LexiconNER
This is the implementation of "Distantly Supervised Named Entity Recognition using Positive-Unlabeled Learning" published at ACL 2019. The highlight of this work is it performs NER using only entity dictionaries without any labeled data.
By the way, we recently publish our another work related to Chinese NER. It designs to augment Chinese NER with lexicons. The highlight of this work is that it has high computational efficiency and at the same time, achieves comparative or better performance over existing methods. You can access the source code of that work and a hyper-link of its associated paper at LexiconAugmentedNER.
Set up and run
Download glove.6B.100d.txt
Environment
pytorch 1.1.0 python 3.6.4 cuda 8.0
Instructions for running code
Phrase one <train bnPU model>
Train
Print parameters
run python feature_pu_model.py --h
optional arguments:
-h, --help show this help message and exit
--lr LR learning rate
--beta BETA beta of pu learning (default 0.0)
--gamma GAMMA gamma of pu learning (default 1.0)
--drop_out DROP_OUT dropout rate
--m M class balance rate
--flag FLAG entity type (PER/LOC/ORG/MISC)
--dataset DATASET name of the dataset
--batch_size BATCH_SIZE
batch size for training and testing
--print_time PRINT_TIME
epochs for printing result
--pert PERT percentage of data use for training
--type TYPE pu learning type (bnpu/bpu/upu)e.g.)
Train on PER type of conll2003 dataset:
python feature_pu_model.py --dataset conll2003 --type PER
** Evaluating**
python feature_pu_model_evl.py --model saved_model/bnpu_conll2003_PER_lr_0.0001_prior_0.3_beta_0.0_gamma_1.0_percent_1.0 --flag PER --dataset conll2003 --output 1replace the model name from the training
python final_evl.py Get the final result on all the entity type. Remember to revise the filenames to be the output file name of evaluating.
Phrase two <train adaPU model>
dictionary generation
run python ada_dict_generation.py -h
optional arguments:
-h, --help show this help message and exit
--beta BETA learning rate
--gamma GAMMA gamma of pu learning (default 1.0)
--drop_out DROP_OUT dropout rate
--m M class balance rate
--flag FLAG entity type (PER/LOC/ORG/MISC)
--dataset DATASET name of the dataset
--lr LR learning rate
--batch_size BATCH_SIZE
batch size for training and testing
--iter ITER iteration time
--unlabeled UNLABELED
use unlabeled data or not
--pert PERT percentage of data use for training
--model MODEL saved model namee.g.)
python ada_dict_generation.py --model saved_model/bnpu_conll2003_PER_lr_0.0001_prior_0.3_beta_0.0_gamma_1.0_percent_1.0 --flag PER --iter 1
adaptive training
run python adaptivepumodel.py -h
optional arguments:
-h, --help show this help message and exit
--beta BETA beta of pu learning (default 0.0)
--gamma GAMMA gamma of pu learning (default 1.0)
--drop_out DROP_OUT dropout rate
--m M class balance rate
--p P estimate value of prior
--flag FLAG entity type (PER/LOC/ORG/MISC)
--dataset DATASET name of the dataset
--lr LR learning rate
--batch_size BATCH_SIZE
batch size for training and testing
--output OUTPUT write the test result, set 1 for writing result to
file
--model MODEL saved model name
--iter ITER iteration timee.g.)
python adaptive\_pu\_model.py --model saved\_model/bnpu\_conll2003\_PER\_lr\_0.0001\_prior\_0.3\_beta\_0.0\_gamma\_1.0\_percent\_1.0 --flag PER --iter 1
Replace saved model names and iteration times when doing adaptive learning. And in the same iteration the iter number in dictionary generation and adaptive learning should be same.
Cite
Please cite our ACL 2019 paper:
@article{peng2019distantly,
title={Distantly Supervised Named Entity Recognition using Positive-Unlabeled Learning},
author={Peng, Minlong and Xing, Xiaoyu and Zhang, Qi and Fu, Jinlan and Huang, Xuanjing},
journal={Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL)},
year={2019}
}