中文名詞組的辨識：監督式與半監督式學習法的實驗

Chinese NP Chunking: Experiments with Supervised,and Semisupervised Learning

本文先利用Taku Kudo所發展的SVM工具Yamcha訓練中文名詞組辨識的初始模型，並嘗試以不同於多數文獻的IOB表示法及前二後二位置的語法標記資訊，找到適用於中文的參數。接著利用半監督式學習法中自我學習的概念，利用網路上未標記的資料，強化supervised-learning的模型。實驗結果證明，supervised learning這個步驟裡，我們選用的參數比前人的更合適；而我們所提出的半監督式學習法，可以提昇辨識結果，特別是在動詞修飾名詞的情形，半監督式學習法可以大幅提高辨識的正確率。

This paper utilizes Yamcha, a SVM tool designed by Taku Kudo, to train an NP-chunking model for Chinese. In addition to IOB and two words surrounding the focused word, we experimented on new features and exploited unlabeled data from web pages to enhance the previous model. Our experiments with supervised learning indicate that our chosen feature sets outperform those reported in previous studies. In addition, the proposed method of semisupervised learning is proved to be effective in distinguishing a noun phrase from a verb phrase both consisting of V N combination, thus enhancing the overall accuracy.