All development and testing has been done in Conda Python 3 environments on Linux x86-64 and Windows 10 systems, specifically Python 3.6.x and 3.7.x. To create environment with all dependencies in the root directory run
conda env create -f environment.yml
conda activate pipeline
To download train and test SQuAD 2.0 data run script download_dataset.sh. It loads JSON files into ./data directory
- EDA.ipynb - explaratory data analysis on the data. Here we have tried to see how the data look like and what we are dealing with.
- Baseline.ipynb - simple baseline model that was built to make some conclusions about the complexity of the task and about possbile solution that might work
- Simple Albert - notebook where we have trained just ALBERT to see compare it with next two models.
- ALBERT + Autoencoder.ipynb - notebook where the hypothesis about using Autoencoders with ALBERT was tested
- ALBERT + BiLSTMEncoder + ALBERT-SQuAD-OUT.ipynb - notebook where the hypothesis about using BiLSTMEncoder with ALBERT features was tested.
In the retro_mrc directory you will find the code for retrospective reader and how to run it.