HHGraphSum: Hierarchical heterogeneous graph learning for extractive document summarization

Abstract

Extractive summarization aims to select important sentences from the document to generate a summary. However, current extractive document summarization methods fail to fully consider the semantic information among sentences and the various relations in the entire document. Therefore, a novel end-to-end framework named hierarchical heterogeneous graph learning for document summarization (HHGraphSum) is proposed in this paper. In this framework, a hierarchical heterogeneous graph is constructed for the whole document, where the representation of sentences is learnt by several levels of graph neural network. The combination of single-direction message passing and bidirectional message passing helps graph learning obtain effective relations among sentences and words. For capturing the rich semantic information, space–time collaborative learning is designed to generate the primary features of sentences which are enhanced in graph learning. For generating a less redundant and more precise summary, a LSTM based predictor and a blocking strategy are explored. Evaluations both on a single-document dataset and a multi-document dataset demonstrate the effectiveness of the HHGraphSum. The code of HHGraphSum is available on Github:https://github.com/Devin100086/HHGraphSum.