Counterfactual Interpretation as Search for Coherence In a Learned Model of the World ------------------------------------------------------------------------ [slides][1] "Once there was a little boy who lived in a hot country. One day his mother told him to take some cake to his grandmother. She warned him to hold it carefully so it wouldn't break into crumbs. The little boy put the cake in a leaf under his arm and carried it to his grandmother's. When he got there, the cake had crumbled into tiny pieces. His grandmother told him he was a silly boy" (Stein & Glenn, 1975). Comprehending this elementary-school narrative as a coherent whole requires us to quickly and seamlessly access our background world knowledge to enrich what is literally said with a range of inferences. For example, the grandmother calls the boy silly because he should have known (as we all do) that cakes are brittle and they crumble when squeezed under one's arm. Counterfactually, __if he had carried it on top of his head, the cake wouldn't have crumbled.__ I propose that, just as discourse comprehension leverages our world knowledge to make inferences and establish narrative coherence as we construct a mental model of the story, the work of constructing a hypothetical situation that both satisfies the counter-to-fact antecedent and is otherwise as similar as possible to the actual situation involves a similar search for temporal coherence. I thus follow Thomason, Ippolito, Khoo a.o. in assuming that we need a historically-structured interpretation for counterfactuals going beyond the Stalnaker / Lewis pure similarity account. I go further and show how we can leverage the model of discourse comprehension in Frank et al (2003) to construct the similarity / accessibility relation that is a theoretical primitive in these accounts, and largely relegated to the _deus ex machina_ of pragmatics, out of a training sample of 250 temporally ordered situations in a microworld involving two characters Bob and Jane engaging in various activities. From this sample "movie" of 250 "frames", we learn a first-order Markov model of how events typically unfold in the microworld. This simple model is sufficient to both make gradient predictions about counterfactual interpretation (Kaufmann, Lassiter a.o.) and predict processing difficulty profiles for counterfactuals that depend on the size of the similarity gap between the actual context and the counter-to-fact antecedent, and therefore on the amount of inference work needed to bridge this gap and establish temporal coherence. [1]: https://drive.google.com/file/d/177fh4pWTK8eMpyB9B3Q6CZpQcS3OyfIV/view?usp=drive_link