Liang, Percy and Klein, Dan and Jordan, Michael
Given a model family and unlabeled examples, what is the most cost-effective way of estimating model parameters? To address this question, we present a Bayesian decision-theoretic framework for learning from measurements, which unifies various labeling schemes with more general constraints and preferences, as well as providing a principle for actively choosing measurements. We introduce variational and stochastic approximations for inference, which allow us to scale up to real-world sequence labeling tasks.
Discussion