언어모델과 재순위화를 이용한 형태소 기반 한국어 품사 태거

Abstract

In Korean Part-of-Speech (POS) tagging task, morpheme-based tagging is linguistically standard approach compared to character-based tagging, because character itself means nothing. Moreover, despite of its high speed and easy customizing, morpheme-based approach is not general in recent Korean POS tagging, because of its complex architecture and poor precision. Because segmentation of morpheme is also determined in decoding step, it is hard to use broad context features such as higher n-gram feature. This thesis applied re-ranking framework into POS tagging system to utilize global feature. The baseline POS tagger generates K-best segmented and tagged candidates and a higher n-gram language model re-scores them. By using re-ranking with language model, it improved word precision by 1.2\% points compared to baseline tagger.