Our seminar series is free and available for anyone to attend. Unless otherwise stated, seminars take place on Wednesday afternoons at 2pm in the Kilburn Building during teaching season.

If you wish to propose a seminar speaker please contact Antoniu Pop.


Learning and context in spoken dialogue systems: recent results from the TALK project

  • Speaker:   Dr  Oliver Lemon  (University of Edinburgh)
  • Host:   Ian Pratt-Hartmann
  • 10th May 2006 at 14:15 in 1.5
We introduce the work with an overview of current research on dialogue systems in the TALK project ( and then focus on the use of computational learning techniques for optimizing dialogue strategies, in particular, modelling human-machine dialogues as Markov Decision Processes (e.g. Levin and Pieraccini 1997, Young 2000, Litman et al. 2000, Henderson et al. 2005). We show how to use a variety of statistical learning techniques over rich representations of dialogue context to learn optimal system dialogue moves (e.g. request(departure-city)) or implicit-confirm(arrival-date)) in each possible context. We present results for learned dialogue strategies which perform better than the original hand-coded DARPA COMMUNICATOR dialogue systems (Henderson et al. 2005), and results for testing the learned strategies in spoken dialogue systems with real users. We also present related techniques for context-sensitive speech recognition and parsing (Gabsdil and Lemon 2004), leading to significant reductions in error rates. In addition we discuss issues of tractability when using very large state spaces in Reinforcement Learning for dialogue systems, and the use of user simulations when learning dialogue strategies. We conclude by outlining future work involving "Partially Observable" dialogue states.

See and
▲ Up to the top