ConNER: Consistency Training
for Cross-lingual Named Entity Recognition

This repository contains the code for EMNLP 22 paper "ConNER: Consistency Training for Cross-lingual Named Entity Recognition"

Requirements

Python >= 3.7
PyTorch >= 1.7
Transformers >= 3.5
conlleval = 0.2

Our experiments are run on a single Nvidia V100 32GB GPU.

Data Format

We have released the preprocessed data used in our experiments under the data directory. To train on your customized datasets, follow the data format as below.

Training train.txt, validation dev.txt and testing test.txt files adopt the CoNLL two-column format. Unlabeled data file unlabel.txt and its translation trans.txt use an extra column in between to indicate span-alignment, which can be automatically obtained using our alignment-free translation.

Training

To train ConNER on cross-lingual NER (for example, English to German transfer), use the following commands:

cd model/En2De
bash 01_xlmr_train_eval.sh

All hyperparameters, such as loss weight coefficients UDA_W, RDROP_W, batch size BS are specified in the bash script.

Alignment-free Translation

Place your Google API key google_translate.json under the translate directory.

Run the following commands to conduct alignment-free translation on translate/sample.txt

cd translate
python 01_translate.py  # translate unlabeled data
python 02_clean.py      # cleaning the raw translated data
python 03_lin2cols.py   # convert inline sentences into CoNLL's two column format
python 04_col2align.py  # add span alignment indicator

The generated files unlabel.txt and trans.txt will be used for ConNER training.

Citation

If you find this repository useful, please cite our paper:

@inproceedings{zhou2022conner,
  title={ConNER: Consistency Training for Cross-lingual Named Entity Recognition},
  author={Zhou, Ran and Li, Xin and Bing, Lidong and Cambria, Erik and Si, Luo and Miao, Chunyan},
  booktitle={Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing},
  year={2022}
}

Acknowledgements

The training code in this repo has been built upon pytorch-neural-crf

RandyZhouRan/ConNER