Abduction allows us to model interpretation of discourse as the explanation of observables, given additional knowledge about the world. In an abductive framework, many explanations can be constructed for the same observation, requiring an approach to estimate the likelihood of these alternative explanations. We show that, for discourse interpretation, weighted abduction has advantages over alternative approaches to estimating the likelihood of hypotheses. However, weighted abduction has no probabilistic interpretation, which makes the estimation and learning of weights difficult. To address this, we propose a formal probabilistic abductive framework that captures the advantages weighted abduction when applied to discourse interpretation.
展开▼
