DISCOURSE ANALYSIS OF LYRIC AND LYRIC-BASED CLASSIFICATION OF MUSIC

Abstract

Lyrics play an important role in the semantics and the structure of many pieces of music. However, while many existing lyric analysis systems consider each sentence of a given set of lyrics separately, lyrics are more naturally understood as multi-sentence units, where the relations between sentences is a key factor. Here we describe three experiments using discourse-based features, which describe the relations between different sentences within a set of lyrics, for common Music Information Retrieval tasks. We investigate genre recognition and release date estimation, and present evidence that collaborating discourse features allow for more accurate results than single-sentence lyric features do. Then we perform popularity analysis by comparing Billboard Magazine's "All-Time Top 100" songs and non-top songs, and show results indicating that the patterns of discourse features are different in the two song sets. These results suggest that discourse-based features are potentially useful for Music Information Retrieval tasks.