SummaReranker: A Multi-Task Mixture-of-Experts Re-ranking Framework for Abstractive Summarization
This work addresses a key bottleneck in summarization for NLP researchers, offering a novel re-ranking approach that significantly boosts performance on multiple benchmarks.
The paper tackles the problem of suboptimal decoding in abstractive summarization by introducing SummaReranker, a second-stage re-ranking model that improves base model performance, achieving ROUGE score increases of up to 9.34% on datasets like Reddit TIFU and setting new state-of-the-art results.
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.