Part-of-Speech Tagging and PP Attachment Disambiguation Using a Boosted Maximum Entropy Model

Abstract

We have proposed previously a boosted maximum entropy model to overcome three major problems in applying the maximum entropy models to text chunking [1]: (i) feature selection, (ii) high computational complexity, and (iii) highly-imbalanced data. To cope with the first problem, the boosted ME models adopt a decision tree as a constructor of the high-order features. Because decision trees can be represented as a set of if-then rules, the features for ME models are automatically constructed by transforming a decision tree into if-then rules. Active learning is adopted to solve the high computational complexity, and the AdaBoost is used to overcome the highly imbalance in natural language resources. In this paper, we apply the boosted maximum entropy model to two major tasks in natural language processing: POS tagging and PP attachment.