COV Model and its Application in Chinese Part-of-Speech Tagging

This article presents a new sequence labeling model named Context OVerlapping (COV) model, which expands observation from single word to n-gram unit and there is an overlapping part between the neighboring units. Due to the co-occurrence constraint and transition constraint, COV model reduces the search space and improves tagging accuracy. The 2-gram COV is applied to Chinese PoS tagging and the precision rate of the open test is as high as 96.83%, which is higher than the second order HMM, which is 95.73%. The result is also comparable to the discriminative models but COV takes much less training time than them. With symbol decoding COV prunes many nodes before statistics decoding and the search space of COV is about10-20% less than that of HMM.