This repository contains an open-source Kazakh named entity recognition dataset (KazNERD), named entity annotation guidelines (in Kazakh), and NER model training codes (CRF, BiLSTM-CNN-CRF, BERT and XLM-RoBERTa).

1. KazNERD Corpus
2. Annotation Guidelines
3. NER Models
   1. CRF
   2. BiLSTM-CNN-CRF
   3. BERT and XLM-RoBERTa
4. Citation

KazNERD contains 112,702 sentences, extracted from the television news text, and 136,333 annotations for 25 entity classes. All sentences in the dataset were manually annotated by two native Kazakh-speaking linguists, supervised by an ISSAI researcher. The IOB2 scheme was used for annotation. The dataset, in CoNLL 2002 format, is located here.

The annotation guidelines followed to build KazNERD are located here. The guidelines contain rules for annotating 25 named entity classes and their examples. The guidelines are in the Kazakh language.

The CRF-based NER model training codes are based on Python 3.8. To ease the experiment replication experience, we recommend setting up a Conda environment.

conda create --name knerdCRF python=3.8
conda activate knerdCRF
conda install -c anaconda nltk scikit-learn
conda install -c conda-forge sklearn-crfsuite seqeval

$ cd crf
$ python runCRF_KazNERD.py

The BiLSTM-CNN-CRF-based NER model training codes are based on Python 3.8 and PyTorch 1.7.1. To ease the experiment replication experience, we recommend setting up a Conda environment.

conda create --name knerdLSTM python=3.8
conda activate knerdLSTM
# Check https://pytorch.org/get-started/previous-versions/#v171
# to install a PyTorch version suitable for your OS and CUDA
# or feel free to adapt the code to a newer PyTorch version
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=10.1 -c pytorch   # we used this version
conda install -c conda-forge tqdm seqeval

$ cd BiLSTM_CNN_CRF
$ bash run_train_p.sh

The BERT- and XLM-RoBERTa-based NER models training codes are based on Python 3.8 and PyTorch 1.7.1. To ease the experiment replication experience, we recommend setting up a Conda environment.

conda create --name knerdBERT python=3.8
conda activate knerdBERT
# Check https://pytorch.org/get-started/previous-versions/#v171
# to install a PyTorch version suitable for your OS and CUDA
# or feel free to adapt the code to a newer PyTorch version
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=10.1 -c pytorch   # we used this version
conda install -c anaconda numpy
conda install -c conda-forge seqeval
pip install transformers
pip install datasets

$ cd bert
$ python run_finetune_kaznerd.py bert

$ cd bert
$ python run_finetune_kaznerd.py roberta

Once the training is complete, you can use the following script to tag an input sentence:

$ python run_predict_kaznerd.py model model_checkpoint 'kazakh sentence'

where model is either 'bert' or 'roberta', model_checkpoint is a path to the pre-trained model, 'kazakh sentence' is an input sentence. For example:

$ python run_predict_kaznerd.py bert bert-base-multilingual-cased-finetuned-ner-6/checkpoint-705/ 'Кеше Әйгерім Әбдібекова Абайдың «Қара сөздерінің» аудиодискісін 1000 теңгеге алды.'

@misc{yeshpanov2021kaznerd,
      title={KazNERD: Kazakh Named Entity Recognition Dataset}, 
      author={Rustem Yeshpanov and Yerbolat Khassanov and Huseyin Atakan Varol},
      year={2021},
      eprint={2111.13419},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}