Adversarial hard negative samples for continual relation extraction

Abstract

Continual relation extraction (CRE) is a crucial task in continuous learning, aiming to train a model continually on data of new relations to extract relations between entities from unstructured text. This process involves learning newly emerged relations while avoiding catastrophic forgetting of previously learned relations. Existing works have demonstrated that storing a few typical samples of old relations in memory and replaying them during subsequent training for new relations is helpful. It can assist the model to maintain a stable understanding of old relations, thus effectively avoiding forgetting them. However, most prior research has focused on efficiently utilizing memory samples while neglecting their learning difficulty, which refers to the challenge in mastering these samples. In this paper, we propose an adversarial hard negative samples selection mechanism to increase the diversity of memory samples and dynamically adjust the number of samples among relations according to the performance of the model. Experimental results show that our method consistently improves the performance of state-of-the-art CRE models without increasing the number of training samples on mainstream benchmarks.