Mandarin prosodic word prediction using dependency relationships

Previous research demonstrated that the dependency structure of a sentence is helpful for prosodic phrase boundary prediction in mandarin Text-To-Speech systems. However, no experimental results proved that the dependency relations are important to prosodic word boundary detection. Also, most of the published methods use machine learning technologies, which require people to label the prosodic boundaries manually for training purpose. In this paper, we propose a rule based method for prosodic word boundary prediction based on two observations. First, in most of the cases, a prosodic word is a lexical word, or it is a combination of adjacent lexical words. Second, the combination of lexical words relies on semantic relationships. The dependency tree of a sentence can describe the semantic relations between words. Hence, we combine adjacent words which have dependent relationships into a prosodic word. Some other restrictions are added to fine-tune the method. Experimental results demonstrate that the method achieved 0.918 and 0.901 on two corpora in terms of F-score.