FIG. 7 illustrates a block diagram of an example of the word representation generation module 520 and the sentence representation generation module 530 of FIG. 5. For purpose of discussion, the directed sentence graph 512 in the example of FIG. 6B is provided as the input to the neighbor determination module 522. As mentioned above, the neighbor determination module 522 may determine the set of neighbor nodes for each of the nodes in the sentence graph 512.
As shown in FIG. 7, the graph convolution module 526 includes graph aggregation layers 710, . . . , 720 to perform the graph convolution operation on each of the nodes 610. It is supposed that K graph aggregation layers are included and K may be greater than or equal to one. In the example of FIG. 7, a plurality of parallel graph aggregation layers 710, . . . , 720 may be configured in the graph convolution module 526 for the graph convolution operation of the respective nodes. The number of the parallel layers may be equal to the number of the node in the sentence graph. Of course, the convolution operations for one or more nodes may not be parallel.
The aggregation at each layer 710, . . . , 720 for a given node may be represented as follows:
hi(k)=fA(k)(hi(k-1),{hj(k-1),rij:vj∈![custom character]()
(vi)})??Equation (1)
