An important challenge in extracting useful structured information from text collections is relation extraction, i.e. identifying the types of relations between entities that are expressed in text. Due to the variety in how relations are rendered in language, labeling data for relation extraction is unfortunately notoriously time-consuming and expensive. Recently proposed embedding-based extractors that utilize unlabeled data and use noisy KB alignments as “distant labels” partially address this concern. However not only are these models inaccurate for relations that do not have large KB, but further, cannot be improved upon without annotating data. Purely rule-based systems, on the other hand, provide an attractive alternative as they allow users to directly inject symbolic domain knowledge, however require a large number of formulae to achieve reasonable generalization.

In this talk, I introduce an interactive training paradigm that combines embedding-based models of relation extraction with symbolic domain knowledge. I first describe how symbolic domain knowledge, if provided by the user as first-order logic statements, can be injected into the embeddings to improve the predictions. In the second part of the talk, I present an approach to “explain” the embedding-based model predictions using a symbolic representation, which the user can annotate directly for more effective supervision. I present experiments that demonstrate the potential of symbolic knowledge as supervision in reducing annotation effort and in quickly training accurate relation extraction systems.

This work is a collaboration with Tim Rocktaschel, Sebastian Riedel, Luke Zettlemoyer, and Carlos Guestrin.