Pointwise Prediction and Sequence-Based Reranking for Adaptable Part-of-Speech Tagging

This paper proposes an accurate method for partof-speech (POS) tagging that is highly domain-adaptable. The method is based on an assumption that the POS transition tendencies do not depend on domains, and has the following three characteristics: 1) it is trainable from partially annotated data, 2) it uses efficiently trainable pointwise POS taggers to allow for active learning, and 3) is more accurate than the pointwise or sequence-based POS taggers. The proposed method estimates POS tags by stacking pointwise and sequence-based predictors. In the experiments we deal with the joint problem of word segmentation and POS tagging in Japanese. We show that our proposed stacking process improves over pointwise and sequence-based methods (hidden Markov models and conditional random fields) both in the general domain and the target domain. In addition we show the learning curve in a domain adaptation scenario. The result shows that our method outperforms state-of-the-art methods in the same domain as the training data and is better than them in domain adaptation situations as well. Keywords-Active learning; Reranking; Word segmentation; Part-of-speech tagging; Pointwise prediction;