Counterfactual interpretation as search for coherence in a learned model of the world

Abstract

We propose a semantics for counterfactuals that explicitly connects it to cognitive models of temporal / causal inferences in narrative discourse comprehension. Both narrative discourses and counterfactuals are understood by (i) constructing temporally-indexed sequences of situation models – analog, vector-space representations of meaning grounded in our experience perceiving and understanding the world (our learned model of the world) – and then (ii) enriching these representations with temporal inferences through a temporal coherence-seeking process. The account builds on the model of narrative discourse comprehension in Frank et al. (2003), showing how we can use it to move beyond theoretically primitive notions of counterfactual similarity by computing / constructing a historically-structured similarity relation through this coherence-seeking interpretation process. We formalize a semantics that integrates linguistic input with rich temporal world knowledge, interprets counterfactuals in a graded context-sensitive manner, and inherently links interpretation to processing time.