Generalized graphical abstractions for statistical machine translation
Karim Filali, Jeff Bilmes · 2007
We introduce a novel framework for the expression, rapid-prototyping, and evaluation of statistical machine-translation (MT) systems using graphical models. The framework extends dynamic Bayesian networks with multiple connected different-length streams, switching variable existence and dependence mechanisms, and constraint factors. We have implemented a new general-purpose MT training/decoding system in this framework, and have tested this on a variety of existing MT models (including the 4 IBM models), and some novel ones as well, all using Europarl as a test corpus. We describe the semantics of our representation, and present preliminary evaluations, showing that it is possible to prototype novel MT ideas in a short amount of time.