The project is no longer available

Efficient Probabilistic Inference for Spoken Dialogue Systems

Spoken Dialogue Systems (SDS) are typically plagued with noise and uncertainties, at all levels of processing.  These uncertainties might arise from error-prone speech recognition, linguistic ambiguities, unknown user intentions, as well as many other sources.  To handle these uncertainties adequately, some sort of probabilistic reasoning is usually required.  

There is compelling evidence in the scientific literature that the use of probabilistic reasoning (usually combined with machine learning techniques to learn the model parameters) can deliver significantly better performance than non-probabilistic methods for spoken dialogue systems.  However, probabilistic inference is a computationally complex problem, and can easily become intractable without strong domain-specific constraints.  It has been shown for instance that exact inference on Bayesian Networks is NP-Hard (Cooper 1988).

Many techniques have been devised in the last decade to address this computational limitation, usually based the idea of approximate inference: instead of directly computing exact probabilities, these methods seek to approximate these probabilities to a certain degree of accuracy, and incrementally refine these estimates over time.  This family of frameworks notably include particle-based inference (Markov Chain Monte Carlo etc.), belief propagation, and variational methods.

The goal of the thesis would be to investigate the use of these approximate inference techniques for the specific domain of spoken dialogue systems.  To this end, various inference algorithms will be implemented and integrated as part of a generic architecture for spoken dialogue systems, currently in development in our research group.  These algorithms might of course be adapted/optimised to take advantage of the specific problem structure.  The performance of these algorithms will then be empirically evaluated on a few dialogue domains.

Prerequisites: basic knowledge of probabilistic inference (Bayesian Networks etc.), as well as programming skills in Java, C++, Lisp or Python.

If you are interested in this topic, please contact Erik and Pierre to discuss the practical details of supervision. 

Tags: Artificial Intelligence, Computational Linguistics, Spoken Dialogue Systems, Natural Language Processing
Published Sep. 27, 2011 1:26 PM - Last modified Jan. 4, 2013 11:53 AM

Supervisor(s)

Scope (credits)

60