Hierarchical Heterogeneous Graph Attention Network for Syntax-Aware Summarization

he task of summarization often requires a non-trivial understanding of the given text at the semantic level. In thiswork, we essentially incorporate the constituent structure intothe single document summarization via the Graph NeuralNetworks to learn the semantic meaning of tokens. Morespecifically, we propose a novel hierarchical heterogeneousgraph attention network over constituency-based parse treesfor syntax-aware summarization. This approach reflects psychological findings that humans will pinpoint specific selection patterns to construct summaries hierarchically. Extensiveexperiments demonstrate that our model is effective for boththe abstractive and extractive summarization tasks on fivebenchmark datasets from various domains. Moreover, furtherperformance improvement can be obtained by virtue of stateof-the-art pre-trained models.