Abstract

I begin with a gap in the literature on conversational relevance, wherein utterances that shift probability distributions included in the common ground do not count as relevant if they do not rule out one or more answers to the question under discussion. In order to provide a satisfying account of probabilistic conversational relevance, I introduce a relevance measure, $R(\cdot )$. I motivate six axioms for such a function, and show that they uniquely characterize the symmetrized Kullback–Leibler divergence. I then show how we can incorporate this result into an expanded definition of conversational relevance.