Unsupervised Part-Of-Speech Tagging Supporting Supervised Methods

This paper investigates the utility of an unsupervised partof-speech (PoS) system in a task oriented way. We use PoS labels as features for different supervised NLP tasks: Word Sense Disambiguation, Named Entity Recognition and Chunking. Further we explore, how much supervised tagging can gain from unsupervised tagging. A comparative evaluation between variants of systems using standard PoS, unsupervised PoS and no PoS at all reveals that supervised tagging gains substantially from unsupervised tagging. Further, unsupervised PoS tagging behaves similarly to supervised PoS in Word Sense Disambiguation and Named Entity Recognition, while only chunking benefits more from supervised PoS. Overall results indicate that unsupervised PoS tagging is useful for many applications and a veritable low-cost alternative, if none or very little PoS training data is available for the target language or domain.