對使用者而言,無法在檢索結果的前幾筆資料中,找到所需的資訊是一件相當困擾的事情。資訊檢索系統一直以來存在的問題就是回傳的檢索結果中包含過多不相關的文件,增加使用者查詢上不必要的負擔。在過去的研究中,雖有不少學者以相關回饋的方式來改善檢索的效率,卻未見以負相關回饋的資訊做進一步的探討。因此本研究將相關回饋之負相關文件所隱含之資訊做運用,結合TREC 6資料的特性與字詞的分佈,在傳統文件字詞權重計算方法加入字詞敏感度的概念。利用相關回饋以及負相關回饋文件之字詞出現的情況和頻率找出只在負相關的高頻特徵來建立字典,再利用字典對檢索文件中之字詞進行加權,藉此降低負相關文件與相關資訊的相似度,以幫助檢索結果的重排序,協助使用者快速找到所需的資訊。實驗結果顯示,利用本方法對文件之字詞進行權重調整,能提高檢索結果之P@10至P@100的平均準確率,優於Rocchio演算法的效能。因此,本研究驗證負相關回饋的資訊對於文件重排序是有用的。 Too much non-relevant information retrieved put burden on the user. In past studies, many scholars used relevance feedback to improve document retrieval performance. This paper proposes a method to apply the information of non-relevant documents with TREC 6 data characteristics and term distributions. The proposed method uses the term-appearance situation and term frequency of non-relevant documents to find negative features to create a dictionary, and then uses features of the negative dictionary to adjust term weights of retrieved documents to reduce the weights of non-relevant documents for re-ranking the search result. We compare the proposed method to Rocchio algorithm, the results of our experiment show that P@10 to P@l00 of our purpose method significantly outperforms Rocchio. Therefore, our study verifies that the information of non-relevance feedback can be useful for document re-ranking.