This repository contains code used to run experiments in the above paper.

The bash files run_crisp.sh trains using the best curricula and GRUs. By default we use the hyperparameters that gave us decent performance for a reasonable training time. They may be changed inside the file.

The bash files run_alt.sh trains using the best curricula and CNNs. Other models can be trained by changing the --model option (can choose between conv(CNNs), gpt(GPT), encoder(BERT)). By default we use the hyperparameters that gave us decent performance for a reasonable training time. They may be changed inside the file. The curriculum can be changed with --curriculum option (choose between l2r, r2l, c2n and n2c).

The required Python packages can be installed by using:

pip install -r requirements.txt