German→English, for English→German we propose a method to inte- grate the English parse into the reordering model focusing on the issues in English→German reordering. We compare the performance of both methods against various baselines and across multiple evaluation domains for both English→German and German→English directions. Our results indicate that both methods achieve significant improvements in translation quality, as measured by automatic …

Content truncated to 500 words