以主題事件追蹤為基礎之摘要擷取

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/68810

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/68810

Title:	以主題事件追蹤為基礎之摘要擷取
Authors:	王蓮淨;Wang,Lian-Jing
Contributors:	資訊管理學系
Keywords:	查詢式摘要;擷取式摘要;K-Medoids;遺忘因子;多文件摘要;主題事件追蹤;Query-oriented Summarization;Extractive Summarization;K-Medoids;damping factor;Multi-document Summarization;tracing topic event
Date:	2015-07-27
Issue Date:	2015-09-23 14:30:28 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年由於網路發展迅速，使用者只要透過網路即可以取得所需資訊，但過多的資訊造成資訊過載之問題，因此如何在眾多資訊中擷取出重要的資訊供使用者閱讀已成為當今重要議題。然而傳統的摘要模式通常為靜態摘要，並無法針對特定主題做每日摘要的動態更新，因此本研究加入了遺忘因子，每日可更新摘要內容。並採用以主題關鍵字為基礎的方式，產出特定主題摘要內容，本研究將使用查詢式摘要(Query-oriented Summarization)法來進行多文件摘要之擷取。本研究將採用圖形網路分群架構分析文句之間潛在語意關聯性，分群方式為K-Medoids分群，探討圖形網路中所有文句節點之間的相似度，並將之做分群，得出文句間潛在語意，以提升摘要品質。實驗採用DUC 2002資料集，並以ROUGE衡量摘要品質，和自行蒐集之CNN新聞文章，其主題分別為尼泊爾大地震、伊斯蘭國及MERS，並觀察摘要結果是否能達到主題事件追蹤的功效；經實驗證明，本研究採用K-Medoids分群架構之多文件摘要方法在DUC 2002之50字、100字和200字多文件摘要，ROUGE-1值分別可達到0.2948、0.3435與0.4375，此結果在50字與100字摘要品質幾乎優於全數當年研討會之參賽者之摘要品質，另外200字摘要結果也與當年參賽者勢均力敵；而在主題事件追蹤之摘要實驗，也證實本系統可以達到主題事件追蹤摘要的功效。關鍵字：查詢式摘要、擷取式摘要、K-Medoids、遺忘因子、多文件摘要、主題事件追蹤。 ;In recent year, the developing technology of Network is getting soon. User can get information through the Internet, but it generates a problem that is information overload. Therefore, how to get some important information to user is really important now. However, the traditional technology of summarization is static, and it can′t trace the specific topic and update the summary everyday. That is why there is a damping factor in this research, and it can update the summary everyday. Also, in this research, using a way which based on topic term, and created the summary of the specific topic. In this research, using the Query-oriented Summarization way is to get Multi-document Summarization. Using the clustering architecture of graph network is to analyze the hiding semantic relation between sentences in this research. The clustering way is K-Medoids Clustering. Discussing the similarly between all sentences in graph network, and clustering these sentences are to get hiding semantic relation between sentences to rise the quality of summary. In experiment, using DUC 2002 data set and analyzing quality of summary by ROUGE, and the other data set is CNN news which topics are Nepal earthquake, Islamic State, and MERS. Observing the result of summaries is achieving the efficacy which is tracing topic event or not. The result show that using K-Medoids clustering architecture is to create Multi-document Summarizations which are 50, 100 and 200 words by DUC 2002 data set. The results of ROUGE-1 are 0.2948, 0.3435 and 0.4375. Also, the quality of summaries which are 50 and 100 words are higher than participants in DUC 2002. In addition, the result of summary of 200 words is good as participants in DUC 2002. Furthermore, in experiment of summary of tracing topic event, also proving the system in this research can achieve the efficacy which is tracing topic event. Keywords: Query-oriented Summarization, Extractive Summarization, K-Medoids, damping factor, Multi-document Summarization and tracing topic event
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	538	View/Open

社群 sharing

Loading...