Few-Shot Relation Extraction Based on Prompt Learning: A Taxonomy, Survey, Challenges and Future Directions

Abstract

Relation extraction (RE) is critical in information extraction (IE) and knowledge graph construction. RE aims to identify the semantic relations between entities from natural language texts. Traditional RE models often rely on many manually annotated training samples, which are limited when data is scarce. Therefore, exploring how to perform relation extraction under few-shot conditions has become a research focus. Recently, prompt learning has attracted attention from researchers due to its ability to fully activate the potential of Pre-trained Language Models (PLMs), especially making significant progress in Few-Shot Relation Extraction (FSRE). This paper comprehensively reviews FSRE based on prompt learning. We first introduce the fundamental concepts of FSRE and prompt learning. Then, we systematically review recent research advances in FSRE with prompt learning, focusing on two perspectives: template construction and model fine-tuning strategies. Next, we summarize the benchmark datasets, evaluation metrics, and experimental results of representative works in FSRE. Afterward, we present practical applications of prompt-based FSRE in specialized domains. Finally, we discuss the critical challenges and future research directions of FSRE tasks based on prompt learning.