A Hybrid Modeling Approach for an Automated Lyrics-Rating System for Adolescents

The South Korean government operates human-based lyrics-rating systems to reduce adolescents’ exposure to inappropriate songs. In this study, we developed lyrics classification models for an automated lyrics-rating system for adolescents. There are two kinds of inappropriate lyrics for adolescents: (1) lyrics with inappropriate words and (2) lyrics with inappropriate content based on the semantic context. To tackle the first issue, we propose ( {text{logCD}}_{alpha } ) as a method for generating a lexicon of inappropriate words. It attained the highest performance among the lexicon-based filtering methods examined. Further, to deal with the second issue, we propose a hybrid classification model that combines ( {text{logCD}}_{alpha } ) with an RNN based model. The hybrid model composed of a ‘lexicon-checking model’ and a ‘context-checking model’ achieved the highest performance among all of the models examined, highlighting the effectiveness of combining the models to specifically target each of the two types of inappropriate lyrics.