Character-based Bidirectional LSTM-CRF with words and characters for Japanese Named Entity Recognition
Shotaro Misawa, Motoki Taniguchi, Yasuhide Miura, Tomoko Ohkuma · 2017
Recently, neural models have shown superior performance over conventional models in NER tasks.These models use CNN to extract sub-word information along with RNN to predict a tag for each word.However, these models have been tested almost entirely on English texts.It remains unclear whether they perform similarly in other languages.We worked on Japanese NER using neural models and discovered two obstacles of the state-ofthe-art model.First, CNN is unsuitable for extracting Japanese sub-word information.Secondly, a model predicting a tag for each word cannot extract an entity when a part of a word composes an entity.The contributions of this work are (i) verifying the effectiveness of the state-of-theart NER model for Japanese, (ii) proposing a neural model for predicting a tag for each character using word and character information.Experimentally obtained results demonstrate that our model outperforms the state-of-the-art neural English NER model in Japanese.2016).Rei et al. (2016) proposed the model using an attention mechanism whose inputs are words and characters.Above all, BLSTM-CNNs-CRF (Ma and Hovy, 2016) achieved state-of-theart performance on the standard English corpus: CoNLL2003 (Tjong Kim Sang and De Meulder, 2003).Character-based Neural Models: Kuru et al. (2016) proposed a character-based neural model.This model, which inputs only characters, exhibits good performance on the condition that no external knowledge is used.This model predicts a tag for each character and forces that predicted tags in a word are the same.Therefore, it is unsuitable for languages in which boundary conflicts occur.Japanese NER: For Japanese NER, many models using conventional algorithms have been proposed (Iwakura, 2011;Sasano and Kurohashi, 2008).Most such models are character-based models to deal with boundary conflicts.Tomori et al. ( 2016) applied a neural model to Japanese NER.This study uses non-sequential neural networks with inputs that are hand-crafted features.This model uses no recent advanced approaches for NER, such as word embedding or CNN to extract sub-word information.Therefore, the effectiveness of recent neural models for Japanese NER has not been evaluated.