In this paper, by relabeling nodes generated during binarization of syntactic trees, contexts can be easily and systematically integrated. This not only helps to restructure syntactic trees to obtain smaller rules, that can be acquired and exploited for translation, also helps to determine which rules are most suitable for translation. By contextual binarization, high-quality translation could be easily generated from the contextual rules, if available; otherwise the translation just falls back on original syntax-based model without performance loss. Experimental results on the NIST Chinese-to-English corpus show promising improvements, the system applying contextual binarization outperforms over both the original syntax-based system and the original one with right binarization.