Adversarial Rap Lyric Generation

In this paper, we design a novel text generation method called RapGAN for RLG (Rap Lyric Generation). RapGAN proposes two new schemes, PRO (Phrase Roll-Out) and AREGS (Attention Reward at Every Generation Step), to achieve efficient and effective RLG. The conventional text generation methods need to complete the entire-generated sentence in GAN (Generative Adversarial Network) training. RapGAN uses PRO as selective attention to reward only meaningful phrases in every generation step. Due to the varying length of phrases segmented by PRO, we employ an attention-based neural network that can capture the local feature of phrases and the semantics of global sentences as our discriminator. Therefore, we propose AREGS that reward the alignment weights of the attention-based discriminator with feature matching. Our method can provide different weights on different words to improve RLG's diversity, originality and fluency. Experimental results s how that our RapGAN outperforms the state-of-the-art RLG methods on many essential metrics of text generation. We also show the human evaluation to illustrate the relations between human perception and typical NLP metrics in diversity, originality and fluency. Our proposed method can facilitate real-world applications in content open RLG dataset in this paper.