A0593
Title: Modeling Eurovision 2025 songs with a flexible topic modeling approach
Authors: Alice Giampino - University of Milano-Bicocca (Italy) [presenting]
Roberto Ascari - University of Milano-Bicocca (Italy)
Sonia Migliorati - University of Milano Bicocca (Italy)
Abstract: Understanding the main topics of song lyrics in large-scale music competitions, such as the Eurovision Song Contest, is key to analyzing cultural trends, audience resonance, and artistic strategies. In 2025, with an increasing number of participating entries written in English, automatic methods for identifying underlying themes are especially relevant. However, lyrical data present significant challenges due to brevity, stylistic variability, and lexical diversity. Latent Dirichlet Allocation (LDA) has been widely used for uncovering latent topics by modeling word co-occurrence patterns. LDA's standard Dirichlet prior imposes only negative dependence among topics, limiting its ability to reflect nuanced thematic relationships or overlapping emotional tones in song lyrics. To overcome these constraints, a generalization of the LDA model is proposed through an extended flexible Dirichlet mixture prior tailored for topic distribution in musical texts. This enriched prior allows for positive correlations among topics, capturing common thematic clusters such as empowerment or political commentary that frequently co-occur across entries. The model maintains computational efficiency via conjugacy with the multinomial likelihood, enabling scalable inference through a collapsed Gibbs sampler. This approach yields a more expressive representation of lyrical themes, offering deeper insight into the collective voice of Eurovision 2025's songs and the cultural narratives they convey.
