SummaReranker: A Multi-Task Mixture-of-Experts Re-ranking Framework for Abstractive Summarization

TL;DR

SummaReranker introduces a multi-task mixture-of-experts re-ranking framework that enhances abstractive summarization by selecting superior candidates, significantly improving ROUGE scores across multiple datasets and setting new state-of-the-art results.

Contribution

The paper proposes a novel second-stage re-ranking model using mixture-of-experts that improves summarization quality beyond standard fine-tuning methods.

Findings

ROUGE scores improved by 5.44% on CNN-DailyMail

ROUGE scores improved by 1.31% on XSum

ROUGE scores improved by 9.34% on Reddit TIFU

Abstract

Sequence-to-sequence neural networks have recently achieved great success in abstractive summarization, especially through fine-tuning large pre-trained language models on the downstream dataset. These models are typically decoded with beam search to generate a unique summary. However, the search space is very large, and with the exposure bias, such decoding is not optimal. In this paper, we show that it is possible to directly train a second-stage model performing re-ranking on a set of summary candidates. Our mixture-of-experts SummaReranker learns to select a better candidate and consistently improves the performance of the base model. With a base PEGASUS, we push ROUGE scores by 5.44% on CNN-DailyMail (47.16 ROUGE-1), 1.31% on XSum (48.12 ROUGE-1) and 9.34% on Reddit TIFU (29.83 ROUGE-1), reaching a new state-of-the-art. Our code and checkpoints will be available at https://github.com/ntunlp/SummaReranker.