Salient Information Prompting to Steer Content in Prompt-based Abstractive Summarization

TL;DR

The paper addresses steering prompt-based abstractive summarization by injecting salient information drawn from the source document. It introduces SigExt, a Longformer-based keyphrase extractor, and GP-SigExt for cross-domain generalization, and demonstrates that phrase-level keyphrases in prompts improve ROUGE scores across multiple datasets and LLMs. The results reveal consistent ROUGE and recall gains, with notable improvements for zero-shot use and with supervised fine-tuning, along with a trade-off between precision and recall and model-dependent effects on factuality. The work provides practical insights for building prompt-based summarization systems, including a code release for SigExt.

Abstract

Large language models (LLMs) can generate fluent summaries across domains using prompting techniques, reducing the need to train models for summarization applications. However, crafting effective prompts that guide LLMs to generate summaries with the appropriate level of detail and writing style remains a challenge. In this paper, we explore the use of salient information extracted from the source document to enhance summarization prompts. We show that adding keyphrases in prompts can improve ROUGE F1 and recall, making the generated summaries more similar to the reference and more complete. The number of keyphrases can control the precision-recall trade-off. Furthermore, our analysis reveals that incorporating phrase-level salient information is superior to word- or sentence-level. However, the impact on hallucination is not universally positive across LLMs. To conduct this analysis, we introduce Keyphrase Signal Extractor (SigExt), a lightweight model that can be finetuned to extract salient keyphrases. By using SigExt, we achieve consistent ROUGE improvements across datasets and open-weight and proprietary LLMs without any LLM customization. Our findings provide insights into leveraging salient information in building prompt-based summarization systems. We release our code at \url{https://github.com/amazon-science/SigExt}

Figures (2)

• Figure 1: SigExt -- a finetuned Longformer to extract keyphrases from an article. We construct labels by thresholding the character-level fuzzy matching score between phrases in the article and the summary.
• Figure 2: Effect of using different number of keyphrases on the precision-recall trade off.