Skip to content

Our model implements two components, namely a prompt-independent and a prompt-dependent phase. The prompt-independent phase implements BERT, in a manner contrasting that of the authors’.

Notifications You must be signed in to change notification settings

AngadSethi/AutomatedEssayScoring

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

56 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Automated Essay Scoring with an Ensemble of Deep-Learning Models

We propose a two-stage model, encompassing two components, namely a prompt-independent and a prompt-dependent component. The prompt-independent phase implements BERT in a manner contrasting that of the authors’. We use BERT simply as an encoder, passing the tokenized essay into BERT and passing the output into feed-forward networks. In the prompt-independent phase, we are able to achieve quadratic weighted kappa (QWK) of TO-DO. For the prompt dependent phase, we experiment with two distinct neural networks: Bi-directional Attention Flow (BiDAF), and Hierarchical Attention Networks (HAN), and are able to achieve a QWK of TO-DO. We vary the hyperparameters, including the learning rates, sequence lengths, and tokenization methods, and show the increased robustness of our model. Towards the end of the paper, we ensemble our models and achieve a quadratic weighted kappa, the official metric, of TO-DO, which is 0.1 points less than the highest score of TO-DO.

Our major contributions in the development of this systems are:

  • Perused previous literature, handpicking best performing models
  • Developed a model with both prompt-independent and prompt-dependent components
  • Compared our model's performance against previously developed models
  • Incorporated modern encoders, like BERT, and modern concepts, like attention, into our models
  • Performed ablation studies to prove the efficacy of our models

Steps to re-create

  1. Create Conda environment
conda env create -n aes -f requirements.txt
  1. Activate environment
conda activate aes
  1. Begin training of model Since we utilise PyTorch Lightning, you can pass all arguments listed here. Additionally, there are a few generic args listed here and model specific args listed here and here. A typical train command looks something like this:
python train.py --model=original --bert_model=bert-base-uncased --batch_size=8 --max_seq_length=512

About

Our model implements two components, namely a prompt-independent and a prompt-dependent phase. The prompt-independent phase implements BERT, in a manner contrasting that of the authors’.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages