Statistical Spoken Language Generation of Stylistic Variation

Опубликовано 6 сентября 2016, 17:06

Utterance generation for dialogue applications such as spoken dialogue systems and some types of computer games is typically highly handcrafted and domain dependent. This leads to problems of portability and scalability, or what has been called the 'authoring bottleneck' in the case of interactive drama systems. In this talk, I will describe a novel statistical spoken language generation framework, and our research investigating the performance of this approach across three successively more challenging application domains. I will focus in particular on our recent work on statistical generation techniques capable of generating a wide range of stylistic variants for an input communicative goal, based on learning the settings for linguistic parameters identified by research on personality psychology.