Vietnamese Multidocument Summarization Using Subgraph Selection-Based Approach with Graph-Informed Self-attention Mechanism

In Multi-Document Summarization (MDS), the exceedingly large input length is a significant difficulty. In this paper, we designed a graph-based model with Graph-informed Self-attention to capture intersentence and inter-document relations in graph. Our models represent the MDS task as a sub-graph selection problem, where the candidate summaries are its subgraphs and the source documents are thought of as a similarity a graph. We add a pair of weight matrix learning parameters that represent the relationship between sentences and the relationship between documents in the cluster, thereby seeing the close relationship and the role of relation types on the graph. Empirical results on the MultiNews and Vietnamese AbMuSu dataset demonstrate that our approach substantially outperforms some baseline models.