This paper proposes and estimates a dynamic model of household inflation expectations with information frictions and time-varying parameters, where households use a Bayesian learning model to form and update inflation expectations. The model decomposes households' inflation expectation formation process into a learning component, a noisy signal component, and a measurement component. Model-implied household inflation expectations provide a robust fit for the expectation-augmented Phillips curve. As a result of time-varying inflation dynamics, households' attention to inflation is endogenous to its volatility. This insight offers explanations for the anchoring of inflation expectations during the Great Moderation.