Utilizing Deep Learning Language Models to Analyze Preservice Teachers' Written Reflections

• In Event: Interdisciplinary Techniques in Educational Assessments

Abstract

Educational researchers have successfully utilized machine learning algorithms to automatically label educational data such as learners’ texts. There is mixed evidence, however, to what extent more recent deep learning algorithms outperform shallow algorithms such as logistic regression in these tasks. This study sought to compare deep learning language models for teachers’ written reflections (N=217) with logistic regression models. Results indicate that the deep learning language model outperformed the logistic regression model (86% versus 82%) for classifying written reflections. In addressing the call for interpretable models, we identified relevant features that the model used to make the classification decisions. The superior performance and interpretability of the deep learning model indicate that it can be used to advance research on written reflections.

Authors

• Peter Wulff, University of Education - Heidelberg

• Anna Nowak, University of Potsdam

• Lukas Mientus, University of Potsdam

• Andreas Borowski, University of Potsdam