TL;DR
This paper introduces a novel approach to analyzing musical harmony by leveraging Description Logics (DL) and the OWL 2 RL profile. By formally representing music theory concepts, the authors enable automated harmonic analysis of chord progressions, offering a more robust and extensible alternative to existing methods.
What Problem Does This Paper Address?
Algorithmic analysis of music, particularly in areas like recommender systems and music production tools, often relies on signal processing and machine learning applied to audio data. While symbolic music representations (like MIDI or MusicXML) are abundant and well-suited for theoretical analysis, they are underutilized in current MIR research. Specifically, analyzing the functional harmony of music – how chords progress and create tension/resolution – is a complex task. Existing approaches, like HarmTrace, often use custom grammars and may not fully capture the nuances of tonal and modal harmony. This paper aims to bridge this gap by using knowledge representation techniques to provide a formal, semantically rich, and extensible framework for functional harmonic analysis.
Methods
The core of this work is the development of a Description Logic (DL) ontology that formalizes the theory of tonal and modal harmony. The authors translate music theory concepts, such as scales, modes, chords, and their functions (tonic, dominant, subdominant), into DL classes and properties. This ontology is designed to be compatible with the OWL 2 RL profile, allowing for efficient reasoning.
Key aspects of their method include:
- Ontology Engineering: Building a TBox (terminological box) that defines the relationships and hierarchy of musical concepts.
- Functional Axioms: Defining axioms that capture the rules of tonal and modal harmony, enabling the inference of chord functions within a progression.
- Semantic Web Technologies: Utilizing OWL 2 for ontology representation and SPARQL for querying the knowledge graph.
- Extensibility: The DL approach allows for easy extension of the ontology to include more music theory concepts or different cultural harmonic systems.
Datasets
The authors evaluated their approach using four datasets:
- Isophonics Beatles Reference Annotations: Contains chord annotations for The Beatles' discography.
- McGill Billboard Annotations: Provides chord annotations for a large number of songs from Billboard charts.
- The Weimar Jazz Database: Contains transcriptions and chord information from jazz improvisations.
- Hooktheory Data: MusicXML data used for a web application, including information on diatonic mode and scale degrees of chords.
The hooktheory dataset was particularly useful for comparison with HarmTrace as it provided known song keys.
Key Insights and Ideas
- Formalizing Music Theory: Representing complex music theory concepts like functional harmony in a formal language like DL allows for precise, unambiguous, and machine-interpretable analysis.
- Bridging Symbolic and Theoretical Analysis: This work demonstrates how symbolic music data can be enriched with theoretical knowledge, enabling deeper insights than purely symbolic or audio-based analysis.
- Extensibility and Interoperability: The use of OWL 2 and semantic web standards makes the framework highly extensible and interoperable with other data sources.
- Tonal Harmony as a Subset of Modal Harmony: The ontology naturally represents tonal harmony as a specific case within the broader framework of modal harmony, reflecting theoretical relationships.
- Local vs. Global Tonic: The introduction of "Local Functions" (Local Tonic, Local Dominant, LocalSubdominant) acknowledges that the tonic center can shift within a piece, a crucial aspect of musical analysis not always captured by simpler models.
Key Results
The paper presents a comparison between their ontology-based reasoning and the HarmTrace system.
- Functional Analysis: Both systems were able to perform functional harmonic analysis. For the hooktheory dataset, their system identified 2767 tonic functions and 7529 dominant functions, while HarmTrace found 3264 tonic and 7604 dominant functions.
- Qualitative Comparison: Despite differences in methodology (HarmTrace's insertions/deletions vs. the ontology's direct reasoning), the analyses were largely identical in most cases.
- Genre Differentiation: Using SPARQL queries on chord functions, the authors were able to identify distinct harmonic differences between music genres (e.g., Jazz vs. Pop-Rock), demonstrating the practical utility of their approach. For instance, in the Jazz genre, over half the chords served a dominant function, indicating a greater use of tension-building chords.
Limitations and Future Work
- HarmTrace Comparison Nuances: The authors acknowledge that a direct comparison with HarmTrace is challenging due to differing theoretical frameworks (tonal vs. modal functional harmony) and HarmTrace's use of chord insertions/deletions.
- Musician Insight: While HarmTrace's hierarchical representation might offer more direct insight to musicians, the authors argue their DL approach is theoretically more thorough and more usable due to standardization.
- Future Work:
- Exploring the impact of these enriched symbolic representations for machine learning pipelines.
- Developing educational tools for music students based on this knowledge representation.
- Converting MIDI and MusicXML files directly into RDF using their ontology for broader accessibility.
- Extending the ontology to represent other aspects of music theory like rhythm, melody, and structure, and incorporating harmonic systems from different cultures.
Summary updated