Features, Bagging, and System Combination for the Chinese POS Tagging Task
Fei Xia, Lap Pong Cheung · 2006
In recent years more and more NLP packages become available to the pub-lic, and many of them are implementa-tions of general machine learning meth-ods. A natural question is how one can quickly build a good system using those packages. To address this issue, we built three part-of-speech taggers (i.e., trigram, TBL, and MaxEnt tag-gers) for Chinese using existing pack-ages. Our experiments showed that adapting and extending a package is relative easy if the package is well-written and source code is available. We studied the contribution of each type of feature templates to the tag-ging accuracy and showed that adding some templates could help one tagger but hurt another one. Furthermore, we demonstrated that bagging (Breiman, 1996) provides a moderate gain for the TBL tagger, and combining TBL and MaxEnt taggers work better than using all three taggers. 1