Comparison of the MT performance between the sequence-to-sequence NMT models and the tree-to-sequence NMT models with the same model sizes. We mark a score with † and ‡ to show the significant improvement over the Sequence-to-Sequence NMT model with reverse inputs and both Sequence-to-Sequence NMT models (p < 0.05), respectively. A statistical significance test was performed by using the bootstrap re-sampling method (Koehn 2004).