This paper presents a new approach to part-of-speech (POS) tagging in which the basic unit being tagged is a contiguous sequence of words rather than a single word. We run experiments on two different tagsets: the UPENN treebank and a treebank annotated with more ambiguous tags that have a semantic component. We show that the phrase-based system alone is a respectable tagger that exceeds the performance of the ME tagger on the ambiguous tagset. Moreover, when a log-linear model is built using features from both phrase-and word-based techniques, the tagging accuracy improved on both of our data sets yielding the highest reported performance to date on the more ambiguous tagset.