ChoCo: a Chord Corpus and a Data Transformation Workflow for Musical Harmony Knowledge Graphs

Abstract

Various disconnected chord datasets are currently available for music analysis and information retrieval, but they are often limited by either their size, non-openness, lack of timed information, and interoperability. Together with the lack of overlapping repertoire coverage, this limits cross-corpus studies on harmony over time and across genres, and hampers research in computational music analysis (chord recognition, pattern mining, computational creativity), which needs access to large datasets. We contribute to address this gap, by releasing the Chord Corpus (ChoCo), a large-scale dataset that semantically integrates harmonic data from 18 different sources using heterogeneous representations and formats (Harte, Leadsheet, Roman numerals, ABC, etc.). We rely on JAMS (JSON Annotated Music Specification), a popular data structure for annotations in Music Information Retrieval, to represent and enrich chord-related information (chord, key, mode, etc.) in a uniform way. To achieve semantic integration, we design a novel ontology for modelling music annotations and the entities they involve (artists, scores, etc.), and we build a 30M-triple knowledge graph, including 4 K+ links to other datasets (MIDI-LD, LED).

When in Rome: A Meta-corpus of Functional Harmony

“…We note with enthusiasm certain complementary, recent efforts to integrate all of the above, including this WiR meta-corpus. Notable here is the 'Choco' chord corpus (Berardinis et al (2023), https://github.com/ smashub/choco). This kind of effort effectively 'raises the stakes': the gain in increased scale is counterbalanced by the corresponding detractions of handling such disparate repertoire, syntaxes, and analyses into a single system.…”