A Bayesian Approach to Reasoning under Uncertainty in Natural Language Dialogues

Description

In the talk I would like to present a probabilistic approach to
dealing with various kinds of uncertainty in modelling natural
language dialogues. The uncertainties a conversational agent has to
deal with arise in:
1. the inaccuracy and incompleteness of the linguistic information
extracted from a user utterance by a parser,
2. the interpretation of this information in terms of communicative
actions or dialogue acts,
3. updating the agent's mental model of the user on which the user's
behaviour is based, and
4. updating the agent's mental model of the user when the agent itself
has performed a certain action.

I will give a sketch of how Dynamic Decision Networks (an extension
of Bayesian Networks) could be used for modelling a conversational
agent which can deal with incomplete information, make 'educated
guesses' concerning interpretation, and 'do the right thing', based on
Bayesian updating and utility-directed action. I will indicate which
problems arise when using this approach and what work needs to be done
in order to solve these problems.

Period	3 Nov 2000
Event title	11th Meeting on Computational Linguistics in the Netherlands, CLIN 2000
Event type	Conference
Conference number	11
Location	Tilburg, Netherlands
Degree of Recognition	International