[Predictive Benefit of N-Grams](https://www.proquest.com/openview/8a8d2fc9ae88ea63bbd2659c576596a6/1?cbl=18750&diss=y&pq-origsite=gscholar
class CoAuthor
def __init__(self, name):
self.name = "David Temperley"Format (When): Where
- International Conference on Math and Computation in Music (Proceedings): see here
TL;DR/Abstract
An important question about n-gram models is, how much gain in predictive power is achieved as n-gram order is increased? We offer a new method for approaching this question—one that avoids the difficulties that arise from testing on unseen data, and also avoids the overestimation of predictive benefit that can occur with small data sets. For each nth-order context, we choose a sample of tokens of the “parent” (n-1)-order context whose count exactly matches that of the nth-order context, and compare the conditional entropy of the two contexts. Using corpora of folk songs, classical themes, and hymn tunes, we provide the first unbiased estimate of the predictive benefit of 1st- through fourth-order melodic n-gram models. Comparing predictive benefits for different intervals within a single n-gram order, we find patterns that reflect the influence of basic principles of melodic structure: pitch proximity, step inertia, and range constraints.