Two-Neighbor Orientation Model with Cross-Boundary Global Contexts

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1264–1274, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Two-Neighbor Orientation Model with Cross-Boundary Global Contexts Hendra Setiawan, Bowen Zhou, Bing Xiang and Libin Shen IBM T.J.Watson Research Center 1101 Kitchawan Road Yorktown Heights, NY 10598, USA {hendras,zhou,bxiang,lshen}@us.ibm.com Abstract Long distance reordering remains one of the greatest challenges in statistical ma- chine translation research as the key con- textual information may well be beyond the confine of translation units. In this paper, we propose Two-Neighbor Orien- tation (TNO) model that jointly models the orientation decisions between anchors and two neighboring multi-unit chunks which may cross phrase or rule bound- aries. We explicitly model the longest span of such chunks, referred to as Max- imal Orientation Span, to serve as a global parameter that constrains under- lying local decisions. We integrate our proposed model into a state-of-the-art string-to-dependency translation system and demonstrate the efficacy of our pro- posal in a large-scale Chinese-to-English translation task. On NIST MT08 set, our most advanced model brings around +2.0 BLEU and -1.0 TER improvement. 1 Introduction Long distance reordering remains one of the great- est challenges in Statistical Machine Translation (SMT) research. The challenge stems from the fact that an accurate reordering hinges upon the model’s ability to make many local and global reordering decisions accurately. Often, such reordering decisions require contexts that span across multiple translation units.1 Unfortunately, previous approaches fall short in capturing such cross-unit contextual information that could be 1We define translation units as phrases in phrase-based SMT, and as translation rules in syntax-based SMT. critical in reordering. Specifically, the popular dis- tortion or lexicalized reordering models in phrase- based SMT focus only on making good local pre- diction (i.e. predicting the orientation of imme- diate neighboring translation units), while transla- tion rules in syntax-based SMT come with a strong context-free assumption, which model only the re- ordering within the confine of the rules. In this paper, we argue that reordering modeling would greatly benefit from richer cross-boundary contex- tual information We introduce a reordering model that incorpo- rates such contextual information, named the Two- Neighbor Orientation (TNO) model. We first iden- tify anchors as regions in the source sentences around which ambiguous reordering patterns fre- quently occur and chunks as regions that are con- sistent with word alignment which may span mul- tiple translation units at decoding time. Most no- tably, anchors and chunks in our model may not necessarily respect the boundaries of translation units. Then, we jointly model the orientations of chunks that immediately precede and follow the anchors (hence, the name “two-neighbor”) along with the maximal span of these chunks, to which we refer as Maximal Orientation Span (MOS). As we will elaborate further in next sections, our models provide a stronger mechanism to make more accurate global reordering decisions for the following reasons. First of all, we consider the orientation decisions on both sides of the anchors simultaneously, in contrast to existing …

Content truncated to 500 words