Fine-Grained Common Knowledge Learning for Domain Adaptive Few-Shot Relation Extraction

Minghui Zhai, Feifei Dai, Xiaoyan Gu, Chuanrong Li, Bo Li, Weiping Wang

Abstract: As few-shot relation extraction (FSRE) can accurately extract relations in the scenario of lacking labeled data, it has been widely used in natural language processing. However, their effectiveness drops dramatically when models need to be adapted to new domains. Then, domain adaptive FSRE is proposed. Nevertheless, existing works learn too much knowledge that is specific to the source domain thus weakening the common knowledge between domains. Moreover, the task-level common knowledge they learned is coarse-grained. Therefore, they cannot effectively use the learned knowledge to guide relation extraction in the target domain. In this paper, we propose a fine-grained common knowledge learning method. It focuses on learning the explicit and implicit common knowledge between domains and transforming them into fine-grained knowledge. This can effectively use the common knowledge to guide the relation extraction in the target domain. We conduct domain adaptive few-shot experiments on the FewRel 2.0 dataset. The results show an average gain of 2.81% on strong baselines for the accuracy of our model. Our code is available at: https://anonymous.4open.science/r/FCKL-BF7F.