Reproduce “Context Matters: A Strategy to Pre-train Language Model for Science Education” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhengliang Liu, Xinyu He, Lei Liu, Tianming Liu, Xiaoming Zhai
Researchers at the University of Georgia developed a strategy to continually pre-train language models using student-generated text, enhancing the accuracy of automated scoring for student science responses. Their SciEdBERT models achieved up to 0.866 accuracy on scientific argumentation tasks, outperforming models pre-trained on formal academic literature.
Key insights
- Contextualizing language models with actual student-generated text significantly improves performance for automated scoring of student science responses.
- Pre-training with formal academic science education articles did not yield performance improvements and sometimes lowered accuracy compared to baselines, suggesting a mismatch with student language.
- Combining an initial science-aware model (SciBERT) with further pre-training on specific student argumentation responses (SR2-SciBERT) proved to be the most effective strategy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.