Muzeeglot : annotation multilingue et multi-sources d'entités musicales à partir de représentations de genres musicaux (Muzeeglot : cross-lingual multi-source music item annotation from music genre embeddings)

Elena V. Epure, Guillaume Salha, Félix Voituret, Marion Baranes, Romain HennequinJEP-TALN-RECITAL (Demos) 2020 · OpenReview

Abstract

Elena V. Epure, Guillaume Salha, Félix Voituret, Marion Baranes, Romain Hennequin. Actes de la 6e conférence conjointe Journées d'Études sur la Parole (JEP, 33e édition), Traitement Automatique des Langues Naturelles (TALN, 27e édition), Rencontre des Étudiants Chercheurs en Informatique pour le Traitement Automatique des Langues (RÉCITAL, 22e édition). Volume 4 : Démonstrations et résumés d'articles internationaux.

2020.

Introduction

Music genre annotation is a critical component of music streaming platforms, underpinning recommendation engines and catalog organization. However, these annotations are often inconsistent across different languages and taxonomies. A genre like "heavy metal" remains "heavy metal" in French rather than being translated, and different cultures may categorize the same artist differently based on local musicological perspectives.

To address these challenges, researchers at Deezer developed Muzeeglot, an unsupervised system and web interface designed to harmonize music genre annotations across languages and data sources. Muzeeglot maps genres using semantic embeddings, allowing for accurate genre prediction in a target language based on annotations from entirely different source taxonomies.

Methodology

The Muzeeglot pipeline integrates general-purpose natural language processing with domain-specific music knowledge through a multi-step process:

  1. Initial Embeddings: The system utilizes multilingual fastText word embeddings, representing words as vectors based on their usage in large text corpora like Wikipedia.
  2. Knowledge Graph Integration: To correct for the "music gap"—where general embeddings might associate "house" with buildings rather than music—the authors incorporate a Knowledge Graph from DBpedia. This graph defines musicological relationships such as sub-genre or fusion categories.
  3. Retrofitting: Using a retrofitting algorithm, the initial word vectors are adjusted. Vectors for related genres (e.g., "Jazz" and "Bebop") are pulled closer together in the embedding space, regardless of their proximity in the original text corpus.
  4. Unsupervised Inference: To predict genres for a target language, the system calculates the average cosine similarity between the embeddings of source-language genres and the available genres in the target taxonomy.

The system was evaluated using a dataset of over 63,000 music entities (tracks, artists, and albums) from DBpedia, spanning six languages: English, Spanish, French, Japanese, Dutch, and Czech.

Results and Discussion

The core innovation of Muzeeglot is treating genre alignment as a vector-space proximity problem rather than a literal translation task. This approach allows the system to handle genres that lack direct linguistic equivalents. By "nudging" general models with structured domain knowledge, the researchers significantly improved the relevance of the resulting embeddings.

The system's effectiveness is demonstrated through a case study of the band Pink Floyd. By analyzing source annotations in English and Japanese, Muzeeglot predicted the top 10 French genre tags.

The Muzeeglot interface displaying source genres in English and Japanese for the band Pink Floyd.

The system successfully predicted relevant tags such as Rock psychédélique and Rock progressif. When validated against actual French annotations in DBpedia, the top predictions showed high coincidence, indicated by green highlights in the interface.

The Muzeeglot interface showing predicted French genres for Pink Floyd, with green highlights indicating matches with ground-truth DBpedia data.

The Muzeeglot interface, built with VueJS and Plotly, provides a 3D visualization of these genre clusters, making the latent embedding space interpretable for musicologists and data scientists.

Conclusion

Muzeeglot offers a robust framework for cross-lingual genre annotation without the need for manually labeled parallel datasets. While currently dependent on the breadth of the DBpedia knowledge graph, the system demonstrates the power of combining unsupervised embeddings with domain-specific retrofitting. Future work may involve expanding the system to more languages and leveraging harmonized genres to improve cold-start recommendations for global audiences.

Tech Stack Note: The Muzeeglot backend is powered by Python 3, FastAPI, and Redis, while the frontend utilizes VueJS, Vuetify, and WebGL for 3D rendering. The code is available on Deezer's GitHub.

Summary updated

Abstract

Elena V. Epure, Guillaume Salha, Félix Voituret, Marion Baranes, Romain Hennequin. Actes de la 6e conférence conjointe Journées d'Études sur la Parole (JEP, 33e édition), Traitement Automatique des Langues Naturelles (TALN, 27e édition), Rencontre des Étudiants Chercheurs en Informatique pour le Traitement Automatique des Langues (RÉCITAL, 22e édition). Volume 4 : Démonstrations et résumés d'articles internationaux. 2020.
BibTeX
@inproceedings{epure2020muzeeglot,
  author = {Elena V. Epure and Guillaume Salha and Félix Voituret and Marion Baranes and Romain Hennequin},
  title = {{Muzeeglot : annotation multilingue et multi-sources d'entités musicales à partir de représentations de genres musicaux (Muzeeglot : cross-lingual multi-source music item annotation from music genre embeddings)}},
  booktitle = {JEP-TALN-RECITAL (Demos)},
  year = {2020},
  url = {https://openreview.net/forum?id=cYGKKyR2sgX},
}
Figure