The HuMus Project: Corpus Compilation and Multi-experiment Study of Hungarian Popular Music Lyrics

Recent advances in natural language processing (NLP) have enabled sophisticated analyses of textual data, including the lyrics of popular music. This study applies statistical NLP methods to a large corpus of Hungarian popular music lyrics to examine differences and similarities across 10 musical genres. We investigate whether genre classification can be achieved using lyrics alone through multilabel and pairwise classification experiments, employing multinomial naive Bayes and huBERT models. In addition, we assess textual complexity using two approaches, namely the Flesch–Kincaid formula and gzip compression ratio, exploring potential correlations with genre popularity. Sentiment analysis further provides a positivity ranking of genres, allowing us to confirm or challenge common assumptions about their thematic characteristics, such as the prevalence of negative sentiment in metal music. The findings are compared with prior studies on English-language lyrics, offering insights into cross-linguistic and cultural patterns in musical expression, and we hope that this work will pave the way for further similar research. The primary objective of the present study is to demonstrate the feasibility of constructing corpora of lyrical texts for languages that are not widely utilized, and of concomitantly accumulating metadata for these corpora. The overarching ambition of these endeavors is to enable the conduct of an extensive array of experiments, thereby advancing research in the field.