vishpillai123 / w266_aes

Final Project for Automated Essay Scoring

W266 Automated Essay Scoring Final Project

Group Members: Sharon Wu, Candice Sener, Vish Pillai

Abstract:

Our project aims to develop AES by using BERT base and BERT-LSTM models in order to assess whether an RNN layer is needed for an effective two-stage learning framework. We found that our BERT base model is competitive to the original TSLF-1 (https://arxiv.org/pdf/1901.07744v2.pdf). On another note, our implementation of clique-based coherence based on BERT’s sentence pairing feature was ineffective in capturing coherence of longer passages. Overall, we found that TSLF-ALL demonstrated robustness with small sample sizes and adversarial samples.

Dataset: Kaggle ASAP Dataset by The Hewlett Foundation

GitHub Repo Organization:

Each folder includes a .ipynb notebook that corresponds to the name of the folder.

Folders:

• EDA
• Semantic Score
• Prompt-Relevance Score
• Coherence Score
• Second Stage