Using ACE XML 2.0 to Store and Share Feature, Instance and Class Data for Musical Classification

Jessica A F Thompson, Cory McKay, John Ashley Burgoyne, Jessica Thompson 0001, Ichiro FujinagaIsmir 2009

Introduction

In the domain of Music Information Retrieval (MIR), the lack of standardized data formats has long hindered collaborative research. Many laboratories rely on proprietary "in-house" formats or general-purpose tools like Weka’s ARFF, which often fail to capture the nuances of musical data. These legacy formats struggle with multi-label classifications (e.g., a track belonging to both "Jazz" and "Blues"), lack comprehensive metadata, and cannot represent complex hierarchical relationships between categories. ACE XML 2.0 addresses this "data-sharing crisis" by providing a modular, standardized framework for storing everything from raw feature vectors to intricate class ontologies.

Methodology: A Modular Framework

ACE XML 2.0 introduces a suite of four distinct XML-based file formats designed to be used independently or in combination:

  1. Feature Value Files: These store the actual data points extracted from audio or symbolic music. The format supports diverse structures, including sparse arrays and feature vectors of varying dimensions.
  2. Feature Description Files: These serve as metadata, detailing how features were calculated, their dimensionality, and any pre-processing steps applied.
  3. Instance Label Files: These store ground-truth or predicted classifications. A key innovation is the support for weighted multi-class membership and the ability to associate labels with specific temporal windows within a track.
  4. Class Ontology Files: These define the relationships between categories, such as sub-genre hierarchies or artist influences.
ACE XML 2.0 Feature Value file excerpt showing support for single values and multi-dimensional arrays.

The framework also facilitates links to external Semantic Web resources via RDF-like triples, integrating MIR data into the broader web of linked data.

Discussion and Key Insights

The design of ACE XML 2.0 emphasizes the separation of concerns. By decoupling feature values from their descriptions and labels, researchers can reuse feature sets across multiple classification tasks (e.g., genre, mood, or instrumentation) simply by swapping the Label and Ontology files.

Furthermore, the framework embraces the inherent ambiguity of music. Rather than forcing a single label onto a track, ACE XML 2.0 allows for weighted multi-labeling, enabling models to express nuanced classifications (e.g., a track being 70% "Folk" and 30% "Rock"). The use of XML ensures human-readability and long-term data portability, avoiding the obsolescence risks associated with binary formats.

ACE XML 2.0 Class Ontology file illustrating weighted relationships between musical artists.

Implementation and Future Directions

ACE XML 2.0 has been successfully integrated into the jMIR software suite, supporting major research datasets such as Codaich, Bodhidharma MIDI, and SAC. The Class Ontology format, in particular, allows machine learning models to utilize "near-miss" penalties based on defined relationships (e.g., mistaking a student for their teacher), providing a more sophisticated evaluation than binary correct/incorrect metrics.

Future work focuses on expanding language support beyond Java to Python, C++, and Matlab. Additionally, the authors aim to transition from DTDs to more robust XSD or Relax NG schemas to enhance data validation and eventually establish a centralized global repository for sharing musical features and annotations.

Summary updated

BibTeX
@inproceedings{thompson2009using,
  author = {Jessica A F Thompson and Cory McKay and John Ashley Burgoyne and Jessica Thompson and Ichiro Fujinaga},
  title = {{Using ACE XML 2.0 to Store and Share Feature, Instance and Class Data for Musical Classification}},
  booktitle = {Ismir},
  year = {2009},
}
Figure