Tree as a Pivot: Syntactic Matching Methods in Pivot Translation

Proceedings of the Conference on Machine Translation (WMT), Volume 1: Research Papers, pages 90–98 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics Tree as a Pivot: Syntactic Matching Methods in Pivot Translation Akiva Miura†, Graham Neubig‡,†, Katsuhito Sudoh†, Satoshi Nakamura† † Nara Institute of Science and Technology, Japan ‡ Carnegie Mellon University, USA miura.akiba.lr9@is.naist.jp gneubig@cs.cmu.edu sudoh@is.naist.jp s-nakamura@is.naist.jp Abstract Pivot translation is a useful method for translating between languages with little or no parallel data by utilizing parallel data in an intermediate language such as En- glish. A popular approach for pivot trans- lation used in phrase-based or tree-based translation models combines source-pivot and pivot-target translation models into a source-target model, as known as triangu- lation. However, this combination is based on the constituent words’ surface forms and often produces incorrect source-target phrase pairs due to semantic ambiguity in the pivot language, and interlingual differ- ences. This degrades translation accuracy. In this paper, we propose a approach for the triangulation using syntactic subtrees in the pivot language to distinguish pivot language words by their syntactic roles to avoid incorrect phrase combinations. Ex- perimental results on the United Nations Parallel Corpus show the proposed method gains in all tested combinations of lan- guage, up to 2.3 BLEU points.1 1 Introduction In statistical machine translation (SMT) (Brown et al., 1993), it is known that translation with mod- els trained on larger parallel corpora can achieve greater accuracy (Dyer et al., 2008). Unfor- tunately, large bilingual corpora are not readily available for many language pairs, particularly those that do not include English. One effective so- lution to overcome the scarceness of bilingual data is to introduce a pivot language for which paral- 1Code to replicate the experiments can be found at https://github.com/akivajp/wmt2017 [X1] dossier [X2]� [X1] enregistrer [X2]� [X2] [X1] 记录� [X1] 记录 [X2]� [X1] record [X2]� (a) Standard triangulation method matching phrases NP� NP� DT� record� NP� NN� [X1]� [X2]� [X1] dossier [X2]� VP� VP� TO� record� NP� VB� [X1]� [X2]� [X1] enregistrer [X2]� [X2] [X1] 记录� [X1] 记录 [X2]� (b) Proposed triangulation method matching subtrees Figure 1: Example of disambiguation by parse subtree matching (Fr-En-Zh), [X1] and [X2] are non-terminals for sub-phrases. lel data with the source and target languages exists (de Gispert and Mariño, 2006). Among various methods using pivot languages, one popular and effective method is the triangu- lation method (Utiyama and Isahara, 2007; Cohn and Lapata, 2007), which first combines source- pivot and pivot-target translation models (TMs) into a source-target model, then translates using this combined model. The procedure of triangu- lating two TMs into one has been examined for different frameworks of SMT and its effectiveness has been confirmed both in Phrase-Based SMT (PBMT) (Koehn et al., 2003; Utiyama and Isahara, 2007) and in Hierarchical Phrase-Based SMT (Hi- ero) (Chiang, 2007; Miura et al., 2015). How- ever, word sense ambiguity and interlingual dif- ferences of word usage cause difficulty in accu- rately learning correspondences between source and target phrases, and thus the accuracy obtained by triangulated …

Content truncated to 500 words