I'm currently trying to understand this paper but I struggle with the input into the NN. What I don't understand is what the input vectors should look like for the network described in b) in the image below.
As described in the paper the inputs are two sub-paths of a tree which are connected at a single node. Each node represents a feature vector, for example the word embedding of the word. My Question now is do I concatenate the vectors for two nodes that are at the same level in the tree? And if this is the case what do I do with vectors of sub-paths of unequal length?
To give an example lets say we have the following paths and the corresponding subtree of my dependency tree:
v_root --> v_l1 --> v_l2 --> v_l3 --> v_l4 v_root --> v_r1 --> v_r2 --> v_r3 v_root |--v_l1 |--v_l2 |--v_l3 |--v_l4 |--v_r1 |--v_r2 |--v_r3
The input-sequence is bottom up so I would assume, that the input for the tree looks like the table below. But if this is the case what about
v_l4, there is no
v_r4 or do I simply use a zero-vector where all values are 0 for
Timestep | Input-Vectors (concatenated) --------------------------------------- 0 | v_l4:?? 1 | v_l3:v_r3 2 | v_l3:v_r2 3 | v_l3:v_r1 4 | v_root:v_root