The University of Maryland Statistical Machine Translation System for the Fifth Workshop on Machine Translation

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 72–76, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics The University of Maryland Statistical Machine Translation System for the Fifth Workshop on Machine Translation Vladimir Eidelman†, Chris Dyer†‡, and Philip Resnik†‡ †UMIACS Laboratory for Computational Linguistics and Information Processing ‡Department of Linguistics University of Maryland, College Park {vlad,redpony,resnik}@umiacs.umd.edu Abstract This paper describes the system we devel- oped to improve German-English transla- tion of News text for the shared task of the Fifth Workshop on Statistical Machine Translation. Working within cdec, an open source modular framework for ma- chine translation, we explore the benefits of several modifications to our hierarchical phrase-based model, including segmenta- tion lattices, minimum Bayes Risk de- coding, grammar extraction methods, and varying language models. Furthermore, we analyze decoder speed and memory performance across our set of models and show there is an important trade-off that needs to be made. 1 Introduction For the shared translation task of the Fifth Work- shop on Machine Translation (WMT10), we par- ticipated in German to English translation under the constraint setting. We were especially inter- ested in translating from German due to set of challenges it poses for translation. Namely, Ger- man possesses a rich inflectional morphology, pro- ductive compounding, and significant word re- ordering with respect to English. Therefore, we directed our system design and experimentation toward addressing these complications and mini- mizing their negative impact on translation qual- ity. The rest of this paper is structured as follows. After a brief description of the baseline system in Section 2, we detail the steps taken to improve upon it in Section 3, followed by experimental re- sults and analysis of decoder performance metrics. 2 Baseline system As our baseline system, we employ a hierarchical phrase-based translation model, which is formally based on the notion of a synchronous context-free grammar (SCFG) (Chiang, 2007). These gram- mars contain pairs of CFG rules with aligned non- terminals, and by introducing these nonterminals into the grammar, such a system is able to uti- lize both word and phrase level reordering to cap- ture the hierarchical structure of language. SCFG translation models have been shown to be well suited for German-English translation, as they are able to both exploit lexical information for and ef- ficiently compute all possible reorderings using a CKY-based decoder (Dyer et al., 2009). Our system is implemented within cdec, an ef- ficient and modular open source framework for aligning, training, and decoding with a num- ber of different translation models, including SCFGs (Dyer et al., 2010).1 cdec’s modular framework facilitates seamless integration of a translation model with different language models, pruning strategies and inference algorithms. As input, cdec expects a string, lattice, or context-free forest, and uses it to generate a hypergraph repre- sentation, which represents the full translation for- est without any pruning. The forest can now be rescored, by intersecting it with a language model for instance, to obtain output translations. The above capabilities …

Content truncated to 500 words