Chinese Spelling Check via Bidirectional LSTM-CRF

Chinese spelling check is an automatic mechanism to detect Chinese spelling errors in the text. State-of-the-art Chinese Spelling Check systems using bidirectional LSTM, which does not consider the constraint between output labels. In this paper, we introduce a novel neural network architecture that combines the bidirectional LSTM and CRF. This model takes char sequence of a sentence as input, and BLSTM learns the order information, feeds the probability vector to the CRF layer, then CRF output predict best tag sequences. Our system is effective and efficient, requiring no feature engineering or data preprocessing, thus making it applicable to many tasks. We evaluate our system on two datasets - the Sogou News dataset and the Literary Novels dataset. Compared with BLSTM models, we achieve a 5% and 3% relative f1-score improvement in Sogou News and Literary Novels, respectively.