This repo contains PyTorch implementation of an LSTM-CRF (Lample et al.) and a CNN-LSTM-CRF (Ma et al.) model with combined character-level and word-level representations for named entity recognition (NER).

• Python 3
• PyTorch (>= v1.0.0)
• tqdm

To train an English NER tagger on the CoNLL03 dataset, please follow the procedures below:

The repo already contains preprocessed CoNLL03 data in both IOB and IOBES tagging scheme (see dataset/conll03 folder). However, you need to convert the corresponding data into json-line format to run the training code. To use the IOBES scheme, run:

python -m prepro.process_data dataset/conll03/iobes dataset/conll_iobes --scheme iobes

Download the 100d GloVe vector file from here. Copy the file to a folder dataset/glove/, and run the following script:

python prepare_vocab.py dataset/conll_iobes dataset/vocab --all --lower

This will create vocab and embedding files in the directory dataset/vocab.

To train a CNN-LSTM-CRF model, run the script:

bash train_cnn_lstm_crf.sh 0

where 0 is the training ID. The model files will then be saved into a folder saved_models/00/.

Run evaluation using a trained model on the test split of the dataset with:

python eval.py saved_models/00 --dataset testb

where testb is the name of the test split.

Performance of the models is evaluated with entity-level F1 scores on the CoNLL03 English test set (testb). For each model and each tagging scheme (IOB vs IOBES), median and max scores across 5 random seeds are reported in the format of median (max). Note that the hyperparameters used to produce these scores (see training scripts) are different from those reported in the original papers.