New Parameterizations and Features for PSCFG-Based Machine Translation

We propose several improvements to the hierarchical phrase-based MT model of Chiang (2005) and its syntax-based extension by Zollmann and Venugopal (2006). We add a source-span variance model that, for each rule utilized in a probabilistic synchronous context-free grammar (PSCFG) derivation, gives a confidence estimate in the rule based on the number of source words spanned by the rule and its substituted child rules, with the distributions of these source span sizes estimated during training time. We further propose different methods of combining hierarchical and syntax-based PSCFG models, by merging the grammars as well as by interpolating the translation models. Finally, we compare syntax-augmented MT, which extracts rules based on targetside syntax, to a corresponding variant based on source-side syntax, and experiment with a model extension that jointly takes source and target syntax into account.