Dynamic Coattention Networks for Reading Comprehension

We investigate the reading comprehension task using SQaAD dataset. The reading comprehension task is formulated as finding the answer span in the given paragraph for the given question. We describe multiple approaches to solving this problem starting from very simple baseline that gradually evolves into a much more complicated architecture. For each model we describe the underlying architecture and analyze its performance on the training and validation data. Our final model is a modification of the dynamic Coattention network described in the work of Socher et al. [1] and it achieves F1 = 58.6% and EM = 44.6% scores on the test set.