Paying Attention to SQuAD: Exploring Bidirectional Attention Flow

With the goal of automated reading comprehension, we apply a neural network with Bidirectional Attention Flow (BiDAF) to the Stanford Question Answering Dataset (SQuAD) and achieve F1 and Exact Match (EM) scores close to the original paper with a single model. We obtain a test F1 score of 76.037 and test EM score of 66.663. Our model includes Character-level CNN embeddings, a Highway Network layer, a Phrase Embedding layer, a Modeling layer, and smart span selection. We also explored expanding the model with feature engineering and an Answer Pointer output layer, which did not further improve our best model. We analyze our model’s performance across categories of contexts, questions, and answers, and compare baseline attention with BiDAF.