From Composition to Compositionality: Discovering Reusable Structure in Polyphonic Music Embeddings

\emph{Composer}, \emph{composition}, \emph{compositional}, \emph{compositionality}: four words associated with the same Latin root \emph{componere}, ``to put together.'' We take this overlap literally and study Renaissance polyphony as a naturalistic testbed for compositional representation learning: composers build works by recombining reusable melodic units under constraints of composer, genre, and voice. We model polyphonic music as a multi-relational graph over interval $n$-grams, with sequential, vertical, and thematic relations capturing how musical units are combined melodically, contrapuntally, and motivically. Embeddings trained only from these relations, without attribute supervision, recover interpretable additive factors corresponding to externally documented stylistic categories. Held-out (composer, genre, voice) combinations are recoverable bidirectionally, and the real attribute--embedding correspondence outperforms 100 permutation baselines across reconstruction, cosine, and retrieval metrics. Finally, projecting out the genre subspace selectively reduces genre information while preserving much of the relational structure, suggesting that relational musical context can induce partially separable compositional representations.