LyricsRadar: A Lyrics Retrieval System Based on Latent Topics of Lyrics

This paper presents a lyrics retrieval system called LyricsRadar that enables users to interactively browse song lyrics by visualizing their topics. Since conventional lyrics retrieval systems are based on simple word search, those systems often fail to reflect user’s intention behind a query when a word given as a query can be used in different contexts. For example, the word“tears”can appear not only in sad songs (e.g., feel heartrending), but also in happy songs (e.g., weep for joy). To overcome this limitation, we propose to automatically analyze and visualize topics of lyrics by using a well-known text analysis method called latent Dirichlet allocation (LDA). This enables LyricsRadar to offer two types of topic visualization. One is the topic radar chart that visualizes the relative weights of five latent topics of each song on a pentagon-shaped chart. The other is radar-like arrangement of all songs in a two-dimensional space in which song lyrics having similar topics are arranged close to each other. The subjective experiments using 6,902 Japanese popular songs showed that our system can appropriately navigate users to lyrics of interests.