Rule learning based Chinese prosodic phrase prediction

We describe a rule-learning approach towards Chinese prosodic phrase prediction for TTS systems. 3167 sentences with two-level prosodic phrase labeling information was prepared for analysis. Candidate features related to prosodic phrasing were extracted from the corpus to establish an example database. Based on this, a series of comparative experiments is conducted to collect the most effective features from the candidates. Two typical rule learning algorithms (C4.5 and TBL) were applied on the example database to induce prediction rules. To compare the results with others, the general evaluation parameters were introduced in the paper. With these parameters, the methods were compared to RNN and bigram based methods. Results show that the rule-learning approach introduced here can achieve better prediction accuracy than the nonrule based methods and yet retain the advantage of the simplicity and understandability.