Using Multiple Taggers to Improve English Part-of-Speech Tagging by Error Learning

Many approaches to Part-of-Speech tagging have reached accuracy about 96-97% which is close to the upper bound, and little improvement was made in recent years. In this project, we propose an idea of increasing tagging accuracy by learning the di erence of results of multiple taggers. Since the errors of di erent taggers might be complementary, we can avoid some of them by using di erent taggers under di erent contexts. Context patterns that tell which tagging result should be used for certain circumstances are learned from the tagging results of training data, and are applied to the testing one. We combine trigram, unigram and rule-based taggers together by this means, and reduce 11% of tagging errors of trigram tagger. Many experiments are done in this project, and the possible ways to improving tagging result are tested and compared.