Unsupervised Bayesian Parameter Estimation for Dependency Parsing

We explore a new Bayesian model for probabilistic grammars, a family of distributions over discrete structures that includes hidden Markov models and probabilitsic context-free grammars. Our model extends the correlated topic model framework to probabilistic grammars, exploiting the logistic normal prior as a prior over the grammar parameters. We derive a variational EM algorithm for that model, and then experiment with the task of unsupervised grammar induction for natural language dependency pasring, and show that our model achieves superior results over previous models that differ mostly in the choice of a prior different than the logistic-normal prior. For a full version of this paper, see [5].