深度神經網路架構之跨語言線上百科連結;Cross-lingual Encyclopedia Linking in DNN Framework

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/79568

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/79568

Title:	深度神經網路架構之跨語言線上百科連結;Cross-lingual Encyclopedia Linking in DNN Framework
Authors:	潘照霖;Pan, Chao-Lin
Contributors:	資訊工程學系
Keywords:	線上百科;維基百科;百度百科;跨語言;連結;文章表示向量;詞向量;深度學習;卷積類神經網路;長短期記憶模型;注意力機制
Date:	2018-10-01
Issue Date:	2019-04-02 15:03:08 (UTC+8)
Publisher:	國立中央大學
Abstract:	維基百科的出現，徹底改變人們學習新知識的習慣。其發展了超過298種不同語言的版本，在不同語言之間，卻存在文章數量不平衡的情形，英文維基的文章數量遙遙領先其他語言版本。以中文維基為例，其文章數量僅英文維基的六分之一。除此之外，維基百科在不同語言版本之間的跨語言間連結亦嚴重缺乏，根據統計，英文維基百科僅有2.3%的文章具有跨語言連結至其中文版本。除維基百科外，在各國仍有提供特定語言的線上百科存在，其內容更豐富於維基百科的其他語言版本。因此我們以《英文維基》和《百度百科》為目標，建置「跨語言線上百科跨連結」，除有助於全球知識共享，更有助於跨語言相關研究。以往跨語言百科連結的作法，常需要依賴語言特性。因此我們提出一深度學習模型，不需要依靠語言特性與線上百科的架構產生特徵，僅以文章內文作為訓練資料的依據，運用各種神經網路，辨別跨語文章的語意上的相似程度。在面對不同語言版本的資料時，僅需替換預訓練詞向量即可。 ;The emergence of Wikipedia has completely changed the habit of learning new knowledge. It has developed more than 298 different language versions. However, it exists imbalances in the number of articles between different language versions of Wikipedia. The number of articles in English Wikipedia is far ahead over other languages. Take Chinese Wikipedia as an example, the number of articles is only one-sixth of the English Wikipedia. In addition, Wikipedia′s cross-language links between different language versions are also seriously lacking. According to the statistics, only 2.3% of English Wikipedia articles have cross-language links to its Chinese versions. Despite of Wikipedia, some other countries has its own online encyclopedias and its content is much more abundant than its language versions of Wikipedia. Therefore, we aim to build a "cross-language online encyclopedia between "English Wikipedia" and "Baidu Baike". It is not only contributing to global knowledge sharing, but more conducive to cross-language related research. In previous CLAL works, their methods usually depend on the language characteristics and the structure of encyclopedia. Therefore, we propose a deep learning model, which only uses the textual main content as the basis of the training data, and various neural networks to distinguish the semantic similarity of the cross-language articles. When facing data in different language versions, the only thing to do is replacing the pre-training word embedding.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	130	View/Open

社群 sharing

Loading...