questions about GCN when training and testing
hao920293897 opened this issue · comments
Each sentence have their own dependency tree, so the dependency trees are different among sentences. I wonder how to share and learn the parameters of GCNs among different sentences.
The GCN parameters include
W
matrix and biasb
which are shared across the dependency trees of all the sentences. Similar models [1, 2] can also be found in the literature.
Thank you . I wonder whether the shared parameters W matrix and bias b both require the same size of adjacency matrix A among different sentences. We note the number of words are different among different sentences.
Hi @hao920293897,
The parameters W
and b
are involved in the operation with input node features which are the same for all the documents. Changing the size of adjacency matrices doesn't require to change W
and b
. The adjacency matrix is involved only during the aggregation step in GCN. Please refer to Section 3 of the paper again. I think it will clear everything.
Thank you very much. Maybe I have understood this algorithm. And there is no doubt.
Great! Let me know if you have any other query.