The paper presents a statistical phrase-based translation model that uses *hierarchical phrases*, which are phrases containing subphrases. The model is formally a synchronous context-free grammar but is learned from bitext without syntactic information, making it formally syntax-based without linguistic commitment. The model improves translation accuracy by capturing larger-scale translations beyond the scope of individual phrases. Experiments using BLEU as a metric show that the hierarchical phrase-based model achieves a relative improvement of 7.5% over Pharaoh, a state-of-the-art phrase-based system. The model is trained on a word-aligned corpus and uses a log-linear model with features such as lexical weights, phrase penalties, and language model integration. The decoder is a CKY parser with beam search, and the system is evaluated on a Mandarin-to-English translation task. The hierarchical model outperforms the baseline system without additional training data, and adding a constituent feature does not provide a statistically significant improvement. Future work aims to incorporate more syntactic information and reduce grammar size to improve accuracy and efficiency.The paper presents a statistical phrase-based translation model that uses *hierarchical phrases*, which are phrases containing subphrases. The model is formally a synchronous context-free grammar but is learned from bitext without syntactic information, making it formally syntax-based without linguistic commitment. The model improves translation accuracy by capturing larger-scale translations beyond the scope of individual phrases. Experiments using BLEU as a metric show that the hierarchical phrase-based model achieves a relative improvement of 7.5% over Pharaoh, a state-of-the-art phrase-based system. The model is trained on a word-aligned corpus and uses a log-linear model with features such as lexical weights, phrase penalties, and language model integration. The decoder is a CKY parser with beam search, and the system is evaluated on a Mandarin-to-English translation task. The hierarchical model outperforms the baseline system without additional training data, and adding a constituent feature does not provide a statistically significant improvement. Future work aims to incorporate more syntactic information and reduce grammar size to improve accuracy and efficiency.