Creative Tracklist Generation

Final Project for MIT 6.8610 Quantitative Methods for Natural Language Processing


This study proposes and evaluates models capable of generating song titles given the name of an album. Our goal is to contribute to the evolving landscape of NLP applications in the music industry, offering a creative tool for content generation and exploring novel possibilities for enhancing the artistic process. 

Our research adopts a multifaceted approach to address the challenge of generating song titles from album names, leveraging state-of-the-art language models for creative text generation.

  1. GPT-2 with Prompt Engineering:
    The initial phase employs the GPT-2 language model with prompt engineering to directly generate song titles. Training data consists of album name-track name tuples obtained from Last.FM APIs. The GPT-2 model is fine-tuned on this dataset, and various prompts are experimented with to enhance title diversity.

  2. T5-GPT-2 Narrative Framework:
    Recognizing the limitations of recurring themes in GPT-2, the second phase attempts to enrich the narrative context by leveraging a T5 model fine-tuned on commongen. With T5, we transform track listings into sentences, which serves as a narrative context mapped with each album’s name. A GPT-2 model is then trained on this dataset, and prompted to produce more varied and contextually rich outputs that are decoded back into a list of song titles.