Quesiton Answering on SQuAD v1.1

This repository contains code for the Question Answering part of our project work for NLP2020 class by Paolo Torroni @unibo.
It implements 2 main architectures: a BiLSTM fully trained model and one based on BERT fine-tuning.

This folder contains:

compute_answer.py: given the question file, this script will download the best model and save the predictions
QA.ipynb: the notebook used for train and evaluate the best model

Experiment plot:

Branches

main: merged from the huggingface branch
rnn: baseline model based on RNN
rnn-regression: experiment with RNN + regression heads
huggingface: transformer-base models, comparation between BERT, ELECTRA, RoBERTa, Longformer
huggingface-regression: experiment with RoBERTa + regression head