Recurrent Neural Networks for Dialog State Tracking

Опубликовано 21 июня 2016, 20:37

The recent Dialog State Tracking Challenges have brought into focus the problem of accurately estimating the dialog state throughout a sequence of spoken interactions with a dialog system. Recurrent Neural Networks (RNNs) can be effectively applied to this problem, giving results which perform competitively in the shared challenge tasks. The proposed model is unique in that it eliminates the need for a spoken language understanding component, and the inherent bottleneck, by mapping directly from words to the dialog state. No intermediate semantic representation is required. Latest results are presented in adapting the RNN tracker to extended domains. Initial results show that state tracking accuracy can be improved further with online and unsupervised adaptation while being exposed to unlabeled dialogs in an extended domain. Paper: Word-based Dialog State Tracking with Recurrent Neural Networks mi.eng.cam.ac.uk/~mh521/papers... Matthew Henderson, Blaise Thomson and Steve Young. SIGdial 2014