Low-resource extraction with knowledge-aware pairwise prototype learning

Abstract Knowledge Extraction (KE) aims at extracting structured information from raw texts, such as relation extraction and event extraction. One of the major issues for KE is the low-resource problem due to deficient samples. Previous work addresses the low-resource issue mostly via data-driven methods, such as transfer learning, while neglecting correlation knowledge among classes. For example, inherent correlation of entailment between the relation pair and causality between the event pair can also be utilized for low-resource KE. Consequently, we propose to leverage correlation knowledge via pairwise prototype learning on the hypersphere with a novel framework called Knowledge-aware Hyperspherical Prototype Network (K-HPN). K-HPN is able to recognize inherent correlation among classes, where each class is represented as a prototype on the hypersphere. The experimental results demonstrate that K-HPN outperforms previous methods of KE, particularly with low-resource training data regimes.