著者

行野 顕正 青木 さやか 谷川 龍司 [ 他 ]

出版者

九州大学

雑誌

Research reports on information science and electrical engineering of Kyushu University (ISSN:13423819)

巻号頁・発行日

vol.11, no.2, pp.115-119, 2006-09

---

We propose using long and low-frequency part of speech (POS) strings for document separation between native English documents and non-native English documents. The long POS strings were ignored in previous works because their frequencies in training data are too small to estimate their probabilities. Meanwhile, a research of language identification showed that the long and low-frequency byte strings were useful for language identification among similar languages. There are some similarity between language identification and document separation between native English documents and non-native English documents, for example long POS strings are more peculiar to one class than short ones, though there is a difference between POS and byte. Therefore, we can expect higher accuracy by using long and low-frequency POS strings. Some experiments are described in this paper. These experiments show that the proposed method has higher accuracy than previous ones.