Lyric Jumper: A Lyrics-Based Music Exploratory Web Service by Modeling Lyrics Generative Process

Each artist has their own taste for topics of lyrics such as “love” and “friendship.” Considering such artist’s taste brings new applications in music information retrieval: choosing an artist based on topics of lyrics and finding unfamiliar artists who have similar taste to a favorite artist. Although previous studies applied latent Dirichlet allocation (LDA) to lyrics to analyze topics, LDA was not able to capture the artist’s taste. In this paper, we propose a topic model that can deal with the artist’s taste for topics of lyrics. Our model assumes each artist has a topic distribution and a topic is assigned to each song according to the distribution. Our experimental results using a realworld dataset show that our model outperforms LDA in terms of the perplexity. By applying our model to estimate topics of 147,990 lyrics by 3,722 artists, we implement a web service called Lyric Jumper that enables users to explore lyrics based on the estimated topics. Lyric Jumper provides functions such as artist’s topic taste visualization and topic-similarity-based artist recommendation. We also analyze operation logs obtained from 12,353 users on Lyric Jumper and show the usefulness of Lyric Jumper especially in recommending topic-related phrases in lyrics.