An Encoder-decoder Architecture with Graph Convolutional Networks for Abstractive Summarization
Qiao Yuan, Pin Ni, Junru Liu, Xiangzhi Tong, Hanzhe Lu, Gangmin Li, Steven Guan · 2021
We propose a single-document abstractive summarization system that integrates token relation into a traditional RNN-based encoder-decoder architecture. We employ pointer-wise mutual information to represent the token relation and adopt Graph Convolutional Networks (GCN) to extract token representation from the relation graph. In our experiment on Gigaword, we consider importing two kinds of structural information: token (node) representation from the relation graph. Also, we implement two kinds of GCNs, a spectral-based one and a spatial-based one, to extract structural information. The result shows that the spatial based GCN-enhanced model with node representation outperforms the classical RNN-based encoder-decoder model.