Unsupervised Multi-Granularity Summarization

Anonymous

Abstract: Text summarization is a user-preference based task. For one document, users often have different priorities for summary. Granularity level of the summary is a core component of these preferences. However, most existing studies focus solely on single-granularity scenarios, resulting in models that are limited to producing summaries with similar semantic coverage and are not customizable. In this paper, we propose the first unsupervised multi-granularity summarization framework, GranuSum. We regard events as basic semantic units of the original text and design a model that can take these events as anchors when generating summary. Meanwhile, by ranking these hint events and controlling the number of events, GranuSum is capable of generating summaries at different granularities in an unsupervised manner. We develop a testbed for the multi-granularity summarization task, including a new human-annotated benchmark GranuDUC where each document is paired with multiple summaries with different granularities. Extensive experiments on this benchmark and other large-scale datasets show that GranuSum substantially outperforms previous baselines. We also find that GranuSum exhibits impressive performance on conventional unsupervised abstractive summarization tasks via exploiting the event information, achieving new state-of-the-art results on three summarization datasets.