In this article, we evaluate features and algorithms for the task of prosodic boundary prediction for Greek. For this purpose a prosodic corpus composed of generic domain text was constructed. Feature contribution was evaluated and ranked with the application of information gain ranking and correlation-based feature selection filtering methods. Resulted datasets were applied to C4.5 decision tree, one-neighbour instance based learner and Bayesian learning methods. Models performance exploitation led as to the construction of a practically optimal feature set whose prediction effectiveness was evaluated with two prosodic databases. In terms of total accuracy and F-measure, evaluation results established the decision tree effectiveness in learning rules for prosodic boundary prediction.

prosody, phrase breaks, ToBI, C4.5, IB1, bayesian learning