This paper describes a system for unsuper- vised dependency parsing based on Gibbs sampling algorithm. The novel approach in- troduces a fertility model and reducibility model, which assumes that dependent words can be removed from a sentence without vio- lating its syntactic correctness.