Minimally Supervised Japanese Named Entity Recognition Resources and Evaluation.pdfVIP

Minimally Sup ervised Japanese Named Entity Recognition: Resources and Evaluation   Takehito Utsuro , Manabu Sassano  Department of Information and Computer Sciences, Toyohashi University of Technology Tenpaku-cho, Toyohashi, 441-8580, Japan utsuro@ics.tut.ac.jp  Fujitsu Lab oratories, Ltd. 4-4-1, Kamiko danaka, Nakahara-ku, Kawasaki 211-8588, Japan sassano@ ab.fujitsu.co.jp Abstract Approaches to named entity recognition that rely on hand-crafted rules and/or sup ervised learning techniques have limitations in terms of their p ortability into new domains as well as in the robustness over time. For the purp ose of overcoming those limitations, this pap er evaluates named entity chunking and classication techniques in Japanese named entity recognition in the context of minimally sup ervised learning. This exp erimental evaluation demonstrates that the minimally sup ervised learning metho d prop osed here improved the p erformance of the seed knowledge on named entity chunking and classication. We also investigated the correlation b etween p erformance of the minimally sup ervised learning and the sizes of the training resources such as the seed set as well as the unlab eled training data. 1. Intro duction recognition prop osed by (Collins and Singer, 1999;