FSRE meta dataset

FSRE meta dataset是一个专为少样本关系抽取设计的综合数据集，包含三个子数据集：基于NYT29的少样本数据集，基于WIKIDATA的少样本数据集，以及TACRED的少样本变体。这些数据集通过严格的转换过程，确保测试关系与模型可能见过的任何关系不同，训练数据有限，并且大多数候选关系提及不对应于任何感兴趣的关系。该数据集旨在为少样本关系抽取提供一个真实的评估基准，适用于自然语言处理中的问题回答、事件抽取和知识库填充等任务。

The FSRE Meta Dataset is a comprehensive benchmark dataset specifically designed for few-shot relation extraction. It consists of three sub-datasets: a few-shot dataset based on NYT29, a few-shot dataset based on WIKIDATA, and a few-shot variant of TACRED. Through a strict conversion process, these datasets ensure that the test relations are completely distinct from any relations the model may have encountered before, with limited training data, and that most candidate relation mentions do not correspond to any target relations of interest. This dataset aims to provide a realistic evaluation benchmark for few-shot relation extraction, and is applicable to tasks including question answering, event extraction, and knowledge base filling in the field of natural language processing.

搜集汇总

背景与挑战

背景概述

FSRE meta dataset是一个专为少样本关系抽取设计的综合数据集，包含基于NYT29、WIKIDATA和TACRED的三个子数据集。其特点是通过严格转换确保测试关系新颖、训练数据有限且多数候选关系不直接对应目标关系，旨在为自然语言处理任务如问题回答、事件抽取和知识库填充提供真实评估基准。