中文名词组的辨识ChineseNPChunking.ppt

Semi-supervised method 導遊 發 給 每人 一 本 導覽 手冊。 這個 報導 給予 我 們 無限 的 想像 空間。 大家 看 了 宣導 短片 之後 有 什麼 感想 呢？ Discussion SSL vs Supervised Raise acc. of nominalized verbs Unsupervised outperforms other Supervised and SSL are limited to training data Rule-based one cant tell boundary and nominalized part * * * * * * What is the motivation of this work? In VLC 2000, I presented a state-of-the-art Japanese dependency parser using SVMs, which archived around 89.09% for standard dataset. This accuracy was the best among the previous statistical approaches for Japanese Dependency analysis. In that work, we could show the high generalization performance and feature selection abilities of SVMs This parser performed well in terms of accuracy, however there are two serious problems. The one is it is not scalable. Actually, It took about 2weeks to train SVMs even using 8000 sentences. Now about 40,000 sentences are available for training, however, it is too difficult to apply it to such a larger data. The other problem is it is too slow to use it for actual NL applications. Our current parser took about 2 or 3 seconds / sentences. It is too slow * * * * * * * * Open test III(Overall) Open test II + sentences from web news Comparison Supervised learning approach ?Language independent Model training is time consuming extremely relys on training data Semi-supervised learning ? A help to solve some unseen event Relies on labels of training data Requires more time than supervised learning Comparison Rule-based approach ?Simple and fast Language specific Linguistic knowledge, good design of rules needed Structual problem Unsupervised learning approach ?Large amount of available webpages (unlabeled data) (time consumption:incomparable to others) Demerit : same as rule-based Outline Introduction Methods Experiment result Conclusion and future work Conclusion New frameworks of Chinese NP chunking Brand new ideas Effective features Furthermore to parsing and other NLP tasks Future work Bring i