Reranking CCG parser for Jazz chord sequences

摘要

When listen to the music, specifically chord progression, each person can have more than one understanding or feeling about what they heard which we call musical intuitions of listeners or human capacity of musical understanding. To disambiguate the chord progression, Granroth-Wilding et al. have employed Probabilistic Combinatory Categorial Grammar (PCCG), and they acquired the recall value of 88.78%, and the precision value of 90.18%. Because this chord progression parser only outputs the one with the highest probability, the correct solution may still reside in the following candidates. In this paper, we use the reranking model to improve the performance of the parser. By selecting a set of simple n-gram features and configuring perceptron algorithm for finding optimizing parameters, we have improved performance of the system by 2.2%, and even 6.57% when we could perfectly pick up correct candidates from 5000-best results.