POS Tagging using Dependency Information

概要

論文の詳細を見る
This paper presents a POS tagging approach that makes use of dependency information of a word as feature to condition a model. A part-of-speech tagger for Tagalog makes use of morphological information such as affixes and reduplication as features. However, state-of-the art sequential labeling technique cannot achieve high accuracy for Tagalog. In this work, we investigate the use of dependency head information of the words to help predict the POS tag of the word. Most existing dependency parsing assumes POS tagging as a preprocess. In this paper, we did the reverse. We apply dependency parsing without POS information, and the POS tagger tested using the output of the dependency parser. Experiments show that this approach improves the baseline scores of the POS taggers of about 1.5% for POS unigram model and 2% for POS bigram model.
2011-07-08