This repository contains the code and data for our paper:

Meta-Learning Triplet Network with Adaptive Margins for Few-Shot Named Entity Recognition

If you find this work useful and use it on your own research, please cite our paper.

@article{han2023meta,
  title={Meta-Learning Triplet Network with Adaptive Margins for Few-Shot Named Entity Recognition},
  author={Chengcheng Han and 
          Renyu Zhu and 
          Jun Kuang and 
          FengJiao Chen and
          Xiang Li and
          Ming Gao and
          Xuezhi Cao and
          Wei Wu},
  journal={arXiv preprint arXiv:2302.07739},
  year={2023}
}

We propose an improved triplet network with adaptive margins (MeTNet) and a new inference procedure for few-shot NER.

We release the first Chinese few-shot NER dataset FEW-COMM.

The datasets used by our experiments are in the data/ folder, including FEW-COMM, FEW-NERD, WNUT17, Restaurant and Multiwoz.

FEW-COMM is a Chinese few-shot NER dataset we released, which consists of 66,165 product description texts that merchants display on a large e-commerce platform, including 140,936 entities and 92 pre-defined entity types. These entity types are various commodity attributes that are manually defined by domain experts, such as "material", "color" and "origin". Please see Appendix C of our paper for more details on the dataset.

Our code will be open-sourced soon. You can use Few-COMM in your experiments. The data format of FEW-COMM is the same as that of FEW-NERD, so you can use the DataLoader in FEW-NERD. Currently, the benchmarks on the FEW-COMM dataset are as follows:

If you have the latest experimental results on the FEW-COMM dataset, please contact us to update the benchmark.

• Python 3.8
• nltk>=3.6.4
• numpy==1.21.0
• pandas==1.3.5
• torch==1.7.1
• transformers==4.0.1
• apex==0.9.10dev
• scikit_learn==0.24.1
• seqeval

Parts of this code are based on the following repositories:

• FewNERD
• MLADA