University of Glasgow
Text augmentations from music artist biographies improve performance on genre classification, emotion recognition and music tagging
Abstract
dc:description.abstractMusic Information Retrieval (MIR) aims to develop automated systems to analyse, categorise, and recommend music. Multimodal learning plays an important but challenging role in MIR. Compared to more common multimodal settings such as speech transcription or image captioning, aligning audio with textual descriptions of music is particularly difficult because many musical concepts are abstract, culturally grounded, and not directly observable in the signal. Several approaches have explored incorporating textual information into MIR systems, leveraging lyrics, editorial metadata, or specialised audio-text corpora. However, these approaches exhibit a common trade-off between the semantic richness of the textual source and its availability and reproducibility, often due to coverage limitations or copyright restrictions. In this thesis, we investigate artist biographies as an alternative source of textual supervision. We show that artist biographies provide a scalable and reproducible source of high-level semantic information. To this end, we construct two corpora: one derived from web-scraped, community-authored content, and one generated using a large language model. Both are made available for reuse, with a documented reconstruction procedure for the web-scraped text, whose licensing prevents direct redistribution. We further introduce a unified evaluation framework that enables controlled comparisons of representations across multiple semantic MIR tasks and allows us to assess the contributions of textual, audio, and multimodal inputs, both in isolation and in combination. Experimental results show that incorporating artist biographies improves performance across most evaluated tasks. Relative to audio-only baselines, multimodal representations yield improvements of up to +38% in genre classification, +28% in emotion recognition, and +14% in music tagging, although one smaller dataset remains inconclusive due to limited training data. Additionally, we provide initial evidence that models can internalise the semantic information in biographies through generative pretraining, leading to improved performance even when textual inputs are unavailable at inference time. These findings establish artist biographies as an effective, scalable, and reproducible source of semantic supervision for multimodal MIR and highlight the broader potential of structured textual knowledge for enhancing music representation learning.
Degree
thesis:*- Level dc:type.qualificationlevel
- PhD
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Rossetto, Federico
Subjects
dc:subject × 3Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.86297