Wojood is a corpus for Arabic nested Named Entity Recognition (NER). Nested entities occur when one entity mention is embedded inside another entity mention. 550K tokens (MSA and dialect) This repo contains the source-code to train Wojood nested NER.
You can try our model using the demo link below
https://ontology.birzeit.edu/Wojood/
A corpus and model for nested Arabic Named Entity Recognition Version: 1.0 (updated on 20/1/2022)
Wojood consists of about 550K tokens (MSA and dialect) that are manually annotated with 21 entity types (e.g., person, organization, location, event, date, etc). It covers multiple domains and was annotated with nested entities. The corpus contains about 75K entities and 22.5% of which are nested. A nested named entity recognition (NER) model based on BERT was trained (F1-score 88.4%).
Corpus size: 550K tokens (MSA and dialects)
Richness: 21 entity classes, contains ~75K entities and 22.5% of them are nested entities
Domains: Media, History, Culture, Health, Finance, ICT, Law, Elections, Politics, Migration, Terrorism, social media
Inter-annotator agreement: 97.9% (Cohen's Kappa)
NER Model: AraBERTV2 (88.4% F1-score)
Entity Classes (21): | ||
---|---|---|
PERS (person) | EVENT | CARDINAL |
NORP (group of people) | DATE | ORDINAL |
OCC (occupation) | TIME | PERCENT |
ORG (organization) | LANGUAGE | QUANTITY |
GPE (geopolitical entity) | WEBSITE | UNIT |
LOC (geographical location) | LAW | MONEY |
FAC (facility: landmarks places) | PRODUCT | CURR (currency) |
Please email Prof. Jarrar (mjarrar AT birzeit.edu) for the annotation guidelines
A sample data is available in the data
directory. But the entire Wojood NER corpus is
available to download upon request for academic and commercial use. Request to download
Wojood (corpus and the model).
https://ontology.birzeit.edu/Wojood/
huggingface: https://huggingface.co/SinaLab/ArabicNER-Wojood
At this point, the code is compatible with Python 3.9
and torchtext==0.9.0
.
Clone this repo
git clone https://github.com/SinaLab/ArabicNER.git
This package has dependencies on multiple Python packages. It is recommended to Conda to create a new environment
that mimics the same environment the model was trained in. Provided in this repo environment.yml
from which you
can create a new conda environment using the command below.
conda env create -f environment.yml
Update your PYTHONPATH to point to ArabiNER package
export PYTHONPATH=PYTHONPATH:/path/to/ArabiNER
Argument for model traning are listed below. Note that some arguments including data_config
,
trainer_config
, network_config
, optimizer
, lr_scheduler
and loss
take as input JSON
configuration (see examples below).
usage: train.py [-h] --output_path OUTPUT_PATH --train_path TRAIN_PATH
--val_path VAL_PATH --test_path TEST_PATH
[--bert_model BERT_MODEL] [--gpus GPUS [GPUS ...]]
[--log_interval LOG_INTERVAL] [--batch_size BATCH_SIZE]
[--num_workers NUM_WORKERS] [--data_config DATA_CONFIG]
[--trainer_config TRAINER_CONFIG]
[--network_config NETWORK_CONFIG] [--optimizer OPTIMIZER]
[--lr_scheduler LR_SCHEDULER] [--loss LOSS] [--overwrite]
[--seed SEED]
optional arguments:
-h, --help show this help message and exit
--output_path OUTPUT_PATH
Output path (default: None)
--train_path TRAIN_PATH
Path to training data (default: None)
--val_path VAL_PATH
Path to training data (default: None)
--test_path TEST_PATH
Path to training data (default: None)
--bert_model BERT_MODEL
BERT model (default: aubmindlab/bert-base-arabertv2)
--gpus GPUS [GPUS ...]
GPU IDs to train on (default: [0])
--log_interval LOG_INTERVAL
Log results every that many timesteps (default: 10)
--batch_size BATCH_SIZE
Batch size (default: 32)
--num_workers NUM_WORKERS
Dataloader number of workers (default: 0)
--data_config DATA_CONFIG
Dataset configurations (default: {"fn":
"arabiner.data.datasets.DefaultDataset", "kwargs":
{"max_seq_len": 512}})
--trainer_config TRAINER_CONFIG
Trainer configurations (default: {"fn":
"arabiner.trainers.BertTrainer", "kwargs":
{"max_epochs": 50}})
--network_config NETWORK_CONFIG
Network configurations (default: {"fn":
"arabiner.nn.BertSeqTagger", "kwargs": {"dropout":
0.1, "bert_model": "aubmindlab/bert-base-arabertv2"}})
--optimizer OPTIMIZER
Optimizer configurations (default: {"fn":
"torch.optim.AdamW", "kwargs": {"lr": 0.0001}})
--lr_scheduler LR_SCHEDULER
Learning rate scheduler configurations (default:
{"fn": "torch.optim.lr_scheduler.ExponentialLR",
"kwargs": {"gamma": 1}})
--loss LOSS Loss function configurations (default: {"fn":
"torch.nn.CrossEntropyLoss", "kwargs": {}})
--overwrite Overwrite output directory (default: False)
--seed SEED Seed for random initialization (default: 1)
In the case of nested NER we pass NestedTagsDataset
to --data_config
, BertNestedTrainer
to --trainer_config
,
and BertNestedTagger
to --network_config
.
python train.py \
--output_path /path/to/output/dir \
--train_path /path/to/train.txt \
--val_path /path/to/val.txt \
--test_path /path/to/test.txt \
--batch_size 8 \
--data_config '{"fn":"arabiner.data.datasets.NestedTagsDataset","kwargs":{"max_seq_len":512}}' \
--trainer_config '{"fn":"arabiner.trainers.BertNestedTrainer","kwargs":{"max_epochs":50}}' \
--network_config '{"fn":"arabiner.nn.BertNestedTagger","kwargs":{"dropout":0.1,"bert_model":"aubmindlab/bert-base-arabertv2"}}' \
--optimizer '{"fn":"torch.optim.AdamW","kwargs":{"lr":0.0001}}'
In the case of flat NER we pass DefaultDataset
to --data_config
, BertTrainer
to --trainer_config
,
and BertSeqTagger
to --network_config
.
python train.py \
--output_path /path/to/output/dir \
--train_path /path/to/train.txt \
--val_path /path/to/val.txt \
--test_path /path/to/test.txt \
--batch_size 8 \
--data_config '{"fn":"arabiner.data.datasets.DefaultDataset","kwargs":{"max_seq_len":512}}' \
--trainer_config '{"fn":"arabiner.trainers.BertTrainer","kwargs":{"max_epochs":50}}' \
--network_config '{"fn":"arabiner.nn.BertSeqTagger","kwargs":{"dropout":0.1,"bert_model":"aubmindlab/bert-base-arabertv2"}}' \
--optimizer '{"fn":"torch.optim.AdamW","kwargs":{"lr":0.0001}}'
Inference is the process of using a pre-trained model to perform tagging on a new text. To do that, we will need the following:
Note that the model has the following structure and it is important to keep the same structure for inference to work.
.
├── args.json
├── checkpoints
├── predictions.txt
├── tag_vocab.pkl
├── tensorboard
└── train.log
provided in the bin
directory infer.py
script that performs inference.
The infer.py
has the following parameters:
usage: infer.py [-h] --model_path MODEL_PATH --text
TEXT [--batch_size BATCH_SIZE]
optional arguments:
-h, --help show this help message and exit
--model_path MODEL_PATH
Model path for a pre-trained model, for this we you need to download the checkpoint from this repo (default: None)
--text TEXT Text or sequence to tag, segments will be identified based on periods (default: None)
--batch_size BATCH_SIZE
Batch size (default: 32)
Example inference command:
python -u /path/to/ArabiNER/arabiner/bin/infer.py
--model_path /path/to/model
--text "وثائق نفوس شخصية من الفترة العثمانية للسيد نعمان عقل"
Optionally, there is eval.py
script in bin
directory to evaluate NER dataset with ground truth data.
usage: eval.py [-h] --output_path OUTPUT_PATH --model_path MODEL_PATH
--data_paths DATA_PATHS [DATA_PATHS ...] [--batch_size BATCH_SIZE]
optional arguments:
-h, --help show this help message and exit
--output_path OUTPUT_PATH
Path to save results (default: None)
--model_path MODEL_PATH
Model path (default: None)
--data_paths DATA_PATHS [DATA_PATHS ...]
Text or sequence to tag, this is in same format as
training data with 'O' tag for all tokens (default: None)
--batch_size BATCH_SIZE
Batch size (default: 32)
This research is partially funded by the Palestinian Higher Council for Innovation and Excellence.
Mustafa Jarrar, Mohammed Khalilia, Sana Ghanem: Wojood: Nested Arabic Named Entity Corpus and Recognition using BERT. In Proceedings of the International Conference on Language Resources and Evaluation (LREC 2022), Marseille, France. 2022