Assign Stress for Interrogative Sentences via Syntax Structure Mapping

Compared to the prosody study in statement-like sentence, little research has been conducted in interrogative sentence, which will definitely improve the naturalness of spoken dialog system. This paper proposes a computational model to assign stress in interrogative sentences using linguistic knowledge. We firstly investigate the interrogative sentences perceptually to label their stress locations and then summarize a series of syntax-stress mapping rules. The dependency syntax is adopted in this work and automatically obtained by Stanford Parser. All the dependency relations in Stanford Parser are studied and stable stress patterns are assigned. Syntax depth is considered as a factor to weight each dependency relation. Finally, the linear superposition of stress score calculated from each dependency relation is assigned to prosodic words. Experiments show that the weighted accuracy of syntax-stress mapping is 74%. A detailed analysis of results is also provided.