Modeling Global and Local Node Contexts for Text Generation from Knowledge Graphs

Recent graph-to-text models generate text from graph-based data using either\nglobal or local aggregation to learn node representations. Global node encoding\nallows explicit communication between two distant nodes, thereby neglecting\ngraph topology as all nodes are directly connected. In contrast, local node\nencoding considers the relations between neighbor nodes capturing the graph\nstructure, but it can fail to capture long-range relations. In this work, we\ngather both encoding strategies, proposing novel neural models which encode an\ninput graph combining both global and local node contexts, in order to learn\nbetter contextualized node embeddings. In our experiments, we demonstrate that\nour approaches lead to significant improvements on two graph-to-text datasets\nachieving BLEU scores of 18.01 on AGENDA dataset, and 63.69 on the WebNLG\ndataset for seen categories, outperforming state-of-the-art models by 3.7 and\n3.1 points, respectively.\n