基于依存信息融合特征的汉语韵律预测

Chinese Prosodic Prediction Based on Dependency Information Fusion Feature

针对目前基于浅层语法特征和依存句法单特征的汉语韵律层级预测能力较弱的情况,提出一种改进的汉语韵律预测方法。通过从输入文本的依存句法分析结果中自动提取依存句法单特征,并对其中关键特征进行特征融合,得到依存信息融合特征。将依存句法单特征与融合特征进行韵律层级预测实验对比,选取最优的依存特征组合与浅层语法特征相结合,利用决策树C4.5算法实现韵律结构层级的预测。经过大量的语料训练和测试结果表明,依存信息融合特征相比依存句法单特征整体韵律层级的预测准确率均有所提升,相对于浅层语法特征,韵律词和韵律短语的预测准确率分别提高了5.8%和15.4%。

Aiming at the weak ability of predicting the prosodic level based on the characteristics of shallow grammar and dependence-based syntactic single feature, this paper proposes an improved Chinese prosodic prediction method. By extracting the dependency syntactic single feature from the dependency parsing results of the input text and integrating key features, the dependent information fusion feature is obtained. The syntactic single feature is compared with the fusion feature through the prosodic hierarchical predictive experiment. The optimal dependence feature combination is combined with the shallow grammar feature,and the C4.5 algorithm is used to predict the prosodic hierarchy. After a lot of corpus training and testing, results show that the prosodic predictions of interdependent information fusion features are better than the predicative ones based on dependency syntax. Compared with rhyme hierarchical law prediction based on shallow grammar features,the accuracy of prosodic word is increased by 5.8% , and the precision of prosodic phrase is increased by 15.4%.