Korean NER with Pytorch

Korean NER Task with CharCNN + BiLSTM + CRF (with Naver NLP Challenge dataset), implemented with Pytorch

Model

• Character Embedding with CNN
• Concatenate word embedding with character represention
• Put the feature above to BiLSTM + CRF

Dependencies

• python>=3.5
• torch==1.4.0
• seqeval==0.0.12
• pytorch-crf==0.7.2
• gdown==3.10.1

Data

	Train	Test
# of Data	81,000	9,000

• Naver NLP Challenge 2018 NER Dataset (Github link)
• Original github only has train dataset, so test dataset is created by splitting the train dataset. (Data link)

Pretrained Word Vectors

• Use Korean fastText vectors with 300 dimension
• It takes quiet long time to load from original vector, so I take out the word vectors that are only in word vocab.
• It will be downloaded automatically when you run main.py.

Usage

$ python3 main.py --do_train --do_eval

• Evaluation prediction result will be saved in preds dir when you give --write_pred option.

Results

	Slot F1 (%)
CNN+BiLSTM+CRF	73.65
CNN+BiLSTM+CRF (+fastText)	74.57

Reference

• Naver NLP Challenge
• End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF
• NLP Implementation by aisolab