News

Abstract:

Procedural generation algorithms that infer rules based on a dataset of examples require that each example is made up of explicitly labeled components. Musical sequences resist this categorization because they lack explicit structural semantics. To infer a set of rules based on musical sequences, then, a phrase segmentation process is needed to partition a musical sequence into identifiable patterns. We outline a solution to this challenge that uses grammatical induction algorithms, a class of algorithms that infer a context-free grammar from an input sequence. We study five different grammatical induction algorithms on three different datasets, one of which is introduced in this work. Additionally, we test how the performance of each algorithm varies when transforming musical sequences using viewpoint combinations. Our experiments show that LongestFirst achieves the maximum F1 score across all three datasets, and viewpoint combinations that include the duration viewpoint result in the best performance.