Research has shown that linguistic units carrying more information tend to be realized with enhanced speech signals. Most previous studies measure the information that a linguistic unit carries with its surprisal. However, such measurement lacks the ability to model long-distance contextual effects. The current study proposes novel measures of linguistic unit' s information by incorporating the GPT-2 pre-trained language model and mutual information (MI) between text and its phonemic transcription. We examine the prosodic effects of word surprisal and Mi-based information of final and tones in Mandarin Chinese. Results show that more information of both words and finals enhance prosodic prominence, proving the validity of our proposed measurements. Besides, the effects of information are more notable on duration feature compared with pitch and intensity feature.
GPT-2 ；信息量；韵律；音长；互信息
郝韵、解焱陆、林炳怀、张劲松
展开 >
北京语言大学信息科学学院，北京100083
腾讯科技，北京100083
GPT-2 ；信息量；韵律；音长；互信息
Chinese national conference on computational linguistic
Nanchang(CN)
The 21st Chinese national conference on computational linguistic
NETL
