Abstract
Keyphrase generation is the task of generating phrases (keyphrases) thatsummarize the main topics of a given document. The generated keyphrases can beeither present or absent from the text of the given document. While theextraction of present keyphrases has received much attention in the past, onlyrecently a stronger focus has been placed on the generation of absentkeyphrases. However, generating absent keyphrases is very challenging; even thebest methods show only a modest degree of success. In this paper, we propose anapproach, called keyphrase dropout (or KPDrop), to improve absent keyphrasegeneration. We randomly drop present keyphrases from the document and turn theminto artificial absent keyphrases during training. We test our approachextensively and show that it consistently improves the absent performance ofstrong baselines in keyphrase generation.