Identifying Critical Features for Formative Essay Feedback with Artificial Neural Networks and Backward Elimination

Abstract

For predicting and improving the quality of essays, text analytic metrics (surface, syntactic, morphological and semantic features) can be used to provide formative feedback to the students. In this study, the intent was to find a small number of features that exhibit a fair proxy of the scores given by the human raters. 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 extracted features automatically. Irrelevant features were eliminated based on the inter-rater agreement between predicted and human scores calculated using Cohen’s Kappa ($$\kappa $$ ). By using our algorithm, the number of features in this study was reduced from 457 to 23. The selected features were grouped into six different categories. Of these categories, we believe that the features present in the groups “Word Difficulty” and “Lexical Diversity” are most useful for providing automated formative feedback to the students. 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.