Improved Part-of-speech Tagging by Log-linear Models
Ruiqiang Zhang, Atr Spoken · 2008
This paper presented our latest approaches for improving English part-of-speech tagging with a large tagset by using a log-linear model.We found that integration of multiple probability models log-linearly led to significant im- provements on part-of-speech tagging.We compared our proposed approach with the two traditional approaches,hid- den markov model(HMM)and maximum entropy principle.The HMM approach was implemented by using a new smoothing method while for the maximum entropy approach,we integrated detailed local and long range word and tag information as the predictive features.Those predictive features were proved very effective for part-of-speech tagging. The experimental results showed that the maximum entropy model integrating multiple source information achieved higher POS tagging accuracy than the HMM models,however,the maximal improvements were achieved by the pro- posed log-linear models.