Given a weighted bipartite graph, the assignment problem consists of finding the heaviest perfect match. This is a classical problem in combinatorial optimization, which is solvable exactly and efficiently by standard methods such as the Hungarian algorithm, and is widely applicable in real-world scenarios. We give an exponential family model for the assignment problem. Edge weights are obtained from a suitable composition of edge features and a parameter vector, which is learned so as to maximize the likelihood of a sample consisting of training graphs and their labeled matches. The resulting consistent estimator contrasts with existing max-margin structured estimators, which are inconsistent for this problem.