unofficial code: GitHub - Cynwell/Text-Level-GNN: Text Level Graph Neural Network for Text Classification
1 Main Contributions
-
Propose to produce a text level graph for each input text instead of building a single corpus level graph and thus reduce much GPU consumption.
-
The representation of the same nodes and weights of edges are shared globally and can be upd in the text level graphs through a message passing mechanism.
2 Method
Build text graph
build the text graph with a small window size
Message Passing Mechanism
This is the massage passing mechanism of the proposed GNN
Then,the summation of all node vectors is used as the graph-level representaion.
3 Experiment
Overall performance
GPU consumption