Extended Translation Models in Phrase-based Decoding

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 282–293, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. Extended Translation Models in Phrase-based Decoding Andreas Guta, Joern Wuebker, Miguel Grac¸a, Yunsu Kim, Hermann Ney Human Language Technology and Pattern Recognition Group RWTH Aachen University Aachen, Germany {surname}@cs.rwth-aachen.de Abstract We propose a novel extended translation model (ETM) to counteract some prob- lems in phrase-based translation: The lack of translation context when using single- word phrases and uncaptured dependen- cies beyond phrase boundaries. The ETM operates on word-level and augments the IBM models by an additional bilingual word pair and a reordering operation. Its implementation in a phrase-based decoder introduces translation and reordering de- pendencies for single-word phrases and dependencies across phrase boundaries. More, the model incorporates an explicit treatment of multiple and empty align- ments. Its integration outperforms com- petitive systems that include lexical and phrase translation models as well as hier- archical reordering models on 4 language pairs significantly by +0.7% BLEU on av- erage. Although simpler and using fewer dependencies, the ETM proves to be on par with 7-gram operation sequence mod- els (Durrani et al., 2013b). 1 Introduction The first successful steps in Statistical Ma- chine Translation have been taken by applying word-based models in a source-channel approach (Brown et al., 1990; Brown et al., 1993). Within this framework, the language model (LM) is esti- mated on monolingual n-grams, whereas the trans- lation models IBM-1 to IBM-5 are trained on bilingual data using word alignments. The dis- advantage of word-to-word translation is over- come by phrase-based translation (PBT) (Och et al., 1999; Zens et al., 2002; Koehn et al., 2003) and log-linear model combination (Och and Ney, 2002). 0 10 20 30 40 1 2 3 4 5 6 7 relative word frequency [%] phrase length De-En (source) De-En (target) En-Fr (source) En-Fr (target) Figure 1: Relative frequency of words dependent on the length of the phrase they were decoded with for the IWSLT dev2010 German→English and English→French corpora. Nevertheless, phrase-based translation models have several drawbacks: (i) Single-word phrases are translated without any context. (ii) Dependen- cies beyond phrase boundaries are not modelled at all. (iii) Phrase-based translation models have dif- ficulties modelling long-distance dependencies on source words with large gaps inbetween. The open question is how much actual lex- ical context is included in decoding. Figure 1 depicts the relative word frequencies plotted against the length of the phrase they were trans- lated with for the IWSLT 20141 German→English and English→French tasks. For English→French, more than 40% of the words are translated us- ing single- or two-word phrases, i.e. with a lexical context of at most one word. For the German→English task, more reorderings occur and lead to less monotone alignments. Here, even 60% of all words are translated with a lexical con- text of at most one single word and over 20% are translated without any lexical context at all. 1http://www.iwslt2014.org 282 We address this problem by developing two variants of extended translation …

Content truncated to 500 words