概念式自動問答探索系統

Abstract

本論文提出一套以潛在語意分析 (LSA) 為核心技術的概念式自動問答探索系統。自動問答探索系統能讓使用者以自然語言的方式輸入新問題，系統會從歷史問答集中找出符合的答案。此套概念式自動問答探索系統首先學習三種知識庫：問題詞鍵和答案詞鍵間的關係矩陣、概念空間知識庫和答案類型判別知識庫。其中，問題詞鍵和答案詞鍵間的關係矩陣以及建構概念空間知識庫係利用潛在語意分析學習而得，而疑問詞鍵與答案類型的關係則是以一機率模組學習而得。依據此三種知識庫，概念式自動問答探索系統會比對新問題與問答集中的答案之詞鍵相似度、概念描述相似度和答案類型相似度，不僅透過詞鍵比對，更擷取出問題的概念描述，以期將自動問答探索系統提昇至語意層面。實驗中，先人工建構了關於籃球和棒球規則的問答組，概念式自動問答探索系統的效能在排序準確度 (TRDR) 評估上平均可達83.87%，準確率平均可達36.4%，而查全率平均可達44.2%；較非概念式的自動問答系統的效能在排序準確度、準確率和查全率的評估上，平均增幅分別為17.75%、8.53%、18.37%。

In this thesis, we propose an automatic concept-based answer-finding system (ACAF) that exploits LSA as its core technology. Users issue their new questions into the ACAF system by natural language, and the system will return suitable answers from the question-and-answer set (QA set). To accomplish its task, ACAF employs machine learning techniques to learn three kinds of knowledge: the relationship between question keywords and answer keywords, conceptual space, and answer-type knowledge. LSA is used learn the relation matrix between question keywords and answer keywords; LSA is used construct the conceptual space as well. In addition, a probabilistic model is employed to train the relations between interrogatives and answer types. According to these three knowledge bases, ACAF calculates the similarity of a new question and the answers in the QA set. ACAF not only compares keyword-similarity but also retrieves the concepts of the new question. In this manner, an automatic answer-finding system can be promoted to the semantic level. ACAF was evaluated by using a QA set about basketball and baseball rules, and average TRDR of 83.87%, average precision of 36.4% and average recall of 44.2% were achieved. Compared to ACAFW, the average increasing range of TRDR, precision, and recall are 17.75%, 8.53%, and 18.37% respectively.