A joint training network for learning more distinguishable relation features in relation classification

Abstract

Relation classification is an important task in natural language processing, which aims to predict the semantic relation between a given entity pair in a sentence. There are datasets, like TACRED, that contain a large number of “no_relation” type samples. Most existing methods treat “no_relation” and normal relation types equally, and directly apply the softmax function over all relation types. In this paper, we propose a novel joint training network to learn more distinguishable relation features for relation classification. Specially, we convert the original multi-class classification problem into two joint optimized modules, binary classification of whether a relation is “no_relation” and multi-class classification of normal relation types. To further differentiate between similar normal relation types, we introduce a self-supervised contrastive learning method to learn more distinguishable features for them. We jointly optimize the above modules. Experimental results agree well with our design intention and demonstrate that our joint training network not only achieves superior performance against existing competitive models, but also is robust to “no_relation” problem.