Disambiguating the senses of non-text symbols for Mandarin TTS systems with a three-layer classifier

摘要

Various kinds of non-text symbols appear in texts. The oral expressions of these symbols may vary with their senses. This paper proposes a three-layer classifier (TLC) which can disambiguate the senses of these symbols effectively. The layers within TLC are employed in sequence. The 1st layer is composed of two components: pattern table and decision tree. If this layer can disambiguate the sense of the target symbol, the disambiguation task stops. Otherwise the next two layers will be triggered. In such a situation, the procedure will go through the TLC. Based on the Bayesian theory, the 2nd layer adopts the voting scheme to compute the disambiguation score. Several features of token, which may affect the effectiveness of our voting scheme, are analyzed and compared with each other to achieve better accuracy. According to the algorithm confidence of sense disambiguation, the 3rd layer may exploit an alternative model to enhance the performance. Experiments show that our approaches can learn well even with only a small amount of data. The overall accuracies of training and testing sets are 99.8% and 97.5%, respectively.