NCTU and NTUT's Entry to CLP-2014 Chinese Spelling Check Evaluation

Abstract

This paper describes our Chinese spelling check system submitted to SIGHAN Bake-off 2014 evaluation. The system's main components are still the conditional random field (CRF)-based word segmentation/part-ofspeech (POS) tagger and tri-gram language model (LM) used last year. But we tried to refine the misspelling rules, decision-making threshold and improve LM rescoring speed to reduce false alarm rate and improve rescoring speed. Bake-off 2014 evaluation results show that one of our system (Run2) did achieve reasonable performance with about 0.485/0.468 accuracies and 0.226/0.180 F1 scores in the detection/ correction metrics.