Multi-Fact Correction in Abstractive Text Summarization

TL;DR

This paper introduces Span-Fact, a set of models that improve the factual accuracy of neural abstractive summaries by correcting entities through span selection, enhancing consistency without losing summary quality.

Contribution

The paper presents novel factual correction models that leverage question answering knowledge to improve the factual consistency of abstractive summaries.

Findings

Significant improvement in factual consistency metrics.

Maintains summary quality as per automatic and human evaluations.

Effective span-based correction strategies for summarization.

Abstract

Pre-trained neural abstractive summarization systems have dominated extractive strategies on news summarization performance, at least in terms of ROUGE. However, system-generated abstractive summaries often face the pitfall of factual inconsistency: generating incorrect facts with respect to the source text. To address this challenge, we propose Span-Fact, a suite of two factual correction models that leverages knowledge learned from question answering models to make corrections in system-generated summaries via span selection. Our models employ single or multi-masking strategies to either iteratively or auto-regressively replace entities in order to ensure semantic consistency w.r.t. the source text, while retaining the syntactic structure of summaries generated by abstractive summarization models. Experiments show that our models significantly boost the factual consistency of system-generated summaries without sacrificing summary quality in terms of both automatic metrics and human evaluation.