CIS700-BERT-RACE

Setup for Logistic Regresion (Non-Deep Baseline)

1. Download and unzip https://bit.ly/2GE1Frw into Logistic_Regression/Glove/
2. Unzip Data.zip in Logistic_Regression/
3. From Logistic_Regression, run python3 run.py

Setup for CNN (Deep Baseline #1)

To evaluate the model:

1. Download and unzip the data from https://bit.ly/2VvxDyp into CNN/
2. Download and unzip https://bit.ly/2GE1Frw into CNN/ .The files weight_matrix.pickle and word_to_idx.pickle should be directly under the CNN folder.
3. Download https://bit.ly/2GS5idL and put it in CNN/
4. Run cnn_eval.py

To train the model:

1. Download and unzip the data from https://bit.ly/2VvxDyp into CNN/
2. Download and unzip https://bit.ly/2GE1Frw into CNN/ .The files weight_matrix.pickle and word_to_idx.pickle should be directly under the CNN folder.
3. Run cnn.py

Setup for Feed Forward Net (Deep Baseline #2)

To train and evaluate the model:

1. Download and unzip data from https://bit.ly/2IMpAY5
2. Run feed_forward.py

Setup for GRU (Deep Baseline #3)

To evauluate the model:

1. Download and unzip data from https://bit.ly/2VrGPE8 into GRU/
2. Download the trained model https://bit.ly/2DyVB1t into GRU/
3. Run gru_eval.py

To train the model:

1. Download and unzip data from https://bit.ly/2VrGPE8 into GRU/
2. Run gru.py

Setup for BERT (Advanced Deep #1)

To evaluate the model:

1. Download and unzip the trained model from https://bit.ly/2XDOe0k to a folder called large_models/
2. run python3 evaluate_bert.py

To train the model

1. Run python3 race_bert.py it will run your model and save it in a folder called large_models/
2. Run evaluate_bert.py post that to run it in test set.

Setup for DCMN (Advanced Deep #2 / Novel Extra Credit)

To train and evaluate the model:

1. Download train, dev, and test data from https://bit.ly/2PF5v6Q.
2. Ensure pytorch-pretrained-bert is installed.
3. Run python DCMN.py where a GPU is installed and available.

Data Hypotheses

The folder Hypo contains all the hypothesis we tested, details regarding those are in the slides (Attention class.pdf), technical report and the blog post.