Court decisions on music plagiarism and the predictive value of similarity algorithms

Abstract

Tune plagiarism in pop music is a common and often feverishly debated phenomenon which surely has to do with the vast amounts of money that individual melodies are able to generate in today s pop music business. The similarity between melodies is assumed to be a very important factor in a court s decision about whether a new tune is an illegitimate version of a pre-existing melody. Despite the wide-spread belief that there is a fixed and simple limit to the number of corresponding notes between two melodies, actual court decisions are based on far more complex considerations regarding the musical material. This paper first sketches the legal framework and principal features of the legal processing of cases of alleged melodic plagiarism with a focus on US copyright law and discusses selected cases to highlight the corresponding legal practices. In the empirical part of this paper, we model court decisions for cases of alleged melodic plagiarism employing a number of similarity algorithms. As a ground truth dataset we use a collection of 20 publicly available cases from the last 30 years of US jurisdiction. We compare the performance of standard similarity algorithms (edit distance and n-gram similarity measures) to several new similarity algorithms that make use of statistical information about the prevalence of chains of pitch intervals in a large pop music database. Results indicate that these statistically informed algorithms generally outperform the comparison algorithms. In particular, algorithms based on Tversky s (1977) concept of similarity show a high performance of up to 90% of court decisions correctly predicted. We discuss the performance and structure of the algorithms in relation to a few interesting example cases and give an outlook on the potential and intricacies of our approach.