Spoken dialog systems help users achieve a task using natural language. Noisy speech recognition and ambiguity in natural language motivate statistical approaches that exploit distributions over the user's goal at every step in the dialog. The task of tracking these distributions, termed Dialog State Tracking, is therefore an essential component of any Spoken dialog system. In recent years, the Dialog State Tracking Challenges have provided a common test-bed and evaluation framework for this task, as well as labeled dialog data. As a result, a variety of machine-learned methods have been successfully applied to Dialog State Tracking. This paper reviews the machine-learning techniques that have been adapted to Dialog State Tracking, and gives an overview of published evaluations. Discriminative machine-learned methods outperform generative and rule-based methods, the previous state-of-the-art.