Abstract
We propose a multi-task, probabilistic approach to facilitate distantlysupervised relation extraction by bringing closer the representations ofsentences that contain the same Knowledge Base pairs. To achieve this, we biasthe latent space of sentences via a Variational Autoencoder (VAE) that istrained jointly with a relation classifier. The latent code guides the pairrepresentations and influences sentence reconstruction. Experimental results ontwo datasets created via distant supervision indicate that multi-task learningresults in performance benefits. Additional exploration of employing KnowledgeBase priors into the VAE reveals that the sentence space can be shifted towardsthat of the Knowledge Base, offering interpretability and further improvingresults.