The Contribution of Contrast in Context

Contrastive relations are often signalled by cue words like but or although, e.g. Although Sweden is a member of the EU, they don´t use the Euro. Contrast relations are pervasive in discourse, yet we know little about their function in context, i.e. why not just say Sweden is a member of the EU that doesn´t use the Euro? We have theories that address the semantic relationship between the two parts of a contrast relation, but few that discuss how the relation itself then fits with the context, e.g. what type of context (argumentative, narrative) does the first example tend to occur in? Our current state of ignorance is even more surprising given that almost every major work in computational discourse includes contrast as one of its core rhetorical relations. Work on contrast has been approached from two perspectives, rhetorical discourse structure and information structure (e.g. what concepts are new or given in a discourse). Neither perspective has addressed the issue of the function and relation of contrast relations to the context. To remedy this, a more explicit, empirically informed account of contrast is required. I will develop a formal model of contrast by using the results of a corpus study of dialogue to examine its use in context, and the results of experimental study to explore more subtle issues. This model will also be useful for computational applications that need to handle contrast, such as e.g. question-answering and automatic summarization. By using Bidirectional Optimality Theory (BiOT) as a framework, I can handle multiple information sources and model both speakers and hearers so that the same formal description can be used both for systems which require discourse interpretation and discourse generation. Additionally, using BiOT for dialogue modelling is unique, and can represent a significant contribution in its own right.