Understanding and Mitigation of Noise in Crowd Sourced Relation Classification Dataset

Abstract

Relation classification (RC), a task of classifying the relation between a given pair of entities in a sentence to a relation newlinelabel is fundamental to IE systems. The identified structured triple (subject_entity, relation, object_entity) from the newlineunstructured text can vastly help in knowledge base completion. This organized relational knowledge can further be newlineused for other downstream tasks like question-answering, and common-sense reasoning. A large RC dataset TACRED newlinehas been widely used for benchmarking modern deep neural models. However, RC at a large scale is restricted mainly due to the presence of noise in the training dataset. Hence, the performance of such advanced deep neural models, which have shown excellent improvement on other NLP tasks, has been held back for RC.