HGN

Code for Hero-Gang Neural Model For Named Entity Recognition (Accepted in NAACL-2022)

Citations

If you use or extend our work, please cite our paper at NAACL-2022.

@inproceedings{hu-etal-2022-hero,
    title = "Hero-Gang Neural Model For Named Entity Recognition",
    author = "Hu, Jinpeng  and
      Shen, Yaling  and
      Liu, Yang  and
      Wan, Xiang  and
      Chang, Tsung-Hui",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jul,
    year = "2022"
}

Requirements

• Python 3 (tested on 3.7)
• PyTorch (tested on 1.5 and 1.7)

Data

We give an example dataset in data/W16_bio

Training

To start training, run

export MODEL=xlnet-large-cased
epoch=20
lr=5e-5
wis=1qq3qq5qq7
data_type=W16_bio
connect_type=dot-att

CUDA_VISIBLE_DEVICES=0 python run_hgn.py \
--train_data_dir=data/$data_type/train_merge.txt \
--dev_data_dir=data/$data_type/dev.txt \
--test_data_dir=data/$data_type/test.txt \
--bert_model=${MODEL} \
--task_name=ner \
--output_dir=./output/xlnet_multi_window_${lr}_win_size_${wis}_epoch_${epoch}_${connect_type} \
--max_seq_length=128 \
--num_train_epochs ${epoch} \
--do_train \
--gpu_id 0 \
--learning_rate ${lr} \
--warmup_proportion=0.1 \s
--train_batch_size=32 \
--use_bilstm \
--use_multiple_window \
--windows_list=${wis} \
--connect_type=${connect_type}

In this bash, Model is the path to your pre-trained model (such as BERT, XLNET or BioBERT), windows_list is the hyperparameter that control the windows. For example, 1qq3qq5qq7 means that we utilize 4 different windows and their sizes are 1, 3, 5 and 7 respectively. connect_type can be mlp-att or dot-att.

Evaluation

CUDA_VISIBLE_DEVICES=0 python run_hgn.py \
--train_data_dir=data/$data_type/train_merge.txt \
--dev_data_dir=data/$data_type/dev.txt \
--test_data_dir=data/$data_type/test.txt \
--bert_model=${MODEL} \
--task_name=ner \
--output_dir=./saved_model_path \
--max_seq_length=128 \
--num_train_epochs ${epoch} \
--do_predict \
--gpu_id 0 \
--learning_rate ${lr} \
--warmup_proportion=0.1 \
--train_batch_size=32 \
--use_bilstm \
--use_multiple_window \
--windows_list=${wis} \
--connect_type=${connect_type}

Pre-trained model

The pre-trained HGN models on W16 and W17. You can download the models and run them on the corresponding datasets to replicate our results.

Section	BaiduNetDisk	Description
W16	download (Password: jcjp)	HGN model trained on XLNET
W17	download (Password: geic)	HGN model trained on XLNET