DPRL: Labeling Relation Based on Distant Supervision and POS Rule

The lack of high-quality labeled data is one of the main challenges to a lot of relation extraction tasks. Distant supervision introduces many noises, while labeling data manually is costly. In this paper, we design a novel labeling rule called the POS rule. POS rule improves the precision of relation extraction by considering the position, ontology, and semantic of relations. We propose a POS rule mining and optimization framework to generate high-quality labeled data. Experiments on real datasets show that the proposed discovery algorithms can find high-quality POS rules which accurately label training data for relation extraction.