Google's Tensor2Tensor encodes positional information with each word. This positional encoding is a vector of sine function with different frequencies (fixed) or can be learned with the model. Our hypothesis is that dependency parse can add more helpful information and/or replace the current positional encoding scheme.
Our goal is to encode the tree address
or at least depth of each source word in a dependency parse, then compare the enhanced
model with the current one.