Lyrics-based Analysis and Classification of Music

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 620–631, Dublin, Ireland, August 23-29 2014. Lyrics-based Analysis and Classification of Music Michael Fell Computational Linguistics Saarland University D-66123 Saarbr¨ucken mic.fell@gmail.com Caroline Sporleder Computational Linguistics & Digital Humanities Trier University D-54286 Trier sporledc@uni-trier.de Abstract We present a novel approach for analysing and classifying lyrics, experimenting both with n- gram models and more sophisticated features that model different dimensions of a song text, such as vocabulary, style, semantics, orientation towards the world, and song structure. We show that these can be combined with n-gram features to obtain performance gains on three different classification tasks: genre detection, distinguishing the best and the worst songs, and determining the approximate publication time of a song. 1 Introduction The ever growing amount of music available on the internet calls for intelligent tools for browsing and searching music databases. Music recommendation and retrieval systems can aid users in finding music that is relevant to them. This typically requires automatic music analysis, e.g., classification according to genre, content or artist and song similarity. In addition, automatic music (and lyrics) analysis also offers potential benefits for musicology research, for instance, in the field of Sociomusicology where lyrics analysis is used to place a piece of music in its sociocultural context (Frith, 1988). In principle, both the audio signal and the lyrics (if any exist) can be used to analyse a music piece (as well as additional data such as album reviews (Baumann et al., 2004)). In this paper, we focus on the contribution of the lyrics. Songwriters deploy unique stylistic devices to build their lyrics. Some of those can be measured automatically and we hypothesise that these are distinctive enough to identify song classes such as genre, song quality and publication time. There is, in fact, strong empirical evidence that it is worthwhile to look deeper into lyrical properties when analysing and classifying music. For example, it has been shown that classifiers that incorporate textual features outperform audio-only classifiers on most classification tasks (Mayer et al., 2008a; Mayer and Rauber, 2011; Li and Ogihara, 2004). Lyrics are also often easier to obtain and process than audio data, and non-musicians, in particular, often rely strongly on lyrics when interacting with a music retrieval system (Baumann and Kl¨uter, 2002; Bainbridge et al., 2003). Moreover, lyrics do not only add semantic content, they can serve as an (easily observable) proxy for the melodic, structural and rhythmic properties of the audio signal. Melody and rhythm, for example, can often be traced in the stress pattern of the text (Nichols et al., 2009), while a song’s overall structure is reflected in the order of textual elements such as chorus, verse and bridge. Psychological research also provides evidence the audio and textual content are indeed processed independently in the brain and hence are complementary for our appreciation of a song (Besson et al., 1998). We extend previous research on lyrics-based song classification in two important ways: First, while earlier approaches …

Content truncated to 500 words