Expectation propagation  (EP) is becoming a standard tool in the machine learning community as it is able to able to perform inference quickly, often has less bias than alternatives such as variational inference and Laplace approximations, and requires less computation than many sampling methods. This talk will discuss the application of EP to modelling user behaviour in spoken dialogues. EP provides an efficient way to train the parameters and update the beliefs of a spoken dialogue systems based on the partially observable Markov decision process. These parameters can even be learned using noisy observations, and do not require any annotations besides semantic representations of the speech recognition output of a dialogue. The resulting systems are shown to be more robust to errors than standard approaches, largely because the models are able to handle the uncertainty in the dialogue in a principled way. EP also provides a mechanism from training the parameters of Bayesian network-based simulators of the user behaviour in a dialogue. Performance of the resulting simulators will also be discussed.

Blaise Thomson is a Research Fellow at St John's College in the University of Cambridge, where he also received his PhD in 2010. His research papers focus on spoken dialogue systems, reinforcement learning, and collaborative filtering. Several of these papers have received awards at venues such as Interspeech, the International Conference on Acoustics, Speech, and Signal Processing (ICASSP) and the Spoken Language Technology (SLT) conference.