Vincenzo Pallotta
Faculty of Computer and Communication Sciences
Swiss Federal Institute of Technology –
The goal of this tutorial is to acquaint the attendees with the applications for human-machine dialogue systems, and with the problems arising in developing systems that can participate in natural dialogue with humans.
Difficulties in modelling dialogue arise, for instance, from very frequent phenomena such as the use of anaphoric expressions, forms of deixis (e.g. 'now/then', 'here/there', 'I/you', 'this/that'). The use of the contextual information is required for the right interpretation of these linguistic objects. Participants are able to make sense of a dialogue even when little linguistic information is present in the utterances. This is achieved by their cognitive skills: their abilities to perform inferences based on background knowledge and assumptions on the other participants' mental states. There are situations where the literal meaning is not sufficient to understand the role of the utterance in the dialogue, that is, the corresponding dialogue act cannot be directly recognized by simply its linguistic content: it must be inferred.
In the first part of this tutorial, the current state-of-the-art in Dialogue Management Systems will be presented, and particular attention will be paid to semantic and pragmatic models of dialogue based on the representation of participants beliefs, goals, and intentions (i.e. BDI models), and to multimodality.
In the second part, the ViewFinder model is described in detail as an efficient and elegant alternative to plan-based approaches in BDI dialogue models. ViewFinder is a formal framework for representing, ascribing and maintaining nested mental attitudes of interacting agents. Viewpoints on mental attitudes of communicating agents are represented by means of nested typed environments. Operations over typed environment are defined and used to simulate several form of reasoning which are necessary in order to assimilate information into knowledge structures from communication. ViewFinder has been recently implemented as a computational framework, and it is now available as a tool for Computational Dialogue Modelling. A classical example of dialogue problem is modelled and solved within the ViewFinder framework showing how it is possible to avoid the explicit use of Common Ground.