I wanted to implement a tree/pyramid structure for a NLP task, which would look similar to this figure
(Self-Adaptive Hierarchical Sentence Model by Zhao et. al.)
The elements of the next layer consist of their left and right child. I was now wondering if a 1D Convolution with (here) 5 Groups could be appropriate.
My concerns are, that I might not be able to get the overlapping triangles, i.e. here only 3 Groups with 2 elements are possible and next that the contribution of a single child might not be as what is depicted in the picture but would contribute somehow different.
So my actual question is: Can someone please help me understand groups for 1D Convolution and explain how elements in groups are chosen and how each group contributes to the output channels?
Thanks so much in advance!