The University of Cambridge Russian-English System at WMT13

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 200–205, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics The University of Cambridge Russian-English System at WMT13 Juan Pino Aurelien Waite Tong Xiao Adri`a de Gispert Federico Flego William Byrne Department of Engineering, University of Cambridge, Cambridge, CB2 1PZ, UK {jmp84,aaw35,tx212,ad465,ff257,wjb31}@eng.cam.ac.uk Abstract This paper describes the University of Cambridge submission to the Eighth Workshop on Statistical Machine Transla- tion. We report results for the Russian- English translation task. We use mul- tiple segmentations for the Russian in- put language. We employ the Hadoop framework to extract rules. The decoder is HiFST, a hierarchical phrase-based de- coder implemented using weighted finite- state transducers. Lattices are rescored with a higher order language model and minimum Bayes-risk objective. 1 Introduction This paper describes the University of Cam- bridge system submission to the ACL 2013 Eighth Workshop on Statistical Machine Transla- tion (WMT13). Our translation system is HiFST (Iglesias et al., 2009), a hierarchical phrase-based decoder that generates translation lattices directly. Decoding is guided by a CYK parser based on a synchronous context-free grammar induced from automatic word alignments (Chiang, 2007). The decoder is implemented with Weighted Finite State Transducers (WFSTs) using standard op- erations available in the OpenFst libraries (Al- lauzen et al., 2007). The use of WFSTs allows fast and efficient exploration of a vast translation search space, avoiding search errors in decoding. It also allows better integration with other steps in our translation pipeline such as 5-gram lan- guage model (LM) rescoring and lattice minimum Bayes-risk (LMBR) decoding (Blackwood, 2010). We participate in the Russian-English transla- tion shared task in the Russian-English direction. This is the first time we train and evaluate a sys- tem on this language pair. This paper describes the development of the system. The paper is organised as follows. Section 2 describes each step in the development of our sys- tem for submission, from pre-processing to post- processing and Section 3 presents and discusses results. 2 System Development 2.1 Pre-processing We use all the Russian-English parallel data avail- able in the constraint track. We filter out non Russian-English sentence pairs with the language- detection library.2 A sentence pair is filtered out if the language detector detects a different language with probability more than 0.999995 in either the source or the target. This discards 78543 sen- tence pairs. In addition, sentence pairs where the source sentence has no Russian character, defined by the Perl regular expression [\x0400-\x04ff], are discarded. This further discards 19000 sen- tence pairs. The Russian side of the parallel corpus is to- kenised with the Stanford CoreNLP toolkit.3 The Stanford CoreNLP tokenised text is additionally segmented with Morfessor (Creutz and Lagus, 2007) and with the TreeTagger (Schmid, 1995). In the latter case, we replace each token by its stem followed by its part-of-speech. This of- fers various segmentations that can be taken ad- vantage of in hypothesis combination: CoreNLP, CoreNLP+Morfessor and CoreNLP+TreeTagger. The English side of the parallel corpus is tokenised with a …

Content truncated to 500 words