A machine learning approach to rhythmic classification of languages.

Abstract

The Orange machine learning environment was used in four experiments designed to test for distinct language rhythm classes (stress‐ and syllable‐timed) on the basis of scores from three sets of rhythm metrics: ΔC‐%V, PVIs, and Varcos. In experiment 1, naive Bayes and decision tree learners were least accurate within classes, and most accurately distinguished stress‐timed languages (English, German) from syllable‐timed languages (Italian, Spanish); results for unclassified languages (Greek, Korean) were mixed. In experiment 2, a naive Bayes learner compared classification accuracy using each metric against the accuracy obtained using all metrics together. Accuracy did not generally suffer when learning on any single metric, supporting the hypothesis that the metrics overlap in their description of the data. In experiment 3, naive Bayes learners trained on stress‐/syllable‐timed pairs and had to classify Greek and Korean. The learners disagreed in their classifications. In experiment 4, unsupervised k‐means clustering was employed using different combinations of metrics, and all of them at once. The classifier performed poorly: English and German could sometimes be separated out, but not the unclassified languages. Overall, the results add to a growing body of research failing to find evidence of rhythm classes from production.