GoSum: extractive summarization of long documents by reinforcement learning and graph-organized discourse state

Summarizing extensive documents involves selecting sentences, with the organizational structure of document sections playing a pivotal role. However, effectively utilizing discourse information for summary generation poses a significant challenge, especially given the inconsistency between training and evaluation in extractive summarization. In this paper, we introduce GoSum, a novel extractive summarizer that integrates a graph-based model with reinforcement learning techniques to summarize long documents. Specifically, GoSum utilizes a graph neural network to encode sentence states, constructing a heterogeneous graph that represents each document at various discourse levels. The edges of this graph capture hierarchical relationships between different document sections. Furthermore, GoSum incorporates offline reinforcement learning, enabling the model to receive ROUGE score feedback on diverse training samples, thereby enhancing the quality of summary generation. On the two scientific article datasets PubMed and arXiv, GoSum achieved the highest performance among extractive models. Particularly on the PubMed dataset, GoSum outperformed other models with ROUGE-1 and ROUGE-L scores surpassing by 0.45 and 0.26, respectively.