Semantics-Guided Knowledge Integration for Domain Adaptation Few-shot Relation Extraction

Few-shot relation extraction achieves great progress by incorporating meta-learning with knowledge. Most of the recent works focus on integrating different types of knowledge, which ignored that there would be a huge knowledge gap in the application, especially facing domain adaptation. In this paper, we explore the effective way to integrate different forms of knowledge for domain adaptation few-shot relation extraction, and we face up to the problem of knowledge gap between training and testing. To be specific, we take the entity concept as the integrated knowledge, and the concept representation can be obtained from two sources, including textual label and concept graph. For the textual label, we evaluate the model's perception of concepts in different types of text fusion. For the concept graph, we pursue an effective integration method to adapt the modal gap and knowledge gap. Extensive experiments on the Few Rei domain adaptation dataset show that textual knowledge with language models and graph knowledge with a distance scorer are easy for transferring, and semantic information can effectively guide the integration process.