izhx / NER-unlabeled-data-retrieval

[COLING 22] Domain-Specific NER via Retrieving Correlated Samples.

NER-unlabeled-data-retrieval

[COLING 22] Domain-Specific NER via Retrieving Correlated Samples.

arxiv http://arxiv.org/abs/2208.12995 is with updated address devset results.

Usage

This code depends on the AllenNLP library, see requirements.txt.

To train a baseline NEZHA-BiLSTM-CRF model on the address dataset: python main.py --config=add --cuda=0 --name=RUN_NAME .

To train a baseline Cross-Encoder model on the address dataset: python main.py --config=add-ret --cuda=0 --name=RUN_NAME .

The above add and add-ret correspond to the filename in the config/ dictionary. They could be replaced with eco and eco-ret to run the e-commerce dataset experiments.

Notes:

1. In each config file, plm_dir and plm_name are used to set the path of pretrained models from huggingface or local filepath, they can be set to plm_dir="" and plm_name="bert-base-chinese" to load a Chinese BERT from huggingface.

Data

The address domain dataset is comes from CCKS2021中文地址要素解析数据集.

The e-commerce domain dataset is comes from MultiDigraphNER.

For some reasons, we can not provide the testset of the address domain (not the final_test.txt in CCKS2021中文地址要素解析数据集), and all retrieved correlated texts. (I really hope I can...)

All experiments are conducted at Alibaba Damo Academy, the results in the paper are real.

Cite

@inproceedings{zhang-etal-2022-domain,
    title = "Domain-Specific {NER} via Retrieving Correlated Samples",
    author = "Zhang, Xin  and
      Jiang, Yong  and
      Wang, Xiaobin  and
      Hu, Xuming  and
      Sun, Yueheng  and
      Xie, Pengjun  and
      Zhang, Meishan",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2022.coling-1.211",
    pages = "2398--2404",
}