基于最大熵模型的韵律短语边界预测  

Prosody Phrase Break Prediction Based on Maximum Entropy Model

作　　者：李剑锋[1] 胡国平[1] 王仁华[1] 

机构地区：[1]中国科学技术大学讯飞语音实验室,安徽合肥230027

出　　处：《中文信息学报》2004年第5期56-63,共8页Journal of Chinese Information Processing

基　　金：国家自然科学基金资助项目 (5 9975 0 18)

摘　　要：语音合成系统中 ,由于韵律短语边界预测的水平不高 ,阻碍了合成语音自然度的进一步提高。本文根据韵律短语边界预测的特点 ,提出了基于最大熵模型的预测方法。为考察该方法的能力 ,在较大规模的数据集上 ,使用相同的属性集 ,对比了其与主流的决策树方法的预测效果。还考察了词面信息的贡献 ,以及选择特征时的不同阈值对最大熵模型的影响。实验表明 ,使用相同的属性信息 ,最大熵方法比传统的决策树方法在F Score上有 5 5 %的提高 ,加入了词面信息的最大熵模型则有 9 4 %的提高。最后指出 ,最大熵模型相当于一个带权重的规则系统 。In TTS (Text-To-Speech) systems, prosody phrase breaks can not be predicted with high accuracy, which slows down the improvement of naturalness of synthesized speech. In this paper, a maximum entropy based model for prosody phrase break prediction is proposed, and a comparison is conducted on large corpora between the new model and the decision tree based model which is the mainstream method for prosody phrase break prediction. The contribution of lexical feature set and influences of different cutoff values are also investigated. It is demonstrated that, utilizing the same feature set, maximum entropy based model makes an improvement of 5.5% on F-Score over decision tree based model. Integrating lexical information, an improvement of 9.4% over decision tree based model is achieved. In the end, it is pointed out that a maximum entropy model can be considered as a weighted rule system, which solves the problem of rule conflicting in an elegant way.

关 键 词：计算机应用 中文信息处理 韵律短语边界预测 最大熵 决策树 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]