The dominant practice of statistical machine translation (SMT) uses the same Chinese word segmentation specification in both alignment and translation rule induction steps in building Chinese-English SMT system, which may suffer from a suboptimal problem that word segmentation better for alignment is not necessarily better for translation.