To have linguistic tree structures in statistical machine translation?

Statistical approaches are dominating the MT field currently, as they fit the non-deterministic characteristic of NLP naturally making the system implementation more manageable by shifting the complexity control mechanism from human to the computer, and ensure global optimization (over training data) by jointly considering all the data during training stage. Compared to not-linguistically-motivated approaches, linguistically-motivated ones possess some advantages. With those advantages, it is expected that linguistic-tree-based approaches should outperform those not-linguistically-motivated ones. In linguistically-motivated approaches, the hierarchical tree structure is usually adopted for handling long-distance dependency and capturing large translation scope that is bigger than a few consecutive words. We believe that the linguistic-tree-structure is still the solution for handling long-distance dependency within the context and performing large-structures transformation between two languages. It would be required in statistical MT approaches in the long run. However, the compositionality issue mentioned above should be carefully handled in the mechanism for unsupervised learning from the bilingual corpus.