文献类型：期刊文献

英文题名：A Survey of Research on Few-Shot Relation Classification

作者：Zhang, YaChuan[1]; Guo, Yi[1,2,3]

机构：[1] Department of Computer Science and Engineering, East China University of Science and Technology, Shanghai, 200237, China; [2] Shanghai Engineering Research Center of Big Data Internet Audience, Shanghai, China; [3] Business Intelligence and Visualization Research Center, National Engineering Laboratory for Big Data Distribution and Exchange Technologies, Shanghai, China

年份：2024

外文期刊名：SSRN

收录：EI(收录号：20240192236)

语种：英文

外文关键词：Classification (of information)

摘要：Relation classification is a crucial sub-task of information extraction, playing a significant role in extracting structured information from vast texts and constructing knowledge graphs. However, in practical scenarios, the availability of labeled data is often limited, and some datasets exhibit a long-tail distribution. To address these challenges, few-shot learning relation classification (FSRC) has emerged as a promising approach. The purpose of FSRC is to learn high-quality features with insufficient data for distinguishing the relation between given entities in a sentence. This task presents greater challenges compared to traditional relation classification tasks. Recently, FSRC has gained a lot of attention, prompting numerous researchers to provide innovative methods, datasets, and ideas to tackle this challenge. This paper focuses on relation classification based on few-shot learning and provides a comprehensive review of state-of-the-art techniques. Based on distinct research characteristics, we classified FSRC methods into five categories: metric-based method, optimization-based method, pre-trained and LLMs based method, prompt-based method, and lower source situation. We analyze the advantages and disadvantages of each category, provide a summary of the datasets, metrics, and results obtained by representative models. Finally, the paper concludes the current limitations of the few-shot relation classification and proposes future research directions. ? 2024, The Authors. All rights reserved.