Edinburgh’s Syntax-Based Systems at WMT 2015

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 199–209, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. Edinburgh’s Syntax-Based Systems at WMT 2015 Philip Williams1, Rico Sennrich1, Maria Nadejde1, Matthias Huck1, Philipp Koehn1,2 1School of Informatics, University of Edinburgh 2Center for Speech and Language Processing, The Johns Hopkins University Abstract This paper describes the syntax-based sys- tems built at the University of Edinburgh for the WMT 2015 shared translation task. We developed systems for all language pairs except French-English. This year we focused on: translation out of En- glish using tree-to-string models; contin- uing to improve our English-German sys- tem; and source-side morphological seg- mentation of Finnish using Morfessor. 1 Introduction This year’s WMT shared translation task featured five language pairs: English paired with Czech, Finnish, French, German, and Russian. We built syntax-based systems in both translation direc- tions for all language pairs except English-French. For English → German, we continued to de- velop our string-to-tree system, which has proven highly competitive in previous years. Additions this year included the use of a dependency lan- guage model, an alternative tuning metric, and soft source-syntactic constraints. For translation from English into Czech, Finnish, and Russian, we built STSG-based tree- to-string systems. Support for this type of model is a recent addition to the Moses toolkit. In previ- ous years, our systems have all used string-to-tree models and have only translated into English and German. For Finnish → English, we experimented with unsupervised morphological segmentation using Morfessor 2.0 (Virpioja et al., 2013). For the remaining systems (Czech → English, German → English, and Russian → English), our systems were essentially the same as last year’s (Williams et al., 2014) except for the addition of this year’s training data. 2 System Overview 2.1 Pre-processing The training data was pre-processed using scripts from the Moses toolkit. We first normalized the data using the normalize-punctuation.perl script then performed tokenization, parsing, and truecasing. To parse the English data, we used the Berkeley parser (Petrov et al., 2006; Petrov and Klein, 2007). To parse the German data, we used the ParZu dependency parser (Sennrich et al., 2013). 2.2 Word Alignment For word alignment we used either MGIZA++ (Gao and Vogel, 2008), a multi-threaded imple- mentation of GIZA++ (Och and Ney, 2003), or fast_align (Dyer et al., 2013). In preliminary experiments, we found that the tree-to-string sys- tems were particularly sensitive to the choice of word aligner, echoing a previous observation by Neubig and Duh (2014). See the individual tree- to-string system descriptions in Section 3. 2.3 Language Model We used all available monolingual data to train one interpolated 5-gram language model for each sys- tem. Using either lmplz (Heafield et al., 2013) or the SRILM toolkit (Stolcke, 2002), we first trained an individual language model for each of the supplied monolingual training corpora. These models all used modified Kneser-Ney smoothing (Chen and Goodman, 1998). We then interpolated the individual models using SRILM, providing the target-side of the system’s tuning set (Section 2.7) for …

Content truncated to 500 words