最新11,001 New Features for Statistical Machine Translation.pdfVIP

11,001 New Features for Statistical Machine Translation David Chiang and Kevin Knight Wei Wang USC Information Sciences Institute Language Weaver, Inc. 4676 Admiralty Way, Suite 1001 4640 Admiralty Way, Suite 1210 Marina del Rey, CA 90292 USA Marina del Rey, CA 90292 USA Abstract Many of the new features use syntactic informa- tion, and in particular depend on information that We use the Margin Infused Relaxed Algo- is available only inside a syntax-based translation rithm of Crammer et al. to add a large num- ber of new features to two machine transla- model. Thus they widen the advantage that syntax- tion systems: the Hiero hierarchical phrase- based models have over other types of models. based translation system and our syntax-based The models are trained using the Margin Infused translation system. On a large-scale Chinese- Relaxed Algorithm or MIRA (Crammer et al., 2006) English translation task, we obtain statistically instead of the standard minimum-error-rate training significant improvements of 1.5 B and or MERT algorithm (Och, 2003). Our results add 1.1 B, respectively. We analyze the im- to a growing body of evidence (Watanabe et al., pact of the new features and the performance 2007; Chiang et al., 2008) that MIRA is preferable to of the learning algorithm. MERT across languages an