A Conditional Random Field-based Traditional Chinese Base-Phrase Parser for SIGHAN Bake-off 2012 Evaluation

This paper describes our system for the sub-task 1 of traditional Chinese Parsing of SIGHAN Bake-off 2012 evaluation. Since this research mainly focuses on speech recognition and synthesis applications, only base phrase chunking was implemented using three Conditional Random Field (CRF) modules, including word segmentation, POS tagging and base phrase chunking sub-systems. The official evaluation results show that the system achieved 0.5038 (0.7210/0.387) micro- and 0.5301 (0.7343/0.4147) macro-averaging F1 (precision/recall) rates on full sentence parsing task. However, if only the performance of base phrase chunking was considered, the F-measures may be around 0.70 and is somehow good enough for speech recognition and synthesis applications.