Meet Low-resource RE!

An Empirical Benchmark for Low-resource Relation Extraction.

Different Schemes for Low-resource Relation Extraction

Standard PLM Fine-tuning

Supervised Pre-trained Models Fine-tuning with strong baselines, such as OpenNRE.

Prompt-based Tuning

Prompting for Few-shot Instances with KnowPrompt and PTR.

Balancing Methods

Re-sampling Methods and Re-weighting Losses for
Long-tailed Distribution Issues.

Leveraging More Instances

Diverse Data Augmentation Methods and Self-training
to Generate More Labeled Data
from Easy-collected Unlabeled Data.

What is Low-resource Relation Extraction Benchmark?

The Low-resource Relation Extraciton benchmark is a comprehensively empirical study on low-resource RE. It focuses on two challenges:

Few-shot Relation Extraction
RE Data with Long-tailed Distribution

We hope this study can help inspire future research for low-resource RE with more robust models and promote transitioning the technology to real-world industrial scenarios.

Paper

Please cite our paper as below if you use the benchmark or codebase.


@article{LREBench2022,
  title={Towards Realistic Low-resource Relation Extraction: A Benchmark with Empirical Baseline Study},
  author={Xin Xu, Xiang Chen, Ningyu Zhang, Xin Xie, Xi Chen, Huajun Chen},
  journal={EMNLP},
  year={2022}
}