Extending Word Representations with Affix Features for Bidirectional LSTM-CRF-based Korean Named Entity recognition

Abstract

We propose a Korean named entity recognition (NER) model which extracts affix features to augment word representations. We build upon two recently prominently used NER models, namely BiLSTM-BiLSTM-CRF and CNN-BiLSTM-CRF, by extending the word embeddings with approximated affix information. We choose to use an inexpensive character-level frequency filter to infer the affix information. Our experimental results on the HCLT 2016 and ETRI NER datasets show up to a 0.93% increase in F1 score compared to the original models without any dictionary or morphological tools. This increase is a significant improvement in NER considering the recent stagnation following the introduction of neural NER. These results show that Korean predicted affix features are useful in building neural NER models.