Content area
Abstract
More than 150 years ago, humans were debating whether algorithms or machines could produce original content. Ada Lovelace, who is credited to be for creating the world's first computer program alongside Charles Babbage, claimed the computer program had no power to author anything original and could only perform what was ordered of it when speaking of the analytical machine. Although this hypothesis has been contested in the many years since her letters have been published, many people still think of art and music as a domain dominated by the human brain as it requires a degree of creative skill.[1]
Much has been written about the creativity of humans. There is an abundance of works of arts that are thought of as only achievable through human ingenuity. As Albert Einstein wrote, "The true sign of intelligence is not knowledge but imagination." Most people wouldn't consider modern day computers to have imaginations or a even an awareness of the mind. These things are seemingly incomprehensible for computers and do not seem possible to produce through deterministic algorithms. However, this has not stopped programmers and researchers from creating machine learning algorithms that surprise its authors with the results they generate. The programmer is unable trace back how the algorithm produced its results from this process. Considering this, it might be time to re-evaluate what role computers can play in the creative space.
In this paper we will examine what role machine learning can play in the creation of new artistic and creative content, specifically song lyrics. I will include the methodology and data preparation I did to form the training dataset of song lyrics. After data cleanup, I will examine the songs lyrics and artists on a couple of features including musical genre, song length, number of songs written, repetition score, and Term Frequency{Inverse Document Frequency (TF-IDF) vectors. Finally, I will generate novel song lyrics from the GPT-2 model software package using a training subset of data. We can examine the product of the machine generated song lyrics and see how it compares to actual song lyrics.
