Part-of-Speech (POS) Tagging Revisited

Accurate part-of-speech (POS) tagging of natural language text data can add power to automated information retrieval and extraction. Brill's transformation-based learning (TBL) approach to automated POS tagging was introduced in 1992, combining virtues of rule-based and stochastic methods. Brill's innovative idea was to use machine learning techniques to search through all of rule space for the most effective rules for tagging ambiguous tokens based on their contexts, but this is computationally infeasible. Typical heuristic patches used in Brill tagging to reduce rule space to tractable sizes seem arbitrary, unempirical, and weak. Now faster hardware and powerful public domain machine learning software offer an opportunity to revisit POS tagging with improved tools. This paper examines TBL and "shotgun" POS tagging using the Weka J48 classifier on text corpora from different genre domains.