[Elsnet-list] PhD thesis proposal

Stergos Afantenos stergos.afantenos at irit.fr
Mon Feb 18 11:38:59 CET 2013

Research Institute: IRIT, CNRS \& Université Paul Sabatier, France

Supervisors: Leila Amgoud, Nicholas Asher, Stergos Afantenos

TITLE: Argumentation in Dialogue

In a persuasion dialogue, participants often have conflicting opinions and everyone tries to convince others. This involves an exchange of arguments which may be in conflict or not. Such a dialogue ends with either a failure, where the disagreement persists, or agreement.

Persuasion dialogues interest linguists as well as researchers in Artificial Intelligence (AI). The first develop theories of discourse and dialogue. These theories allow fine linguistic analysis of texts, and in particular analysis of the arguments present therein. The linguistic foundations enable empirical analysis of the arguments where the goal is the automatic extraction of the relational structure that underpins those arguments.  As far as AI is concerned, the focus is on developing models of argumentation and criteria for evaluating them. Arguments are often seen as abstract entities whose origin and nature are undefined. 

The PhD thesis we propose puts both expertise together in order to build models of argumentation and in accordance with rich dialogues in natural language. The thesis will first examine the state of the art in both areas, and build bridges between them. Discourse theories will give the argumentation theory rich data with a semantic and logical analysis. These data concern both the arguments and the relationships that exist between them. This in turn helps define appropriate criteria for evaluating arguments. These criteria are validated in the texts. The osmosis might also happen the other way around, where argumentation theory might enhance and inform discourse analysis. 

The final theory that will result will have to be empirically validated using Machine Learning approaches. The problem of identifying the underlying argumentation structure is an extremely attractive one, since it involves not simply the prediction of isolated arguments or relations but instead the prediction of the whole argumentation graph. By consequence methods from structured prediction will need to be exploited. 

More information about the Elsnet-list mailing list