On extracting relations using distributional semantics and a tree generalization

Abstract

Extracting relations out of unstructured text is essential for a wide range of applications. Minimal human effort, scalability and high precision are desirable characteristics. We introduce a distant supervised closed relation extraction approach based on distributional semantics and a tree generalization. Our approach uses training data obtained from a reference knowledge base to derive dependency parse trees that might express a relation. It then uses a novel generalization algorithm to construct dependency tree patterns for the relation. Distributional semantics are used to eliminate false candidate patterns. We evaluate the performance in experiments on a large corpus using ninety target relations. Our evaluation results suggest that our approach achieves a higher precision than two state-of-the-art systems. Moreover, our results also underpin the scalability of our approach. Our open source implementation can be found at https://github.com/dice-group/Ocelot.