Скачать или смотреть SCSE-GSC | SLS Talk #27 | Abstractive Summarization? by Mathieu Ravaut

Abstract: Sequence-to-sequence deep neural networks have achieved great progress in abstractive summarization, especially since the rise of transfer learning with Transformer-based pre-trained models. However, when decoding the model to generate a summary, quality may vary drastically between summary candidates. Recently, a few methods have been proposed to learn to select a better summary output through a second-stage of learning, most notably with a ranking loss, or by explicitly training another model for re-ranking. In this paper, we explore a completely new approach to second-stage abstractive summarization, which consists in fusing together entire summary candidates. Our approach improves on the base summary candidates, and presents interesting properties.

Paper: https://arxiv.org/abs/2210.08779