In this paper, we propose a novel method of reducing the size of translation model for hierarchical phrase-based machine translation systems. Previous approaches try to prune infrequent entries or unreliable entries based on statistics, but cause a problem of reducing the translation coverage.