An Analytical Study of Transformational Tagging for Chinese Text

This work is our initial attempt in using the transformation-based error-driven learning (TEL) procedure for tagging Chinese text. TEL has previously been shown to be effective in POS tagging for English [Brill.1995]. TEL provides several attractions: (i) automation for tagging, (ii) induction of interpretable rules, (iii) learning aimed at error-reduction. Our experimental corpus consist of over 70,000 words of Chinese text, divided into disjoint training and test sets of a 9: 1 ratio. With an unknown word/tag proportion of 13%, we achieved overall tagging accuracies of 94.56% (training) and 86.87% (testing).