The Syntax Augmented MT (SAMT) System at the Shared Task for the 2007 ACL Workshop on Statistical Machine Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 216–219, Prague, June 2007. c⃝2007 Association for Computational Linguistics The Syntax Augmented MT (SAMT) System for the Shared Task in the 2007 ACL Workshop on Statistical Machine Translation Andreas Zollmann and Ashish Venugopal and Matthias Paulik and Stephan Vogel School of Computer Science, Carnegie Mellon University, Pittsburgh interACT Lab, University of Karlsruhe {ashishv,zollmann,paulik,vogel+}@cs.cmu.edu Abstract We describe the CMU-UKA Syntax Augmented Machine Translation system ‘SAMT’ used for the shared task “Machine Translation for European Lan- guages” at the ACL 2007 Workshop on Statistical Machine Translation. Following an overview of syn- tax augmented machine translation, we describe pa- rameters for components in our open-source SAMT toolkit that were used to generate translation results for the Spanish to English in-domain track of the shared task and discuss relative performance against our phrase-based submission. 1 Introduction As Chiang (2005) and Koehn et al. (2003) note, purely lexical “phrase-based” translation models suffer from sparse data effects when translating con- ceptual elements that span or skip across several source language words. Phrase-based models also rely on distance and lexical distortion models to rep- resent the reordering effects across language pairs. However, such models are typically applied over limited source sentence ranges to prevent errors in- troduced by these models and to maintain efficient decoding (Och and Ney, 2004). To address these concerns, hierarchically struc- tured models as in Chiang (2005) define weighted transduction rules, interpretable as components of a probabilistic synchronous grammar (Aho and Ull- man, 1969) that represent translation and reordering operations. In this work, we describe results from the open-source Syntax Augmented Machine Trans- lation (SAMT) toolkit (Zollmann and Venugopal, 2006) applied to the Spanish-to-English in-domain translation task of the ACL’07 workshop on statisti- cal machine translation. We begin by describing the probabilistic model of translation applied by the SAMT toolkit. We then present settings for the pipeline of SAMT tools that we used in our shared task submission. Finally, we compare our translation results to the CMU-UKA phrase-based SMT system and discuss relative per- formance. 2 Synchronous Grammars for SMT Probabilistic synchronous context-free grammars (PSCFGs) are defined by a source terminal set (source vocabulary) TS, a target terminal set (target vocabulary) TT , a shared nonterminal set N and pro- duction rules of the form X → ⟨γ, α, ∼, w⟩ where following (Chiang, 2005) • X ∈ N is a nonterminal • γ ∈ (N ∪TS)∗ : sequence of source nonterminals and terminals • α ∈ (N ∪ TT )∗ : sequence of target nonterminals and terminals • the count #NT(γ) of nonterminal tokens in γ is equal to the count #NT(α) of nonterminal tokens in α, • ∼: {1, . . . , #NT(γ)} → {1, . . . , #NT(α)} one- to-one mapping from nonterminal tokens in γ to nonterminal tokens in α • w ∈ [0, ∞) : nonnegative real-valued weight Chiang (2005) uses a single nonterminal cate- gory, Galley et al. (2004) use syntactic constituents for the PSCFG …

Content truncated to 500 words