本研究旨在探究大學學力測驗與指定科目考試中，有關「連貫性」 的克漏字試題（包括克漏字和文意選填測驗），藉以一窺台灣高中學生在這方面的表現，並進一步透過試題分析，學生表現，和題目鑑別度等細項，深入討論台灣高中英語教育中，老師和學生在課堂上可以加強的部分。 首先，我們將連貫性這個觀念加以分類和說明，根據修改過後的分類，將大學入學考試試題分為連接，指涉，字彙連貫三種。字彙連貫又可以細分為重複和常連用字兩種。我們就題目出現的頻率來討論，分析後發現，在克漏字測驗和文意選填題目中，考字彙連貫出現頻率最高。就學生答對率表現來看，常連用字試題，通過率最高，而指涉試題，通過率最低。這結果顯示台灣學生在局部連貫性上表現較佳，對於整體連貫性較不熟悉。試題出現在整篇文章的主題句，也會造成答對率較低。另就試題的鑑別度而言，低鑑別度往往和低通過率的試題重疊，表示大多數學生通過率低的題目，往往因為困難而無法鑑別學生能力。 / In this study, we analyzed the multiple-choice rational cloze tests used in the Subject Ability Test (SAT) and Appointment Subject Test (AST) from 2002 to 2004. The item types were mainly analyzed according to the five cohesion types that Halliday and Hasan proposed in their 1976 research. Through these classifications, we gained better knowledge about whether the cohesion types affected the students’ performances. Furthermore, we also investigated the possible reasons. First of all, we refined the cohesion types proposed by Halliday and Hasan proposed in their 1976 research. In our study, the cohesion types we looked were conjunction, reference, and lexical cohesion. We wanted to find how the cohesive types were tested in the tests. From the analysis of cloze tests and the WYXT tests, the result we got was that lexical cohesion was tested more often than other cohesive categories. For the passing rate of SAT and AST, we found that the collocation test items were the easiest while the reference test items were the most difficult. According to Bachman (1982, 1985), Test items in Type 2 and 3 (repetition, conjunction, and reference) do have lower passing rates. This means students have difficulties in global cohesion. Furthermore, students had more difficulties in answering the test items located in the first or last sentence, which are usually associated with main ideas of the paragraph – a global concept. Therefore, the locations of the test items also had an influence on the passing rate. Lastly, for the discrimination index, those test items with low D value always co-occurred with low passing rate test items, meaning that difficult items fail to discriminate between different proficiency levels.