Extracting Pre-ordering Rules from Predicate-Argument Structures

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 29–37, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Extracting Pre-ordering Rules from Predicate-Argument Structures Xianchao Wu, Katsuhito Sudoh, Kevin Duh, Hajime Tsukada, Masaaki Nagata NTT Communication Science Laboratories, NTT Corporation 2-4 Hikaridai Seika-cho, Soraku-gun Kyoto 619-0237 Japan {wu.xianchao,sudoh.katsuhito,kevin.duh,tsukada.hajime,nagata.masaaki}@lab.ntt.co.jp Abstract Word ordering remains as an essen- tial problem for translating between lan- guages with substantial structural differ- ences, such as SOV and SVO languages. In this paper, we propose to automatically extract pre-ordering rules from predicate- argument structures. A pre-ordering rule records the relative position mapping of a predicate word and its argument phrases from the source language side to the tar- get language side. We propose 1) a linear- time algorithm to extract the pre-ordering rules from word-aligned HPSG-tree-to- string pairs and 2) a bottom-up algorithm to apply the extracted rules to HPSG trees to yield target language style source sen- tences. Experimental results are reported for large-scale English-to-Japanese trans- lation, showing significant improvements of BLEU score compared with the base- line SMT systems. 1 Introduction Statistical machine translation (SMT) suffers from an essential problem for translating between lan- guages with substantial structural differences, such as between English which is a subject-verb- object (SVO) language and Japanese which is a typical subject-object-verb (SOV) language. Numerous approaches have been consequently proposed to tackle this word-order problem, such as lexicalized reordering methods, syntax-based models, and pre-ordering ways. First, in or- der to overcome the shortages of traditional dis- tance based distortion models (Brown et al., 1993; Koehn et al., 2007), phrase dependent lexical- ized reordering models were proposed by several researchers (Tillman, 2004; Kumar and Byrne, 2005). Lexicalized reordering models learn local orientations (monotone or reordering) with proba- bilities for each bilingual phrase from the training data. For example, by taking lexical information as features, a maximum entropy phrase reordering model was proposed by Xiong et al. (2006). Second, syntax-based models attempt to solve the word ordering problem by employing syntac- tic structures. For example, linguistically syntax- based approaches (Galley et al., 2004; Liu et al., 2006) first parse source and/or target sentences and then learn reordering templates from the sub- tree fragments of the parse trees. In contrast, hier- archical phrase based translation (Chiang, 2005) is a formally syntax-based approach which can auto- matically extract hierarchical ordering rules from aligned string-string pairs without using additional parsers. These approaches have been proved to be both algorithmically appealing and empirically successful. However, most of current syntax-based SMT systems use IBM models (Brown et al., 1993) and hidden Markov model (HMM) (Vogel et al., 1996) to generate word alignments. These models have a penalty parameter associated with long distance jumps, and tend to misalign words which move far from the window sizes of their expected positions (Xu et al., 2009; Genzel, 2010). The third type tackles the word-order prob- lem in pre-ordering ways. Through the usage of a sequence of pre-ordering rules, the word or- der of an original …

Content truncated to 500 words