SloGAN: Character Level Adversarial Lyric Generation
Posted:
Aidan Cookson, Krish Kabra, and Auguste Hirth
Course: CS 263 - Natural Language Processing
Instructor: Prof. Kai-Wei Chang
Quarter: Spring 2020
Abstract
This paper explores the application of a deep learning models for the task of rap lyric generation. We first propose a generative LSTM character-level model that contrasts with recent word-level approaches. We rigorously compare the effectiveness of a character-level models with their word-level counterparts using BLEU, RhymeAnalyzer, and human authenticity prediction metrics. Secondly, we attempt to implement a SeqGAN for the rap lyric generation task in order to combat pitfalls of traditional maximum-likelihood estimation methods. Unfortunately, our SeqGAN model fails to converge. We provide reasons and possible solutions to improve the rap lyric generation task using adversarially trained networks.