The present paper explores the issue of corpus prosodic parsing in terms of prosodic words. This question is of importance in both speech processing and corpus annotation studies. We propose a method grounded on both statistical ans symbolic (phonologicial) representations of tonal phenomena and we have recourse to probabilisitic grammars, within which we implement a minimal prosodic hierarchical structure. Both stages of probabilistic grammar building and its testing in prediction are explored and quantitatively and qualitatively evaluated.