Lyrics Segmentation: Textual Macrostructure Detection using Convolutions

Proceedings of the 27th International Conference on Computational Linguistics, pages 2044–2054 Santa Fe, New Mexico, USA, August 20-26, 2018. 2044 Lyrics Segmentation: Textual Macrostructure Detection using Convolutions Michael Fellp, Yaroslav Nechaevo, Elena Cabrion, Fabien Gandonm Université Côte d’Azur, CNRS, Inria, I3S, Francepnm Fondazione Bruno Kessler, University of Trentoo michael.fell@unice.frp, nechaev@fbk.euo elena.cabrio@unice.frn, fabien.gandon@inria.frm Abstract Lyrics contain repeated patterns that are correlated with the repetitions found in the music they accompany. Repetitions in song texts have been shown to enable lyrics segmentation – a fun- damental prerequisite of automatically detecting the building blocks (e.g. chorus, verse) of a song text. In this article we improve on the state-of-the-art in lyrics segmentation by applying a convolutional neural network to the task, and experiment with novel features as a step towards deeper macrostructure detection of lyrics. Title and Abstract in French Segmenter les paroles de chansons: détection par réseau de neurones convolutif d’une macrostructure textuelle Les paroles de chansons contiennent des passages qui se répètent et sont corrélés aux répétitions trouvé dans la musique qui les accompagne. Ces répétitions dans les textes de chansons ont montré leur utilité pour la segmentation des paroles qui est une étape préalable fondamentale dans la détection automatique des blocs de construction d’une chanson (ex. le refrain, les couplets). Dans cet article, nous améliorons l’état de l’art de la segmentation des paroles en concevant un réseau de neurones convolutif pour cette tâche et expérimentons de nouvelles caractéristiques pour aller vers une détection plus profonde de la macrostructure des paroles. 1 Introduction Among the seas of textual resources available online are the lyrics of songs that are very popular re- sources for professional, cultural and social applications. To support intelligent interactions with these resources (e.g. browsing, visualizing, synchronizing) one of the first needs is to access the structure of the lyrics (e.g. intro, verse, chorus). However, lyrics are essentially provided as flat text files without structural markup. In this article, we address the problem of textual macrostructure detection in lyrics. Lyrics encode an important part of the semantics of a song and a motivating scenario for this work is to improve structural clues that can be used by search engines handling large collections of lyrics. Ideally we would like to be able to detect these structures reliably in different music genres to support new search criteria such as “find all the songs where the chorus talks about freedom”. As a first step, structure segmentation could serve as a front end processor for music content analysis, since it enables a local description of each section rather than a coarse, global representation of the whole song (Cheng et al., 2009; Casey et al., 2008). Music structure discovery is a research field in Music Information Retrieval where the goal is to automatically estimate the temporal structure of a music track by analyzing the characteristics of its audio signal over time. However only a few works have addressed such task lyrics-wise (P. G. Mahedero et al., 2005; Watanabe et al., 2016; Baratè et al., …

Content truncated to 500 words