Tel.: +49 541 969-3387
Fax: +49 541 969-3374
Institute of Cognitive Science,
49090 Osnabrück, Germany
Computational Modeling the Pragmatics of Conditionals
Despite the very long history of research on conditionals, there is no consensus and no prevalent theory that is able to explain the many and varied ways to interpret a conditional utterance. Why does the reply "and what if I don't?" to the conditional "If you want, there are biscuits on the sideboard" appear to be meant as a joke, while it seems totally normal as reply to the conditional "If you mow the lawn, I'll give you $5"?.
With the aid of computational models, we aim to understand how such diverse interpretations arise for different uses of if, then. While most related work on conditionals has focused on qualitative models of their semantics, we investigate their pragmatics with quantitative computational models. As a basis, we use the Rational-Speech-Act (RSA) model, which is a Bayesian formalization of Gricean ideas how speakers choose utterances and listeners are thereby able to infer the speakers' intentions going beyond what is literally said. Besides taking into account the interactivity of the interlocutors, we model the utterance content by explicitly representing the underlying causal structure among world states.
With this approach, on the one hand, our model makes predictions for the interpretations of conditionals that align well with theoretical findings from the literature on the acceptability conditions for conditional utterances. Furthermore, first preliminary data from a behavioral online experiment we performed, showed that the model is also able to generate interpretations that correlate with participants' measured intuitions. In the experiment, we exploit peoples' intuitive understanding of physics and causality in order to inject them probabilistic beliefs about world states and their structures - the basis for the predictions of our model.
In the future, we would like to make a step towards a unified theory of peoples' use and interpretation of various sorts of conditionals.
To this end, we plan to integrate other theoretical ideas into our model, e.g.~how a Question-under-Discussion may lead to a biconditional reading (p iff q), making the inference not p from not q valid, in one context, but not in another.