A data-driven approach for the identification of features for automated feedback on academic essays

Abstract

For predicting and improving the quality of essays, text analytic metrics (surface, syntactic, morphological and se- mantic features) can be used to provide formative feedback to the students in higher education. In this study, the goal was to identify a sufficient number of features that exhibit a fair proxy of the scores given by the human raters via a data-driven approach. Using an existing corpus and a text analysis tool for the Dutch language, a large number of features were extracted. Artificial neural networks, Levenberg Marquardt algorithm and backward elimination were used to reduce the number of features automatically. Irrelevant features were eliminated based on the inter-rater agreement between predicted and human scores calculated using Cohen’s Kappa (κ). The number of features in this study was reduced from 457 to 28 and grouped into different categories. The results reported in this paper are an improvement over a similar previous study. Firstly, the inter- rater reliability between the predicted scores and human raters was increased by tweaking the corpus for overfitting for average scores. The resulting maximum value of κ showed substantial agreement compared to moderate inter-rater reliability in the prior study. Secondly, instead of using a dedicated training and test set, the training and testing phases in the new experiments were performed using k-fold cross validation on the corpus of texts. The approach presented in this research paper is the first step towards our ultimate goal of providing meaningful formative feedback to the students for enhancing their writing skills and capabilities.

Marco Kalz

Professor of Educational Technology

My research interests is on open education, pervasive technologies and formative assessment to support (lifelong) learning and knowledge construction.