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

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}
}

Overview

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.

Data

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.

Setup

This implemetation is based on Python3.7. To run the code, you need the following dependencies:

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

You can simply run

pip install -r requirements.txt

Repository structure

We select some important files for detailed description.

|-- data # experiments for five datasets
    |-- Few-COMM/ # a Chinese few-shot NER dataset we released
|-- model # includes all model implementations
|-- transformer_model # includes BERT pre-trained checkpoints
	|--bert-base-chinese
	|--bert-base-uncased
|-- utils
    |-- config.py # configuration
    |-- data_loader.py # load data
    |-- fewshotsampler.py # construst meta-tasks
    |-- framework.py # includes train\eval\test procedure
    |-- tripletloss.py # an improved triplet loss
|-- main.py
|-- run.sh

Quickstart

Unzip our processed data file data.zip and put the data files under data/ folder.
Please download pretrained BERT files bert-base-chinese and bert-base-uncased and put them under transformer_model/ folder.
sh run.sh

You can also adjust the model by modifying the parameters in the run.sh file.

Currently, the benchmarks on the FEW-COMM dataset are as follows:

FEW-COMM	5-way 1-shot	5-way 5-shot	10-way 1-shot	10-way 5-shot
MAML	28.16	54.38	26.23	44.66
NNShot	48.40	71.55	41.75	67.91
StructShot	48.61	70.62	47.77	65.09
PROTO	22.73	53.95	22.17	45.81
CONTaiNER	57.13	63.38	51.87	60.98
ESD	65.37	73.29	58.32	70.93
DecomMETA	68.01	72.89	62.13	72.14
SpanProto	70.97	76.59	63.94	74.67
MeTNet	71.89	78.14	65.11	77.58

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

For the FewNERD dataset, please download it from its official website.

Attribution

Parts of this code are based on the following repositories: