English | 简体中文

Illustration of LightNER (COLING'22) for few-shot named entity recognition (Details in paper LightNER: A Lightweight Tuning Paradigm for Low-resource NER via Pluggable Prompting).

• ❗NOTE: We have released a follow-up work "One Model for All Domains: Collaborative Domain-Prefix Tuning for Cross-Domain NER" at CP-NER.

python == 3.8

• torch == 1.11
• transformers == 4.26.0
• deepke

git clone https://github.com/zjunlp/DeepKE.git
cd DeepKE/example/ner/few-shot

• Create and enter the python virtual environment.
• Install dependencies: pip install -r requirements.txt.

• Dataset

  • Download the dataset to this directory.

    wget 120.27.214.45/Data/ner/few_shot/data.tar.gz
    tar -xzvf data.tar.gz

  • The datasets are stored in data, including CoNLL-2003, MIT-movie, MIT-restaurant and ATIS.

  • CoNLL-2003

    • train.txt: Training set
    • valid.txt : Validation set
    • test.txt: Test set
    • indomain-train.txt: In-domain training set

  • MIT-movie, MIT-restaurant and ATIS

    • k-shot-train.txt: k=[10, 20, 50, 100, 200, 500], Training set
    • test.txt: Testing set

• Training

  • Parameters, model paths and configuration for training are in the conf folder and users can modify them before training.

  • Training on CoNLL-2003

    python run.py

  • Few-shot Training

    If the model need to be uploaded, modify load_path in few_shot.yaml

    python run.py +train=few_shot

  • Few-shot Training for Chinese

    Full data fine-tuning can achieve the best performance.

    Pretrained weights need to be provided in the directory defined in few_shot_cn.yaml

    python run.py +train=few_shot_cn

  • Logs for training are in the log folder. The path of the trained model can be customized.

• Prediction

  • Add - predict in config.yaml

  • Modify load_path as the path of the trained model and write_path as the path of predicted results in predict.yaml

  • python predict.py

If you need to customize your own Tokenizer (eg MBartTokenizer for multilingual processing).

You can customize the tokenizer in tokenizer

If you use or extend our work, please cite the following paper:

@inproceedings{DBLP:conf/coling/00160DTXHSCZ22,
  author    = {Xiang Chen and
               Lei Li and
               Shumin Deng and
               Chuanqi Tan and
               Changliang Xu and
               Fei Huang and
               Luo Si and
               Huajun Chen and
               Ningyu Zhang},
  editor    = {Nicoletta Calzolari and
               Chu{-}Ren Huang and
               Hansaem Kim and
               James Pustejovsky and
               Leo Wanner and
               Key{-}Sun Choi and
               Pum{-}Mo Ryu and
               Hsin{-}Hsi Chen and
               Lucia Donatelli and
               Heng Ji and
               Sadao Kurohashi and
               Patrizia Paggio and
               Nianwen Xue and
               Seokhwan Kim and
               Younggyun Hahm and
               Zhong He and
               Tony Kyungil Lee and
               Enrico Santus and
               Francis Bond and
               Seung{-}Hoon Na},
  title     = {LightNER: {A} Lightweight Tuning Paradigm for Low-resource {NER} via
               Pluggable Prompting},
  booktitle = {Proceedings of the 29th International Conference on Computational
               Linguistics, {COLING} 2022, Gyeongju, Republic of Korea, October 12-17,
               2022},
  pages     = {2374--2387},
  publisher = {International Committee on Computational Linguistics},
  year      = {2022},
  url       = {https://aclanthology.org/2022.coling-1.209},
  timestamp = {Mon, 13 Mar 2023 11:20:33 +0100},
  biburl    = {https://dblp.org/rec/conf/coling/00160DTXHSCZ22.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}