Spans, Not Tokens: A Span-Centric Model for Multi-Span Reading Comprehension

Anonymous

Abstract: Multi-span reading comprehension (MSRC) requires machines to extract multiple non-contiguous spans from a given context to answer a question. Existing MSRC methods either predict the positions of the start and end tokens of answer spans, or predict the BIO tag of each token. Such token-centric paradigms can hardly capture dependencies among spans which are critical to MSRC. In this paper, we propose a span-centric scheme where spans, as opposed to tokens, are directly represented and scored to qualify as answers. Thanks to the explicit representation of spans in the scheme, our implementation called SpanQualifier beneficially models intra-span and inter-span interactions. Our extensive experiments on three MSRC datasets demonstrate the effectiveness of our span-centric scheme and show that SpanQualifier achieves state-of-the-art results.

Paper Type: long

Research Area: Question Answering