Machine Comprehension using SQuAD and Deep Learning

There are many state-of-the-art deep learning architectures that have been used to develop reading comprehension systems using the Stanford Question Answering Dataset (SQuAD). In this work, we provide an overview of some of the state-ofthe-art models and compare their performance across several variations of encoder and decoder architectures and different attention mechanisms. We take inspiration from these state-of-the-art models and develop our own hybrid architectures, which have performed very similarly to previous models in the literature such as Match-LSTM and Dynamic Coattention when excluding extensive hyperparameter searches. Across all the architectures experimented with, our custom Stackson-Stacks (SoS) model yielded the best results of 0.537 F1 and 0.407 EM on the SQuAD test set.