(A) natural language processing approach to quantitative measurement of symbolic melodic similarity

Abstract

The melodic similarity is a key concept that helps with analysis and understanding in theoretical areas such as musicology, music cognition, and music psychology, as well as in applied areas such as music copyright, music classification and recommendation, and various fields. The process of determining melodic similarity is inherently intuitive and subjective. Psychological approaches to evaluating melodic similarity have relied on cognitive experimental evaluations, expertise, or music theory-based models, while computational methods derived from natural language processing have generally provided a single value. These fragmented approaches may only reveal task-specific information. Thus, this study aims to develop a quantitative framework for measuring the semantic and qualitative aspects of melodic similarity. The research is organized around three general steps developed for the similarity analysis of symbolic melodies: 1) Representation, which converts the melodic features into a form optimized for computational analysis

and 3) Computation, which measures and visualizes the semantic similarity of melodies.For melodic representation, this dissertation proposes a text-based representation, Mel2Word. Mel2Word is developed with the intention of facilitating the analysis of melodies through the application of Natural Language Processing. As a textual representation of a melody optimized for NLP technology, it includes important musical features like pitch and rhythm, allowing practical application of music as a language and analytical understanding of various musical features. For melodic segmentation, this dissertation proposes Byte-Pair Encoding (BPE) based melody segmentation using NLP techniques. This is a data-driven method in which a melody is considered a sentence and tokenized into the meaningful melodic vocabulary. This can enable music analysis to be performed as semantic terms with meaning and context, rather than as fragmented melodies with single characters. For similarity calculation, this dissertation proposes two approaches to calculating melodic similarity: 1) Multi-segmental analysis and visualization to understand the hierarchical and structural similarities of melodies

2) A method of embedding and vectorizing melodies that can quantitatively comprehend the contextual meaning of melodies and qualitative features of similarity. For the former, a Cross-Scape Plot is proposed, which provides a hierarchical visual representation of where and how similar the two melodies are. For the latter, we propose a TV-TF-IDF weighting function that analyzes the word salience, importance, and uniqueness as redefined by psychological models and NLP methods. Through the similarity calculation by applying this weighting method to the vectorized word embedding, we assess the qualitative aspect of "substantial similarity" of melodies for copyright infringement cases.The ultimate goal of this study is to identify the semantic and qualitative meaning of melodic similarity through the integration of MIR, NLP, and psychological models. In particular, it is to build a scientific framework that can treat music as a language so that computational analysis of music can be readily performed using an NLP approach. Beyond melodic similarity, we hope that this contribution will advance the field of music analysis by presenting a qualitative yet quantitative approach to assessing a wider range of high-level musical aspects.

2) Segmentation, which divides the melodic representation into meaningful units