Augmenting variation of system utterances using corpora in spoken dialogue systems

Compared to the variation in utterances that users may exhibit in conversation with spoken dialogue systems, system utterances can be very rigid with little variation. One recent approach to dealing with this problem is a trainable sentence planner, which uses natural language generation techniques to create a large number of alternative utterances for a given content, by randomly combining an initial set of basic syntactic structures. However, the amount of variation achieved is limited by the size of the initial set, which is usually specified by hand. We propose augmenting the variation of system utterances by automatically incorporating useful sentences obtained from corpora into the initial set used by the generation process. Experimental results show that this approach can successfully create a generator with augmented variation