BottleSum: Unsupervised and Self-supervised Sentence Summarization using the Information Bottleneck Principle

EMNLP 2019

by Peter West, Ari Holtzman, Jan Buys, Yejin Choi

Information Bottleneck gives a more natural path to information selection than autoencoding


We use the Informaiton Bottleneck Principle for completely unsupervised sentence summarization. We release 2 methods:

  1. BottleSumEx: An extractive approach utilizing pretrained language models. This works right out of the box!
  2. BottleSumSelf: A self-supervised abstractive approach utilizing BottleSumEx for example generation.
Our models are simple to use, leverage advantages of large pretrained language models, and offer an exciting step in unsupervised NLP with information theory.

Read the paper!


View some examples!

Available soon!


Available soon on Github!


Get in touch!