利用WordNet 判斷字詞包含關係─ 應用於動態階層文件分群; Using WordNet to Infer Containment Relationship─ Applied to Dynamic Hierarchical Clustering

NCU Institutional Repository > 管理學院 > 資訊管理研究所 > 博碩士論文 > Item 987654321/61062

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/61062

题名:	利用WordNet 判斷字詞包含關係─ 應用於動態階層文件分群;Using WordNet to Infer Containment Relationship─ Applied to Dynamic Hierarchical Clustering
作者:	詹欣逸;Chan,Hsin-Yi
贡献者:	資訊管理學系
关键词:	字詞包含關係;動態分群;分類學;階層分群;WordNet;Containment relationship;Taxonomy;Dynamic clustering;Hierarchical clustering
日期:	2013-07-24
上传时间:	2013-08-22 12:11:01 (UTC+8)
出版者:	國立中央大學
摘要:	資訊爆炸的現在及未來，不管企業或個人都需要方法有效組織資訊。本研究目標在於將文件動態進行階層分群，以便使用者整理、瀏覽及搜尋日積月累的海量資訊。其做法結合了分群、分類及分類學。本研究以Dynamic Hierarchical Clustering Based on Taxonomy (DHCT) 之動態分群架構為基礎去改善內部方法。其中我們改變了文件相似度的計算以降低分群複雜度，而分類學方面本研究提出以WordNet 中兩字詞路徑來判斷包含關係的方法，透過使用MeSH_2011 資料集的實驗證明能有效建立分類學，以分類學為基礎的做法則能成功建立上下包含的階層分群目錄標籤，解決階層分群無法產生有意義標籤之瓶頸。此外本研究改善DHCT 所提出推論字詞包含關係之Conditional Probability (CP) 計算，發現加入適當標籤確實能幫助CP 正確推論有一字多義的字詞之包含關係，而結合WordNet 及Google 建置分類學的做法不僅提高整體正確率，更能突破WordNet 本身字詞不足之限制。DHCT 已解決文獻中面臨的許多問題並證實優於傳統的UPGMA 及BKM 階層分群法，經過本研究的改良更將分群複雜度由O(nm2)降至O(nm)，並透過使用Wikipedia 資料集的實驗證明能改善DHCT 約20%的F1Score，產生更正確、更能幫助使用者瀏覽之階層分群結構。 The number of text document managed in business and personal computer continues to grow at an amazing speed. We need an efficient way to organize, manage, access, search and browse such large repositories of documents. One of the popular techniques is dynamic hierarchical clustering, which is our focus. This study improves the taxonomy method in Dynamic Hierarchical Clustering Based on Taxonomy (DHCT), and our framework is mainly based on it, which combines the techniques of clustering, classification, and taxonomy. In order to reduce the time complexity, we use Ward’s minimum variance and NGD to calculate document similarity. We also proposed two methods to infer containment relationships between terms for building taxonomy. One is called CR, which makes use of term paths in the WordNet. Another called CP+Label, which improves the Conditional Probability (CP) (proposed by DHCT) by adding an appropriate label when the term is polysemy. These taxonomies are later used as the cluster label to let users easier to browse and search. DHCT had been proved that it is better than the traditional method: UPGMA and BKM, and our experiment results on MeSH_2011 show that both methods we proposed are suitable for producing a meaningful taxonomy and are better than DHCT as well. Moreover, by merging the taxonomies constructed by WordNet and Google, our method not only improves about 20% of the overall F1 score on Wikipedia text collection, but also breaks through the limits while using the WordNet.
显示于类别:	[資訊管理研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	622	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....