Compressed Heterogeneous Graph for Abstractive Multi-Document Summarization

Multi-document summarization (MDS) aims to generate a summary for number of related documents. We propose HGSum — an MDS model that extends encoder-decoder architecture incorporate heterogeneous graph represent different semantic units (e.g., words and sentences) the This contrasts with existing models which do not consider edge types graphs as such capture diversity relationships in To preserve only key information documents graph, uses pooling compress input graph. And guide learn compression, we introduce additional objective maximizes similarity between compressed constructed from ground-truth during training. is trained end-to-end standard cross-entropy objectives. Experimental results over Multi-News, WCEP-100, Arxiv show outperforms state-of-the-art models. The code our experiments available at: https://github.com/oaimli/HGSum.