We conduct a series of Part-of-Speech (POS) Tagging experiments using Ex- pectation Maximization (EM), Varia- tional Bayes (VB) and Gibbs Sampling (GS) against the Chinese Penn Tree- bank. We want to first establish a base- line for unsupervised POS tagging in Chinese, which will facilitate future re- search in this area. Secondly, by com- paring and analyzing the results between Chinese and English, we highlight some of the strengths and weaknesses of each of the algorithms in POS tagging task and attempt to explain the differences based on some preliminary linguistics analysis. Comparing to English, we find that all algorithms perform rather poorly in Chinese in 1-to-1 accuracy result but are more competitive in many-to-1 accu- racy. We attribute one possible explana- tion of this to the algorithms ’ inability to correctly produce tags that match the desired tag count distribution.
展开▼
