Improving Markov Logic Network learning using unlabeled data

Abstract

Existing Markov Logic Network (MLN) learning methods aim at learning an MLN from a set of training examples. To reduce the human effort in preparing training examples, we have developed a semi-supervised framework for learning an MLN from unlabeled data and a limited number of training examples. One characteristic of our approach is that instead of maximizing the pseudo-log-likelihood function of the labeled training examples, we aim at optimizing the pseudo-log-likelihood function of the observation from the set of unlabeled data. The learned MLN can then be applied to the unlabeled data for conducting inference in a more precise manner. We have conducted experiments and the empirical results demonstrate that our framework is effective, outperforming existing approach which considers labeled training examples alone.