Comparing Automated Methods to Detect Explicit Content in Song Lyrics

Proceedings of Recent Advances in Natural Language Processing, pages 338–344, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_039 338 Comparing Automated Methods to Detect Explicit Content in Song Lyrics Michael Fell, Elena Cabrio, Michele Corazza, Fabien Gandon Universit´e Cˆote d’Azur, CNRS, Inria, I3S, France {firstname.lastname}@inria.fr Abstract The Parental Advisory Label (PAL) is a warn- ing label that is placed on audio recordings in recognition of profanity or inappropriate ref- erences, with the intention of alerting parents of material potentially unsuitable for children. Since 2015, digital providers – such as iTunes, Spotify, Amazon Music and Deezer – also fol- low PAL guidelines and tag such tracks as “ex- plicit”. Nowadays, such labelling is carried out mainly manually on voluntary basis, with the drawbacks of being time consuming and therefore costly, error prone and partly a sub- jective task. In this paper, we compare auto- mated methods ranging from dictionary-based lookup to state-of-the-art deep neural networks to automatically detect explicit contents in En- glish lyrics. We show that more complex mod- els perform only slightly better on this task, and relying on a qualitative analysis of the data, we discuss the inherent hardness and sub- jectivity of the task. 1 Introduction All content is not always appropriate for all ages and music is no exception. Content industries have been actively searching for means to help adults determine what is and is not appropriate for chil- dren. In USA, in 1985, the Recording Industry Association of America (RIAA) introduced the Parental Advisory label (PAL) in order to alert par- ents of content unsuitable for children because of profanity or inappropriate references1. PAL is “a notice to consumers that recordings identified by this mark may contain strong language or depic- tions of violence, sex or substance abuse”2 and that parental discretion is advised. In UK, the British Phonographic Industry (BPI) adds to this 1Parental Advisory https://en.wikipedia.org/ wiki/Parental_Advisory 2RIAA PAL https://www.riaa.com/ resources-learning/pal-standards/ list “racist, homophobic, misogynistic or other dis- criminatory language or behavior; or dangerous or criminal behavior”3. In the case of a song, the explicit logo is applied when the lyrics or content of a song matches one of these criteria, raising the problem of detecting and labelling explicit songs in a scalable way. Within the Natural Language Processing (NLP) community, there have been several efforts to deal with the problem of online abusive language de- tection, since the computational analysis of lan- guage can be used to quickly identify offenses and ease the removal of abusive messages. Sev- eral workshops (Park and Fung, 2017; Fiˇser et al., 2018) and evaluation campaigns (Fersini et al., 2018; Bosco et al., 2018; Wiegand et al., 2018) have been recently organized to discuss existing approaches to abusive language detection, propose shared tasks and foster the development of bench- marks for system evaluation. These have led to the creation of a number of datasets for abusive lan- guage detection in different languages, that have been shared within the NLP research community. The SemEval 2019 tasks HatEval (Basile et al., 2019) …

Content truncated to 500 words