以語料庫驅動之隱喻源域界定研究

Abstract

一般來說，領域訊息在隱喻形成的時候會產生互換的情形，而在概念隱喻裡 (Lakoff and Johnson 1980; Lakoff, 1993)，互換的兩個領域訊息分別為源域與目標域。源域通常比目標域較為具體，所以源域提供的訊息對於理解目標域的概念來說是十分關鍵的要素。因此，在過去曾經有相關的研究嘗試定義源域，可是這些研究通常會遇到以下兩個問題：第一，對源域定義的差異、第二，對源域界定的主觀程度不一致。 本論文為了克服這些困難，使用了兩個面向的方式將隱喻之源域做系統化地界定：一是由上而下、二則為由下而上。一般來說，由上而下的方式通常會假設一個上層的概念，而源域就是這些概念裡的一部份。唯有當源域已決定之後，才會把各種的隱喻例子歸類到不同的源域。而隱喻概念這個理論 (Lakoff and Johnson 1980; Lakoff, 1993) 用的就是這一種方式，例如：在LOVE IS JOURNEY 的隱喻概念裡，當決定JOURNEY為源域之後，一些相關的例子才會接著分類到該源域內。相反地，由下而上的方式則會先累積人類的用語，而這些資訊會構成較上層的知識。此種方式通常會以頻率為基礎，如：(Rosch and Mervis, 1975; Labov, 1973) 所舉出的典型模範 (prototype) 指的就是共現度高的搭配詞組。在隱喻相關研究中，這兩種方式都曾有學者使用過，但是，在文獻中卻少有研究對這兩種方式做全面性的比較。本論文使用了WordNet、SUMO (Suggested Upper Merged Ontology) (知識本體) 與Chinese Sketch Engine (中文詞彙特性速描系統) 等計算語言學的工具來界定源域，是第一個整合從由上而下及由下而上兩種方式來比較領域訊息的研究。並且，我們透過計算語言學工具來驗證領域界定的理論架構，得到的結果有助於我們進一步了解人類對事物的歸類。而更重要的是，本論文結合了計算語言學、語言學理論及心理語言學的知識來界定領域概念，因此其成果將具有跨領域的貢獻。

Domain information is mapped in the formulation of metaphors. In conceptual metaphors (Lakoff and Johnson 1980; Lakoff, 1993), the two domains mapped are the target domain and source domain. Target domains are usually more abstract, for the most part borrowing information from the source domains, whereas source domains are more concrete, and the information they provide is an important key to understanding the concept of the target domains. Due to the importance of source domains, studies attempting to operationally define source domains involved in conceptual metaphors have been undertaken in the past. However, authors of previous studies encountered difficulties due to two main reasons: (a) the variation in the specificity of the source domains; and (b) the subjectivity of individuals when determining source domains. This thesis overcomes these difficulties by employing both top-down and bottom-up approaches to determine the source domain of a conceptual metaphor. A top-down approach to metaphor identification usually sorts metaphorical instances according to pre-determined source domains which come from a conceptual knowledge system (such as an ontology). An example of this is the prediction of metaphorical instances based on a general metaphor such as LOVE IS JOURNEY, where JOURNEY is pre-determined (Lakoff and Johnson, 1980). A bottom-up approach, on the other hand, accumulates knowledge about domain though language use (or collocation). The bottom-up approach has been used in frequency-based or prototypical (Rosch and Mervis, 1975; Labov, 1973) definitions of source domains. Both top-down and bottom-up approaches have been used to identify source domains in metaphor studies, however the precision of the two approaches has not previously been compared. It is therefore the intention of this thesis to compare these two approaches employed to determine source domains, using computational tools such as WordNet, SUMO (Suggested Upper Merged Ontology) and Sketch Engine. The outcome of the thesis will be a comprehensive comparison of these two approaches regarding the determination of source domains. This thesis is the first study to compare both top-down and bottom-up approaches in defining domain information. Furthermore, our use of computational tools with a linguistic framework of domain identification, as well as evidence from psycholinguistic experiments, will contribute not only to computational linguistics, but will also enhance our ability to define domains. This in turn will make a substantial contribution to understanding the human categorization of concepts through linguistic evidence. Based on the combined results of the two approaches, this thesis also proposes a hierarchical model for source domain definition. This model facilitates automatic identification of metaphors and explains why certain domains overlap in conceptualization.