過去在向量空間模型上的相關回饋研究，是以使用者對於系統所回傳的相關文件清單，萃取其字頻資訊作為回饋的特徵值，該模型以Rocchio查詢擴張最被廣泛使用，然而Rocchio是將相關文件字頻扣除不相關文件字頻，得出字頻高低排序後，用以當作查詢擴張的字詞來源，其演算法效能雖具一定水平，但是否有其它更好過濾不相關字詞的機制，仍然是一個有趣的議題。然而，近年來語意搜索(Semantic search)的概念逐漸形成，線上的搜尋引擎開始提倡以使用者查詢的語意做為搜尋依據，主要考量的是關鍵字上面所涵蓋的語意概念，而非單純使用關鍵字本身。因此，本研究基於現有的自然語言(Natural language processing)相關研究，運用概念萃取演算法LDA，將使用者所提供的相關文件資訊，萃取其概念特徵值，並且透過萃取概念的過程，間接將不具語意的字詞排除，當使用者回饋的概念萃取完成後，予以整體概念脈絡的回饋。最後透過實驗驗證，證明不論在P@K、MAP或者PR curve等指標，概念處理的回饋效果相較於Rocchio關鍵字處理的回饋效果來得好，因此以概念做為回饋特徵會更貼近於使用者的資訊需求。In the past, the main method in the application of relevance feedback was to aggregate the term frequencies in the feedback documents that the user provided as the feedback characteristics in the vector space model. Rocchio’s query expansion was the most popular one. It reduced the term frequencies of non-relevant documents from the relevant ones first. Then, it ordered the terms by the frequency and kept the top ones as the source for query expansion. Rocchio’s method has been well-performed. Nevertheless, it still is an interesting question: “Is there any better mechanism to filter the non-relevant terms from relevant ones?” Recently, the idea of semantic search is getting more and more popular. Instead of using term-matching to search documents, many on-line search engines promote itself by using the semantic meaning of the user’s query. It is concerned with the semantic meaning that the key words covered. Based on the NLP technique, this research is interested in the application of a concept-retrieval algorithm, LDA, to collect the concept characteristics from relevant documents and exclude the non-relevant terms by the process of concept agglomeration. The performance of our method has been evaluated in the experiments. In P@K, MAP and PR curve, the effect of concept-based feedback is better than the term-matching-based one.
