Açık Uçlu Maddelerin Puanlanmasında ChatGPT ve Gerçek Puanlayıcıların Puanlayıcılar Arası Güvenirlik Bakımından İncelenmesi

Abstract

The aim of this study is to examine the inter-rater reliability of the responses to open-ended items scored by ChatGPT, an artificial intelligence-based tool, and two real raters according to the scoring keys. The study group consists of 30 students, aged between 13 and 15, studying in Eskişehir province in the 2022-2023 academic year. The data of the study were collected face-to-face with the help of 16 open-ended items selected from the sample questions published in the International Student Assessment Program-PISA Reading Skills. Correlation, percentage of agreement and the Generalizability theory were used to determine inter-rater reliability. SPSS 25 was used for correlation analysis, Excel for percentage of agreement analysis, and EduG 6.1 for the Generalizability theory analysis. The results of the study showed that there was a positive and high level of correlation between the raters, the raters showed a high level of agreement, and the reliability (G) coefficients calculated using the Generalizability theory were lower than the correlation values and percentage of agreement. In addition, it was determined that all raters showed excellent positive correlation and full agreement with each other in the scoring of the answers given to the short-answer items whose answers were directly in the text. In addition, according to the results of the Generalizability theory, it was found out that the items (i) explained the total variance the most among the main effects and the student-item interaction (sxi) explained the most among the interaction effects. As a result, it can be suggested to educators to get support from artificial intelligence-based tools such as ChatGPT when scoring open-ended items that take a long time to score, especially in crowded classes or when time is limited.