The CMU Machine Translation Systems at WMT 2014

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 142–149, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics The CMU Machine Translation Systems at WMT 2014 Austin Matthews Waleed Ammar Archna Bhatia Weston Feely Greg Hanneman Eva Schlinger Swabha Swayamdipta Yulia Tsvetkov Alon Lavie Chris Dyer∗ Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 USA ∗Corresponding author: cdyer@cs.cmu.edu Abstract We describe the CMU systems submitted to the 2014 WMT shared translation task. We participated in two language pairs, German–English and Hindi–English. Our innovations include: a label coarsening scheme for syntactic tree-to-tree transla- tion, a host of new discriminative features, several modules to create “synthetic trans- lation options” that can generalize beyond what is directly observed in the training data, and a method of combining the out- put of multiple word aligners to uncover extra phrase pairs and grammar rules. 1 Introduction The MT research group at Carnegie Mellon Uni- versity’s Language Technologies Institute partici- pated in two language pairs for the 2014 Workshop on Machine Translation shared translation task: German–English and Hindi–English. Our systems showcase our multi-phase approach to translation, in which synthetic translation options supple- ment the default translation rule inventory that is extracted from word-aligned training data. In the German–English system, we used our compound splitter (Dyer, 2009) to reduce data sparsity, and we allowed the translator to back off to translating lemmas when it detected case- inflected OOVs. We also demonstrate our group’s syntactic system with coarsened nonterminal types (Hanneman and Lavie, 2011) as a contrastive German–English submission. In both the German–English and Hindi–English systems, we used an array of supplemental ideas to enhance translation quality, ranging from lemma- tization and synthesis of inflected phrase pairs to novel reordering and rule preference features. 2 Core System Components The decoder infrastructure we used was cdec (Dyer et al., 2010). For our primary systems, all data was tokenized using cdec’s tokenization tool. Only the constrained data resources pro- vided for the shared task were used for training both the translation and language models. Word alignments were generated using both FastAlign (Dyer et al., 2013) and GIZA++ (Och and Ney, 2003). All our language models were estimated using KenLM (Heafield, 2011). Translation model parameters were chosen using MIRA (Eidelman, 2012) to optimize BLEU (Papineni et al., 2002) on a held-out development set. Our data was filtered using qe-clean (Denkowski et al., 2012), with a cutoff of two standard deviations from the mean. All data was left in fully cased form, save the first letter of each segment, which was changed to whichever form the first token more commonly used throughout the data. As such, words like The were lowercased at the beginning of segments, while words like Obama remained capitalized. Our primary German–English and Hindi– English systems were Hiero-based (Chiang, 2007), while our contrastive German–English sys- tem used cdec’s tree-to-tree SCFG formalism. Before submitting, we ran cdec’s implementa- tion of MBR on 500-best lists from each of our systems. For both language pairs, we used …

Content truncated to 500 words