Estimating Lexical Complexity from Document-Level Distributions