Song authorship attribution: a lyrics and rhyme based approach

Abstract In this work, we apply authorship attribution to a large-scale corpus of song lyrics. As sub-category poetry, lyrics embody cultural elements as well stylistic attributes that are not present in prose. We draw attention special characteristics such repetitive sound patterns and rhyme based structures can be key ownership, opportunities cannot employed for other types text tweets, emails, blog posts. first create new balanced, data set 12,000 from 120 different artists. propose CNN models on lyric set, order use structural information included the lyrics, similarly image classification. conduct experiments at character sub-word levels mostly reflect positional information. addition, phoneme level features, which intrinsically involve repetitions, rhyme, meter, represent unique verse-based textual compositions. attempt discover idiosyncratic features consequently author genre associations by working with variants architectures have been successfully used classification domains. Our architecture choice results particular focus residing neighboring regions, since CNNs fail apprehend long term dependencies. Finally, empirically evaluate our comparison findings previous test research