詞風與情境判斷專家系統

Abstract

宋詞為我國傳統的四大韻文之一，其風格與情境抽象，往往必須透過文學專家的分析，方得以理解作品中的情感。本研究首先擷取詩詞專家知識，用以建立分析詞風與情境之知識，並建置成詞風與情境判斷專家系統 (Expert System)。 在本論文中，我們先建置分析宋詞韻文所使用的「宋詞概念階層」，以解決古文詞彙與現代漢語文出入的問題。然後擷取專家對於判斷宋詞詞風與情境的知識，分三個階段建構系統。第一階段為宋詞斷詞：【宋詞斷詞器】將宋詞按照節奏、典故、領字、構詞、對仗、專有名詞六大模組作精確斷詞，並由「斷詞模組規則知識庫」決定模組順序。第二階段是詞風判斷：透過詞彙所代表的語意，擷取專家知識以建置「詞風概念階層」，並依此設計規則集（形式數量、內容特徵、隱含特徵、詞調特徵），建構「詞風判斷規則知識庫」。第三階段為情境判斷：利用五種感官特徵，建置「情境概念階層」，其中包含三個階層(五感識別、感官概念、主體概念)，並且建置為「情境規則知識庫」，以此判斷詞作的情境意涵。 經過實驗證明，在風格判斷上我們有80%左右的正確率，而在情境判斷的正確率上也有超過70%。在未來的工作，我們會持續增加並修訂所各種知識庫，使之能夠有更好的成果。

The SongCi is one of Chinese traditional four major verses and the ancient usually expresses one's emotion with the wonderful words. But the style and scenario of SongCi are usually implicitly described and very difficult to understand. Therefore, to understand the emotion implied in the SongCi always needs the help of literature expert. In our research, we acquire the knowledge of Ci style and scenario from domain experts, and use it as the knowledge base of Expert System. In this thesis, we constructed the SongCi Concept Hierarchy for Ci analysis firstly, because the term used in poem is different from modern Chinese article. Then, three phases are proposed to construct Ci Style and Scenario diagnosis Expert System based on the domain expertise. In first phase, SongCi Parser: uses the six modules to extract the nouns in each SongCi. In addition, we use the Parse Module Rule Knowledge Base to rank the order of the application of these modules. The second phase is to construct Ci style diagnosis rule base: Form and Amount Rule Set, Content Feature Rule Set, Imply Feature Rule Set, and Rhythm Feature Rule Set. We construct different rule sets from concepts of terms and the domain expertise. According to these Rule Sets, Ci Style diagnosis Rules Knowledge Base can be constructed. The Third phase is Ci scenario diagnosis. We use the features of five human senses (Sight, Hearing, Smell, Taste and Dermal) to construct Ci Scenario Diagnosis Rules Knowledge Base. Three steps (Sense Distinguishing, Sense Word Concept and Object Concept) are proposed to find out the Ci scenario. The experimental result shows that the accuracy is 80% at Ci Style Diagnosis and the accuracy is more than 70% at Ci Scenario Diagnosis. In the near future, we will continue to improve proposed knowledge base to increase the accuracy and the scalability of our system.