The purpose of this study was to investigate inter- and intra- rater reliability in an interview and a computerized oral test. It was also examined whether rater characteristics influenced on their reliability and biases, and finally the scores of both tests were compared with those of the Versant test using an automated computer rating system. For the study, the data from 21 Korean university students and 18 Korean or native speakers of English raters with various characteristics were collected. Some of the main findings from the study were as follows. First, rater severity was significantly different in each test, but each rater consistently graded on both tests suggesting lower inter-rater reliability and higher intra-rater reliability. Secondly, rater severity was impacted by the rater characteristics such as mother tongue, gender, age, and major. Lastly, there existed a positive correlation among the scores of the three tests, indicating that the scores of human beings and computers are strongly related.

본 연구에서는 일반화가능도 이론을 이용하여 폴리그라프 검사에 사용된 관련 질문의 개수와 반복측정 횟수 (차트의 수), 채점자 수가 폴리그라프 검사의 신뢰도에 미치는 영향을 평가하였다. 검찰청에서 형사피의자를 대상으로 Backster ZCT를 사용한 폴리그라프 검사자료 중 31명의 폴리그라프 검사자료를 표본추출하였으며, 31명의 검사자료를 13명의 채점자가 수치적 채점방법을 이용하여 채점한 점수에 대하여 일반화가능도 이론을 적용하여 분석하였다. 분석결과, 피검사자의 변량성분이 43.97%로 가장 컸으며, 다음으로 잔여오차변량성분이 16.84%, 피검사자와 반복측정 횟수의 상호작용오차변량성분이 12.17%, 피검사자와 반복측정 횟수, 관련 질문 개수의 삼원상호작용오차변량성분이 10.31%였으며, 나머지 변량성분은 모두 7% 미만이었다. 관련 질문의 개수와 반복측정 횟수, 채점자의 수에 따른 일반화가능도 계수를 산출한 결과, 바람직한 일반화가능도 계수인 0.80 이상을 보이는 조건은 관련 질문 2개 이상과 반복측정 3회 이상, 채점자 2명 이상의 조합인 것으로 나타났다.

Serious inherent problems with practicality, intra-rater and inter-rater reliability overshadow the known positive washback effects of performance assessment in language education. In particular, it has been welldocumented that inter-rater reliability poses a serious threat to overall test validity, since individual raters necessarily measure performance according to their own subjective severity criteria in language proficiency. However, language testing has witnessed a remarkable series of breakthroughs in performance assessment during the recent advent of the information era. One such breakthrough utilizes state-of-the-art automatic speech recognition (ASR) technology for oral proficiency interviews(OPI). Granting that current forms of ASR technologies may not produce results with the reliability needed to accommodate highstakes standardized test administration, they do offer aid in approaching the thorny issues of practicality and inherent human inter-rater subjectivity. Accordingly, this paper is intended to investigate the degree to which ASR-based OPI ratings match similar human-conducted OPI ratings by employing correlational analyses on the basis of degrees of rater severity. Furthermore, this paper attempts to explore a method of enhancing the robustness of ASR-based OPI ratings which capitalizes on suprasegmental information by measuring fluency based principally on the test-takers’ response time length.