Abstract

n this paper, we present a keyphrase generation ap-proach using conditional Generative Adversarial Net-works (GAN). In our GAN model, the generator outputsa sequence of keyphrases based on the title and abstractof a scientific article. The discriminator learns to distin-guish between machine-generated and human-curatedkeyphrases. We evaluate this approach on standardbenchmark datasets. Our model achieves state-of-the-art performance in generation of abstractive keyphrasesand is also comparable to the best performing extractivetechniques. We also demonstrate that our method gener-ates more diverse keyphrases and make our implemen-tation publicly available