請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/36780
標題: | 跨語言文件與圖片檢索之研究 A Study on Cross-Language Text and Image Retrieval |
作者: | Wen-Cheng Lin 林紋正 |
指導教授: | 陳信希 |
關鍵字: | 跨語言跨媒體資訊檢索,跨語言檢索,查詢翻譯,多語言資訊檢索,結果彙整,彙整策略,跨語言圖片檢索,媒體轉換,自動產生影像查詢, cross-language cross-media information retrieval,cross-language information retrieval,query translation,multilingual information retrieval,result merging,merging strategy,ross-language image retrieval,media transformation,generated visual query, |
出版年 : | 2005 |
學位: | 博士 |
摘要: | 在數位化的時代,各種類型的數位資料正不斷的快速增加中。如何在大量的資料中快速且正確地找到我們所需要的資訊是非常重要的課題。本篇論文將探討跨語言跨媒體資訊檢索,並著重在文字和圖片這二種媒體。在進行跨語言跨媒體檢索時,必須要做語言翻譯及媒體轉換,使得查詢和文件有一樣的表示方式,進而進行檢索。
首先研究跨語言資訊檢索,查詢翻譯是研究的重點。我們探討翻譯歧異性、目標多義性及未知詞處理等問題。在中英文跨語檢索的研究中,我們利用一部機器自動建立的中英文雙語詞網及中英文雙語詞典來做查詢翻譯。最好的結果達到0.1010的平均精確度,是英文單語檢索效能的69.23%。另外我們使用相似度為本的反向音譯技術來翻譯具名實體,並且設計了一個檢索系統為本的候選詞篩選器來加快反向音譯系統的速度。 接著再研究多語言檢索中的結果彙整機制。我們探討多語言檢索中的二種架構:集中式及分散式架構,並提出了多個結果彙整的方法。我們利用前k篇文章的分數來做正規化,另外也試著預測個別語言的檢索效能,以調整各語言檢索結果的權重。實驗結果顯示我們的方法可以適用於單一或多種檢索系統的分散式架構。使用前k篇文章做正規化可以避免只用第一篇文章作正規化所產生的問題。加入翻譯損害後,可以避免最後彙整後的檢索效能被一個效能很差的檢索結果拉低。 最後把媒體由文字擴展到圖片,做跨語言圖片檢索之研究。我們研究如何整合文字和影像的資訊來提高圖片檢索的效能。我們提出了一個架構能自動將一個文字的查詢轉換成影像的查詢,並使用文字查詢及自動產生的影像查詢做圖片檢索。在所提的架構中,一個重要的問題是要用哪些查詢詞來產生影像查詢。實驗結果顯示整合文字和影像的資訊可以提高圖片檢索的效能。在查詢轉換時,名詞是適合用來產生影像查詢,而使用具名實體或是動詞則是沒有幫助的。 Various types of digital data have an explosive growth nowadays. Retrieving the information we need effectively from so large amount of data is indispensable. In this dissertation, we investigate cross-language cross-media information retrieval, and consider two types of media, i.e. text and image. In cross-language cross-media information retrieval, language translation and media transformation are necessary to unify the representations of queries and documents. First, we investigate cross-language information retrieval. Query translation is the main issue. Translation ambiguity, target polysemy and unknown words handling are dealt with. We use different linguistic resources to translate query. A Chinese-English WordNet and bilingual dictionary are used to deal with Chinese-English information retrieval. The best model achieves 0.1010 average precision, 69.23% of monolingual information retrieval. For named entity translation, a similarity-based backward transliteration framework is adopted. We propose an IR-based candidate filter to enhance the efficiency of the similarity-based backward transliteration. We then investigate merging mechanisms in multilingual information retrieval. We consider two different MLIR architectures: centralized and distributed architectures. Several merging strategies are proposed. Normalized-by-top-k merging is proposed to normalize similarity scores. We also consider the retrieval effectiveness of each individual run in merging stage. Experimental results show that the proposed approaches are feasible in single and multiple IR system architectures. Normalized-by-top-k merging overcomes the drawback of normalized-score merging. Normalized-by-top-k merging with translation penalty could avoid performance drop down caused by a poor intermediate run. We further extend the media to image and investigate cross-language image retrieval. We explore the integration of textual and visual information in image retrieval and propose a scheme to deal with cross-language image retrieval. An approach that automatically transforms textual queries into visual representations is proposed. Which query terms should be adopted to generate a visual query is investigated. Experimental results show that integrating textual and visual information improves retrieval performance. Nouns are appropriate to generate visual queries, while using named entities and verbs is helpless. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/36780 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-94-1.pdf 目前未授權公開取用 | 1.26 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。