Chinese Chunking Based on Coarse-Grained Part-of-Speech Features

Abstract

Although part-of-speech (POS) is an effective feature for Chinese Chunking, the POS-tagging errors generated by automatic POS tagger leads to almost 10% performance drop in F-score. To solve this problem, this paper presents new features to replace the POS features, namely the coarse-grained part-of-speech features. Combining with the methods of processing out-of-vocabulary words, the new features are utilized in the Chinese chunking model. Experimental results show that the new features can contribute 2.71% performance improvement over the baseline method.