Paper Abstract and Keywords |
Presentation |
2020-09-10 15:50
chiVe: Towards Industrial-strength Japanese Word Vector Resources
-- Constructing and Improving Embedding with Tokenizer "Sudachi" and Ultra-large-scale Web-based Corpus "NWJC" -- Sorami Hisamoto, Takashi Yamamura, Akihiro Katsuta, Yuto Takebayashi, Kazuma Takaoka, Yoshitaka Uchida (WAP), Teruaki Oka, Masayuki Asahara (NINJAL) NLC2020-9 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
Pretrained word embeddings (word vectors, distributed representations) are fundamental and important components for recent neural network based natural language processing. However, there are not many resources available for Japanese language that is useful for the practical purpose. “chiVe” is our attempt to construct useful word embedding resource, using morphological analyzer “Sudachi” that provides multi-granular tokenization results, and 10 billion words scale corpus “NWJC”. We made this resource publicly available under Apache 2.0 license that allows commercial usage ( https://github.com/WorksApplications/chiVe ). In this article, we describe the overview of the project, and introduce our ongoing work to further improve the resources which includes handling Out-of-Vocabulary words, utilizing the synonym resources, and adapting to specific domains. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Distributed Representation / Morphological Analysis / Out-of-Vocabulary / Synonym / Domain Adaptation / / / |
Reference Info. |
IEICE Tech. Rep., vol. 120, no. 166, NLC2020-9, pp. 40-45, Sept. 2020. |
Paper # |
NLC2020-9 |
Date of Issue |
2020-09-03 (NLC) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
Download PDF |
NLC2020-9 |
|