The reading comprehension task, that asks questions about a given evidence
document, is a central problem in natural language understanding. Recent
formulations of this task have typically focused on answer selection from a set of
candidates pre-defined manually or through the use of an external NLP pipeline.
However, Rajpurkar et al. (2016) recently released the SQUAD dataset in which the
answers can be arbitrary strings from the supplied text. In this paper, we focus on
this answer extraction task, presenting a novel model architecture that efficiently
builds fixed length representations of all spans in the evidence document with a
recurrent network. We show that scoring explicit span representations significantly
improves performance over other approaches that factor the prediction into separate
predictions about words or start and end markers. Our approach improves upon the
best published results of Wang & Jiang (2016) by 5% and decreases the error of
Rajpurkar et al.’s baseline by > 50%.