LeKAN: Extracting Long-tail Relations via Layer-Enhanced Knowledge-Aggregation Networks

Long-tailed relation extraction is a crucial task in the information extraction field for extracting the long-tailed, imbalanced relation between two annotated entities based on related context. Although many works have been devoted to distinguishing valid instances from noisy data and have achieved promising performance, such studies still have critical defects: works based on nonhierarchical relations ignore the correlations among the relations, and those based on hierarchical relations neglect the hierarchy of the relation structure, which is unbalanced and causes difficulty in extracting data-poor classes. In this paper, a novel layer-enhanced knowledge aggregation network, named LeKAN, is presented to classify the relations between two annotated entities from text, especially long-tailed relations, which are very common in various corpora. Inspired by the election mechanism, we aggregate the ancestors of long-tailed relation classes into new relation representations to prevent the long-tailed relations from being ignored. Specifically, we use GraphSAGE to learn the relational knowledge from an existing knowledge graph via class embedding. Moreover, we aggregate the acquired relational knowledge into the LeKAN by layer-enhanced knowledge-aggregating attention mechanism. Comprehensive experimental results demonstrate that the new method yields considerable improvement over other relation extraction methods on a large-scale benchmark dataset with a long-tailed distribution.