PolDeepNer recognizes mentions of named entities in text using deep learning methods. The tool won second place in the PolEval 2018 Task 2 on named entity recognition. It contains a pre-trained model trained on the NKJP corpus which recognizes nested annotations of the following types:
Marcińczuk, Michał; Kocoń, Jan; Gawor, Michał. Recognition of Named Entities for Polish-Comparison of Deep Learning and Conditional Random Fields Approaches Ogrodniczuk, Maciej; Kobyliński, Łukasz (Eds.): Proceedings of the PolEval 2018 Workshop, pp. 63-73, Institute of Computer Science, Polish Academy of Science, Warszawa, 2018.
[PDF]
[Bibtex]
@inproceedings{poldeepner2018,
title = "Recognition of Named Entities for Polish-Comparison of Deep Learning and Conditional Random Fields Approaches",
author = "Marcińczuk, Michał and Kocoń, Jan and Gawor, Michał",
year = "2018",
editor = "Ogrodniczuk, Maciej and Kobyliński, Łukasz",
booktitle = "Proceedings of the PolEval 2018 Workshop",
location = "Warsaw, Poland",
pages = "77--92",
publisher = "Institute of Computer Science, Polish Academy of Science"
}
BiLSTM and CRF implementation was based on AnaGo (https://github.com/Hironsan/anago)
- Michał Marcińczuk michal.marcinczuk@pwr.edu.pl
- Jan Kocoń jan.kocon@pwr.edu.pl
- Michał Gawor michal.gawor@pwr.edu.pl
- Python 3.6
- CUDA 10.0+
Download and/or unpack models (word embeddings, NN pre-trained models):
sudo apt-get install p7zip-full./download_unpack.shBuild Docker image:
./build.shRun Docker image in the interactive mode:
./run.shsudo apt-get install python3-pip python3-dev python-virtualenv
sudo pip install -U pip
virtualenv --system-site-packages -p python3.6 venv
source venv/bin/activate
pip install -U pip
pip install -r requirements.txtThe evalation corpus in json format was split into a set of separate files. Then each of the documents was tagged using WCRFT tagger and four files were generated:
- NAME.txt — file content,
- NAME.ini — ini file with document original name,
- NAME.xml — ccl file containing tagging output,
- NAME.iob — ccl file converted to iob format.
Run Docker
./run.shGenerate json file for the evaluation corpus:
python poldeepner/core/process_poleval.py \
-i poldeepner/data/poleval2018ner-data/index_iob.list \
-o poldeepner/data/poldeepner-output.json \
-m nkjp-poleval18Run the official evaluation script:
python poldeepner/core/poleval_ner_test.py \
-g poldeepner/data/POLEVAL-NER_GOLD.json \
-u poldeepner/data/poldeepner-output.jsonOutput:
OVERLAP precision: 0.918 recall: 0.815 F1: 0.864
EXACT precision: 0.860 recall: 0.764 F1: 0.809
Final score: 0.853Disclaimer: the current version of the script uses NLTK tokenizer which is not fully compatible with the training data used to train the model. The final version of the script should use toki tokenizer or WCRFT tagger.
Run:
python poldeepner/core/interactive_test.pyEnter text to process:
Enter text to process:Enter text to process: XXXIII konwent fanów fantastyki Polcon trwał przez 4 dni: od 12 do 15 sierpnia 2018 na terenie Miasteczka Akademickiego Uniwersytetu Mikołaja Kopernika w Toruniu.Output:
[ 32: 38] orgName Polcon
[ 61: 63] date 12
[ 67: 83] date 15 sierpnia 2018
[ 95:151] geogName Miasteczka Akademickiego Uniwersytetu Mikołaja Kopernika
[154:161] placeName_settlement Toruniu
To exit press Enter without typing any text.
Sample file: KPWr: Toronto Dominion Center
Run:
python poldeepner/core/process_file.py -m nkjp-poleval18 -i poldeepner/data/kpwr-toronto.txtExpected output:
[ 0: 16] orgName Toronto Dominion
[ 0: 7] placeName_settlement Toronto
[ 85: 96] placeName_country kanadyjskim
[105:112] placeName_settlement Toronto
[116:134] orgName Financial District
[204:229] persName Ludwiga Mies van der Rohe
[204:211] persName_forename Ludwiga
[212:216] persName_surname Mies
[225:229] persName_surname Rohe
[236:243] orgName Budynki
[293:303] geogName Joe Fafard
[503:550] geogName Inuitów – Toronto Dominion Gallery of Inuit Art
[513:520] placeName_settlement TorontoRun:
python poldeepner/core/process_file.py -m n82-ft-kgr10 -i poldeepner/data/kpwr-toronto.txtExpected output:
[ 0: 23] nam_fac_goe Toronto Dominion Centre
[ 28: 51] nam_fac_goe Toronto Dominion Centre
[ 85: 96] nam_adj_country kanadyjskim
[105:112] nam_loc_gpe_city Toronto
[116:134] nam_loc_gpe_district Financial District
[204:229] nam_liv_person Ludwiga Mies van der Rohe
[293:303] nam_liv_person Joe Fafard
[503:510] nam_org_nation Inuitów
[513:550] nam_fac_goe Toronto Dominion Gallery of Inuit ArtRun virtual environment.
source venv/bin/activateTrain each model separately. The training will replace the pre-trained models.
python poldeepner/core/trainmodel.py \
-i poldeepner/data/nkjp-nested-simplified-v2.iob \
-t poldeepner/data/nkjp-nested-simplified-v2.iob \
-f poldeepner/model/cc.pl.300.bin \
-m poldeepner/model/poldeepner-nkjp-ftcc-bigru \
-e 15 -n GRU
python poldeepner/core/trainmodel.py \
-i poldeepner/data/nkjp-nested-simplified-v2.iob \
-t poldeepner/data/nkjp-nested-simplified-v2.iob \
-f poldeepner/model/kgr10_orths.vec.bin \
-m poldeepner/model/poldeepner-nkjp-ftkgr10orth-bigru \
-e 15 -n GRU -s 100
python poldeepner/core/trainmodel.py \
-i poldeepner/data/nkjp-nested-simplified-v2.iob \
-t poldeepner/data/nkjp-nested-simplified-v2.iob \
-f poldeepner/model/kgr10-plain-sg-300-mC50.bin \
-m poldeepner/model/poldeepner-nkjp-ftkgr10plain-lstm \
-e 15 -n LSTMBefore training the structure of the NN will be printed:
Layer (type) Output Shape Param #
=================================================================
word_input (InputLayer) (None, None, 100) 0
_________________________________________________________________
dropout_1 (Dropout) (None, None, 100) 0
_________________________________________________________________
bidirectional_1 (Bidirection (None, None, 200) 120600
_________________________________________________________________
dense_1 (Dense) (None, None, 100) 20100
_________________________________________________________________
crf_1 (CRF) (None, None, 41) 5904
=================================================================
Total params: 146,604
Trainable params: 146,604
Non-trainable params: 0
_________________________________________________________________After every epoch a table with accuracy of the current model will be printed. Below is a sample table after the first epoch.
2677/2677 [==============================] - 234s 87ms/step - loss: 0.0762
- f1: 74.68
annotation TP FP FN precision recall f1-score support
date 3365 647 1133 83.87 74.81 79.08 4498
geogName 1102 417 3040 72.55 26.61 38.93 4142
orgName 6714 2867 4444 70.08 60.17 64.75 11158
persName 16165 2495 3773 86.63 81.08 83.76 19938
persName_addName 0 0 949 0.00 0.00 0.00 949
persName_forename 11271 1138 1683 90.83 87.01 88.88 12954
persName_surname 11351 2035 1266 84.80 89.97 87.31 12617
placeName 0 0 368 0.00 0.00 0.00 368
placeName_bloc 0 0 100 0.00 0.00 0.00 100
placeName_country 6172 1437 1324 81.11 82.34 81.72 7496
placeName_district 0 0 279 0.00 0.00 0.00 279
placeName_region 0 0 760 0.00 0.00 0.00 760
placeName_settlement 5796 2055 1984 73.82 74.50 74.16 7780
time 147 45 417 76.56 26.06 38.89 564
TOTAL 62083 13136 21520 82.54 74.26 78.18 83603Copyright (C) Wrocław University of Science and Technology (PWr), 2010-2018. All rights reserved.
This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/.