Abstractive Summarizers are Excellent Extractive Summarizers

In this paper, we explore the efficacy of modeling extractive summarization with an abstractive summarization system. We propose three novel inference algorithms for sequence-to-sequence models, evaluate them on established summarization benchmarks, and show that recent advancements in abstractive designs have enabled them to compete directly with extractive systems with custom extractive architectures. We show for the first time that a single model can simultaneously produce both state-of-the-art abstractive and extractive summaries, introducing a unified paradigm for summarization systems. Our results question fundamental concepts of extractive systems and pave the way for a new paradigm - generative modeling for extractive summarization.