The article introduces ‘G-SciEdBERT’, a contextualized German Science Education BERT developed for scoring written responses to science questions in German. It is trained on a corpus of 50K German written science responses to the Programme for International Student Assessment (PISA) 2015. The study found a significant improvement in scoring accuracy with G-SciEdBERT, with a 10% increase compared to the standard German BERT (G-BERT), emphasizing the importance of specialized language models in the field of AI in education.
Publication date: 9 Feb 2024
Project Page: https://arxiv.org/abs/2402.06584
Paper: https://arxiv.org/pdf/2402.06584
