Improving Relation Classification Using Relation Hierarchy

Akshay Parekh, Ashish Anand, Amit Awekar

Abstract: The relation classification (RC) task classifies a relation present between two target entities in a given context. It is an important task of information extraction and plays a significant role in several NLP applications. Most of the existing studies consider relation classes as a flat list of classes and thus ignore hierarchical relation between classes. This study explores the application of relation hierarchy in improving relation classification performance. In particular, we focus on the following two applications of relation hierarchy: (i) detecting noisy instances; and (ii) modifying the cross-entropy (CE) loss function. We use TACRED, the most widely used RC dataset for this purpose. We build a taxonomical relation hierarchy over the relation classes of TACRED and use it for filtering and relabeling ambiguous or noisy instances of TACRED. For better optimization, we also introduce hierarchical-distance scaled cross-entropy loss (HCE Loss), using the shortest-path distance between ground truth and predicted label for scaling cross-entropy loss. Our extensive empirical analyses indicate that relation hierarchy-inspired filtering, relabeling, and the HCE loss help in improving the relation classification.