Learning Recurrent Span Representations for Extractive Question Answering

TL;DR

This paper introduces a recurrent span representation model for extractive question answering, significantly improving accuracy on the SQuAD dataset by directly modeling answer spans.

Contribution

The paper presents a novel recurrent architecture for fixed-length span representations that enhances extractive QA performance over previous methods.

Findings

Improves SQuAD answer extraction accuracy by 5%.

Reduces baseline error by over 50%.

Demonstrates the effectiveness of explicit span modeling.

Abstract

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%.