A study of text summarization with graph attention networks

Abstract

This study aimed to leverage graph information, particularly Rhetorical Structure Theory (RST) and Co-reference (Coref) graphs, to enhance the performance of our baseline sum- marization models. Specifically, we experimented with a Graph Attention Network archi- tecture to incorporate graph information. However, this architecture did not enhance the performance. Subsequently, we used a simple Multi-layer Perceptron architecture, which improved the results in our proposed model on our primary dataset, CNN/DM. Addition- ally, we annotated XSum dataset with RST graph information, establishing a benchmark for future graph-based summarizing models. This secondary dataset posed multiple chal- lenges, revealing both the merits and limitations of our models.