在這數位化的時代醫學文獻、病歷、臨床紀錄等都已朝數位化的方向發展，每日所增長的量非常巨大，要如何有效管理及在必要時刻快速搜尋這些文件已成醫學檢索領域中重要的議題，但大多數使用者在搜尋時常因語意表達不清或是用詞模糊導致系統無法回傳有效的資訊給使用者。使用相關回饋(Relevance Feedback)的查詢擴展(Query Expansion)方法一直解是決這問題的主要方法之一，其中最著名的Rocchio演算法僅以字詞間的頻率來判斷，未考量字詞間其他可利用的資訊及專有名詞的重要性，因此本研究利用原始查詢與查詢結果作為基礎，主要利用Word2Vec模型所建立之醫學字詞向量以及MeSH主題詞表來分析相關回饋以及查詢字詞間所隱含之語意關係，萃取出相關回饋資訊內關鍵字詞，並利用MeSH主題字詞表以及Word2Vec模型進行字詞擴展，將其加入最後的查詢擴展集合，使查詢能更準確的回傳符合使用者需求之結果。本研實驗使用TREC 2007 Genomics資料集進行檢索效能驗證，最終結果統計本研究所提出之應用相關回饋之醫學字詞資訊於醫學查詢擴展之方法相較於Rocchio演算法在評估指標P@5提升20%、P@10提升11%、P@N提升13%、MAP提升17%以及PR Curve提升14%，顯示本研究檢索結果能更符合使用者需求。;The electronization of medical literature, medical record and clinical records are required in the information age. How to efficiently manage and search these huge data volume has become important issue in in medical retrieval domain. However, the search result sometimes is bad because user cannot effectively conversion his requirement to search keyword. Query expansion of Relevance feedback and is main method to solve this problem. Rocchio’s query expansion is most famous in relevance feedback. However, Rocchio’s method only focuses on term frequency and ignores other relationships between terms and medical terms. Therefore, this study is based on the user′s original query and search results, our research uses the medical word embeddings by the Word2Vec model and the MeSH to analyze the semantic relationship between the relevant feedback and the query words, extract the important terms in the relevant feedback information, and use the MeSH and Word2Vec model for query expansion. This study used TREC 2007 Genomics dataset for performance evaluation of retrieval. The experimental results show that the application of the medical term information residing in relevance feedback for medical query expansion can improve the retrieval performance.
