Improving Transformer-Based Automated Scoring of K-12 Science Items

• In Event: Electronic Board Session II

Abstract

Transformer-based language models, such as ELECTRA, are often effective at scoring short, constructed response (SCR) items. Yet, for some times, ELECTRA underperforms. This paper focuses on six, difficult-to-autoscore K-12 science items. We explore four techniques for improving model performance. Results indicate our novel approaches can improve SCR autoscoring with ELECTRA.

Authors

• Alexander Kwako

• Chris Ormerod, Cambium Assessment, Inc.