MARVS理論再探：以量化觀點比較近義詞的詞義頻率與搭配詞共現值

MARVS Revisited: Incorporating Sense Distribution and Mutual Information into Near-Synonym Analyses

在 MARVS (Module-Attribute Representation of Verbal Semantics) 這個理論裡，動詞的分辨是以事件訊息為基礎，而事件訊息主要包括了事件模組和角色模組。黃居仁等 (2000) 曾以 MARVS 檢查近義詞的語意，並建議使用此理論來凸顯近義詞間的差別。本文則針對 MARVS 的理論，加強其分析結果，並加入兩個新的步驟。這兩個步驟分別是：一、量化比較和分析近義詞在語料庫的詞義，二、透過 Mutual Information 的計算比較不同近義詞的搭配詞。這兩個步驟能增加 MARVS 在語義分析的可檢驗性，也更能奠定其理論運用在自動擷取詞彙語意上的基礎。

In MARVS (Module-Attribute Representation of Verbal Semantics), verbs are differentiated based on eventive information, which is comprised of event modules and role modules. Huang et al. (2000) used MARVS to examine near-synonyms and suggested that it can highlight the difference between synonymous sets. This paper suggests that the operational steps underlying a MARVS analysis can be improved by analyzing the sense distribution of the near synonyms and by looking at the Mutual Information values of the collocating words. Both these steps increase the verifiability of the semantic analysis in MARVS and set the groundwork for automatic extraction of lexical meaning.