Learning Markov logic networks with limited number of labeled training examples

Abstract

Markov Logic Networks (MLN) is a unified framework integrating first-order logic and probabilistic inference. Most existing methods of MLN learning are supervised approaches requiring a large amount of training examples, leading to a substantial amount of human effort for preparing these training examples. To reduce such human effort, we have developed a semi-supervised framework for learning an MLN, in particular structure learning of MLN, from a set of unlabeled data and a limited number of labeled training examples. To achieve this, we aim at maximizing the expected pseudo-log-likelihood function of the observation from the set of unlabeled data, instead of maximizing the pseudo-log-likelihood function of the labeled training examples, which is commonly used in supervised learning of MLN. To evaluate our proposed method, we have conducted experiments on two different datasets and the empirical results demonstrate that our framework is effective, outperforming existing approach which considers labeled training examples alone.