A new word usage coefficient, Um, is proposed. It avoids the disadvantages of Juilland's U by (1) taking account of unequally-sized categories, (2) using a superior measure of the dispersion of frequencies over categories, (3) not permitting Um to be smaller than a certain minimum value greater than zero, even when all occurrences are concentrated in a single category, and (4) scaling the coefficient in terms of a corpus of a "standard million" tokens. Computations are given for illustrative data and discussed.