專有詞彙之定義式問題答案句自動擷取系統

Abstract

本論文針對專有詞彙之定義式問題，建立一套以電子書為答案來源之定義式 答案句自動擷取系統雛形。本論文運用資訊檢索的概念由電子書內容中選取候選句子，並提出以維基百科等外部知識來源衡量句中所包含的字詞與查詢專有詞彙關鍵字的關聯權重值，作為系統挑選答案句之評分依據。本論文方法能夠讓答案不受限於特定定義式句型，而找出更多能夠幫助了解該專有詞彙之相關定義解釋說明的內容作為答案。並採用句子間字詞的語意關聯度，綜合評估計算答案句間的相似程度值，以不同聚落分析演算法對答案句進行自動分群處理，使答案句能依所涵蓋概念類似性分群整理呈現給使用者。由實驗結果顯示，本論文研究方法所擷取之答案句及排序順序，與專家人工評分挑選的標準答案結果一致性很高。This thesis proposes a sentences retrieval prototype system for answering definitional questions of domain-specific terms. Our approach select candidate answer sentences from eBooks. We propose a term weighting model using external knowledge (e.g. Wikipedia) to measure the importance of each terms in the sentence toward the querying domain-specific term. We then rank candidate answer sentences according to the sum of its term weights. Retrieved answers are not limited to specific definitional pattern. Any sentences which would be helpful for understanding the definition and explanation of the domain-specific terms can be retrieved by our proposed system. Finally, We summarize the answer result automatically by clustering answer sentences based on their semantic relatedness. Experimental results show that the ranked list of answer sentences retrieved by our proposed system are consistent with the expert voted ground-true answer in most cases.