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.

An in depth description of the work can be found here: QA-IR-report.pdf.

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:

• rnn: https://wandb.ai/veri/SQUAD/reports/RNN--Vmlldzo1Mzk2NTU
• transformers: https://wandb.ai/buio/SQUAD/reports/Transformers-Report--Vmlldzo1Mzk3MjE

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