A Deep-Learning based Approach for Automatic Lyric Generation

Tanmoy Debnath, Suvvari Sai Dileep

Abstract: Writing as a task is characteristic of humans, given its inherent requirement for creativity and grammatical abilities. However, with rapid advancements in artificial intelligence, there has been tremendous progress in automating writing tasks. In this paper, we study the effectiveness of a Long Short-Term Memory (LSTM) model, utilizing a Markov model, in automatically generating lyrics, a form of writing. We further analyze a pre-trained GPT-2 model in its performance of the same task and evaluate its results against those of the Markov-LSTM. For the evaluation, we leverage BLEU scores and assessments by humans. The results of both evaluations show that the Markov-LSTM model delivered better results than the pre-trained GPT-2.

Keywords: deep learning, long short term memory, lyric generation, GPT-2