Code for Two-stage Identifier: "Locate and Label: A Two-stage Identifier for Nested Named Entity Recognition", accepted at ACL 2021. For details of the model and experiments, please see our paper.
conda create --name acl python=3.8
conda activate acl
pip install -r requirements.txtThe datasets used in our experiments:
- ACE04: https://catalog.ldc.upenn.edu/LDC2005T09
- ACE05: https://catalog.ldc.upenn.edu/LDC2006T06
- KBP17: https://catalog.ldc.upenn.edu/LDC2017D55
- GENIA: http://www.geniaproject.org/genia-corpus
Data format:
{
"tokens": ["2004-12-20T15:37:00", "Microscopic", "microcap", "Everlast", ",", "mainly", "a", "maker", "of", "boxing", "equipment", ",", "has", "soared", "over", "the", "last", "several", "days", "thanks", "to", "a", "licensing", "deal", "with", "Jacques", "Moret", "allowing", "Moret", "to", "buy", "out", "their", "women", "'s", "apparel", "license", "for", "$", "30", "million", ",", "on", "top", "of", "a", "$", "12.5", "million", "payment", "now", "."],
"pos": ["JJ", "JJ", "NN", "NNP", ",", "RB", "DT", "NN", "IN", "NN", "NN", ",", "VBZ", "VBN", "IN", "DT", "JJ", "JJ", "NNS", "NNS", "TO", "DT", "NN", "NN", "IN", "NNP", "NNP", "VBG", "NNP", "TO", "VB", "RP", "PRP$", "NNS", "POS", "NN", "NN", "IN", "$", "CD", "CD", ",", "IN", "NN", "IN", "DT", "$", "CD", "CD", "NN", "RB", "."],
"entities": [{"type": "ORG", "start": 1, "end": 4}, {"type": "ORG", "start": 5, "end": 11}, {"type": "ORG", "start": 25, "end": 27}, {"type": "ORG", "start": 28, "end": 29}, {"type": "ORG", "start": 32, "end": 33}, {"type": "PER", "start": 33, "end": 34}],
"ltokens": ["Everlast", "'s", "Rally", "Just", "Might", "Live", "up", "to", "the", "Name", "."],
"rtokens": ["In", "other", "words", ",", "a", "competitor", "has", "decided", "that", "one", "segment", "of", "the", "company", "'s", "business", "is", "potentially", "worth", "$", "42.5", "million", "."],
"org_id": "MARKETVIEW_20041220.1537"
}The ltokens contains the tokens from the previous sentence. And The rtokens contains the tokens from the next sentence.
Due to the license of LDC, we cannot directly release our preprocessed datasets of ACE04, ACE05 and KBP17. We only release the preprocessed GENIA dataset and the corresponding word vectors and dictionary. Download them from here.
If you need other datasets, please contact me (syl@zju.edu.cn) by email. Note that you need to state your identity and prove that you have obtained the LDC license.
The word vectors used in our experiments:
- BioWord2Vec for GENIA: https://github.com/cambridgeltl/BioNLP-2016
- GloVe for other datasets: http://nlp.stanford.edu/data/glove.6B.zip
Download and extract the wordvecs from above links, save GloVe in ../glove and BioWord2Vec in ../biovec.
mkdir ../glove
mkdir ../biovec
mv glove.6B.100d.txt ../glove
mv PubMed-shuffle-win-30.txt ../biovecNote: the BioWord2Vec needs to be converted from binary format to text format. The code is as follows:
from gensim.models.keyedvectors import KeyedVectors
model = KeyedVectors.load_word2vec_format('PubMed-shuffle-win-30.bin', binary=True)
model.save_word2vec_format('./PubMed-shuffle-win-30.txt', binary=False)python identifier.py train --config configs/example.confNote: You should edit this line in config_reader.py according to the actual number of GPUs.
You can download our checkpoints, or train your own model and then evaluate the model.
cd data/
# download checkpoints from https://drive.google.com/drive/folders/1NaoL42N-g1t9jiif427HZ6B8MjbyGTaZ?usp=sharing
unzip checkpoints.zip
cd ../
python identifier.py eval --config configs/eval.confIf you use the checkpoints we provided, you will get the following results:
- ACE05:
-- Entities (named entity recognition (NER)) ---
An entity is considered correct if the entity type and span is predicted correctly
type precision recall f1-score support
WEA 84.62 88.00 86.27 50
ORG 83.23 78.78 80.94 523
PER 88.02 92.05 89.99 1724
FAC 80.65 73.53 76.92 136
GPE 85.13 87.65 86.37 405
VEH 86.36 75.25 80.42 101
LOC 66.04 66.04 66.04 53
micro 86.05 87.20 86.62 2992
macro 82.01 80.19 81.00 2992
- GENIA:
--- Entities (named entity recognition (NER)) ---
An entity is considered correct if the entity type and span is predicted correctly
type precision recall f1-score support
RNA 89.91 89.91 89.91 109
DNA 76.79 79.16 77.96 1262
cell_line 82.35 72.36 77.03 445
protein 81.11 85.18 83.09 3084
cell_type 72.90 75.91 74.37 606
micro 79.46 81.84 80.63 5506
macro 80.61 80.50 80.47 5506
- ACE04:
--- Entities (named entity recognition (NER)) ---
An entity is considered correct if the entity type and span is predicted correctly
type precision recall f1-score support
FAC 72.16 62.50 66.99 112
PER 91.62 91.26 91.44 1498
LOC 74.36 82.86 78.38 105
VEH 94.44 100.00 97.14 17
GPE 89.45 86.09 87.74 719
WEA 79.17 59.38 67.86 32
ORG 83.49 82.43 82.95 552
micro 88.24 86.79 87.51 3035
macro 83.53 80.64 81.78 3035
- KBP17:
--- Entities (named entity recognition (NER)) ---
An entity is considered correct if the entity type and span is predicted correctly
type precision recall f1-score support
LOC 66.75 64.41 65.56 399
FAC 72.62 64.06 68.08 679
PER 87.86 88.30 88.08 7083
ORG 80.06 72.29 75.98 2461
GPE 89.58 87.36 88.46 1978
micro 85.31 82.96 84.12 12600
macro 79.38 75.28 77.23 12600
The preprocessed GENIA dataset is available, so we use it as an example to demonstrate the training and evaluation of the model.
cd identifier
mkdir -p data/datasets
cd data/datasets
# download genia.zip (the preprocessed GENIA dataset, wordvec and vocabulary) from https://drive.google.com/file/d/13Lf_pQ1-QNI94EHlvtcFhUcQeQeUDq8l/view?usp=sharing.
unzip genia.zippython identifier.py train --config configs/example.confoutput:
--- Entities (named entity recognition (NER)) ---
An entity is considered correct if the entity type and span is predicted correctly
type precision recall f1-score support
protein 81.19 85.08 83.09 3084
RNA 90.74 89.91 90.32 109
cell_line 82.35 72.36 77.03 445
DNA 76.83 79.08 77.94 1262
cell_type 72.90 75.91 74.37 606
micro 79.53 81.77 80.63 5506
macro 80.80 80.47 80.55 5506
Best F1 score: 80.63560463237275, achieved at Epoch: 34
2021-01-02 15:07:39,565 [MainThread ] [INFO ] Logged in: data/genia/main/genia_train/2021-01-02_05:32:27.317850
2021-01-02 15:07:39,565 [MainThread ] [INFO ] Saved in: data/genia/main/genia_train/2021-01-02_05:32:27.317850
vim configs/eval.conf
# change model_path to the path of the trained model.
# eg: model_path = data/genia/main/genia_train/2021-01-02_05:32:27.317850/final_model
python identifier.py eval --config configs/eval.confoutput:
--------------------------------------------------
Config:
data/checkpoint/genia_train/2021-01-02_05:32:27.317850/final_model
Namespace(bert_before_lstm=True, cache_path=None, char_lstm_drop=0.2, char_lstm_layers=1, char_size=50, config='configs/eval.conf', cpu=False, dataset_path='data/datasets/genia/genia_test_context.json', debug=False, device_id='0', eval_batch_size=4, example_count=None, freeze_transformer=False, label='2021-01-02_eval', log_path='data/genia/main/', lowercase=False, lstm_drop=0.2, lstm_layers=1, model_path='data/checkpoint/genia_train/2021-01-02_05:32:27.317850/final_model', model_type='identifier', neg_entity_count=5, nms=0.65, no_filter='sigmoid', no_overlapping=False, no_regressor=False, no_times_count=False, norm='sigmoid', pool_type='max', pos_size=25, prop_drop=0.5, reduce_dim=True, sampling_processes=4, seed=47, size_embedding=25, spn_filter=5, store_examples=True, store_predictions=True, tokenizer_path='data/checkpoint/genia_train/2021-01-02_05:32:27.317850/final_model', types_path='data/datasets/genia/genia_types.json', use_char_lstm=True, use_entity_ctx=True, use_glove=True, use_pos=True, use_size_embedding=False, weight_decay=0.01, window_size=[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10], wordvec_path='../biovec/PubMed-shuffle-win-30.txt')
Repeat 1 times
--------------------------------------------------
Iteration 0
--------------------------------------------------
Avaliable devices: [3]
Using Random Seed 47
2021-05-22 17:52:44,101 [MainThread ] [INFO ] Dataset: data/datasets/genia/genia_test_context.json
2021-05-22 17:52:44,101 [MainThread ] [INFO ] Model: identifier
Reused vocab!
Parse dataset 'test': 100%|███████████████████████████████████████| 1854/1854 [00:09<00:00, 202.86it/s]
2021-05-22 17:52:53,507 [MainThread ] [INFO ] Relation type count: 1
2021-05-22 17:52:53,507 [MainThread ] [INFO ] Entity type count: 6
2021-05-22 17:52:53,507 [MainThread ] [INFO ] Entities:
2021-05-22 17:52:53,507 [MainThread ] [INFO ] No Entity=0
2021-05-22 17:52:53,508 [MainThread ] [INFO ] DNA=1
2021-05-22 17:52:53,508 [MainThread ] [INFO ] RNA=2
2021-05-22 17:52:53,508 [MainThread ] [INFO ] cell_type=3
2021-05-22 17:52:53,508 [MainThread ] [INFO ] protein=4
2021-05-22 17:52:53,508 [MainThread ] [INFO ] cell_line=5
2021-05-22 17:52:53,508 [MainThread ] [INFO ] Relations:
2021-05-22 17:52:53,508 [MainThread ] [INFO ] No Relation=0
2021-05-22 17:52:53,508 [MainThread ] [INFO ] Dataset: test
2021-05-22 17:52:53,508 [MainThread ] [INFO ] Document count: 1854
2021-05-22 17:52:53,508 [MainThread ] [INFO ] Relation count: 0
2021-05-22 17:52:53,508 [MainThread ] [INFO ] Entity count: 5506
2021-05-22 17:52:53,508 [MainThread ] [INFO ] Context size: 242
2021-05-22 17:53:10,348 [MainThread ] [INFO ] Evaluate: test
Evaluate epoch 0: 100%|██████████████████████████████████████████| 464/464 [01:14<00:00, 6.26it/s]
Enmuberated Spans: 0
Evaluation
--- Entities (named entity recognition (NER)) ---
An entity is considered correct if the entity type and span is predicted correctly
type precision recall f1-score support
cell_line 82.60 71.46 76.63 445
DNA 77.67 78.29 77.98 1262
RNA 92.45 89.91 91.16 109
cell_type 73.79 75.25 74.51 606
protein 81.99 84.57 83.26 3084
micro 80.33 81.15 80.74 5506
macro 81.70 79.89 80.71 5506
2021-05-22 17:54:28,943 [MainThread ] [INFO ] Logged in: data/genia/main/genia_eval/2021-05-22_17:52:43.991876
If you have any questions related to the code or the paper, feel free to email syl@zju.edu.cn.
@inproceedings{shen-etal-2021-locate,
title = "Locate and Label: A Two-stage Identifier for Nested Named Entity Recognition",
author = "Shen, Yongliang and
Ma, Xinyin and
Tan, Zeqi and
Zhang, Shuai and
Wang, Wen and
Lu, Weiming",
booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
month = aug,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.acl-long.216",
doi = "10.18653/v1/2021.acl-long.216",
pages = "2782--2794",
}