Reproduce “The effects of data size on Automated Essay Scoring engines” 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

Christopher Ormerod, Amir Jafari, Susan Lottridge, Milan Patel, Amy Harris, Paul van Wamelen

This research investigates how training data size and quality impact Automated Essay Scoring (AES) engines, comparing traditional feature-based (BOW), LSTM, and BERT models. It finds that large quantities of single-scored data, traditionally disregarded, can significantly improve neural network performance, particularly for LSTMs, when used for pre-initialization before fine-tuning with higher-quality double-scored data.

Key insights

  • LSTM models exhibit strong data dependency, requiring significant quantities of single-scored data to achieve stable and competitive performance.
  • BERT models, benefiting from pretraining, perform robustly even with limited data, consistently improving with more data and demonstrating particular strength in assessing essay conventions.
  • Single-scored data, traditionally undervalued, can be effectively leveraged through transfer learning to pre-initialize neural networks, leading to performance improvements when subsequently fine-tuned on high-quality double-scored data.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.