過去在資訊檢索 (Information Retrieval) 領域中，往往都是利用字詞出現頻率來衡量使用者下達之查詢字詞與系統中文件間之關聯性，此方式存在一個問題便是系統的檢索效能決定於使用者組成的查詢字詞品質，後來有研究者提出了以相關回饋(Relevance feedback)來解決此問題，相關回饋領域在向量空間回饋模型中又以Rocchio演算法為其指標，Rocchio演算法在運作時需要正相關文件與負相關文件來達成回饋動作，本研究提出一套方法利用原始查詢字詞所隱含之語意連結關係過濾回饋回來之正相關文件，達成語意式的負相關回饋，擷取出Rocchio演算法所忽略之語意資訊。本研究方法可使回饋文件所隱含之正相關資訊更精確，並於實驗結果證實，本方法不管是在MAP、P@N、PR Curve上等等評估指標都可以與Rocchio演算法不分軒輊，並且在某些情況上甚至更佳。 In the past, Information retrieval system often uses term frequency to measure the correlation between user query and corpus. The main problem is that quality of the user query can affect the retrieval efficiency. Recently, researchers have proposed the using of relevance feedback in the solving of this problem. One of the popular method is Rocchio algorithm. In the relevance feedback process, Rocchio algorithm uses positive and negative document to modify the user query. Our research proposes a method to retrieve the original query’s semantic information that Rocchio algorithm was ignored to filter irrelevant terms from positive relevance feedback. The performance of our method has been evaluated in experiment. In MAP, P@N and PR Curve show that our method is as good as Rocchio algorithm, in some case even better.
