在資訊科技蓬勃發展的今日,資訊化與多元化時代儼然來臨;眾多線上資訊服務的崛起,整合服務與人機互動介面成為矚目焦點。對話系統是發展長久的一項研究,其分支眾多,其中一類即是以系統代理為主要目標,是一種目標導向性的對話系統,而本文即是在這樣目的下所做的研究。 本文系統MAGEN是一強調適應性的目標導向對話系統,採用以字詞為基礎,詞類為輔助的方式,捨棄分類器與文法資訊,僅就字詞內容當作對話之依據,解決以往系統採用分類器、文法資訊所造成擴充性不足,成長性受限的窘境。在此情況下,MAGEN將適應性充分發揮在三個方面。首先,在對話領域上,由於採用詞庫與詞類的設計,可以在僅變動存放在資料庫中的知識庫即可達成領域移轉,低門檻、低成本,讓MAGEN易於適用各種對話領域。其次,對話過程中,不同的對象會有不同的詞彙用語,透過線上學習的機制,系統將可學習這些詞彙,下次使用者再度使用這些詞彙時,系統將可有效辨識,達到適應使用者的對話習慣。最後,系統本身核心相當輕量,對話皆以文字方式進行,無須圖形化介面之輔助,因此可輕易移轉成各種型態,例如:Web Service、手持式系統等。 為驗證三項適應性,設計有實驗項目,以不同類別之主題、雙回合的方式驗證適應性的情況,並實作三種應用形態的系統,更突顯實際用途上確實存有其經濟價值。 In this thesis, we present MAGEN, a light-weight dialogue system, which can be adapted to act for variant applications. It uses shallow parsing as the Natural Language Understanding component, and use classified terms as the knowledge base. In the situation where grammar is not followed strictly and mixed with Chinese and English, shallow parser can be better than full parsing and semantic analyzer. Using classified terms on the Knowledge base makes the growth of knowledge much easier and simpler. MAGEN is a frame-based conversational system where the control of dialogue is shared by users and the system. To be more specific, the user has the initial control; once the goal of the user is identified, the control is transferred to the user. New purposes/conversation can be added to MAGEN by inserting new scripts which can describe the necessary information for such conversation. The term-based knowledge and script-based goals make MAGEN very adaptive and easily transform to various application domains, such as hand-held devices, integration, systems etc.