Presentation 2018-09-07
Short text categorization with fine-tuning
Kazuya Shimura, Jiyi Li, Fumiyo Fukumoto,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) We propose an approach for multi-label categorization of short texts and explore the use of a hierarchical structure (HS) of categories. The lower the HS level, the worse the categorization performance because the number of training data per category in the lower level is much smaller than that in the upper level. We use a transfer learning technique by fine-tuning to learn the hierarchical structure of categories. We applied the Convolutional Neural Network (CNN) with fine-tuning. By transferring and finely tuning the trained parameters of CNN at each level of HS, we aim to improve the accuracy at the lower level of the HS. The results using a benchmark dataset show that the proposed method is competitive with the state-of-the-art CNN based multi-label categorization method XML-CNN, as the improvement of our method attained at 1.2% in Micro-F1 and 7.4% in Macro-F1 compared with XML-CNN.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) short text categorization / multi-label categorization / convolutional neural network / fine-tuning
Paper # NLC2018-20
Date of Issue 2018-08-30 (NLC)

Conference Information
Committee NLC / IPSJ-DC
Conference Date 2018/9/6(2days)
Place (in Japanese) (See Japanese page)
Place (in English) Seikei University
Topics (in Japanese) (See Japanese page)
Topics (in English) The Thirteenth Text Analytics Symposium
Chair Takeshi Sakaki(Hottolink) / Michiko Oba(Hitachi)
Vice Chair Mitsuo Yoshida(Toyohashi Univ. of Tech.) / Kazutaka Shimada(Kyushu Inst. of Tech.)
Secretary Mitsuo Yoshida(Ryukoku Univ.) / Kazutaka Shimada(NTT) / (Kyushu Univ.)
Assistant Takeshi Kobayakawa(NHK) / Hiroki Sakaji(Univ. of Tokyo)

Paper Information
Registration To Technical Committee on Natural Language Understanding and Models of Communication / Special Interest Group on Document Communication
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Short text categorization with fine-tuning
Sub Title (in English)
Keyword(1) short text categorization
Keyword(2) multi-label categorization
Keyword(3) convolutional neural network
Keyword(4) fine-tuning
1st Author's Name Kazuya Shimura
1st Author's Affiliation University of Yamanashi(Univ of Yamanashi)
2nd Author's Name Jiyi Li
2nd Author's Affiliation University of Yamanashi(Univ of Yamanashi)
3rd Author's Name Fumiyo Fukumoto
3rd Author's Affiliation University of Yamanashi(Univ of Yamanashi)
Date 2018-09-07
Paper # NLC2018-20
Volume (vol) vol.118
Number (no) NLC-210
Page pp.pp.73-78(NLC),
#Pages 6
Date of Issue 2018-08-30 (NLC)