Relabeling Syntax Trees to Improve Syntax-Based Machine Translation Quality

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 240–247, New York, June 2006. c⃝2006 Association for Computational Linguistics Relabeling Syntax Trees to Improve Syntax-Based Machine Translation Quality Bryant Huang Language Weaver, Inc. 4640 Admiralty Way, Suite 1210 Marina del Rey, CA 90292 bhuang@languageweaver.com Kevin Knight Information Sciences Institute University of Southern California 4676 Admiralty Way Marina del Rey, CA 90292 knight@isi.edu Abstract We identify problems with the Penn Tree- bank that render it imperfect for syntax- based machine translation and propose methods of relabeling the syntax trees to improve translation quality. We develop a system incorporating a handful of relabel- ing strategies that yields a statistically sig- nificant improvement of 2.3 BLEU points over a baseline syntax-based system. 1 Introduction Recent work in statistical machine translation (MT) has sought to overcome the limitations of phrase- based models (Marcu and Wong, 2002; Koehn et al., 2003; Och and Ney, 2004) by making use of syntactic information. Syntax-based MT of- fers the potential advantages of enforcing syntax- motivated constraints in translation and capturing long-distance/non-contiguous dependencies. Some approaches have used syntax at the core (Wu, 1997; Alshawi et al., 2000; Yamada and Knight, 2001; Gildea, 2003; Eisner, 2003; Hearne and Way, 2003; Melamed, 2004) while others have integrated syn- tax into existing phrase-based frameworks (Xia and McCord, 2004; Chiang, 2005; Collins et al., 2005; Quirk et al., 2005). In this work, we employ a syntax-based model that applies a series of tree/string (xRS) rules (Gal- ley et al., 2004; Graehl and Knight, 2004) to a source language string to produce a target language phrase structure tree. Figure 1 exemplifies the translation process, which is called a derivation, from Chinese into English. The source string to translate ( � � �� � � � � �� � �� � �� � � .) is shown at the top left. Rule 1⃝ replaces the Chinese word � � �� � � (shaded) with the English NP-C police. Rule 2⃝ then builds a VP over the � � � NP-C � � � sequence. Next, � � � � � � is translated as the NP-C the gunman by rule 3⃝. Finally, rule 4⃝ combines the sequence of NP-C VP . into an S, denot- ing a complete tree. The yield of this tree gives the target translation: the gunman was killed by police . The Penn English Treebank (PTB) (Marcus et al., 1993) is our source of syntactic information, largely due to the availability of reliable parsers. It is not clear, however, whether this resource is suitable, as is, for the task of MT. In this paper, we argue that the overly-general tagset of the PTB is problematic for MT because it fails to capture important grammati- cal distinctions that are critical in translation. As a solution, we propose methods of relabeling the syn- tax trees that effectively improve translation quality. Consider the derivation in Figure 2. The output translation has two salient errors: determiner/noun number disagreement …

Content truncated to 500 words