Abstract

We show how eye-tracking corpora can be used to improve sentence compression models, presenting a novel multi-task learning algorithm based on multi-layer LSTMs.We obtain performance competitive with or better than state-of-the-art approaches.