Question Answering with Recurrent Span Representations

Using the Stanford Question and Answering Dataset (SQuAD) released by Rajpurkar et al. [1], we leverage a modified Recurrent Span Representation (RaSoR) model by Lee et al. [2] but with linear space complexity. This contrasts the original model by Lee et. al which requires a demanding quadratic space requirement to enumerate all answer spans. In addition, we introduce a novel architecture, Sequence Question-Induced Distances (SQuID) as an extension to our modified RaSoR implementation, which allows us to impose a prior on the answer span based on a question representation. Using our modified RaSoR implementation, we were still capable of obtaining good model accuracy, getting F1 and Exact Match (EM) scores of around 60% and 45% respectively on both the CodaLab dev and test set, with less memory and shorter time requirements. However, the results from our extension SQuID model are inconclusive; we believe that while imposing a question-induced prior on what the answer span may allow us to encode better answer distances, a generic question representation may not in itself be a very reliable indicator for the length of an answer span.