This repository consists of the code for proccessing the data and training NER for Russian language at BSNLP SlavNER 2021 using Spacy

• pip install -r requirements.txt

• Look at this article if you're experiencing troubles with gpu to fine-tune Bert
• Add pre-trained vectors
• Add pretraining corpus

1. Clone this repository
2. Download the data from BSNLP Shared Task page and put it into data/bsnlp2021_train_r1
3. Run python save_data.py $folder1 $folder2 $folder3 split_train split_dev folder4 train_file dev_file test_file where folder1, folder2, folder3 are the names of the folders for training and development sets, split_train and split_dev are percentage of data used in training and dev sets, folder4 is the name of the folder for test_set, train_file, dev_file, test_file are the file names of the resulting spacy binary data sets. Now you have your training data for Spacy v.3
4. (optional) Run save_pretraining.py $folder_names where folder_names are the different folder names you want to use for the pretraining corpus. Check it out here how pretraining might help you to obtain better results.
5. Run python -m spacy train config_ner_ruVec_pretrain.cfg --output ./tok2vec_output to train the model. Specify paths to the training and development sets inside the config file before training. Also you may use a pretraining corpus and choose different pretrained vectors to potentially obtain better results. By default there are vectors from ru_core_news_lg Russian model. You may find more info on training command and config editting here
6. Run python -m spacy train config_spacy_trans.cfg --output ./multilingual_output to train the model. Specify paths to the training and development sets inside the config file before training. You can change hyperparameters there and choose different models from https://huggingface.co/models. By default the model is "bert-base-multilingual-uncased", which was used for the training.