Reproduce “Musical Word Embedding: Bridging the Gap between Listening Contexts and Music” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seungheon Doh, Jongpil Lee, Tae Hong Park, Juhan Nam
Researchers from KAIST and NYU developed word embeddings that bridge natural language descriptions of music contexts with musical compositions by systematically combining general text data (Wikipedia) and music-specific corpora. This approach facilitates more intuitive music search and recommendation.
Key insights
- Integrating music-specific data into word embedding training is crucial, as models trained with any music data consistently outperformed Wikipedia-only models.
- Hybrid approaches combining general and music-specific data effectively bridge concepts, performing best on balanced datasets containing both music and general context terms.
- A data augmentation technique involving word shuffling within track-specific clusters proved highly beneficial, improving embedding quality by capturing diverse word relationships.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.