Two-Stage Pre-ordering for Japanese-to-English Statistical Machine Translation

International Joint Conference on Natural Language Processing, pages 1062–1066, Nagoya, Japan, 14-18 October 2013. Two-Stage Pre-ordering for Japanese-to-English Statistical Machine Translation Sho Hoshino1 and Yusuke Miyao1,2 Katsuhito Sudoh3 and Masaaki Nagata3 1The Graduate University for Advanced Studies 2National Institute of Informatics {hoshino, yusuke}@nii.ac.jp 3NTT Communication Science Laboratories, NTT Corporation {sudoh.katsuhito, nagata.masaaki}@lab.ntt.co.jp Abstract We propose a new rule-based pre-ordering method for Japanese-to-English statistical machine translation that employs heuristic rules in two-stages. This two-stage frame- work contributes to experimental results that our method outperforms conventional rule-based methods in BLEU and RIBES. 1 Introduction Reordering is an important strategy in statistical machine translation (SMT) to achieve high qual- ity translation. While many reordering meth- ods often fail in long distance reordering due to computational complexity, a promising technol- ogy called pre-ordering (Xia and McCord, 2004; Collins et al., 2005) has been successful for dis- tant English-to-Japanese translation (Isozaki et al., 2010b). However, this strong effectiveness has not been shown for Japanese-to-English translation. In this paper, we propose a novel rule-based pre-ordering method for the Japanese-to-English translation. The method utilizes simple heuristic rules in two-stages1: the inter-chunk and intra- chunk levels. Thus the method can achieve more accurate reorderings in Japanese. The transla- tion experiments in patent domain showed that our method outperformed conventional rule-based methods, especially on the word reorderings. Our claims in this paper are summarized as follows: 1. The inter-chunk pre-ordering that relies on PAS analysis contributes to improvements in translation quality. 2. The intra-chunk pre-ordering which con- verts postpositional phrases into prepositional phrases further improves translation quality. 3. Thus, our two-stage framework is more effec- tive than other pre-ordering methods. 1In this paper, we refer to Japanese bunsetsu as a “chunk”, a grammatical and phonological unit consists of noun, verb, or adverb followed by dependents such as particles. 2 Related Work Japanese-to-English is challenging because the grammatical forms of the two languages are totally dissimilar. For instance, English is a head-initial language, and utilizes subject-verb-object (SVO) word orders, while Japanese is a pure head-final language, and utilizes subject-object-verb (SOV). Komachi et al. (2006) proposed a rule-based pre-ordering method to convert SOV into SVO via a PAS analyzer. This method pre-orders inter- chunk level word orders in a single-stage, via the PAS analyzer which produced dependency trees and tagged each S, O, and V label. Then SOV se- quences are converted into SVO. However, since the non-labeled words are left untouched, the ef- fectiveness of this method is limited to simple SOV labeled matrix sentences without multiple clauses. Katz-Brown and Collins (2008) proposed a two- stage rule-based pre-ordering method. In the first stage, SOV sequences are converted into SVO via the dependency analyzer. In the second stage, each chunk word order is naively reversed2. Neubig et al. (2012) proposed a statistical model that was capable of learning how to pre-order word sequences from human annotated or automatically generated alignment data. However, this method has very large computational complexity to model long distance reordering. 3 Two-stage Pre-ordering Method Here, we describe a …

Content truncated to 500 words