Minimum-error-rate training (MERT) is a bottleneck for current development in statistical machine translation (MT) because it has difficulty estimating more than a dozen or two parameters. I will present two classes of features that address deficiencies in the Hiero hierarchical phrase-based translation model but cannot practically be trained using MERT. Instead, we use the MIRA algorithm, introduced by Crammer et al and previously applied to MT by Watanabe et al. Building on their work, we show that by parallel processing and utilizing more of the parse forest, we can obtain results using MIRA that match those of MERT in terms of both translation quality and computational requirements. We then test the method on the new features: first, simultaneously training a large number of Marton and Resnik's soft syntactic constraints, and, second, introducing a novel structural distortion model based on a large number of features. In both cases we obtain significant improvements in translation performance over the baseline.

This talk represents joint work with Yuval Marton and Philip Resnik of the University of Maryland.