The objective of hypothesis refinement is to modify the scope of a rule to more accurately model the data. In this paper we examine the relation between data summarization and hypothesis refinement in association rules with fuzzy temporal constraints. We then present two refinement strategies based on disjunctive constraint generalization and constraint specialization. Disjunctive generalization produces more general rules by merging adjacent constraints in the partition of the window of relevance. Temporal specification uses linguistic hedges to reduce the constraint window while maintaining the interpretability of the rule. The refinement strategies are developed to maintain or enhance the linguistic interpretability of the rules.