Hierarchical graph summarization: leveraging hybrid information through visible and invisible linkage

Graph-based ranking algorithm has been recently exploited for summarization by using sentence-to-sentence relationships. Given a document set with linkage information to summarize, different sentences belong to different documents or clusters (either visible cluster via anchor texts or invisible cluster by semantics), which enables a hierarchical structure. It is challenging and interesting to investigate the impacts and weights of source documents/clusters: sentence from important ones are deemed more salient than the others. This paper aims to integrate three types of hierarchical linkage into traditional graph-based methods by proposing Hierarchical Graph Summarization (HGS). We utilize a hierarchical language model to measure the sentence relationships in HGS. We develop experimental systems to compare 5 rival algorithms on 4 instinctively different datasets which amount to 5197 documents. Performance comparisons between different system-generated summaries and manually created ones by human editors demonstrate the effectiveness of our approach in ROUGE metrics.