Exploiting unannotated natural language data is hard largely because unsupervised parameter estimation is hard. We describe deterministic annealing (Rose et al., 1990) as an appealing alternative to the ExpectationMaximization algorithm (Dempster et al., 1977).